CN116361559A

CN116361559A - User resource anti-fraud strategy generation method and server adopting artificial intelligence

Info

Publication number: CN116361559A
Application number: CN202310362471.6A
Authority: CN
Inventors: 王星
Original assignee: Jiuquan Transfinite Information Technology Co ltd
Current assignee: Aileyun Shenzhen Technology Co ltd
Priority date: 2023-04-07
Filing date: 2023-04-07
Publication date: 2023-06-30
Anticipated expiration: 2043-04-07
Also published as: CN116361559B

Abstract

The invention provides a user resource anti-fraud strategy generation method and a server adopting artificial intelligence, which can not only obtain target user resource description knowledge recording more resource details, but also improve the anti-fraud strategy matching precision aiming at the user resource information to be processed, so as to provide anti-fraud guidance for the user corresponding to the user resource information to be processed through a final anti-fraud strategy text, and avoid the loss of the user as much as possible; the reference linkage anti-fraud strategy text can be matched in an auxiliary mode through the target knowledge segments, so that modification of the reference strategy text vector in the reference anti-fraud strategy pool can be avoided under the condition that the reference anti-fraud strategy pool is provided, and additional processing expenditure is avoided. Thus, the problem that the conventional technology is difficult to accurately match the anti-fraud strategy for the user resource information to be processed can be solved.

Description

User resource anti-fraud strategy generation method and server adopting artificial intelligence

Technical Field

The invention relates to the technical field of artificial intelligence and big data anti-fraud, in particular to a user resource anti-fraud strategy generation method and a server adopting the artificial intelligence.

Background

The new generation of information technology (big data, cloud computing, artificial intelligence and the like) accelerates the energized industry and the stimulated society to change, and simultaneously, as the feeler of technical innovation is continuously explored, the roles played by the new generation of information technology are more diverse, the roles played by the new generation of information technology are more huge, and the call from public interests is continuously emerging. In recent years, with the continuous development of networking technology, the "recruitment" of fraud by the network is also being upgraded and coded, which is not well-controlled. For this reason, it is imperative to introduce artificial intelligence and big data mining for anti-fraud processing.

Disclosure of Invention

The invention provides a method and a server for generating a user resource anti-fraud strategy by adopting artificial intelligence, and the invention adopts the following technical scheme for realizing the technical purposes.

The first aspect is a user resource anti-fraud policy generation method adopting artificial intelligence, applied to an AI anti-fraud server, the method comprising:

obtaining user resource information to be processed, and generating target user resource description knowledge corresponding to the user resource information to be processed through a target anti-fraud analysis network; the target anti-fraud analysis network is obtained by debugging according to a reference anti-fraud analysis network for completing debugging; the target user resource description knowledge has a target knowledge segment which has knowledge commonality relation with the reference user resource description knowledge, the scale variable corresponding to the target user resource description knowledge is larger than the scale variable corresponding to the reference user resource description knowledge, and the reference user resource description knowledge is description knowledge corresponding to the user resource information to be processed, which is output through the reference anti-fraud analysis network;

Obtaining target linkage anti-fraud strategy text which has correlation with the target user resource description knowledge from the target anti-fraud strategy pool according to a target strategy text vector corresponding to the target anti-fraud strategy text in the target anti-fraud strategy pool; the target strategy text vector is descriptive knowledge corresponding to the target anti-fraud strategy text output through the target anti-fraud analysis network;

obtaining a reference linkage anti-fraud strategy text which has relevance with the target knowledge segment from the reference anti-fraud strategy pool according to a reference strategy text vector corresponding to the reference anti-fraud strategy text in the reference anti-fraud strategy pool; the reference strategy text vector is descriptive knowledge corresponding to the reference anti-fraud strategy text output through the reference anti-fraud analysis network;

a final anti-fraud policy text for matching the pending user resource information is determined in the reference linked anti-fraud policy text and the target linked anti-fraud policy text.

In some optional embodiments, the generating, by the target anti-fraud analysis network, target user resource description knowledge corresponding to the user resource information to be processed includes:

Loading the user resource information to be processed into the target anti-fraud analysis network; the target anti-fraud analysis network comprises a description knowledge mining unit, a description knowledge downsampling unit and a target description knowledge aggregation unit; the knowledge aggregation level number corresponding to the target description knowledge aggregation unit is larger than the knowledge aggregation level number corresponding to the reference description knowledge aggregation unit in the reference anti-fraud analysis network;

performing description knowledge mining processing on the user resource information to be processed through the description knowledge mining unit, generating initial user resource description knowledge corresponding to the user resource information to be processed, and loading the initial user resource description knowledge into the description knowledge downsampling unit;

performing description knowledge downsampling operation on the initial user resource description knowledge through the description knowledge downsampling unit, generating user resource description knowledge to be aggregated corresponding to the user resource information to be processed, and loading the user resource description knowledge to be aggregated into the target description knowledge aggregation unit;

and carrying out description knowledge aggregation operation on the user resource description knowledge to be aggregated through the target description knowledge aggregation unit to generate the target user resource description knowledge.

In some alternative embodiments, the method further comprises:

obtaining the target anti-fraud strategy text and obtaining a text theme corresponding to the target anti-fraud strategy text;

loading the target anti-fraud strategy text into the target anti-fraud analysis network, and performing text word vector extraction on the target anti-fraud strategy text through the target anti-fraud analysis network to generate the target strategy text vector;

and taking a text theme corresponding to the target anti-fraud strategy text as a request characteristic, taking the target strategy text vector as a response characteristic, generating a request response relation chain by combining the request characteristic and the response characteristic, and recording the request response relation chain in the target anti-fraud strategy pool.

In some alternative embodiments, the obtaining, from the target anti-fraud policy pool, target linked anti-fraud policy text having relevance to the target user resource description knowledge according to target policy text vectors corresponding to target anti-fraud policy text in the target anti-fraud policy pool includes:

obtaining a request response relation chain in the target anti-fraud policy pool; the request response relation chain comprises request features and response features, wherein the request features are generated by text topics corresponding to the target anti-fraud strategy text, and the response features are generated by the target strategy text vector;

Acquiring cosine similarity between the target user resource description knowledge and the response characteristics, and determining the response characteristics with the cosine similarity larger than or equal to a cosine similarity limit value as similar response characteristics;

and determining the request features corresponding to the similar response features as similar request features, and acquiring the target linkage anti-fraud strategy text from the target anti-fraud strategy pool by combining the similar request features.

In some optional embodiments, the target anti-fraud policy text comprises at least two target anti-fraud policy texts, and the target policy text vector comprises target policy text vectors respectively corresponding to the at least two target anti-fraud policy texts; the method further comprises the steps of:

obtaining the at least two target anti-fraud policy texts, and respectively loading the at least two target anti-fraud policy texts into the target anti-fraud analysis network;

respectively extracting text word vectors of the at least two target anti-fraud strategy texts through the target anti-fraud analysis network to generate target strategy text vectors respectively corresponding to the at least two target anti-fraud strategy texts;

grouping operation is carried out on at least two target strategy text vectors, and Q target strategy description grouping results and target grouping guide vectors respectively corresponding to the Q target strategy description grouping results are generated; q is a positive integer greater than 1, and Q is not greater than the total number of the at least two target policy text vectors;

And storing the Q target strategy description grouping results and Q target grouping guide vector associations in the target anti-fraud strategy pool.

