Abstract
Graph Convolutional Networks (GCNs) have become a cornerstone of modern recommendation systems due to their ability to capture collaborative signals. However, existing models often suffer from information loss due to complex hypergraph constructions or over-engineering with components borrowed from computer vision. We present the Low-pass Graph Convolutional Network (LGCN), a streamlined model that treats user-item interactions as a 1D signal on a bipartite graph. By employing a trainable spectral filter that retains low-frequency user preferences while discarding high-frequency noise, LGCN achieves significant performance gains over state-of-the-art models like LightGCN and LCFN.
Introduction
The primary challenge in graph-based recommendation is effectively capturing collaborative signals from user-item interaction data. While GCNs are adept at this, current approaches face two major hurdles:
- Information Loss: Models like LCFN utilize 2D convolutions on hypergraphs constructed from 2-hop connections, which often ignore direct 1-hop user-item interactions, leading to significant data loss.
- Over-Engineering: Many GCNs include non-linear activations and complex feature transformations that are suboptimal for the sparse nature of recommendation tasks, where "lighter" models frequently perform better.
LGCN addresses these by simplifying the convolution process and focusing on the spectral properties of user preferences.
Methodology
LGCN introduces three core innovations to refine the graph convolution process:
1D Bipartite Graph Convolution
Unlike previous models that treat users and items as separate dimensions (2D), LGCN treats the bipartite graph as a single homogeneous graph. This 1D approach allows the model to leverage direct 1-hop connections, which provide the most critical collaborative signals.
1D Low-pass Collaborative Filter (LCF)
In the spectral domain, user preferences are identified as low-frequency signals, whereas random interactions constitute high-frequency noise. LGCN utilizes a "gate function" filter that retains only the top $\Phi$ low-frequency components. This acts as a sharp cutoff, providing a cleaner signal than the "leaky" low-pass filters found in standard GCNs. Training is made efficient via the Lanczos algorithm, which allows the model to compute only the necessary eigenvectors.
Structural Optimization
The architecture is intentionally "light." Through ablation studies, the authors removed unnecessary components such as feature transformations and non-linear activation functions (e.g., ReLU). The final model employs sum pooling and an inner product for prediction, ensuring efficiency and preventing over-fitting.
Key Results
LGCN was evaluated on the Amazon and Movielens datasets against baselines including MF, GCMC, NGCF, and LightGCN.
- Accuracy: On the Amazon dataset, LGCN improved the $F_1$-score by 9.38% and NDCG by 8.35% over LightGCN.
- Superiority over LCFN: LGCN outperformed LCFN by 12.57% on Amazon, validating the 1D bipartite approach over 2D hypergraphs.
- Stability: As shown in the sensitivity analysis, LGCN maintains stable performance even as the number of layers ($L$) increases, avoiding the common "over-smoothing" problem that plagues many deep GCNs.
Discussion and Future Work
The core insight of LGCN is the formalization of user preference as a low-frequency signal. By "denoising" the interaction matrix in the spectral domain, the model achieves high accuracy with a simpler architecture.
Future research will explore the integration of side information (e.g., user demographics or item descriptions) and the applicability of LGCN to other graph domains like knowledge graphs or social networks. Additionally, the authors aim to investigate further adaptations of CNN strategies for 1D spectral graph tasks.
Summary updated