BLADE: A Behavior-Level Data Augmentation Framework with Dual Fusion Modeling for Multi-Behavior Sequential Recommendation

TL;DR

The paper addresses multi-behavior sequential recommendation by tackling behavior heterogeneity and data sparsity. It introduces BLADE, a framework that combines dual item-behavior fusion (early and intermediate) with behavior-level data augmentation and a sequence-level contrastive loss. Key contributions include a dual fusion architecture, three behavior-level augmentation techniques (co-occurrence addition, frequency masking, auxiliary flipping), and a behavior-richness loss weighting, all validated on three real-world datasets where BLADE consistently outperforms baselines. The results demonstrate improved modeling of complex user preferences and enhanced robustness to long-tail behaviors, with practical implications for real-world recommendation systems.

Abstract

Multi-behavior sequential recommendation aims to capture users' dynamic interests by modeling diverse types of user interactions over time. Although several studies have explored this setting, the recommendation performance remains suboptimal, mainly due to two fundamental challenges: the heterogeneity of user behaviors and data sparsity. To address these challenges, we propose BLADE, a framework that enhances multi-behavior modeling while mitigating data sparsity. Specifically, to handle behavior heterogeneity, we introduce a dual item-behavior fusion architecture that incorporates behavior information at both the input and intermediate levels, enabling preference modeling from multiple perspectives. To mitigate data sparsity, we design three behavior-level data augmentation methods that operate directly on behavior sequences rather than core item sequences. These methods generate diverse augmented views while preserving the semantic consistency of item sequences. These augmented views further enhance representation learning and generalization via contrastive learning. Experiments on three real-world datasets demonstrate the effectiveness of our approach.

Figures (6)

• Figure 1: Overview of the proposed dual item-behavior fusion architecture, which integrates behavior information at both the input and intermediate layers. For clarity, residual connections and normalization layers are omitted in the illustration.
• Figure 2: Overview of the proposed behavior-level data augmentation methods: (a) Co-occurrence behavior addition adds an extra behavior based on co-occurrence frequency; (b) Frequency-based behavior masking masks behaviors according to their occurrence frequency; (c) Auxiliary behavior flipping removes or adds an auxiliary behavior.
• Figure 3: Performance impact of different augmentation methods and operation ratio $p$.
• Figure 4: Performance comparison of three data augmentation methods based on the long-tail degree of user interaction behaviors on the QK-Article dataset.
• Figure 5: Performance with varying contrastive loss balancing hyperparameter $\lambda$.