In some optional embodiments, the Q target policy description grouping results include a target policy description grouping result cluster_q, Q being a positive integer and Q being no greater than Q; the storing the Q target policy description grouping results and Q target grouping guide vector associations in the target anti-fraud policy pool includes:

obtaining a target strategy text vector vec_w from the target strategy description grouping result cluster_q, wherein w is a positive integer, and w is not more than the total number of target strategy text vectors in the target strategy description grouping result cluster_q;

obtaining a target anti-fraud strategy text_w corresponding to the target strategy text vector vec_w, and obtaining a text theme corresponding to the target anti-fraud strategy text_w;

taking a text theme corresponding to the target anti-fraud strategy text text_w as a request feature, taking the target strategy text vector vec_w as a response feature, and combining the request feature and the response feature to generate a request response relation chain which has correlation with the target strategy description grouping result cluster_q;

And storing each target grouping guide vector and a request response relation chain corresponding to each target grouping guide vector in the target anti-fraud strategy pool in a correlated manner.

obtaining the Q target grouping guide vectors in the target anti-fraud policy pool;

the Q target grouping guide vectors comprise a target grouping guide vector guidance_e, e is a positive integer, and e is not more than Q;

obtaining the cosine similarity S_e between the target user resource description knowledge and the target grouping guide vector guidance_e;

sequentially adjusting cosine similarities corresponding to the Q target grouping guide vectors respectively, and determining target cosine similarities from the cosine similarities subjected to sequential adjustment;

determining a target grouping guide vector corresponding to the target cosine similarity as a target similarity grouping guide vector, and obtaining a similar request response relation chain which is connected with the target similarity grouping guide vector;

And combining the similar request response relation chain to obtain the target linkage anti-fraud strategy text which has correlation with the target user resource description knowledge.

In some optional embodiments, the determining a final anti-fraud policy text for matching the pending user resource information in the reference linked anti-fraud policy text and the target linked anti-fraud policy text comprises:

determining the reference linked anti-fraud policy text and the target linked anti-fraud policy text as alternative anti-fraud policy texts; the number of the candidate anti-fraud strategy texts is at least two;

obtaining cosine similarity between similar strategy text vectors of each alternative anti-fraud strategy text and the target user resource description knowledge respectively;

sequentially adjusting the at least two candidate anti-fraud strategy texts by combining cosine similarity corresponding to each candidate anti-fraud strategy text;

from at least two alternative anti-fraud policy texts completing the order adjustment, the final anti-fraud policy text for matching the pending user resource information is determined.

The second aspect is an AI anti-fraud server comprising a memory and a processor; the memory is coupled to the processor; the memory is used for storing computer program codes, and the computer program codes comprise computer instructions; wherein the processor, when executing the computer instructions, causes the AI anti-fraud server to perform the method of the first aspect.

A third aspect is a computer readable storage medium having stored thereon a computer program which, when run, performs the method of the first aspect.

In the embodiment of the invention, the AI anti-fraud server can generate target user resource description knowledge corresponding to the user resource information to be processed through a target anti-fraud analysis network, wherein the target anti-fraud analysis network is obtained by debugging according to a reference anti-fraud analysis network which completes debugging, the network performance of the target anti-fraud analysis network can comprise the network performance of the reference anti-fraud analysis network, and the quality score corresponding to the network performance of the target anti-fraud analysis network can be superior to the quality score of the network performance of the reference anti-fraud analysis network; the resource details recorded by the target user resource description knowledge can be known to be better than the resource details recorded by the reference user resource description knowledge by the fact that the scale variable corresponding to the target user resource description knowledge is larger than the scale variable corresponding to the reference user resource description knowledge; through the existence of the target knowledge segments with knowledge commonalities with the reference user resource description knowledge in the target user resource description knowledge, the target knowledge segments can replace the reference user resource description knowledge to perform anti-fraud strategy matching; further, according to the target policy text vector, the AI anti-fraud server can obtain target linkage anti-fraud policy text which has correlation with target user resource description knowledge from a target anti-fraud policy pool; in addition, according to the reference strategy text vector, a reference linkage anti-fraud strategy text which has relevance with the target knowledge segment can be obtained from the reference anti-fraud strategy pool; further, in the reference linkage anti-fraud policy text and the target linkage anti-fraud policy text, a final anti-fraud policy text for matching the user resource information to be processed may be determined.

Therefore, the embodiment of the invention not only can acquire the target user resource description knowledge recording more resource details, but also can improve the anti-fraud strategy matching precision aiming at the resource information of the user to be processed, so as to provide anti-fraud guidance for the user corresponding to the resource information of the user to be processed through the final anti-fraud strategy text, thereby avoiding the loss of the user as much as possible; the reference linkage anti-fraud strategy text can be matched in an auxiliary mode through the target knowledge segments, so that modification of the reference strategy text vector in the reference anti-fraud strategy pool can be avoided under the condition that the reference anti-fraud strategy pool is provided, and additional processing expenditure is avoided. Thus, the problem that the conventional technology is difficult to accurately match the anti-fraud strategy for the user resource information to be processed can be solved.

Drawings

FIG. 1 is a flow chart of a method for generating user resource anti-fraud policies using artificial intelligence according to an embodiment of the present invention.

Detailed Description

Hereinafter, the terms "first," "second," and "third," etc. are used for descriptive purposes only and are not to be construed as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defining "a first", "a second", or "a third", etc., may explicitly or implicitly include one or more such feature.

FIG. 1 shows a flow diagram of a user resource anti-fraud policy generation method employing artificial intelligence, which may be implemented by an AI anti-fraud server, which may include a memory and a processor, provided by an embodiment of the present invention; the memory is coupled to the processor; the memory is used for storing computer program codes, and the computer program codes comprise computer instructions; wherein the processor, when executing the computer instructions, causes the AI anti-fraud server to execute STEP101-STEP104.

STEP101 obtains the user resource information to be processed, and generates target user resource description knowledge corresponding to the user resource information to be processed through a target anti-fraud analysis network.

In the embodiment of the invention, the target anti-fraud analysis network is obtained by debugging according to a reference anti-fraud analysis network for completing debugging. The target user resource description knowledge has a target knowledge segment which is in knowledge commonality connection with the reference user resource description knowledge, the scale variable corresponding to the target user resource description knowledge is larger than the scale variable corresponding to the reference user resource description knowledge, and the reference user resource description knowledge is description knowledge corresponding to the user resource information to be processed, which is output through the reference anti-fraud analysis network.

Further, for the anti-fraud analysis network with the debugging completed, adaptive optimization or upgrading can be performed to obtain the anti-fraud analysis network with better performance. Thus, in the embodiment of the present invention, the reference anti-fraud analysis network may be a historical anti-fraud analysis network, and the target anti-fraud analysis network may be an anti-fraud analysis network with better performance obtained after optimizing/upgrading the historical anti-fraud analysis network (i.e., the reference anti-fraud analysis network).

Further, the scale variable can be understood as a feature dimension, and for a neural network, if the dimension of the feature (the above description knowledge) reflecting the resource detail is larger, the detail output quality is better, and the required processing overhead is relatively more.

According to the embodiment of the invention, the AI anti-fraud server obtains the user resource information 001 to be processed according to the optimization/upgrading and knowledge derivative development description, and loads the user resource information 001 to be processed into the target anti-fraud analysis network 002, and generates target user resource description knowledge 003 corresponding to the user resource information 001 to be processed through the target anti-fraud analysis network 002. The target anti-fraud analysis network 002 is obtained by debugging according to a reference anti-fraud analysis network 004 for completing debugging, in the embodiment of the present invention, the neural network for identifying and processing the user resource information is called an anti-fraud analysis network, the target anti-fraud analysis network 002 and the reference anti-fraud analysis network 004 belong to the anti-fraud analysis network, and the network configurations respectively corresponding to the target anti-fraud analysis network 002 and the reference anti-fraud analysis network 004 are consistent, which is different in that the knowledge aggregation level number of the target description knowledge aggregation units contained in the target anti-fraud analysis network 002 is greater than the knowledge aggregation level number of the description knowledge aggregation units contained in the reference anti-fraud analysis network 004, so that the scale variables corresponding to the target user resource characteristics (including the target user resource description knowledge 003 and the target policy text vector) obtained through the target anti-fraud analysis network 002 are greater than the scale variables corresponding to the reference user resource characteristics (including the reference user resource description and the reference policy text vector) obtained through the reference anti-fraud analysis network 004.

Those skilled in the art will appreciate that the present embodiment does not limit the network types of the anti-fraud analysis network (including the target anti-fraud analysis network 002 and the reference anti-fraud analysis network 004), and may be any neural network capable of performing user resource information analysis processing, and the present embodiment does not limit the network configuration of the anti-fraud analysis network.

