Multi-Interest Network with Dynamic Routing for Recommendation at Tmall (1904.08030v1)

Abstract: Industrial recommender systems usually consist of the matching stage and the ranking stage, in order to handle the billion-scale of users and items. The matching stage retrieves candidate items relevant to user interests, while the ranking stage sorts candidate items by user interests. Thus, the most critical ability is to model and represent user interests for either stage. Most of the existing deep learning-based models represent one user as a single vector which is insufficient to capture the varying nature of user's interests. In this paper, we approach this problem from a different view, to represent one user with multiple vectors encoding the different aspects of the user's interests. We propose the Multi-Interest Network with Dynamic routing (MIND) for dealing with user's diverse interests in the matching stage. Specifically, we design a multi-interest extractor layer based on capsule routing mechanism, which is applicable for clustering historical behaviors and extracting diverse interests. Furthermore, we develop a technique named label-aware attention to help learn a user representation with multiple vectors. Through extensive experiments on several public benchmarks and one large-scale industrial dataset from Tmall, we demonstrate that MIND can achieve superior performance than state-of-the-art methods for recommendation. Currently, MIND has been deployed for handling major online traffic at the homepage on Mobile Tmall App.

Multi-Interest Network with Dynamic Routing for Recommendation at Tmall

The paper introduces a novel approach for enhancing recommender systems, specifically targeting the matching stage in large-scale e-commerce settings, such as Tmall. The proposed solution, termed the Multi-Interest Network with Dynamic Routing (MIND), aims to effectively model and represent the diverse interests of users beyond the conventional single-vector representations utilized by many deep learning-based recommender systems.

Overview of the Proposed System

The MIND framework is structured to address the limitations of traditional user representation in recommender systems by employing multiple embeddings to encapsulate various facets of user interests. This is achieved through a multi-interest extractor layer that utilizes a capsule routing mechanism, inspired by the dynamic routing approach used in capsule networks. This method enables clustering of user historical behaviors and extraction of diverse interests in a manner that contrasts sharply with existing approaches that rely on singular or static representations of user interests.

Core Components and Methodology

• Multi-Interest Extractor Layer: This layer is the cornerstone of MIND, using a modified version of dynamic routing to lever multiple vector representations for user interests. This soft-clustering of user behavior data allows for the encapsulation of distinct interest profiles.
• Label-Aware Attention Mechanism: To refine the user representation, the system employs a label-aware attention layer, which helps align and tune the multiple interest vectors in accordance with the target items’ features.
• Operational Deployment: The paper highlights the practical application of MIND at Tmall, demonstrating a significant improvement in Click-Through Rates (CTR) compared to baseline models such as item-based collaborative filtering and the YouTube DNN model.

Evaluation and Results

The efficacy of the MIND model is validated through extensive experiments on both public benchmark datasets and a large-scale proprietary dataset from Tmall. Notably, MIND outperforms its baselines on key metrics such as Hit Rate at top-N recommendations, confirming the effectiveness of its multi-vector user representation approach. In practical applications, MIND has been deployed successfully to serve billion-scale users, illustrating its scalability and operational feasibility.

Implications and Future Directions

The development of MIND has significant implications for both theoretical and practical aspects of recommender systems. By effectively capturing diverse user interests with multiple embeddings, MIND sets a precedent for future architectures in the recommendation domain, especially in systems catering to large user bases with varied interaction histories. The authors propose future exploration into incorporating temporal dynamics of user behavior and optimizing initialization schemes for dynamic routing, which could further enhance model robustness and accuracy.

In conclusion, while MIND offers clear advancements in terms of accuracy and scalability over existing models, its potential to reshape methodologies in recommendation systems lies in its flexible, dynamic modeling of user interests—a feature increasingly vital in today’s expansive e-commerce environments. The ability to represent users through multi-dimensional interest profiles not only enriches the matching stage but also paves the way for developing more nuanced, user-centered recommendation strategies in the future.