Artifacts related to the SIGKDD 2022 paper "ProActive: Self-Attentive Temporal Point Process Flows for Activity Sequences".

Use a python-3.7 environment and install PyTorch v1.6.0 and TorchVision v0.4.2. More details are given in requirements.txt file.

We have also provided a dataset used in the paper. For any new dataset, you need to structure it to different files containing the action sequences and the sequence goals as follows:

test_ev.txt test_go.txt test_ti.txt train_ev.txt train_go.txt train_ti.txt

Here, we provide the description of the files that are given in the dataset:

• train_ev.txt = The types of the actions done by the users.
• train_ti.txt = The times of the actions done by the users.
• train_go.txt = The final goal of the activity.

Similarly, we have the exact three files for the test dataset.

PROACTIVE requires you to create dumps before running the model. Thus, you must run the develop_dumps.py, for example, run the following command to get pickle dumps for Breakfast dataset:

python develop_dumps.py Breakfast

The develop_dumps.py is used to combine all the different files, normalize the time of actions, and obtain ".p" dumps for training/testing data.

To train PROACTIVE, use the provided bash script "run.sh". For example, to run the model on Breakfast dataset, use the command:

sh run.sh

The contents of the script are self-explanatory. Once you run the code, PROACTIVE loads the dataset and predicts the accuracy, MAE, and GPA. In detail, the metrics used are as follows

• Acc = The accuracy of predicting the event types.
• MAE = Mean absolute error between the true and predicted action times.
• GPA = The goal prediction accuracy over the test data.
• Itv. GPA = This reflects the ability of PROACTIVE to predict the correct goal at every 'index', i.e, at the incoming of any new action. This is NOT similar to the goal prediction accuracy and is used to keep track of the performance variation due to gamma (RL-trick).

If you use this code in your research, please cite:

@inproceedings{kdd22,
 author = {Vinayak Gupta and Srikanta Bedathur},
 booktitle = {Proc. of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD)},
 title = {ProActive: Self-Attentive Temporal Point Process Flows for Activity Sequences},
 year = {2022}
}

In case of any issues, please send a mail to guptavinayak51 (at) gmail (dot) com