Hierarchical Graph Neural Network Improves Recommendation Accuracy Across Multiple Temporal Scales

A newly designed Hierarchical Time-aware Graph Neural Network, known as HTGNN, has significantly improved recommendation accuracy across diverse online domains by structuring user-item interactions across multiple temporal granularities. Developed to address the rapid fluctuations and enduring habits that characterize user behavior, the architecture constructs a multi-layered representation of evolving preferences rather than relying on flat or static temporal assumptions. In comprehensive experimental evaluations spanning five public benchmark datasets—MovieLens, Amazon, Houses, Last.FM, and Retailrocket—the HTGNN framework consistently surpassed established baseline architectures, achieving superior ranking quality and classification performance. By systematically bridging the gap between multi-scale temporal modeling and hierarchical graph learning, the model demonstrates that incorporating both high-resolution short-term shifts and broad historical trajectories substantially refines top-K recommendation lists. The research establishes that dynamic preference modeling benefits from explicitly embedding temporal dependencies directly into graph structures, enabling recommender systems to deliver more relevant suggestions as interaction patterns evolve over continuous operational timelines.

Modern recommender systems have emerged as vital components across electronic commerce platforms, entertainment streaming services, and online applications, where personalizing item suggestions directly impacts user engagement. However, traditional modeling paradigms frequently fail to accommodate the continuously shifting nature of consumer preferences and content relevance over time. A core technical challenge within time-aware recommendation involves modeling non-linear, complex relationships between users, catalog items, and timestamped activities. Real-world temporal interaction data is inherently noisy, sparse, and susceptible to sudden behavioral transitions triggered by viral trends or seasonal cycles. Furthermore, the operational time scales of modern recommendation environments vary extensively across domains, with certain systems demanding minute-by-minute behavioral updates while others function effectively along daily or weekly aggregation intervals. Conventional recommendation methods historically addressed this challenge by treating time as an invariant static attribute or applying rudimentary decay functions, both of which fail to capture the multi-layered spectrum of human behavioral evolution. Graph-based architectures successfully represent intricate relational connections among users and items, yet existing graph approaches predominantly overlook temporal variations or restrict analysis to a single temporal resolution.

To situate the HTGNN framework within current scientific literature, the researchers contrasted its architectural design against prevailing graph neural networks and dynamic recommendation models. Foundational architectures such as Graph Convolutional Networks and GraphSAGE demonstrate substantial capacity for learning structural relationships across user-item interaction graphs, yet they operate primarily upon static topology without temporal reasoning. Conversely, specialized temporal graph approaches such as Temporal Graph Networks, or TGN, integrate continuous-time updates but lack structured multi-scale hierarchical mechanisms. Scalable paradigms like LightGCN simplify message passing across structural layers, but they treat historical connections uniformly over time, neglecting the distinction between immediate user actions and distant historical engagements. While contemporaneous research efforts like DR-GNN and temporal distributionally robust optimization emphasize resilience against distributional and feature shifts in cold-start scenarios, HTGNN pursues a complementary focus by encoding multi-scale temporal patterns directly. Furthermore, although models such as CM-GNN apply contrastive multi-level strategies for session dynamics, BoxGNN utilizes box embeddings for tag diversity, and Ripple-KGCN aggregates multi-hop knowledge graphs, HTGNN uniquely unifies multi-scale encoding, attention, and hierarchical aggregation.

The technical foundation of the HTGNN framework begins with a dynamic graph representation layer that structures the interaction space. Within this dynamic network, nodes explicitly designate individual users and items, while edges represent recorded user-item interactions enriched with precise chronological timestamps. Rather than treating time as an auxiliary metadata tag appended after graph formation, the architecture embeds temporal dependencies directly into the underlying graph topology, enabling the system to track structural and behavioral evolution simultaneously. Building upon this graph foundation, the model implements a dedicated multi-scale temporal encoding mechanism designed to process timestamps across a broad spectrum of temporal granularities. By discretizing chronological data into distinct operational windows spanning minutes, hours, days, and months, HTGNN produces rich temporal embeddings. These mathematical representations capture the contextual significance of interactions relative to when they occurred. Consequently, the temporal encoder preserves both the immediate operational context of recent actions and the broader contextual background of older interactions, establishing a multi-resolution basis for subsequent recommendation calculations.

Beyond multi-scale encoding, HTGNN incorporates a specialized temporal attention mechanism and hierarchical temporal aggregation layers to synthesize temporal dynamics. The temporal attention mechanism functions by dynamically allocating computational attention weights across distinct time intervals, evaluating the contextual relevance of each historical event relative to the current recommendation target. This adaptive weighting enables the model to amplify short-term interactions when a user exhibits rapid shifts in taste, while elevating long-term patterns when steady habits dominate. Working alongside attention, the hierarchical temporal aggregation component partitions user-item interactions across multiple temporal resolutions instead of aggregating data uniformly. Individual hierarchical layers specialize in extracting behavioral dependencies at specific operational frequencies, from short-term minutes and hours to extended days and months. The aggregation framework then synthesizes these layered representations, balancing the volatile influence of temporary interests against the foundational stability of enduring historical habits. Through this integrated multi-stage processing pipeline, HTGNN generates expressive, time-sensitive graph representations that adjust fluidly to dynamic user behavior.

