An improved recommendation based on graph convolutional network

Yichen He, Yijun Mao, Xianfen Xie, Wanrong Gu

Abstract

Graph convolutional networks (GCNs) have emerged as a powerful tool for recommendation systems. However, existing models often fail to fully exploit attribute interactions, are susceptible to noise in high-order connectivity, and ignore the nuances of user opinions (ratings). This paper introduces the Improved Graph Convolutional Network-based Collaborative Filtering (IGCN-CF) model. By modeling inner and cross-attribute interactions, aggregating second-order collaborative signals, and incorporating user ratings as message weights, IGCN-CF achieves a 2.1% to 4.3% performance improvement over state-of-the-art baselines across three real-world datasets.

Introduction

Collaborative filtering (CF) is a cornerstone of modern recommendation systems. While GCNs have improved CF by capturing high-order relationships in user-item graphs, three significant challenges remain in current implementations:

  1. Attribute Underutilization: Most models use simplistic encodings for node attributes, missing the complex "inner" and "cross" interactions between user and item traits.
  2. Noise Propagation: Deep GCN layers often introduce "noise" from accidental interactions (e.g., a random click), which degrades the quality of node embeddings as it propagates through the network.
  3. Uniform Edge Weighting: Standard GCNs treat all user-item interactions equally, failing to distinguish between a strong "like" (high rating) and a "dislike" (low rating).

The IGCN-CF Framework

To address these gaps, we propose IGCN-CF, a multi-layered framework designed for more nuanced preference modeling.

Architecture of the IGCN-CF model, illustrating the flow from attribute interaction to the final prediction layer.

1. Improved Input Layer

Instead of relying on raw embeddings or one-hot encodings, IGCN-CF constructs "attribute graphs." It employs a Multilayer Perceptron (MLP) to capture inner interactions (e.g., the relationship between a user's age and occupation) and Bi-interaction to capture cross interactions between users and items. These signals are fused using a Gated Recurrent Unit (GRU) to produce a semantically rich initial representation.

2. Enhanced Convolution Layer

The convolution layer introduces two primary innovations to refine the collaborative signal:

  • Second-order Signal Aggregation: Based on the "Triadic Closure" theory—the idea that two people with a common friend are likely to become friends—the model explicitly identifies and aggregates signals from second-order neighbors. A mathematical suppression term is applied to filter out noise from weak or accidental connections.
  • Opinion Integration: Unlike traditional GCNs, IGCN-CF weights the messages passed along edges based on normalized user ratings. This ensures that strong preferences (high ratings) have a greater influence on the final embeddings than weak or negative ones.

3. Nonlinear Prediction Layer

While many models use a simple inner product for final scoring, IGCN-CF utilizes a 2-layer MLP. This allows the model to capture complex, non-linear relationships between the refined user and item embeddings, leading to more accurate recommendations.

Experimental Evaluation

The model was evaluated on three diverse, real-world datasets: MovieLens-1M (enriched with IMDB data), Book-crossing, and Yelp2018.

Performance Comparison

IGCN-CF consistently outperformed state-of-the-art models such as NGCF, GCMC, and GraphRec. Experimental results showed an improvement of 2.1% to 4.3% in Recall@10 and NDCG@10. Notably, the model achieved these results with fewer layers than its competitors, demonstrating the efficiency of its signal aggregation strategy.

Ablation Study

Ablation tests confirmed that each component—attribute interaction, second-order signals, and opinion weighting—contributes significantly to the overall performance. Removing any of these features resulted in a measurable decline in recommendation accuracy, confirming that all three innovations are necessary for the model's success.

Conclusion and Future Work

IGCN-CF represents a significant step forward in GCN-based recommendation by better utilizing attribute data and refining the way collaborative signals are processed. While the model shows superior accuracy, its current implementation involves high memory overhead due to the processing of large adjacency matrices. Future research will focus on implementing node sampling techniques, such as those based on random walks, to improve scalability for industrial-scale applications without sacrificing the benefits of high-order signal processing.

Summary updated

BibTeX
@inproceedings{he2022improved,
  author = {Yichen He and Yijun Mao and Xianfen Xie and Wanrong Gu},
  title = {{An improved recommendation based on graph convolutional network}},
  booktitle = {Research},
  year = {2022},
  doi = {10.21203/rs.3.rs-1737751/v1},
}
Figure