ASPIRE: Make Spectral Graph Collaborative Filtering Great Again via Adaptive Filter Learning

Researchers from China and Australia have introduced ASPIRE, a new adaptive spectral graph collaborative filtering framework that addresses a fundamental flaw in traditional recommendation systems. The team identified a spectral phenomenon called "low-frequency explosion," where conventional objectives overemphasize low-frequency signals, hindering the learning of effective graph filters. ASPIRE uses a bi-level optimization objective to disentangle filter learning from the recommendation task, enabling fully adaptive filters without manual hyperparameter tuning. This theoretical innovation translates into practical gains: ASPIRE matches the performance of carefully engineered, task-specific filter designs across diverse datasets.

Extensive experiments confirm ASPIRE's effectiveness in both standard collaborative filtering and LLM-powered scenarios, demonstrating its generalizability. The framework achieves excellent recommendation performance, spectral adaptivity, and training stability. By making graph filter learning viable and generalizable, ASPIRE paves the way for more expressive graph neural networks in recommendation systems. This work represents a significant step forward in spectral collaborative filtering, potentially reducing the need for laborious manual filter engineering in production systems.