To validate the architecture empirically, the authors conducted extensive experiments across five public benchmark datasets representing varied application domains. These collections included MovieLens for cinematic ratings, Amazon for product reviews, Houses for real estate browsing activity, Last.FM for musical consumption histories, and Retailrocket for commercial e-commerce event streams. The experimental protocol utilized four standardized performance metrics: Normalized Discounted Cumulative Gain, Precision, Recall, and the harmonic F1-score, all computed under top-K ranking constraints. NDCG evaluates the quality of ranked recommendation lists by penalizing relevant items placed lower in the ranking order through logarithmic discounting. Precision measures the proportion of recommended items within the top-K list that match verified user interactions, whereas Recall quantifies the proportion of all relevant items successfully retrieved. The F1 metric combines Precision and Recall into a balanced harmonic mean to evaluate overall trade-offs. By conducting full dataset evaluations rather than subsampled approximations, the experimental framework ensured an unbiased assessment of model capability across both ranking accuracy and classification coverage.

Experimental comparisons against recurrent neural networks, temporal graph networks, and GraphSAGE demonstrated clear performance advantages for the HTGNN framework across all examined datasets. In top-10 ranking assessments, HTGNN recorded an NDCG@10 score of 97.4 percent on the MovieLens benchmark, 89 on the Amazon collection, 72 percent on the Houses dataset, 88 percent on Last.FM, and 76.3 percent on Retailrocket. These empirical measurements reflect marked improvements over baseline models that lack multi-scale temporal structuring. The advantage of hierarchical time-aware processing proved particularly pronounced in scenarios characterized by rapid interaction shifts and high user volatility. In the Retailrocket dataset, where user behaviors fluctuate rapidly within short browsing sessions, HTGNN attained a leading F1-score of 50 percent. This outcome illustrates the model’s robustness in adapting to swift behavioral transitions, confirming that combining multi-scale temporal features allows recommender architectures to maintain ranking fidelity even when interaction patterns shift unpredictably over compressed observation windows.

To confirm the reliability of these observed performance gains, the researchers conducted statistical significance testing using paired t-tests across the evaluation runs. The analysis yielded a p-value of 0.04 when comparing HTGNN against TGN, and 0.02 against recurrent neural networks, establishing that the performance advantages over these two baselines are statistically significant at the 95 percent confidence level. When tested against GraphSAGE, the comparison produced a p-value of 0.07, indicating marginal statistical significance. Alongside significance validation, the authors performed a parameter sensitivity analysis to examine model resilience across varying temporal resolutions. Evaluating HTGNN on Last.FM and MovieLens across day-level, hour-level, and minute-level groupings revealed that the architecture maintains robust recommendation quality across diverse aggregation settings. On Last.FM, the model reached its peak NDCG@10 score of 98.4 under day-level granularity. Conversely, on MovieLens, minute-level granularity achieved the highest NDCG@10 score of 79.2, indicating that datasets featuring denser timestamp annotations derive measurable performance gains from finer temporal partitioning.

A focused frequency impact evaluation further clarified how data granularity influences model behavior by directly comparing HTGNN and TGN across broad and fine temporal resolutions. Because both frameworks leverage graph neural networks to model structural interaction graphs, testing them across progressively adjusted granularities—from year-level down to months, days, hours, and minutes—isolated their ability to process temporal dynamics. At the broadest year-level resolution, HTGNN recorded an NDCG of 86.9 percent, noticeably surpassing TGN’s 81.2 percent. At the intermediate month-level scale, both models delivered identical performance, achieving an NDCG of 83.3 percent. However, as timestamps were refined to capture shorter operational windows, HTGNN demonstrated a consistent competitive edge. At the hour-level resolution, HTGNN attained an NDCG of 84.9 percent compared to 79.7 percent for TGN. Even at the finest minute-level granularity, HTGNN maintained a leading score of 78.9 percent against TGN’s 75.8 percent, underscoring its capacity to deliver accurate recommendations in time-sensitive environments.

The researchers illustrated the dynamic mechanics of the framework through a case study examining temporal preference evolution in movie recommendations across distinct dates, comparing an initial interaction state on April 21, 2024, with subsequent behavior on August 23, 2024. Over this progression, user connections transitioned from genres such as Romance and History toward Thriller and Comedy, demonstrating the necessity of tracking graph structural modifications across time. The model’s temporal attention mechanism provided interpretability by assigning quantifiable weights to historical interactions, revealing whether recent shifts toward Thriller outweighed older, established preferences for Animation. Despite these demonstrated capabilities, the authors explicitly acknowledged that model interpretability remains a key limitation of the current architecture. To address this constraint, future research directions will focus on developing adaptive timestamp granularity and integrating contrastive learning mechanisms to improve transparency and operational scalability. The experimental implementation and code have been made publicly available on GitHub, providing an empirical foundation for continued research into multi-scale time-aware graph modeling.

Subject of Research: Hierarchical time-aware graph neural networks for dynamic user preference modeling in recommender systems

Article Title: Dynamic user preference modeling in graph-based recommender systems

Article References: Abidi, R., Inoubli, W., Ayadi, M. G., & Abdessalem, W. B. (2026). Dynamic user preference modeling in graph-based recommender systems. Applied Network Science, 11(1), Article 47. https://doi.org/10.1007/s41109-025-00730-x

Image Credits: AI Generated

DOI: 10.1007/s41109-025-00730-x

Keywords: recommender systems, graph neural networks, hierarchical time-aware graph neural networks, temporal dynamics, user preference modeling, multi-scale temporal encoding, temporal attention mechanism, hierarchical temporal aggregation, normalized discounted cumulative gain, MovieLens dataset, Retailrocket dataset, temporal-graph networks, precision and recall, interaction graphs, temporal granularity, paired t-tests, dynamic user-item interactions