MFGAN

Code for our SIGIR 2020 paper:

Sequential Recommendation with Self-Attentive Multi-Adversarial Network.
Paper

Introduction

In this paper, we have proposed a Multi-Factor Generative Adversarial Network (MFGAN) for sequential recommendation. In our framework, the generator taking user behavior sequences as input is used to generate possible next items, and multiple factor-specific discriminators are used to evaluate the generated sub-sequence from the perspectives of different factors.

Setup

This code is based on SASREC.

Requirements

• Tensorflow 1.2.0
• Python 3.6

Data

For the training data of the generator, each line contains an user id and item id meaning an interaction.

As for the training data of the discriminators, it contains n files if the dataset have n attributes. In an attribute file, each line contains an attribute of a item, which depends on how the attribute information defined.

Model training

python main.py --dataset="generator training set name" --train_dir=default

Reference

@inproceedings{Ren2020Sequential,
  author    = {Ruiyang Ren,
               Zhaoyang Liu,
               Yaliang Li,
               Wayne Xin Zhao,
               Hui Wang,
               Bolin Ding,
               Ji{-}Rong Wen},                 
  title     = {Sequential Recommendation with Self-Attentive Multi-Adversarial Network},  
  booktitle = {Proceedings of the 43rd International {ACM} {SIGIR} conference on
               research and development in Information Retrieval, {SIGIR} 2020, Virtual
               Event, China, July 25-30, 2020},                 
  year      = {2020},    
}