The repository holds the code for mining logical queries/creating user intention KG from session data.
Just download the data by clicking this link, and extract them in root of this repository. Three datasets are all included. There is more pre-processing for LSGT and some of the baselines, so please remember that using the correct corresponding files otherwise will cause data-loading errors. It is too abstract to describe how to run the code because there are many combinations of logical encoder, session encoder, and datasets. We make it clearer by the following example scripts, and you can substitute the flags.
You can run the following script for logical encoder + session encoder on the Amazon dataset
python ./deduction_model/train.py \
-dn hyper_graph_data_en \
-m fuzzqe \
--train_query_dir ./sampled_hyper_train_merged \
--valid_query_dir ./sampled_hyper_valid \
--test_query_dir ./sampled_hyper_test \
--checkpoint_path ./logs \
-fol \
-b 512 \
--log_steps 120000 \
-lr 0.001 \
--session_encoder AttnMixer
The -m fuzzqe denotes the logical encoder which we use here. The ```--session_encoder''' denotes the session encoder here. In this implementation, we implemented the combinations of these two flags, which have the following options.
The baseline logical encoder implementations are adopted from the SQE repo, a general code base. Here is the code for the
Here are the options for the session encoder flags.
| Session Encoder Flag (--session_encoder) | Paper |
|---|---|
| GRURec | Empirical Evaluation of Gated Recurrent Neural Networks on Sequence Modeling |
| SRGNN | Session-based Recommendation with Graph Neural Networks |
| AttnMixer | Efficiently Leveraging Multi-level User Intent for Session-based Recommendation via Atten-Mixer Network |
python ./deduction_model/train.py \
-dn hyper_graph_dressipi \
-m nqe \
--train_query_dir ./sampled_hyper_train_dressipi_merged \
--valid_query_dir ./sampled_hyper_valid_dressipi \
--test_query_dir ./sampled_hyper_test_dressipi \
--checkpoint_path ./logs \
-fol \
-b 512 \
--log_steps 120000 \
-lr 0.0001
Note that there is a trailing _dgl in the queries dirs, in which we use dgl to convert the query graph into graph tokens.
python ./deduction_model/train.py \
-dn hyper_graph_diginetica \
-m lsgt \
--train_query_dir ./sampled_hyper_train_diginetica_merged_graph_dgl \
--valid_query_dir ./sampled_hyper_valid_dininetica_graph_dgl \
--test_query_dir ./sampled_hyper_test_diginetica_graph_dgl \
--checkpoint_path ./logs \
--num_layers 2 \
-fol \
-b 1024 \
--log_steps 120000 \
-lr 0.001
We also include the shell scripts that we run to reproduce the number shown in the paper.
If anything is unclear, please get in touch with Jiaxin Bai via [email protected]. If you find this useful, please cite:
@article{DBLP:journals/corr/abs-2312-13866,
author = {Jiaxin Bai and
Chen Luo and
Zheng Li and
Qingyu Yin and
Yangqiu Song},
title = {Understanding Inter-Session Intentions via Complex Logical Reasoning},
journal = {CoRR},
volume = {abs/2312.13866},
year = {2023},
url = {https://doi.org/10.48550/arXiv.2312.13866},
doi = {10.48550/ARXIV.2312.13866},
eprinttype = {arXiv},
eprint = {2312.13866},
timestamp = {Wed, 17 Jan 2024 17:14:38 +0100},
biburl = {https://dblp.org/rec/journals/corr/abs-2312-13866.bib},
bibsource = {dblp computer science bibliography, https://dblp.org}
}