In the embodiment of the invention, the knowledge aggregation level included in the target anti-fraud analysis network 002 is disassembled into the reference knowledge aggregation level and the target knowledge aggregation level, wherein the reference knowledge aggregation level corresponds to the (global) knowledge aggregation level (the channel for connecting the features) in the reference anti-fraud analysis network 004, and since the target anti-fraud analysis network 002 is obtained by debugging according to the reference anti-fraud analysis network 004, the target knowledge segments 005 (local features) generated by the reference knowledge aggregation level have knowledge commonalities (similar association relations) with the reference user resource description knowledge, namely cosine similarity corresponding to the two features is higher. In the embodiment of the invention, the knowledge aggregation level for scale variable derivation (expansion) is called a target knowledge aggregation level, that is, the dimension difference between the scale variable corresponding to the target user resource feature and the scale variable corresponding to the reference user resource feature is equal to the number of target knowledge aggregation levels. The embodiment of the invention does not limit the reference knowledge aggregation level number and the target knowledge aggregation level number.

Those skilled in the art will appreciate that the embodiment of the present invention does not set the hierarchical priorities of the reference knowledge aggregation level and the target knowledge aggregation level, and only needs that the reference knowledge aggregation level be an uninterrupted feature channel.

Illustratively, the embodiment of the present invention sets the knowledge aggregation level number of the target description knowledge aggregation unit of the target anti-fraud analysis network 002 to 32, and sets the knowledge aggregation level number of the description knowledge aggregation unit of the reference anti-fraud analysis network 004 to 16. Further, the first 16 levels of the target description knowledge aggregation units of the target anti-fraud analysis network 002 may be set to a reference knowledge aggregation level corresponding to the description knowledge aggregation units of the reference anti-fraud analysis network 004, at this time, the scale variable of the target knowledge segment 005 is 16-dimensional, the scale variable of the target user resource description knowledge 003 is 32-dimensional, and the target knowledge segment 005 is the first 16-dimensional description knowledge of the target user resource description 003, and at the same time, the last 16-dimensional description knowledge in the target user resource description 003 is the description knowledge generated by the target knowledge aggregation level.

Further, the target user resource description knowledge 003 is a high-order feature for characterizing the user resource information 001 to be processed, and the reference user resource description knowledge is a low-order feature for characterizing the user resource information 001 to be processed, so that the resource detail output quality of the target user resource description knowledge 003 is better than the resource detail output quality of the reference user resource description knowledge.

STEP102, obtaining target linkage anti-fraud strategy text which has correlation with target user resource description knowledge from the target anti-fraud strategy pool according to the target strategy text vector corresponding to the target anti-fraud strategy text in the target anti-fraud strategy pool.

The target strategy text vector is descriptive knowledge corresponding to target anti-fraud strategy text output through the target anti-fraud analysis network.

Illustratively, the target anti-fraud policy pool in embodiments of the present invention may understand a database for storing anti-fraud policy texts that may provide anti-fraud guidance and information protection guidance for users corresponding to the user resource information.

Further, after generating the target user resource description knowledge 003, the AI anti-fraud server may perform a commonality analysis on the target user resource description knowledge 003 and the target policy text vector to obtain a commonality score (similarity) corresponding to the two features. It can be understood that the larger the commonality score between the target user resource description knowledge 003 and the target policy text vector is, the larger the cosine similarity between the target user resource description knowledge 003 and the target policy text vector is, so that it can be explained that the text resource similarity between the user resource information to be processed and the target anti-fraud policy text is high, that is, the user resource information to be processed and the target anti-fraud policy text can be text information matched with each other; otherwise, the smaller the commonality score between the target user resource description knowledge 003 and the target policy text vector is, the smaller the cosine similarity between the target user resource description knowledge 003 and the target policy text vector is, and further, the lower the text resource similarity between the to-be-processed user resource information and the target anti-fraud policy text can be indicated, namely, the to-be-processed user resource information and the target anti-fraud policy text are unmatched text information. It should be noted that whether the user resource information to be processed is matched with the target anti-fraud policy text can be understood as whether the target anti-fraud policy text can provide anti-fraud guidance for the user resource information to be processed, that is, whether the target anti-fraud policy text is adapted to the user resource information to be processed.

Further, the AI anti-fraud server may determine the target anti-fraud policy text with cosine similarity greater than or equal to the cosine similarity limit as the target linkage anti-fraud policy text, e.g., determine the target policy text vector with cosine similarity greater than or equal to 0.8 (cosine similarity limit) from the target anti-fraud policy pool 006 as the target linkage anti-fraud policy text; the target linkage anti-fraud policy texts which match the preset number can also be obtained from the target anti-fraud policy texts in the descending order of cosine similarity, for example, the target anti-fraud policy texts with the cosine similarity of 10 being determined from the target anti-fraud policy pool 006 are used as the target linkage anti-fraud policy texts.

STEP103, obtaining a reference linkage anti-fraud strategy text with correlation with the target knowledge piece from the reference anti-fraud strategy pool according to a reference strategy text vector corresponding to the reference anti-fraud strategy text in the reference anti-fraud strategy pool; the reference policy text vector is descriptive knowledge corresponding to the reference anti-fraud policy text output by the reference anti-fraud analysis network.

After the target user resource description knowledge 003 is generated, the AI anti-fraud server can obtain a target knowledge segment 005 which has knowledge commonality relation with the reference user resource description knowledge, and further, the commonality analysis can be carried out on the target knowledge segment 005 and the reference strategy text vector to obtain the commonality scores corresponding to the two features. It can be understood that the larger the commonality score between the target knowledge segment 005 and the reference strategy text vector is, the larger the cosine similarity between the target knowledge segment 005 and the reference strategy text vector is, so that it can be explained that the similarity of text resources between the user resource information to be processed and the reference anti-fraud strategy text is high, that is, the user resource information to be processed and the reference anti-fraud strategy text can be text information matched with each other; otherwise, the smaller the commonality score between the target knowledge segment 005 and the reference strategy text vector is, the smaller the cosine similarity between the target knowledge segment 005 and the reference strategy text vector is, and further, the fact that the text resource similarity between the user resource information to be processed and the reference anti-fraud strategy text is low can be indicated, namely, the user resource information to be processed and the reference anti-fraud strategy text are unmatched text information.

Further, the AI anti-fraud server may determine a reference anti-fraud policy text having a cosine similarity greater than or equal to a cosine similarity limit as a reference linkage anti-fraud policy text, e.g., determine a reference policy text vector having a cosine similarity greater than or equal to 0.7 (cosine similarity limit) from the reference anti-fraud policy pool 007 as a reference linkage anti-fraud policy text. The reference linked anti-fraud policy texts that match the predetermined number may also be obtained from the reference anti-fraud policy texts in descending order of cosine similarity, for example, 20 reference anti-fraud policy texts with the front cosine similarity position determined from the reference anti-fraud policy pool 007 are reference linked anti-fraud policy texts.

It will be appreciated that embodiments of the present invention do not limit the manner in which the reference linked anti-fraud policy text is matched from the reference anti-fraud policy pool 007 nor the manner in which cosine similarity between the target knowledge segments 005 and the reference policy text vectors.

STEP104, in the reference linked anti-fraud policy text and the target linked anti-fraud policy text, determines a final anti-fraud policy text for matching the user resource information to be processed.

By way of example, the embodiment of the invention adopts two anti-fraud policy pools (including a reference anti-fraud policy pool and a target anti-fraud policy pool) to match final anti-fraud policy texts for the user resource information to be processed, so after obtaining the matching results (including the target linkage anti-fraud policy texts and the reference linkage anti-fraud policy texts) returned by the two matching threads, the matching results can be combined first, the cosine similarity corresponding to the matching results respectively is used for sorting all the returned matching results according to the cosine similarity descending order, and the final anti-fraud policy texts aiming at the user resource information to be processed are determined according to the sorting results.

