Adaptive Multi-Intent Model Improves Accuracy in Session-Based Recommendation Systems

A newly developed recommendation model known as DC-MISR improves next-item prediction accuracy in session-based recommendation by deeply integrating adaptive multi-intent modeling with global prior guidance, according to an original article published in the International Journal of Machine Learning and Cybernetics. In extensive experimental trials conducted across five public benchmark datasets—Diginetica, Tmall, RetailRocket, Yoochoose 1/4, and Last.FM—the proposed framework achieved the top ranking in eight out of ten tested evaluation settings and finished second in the remaining two settings. The experimental evaluations demonstrate that the DC-MISR architecture delivers particularly strong performance in long user sessions, addressing persistent performance bottlenecks associated with accurately tracking and anticipating dynamically changing user intents.

Session-based recommendation systems operate under the objective of predicting the next interacted item relying solely on a user’s current interaction session. In this recommendation domain, the primary operational challenge centers on the accurate modeling of user intents that shift dynamically as an interaction unfolds. Accurately characterizing these fluctuating intents remains the fundamental requirement for generating timely, context-sensitive item predictions within ongoing interaction sessions.

According to the researchers, existing session-based recommendation methods suffer from two critical deficiencies in their approach to intent modeling. First, current techniques consistently employ a static, fixed number of intent representations across all interaction sessions regardless of their actual duration or complexity. This uniform representation introduces redundant noise into short sessions where user goals are simple and focused, while simultaneously failing to provide sufficient representational capacity to capture the wide variety of diverse interests that emerge across extended, long sessions. Second, the intent identification workflows in conventional models operate without the guidance of global relational knowledge, leaving the intent extraction process highly vulnerable to interference from irrelevant local noise.

To resolve these structural shortcomings, the DC-MISR framework incorporates global prior guidance to anchor the recommendation process in broader contextual relationships among items. The researchers construct and integrate three distinct global graph structures that serve as informational priors: a global sequential graph, a global co-occurrence graph, and a global incompatibility graph. Each graph structure is designed to capture global item relations across different dimensions of interaction. Specifically, the global sequential graph models temporal dependencies across sequential transitions, the global co-occurrence graph captures collaborative relationships reflecting items that appear together, and the global incompatibility graph encodes mutual-exclusion constraints to formally account for items that are incompatible or mutually exclusive.

Guided by these global item relational priors, the framework introduces an adaptive sparsification mechanism designed to overcome the rigidness of traditional fixed intent allocations. The adaptive sparsification mechanism dynamically selects effective intents from a fixed set of candidate intents by evaluating both individual session characteristics and global item relations. This dynamic selection process allows short sessions to retain only their core intents while discarding irrelevant dimensions, and it simultaneously enables long sessions to expand their representational scope to capture diverse user interests. Furthermore, the selected multi-intent representations are deployed in similarity computations for cross-session collaboration, allowing the framework to identify other sessions exhibiting similar intent patterns and incorporate collaborative signals into the recommendation process.

In the final stage of the architecture, DC-MISR incorporates a dual-channel multi-head attention mechanism that adaptively fuses local session representations with global relational information based on session characteristics. This dual-channel mechanism improves the model’s overall adaptability across varied session scenarios, ensuring that local interactions and global graph priors are combined effectively whether a session is short or extensive. Empirical evaluations across the Diginetica, Tmall, RetailRocket, Yoochoose 1/4, and Last.FM datasets validated this design, showing that DC-MISR ranked first in eight settings and second in two settings out of ten total benchmarks. The authors noted that the framework demonstrated especially strong performance in long sessions, confirming the benefit of combining global priors with adaptive multi-intent modeling.

Subject of Research: Session-based recommendation algorithms and dynamic multi-intent user modeling

Article Title: Dynamic context-aware multi-intent session recommendation

Article References: Li, Y., Qin, Z., Ma, R., & Liu, C. (2026). Dynamic context-aware multi-intent session recommendation. International Journal of Machine Learning and Cybernetics, 17(11), Article 523. https://doi.org/10.1007/s13042-026-03343-7

Image Credits: AI Generated

DOI: 10.1007/s13042-026-03343-7

Keywords: session-based recommendation, intent modeling, adaptive sparsification, graph priors, attention mechanisms, machine learning, collaborative filtering, item prediction, Dynamic, context-aware, multi-intent, session