TimeMM: Time-as-Operator Spectral Filtering for Dynamic Multimodal Recommendation

TimeMM, developed by Wei Yang and colleagues, introduces a novel approach to multimodal recommendation by integrating time as a core operator. Instead of relying on static interaction graphs or coarse temporal heuristics, it uses parametric temporal kernels to reweight edges on user-item graphs based on interaction recency. This allows the model to generate component-specific representations without explicit eigendecomposition, enabling it to capture the continuous evolution of user preferences.

The framework further enhances performance through adaptive spectral filtering, which mixes different operator banks according to temporal context to produce prediction-specific spectral responses. It also introduces spectral-aware modality routing to calibrate the contributions of visual and textual features based on the same temporal context, addressing the challenge that different modalities dominate at different times. A spectral diversity regularization prevents filter-bank collapse, ensuring diverse expert behaviors. Tests on real-world benchmarks show TimeMM consistently outperforms existing state-of-the-art multimodal recommenders, all while maintaining linear-time scalability.