Open Access

ARTICLE

Diverse Behavior Path Graphs for Multi-Behavior Recommendation

Qian Hu, Lei Tan^*, Qingjun Yuan, Zong Zuo, Yan Li

Henan Key Laboratory of Cyberspace Situation Awareness, Zhengzhou, China

* Corresponding Author: Lei Tan. Email:

Computers, Materials & Continua 2026, 87(3), 22 https://doi.org/10.32604/cmc.2026.076137

Received 14 November 2025; Accepted 15 January 2026; Issue published 09 April 2026

Abstract

Multi-behavior recommendation methods leverage various types of user interaction behaviors to make personalized recommendations. Behavior paths formed by diverse user interactions reveal distinctive patterns between users and items. Modeling these behavioral paths captures multidimensional behavioral features, which enables accurate learning of user preferences and improves recommendation accuracy. However, existing methods share two critical limitations: (1) Lack of modeling for the diversity of behavior paths; (2) Ignoring the impact of item attribute information on user behavior paths. To address these issues, we propose a Directed Behavior path graph-based Multi-behavior Recommendation method (DBMR). Specifically, we first construct a directed user-item behavior path graph based on diverse behavior chains. For each behavior, we then build a user-item interaction graph and use a LightGCN model with residual design to learn user and item embeddings. Next, we introduce a graph attention message aggregator that integrates features from previous behaviors into the learning of the next behavior, weighted by the transition strength between behaviors. Finally, we compute the recommendation score from the user preference and item representations under the target behavior. We adopt a joint optimization framework with a multi-task learning strategy, which accounts for each auxiliary behavior’s contribution to target behavior prediction. Additionally, an auxiliary loss measures the difference between item embeddings from behavior paths and those from an attribute-feature encoder, thereby capturing multidimensional item features and refining recommendation results. Experiments on two real-world datasets demonstrate the effectiveness of our method in utilizing multi-behavior data.

---

Keywords

---