Abstract
Traditional Collaborative Filtering (CF) models based on Multi-Layer Perceptrons (MLPs) often suffer from "catastrophic forgetting" when updated with new data, where global weight updates overwrite previously learned patterns. This paper introduces CF-KAN, a novel autoencoder-based architecture that leverages Kolmogorov-Arnold Networks (KANs). By placing learnable activation functions on the edges rather than nodes, CF-KAN achieves superior stability and plasticity. Experimental results on benchmark datasets like MovieLens-1M, Yelp, and Anime demonstrate that CF-KAN outperforms state-of-the-art baselines in both accuracy and training efficiency while providing enhanced interpretability.
Introduction
Modern recommender systems frequently utilize MLPs for collaborative filtering. However, MLPs are prone to catastrophic forgetting because they use fixed activation functions on nodes and update weights globally. In the sparse world of user-item interactions, these global updates can drown out subtle signals from past interactions. The challenge lies in balancing "plasticity" (learning new information) with "stability" (retaining old knowledge). The authors propose CF-KAN to address this stability-plasticity dilemma.
Methods: The CF-KAN Architecture
CF-KAN addresses the limitations of MLPs by replacing them with Kolmogorov-Arnold Networks. Unlike standard neural networks that apply fixed activation functions (e.g., ReLU) to weighted sums at nodes, KANs utilize learnable nonlinear activation functions on the edges (weights), typically implemented using B-splines.
This architectural shift allows for localized updates; learning a new interaction only affects specific spline coefficients rather than the entire network's weights. CF-KAN is structured as an autoencoder, mapping user interaction vectors to a latent space and reconstructing them. This design facilitates faster training and allows for model pruning via $L_1$ and entropy regularization, which enhances interpretability by tracing recommendations back to specific historical interactions.
Experimental Evaluation
The authors evaluated CF-KAN on three benchmark datasets: MovieLens-1M, Yelp, and the large-scale Anime dataset (7.8M interactions). To test continual learning, data was split chronologically into a base block followed by five incremental update blocks.
Key Results and Insights
- Superior Accuracy: CF-KAN outperformed 13 state-of-the-art baselines, showing gains of up to 8.2% in Recall@20 compared to the best competitors.
- Mitigating Forgetting: CF-KAN achieved a 31.3% improvement in Retained Average (RA) compared to MLP-based models, maintaining high performance across incremental data updates.
- The "Shallow" Advantage: Interestingly, a single-layer KAN ($L=1$) performed best, suggesting that the complex edge-level functions in a shallow KAN are sufficient to capture user preference compositions.
- Efficiency at Scale: On the Anime dataset, CF-KAN demonstrated remarkable training speed and low memory overhead.
As shown in the figure above, CF-KAN trains in approximately 5.4 seconds per epoch, whereas graph-based competitors like LightGCN and SGL require thousands of seconds (4828s and 18711s respectively). This efficiency is attributed to the autoencoder design, which avoids the expensive pairwise optimization required by graph methods.
Conclusion
CF-KAN represents a significant step forward in collaborative filtering by successfully applying Kolmogorov-Arnold Networks to solve the stability-plasticity dilemma. By moving activation functions to the edges, the model gains localized update capabilities that prevent catastrophic forgetting. Future work will explore the application of KANs in more complex scenarios, such as cross-domain and multi-modal recommendation systems where forgetting is an even larger hurdle.
Summary updated