Multi-Agent Reinforcement Knowledge Graph Reasoning for Explainable Recommendation

Datasets

Two Amazon datasets (Amazon_Beauty, Amazon_Cellphones) are available in the "data/" directory and the split is consistent with [1]. All four datasets used in this paper can be downloaded here.

Requirements

Python >= 3.6
PyTorch = 1.0

Skip the training phase

We've included all the required files for the Cellphone dataset so you can simply evaluate performance and play around with the recommendation parameters

Copy the contents of this folder ( https://drive.google.com/open?id=1ZbR7nWmpmDjMYmv56ChxonB5q3gKEsAI ) to the root directory of your project
Run the multi-agent evaluation

python test_multi_agent.py --dataset cell --run_path False --run_eval True --re_rank True

How to run the code from scratch

Pre-process the data:

python preprocess.py --dataset <dataset_name>

"<dataset_name>" should be one of "cd", "beauty", "cloth", "cell" (refer to utils.py).

Train knowledge graph embeddings (TransE in this case):

python train_transe_model.py --dataset <dataset_name>

Train User RL agent:

python train_agent.py --dataset <dataset_name>

Evaluate the User agent

python test_agent.py --dataset <dataset_name> --run_path True --run_eval True

If "run_path" is True, the program will generate paths for recommendation according to the trained policy. Note that for subsequent runs you should set this to 'False'

If "run_eval" is True, the program will evaluate the recommendation performance based on the resulting paths.

Train Product RL agent:

python train_product_agent.py --dataset <dataset_name>

Evaluate the Product agent

python test_product_agent.py --dataset <dataset_name> --run_path True --run_eval True

Evaluate the final multi-agent setup

python test_multi_agent.py --dataset <dataset_name> --run_path False --run_eval True --re_rank True

References

[1] Yongfeng Zhang, Qingyao Ai, Xu Chen, W. Bruce Croft. "Joint Representation Learning for Top-N Recommendation with Heterogeneous Information Sources". In Proceedings of CIKM. 2017.