ELECRec

Source code for paper: ELECRec: Training Sequential Recommenders as Discriminators

Introduction

Motivation:

Sequential recommendation is often considered as a generative task, i.e., training a sequential encoder to generate the next item of a user’s interests based on her historical interacted items. Despite their prevalence, these methods usually require training with more meaningful samples to be effective, which otherwise will lead to a poorly trained model.

Model Architecture:

Result Summary

Following table shows the overall comparison with other sequential models.

See our paper for more details.

Implementation

Requirements

Python >= 3.7
Pytorch >= 1.2.0
tqdm == 4.26.0

Datasets

Four prepared datasets are included in data folder.

Evaluate Model

We provide the trained models on Yelp, Beauty, Sports_and_Games, and Toys_and_Games datasets in ./src/output folder. You can directly evaluate the trained models on test set by running:

python main.py --data_name <Data_name> --model_idx 0 --do_eval

You are expected following results:

On Beauty:

{'Epoch': 0, 'HIT@5': '0.0703', 'NDCG@5': '0.0503', 'HIT@10': '0.0964', 'NDCG@10': '0.0587', 'HIT@20': '0.1311', 'NDCG@20': '0.0674'}

On Sports:

{'Epoch': 0, 'HIT@5': '0.0380', 'NDCG@5': '0.0268', 'HIT@10': '0.0548', 'NDCG@10': '0.0322', 'HIT@20': '0.0766', 'NDCG@20': '0.0377'}

On Toys:

{'Epoch': 0, 'HIT@5': '0.0756', 'NDCG@5': '0.0557', 'HIT@10': '0.0997', 'NDCG@10': '0.0634', 'HIT@20': '0.1316', 'NDCG@20': '0.0714'}

On Yelp:

{'Epoch': 0, 'HIT@5': '0.0434', 'NDCG@5': '0.0218', 'HIT@10': '0.0593', 'NDCG@10': '0.0270', 'HIT@20': '0.0649', 'NDCG@20': '0.0285'}

Please feel free to test is out!

Train Model

To train ELECRec on a specific dataset, change to the src folder and run following command:

bash scripts/run_<data_name>.sh

The script will automatically train ELECRec and save the best model found in validation set, and then evaluate on test set.

Explanation of key arguments in the script:

--gen_loss_weight: strength of generator loss    
--sample_ratio: percentage of target items in a sequences will be replaced by the predict items from generator.     
--dis_loss_weight: strength of discriminator loss
--model_shared_type: model parameter sharing strategy between generator and discriminator

Reference

@inproceedings{chen22elecrec,
  title = "ELECRec: Training Sequential Recommenders as Discriminators",
  author = "Yongjun Chen and Jia Li and Caiming Xiong",
  year = "2022",
  booktitle = "SIGIR"
}

PALIN2018 / ELECRec