Controllable Multi-Interest Framework for Recommendation (2005.09347v2)

Abstract: Recently, neural networks have been widely used in e-commerce recommender systems, owing to the rapid development of deep learning. We formalize the recommender system as a sequential recommendation problem, intending to predict the next items that the user might be interacted with. Recent works usually give an overall embedding from a user's behavior sequence. However, a unified user embedding cannot reflect the user's multiple interests during a period. In this paper, we propose a novel controllable multi-interest framework for the sequential recommendation, called ComiRec. Our multi-interest module captures multiple interests from user behavior sequences, which can be exploited for retrieving candidate items from the large-scale item pool. These items are then fed into an aggregation module to obtain the overall recommendation. The aggregation module leverages a controllable factor to balance the recommendation accuracy and diversity. We conduct experiments for the sequential recommendation on two real-world datasets, Amazon and Taobao. Experimental results demonstrate that our framework achieves significant improvements over state-of-the-art models. Our framework has also been successfully deployed on the offline Alibaba distributed cloud platform.

• The paper introduces ComiRec, a framework that extracts multiple distinct interest vectors from user behavior to improve recommendation accuracy.
• It employs dynamic routing and self-attentive mechanisms along with a controllable module to balance precision and diversity in recommendations.
• Empirical results on Amazon and Taobao datasets and successful deployment on Alibaba’s cloud platform validate its scalability and superior Recall and NDCG performance.

Controllable Multi-Interest Framework for Recommendation

The paper presents a novel approach to improving recommendations in e-commerce by addressing the multi-interest nature of user behavior. Current models often rely on a single user embedding derived from user behavior sequences, which can be insufficient for capturing the diverse interests a user may have over time. To overcome this limitation, this research introduces ComiRec, a controllable multi-interest framework that harnesses deep learning techniques to extract multiple interests from user behavior data.

Core Contributions

1. Multi-Interest Extraction: Unlike traditional methods that generate a unified user embedding, ComiRec employs a multi-interest module capable of capturing multiple distinct interest vectors from a user's interaction history. This is achieved through either a dynamic routing method akin to capsule networks or a self-attentive mechanism, allowing for nuanced interest representation and improved recommendation specificity.
2. Aggregation with Controllability: ComiRec includes an aggregation module that uses a controllable factor to balance between recommendation accuracy and item diversity. This module can adapt the level of interest diversity in recommendations, allowing practitioners to fine-tune the trade-off between strict accuracy and a more exploratory, diverse recommendation set.
3. Performance Validation on Large Datasets: The framework is empirically tested on two substantial datasets from Amazon and Taobao, demonstrating significant improvements over existing state-of-the-art methods. Notably, ComiRec achieves superior Recall and NDCG scores, emphasizing its effectiveness in practical industrial scenarios.
4. Deployment and Scalability: The framework has been successfully deployed on Alibaba's distributed cloud platform, showcasing its scalability and effectiveness on a billion-scale dataset, validating its industrial applicability.

Experimental Insights

• Robustness Across Scenarios: ComiRec's ability to handle diverse and large-scale datasets makes it suitable for real-world applications where user interests are complex and multifaceted.
• Parameter Sensitivity: The experiments reveal that ComiRec's performance varies with the number of interest vectors, suggesting that model tuning is essential for optimal results depending on the specific dataset characteristics and application needs.
• Impact of Control Factor: By adjusting the control factor, the framework allows stakeholders to manage the diversity-accuracy trade-off tailored to business goals, such as enhancing exploration while maintaining the precision of recommendations.

Implications and Future Work

This paper's framework has potential implications for enhancing user satisfaction and engagement in recommendation systems by acknowledging and leveraging the variability and richness of user interests. The paper hints at the possibility of integrating memory networks and cognitive theories to further capture evolving user behaviors, thus paving the way for more sophisticated, personalized recommendation mechanisms. Future research may explore these avenues to refine and expand upon the foundational work laid out by ComiRec, examining the impact of dynamic adjustments and cognitive insights in recommendation contexts.