Multi-Interest Recommendation: A Survey

Abstract: Existing recommendation methods often struggle to model users' multifaceted preferences due to the diversity and volatility of user behavior, as well as the inherent uncertainty and ambiguity of item attributes in practical scenarios. Multi-interest recommendation addresses this challenge by extracting multiple interest representations from users' historical interactions, enabling fine-grained preference modeling and more accurate recommendations. It has drawn broad interest in recommendation research. However, current recommendation surveys have either specialized in frontier recommendation methods or delved into specific tasks and downstream applications. In this work, we systematically review the progress, solutions, challenges, and future directions of multi-interest recommendation by answering the following three questions: (1) Why is multi-interest modeling significantly important for recommendation? (2) What aspects are focused on by multi-interest modeling in recommendation? and (3) How can multi-interest modeling be applied, along with the technical details of the representative modules? We hope that this survey establishes a fundamental framework and delivers a preliminary overview for researchers interested in this field and committed to further exploration. The implementation of multi-interest recommendation summarized in this survey is maintained at https://github.com/WHUIR/Multi-Interest-Recommendation-A-Survey.

• The paper establishes a foundational framework for multi-interest recommendation systems by comprehensively reviewing evolution, extraction methods, and application domains.
• The paper highlights advanced extraction methods like dynamic routing and attention mechanisms to capture multifaceted user interests with greater accuracy.
• The paper addresses challenges such as cold-start and long-tail issues while proposing future research directions including reinforcement learning integrations.

Multi-Interest Recommendation: An In-Depth Examination

The paper "Multi-Interest Recommendation: A Survey" (2506.15284) provides a comprehensive overview of multi-interest recommendation systems, outlining their evolution, current methodologies, and application domains. This survey aims to establish a foundational framework for researchers entering this field and offers insights into the diverse approaches and potential future directions in multi-interest recommendation systems.

Introduction to Multi-Interest Recommendations

Recommendation systems are pivotal in assisting users with decision-making by filtering information based on historical interactions. However, traditional systems often fall short in capturing the complex and multifaceted preferences of users due to the volatility and diversity of user behaviors and the ambiguity of item attributes. Multi-interest recommendation addresses these limitations by extracting multiple interest representations from users' historical interactions, thereby enabling finer-grained preference modeling and more accurate recommendations.

Figure 1: A toy example of movie recommendation, where a user interacts with movies spanning multiple themes or genres, revealing her multiple interests. Also, each movie encompasses several genres and themes.

Initially introduced in 2005 for e-commerce, this concept has gained traction in recent years, exemplified by MIND's dynamic routing mechanism used in processing online traffic in commerce applications. The surge in related publications (Figure 2) indicates the growing interest and development milestones in this research area.

Figure 2: The left line chart illustrates the cumulative number of multi-interest recommendation publications, retrieved from the DBLP platform. The related works increased year by year and reached a total of 172 by March 4, 2025. The size of the circles is proportional to the number of publications for the given year. The middle bar presents the representative works and their citations. The word cloud on the right displays the frequency of keywords extracted from the 172 papers.

Motivations and Aspects of Multi-Interest Modeling

The primary motivations for adopting multi-interest recommendation systems include enhancing the granularity of both user preferences and item attributes, increasing recommendation diversity, and improving the explainability of recommendations.

User-Oriented and Item-Oriented Modeling

The framework of multi-interest recommendations involves both user-oriented and item-oriented approaches. User-oriented modeling extracts multiple interests from behaviors, spatial-temporal aspects, social group information, and interaction types. Complementarily, item-oriented modeling covers facets such as item attributes, reviews, multi-modal data, and domain information (Figure 3).

Figure 3: The diagram of explicit multi-interest modeling aspects. It can be divided into item-oriented multi-interest modeling (top) and user-oriented multi-interest modeling (bottom).

Implicit methods bypass external information, relying solely on historical interactions to model multi-interest representations dynamically.

Methodology Framework

The methodology of multi-interest recommendation systems is bifurcated into multi-interest extractor and aggregator components (Figure 4). Extractors, such as dynamic routing and attention mechanisms, handle interest representation learning, described in Figure 5. Aggregators fuse multiple interests into a coherent recommendation, through either representation or recommendation aggregation strategies (Figure 6).

Figure 4: The framework of multi-interest modeling and recommendation. It includes two main components: an interest extractor and an interest aggregator, circled by the dotted line boxes respectively.

Regularization Techniques

Due to the potential collapse of multi-interest representations during training, regularization techniques are employed to enhance diversity. These include representation regularization through cosine similarity and contrastive learning, and distribution regularization via covariance methods (Figure 7).

Figure 7: The categorization of multi-interest representation regularization for diversity. By increasing the dissimilarity of each representation during optimization, the multiple interest representation is ensured to capture distinct aspects of user behaviors, thereby enhancing the overall performance and effectiveness of the model.

Applications and Datasets

The applications of multi-interest recommendations span across domains such as movie and video recommendations, news personalization, travel services, online shopping, and education (Figure 8). Public datasets like MovieLens and MIND provide the real-world backbone for developing and testing these systems.

Figure 8: Application scenarios of multi-interest recommendation.

Challenges and Future Directions

Multi-interest recommendation systems face challenges in adaptive interest extraction, efficiency in handling large-scale data, and mitigating long-tail and cold-start issues. Future directions include leveraging reinforcement learning, enhancing explainability, and integrating large language and diffusion models into multi-interest frameworks.

Conclusion

This survey defines the landscape of multi-interest recommendation systems, detailing the evolution of methodologies, challenges encountered, and the potential future scope of research. Researchers are poised to further refine these systems for more personalized and efficient application in diverse domains, guided by insights from this comprehensive review.