[Paper Review] Learning Recommender Systems from Multi-Behavior Data.

Chen Gao, Xiangnan He+6 more|arXiv (Cornell University)|Sep 21, 2018

Recommender Systems and Techniques38 references5 citations

TL;DR

This paper proposes NMTR, a neural multi-task recommendation model that jointly learns from multiple user behaviors—such as clicks, views, and purchases—by modeling their cascading dependencies. By leveraging multi-task learning to optimize all behaviors simultaneously, NMTR significantly improves recommendation quality, especially for users with sparse interaction histories, outperforming state-of-the-art methods on real-world datasets.

ABSTRACT

Most existing recommender systems leverage user behavior data of one type only, such as the purchase behavior in E-commerce that is directly related to the business KPI (Key Performance Indicator) of conversion rate. Besides the key behavioral data, we argue that other forms of user behaviors also provide valuable signal, such as views, clicks, adding a product to shop carts and so on. They should be taken into account properly to provide quality recommendation for users. In this work, we contribute a new solution named NMTR (short for Neural Multi-Task Recommendation) for learning recommender systems from user multi-behavior data. We develop a neural network model to capture the complicated and multi-type interactions between users and items. In particular, our model accounts for the cascading relationship among different types of behaviors (e.g., a user must click on a product before purchasing it). To fully exploit the signal in the data of multiple types of behaviors, we perform a joint optimization based on the multi-task learning framework, where the optimization on a behavior is treated as a task. Extensive experiments on two real-world datasets demonstrate that NMTR significantly outperforms state-of-the-art recommender systems that are designed to learn from both single-behavior data and multi-behavior data. Further analysis shows that modeling multiple behaviors is particularly useful for providing recommendation for sparse users that have very few interactions.

Motivation & Objective

• To address the limitation of existing recommender systems that rely on only one type of user behavior, such as purchases.
• To explore how multiple user behaviors—like clicks, views, and cart additions—can provide complementary signals for better recommendation quality.
• To model the inherent cascading relationship among behaviors (e.g., click before purchase) in a unified framework.
• To improve recommendation performance for users with sparse interaction histories by leveraging richer behavioral signals.
• To develop a joint optimization framework using multi-task learning to simultaneously learn from diverse behavioral data types.

Proposed method

• Design a neural network architecture that captures complex, multi-type interactions between users and items across multiple behavior types.
• Model the cascading dependency among behaviors (e.g., view → click → add-to-cart → purchase) using a structured sequence-aware design.
• Treat each behavior type as a separate but interrelated task within a multi-task learning framework to enable joint optimization.
• Use shared representations across tasks to improve generalization while allowing task-specific adaptation for behavior-specific patterns.
• Apply end-to-end training with a unified loss function that combines supervision signals from all behavior types.
• Incorporate attention mechanisms or sequence modeling components to weigh the importance of prior behaviors in predicting downstream actions.

Experimental results

Research questions

• RQ1 Can jointly modeling multiple user behaviors improve recommendation performance compared to single-behavior baselines?
• RQ2 How does modeling the cascading relationship among behaviors affect recommendation accuracy and robustness?
• RQ3 To what extent does multi-behavior learning enhance recommendations for users with limited interaction history?
• RQ4 Does the multi-task learning framework in NMTR lead to better generalization and representation learning than single-task or heuristic multi-behavior approaches?
• RQ5 How do different behavioral signals (e.g., views vs. purchases) contribute to the final recommendation quality in a joint learning setup?

Key findings

• NMTR significantly outperforms state-of-the-art recommender systems that are trained on either single-behavior or multi-behavior data.
• The model achieves superior performance on both dense and sparse user interaction scenarios, with notable gains for users who have few interactions.
• Modeling the cascading nature of behaviors leads to better representation learning and improved prediction accuracy across all behavior types.
• The multi-task learning framework enables effective knowledge transfer between behavior types, enhancing overall system robustness.
• The joint optimization of multiple behaviors leads to more stable and generalizable recommendations compared to training separate models per behavior.
• Empirical results on two real-world datasets confirm that NMTR consistently improves key recommendation metrics such as precision, recall, and NDCG.