Graph-Refined Convolutional Network for Multimedia Recommendation with Implicit Feedback

This is our Pytorch implementation for the paper:

Yinwei Wei, Xiang Wang, Liqiang Nie, Xiangnan He and Tat-Seng Chua. Graph-Refined Convolutional Network for Multimedia Recommendation with Implicit Feedback. In ACM MM`20, Seattle, United States, Oct. 12-16, 2020
Author: Dr. Yinwei Wei (weiyinwei at hotmail.com)

Introduction

In this work, we focus on adaptively refining the structure of interaction graph to discover and prune potential false-positive edges. Towards this end, we devise a new GCN-based recommendermodel, Graph-Refined Convolutional Network(GRCN), which adjusts the structure of interaction graph adaptively based on status of mode training, instead of remaining the fixed structure.

Environment Requirement

The code has been tested running under Python 3.5.2. The required packages are as follows:

• Pytorch == 1.4.0
• torch-cluster == 1.4.2
• torch-geometric == 1.2.1
• torch-scatter == 1.2.0
• torch-sparse == 0.4.0
• numpy == 1.16.0

Example to Run the Codes

The instruction of commands has been clearly stated in the codes.

• Kwai dataset
  python main.py --l_r=0.0001 --weight_decay=0.1 --dropout=0 --weight_mode=confid --num_routing=3 --is_pruning=False --data_path=Kwai --has_a=False --has_t=False
• Tiktok dataset
  python main.py --l_r=0.0001 --weight_decay=0.001 --dropout=0 --weight_mode=confid --num_routing=3 --is_pruning=False --data_path=Tiktok
• Movielens dataset
  python main.py --l_r=0.0001 --weight_decay=0.0001 --dropout=0 --weight_mode=confid --num_routing=3 --is_pruning=False

Some important arguments:

• weight_model It specifics the type of multimodal correlation integration. Here we provide three options:

  1. mean implements the mean integration without confidence vectors. Usage --weight_model 'mean'
  2. max implements the max integration without confidence vectors. Usage --weight_model 'max'
  3. confid (by default) implements the max integration with confidence vectors. Usage --weight_model 'confid'

• fusion_mode It specifics the type of user and item representation in the prediction layer. Here we provide three options:

  1. concat (by default) implements the concatenation of multimodal features. Usage --fusion_mode 'concat'
  2. mean implements the mean pooling of multimodal features. Usage --fusion_mode 'max'
  3. id implements the representation with only the id embeddings. Usage --fusion_mode 'id'

• is_pruning It specifics the type of pruning operation. Here we provide three options:

  1. Ture (by default) implements the hard pruning operations. Usage --is_pruning 'True'
  2. False implements the soft pruning operations. Usage --is_pruning 'False'

• 'has_v', 'has_a', and 'has_t' indicate the modality used in the model.

Dataset

Please check MMGCN for the datasets: Kwai, Tiktok, and Movielens.

Due to the copyright, we could only provide some toy datasets for validation. If you need the complete ones, please contact the owners of the datasets.

	#Interactions	#Users	#Items	Visual	Acoustic	Textual
Movielens	1,239,508	55,485	5,986	2,048	128	100
Tiktok	726,065	36,656	76,085	128	128	128
Kwai	298,492	86,483	7,010	2,048	-	-

-train.npy Train file. Each line is a user with her/his positive interactions with items: (userID and micro-video ID)
-val.npy Validation file. Each line is a user with her/his several positive interactions with items: (userID and micro-video ID)
-test.npy Test file. Each line is a user with her/his several positive interactions with items: (userID and micro-video ID)