TimeMM: Time-as-Operator Spectral Filtering for Dynamic Multimodal Recommendation Conference: Proceedings of the 49th International ACM SIGIR Conference on Research and Development in Information Retrieval; July 20–24, 2026; Melbourne, VIC, Australia Proceedings of the 49th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR ’26), July 20–24, 2026, Melbourne, VIC, Australia DOI: 10.1145/3805712.3809616 ISBN: 979-8-4007-2599-9/2026/07 CCS: Information systems Recommender systems CCS: Information systems Multimedia and multimodal retrieval
Wei Yang, Rui Zhong, Zihan Lin, Xiaodan Wang, Cheng Chen, Huan Ren, Yao Hu · 2026
Read TimeMM: Time-as-Operator Spectral Filtering for Dynamic Multimodal Recommendation Conference: Proceedings of the 49th International ACM SIGIR Conference on Research and Development in Information Retrieval; July 20–24, 2026; Melbourne, VIC, Australia Proceedings of the 49th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR ’26), July 20–24, 2026, Melbourne, VIC, Australia DOI: 10.1145/3805712.3809616 ISBN: 979-8-4007-2599-9/2026/07 CCS: Information systems Recommender systems CCS: Information systems Multimedia and multimodal retrieval by Wei Yang, Rui Zhong, Zihan Lin, Xiaodan Wang, Cheng Chen, Huan Ren, Yao Hu (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-26247.txt.
Introduction
Yang and six colleagues proposed TimeMM, a recommender that lets time reshape the user-item graph. How recent each link is sets a set of time kernels that reweight the graph edges, and a spectral filtering step blends those kernels by context to fit each guess. A routing step then tunes visual and textual signals to the same moment. On real test sets, it beat top multimodal rivals while keeping run time linear in scale.
Catalog record
| CREATOR | Wei Yang, Rui Zhong, Zihan Lin, Xiaodan Wang, Cheng Chen, Huan Ren, Yao Hu |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 56356 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2604.26247
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
TimeMM: Time-as-Operator Spectral Filtering for Dynamic Multimodal Recommendation Conference: Proceedings of the 49th International ACM SIGIR Conference on Research and Development in Information Retrieval; July 20–24, 2026; Melbourne, VIC, Australia Proceedings of the 49th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR ’26), July 20–24, 2026, Melbourne, VIC, Australia DOI: 10.1145/3805712.3809616 ISBN: 979-8-4007-2599-9/2026/07 CCS: Information systems Recommender systems CCS: Information systems Multimedia and multimodal retrieval Wei Yang, Rui Zhong, Zihan Lin, Xiaodan Wang, Cheng Chen, Huan Ren, Yao Hu Abstract . Multimodal recommendation improves user modeling by integrating collaborative signals with heterogeneous item content. In real applications, user interests evolve over time and exhibit nonstationary dynamics, where different preference factors change at different rates. This challenge is amplified in multimodal settings because visual and textual cues can dominate decisions under different temporal regimes. Despite strong progress, most multimodal recommenders still rely on static interaction graphs or coarse temporal heuristics, which limits their ability to model continuous preference evolution with fine-grained temporal adaptation. To address these limitations, we propose TimeMM , a time-conditioned spectral filtering framework for dynamic multimodal recommendation. TimeMM instantiates Time-as-Operator by mapping interaction recency to a family of parametric temporal kernels that reweight edges on the user–item graph, producing component-specific representations without explicit eigendecomposition. To capture non-stationary interests, we introduce Adaptive Spectral Filtering that mixes the operator bank according to temporal context, yielding prediction-specific effective spectral responses. To account for modality-specific temporal sensitivity, we further propose Spectral-Aware Modality Routing that calibrates visual and textual contributions conditioned on the same temporal context. Finally, a ranking-space Spectral Diversity Regularization encourages complementary expert behaviors and prevents filter-bank collapse. Extensive experiments on real-world benchmarks demonstrate that TimeMM consistently outperforms state-of-the-art multimodal recommenders while maintaining linear-time scalability. The code is available at https://github.com/llm-ml/TimeMM . † † cc-license: by 1. Introduction Multimodal recommendation improves collaborative filtering by incorporating rich item side information, such as visual content and textual descriptions, to alleviate interaction sparsity and strengthen preference understanding ( Wei et al. 2019 ; Valencia et al. 2026 ; Li et al. 2025b ; Kan et al. 2026 ) . By integrating collaborative signals with heterogeneous content features, it can represent not only what users interacted with, but also why an item matches their interests ( Deldjoo et al. 2021 ; Han et al. 2022 ; Yang et al. 2023b ) . In real applications, however, this matching process is inherently time-dependent ( Fan et al. 2021 ; Li et al. 