Optionally, the AI anti-fraud server may determine, based on the similarity of the text resources, a final anti-fraud policy text for matching the user resource information to be processed according to the binding weights (screening weights) corresponding to the reference linkage anti-fraud policy text and the target linkage anti-fraud policy text, for example, the AI anti-fraud server preferentially binds the target linkage anti-fraud policy text, and may set the binding weight corresponding to the target linkage anti-fraud policy text to be a variable in1 smaller than 0.5 and set the binding weight corresponding to the reference linkage anti-fraud policy text to be a variable in2 larger than 0.5, where the AI anti-fraud server determines, based on the similarity of the text resources and the binding weights, the final anti-fraud policy text for matching the user resource information to be processed in the reference linkage anti-fraud policy text and the target linkage anti-fraud policy text. For example, the AI anti-fraud server weights the text resource similarity corresponding to the reference linkage anti-fraud strategy text and the binding weight corresponding to the reference linkage anti-fraud strategy text to obtain a reference matching result for the reference linkage anti-fraud strategy text, weights the text resource similarity corresponding to the target linkage anti-fraud strategy text and the binding weight corresponding to the target linkage anti-fraud strategy text to obtain a target matching result for the target linkage anti-fraud strategy text, compares the target matching result with the reference matching result, binds the target linkage anti-fraud strategy text if the target matching result is smaller than the reference matching result, and binds the reference linkage anti-fraud strategy text if the target matching result is equal to or larger than the reference matching result.

For example, if the target linked anti-fraud policy text is preferentially bound, the AI anti-fraud server may set a binding weight corresponding to the target linked anti-fraud policy text to a coefficient in3 smaller than 1, and set a binding weight corresponding to the reference linked anti-fraud policy text to a coefficient in4 larger than 1, at which time the AI anti-fraud server determines a final anti-fraud policy text for matching the user resource information to be processed in the reference linked anti-fraud policy text and the target linked anti-fraud policy text according to the text resource similarity and the binding weight. For example, the AI anti-fraud server multiplies the text resource similarity corresponding to the reference linkage anti-fraud policy text by the binding weight corresponding to the reference linkage anti-fraud policy text to obtain a reference matching result for the reference linkage anti-fraud policy text, multiplies the text resource similarity corresponding to the target linkage anti-fraud policy text by the binding weight corresponding to the target linkage anti-fraud policy text to obtain a target matching result for the target linkage anti-fraud policy text, and further, the AI anti-fraud server compares the target matching result with the reference matching result, binds the target linkage anti-fraud policy text if the target matching result is smaller than the reference matching result, and binds the reference linkage anti-fraud policy text if the target matching result is equal to or larger than the reference matching result. Optionally, the AI anti-fraud server may further weight the text resource similarity corresponding to the reference linkage anti-fraud policy text and the binding weight corresponding to the reference linkage anti-fraud policy text to obtain a reference matching result for the reference linkage anti-fraud policy text, weight the text resource similarity corresponding to the target linkage anti-fraud policy text and the binding weight corresponding to the target linkage anti-fraud policy text to obtain a target matching result for the target linkage anti-fraud policy text, and then compare the reference matching result with the target matching result, where the subsequent process is similar to the above.

In the embodiment of the invention, the AI anti-fraud server can generate target user resource description knowledge corresponding to the user resource information to be processed through a target anti-fraud analysis network, wherein the target anti-fraud analysis network is obtained by debugging according to a reference anti-fraud analysis network which completes debugging, the network performance of the target anti-fraud analysis network can comprise the network performance of the reference anti-fraud analysis network, and the quality score corresponding to the network performance of the target anti-fraud analysis network can be superior to the quality score of the network performance of the reference anti-fraud analysis network; the resource details recorded by the target user resource description knowledge can be known to be better than the resource details recorded by the reference user resource description knowledge by the fact that the scale variable corresponding to the target user resource description knowledge is larger than the scale variable corresponding to the reference user resource description knowledge; through the existence of the target knowledge segments with knowledge commonalities with the reference user resource description knowledge in the target user resource description knowledge, the target knowledge segments can replace the reference user resource description knowledge to perform anti-fraud strategy matching; further, according to the target policy text vector, the AI anti-fraud server can obtain target linkage anti-fraud policy text which has correlation with target user resource description knowledge from a target anti-fraud policy pool; in addition, according to the reference strategy text vector, a reference linkage anti-fraud strategy text which has relevance with the target knowledge segment can be obtained from the reference anti-fraud strategy pool; further, in the reference linkage anti-fraud policy text and the target linkage anti-fraud policy text, a final anti-fraud policy text for matching the user resource information to be processed may be determined. Therefore, the embodiment of the invention not only can acquire the target user resource description knowledge recording more resource details, but also can improve the anti-fraud strategy matching precision aiming at the user resource information to be processed; the reference linkage anti-fraud strategy text can be matched in an auxiliary mode through the target knowledge segments, so that modification of the reference strategy text vector in the reference anti-fraud strategy pool can be avoided under the condition that the reference anti-fraud strategy pool is provided, and additional processing expenditure is avoided.

The other method for generating the user resource anti-fraud strategy by adopting the artificial intelligence provided by the embodiment of the invention at least comprises the following steps.

STEP201, obtaining the target anti-fraud policy text and obtaining the text subject corresponding to the target anti-fraud policy text.

For example, the AI anti-fraud server needs to first establish a policy matching relationship for the user resource information to achieve anti-fraud policy text matching of the user resource information. For example, the AI anti-fraud server obtains a target anti-fraud policy text, and further obtains a text topic (text label) corresponding to the target anti-fraud policy text. It is understood that the text topic may be any information that can be used to distinguish the target anti-fraud policy text, such as a string tag.

STEP202 loads the target anti-fraud policy text into the target anti-fraud analysis network, and performs text word vector extraction on the target anti-fraud policy text through the target anti-fraud analysis network to generate a target policy text vector.

For example, the exemplary idea of STEP202 may be seen from the description below in STEP204, which differs only in the user resource information loaded into the target anti-fraud analysis network, and the target anti-fraud policy text is loaded into the target anti-fraud analysis network in this STEP, so that a target policy text vector corresponding to the target anti-fraud policy text is generated in this STEP. STEP204 loads the user resource information to be processed into the target anti-fraud analysis network, thereby generating target user resource description knowledge corresponding to the user resource information to be processed, which is another 32-dimensional feature vector.

STEP203, taking the text subject corresponding to the target anti-fraud policy text as a request feature, taking the target policy text vector as a response feature, generating a request response relation chain according to the request feature and the response feature, and recording the request response relation chain in the target anti-fraud policy pool.

Illustratively, the AI anti-fraud server takes the text subject corresponding to the target anti-fraud policy text as a request feature (key feature) and takes the target policy text vector as a response feature (value feature), so that a request response feature tuple (which can also be understood as a key value pair) comprising the request feature and the response feature can be generated, and further, the generated request response feature tuple can be stored into the request response relation chain in the target anti-fraud policy pool. It can be appreciated that the embodiment of the invention only takes the target anti-fraud policy text as an example for the process of establishing the policy matching relationship for the user resource information.

The embodiment of the invention does not modify the reference policy text vector in the reference anti-fraud policy pool.

STEP204, obtaining user resource information to be processed, and generating target user resource description knowledge corresponding to the user resource information to be processed through a target anti-fraud analysis network; the target anti-fraud analysis network is obtained by debugging according to a reference anti-fraud analysis network for completing debugging; the target user resource description knowledge has a target knowledge segment which is in knowledge commonality connection with the reference user resource description knowledge, the scale variable corresponding to the target user resource description knowledge is larger than the scale variable corresponding to the reference user resource description knowledge, and the reference user resource description knowledge is description knowledge corresponding to the user resource information to be processed, which is output through the reference anti-fraud analysis network.

Illustratively, the AI anti-fraud server loads the user resource information to be processed into a target anti-fraud analysis network; the target anti-fraud analysis network comprises a description knowledge mining unit, a description knowledge downsampling unit and a target description knowledge aggregation unit; the knowledge aggregation level number corresponding to the target description knowledge aggregation unit is larger than that corresponding to the reference description knowledge aggregation unit in the reference anti-fraud analysis network; performing description knowledge mining processing on the user resource information to be processed through a description knowledge mining unit, generating initial user resource description knowledge corresponding to the user resource information to be processed, and loading the initial user resource description knowledge into a description knowledge downsampling unit; performing description knowledge downsampling operation on the initial user resource description knowledge through a description knowledge downsampling unit, generating user resource description knowledge to be aggregated corresponding to the user resource information to be processed, and loading the user resource description knowledge to be aggregated into a target description knowledge aggregation unit; and carrying out description knowledge aggregation operation on the user resource description knowledge to be aggregated through a target description knowledge aggregation unit to generate target user resource description knowledge.

