The code for TKDE 2024 paper "Reinforced path reasoning for counterfactual explainable recommendation"

Getting Started

Our code contains three important steps:

1. Pretrain recommendation model (optional);
2. Train counterfactual reasoning agent;
3. Generate counterfactual explanations.

Step 1: Pretraining black-box recommendation model (optional)

We provide an example: CliMF in the BaseRecRepo, for training the black-box recommendation model: The CliMF directly optimizes the Top-$K$ recommendation by modeling parameters through maximizing the Mean Reciprocal Rank (MRR).

1. To train the CliMF model, you should first generate .bat train files by

   convert_file(train_file_path, CF.train_user_dict)
   
   convert_file(test_file_path, CF.test_user_dict)
   

2. Then manually add the file head to the two files as required in Orange.data.Table(files) function with:

 user   item  relevance
 c  c  c
 row=1 col=1 class

3. Train CliMF and save recommendation model parameters:

 python train_base_rec.py 

4. Load pretraining embeddings into CERec: in recommender_agent.py, load embeddings with:

 all_embed.data = torch.from_numpy(*your_trained_embedding_file) 

If you do not wish to use pertaining models, apply Xavier on user/item embeddings to initialize embeddings. To do so, simply use:

 nn.init.xavier_uniform_(all_embed) in recommender_agent.py 

Step 2. Counterfactual reasoning agent (mandatory)

1. Edit the configs for training RL agent: The parser file is placed in common/parser; Important args are:

--data_path: Input data path.
--dataset: Choose a dataset.
--emb_size: Embedding size.
--counter_threshold: counterfactual top-k rank threshold.
--slr: Learning rate for sampler.
--rlr: Learning rate recommender.
--inter_threshold: interaction threshold to construct interaction matrix.
--interaction_len: Input vector length of user embedding.
--k_step: k step from current positive items

2. We also provide alternatives for variants; you may choose different graph learning methods, samplers, recommendation models and reward functions by:

  #Graph Learning Module
  --GCN: GCN representation learning operator, options are [SAGE, GCN, GNN, SG, LG, GAT, Trans]

  #Agent
  --reward: reward function, options are [all, R-reward, S-reward]

  #Sampler: 
  --sampler: samplers, options are [CPS, 1-Uniform, 2-Uniform]

  #Recommendation Model
  --recommender: Recommenders, support [MF, BPR, NeuMF, KGAT]

3. After configuring the parser, train the agent with:

python train_RL_agent.py --your-preferred-configs

4. You can get the trained sampler and recommender models saved in ./weights/.

Step 3: Generate counterfactual attribute-level explanations

1. load_state_dict from the trained sampler and recommender models from step 2 by replacing the file names in generate_explanation.py
2. run

python generate_explanation.py
  ```

The output: .txt files of counterfactual_examples_{dataset}.txt; and counterfactual_attributes_{dataset}.txt, which are placed in ./explanations/

Cite information:

If you find our work useful, please cite the following work: @article{wang2024reinforced, title={Reinforced path reasoning for counterfactual explainable recommendation}, author={Wang, Xiangmeng and Li, Qian and Yu, Dianer and Li, Qing and Xu, Guandong}, journal={IEEE Transactions on Knowledge and Data Engineering}, year={2024}, publisher={IEEE} }