2020 ; Ye et al. 2020 ) . Users often maintain stable semantic interests over long horizons, while their visual tastes can shift over shorter horizons ( Zhang et al. 2025a ; Ye et al. 2025 ) . For example, a user may repeatedly purchase running shoes over an extended period, yet the preferred visual style can evolve from minimalist designs to high-contrast colorways as trends change. These observations suggest that effective multimodal recommendation requires modeling the dynamic evolution of user preferences ( Chen et al. 2022 ; Qin et al. 2024 ) . Despite the evident need for temporal modeling, prevailing multimodal approaches largely overlook the time dimension ( Jiang et al. 2024 ; Karra and Tulabandhula 2024 ; Zhao et al. 2025 ) . State-of-the-art methods such as AlignRec ( Liu et al. 2024c ) , MMIL ( Yang and Yang 2024 ) , and SLMRec ( Tao et al. 2022 ) typically conceptualize user–item interactions as a static bipartite graph. In this view, historical interactions are aggregated indiscriminately, compressing a user’s behavioral timeline into a flat structural snapshot. This temporal agnosticism is problematic. An interaction from years ago is treated as equally informative as one from yesterday, making it difficult to distinguish obsolete interests from emerging intents ( Vaswani et al. 2021 ; Wu et al. 2025 ) . Consequently, static models often struggle with concept drift. They tend to favor signals from a user’s global history, which can misalign with the user’s potential visually driven short-term intent ( Bian et al. 2023 ; Lu and Duan 2024 ) . Bridging this gap calls for a mechanism that injects temporal dynamics into the graph propagation operator itself, rather than appending time as an auxiliary feature, which remains largely underexplored in multimodal recommendation. However, recognizing the need for temporal dynamics is only the first step. The key challenge lies in capturing the multi-scale granularity of preference shifts ( Huang et al. 2022 ; Huang et al. 2023 ; Zhang et al. 2025b ) . User decisions are rarely governed by a single temporal scope. Instead, they reflect a superposition of long-term habits and short-term impulses, spanning stable category loyalty and fleeting responses to viral visual trends. The effective scope varies across users and can change for the same user over time ( Jiang et al. 2020 ; Xia et al. 2022 ) . Moreover, temporal sensitivity is modality-dependent: visual preferences often react rapidly to aesthetics, whereas semantic interests tend to evolve more steadily ( Liu et al. 2023 ; Zhang et al. 2025b ) . This naturally suggests a spectral perspective on graph signals, where different temporal scopes correspond to different smoothing strengths and thus different frequency emphases. To capture such dynamics, a promising direction is to examine the spectral properties of user behaviors ( Xu et al. 2025 ; Zhang et al. 2025b ; Yang et al. 2025e ) . High-frequency signals often correspond to transient and impulse-driven behaviors, whereas low-frequency components reflect stable and enduring habits. Prior studies such as FITMM ( Yang et al. 2025d ) and SMORE ( Ong and Khong 2025 ) introduced spectral graph modeling into this domain. By explicitly decomposing representations into orthogonal frequency bands through spectral transformation or eigendecomposition, these methods can separate informative signals from modality-specific noise ( Fan et al. 2024 ; Huang et al. 2025 ; Kim et al. 2025a ) . However, explicit spectral approaches face a key limitation in dynamic settings. They typically rely on a global decomposition of a static graph snapshot, yielding fixed spectral bases and filtering behaviors that cannot adapt to time-varying spectral shifts ( Kim et al. 2025b ; Liu et al. 2024b ; Luo et al. 2024b ) . As a result, it becomes difficult to determine whether a high-frequency component reflects noise or a genuinely emerging short-term interest. In addition, the required matrix factorization introduces substantial computational overhead, which limits scalability. Finally, existing methods often apply uniform filtering across modalities, even though visual signals can be more sensitive to short-range fluctuations than textual semantics. To bridge static spectral modeling with non-stationary user behavior, we propose TimeMM , a time-conditioned spectral filtering framework for dynamic multimodal recommendation. Our core insight follows the duality between time and frequency: rather than performing costly explicit decomposition, we realize frequency-aware modeling implicitly via Time-as-Operator . Specifically, TimeMM maps interaction recency to a family of parametric temporal kernels that reweight edges on the user–item graph, inducing a lightweight spectral filter bank with distinct smoothing profiles. Short-range kernels emphasize rapidly changing signals, whereas long-range kernels retain stable preference evidence. We then perform Adaptive Spectral Filtering by mixing the operator bank according to user and item temporal context to obtain a prediction-specific effective spectral response. Next, Spectral-Aware