geogubd / STAM

Source code and dataset for WWW'22 paper "STAM: A Spatiotemporal Aggregation Method for Graph Neural Network-based Recommendation"

STAM

Source code and dataset for WWW'22 paper "STAM: A Spatiotemporal Aggregation Method for Graph Neural Network-based Recommendation"

STAM: A Spatiotemporal Aggregation Method for Graph Neural Network-based Recommendation

Zhen Yang, Ming Ding, Bin Xu, Hongxia Yang and Jie Tang

In WWW 2022

Introduction

STAM is a novel aggregation method, which generates spatiotemporal neighbor embeddings from the perspectives of spatial structure information and temporal information, facilitating the development of aggregation methods from spatial to spatiotemporal. STAM utilizes the Scaled Dot-Product Attention to capture temporal orders of one-hop neighbors and employs multi-head attention to perform joint attention over different latent subspaces.

Preparation

• Python 3.7
• Tensorflow 1.14.0

Training

Training on the existing datasets

You can use $ ./experiments/***.sh to train STAM model. For example, if you want to train on the ml-1m dataset, you can run $./experiments/train_ml.sh to train STAM model.

Cite