In the embodiment of the invention, the network categories respectively corresponding to the target anti-fraud analysis network and the reference anti-fraud analysis network are not limited, and the network categories can be composed of any one or more neural networks (such as CNN, RNN and the like). In addition, the embodiments of the present invention do not limit the network configurations respectively corresponding to the target anti-fraud analysis network and the reference anti-fraud analysis network.

The AI anti-fraud server generates target user resource description knowledge corresponding to the user resource information to be processed through the target anti-fraud analysis network, and can refer to the text word vector extraction process of the user resource information by the reference anti-fraud analysis network.

STEP205 obtains target linkage anti-fraud strategy text which has correlation with target user resource description knowledge from the target anti-fraud strategy pool according to target strategy text vectors corresponding to the target anti-fraud strategy text in the target anti-fraud strategy pool; the target strategy text vector is descriptive knowledge corresponding to target anti-fraud strategy text output through the target anti-fraud analysis network.

Illustratively, the AI anti-fraud server obtains a request reply relationship chain in a target anti-fraud policy pool; the request response relation chain comprises request features and response features, wherein the request features are generated by text topics corresponding to target anti-fraud strategy texts, and the response features are generated by target strategy text vectors; acquiring the cosine similarity between the target user resource description knowledge and the response characteristics, and determining the response characteristics with the cosine similarity larger than or equal to the cosine similarity limit value as similar response characteristics; and determining the request features corresponding to the similar response features as similar request features, and obtaining target linkage anti-fraud strategy texts in the target anti-fraud strategy pool according to the similar request features.

Optionally, according to the descending order of cosine similarity corresponding to each response feature, the AI anti-fraud server ranks the target strategy text vectors corresponding to each response feature, and determines the target strategy text vectors belonging to the set number as the associated target strategy text vector in the target strategy text vectors subjected to the sequence adjustment, where the set number may represent the preset strategy number of the AI anti-fraud server, for example, set number is 5, and represents that the first 5 target strategy text vectors in the target strategy text vectors subjected to the sequence adjustment are determined as the associated target strategy text vectors; further, the target strategy text vector corresponding to the associated target strategy text vector is determined to be the target linkage anti-fraud strategy text.

STEP206, obtaining a reference linkage anti-fraud strategy text with correlation with the target knowledge piece from the reference anti-fraud strategy pool according to the reference strategy text vector corresponding to the reference anti-fraud strategy text in the reference anti-fraud strategy pool; the reference policy text vector is descriptive knowledge corresponding to the reference anti-fraud policy text output by the reference anti-fraud analysis network.

Illustratively, the embodiment of the invention has a dual request response relation chain link0001 for the target policy text vector corresponding to the target anti-fraud policy text and a request response relation chain link0002 for the reference policy text vector corresponding to the reference anti-fraud policy text, and the request response relation chain link0002 is recorded in the reference anti-fraud policy pool. The information generation process in the request response relation chain link0002 may include: the AI anti-fraud server takes a text subject corresponding to the reference anti-fraud policy text as a request feature and takes the reference policy text vector as a response feature, so that a request response feature doublet comprising the request feature and the response feature can be generated, and further, the generated request response feature doublet can be stored in a request response relation chain link0002 in the reference anti-fraud policy pool.

The AI anti-fraud server obtains a request response relationship chain link0002 in a reference anti-fraud policy pool; the request response relation chain link0002 comprises request features and response features, wherein the request features are generated by text topics corresponding to reference anti-fraud strategy texts, and the response features are generated by reference strategy text vectors; obtaining cosine similarity between the target knowledge segment and the response characteristics, and determining the response characteristics with the cosine similarity larger than or equal to a cosine similarity limit value as similar response characteristics; request features corresponding to the similar response features are determined as similar request features, and according to the similar request features, reference linkage anti-fraud policy text is obtained in the reference anti-fraud policy pool 50 q.

Optionally, according to the descending order of cosine similarity corresponding to each response feature, the AI anti-fraud server ranks the reference strategy text vectors corresponding to each response feature, and determines the reference strategy text vector belonging to the set number as the associated reference strategy text vector from among the reference strategy text vectors subjected to the sequence adjustment, where the set number may represent the number of strategies preset by the AI anti-fraud server, for example, set number is 10, and characterizes that the first 10 reference strategy text vectors in the reference strategy text vectors subjected to the sequence adjustment are determined as the associated reference strategy text vectors; further, a reference policy text vector corresponding to the associated reference policy text vector is determined as a reference linked anti-fraud policy text.

STEP207, in the reference linkage anti-fraud policy text and the target linkage anti-fraud policy text, determines a final anti-fraud policy text for matching the user resource information to be processed.

Illustratively, the AI anti-fraud server determines the reference linked anti-fraud policy text and the target linked anti-fraud policy text as alternative anti-fraud policy texts; the number of alternative anti-fraud policy texts is at least two; obtaining cosine similarity between similar strategy text vectors of each alternative anti-fraud strategy text and target user resource description knowledge respectively; sequentially adjusting at least two candidate anti-fraud strategy texts according to cosine similarity corresponding to each candidate anti-fraud strategy text; from at least two alternative anti-fraud policy texts completing the order adjustment, a final anti-fraud policy text for matching the pending user resource information is determined.

Alternatively, the AI anti-fraud server determines as final anti-fraud policy text an alternative anti-fraud policy text with cosine similarity equal to the cosine similarity limit. It is understood that the cosine similarity limit values in different steps can be flexibly adjusted in the present invention.

Therefore, the embodiment of the invention not only can acquire the target user resource description knowledge recording more resource details, but also can improve the anti-fraud strategy matching precision aiming at the user resource information to be processed; the reference linkage anti-fraud strategy text can be matched in an auxiliary mode through the target knowledge segments, so that modification of the reference strategy text vector in the reference anti-fraud strategy pool can be avoided under the condition that the reference anti-fraud strategy pool is provided, and additional processing expenditure is avoided.

The method for generating the user resource anti-fraud strategy by adopting the artificial intelligence at least comprises the following steps.

STEP301, obtaining at least two target anti-fraud policy texts, and loading the at least two target anti-fraud policy texts into the target anti-fraud analysis network respectively.

STEP302, through the target anti-fraud analysis network, respectively extracting text word vectors from at least two target anti-fraud strategy texts, and generating target strategy text vectors corresponding to the at least two target anti-fraud strategy texts respectively.

Illustratively, in connection with STEP301-STEP302 content, the target anti-fraud policy text includes at least two target anti-fraud policy texts, and the target policy text vector includes target policy text vectors respectively corresponding to the at least two target anti-fraud policy texts.

The AI anti-fraud server generates at least two target policy text vectors corresponding to the target anti-fraud policy texts respectively, and may refer to the description of the target user resource description knowledge corresponding to the generated user resource information to be processed in the above related embodiments.

STEP303, performing grouping operation on at least two target strategy text vectors, and generating Q target strategy description grouping results and target grouping guide vectors respectively corresponding to the Q target strategy description grouping results; q is a positive integer greater than 1, and Q is not greater than the total number of at least two target policy text vectors.

By way of example, the at least two target anti-fraud policy texts 0007 may include user resource information 00071, user resource information 00072, …, user resource information 00073, the at least two target policy text vectors may include target policy text vectors 00074 corresponding to user resource information 00071, target policy text vectors 00075, … corresponding to user resource information 00072, target policy text vectors 00076 corresponding to user resource information 00073, wherein the at least two target anti-fraud policy texts 0007 and the at least two target policy text vectors may each store a target anti-fraud policy pool. Alternatively, at least two target anti-fraud policy texts 0007 may be recorded in other servers, respectively, where the target anti-fraud policy pool stores data (for example, text topics corresponding to target policy text vectors) corresponding to at least two target anti-fraud policy texts 0007 respectively and having relevance to the storage tag.

