Abstract
Collaborative Filtering (CF) using Graph Convolutional Networks (GCNs) often faces challenges with computational depth and over-smoothing. This paper presents GODE-CF, a novel approach that models user-item embedding evolution as a continuous flow using Ordinary Differential Equations (ODEs). By parameterizing the derivative function with GCN layers, GODE-CF achieves higher accuracy and faster convergence than discrete state-of-the-art models.
Introduction
Graph Convolutional Networks (GCNs) have become the standard for Collaborative Filtering (CF), powering modern recommendation systems. However, these models face a significant "Goldilocks" problem regarding depth. Shallow models fail to capture high-order relationships, while deep models suffer from high computational costs and "over-smoothing," where node embeddings become indistinguishable.
Researchers identify the discrete nature of traditional GCN layers as a primary limitation. Instead of viewing embedding refinement as a series of discrete steps, this work proposes treating the process as a continuous evolution over time, governed by Ordinary Differential Equations (ODEs).
Methodology: GODE-CF
The proposed framework, GODE-CF (Graph Neural Ordinary Differential Equations-based Collaborative Filtering), shifts the paradigm from discrete stacks to continuous integration.
Continuous Evolution and ODE Integration
In GODE-CF, the change in user and item embeddings is modeled as a derivative function. The final embedding state is determined by solving an integral from time $t_0$ to $t_1$. This allows the model to capture deep structural information without the overhead of many intermediate discrete layers.
GCN-as-Derivative
The model utilizes one or two GCN layers to parameterize the derivative function itself. This ensures the ODE solver leverages the graph's underlying structure to guide the evolution of embeddings. The authors tested two numerical solvers: the Explicit Euler method and the Fourth-order Runge-Kutta (RK4). Surprisingly, the simpler Euler solver proved more effective, as the complex RK4 solver tended to overfit the sparse and noisy data typical of recommendation graphs.
Experimental Setup
The model was evaluated using the public Amazon Reviews dataset, specifically focusing on four benchmark categories: Beauty, Health, Cell Phones, and Office Products. Following standard protocols, the authors applied "5-core" filtering and a leave-one-out evaluation strategy to ensure robust results.
Key Results and Insights
Superior Accuracy and Efficiency
GODE-CF consistently outperformed competitive baselines, including LightGCN, UltraGCN, and the previous ODE-based model LT-OCF. On the Health dataset, GODE-CF improved upon the strongest GCN baseline (GTN) by 11% in Recall@20.
Furthermore, GODE-CF demonstrated superior training efficiency. On the Beauty dataset, it reached peak performance in just 35 minutes, significantly faster than LightGCN (40 minutes) and LT-OCF (98 minutes). The continuous framework allows the model to estimate final embedding states using information from just 1 or 2 internal GCN layers, bypassing the memory constraints of deep architectures.
The Role of Time ($t$)
In this framework, the integration time $t_1$ serves as a continuous analogue to the number of layers in a traditional GCN. Unlike discrete layers, $t$ can be any non-negative real number, providing researchers with much finer control over the model's effective "depth."
Conclusion and Future Work
GODE-CF represents a significant step forward in making recommendation systems both faster and more accurate. While the model is powerful, the authors note it remains susceptible to overfitting when using overly complex solvers or excessive internal layers. Future research will focus on exploring a wider variety of ODE solvers (such as dopri5), investigating non-linear performance fluctuations across different time intervals, and incorporating additional node attributes like user demographics into the continuous framework.
Summary updated