TimeMM: Dynamic Multimodal Recommendation via Spectral Filtering
A new framework called TimeMM addresses the challenge of evolving user preferences in multimodal recommendation systems. Unlike static models, TimeMM treats time as an operator, using parametric temporal kernels to reweight edges on user-item graphs based on interaction recency. This enables fine-grained adaptation to nonstationary dynamics where visual and textual cues dominate at different rates. The approach overcomes limitations of static graphs and coarse temporal heuristics, offering continuous preference modeling.
Key facts
- TimeMM is a time-conditioned spectral filtering framework for dynamic multimodal recommendation.
- It maps interaction recency to parametric temporal kernels that reweight user-item graph edges.
- The framework addresses nonstationary dynamics where different preference factors change at different rates.
- Visual and textual cues can dominate decisions under different temporal regimes.
- Most multimodal recommenders rely on static interaction graphs or coarse temporal heuristics.
- TimeMM instantiates Time-as-Operator for continuous preference evolution modeling.
- The paper is available on arXiv with ID 2604.26247.
- The approach aims to improve user modeling by integrating collaborative signals with heterogeneous item content.
Entities
Institutions
- arXiv