Further, the AI anti-fraud server performs a clustering operation on at least two target policy text vectors to generate at least one target policy description clustering result and at least one target policy description clustering guide vector corresponding to the at least one target policy description clustering result, for example, the at least one target policy description clustering result may include a first target policy description clustering result composed of target policy text vectors 00074, … and target policy text vector 00076, …, and the at least one target clustering guide vector may include a target clustering guide vector 00081, … corresponding to the first target policy description clustering result, and a target clustering guide vector 00082 corresponding to the second target policy description clustering result, for example, it may be understood that the target clustering guide vector 00081 and the target clustering guide vector 00082 respectively correspond to a scale variable equal to a scale variable corresponding to the target policy text vector.

For example, the AI anti-fraud server may obtain at least two reference anti-fraud policy texts 009, the at least two reference anti-fraud policy texts 009 may include user resource information 0091, user resource information 0092, …, and user resource information 0093, the at least two reference anti-fraud policy texts 009 may include reference policy text vectors 0094 corresponding to user resource information 0091, reference policy text vectors 0095, … corresponding to user resource information 0092, and reference policy text vectors 0096 corresponding to user resource information 0093, respectively, wherein the at least two reference anti-fraud policy texts 009 and the at least two reference policy text vectors may each store a reference anti-fraud policy pool 0097. Alternatively, at least two reference anti-fraud policy texts 009 may be recorded in other servers, respectively, and at this time, the reference anti-fraud policy pool 0097 may store data (for example, text topics corresponding to reference policy text vectors) corresponding to at least two reference anti-fraud policy texts 009 respectively and having relevance to the stored tag.

Further, the AI anti-fraud server performs a clustering operation on at least two reference policy text vectors to generate at least one reference policy description clustering result and at least one reference clustering guide vector corresponding to the reference policy description clustering result respectively, the at least one reference policy description clustering result may include a first reference policy description clustering result composed of reference policy text vectors 0094, … and reference policy text vector 0095, …, and at least a second reference policy description clustering result composed of reference policy text vector 0096, the at least one reference clustering guide vector may include reference clustering guide vectors 0100, … corresponding to the first reference policy description clustering result, and the reference clustering guide vector 0101 corresponding to the second reference policy description clustering result, it may be understood that variables corresponding to the reference clustering guide vector 0100 and the reference clustering guide vector 0101 respectively are equal to scale variables corresponding to the reference policy text vector.

Those skilled in the art will appreciate that the embodiment of the present invention does not limit the clustering method, and may be implemented by a K-means clustering algorithm, for example.

STEP304 stores Q target policy description grouping results and Q target grouping guide vector associations in a target anti-fraud policy pool.

Illustratively, the Q target policy description split results include a target policy description split result cluster_q, Q being a positive integer and Q being no greater than Q. Obtaining target strategy text vectors vec_w from the target strategy description grouping result cluster_q, wherein w is a positive integer, and w is not more than the total number of the target strategy text vectors in the target strategy description grouping result cluster_q; obtaining a target anti-fraud strategy text_w corresponding to the target strategy text vector vec_w, and obtaining a text topic corresponding to the target anti-fraud strategy text_w; taking a text theme corresponding to the target anti-fraud strategy text text_w as a request feature, taking a target strategy text vector vec_w as a response feature, and generating a request response relation chain which has relevance with a target strategy description clustering result cluster_q according to the request feature and the response feature; and storing each target grouping guide vector and the request response relation chain corresponding to each target grouping guide vector in the target anti-fraud strategy pool in a correlated manner.

In combination with STEP303, the AI anti-fraud server may generate a request response relation link72 having a correlation with the first target policy description grouping result, where the request response relation link72 may include a request response feature doublet with a target policy text vector 00074 as a response feature, a request response feature doublet with a text topic corresponding to user resource information 00071 as a request feature, … with a target policy text vector 00076 as a response feature, and a text topic corresponding to user resource information 00073 as a request response feature doublet; the AI anti-fraud server may generate a request response relation chain link73 having a correlation with the second target policy description grouping result, and the request response relation chain link73 may include at least a request response feature binary group having the target policy text vector 00075 as a response feature and a text topic corresponding to the user resource information 00072 as a request feature. Further, the AI anti-fraud server saves the target churn guide vector 00081 and the request response relationship chain link72 association in the target anti-fraud policy pool, …, and saves the target churn guide vector 00082 and the request response relationship chain link73 association in the target anti-fraud policy pool.

In combination with STEP303, the AI anti-fraud server may generate a request response relation link74 having a correlation with the first reference policy description grouping result, where the request response relation link74 may include a request response feature doublet with a reference policy text vector 0094 as a response feature, a request response feature doublet with a text topic corresponding to user resource information 0091 as a request feature, … with a reference policy text vector 0096 as a response feature, and a text topic corresponding to user resource information 0093 as a request response feature doublet; the AI anti-fraud server may generate a request response relation chain link75 having a correlation with the second reference policy description grouping result, and the request response relation chain link75 may include at least a request response feature binary group with reference policy text vector 0095 as a response feature and a text topic corresponding to user resource information 0092 as a request feature. Further, the AI anti-fraud server saves the reference grouping guide vector 0100 and the request response relationship chain link74 association in the reference anti-fraud policy pool 0097, …, and saves the reference grouping guide vector 0101 and the request response relationship chain link75 association in the reference anti-fraud policy pool 0097.

STEP305 obtains the user resource information to be processed, and generates target user resource description knowledge corresponding to the user resource information to be processed through a target anti-fraud analysis network; the target anti-fraud analysis network is obtained by debugging according to a reference anti-fraud analysis network for completing debugging; the target user resource description knowledge has a target knowledge segment which is in knowledge commonality connection with the reference user resource description knowledge, the scale variable corresponding to the target user resource description knowledge is larger than the scale variable corresponding to the reference user resource description knowledge, and the reference user resource description knowledge is description knowledge corresponding to the user resource information to be processed, which is output through the reference anti-fraud analysis network.

It can be appreciated that, compared with the reference anti-fraud analysis network, the target anti-fraud analysis network in the embodiment of the present invention can generate high-order description knowledge (including the high-order strategy text vector and the high-order user resource description knowledge), and a part of dimension vectors (i.e. target knowledge segments) of the high-order user resource description knowledge (which is equal to the target user resource description knowledge) and the reference user resource description knowledge remain relatively consistent, so that the present invention provides a matching method for assisting the reference strategy text vector while ensuring the matching quality for the target user resource description knowledge in scale variable optimization.

For specific implementation of STEP305, please bind STEP101.

STEP306, obtaining target linkage anti-fraud strategy text which has correlation with target user resource description knowledge from the target anti-fraud strategy pool according to target strategy text vectors corresponding to the target anti-fraud strategy text in the target anti-fraud strategy pool; the target strategy text vector is descriptive knowledge corresponding to target anti-fraud strategy text output through the target anti-fraud analysis network.

Illustratively, Q target grouping guide vectors are obtained in a target anti-fraud policy pool; the Q target grouping guide vectors comprise a target grouping guide vector guidance_e, e is a positive integer and e is not more than Q; obtaining the cosine similarity S_e between the target user resource description knowledge and the target grouping guide vector guidance_e; sequentially adjusting cosine similarities corresponding to the Q target grouping guide vectors respectively, and determining the target cosine similarities from the cosine similarities subjected to sequential adjustment; determining a target grouping guide vector corresponding to the target cosine similarity as a target similarity grouping guide vector, and obtaining a similar request response relation chain which is connected with the target similarity grouping guide vector; and obtaining target linkage anti-fraud strategy text which has correlation with target user resource description knowledge according to the similar request response relation chain.

STEP307, obtaining a reference linkage anti-fraud strategy text with correlation with the target knowledge piece from the reference anti-fraud strategy pool according to the reference strategy text vector corresponding to the reference anti-fraud strategy text in the reference anti-fraud strategy pool; the reference policy text vector is descriptive knowledge corresponding to the reference anti-fraud policy text output by the reference anti-fraud analysis network.

