<p>Sequential recommender systems (SRSs) investigate users’ behavioral patterns to ascertain their next favored item. User behavior sequences inherently offer personalized insights; however, extracting interpretable and motivation-oriented behavioral patterns remains a significant challenge. Although attention-based models in SRSs perform well by learning sequential dependencies, they fail to effectively exploit item–attribute relationships and have difficulty in differentiating between users’ short-term preferences and long-term intended actions. Additionally, traditional multi-head self-attention assumes that each head operates independently, which results in redundant representations and limits the model’s ability to capture complex cross-head interactions. To address these limitations, we propose Mixed-Weighted Head Attention Networks (MWHAN), a comprehensive framework for context-aware attribute-based session recommendation that integrates real-time user purpose and contextual attributes within a unified model. Specifically, we develop a Dynamic Purpose-Aware module thats captures the hierarchical relationships between items and their associated attributes, providing insights into how users’ goals evolve. Furthermore, we introduce the Mixed-Weighted Multi-Head Attention Network (MWHAN), in which multiple attention heads share projections and learn mixing vectors, thereby enabling more diverse preference representations while reducing redundancy. Extensive experiments on benchmark datasets demonstrate that MWHAN consistently outperforms recent self-supervised Transformer-based models, while offering meaningful interpretability regarding how user behaviors are driven by underlying motivations.</p>