CAMMSR: Category-Guided Attentive Mixture of Experts for Multimodal Sequential Recommendation

There is no available information in the provided context about a system named "CAMMSR: Category-Guided Attentive Mixture of Experts for Multimodal Sequential Recommendation." Therefore, it cannot be confirmed or discussed based on the given sources. However, related research in multimodal sequential recommendation addresses similar challenges.

Several recent approaches focus on improving multimodal fusion in sequential recommendation systems by moving beyond heuristic or fixed fusion strategies. For instance, the Mixture of Conditional Attention (MOCA) framework learns diverse and flexible attention patterns directly from data, utilizing conditional attention and a mixture-of-experts approach to better capture cross-modal relationships and dynamic user preferences . Similarly, the Hierarchical time-aware Mixture of experts for multi-modal Sequential Recommendation (HM4SR) introduces a two-level MoE structure—Interactive MoE and Temporal MoE—to extract interest-relevant information and model dynamic user interests using explicit temporal signals .

Other works emphasize adaptive fusion mechanisms. The MMSR model proposes a graph-based method with dual attention to adaptively determine fusion order, enabling flexible integration ranging from early to late fusion based on modality interdependence . Meanwhile, models like MR-CSAF employ cross-self-attention fusion and adaptive modality selectors to dynamically adjust modality weights, improving robustness under data imbalance and capturing complex inter-modal interactions .

Additionally, large language model (LLM)-based approaches, such as MLLM-SRec, leverage multimodal prompts and instruction tuning to unify visual and textual understanding, generating comprehensive item summaries and modeling user preferences more effectively . These advancements collectively aim to overcome the limitations of predefined fusion rules and enhance personalization in multimodal recommendation systems.

The paper explores the challenge of Multimodal Sequential Recommendation (MSR), where systems must predict user preferences by leveraging both historical interaction patterns and rich item content such as images and text. It identifies a critical bottleneck in existing approaches: the reliance on heuristic fusion strategies—such as simple concatenation or element-wise addition—to combine sequential dynamics with multimodal features. These static methods fail to adaptively weigh the importance of different modalities across varying contexts, often leading to suboptimal performance due to an inability to filter out noise or capture complex, non-linear dependencies between user behavior and item attributes.

To address these limitations, the authors propose CAMMSR, a novel architecture that introduces a Category-Guided Attentive Mixture of Experts (MoE) framework. The core innovation lies in decoupling the modeling of user preferences into distinct components—specifically sequential affinity and multimodal affinity—which are processed by specialized expert networks. The system utilizes item category information as a semantic guide to govern an attentive gating mechanism, which dynamically routes inputs to the most relevant experts based on the current context. This allows the model to selectively attend to specific modalities (visual or textual) depending on the item category, effectively learning a data-dependent fusion strategy rather than relying on fixed, heuristic rules.

This research matters significantly because it offers a robust solution to the semantic gap and noise often prevalent in raw multimodal data within recommendation systems. By enabling adaptive, category-aware fusion, CAMMSR achieves superior recommendation accuracy compared to state-of-the-art baselines, demonstrating that expert routing can better capture the nuanced heterogeneity of user interests. The work highlights the value of incorporating structural metadata (categories) to refine the processing of unstructured data (images and text), providing a more scalable and effective pathway for personalized content discovery.

# CAMMSR: Category-Guided Attentive Mixture of Experts for Multimodal Sequential Recommendation

This paper introduces CAMMSR, a novel framework for multimodal sequential recommendation that leverages text and image modalities to enhance personalized content discovery. Traditional sequential recommenders often rely on heuristic fusion methods (e.g., concatenation or averaging) to combine multimodal signals, which can lead to suboptimal performance due to their inability to adaptively weigh or integrate complementary information. CAMMSR addresses this limitation by employing a category-guided attentive mixture-of-experts (MoE) architecture, where different experts specialize in processing distinct item categories (e.g., news, products, or entertainment) while a gating mechanism dynamically selects and fuses relevant experts based on context.

The key contributions of CAMMSR include: 1. Modality-Aware Expert Specialization: Each expert in the MoE is trained to focus on specific item categories, improving the model’s ability to capture fine-grained multimodal interactions (e.g., aligning textual descriptions with visual features for fashion items). 2. Attentive Gating Mechanism: A learned gating network dynamically weights expert contributions, ensuring that the most relevant modalities and categories influence recommendations for a given user sequence. 3. Empirical Validation: Experiments on benchmark datasets (e.g., Amazon Reviews and MMIMN) demonstrate that CAMMSR outperforms baseline methods, including pure text-based, image-based, and heuristic fusion approaches, in terms of hit rate (HR) and normalized discounted cumulative gain (nDCG).

This work is significant because it provides a scalable and adaptable solution for multimodal sequential recommendation, particularly in domains where items have rich multimodal attributes (e.g., e-commerce, social media). By moving beyond static fusion strategies, CAMMSR enables more nuanced, context-aware recommendations, which can improve user engagement and satisfaction in real-world applications.

Source: [arXiv:2603.04320](https://arxiv.org/abs/2603.04320)