A multi-view GNN-based network representation learning framework for recommendation systems

doi:10.1016/j.neucom.2024.129001

A multi-view GNN-based network representation learning framework for recommendation systems

2025-02-01

0

PRE

AI

A

Amina Amara *

M

Mohamed Ali Hadj Taieb

M

Mohamed Ben Aouicha

DOI：10.1016/j.neucom.2024.129001

Original

Original request for help

Save

Abstract

En 中文

Graph neural network-based methods have been widely adopted by recommendation systems to provide robust backbones for learning embeddings from multi-view graphs. They have the potential to enhance the performance of recommendation systems by transferring abundant knowledge across different views. However, current works still have deficiencies in capturing the various types of relationships present in multi-view data. They overlook the semantic consistency between fused and view representations and have difficulty in modeling the complementary information between different views. To address these limitations, we propose a novel solution named CrossViRec, a Cross-View graph neural network-based model to learn low-dimensional node representations for link Rec ommendation. The model leverages two key aspects of multi-view networks: diversity and collaboration. Diversity preserves the diverse semantics of different views, while collaboration facilitates the cooperation between these views. Therefore, the model defines two types of embeddings: a high- dimensional shared embedding and a low-dimensional edge embedding for each view. In particular, a function is learned to generate edge embeddings by aggregating features from node's neighborhood inside various views through a BiLSTM aggregator followed by a self-attention mechanism. To get the overall vector representation on each view, the self-attention mechanism combines different edge embeddings of a given node and then merges them with the shared embedding vector that encodes the cross-view structural features. Experimental evaluations for the proposed model were conducted on four different challenging datasets: YouTube, Amazon, Twitter, and a private cross-social network dataset. The extensive experiments demonstrate that the proposed model can significantly outperform the existing baseline methods for the link recommendation task, and also achieve better performance for cross-social networks link recommendation task.

Keywords:

Multi-view network representation learning

Recommendation systems

Graph neural networks

BiLSTM

Self-attention mechanism

A multi-view GNN-based network representation learning framework for recommendation systems

Abstract

Journal

Organization

Cited Papers