CSREC

Overview

This GitHub repo is for paper CSRec: Rethinking Sequential Recommendation from A Causal Perspective

Running

Environment setting

conda env create -f CSRec_env.yaml
conda activate csrec

training and testing CSRec

• Step 1: pretrain a sequential recommender system on observational data.

Notice that training at this step does not require the recommendation and user decision data under intervention. They are only used for evaluation purposes.

python run_rec.py --model_type [model type] --data_dir [data directory] --obs_data_name [observational dataset name]\
 --rec_data_name [recommendation dataset name] --dec_data_name [user's decision dataset] --lr [learning rate] \
--num_attention_heads 1 --train_name [saved log name] --epochs [epochs]

• Step 2: train CSRec on interventional data with the given pretrained model

  python run_rec.py --model_type [pretrained model type] --data_dir [data directory] \
  --obs_data_name [observational dataset name] --rec_data_name [recommendation dataset name] \
  --dec_data_name [user's decision dataset] --lr [learning rate] --num_attention_heads 1 \
  --load_pretrain_model [name for the pretrained model] --train_name [saved log name] --epochs [epochs]

All models after training will be saved in the folder src/output.

To test the CSRec after training, add do_eval boolean argument and provide the model name without .pt.

Example on GPT4 books dataset

python main.py --model_type [pretrained model type] --data_dir [data directory] \
--obs_data_name [observational dataset name] --rec_data_name [recommendation dataset name] \
 --dec_data_name [user's decision dataset] --lr [learning rate] --num_attention_heads 1 \
--load_pretrain_model [name for the pretrained model] --load_model [name for CSRec model] --do_eval

Train and test other baselines

other baselines can be trained and tested by changing the model_typeargument to the following ones:

• BSARec Caser, GRU4Rec, SASRec, BERT4Rec, FMLPRec

Example of training SASRec on GPT4 books dataset:

python run_rec.py --model_type SASRec --obs_data_name [observational dataset name] \
 --rec_data_name [recommendation dataset name] --dec_data_name [user's decision dataset] --lr [learning rate] \
--num_attention_heads 1 --train_name [saved log name] --epochs [epochs]

Testing the model

python run_rec.py --model_type SASRec --obs_data_name [observational dataset name] \
 --rec_data_name [recommendation dataset name] --dec_data_name [user's decision dataset] \
--num_attention_heads 1 --load_model [model name] --do_eval

Acknowledgement

This repository is based on BSARec.