GraphTransfer: A Generic Feature Fusion Framework for Collaborative Filtering

TL;DR

Theoretical analysis and experiments on public datasets show that GraphTransfer outperforms other feature fusion methods in CF tasks and demonstrates the universality of the framework via empirical studies in three other scenarios, showing that GraphTransfer leads to significant improvements in the performance of CF algorithms.

Abstract

Graph Neural Networks (GNNs) have demonstrated effectiveness in collaborative filtering tasks due to their ability to extract powerful structural features. However, combining the graph features extracted from user-item interactions and auxiliary features extracted from user genres and item properties remains a challenge. Currently available fusion methods face two major issues: 1) simple methods such as concatenation and summation are generic, but not accurate in capturing feature relationships; 2) task-specific methods like attention mechanisms and meta paths may not be suitable for general feature fusion. To address these challenges, we present GraphTransfer, a simple but universal feature fusion framework for GNN-based collaborative filtering. Our method accurately fuses different types of features by first extracting graph features from the user-item interaction graph and auxiliary features from users and items using GCN. The proposed cross fusion module then effectively bridges the semantic gaps between the interaction scores of different features. Theoretical analysis and experiments on public datasets show that GraphTransfer outperforms other feature fusion methods in CF tasks. Additionally, we demonstrate the universality of our framework via empirical studies in three other scenarios, showing that GraphTransfer leads to significant improvements in the performance of CF algorithms.

Figures (5)

• Figure 1: The architecture of the proposed GraphTransfer framework, which consists of three modules: Graph Feature Extraction Module, Auxiliary Feature Extraction Module and Cross Fusion Module.
• Figure 2: The ablation study results for GraphTransfer-GIN (the upper two figures) and GraphTransfer-LightGCN (the lower two figures) under eight different settings on ML1M dataset. Setting 1--Setting 4 represent replacing cross fusion module in GraphTransfer with concatenation, plain summation, weighted summation and attention mechanism respectively. Setting 5 represents removing user and item GCNs in GraphTransfer.
• Figure 3: The sensitivity analysis results with four hyper-parameters for GraphTransfer-GIN ((a)--(d)) and GraphTransfer-LightGCN ((e)--(h)) on ML1M: learning rate of auxiliary feature extraction module $\eta_1$, user threshold $\epsilon_1$, weight $\lambda_1$ and $\lambda_2$ in Equation (\ref{['eq:loss']}), respectively.
• Figure 4: The true distributions of the top 6 categories favored by user 1377 and user 4261.
• Figure 5: The distribution of node embeddings (including users and items) after t-SNE in three different settings. The figure (a) and (d) show the distributions of node embeddings learned by the auxiliary feature extraction module, the figure (b) and (e) show the distributions of node embeddings from the last layer of LightGCN and IMP-GCN, and the figure (c) and (f) show the distributions of node embeddings from the last layer of GraphTransfer-LightGCN and GraphTransfer-IMP-GCN.