In combination with STEP306 and STEP307, the AI anti-fraud server obtains the user resource information to be processed, and generates target user resource description knowledge 0007 corresponding to the user resource information to be processed through the target anti-fraud analysis network, where the target user resource description knowledge 0007 includes a target knowledge segment having knowledge commonality with the reference user resource description knowledge. Further, the AI anti-fraud server obtains at least one target grouping guide vector, such as target grouping guide vectors 00081, …, target grouping guide vector 00082, in the target anti-fraud policy pool, and then obtains cosine similarity between target grouping guide vector 00081 and target user resource description knowledge 0007, …, and cosine similarity between target grouping guide vector 00082 and target user resource description knowledge 0007; further, sequentially adjusting the plurality of cosine similarities, and determining the target cosine similarity from the cosine similarities subjected to the sequential adjustment; the AI anti-fraud server determines a target grouping guide vector corresponding to the target cosine similarity as a target similar grouping guide vector, for example, determines a target grouping guide vector 00081 as a target similar grouping guide vector, and further obtains a similar request response relationship chain link72 which is connected with the target similar grouping guide vector. Subsequently, the AI anti-fraud server performs cosine similarity determination on the target strategy text vector and the target user resource description knowledge contained in the request response relation chain, namely, determines a commonality score between the target strategy text vector and the target user resource description knowledge.

Further, the AI anti-fraud server obtains at least one reference grouping guide vector, such as reference grouping guide vectors 0100, …, reference grouping guide vector 0101, and then obtains cosine similarity between reference grouping guide vector 0100 and the target knowledge piece, …, reference grouping guide vector 0101 and cosine similarity between the target knowledge piece; further, sequentially adjusting the plurality of cosine similarities, and determining the target cosine similarity from the cosine similarities subjected to the sequential adjustment; the AI anti-fraud server determines the reference cluster guide vector corresponding to the target cosine similarity as an associated reference cluster guide vector, and for example, determines the reference cluster guide vector 0100 as an associated reference cluster guide vector, and further obtains a similar request response relation chain link74 in contact with the associated reference cluster guide vector. Subsequently, the AI anti-fraud server performs cosine similarity determination on the reference strategy text vector and the target knowledge segment contained in the request response relation chain, namely, determines a commonality score between the reference strategy text vector and the target knowledge segment.

STEP308, in the reference linked anti-fraud policy text and the target linked anti-fraud policy text, determines a final anti-fraud policy text for matching the user resource information to be processed.

The embodiment of the invention aims at the prior matching performance of the reference anti-fraud policy pool and forms a double matching thread with the target anti-fraud policy pool, when the target anti-fraud policy pool is filled to a sufficient policy amount (such as the volume of the time dimension or the volume of the number dimension), the reference anti-fraud policy pool is filtered, and after filtering, the matching thread only remains the target anti-fraud policy pool.

The embodiment of the invention provides a neural network for assisting target user resource description knowledge to contain target knowledge fragments which are in knowledge commonality connection with reference user resource description knowledge, and the target user resource description knowledge (compared with the reference user resource description knowledge, belonging to high-order description knowledge) can be ensured to have partial dimensionality (namely target knowledge fragments) which can be matched with low-order strategy text vectors by designating the prediction knowledge obtained by a reference knowledge aggregation layer in a general target description knowledge aggregation unit in debugging and assisting the commonality analysis task of the prediction knowledge obtained by the reference description knowledge aggregation unit and simultaneously maintaining the commonality analysis capability of the prediction knowledge obtained by the general target description knowledge aggregation unit.

When the method is applied, the target user resource description knowledge or the target knowledge fragment can be bound according to the dimension of the strategy text vector for carrying out commonality analysis, so that the matching of the reference anti-fraud strategy pool is assisted after dimension derivation, the dynamic updating and optimization of the anti-fraud strategy pool can be realized, and the unnecessary system overhead is reduced. Therefore, by adopting the method and the device, the matching precision of the anti-fraud strategy for the user resource information to be processed can be improved, and the system overhead can be saved.

Based on the above, under some design ideas that can be implemented independently, the debugging steps of the target anti-fraud analysis network are as follows: obtaining a user resource sample information set; the user resource sample information set comprises original user resource sample information, first user resource sample information corresponding to the original user resource sample information and second user resource sample information corresponding to the original user resource sample information; generating original reference prediction knowledge corresponding to the original user resource sample information, first reference prediction knowledge corresponding to the first user resource sample information and second reference prediction knowledge corresponding to the second user resource sample information through a reference anti-fraud analysis network after debugging is completed; generating target original prediction knowledge corresponding to the original user resource sample information, first target prediction knowledge corresponding to the first user resource sample information and second target prediction knowledge corresponding to the second user resource sample information through a general anti-fraud analysis network; the phishing variables in the generic anti-fraud analysis network include phishing variables in the reference anti-fraud analysis network; combining the original reference prediction knowledge, the first reference prediction knowledge, the second reference prediction knowledge, the target original prediction knowledge, the first target prediction knowledge and the second target prediction knowledge, and adjusting the network variables in the general anti-fraud analysis network to generate a target anti-fraud analysis network; the target anti-fraud analysis network is used for generating target user resource description knowledge corresponding to the user resource information to be processed; the target user resource description knowledge has knowledge segments which are in knowledge commonality connection with the reference user resource description knowledge, the scale variable corresponding to the target user resource description knowledge is larger than the scale variable corresponding to the reference user resource description knowledge, and the reference user resource description knowledge is the description knowledge corresponding to the to-be-processed user resource information output through the reference anti-fraud analysis network.

The sample can be understood as a training sample, the universal network can be understood as an untrained network, and based on the scheme, the debugging quality of the target anti-fraud analysis network can be improved, and the running performance of the target anti-fraud analysis network is ensured.

Based on the above, under some design ideas that can be implemented independently, the obtaining the user resource sample information set includes: obtaining at least two first user resource sample tuples, and disassembling the at least two first user resource sample tuples to obtain R first user resource sample tuple sets; r is a positive integer, and R is not greater than the total number of the at least two first user resource sample tuples; the R first user resource sample tuple sets comprise a first user resource sample tuple set H_a, a is a positive integer, and a is not more than R; the first user resource sample set H_a comprises a target first user resource sample set, wherein the target first user resource sample set comprises target original user resource sample information and target first user resource sample information corresponding to the target original user resource sample information; the user resource sample information set comprises user resource sample information sets eg_p in the first user resource sample binary set H_a, p is a positive integer, and p is not more than the total number of the user resource sample information sets in the first user resource sample binary set H_a; obtaining text resource similarity between the target original user resource sample information and the residual user resource sample information; the residual user resource sample information is residual original user resource sample information in a residual first user resource sample binary group or residual first user resource sample information in the residual first user resource sample binary group; the remaining first user resource sample tuples comprise first user resource sample tuples of the first set of user resource sample tuples h_a other than the target first user resource sample tuple; sequentially adjusting the obtained text resource similarity, and determining target residual user resource sample information from the residual user resource sample information according to the text resource similarity subjected to the sequential adjustment; and determining the target residual user resource sample information as target second user resource sample information corresponding to the target original user resource sample information, and determining the target second user resource sample information and the target first user resource sample doublet as the user resource sample information set eg_p.

Based on the above, under some design ideas which can be implemented independently, the reference anti-fraud analysis network includes a reference description knowledge mining unit, a reference description knowledge downsampling unit, and the reference description knowledge aggregation unit; the generating, by the universal anti-fraud analysis network, target original prediction knowledge corresponding to the original user resource sample information, first target prediction knowledge corresponding to the first user resource sample information, and second target prediction knowledge corresponding to the second user resource sample information includes: determining the original user resource sample information, the first user resource sample information and the second user resource sample information as user resource sample information to be processed; loading the user resource sample information to be processed into the general anti-fraud analysis network; the general anti-fraud analysis network comprises a general description knowledge mining unit, a general description knowledge downsampling unit and a general target description knowledge aggregation unit; wherein the network variable in the generic description knowledge mining unit is equal to the network variable in the reference description knowledge mining unit; the network variable in the general description knowledge downsampling unit is equal to the network variable in the reference description knowledge downsampling unit; the general target description knowledge aggregation unit comprises a reference knowledge aggregation level and a target knowledge aggregation level, wherein network variables in the reference knowledge aggregation level are equal to network variables in the reference description knowledge mining unit; performing description knowledge mining processing on the to-be-processed user resource sample information through the general description knowledge mining unit, generating initial example prediction knowledge corresponding to the to-be-processed user resource sample information, and loading the initial example prediction knowledge into the general description knowledge downsampling unit; performing description knowledge downsampling operation on the initial example prediction knowledge through the general description knowledge downsampling unit, generating example prediction knowledge to be aggregated corresponding to the user resource sample information to be processed, and loading the example prediction knowledge to be aggregated into the general target description knowledge aggregation unit; performing description knowledge aggregation operation on the to-be-aggregated example prediction knowledge through the reference knowledge aggregation layer in the general target description knowledge aggregation unit to generate a reference prediction knowledge block; performing description knowledge aggregation operation on the to-be-aggregated example prediction knowledge through the target knowledge aggregation layer in the general target description knowledge aggregation unit to generate a target prediction knowledge block; performing description knowledge collection operation on the reference prediction knowledge block and the target prediction knowledge block to obtain target example prediction knowledge corresponding to the to-be-processed user resource sample information; the target example prediction knowledge includes the target raw prediction knowledge, the first target prediction knowledge, and the second target prediction knowledge.

