A research team has introduced a hierarchical multi-perspective evolving graph neural network, named HMP-EGNN, designed to improve session-based recommendation systems across modern multimedia platforms. These platforms, which include e-commerce systems, online content portals, video-on-demand services, and music streaming applications, routinely require automated models to predict user engagement when individuals browse anonymously or across brief digital sessions. The investigation focuses on overcoming specific operational constraints inherent in digital multimedia and enterprise information systems, where brief session durations, fast-shifting consumer interests, and sparse interaction histories often degrade recommendation precision.
To capture user behavior under these constraints, the proposed computational architecture constructs both intra-session and inter-session graphs. Generating intra-session graphs allows the model to track immediate, short-term interaction sequences within a single visit, while inter-session graphs capture broader behavioral patterns and cross-session dependencies that emerge across successive user visits over time.
The architecture incorporates Transformer encoders to model temporal user patterns across several distinct operational horizons. Complementing the encoders, the system integrates an adaptive attention mechanism engineered to dynamically blend local graph representations with global graph features.
To address the persistent challenge of interaction sparsity in multimedia datasets, the investigators incorporated a contrastive learning objective into the training framework.
The performance of the HMP-EGNN framework was assessed through extensive empirical experiments conducted on three public benchmark datasets: Diginetica, Tmall, and RetailRocket. In comparative evaluations against established baseline algorithms, the model consistently demonstrated superior performance across standard information retrieval metrics, specifically Recall@20, Mean Reciprocal Rank at twenty positions (MRR@20), and Normalized Discounted Cumulative Gain at twenty positions (NDCG@20).
Based on the experimental outcomes across Diginetica, Tmall, and RetailRocket, the study authors report that the hierarchical evolving graph framework provides an effective, generalizable model architecture for large-scale multimedia recommendation environments. By addressing interaction sparsity and multi-scale temporal dependencies, the method provides a structured approach for processing anonymous user trajectories in modern web services.
Subject of Research: Intelligent multimedia session recommendation using evolving graph neural networks
Article Title: HMP-EGNN: A hierarchical multi-perspective evolving graph neural network for intelligent multimedia session recommendation
Article References: Saffari, V., Bagherifard, K., Parvin, H., Nejatian, S., & Rezaei, V. (2026). HMP-EGNN: A hierarchical multi-perspective evolving graph neural network for intelligent multimedia session recommendation. Multimedia Tools and Applications, 85(11), Article 850. https://doi.org/10.1007/s11042-026-21862-x
Image Credits: AI Generated
DOI: 10.1007/s11042-026-21862-x
Keywords: Session-Based Recommendation, Graph Neural Networks, Multimedia Platforms, Contrastive Learning, Transformer Encoders, Adaptive Attention Mechanism, Interaction Sparsity, Intra-Session Graphs, Inter-Session Graphs, Diginetica Dataset, Tmall Benchmark, RetailRocket Dataset, Recall@20, MRR@20, NDCG@20, Recommender Systems
News Source: Scienmag.
