Abstract
Collaborative Filtering (CF) signals are crucial for a Recommender System (RS) model to learn user and item embeddings. High-order information can alleviate the cold-start issue of CF-based methods, which is modeled through propagating the information over the user-item bipartite graph. Recent Graph Neural Networks (GNNs) propose to stack multiple aggregation layers to propagate high-order signals. However, there are three challenges, the oscillation problem, varying locality of bipartite graphs, and the fixed propagation pattern, which spoil the ability of the multi-layer structure to propagate information. In this paper, we theoretically prove the existence and boundary of the oscillation problem, and empirically study the varying locality and layer-fixed propagation problems. We propose a new RS model, named as Deoscillated adaptive Graph Collaborative Filtering (DGCF), which is constituted by stacking multiple CHP layers and LA layers. We conduct extensive experiments on real-world datasets to verify the effectiveness of DGCF. Detailed analyses indicate that DGCF solves oscillation problems, adaptively learns local factors, and has layer-wise propagation patterns.
1. Introduction
Standard Graph Neural Networks (GNNs) for recommendation often struggle with "oscillation"—where information bounces back and forth between users and items across layers due to the bipartite nature of interaction data. Most modern recommender systems represent user-item interactions as a bipartite graph to capture high-order signals (e.g., "users who liked this also liked that") by stacking multiple GNN layers. However, this approach faces three critical flaws: the oscillation problem, varying locality, and fixed propagation patterns.
Deoscillated Adaptive Graph Collaborative Filtering (DGCF) solves these issues by introducing cross-hop propagation and locality-adaptive layers, allowing the model to learn more stable, high-order representations that significantly outperform existing state-of-the-art methods.
2. The Oscillation Problem in Bipartite Graphs
In a bipartite graph, users only connect to items, and items only to users. This structural constraint means that at layer 1, a user node only receives information from items; at layer 2, it receives information from other users; at layer 3, it returns to items. This "parity" creates a gap in information flow that prevents the model from converging to a stable stationary distribution. Furthermore, existing models use the same aggregation pattern at every layer, leading to massive redundancy where information is sent to a neighbor and immediately returned in the next hop.
3. Methodology: DGCF Architecture
The proposed DGCF model consists of two novel components designed to break the limitations of bipartite propagation:
- Cross-Hop Propagation (CHP) Layer: Instead of just looking at direct neighbors (1-hop), the CHP layer aggregates information from both direct neighbors and "cross-hop" neighbors (2-hops). By including 2-hop neighbors (users-to-users and items-to-items), the bipartite graph effectively becomes a "regular" graph, which eliminates the oscillation problem.
- Locality-Adaptive (LA) Layer: Before propagation, this layer learns a specific "influencing factor" ($\alpha$) for every node at every layer. This weight is learned via a sigmoid function and allows the model to adaptively control how much information a node should propagate based on its local structural density.
4. Experiments and Results
The model was evaluated on four widely used benchmark datasets: ML100K, ML1M, Amazon (Movies and TV), and Gowalla. DGCF consistently outperformed all baselines, including popular models like LightGCN and NGCF. For instance, on the ML1M dataset, DGCF achieved a Recall@20 of 0.3075, a significant improvement over LightGCN's 0.2888.
Key Insights
- Stability: While models like GCN and NGCF show performance (Recall) that fluctuates as more layers are added, DGCF maintains a much smoother, more stable performance curve across depths.
- Density Adaptation: The LA layer learns to assign higher weights to nodes in lower-density areas. This helps alleviate the "cold-start" problem by ensuring sparse nodes contribute meaningful signals. There is a strong positive correlation (up to $\rho = 0.835$ for items) between the reciprocal of the learned weight ($1/\alpha$) and node density.
- Evolutionary Propagation: Unlike fixed-pattern GNNs, DGCF's propagation pattern evolves. In early layers, it spreads information widely; in deeper layers, it concentrates on critical substructures, reducing redundant noise.
5. Conclusion
DGCF provides a robust theoretical and empirical framework for deoscillation in graph-based recommendation. By addressing the inherent structural limitations of bipartite graphs through CHP and LA layers, it achieves superior performance and stability. Future work includes exploring the computational overhead of cross-hop matrices for extreme-scale graphs and adapting these techniques to dynamic or temporal recommendation settings.
Summary updated