Based on the above, under some design ideas that can be implemented independently, the combining the original reference prediction knowledge, the first reference prediction knowledge, the second reference prediction knowledge, the target original prediction knowledge, the first target prediction knowledge, and the second target prediction knowledge adjusts network variables in the general anti-fraud analysis network to generate a target anti-fraud analysis network, including: obtaining first cosine similarity between the original reference prediction knowledge and the first reference prediction knowledge, and obtaining second cosine similarity between the original reference prediction knowledge and the second reference prediction knowledge; combining the first cosine similarity and the second cosine similarity, and determining a reference network debugging cost corresponding to the reference anti-fraud analysis network; obtaining third cosine similarity between the target original prediction knowledge and the first target prediction knowledge, and obtaining fourth cosine similarity between the target original prediction knowledge and the second target prediction knowledge; obtaining a similarity limit value generated according to the reference network debugging cost, and determining a global network debugging cost corresponding to the general anti-fraud analysis network by combining the third cosine similarity, the fourth cosine similarity and the similarity limit value; and adjusting network variables in the general description knowledge mining unit, the general description knowledge downsampling unit and the target knowledge aggregation level in combination with the global network debugging cost to generate the target anti-fraud analysis network.

Based on the above, under some design ideas that can be implemented independently, the combining the original reference prediction knowledge, the first reference prediction knowledge, the second reference prediction knowledge, the target original prediction knowledge, the first target prediction knowledge, and the second target prediction knowledge adjusts network variables in the general anti-fraud analysis network to generate a target anti-fraud analysis network, including: combining the original reference prediction knowledge, the first reference prediction knowledge and the second reference prediction knowledge to determine a reference network debugging cost corresponding to the reference anti-fraud analysis network; determining a first network debugging cost corresponding to the general anti-fraud analysis network by combining the reference network debugging cost, the target original prediction knowledge, the first target prediction knowledge and the second target prediction knowledge; obtaining the reference prediction knowledge block, and determining a second network debugging cost corresponding to the universal anti-fraud analysis network by combining the reference prediction knowledge block and the reference network debugging cost; determining the weighted results of the first network debugging cost and the second network debugging cost as global network debugging cost of the general anti-fraud analysis network; counting reference optimization rounds of the general anti-fraud analysis network; the reference optimization round is used to characterize rounds in the generic anti-fraud analysis network in which network variables have been optimized; and adjusting the network variables in the general anti-fraud analysis network by combining the global network debugging cost and the reference optimization round to generate the target anti-fraud analysis network.

Based on the above, under some design ideas that can be implemented independently, the reference prediction knowledge block includes a reference original prediction knowledge block corresponding to the original user resource sample information, a first reference prediction knowledge block corresponding to the first user resource sample information, and a second reference prediction knowledge block corresponding to the second user resource sample information; the obtaining the reference prediction knowledge block, combining the reference prediction knowledge block and the reference network debugging cost, and determining a second network debugging cost corresponding to the general anti-fraud analysis network includes: determining a reference prediction knowledge block to be updated from the reference original prediction knowledge block, the first reference prediction knowledge block and the second reference prediction knowledge block; when the reference prediction knowledge block to be updated is the reference original prediction knowledge block, obtaining a fifth cosine similarity between the original reference prediction knowledge and the first reference prediction knowledge block, obtaining a sixth cosine similarity between the original reference prediction knowledge and the second reference prediction knowledge block, and determining the second network debugging cost corresponding to the general anti-fraud analysis network by combining the fifth cosine similarity, the sixth cosine similarity and the reference network debugging cost; when the reference prediction knowledge block to be updated is the first reference prediction knowledge block, obtaining a seventh cosine similarity between the reference original prediction knowledge block and the first reference prediction knowledge block, obtaining an eighth cosine similarity between the reference original prediction knowledge block and the second reference prediction knowledge block, and determining the second network debugging cost corresponding to the general anti-fraud analysis network by combining the seventh cosine similarity, the eighth cosine similarity and the reference network debugging cost; and when the reference prediction knowledge block to be updated is the second reference prediction knowledge block, obtaining a ninth cosine similarity between the reference original prediction knowledge block and the first reference prediction knowledge block, obtaining a tenth cosine similarity between the reference original prediction knowledge block and the second reference prediction knowledge, and determining the second network debugging cost corresponding to the general anti-fraud analysis network by combining the ninth cosine similarity, the tenth cosine similarity and the reference network debugging cost.

Based on the above, under some design ideas that can be implemented independently, the adjusting the network variables in the general anti-fraud analysis network by combining the global network debugging cost and the reference optimization round, to generate the target anti-fraud analysis network includes: when the reference optimization turn does not accord with the adjustment step length, the network variables in the general description knowledge mining unit, the network variables in the general description knowledge downsampling unit and the network variables in the target knowledge aggregation level are adjusted by combining with the global network debugging cost, and the target anti-fraud analysis network is obtained according to the adjusted general anti-fraud analysis network; when the reference optimization round meets the adjustment step length, the network variables in the general description knowledge mining unit, the network variables in the general description knowledge downsampling unit, the network variables in the target knowledge aggregation level and the network variables in the reference knowledge aggregation level are adjusted in combination with the global network debugging cost, and the target anti-fraud analysis network is obtained according to the adjusted general anti-fraud analysis network.

The foregoing is only a specific embodiment of the present invention. Variations and alternatives will occur to those skilled in the art based on the detailed description provided herein and are intended to be included within the scope of the invention.

Claims

1. A method for generating a user resource anti-fraud policy by artificial intelligence, which is characterized by being applied to an AI anti-fraud server, the method comprising:

2. The method of claim 1, wherein generating, through a target anti-fraud analysis network, target user resource description knowledge corresponding to the user resource information to be processed, comprises:

3. The method according to claim 1, wherein the method further comprises:

4. The method as recited in claim 1, wherein said obtaining target linked anti-fraud policy text from said target anti-fraud policy pool that has relevance to said target user resource description knowledge in accordance with target policy text vectors corresponding to target anti-fraud policy text in said target anti-fraud policy pool comprises:

5. The method as recited in claim 1, wherein said target anti-fraud policy text comprises at least two target anti-fraud policy texts, said target policy text vector comprising target policy text vectors to which said at least two target anti-fraud policy texts respectively correspond; the method further comprises the steps of:

6. The method of claim 5, wherein the Q target policy description split results include a target policy description split result cluster_q, Q being a positive integer and Q being no greater than Q; the storing the Q target policy description grouping results and Q target grouping guide vector associations in the target anti-fraud policy pool includes:

7. The method as recited in claim 6, wherein said obtaining target linked anti-fraud policy text from said target anti-fraud policy pool that has relevance to said target user resource description knowledge in accordance with target policy text vectors corresponding to target anti-fraud policy text in said target anti-fraud policy pool comprises:

8. The method as recited in claim 1, wherein said determining a final anti-fraud policy text for matching said user resource information to be processed in said reference linked anti-fraud policy text and said target linked anti-fraud policy text comprises:

9. An AI anti-fraud server, comprising: a memory and a processor; the memory is coupled to the processor; the memory is used for storing computer program codes, and the computer program codes comprise computer instructions; wherein the processor, when executing the computer instructions, causes the AI anti-fraud server to perform the method of any of claims 1-8.

10. A computer readable storage medium, characterized in that it has stored thereon a computer program, which, when run, is a method according to any of claims 1-8.