Abstract
Graph Convolutional Networks (GCNs) are widely used in recommendation systems for their ability to capture high-order relationships. However, standard architectures often suffer from over-smoothing and fail to distinguish between the semantic importance of different hierarchical layers. This paper introduces the Self-adaptive Graph Convolutional Network (SAGCN), which employs a distance-based mechanism to adaptively weight hierarchical information. By calculating the distance between embeddings across layers, SAGCN prevents representation collapse and significantly outperforms state-of-the-art baselines like LightGCN.
Introduction
Collaborative filtering has been revolutionized by GCNs, yet two significant "blind spots" persist in current models. First, homogenized aggregation treats all convolutional layers equally, ignoring that direct neighbors (first hop) and distant nodes (higher hops) carry different semantic weights. Second, many models neglect the structure of the representation space, leading to poor uniformity and alignment. These issues culminate in "over-smoothing," where node embeddings become indistinguishable, reducing the model's predictive accuracy.
The SAGCN Framework
The proposed Self-adaptive Graph Convolutional Network (SAGCN) addresses these limitations by innovating how multi-layer information is fused. The architecture consists of three primary stages:
- Nodal Characterization: Standard neighborhood aggregation captures local graph structures.
- Distance-Aware Hierarchical Aggregation: Instead of simple averaging, SAGCN computes the "distance" between the current layer's representation and the previous one. This distance serves as a proxy for information entropy; a larger distance indicates significant new information is being introduced.
- Self-Adaptive Weighting: These distances are processed through a weighting function to determine the optimal contribution of each hierarchy level. This fine-grained approach allows the model to automatically learn the importance of different "hops" in the graph.
The model utilizes metrics such as Euclidean distance, Cosine distance, or KL divergence to quantify these inter-layer differences, ultimately optimizing the embeddings via Bayesian Personalised Ranking (BPR) loss.
Experimental Evaluation
SAGCN was evaluated on four diverse real-world datasets: Yelp2018 (large-scale reviews), MovieLens 1M (classic movie data), Gowalla (social check-ins), and CiteULike (academic tagging).
Key Results
The model demonstrated consistent performance gains over competitive baselines, including LightGCN, DGCF, SGL, and NCL:
- Yelp2018: Achieved over 5% improvement in Recall and NDCG.
- MovieLens 1M: Saw a 5.58% increase in Recall@10.
- Efficiency: SAGCN achieves these results without the high computational cost of complex attention mechanisms, relying instead on efficient distance calculations.
Discussion and Insights
A core insight of this work is viewing distance as a measure of information surplus. By weighting layers based on their novelty relative to previous layers, SAGCN naturally improves the uniformity of the representation space, preventing nodes from collapsing into a single cluster.
The choice of distance metric also plays a role. While Euclidean distance is generally stable, KL divergence showed pronounced benefits on the CiteULike dataset, suggesting that distribution-based metrics may be superior for specific data distributions.
Conclusion and Future Work
SAGCN provides a robust solution to the over-smoothing problem in GCN-based recommendation systems. While highly effective, future research could explore extending this distance-aware logic to dynamic graphs where user interests evolve over time. Additionally, improving robustness against "noisy" interactions remains an open challenge for structural optimization models.
Summary updated