temporal graph forecasting

This paper identifies a temporal granularity inconsistency in the prevailing batch-wise evaluation paradigm for dynamic link prediction, leading to misaligned time windows and spurious temporal dependencies—causing up to 12.7% AUC estimation bias and undermining model comparability and generalizability. To address this, the authors first systematically characterize this distortion mechanism and then propose a novel time-aware evaluation framework centered on: (i) timestamp-aligned sequential modeling, (ii) dynamic graph neural network adaptation, (iii) temporal sliding-window resampling, and (iv) a counterfactual evaluation protocol. Extensive experiments across multiple benchmark datasets demonstrate that the proposed paradigm substantially mitigates evaluation bias, enhances fair model comparison, and improves cross-scenario generalization. The work provides both theoretical foundations and a practical framework for standardizing evaluation in temporal graph learning.

Static snapshot-based approaches fail to capture temporal information flow and the evolution of propagation paths in time-evolving graphs. To address this, we propose the first temporal path tracing system designed for highly dynamic vertices. Our method abandons the conventional static graph snapshot paradigm and instead supports configurable, time-constrained path traversal algorithms, integrated with parallel graph processing and web-based interactive visualization. At the modeling level, we adopt an exact temporal graph representation that preserves fine-grained temporal semantics. Computationally, we enable efficient, low-latency path queries over large-scale evolving graphs. Analytically, our system significantly improves both the accuracy and real-time responsiveness of critical propagation path identification. Empirical evaluation demonstrates that our system outperforms state-of-the-art methods in both query efficiency and interpretability, establishing a new benchmark for temporal path analysis in dynamic networks.

This work addresses the challenge of modeling dynamic community evolution—encompassing community splitting, merging, and node additions or deletions—by proposing a novel generative temporal network model. The approach introduces, for the first time, a mutual information–based similarity measure to guide genetic search, thereby explicitly controlling the evolution of community structure across network snapshots. It further incorporates dynamic edge-generation probabilities conditioned on intra- and inter-community connectivity. The model jointly captures both the temporal evolution of communities and changes in node membership. Experimental results demonstrate that the framework effectively reproduces real-world dynamic community behaviors and successfully quantifies how node insertion and deletion rates influence the performance of dynamic community detection algorithms.

This paper addresses graph representation learning for both static and single-event dynamic networks. Methodologically, it introduces a unified structural-aware embedding framework grounded in latent distance modeling, jointly optimizing homophily, transitivity, and balance within an end-to-end paradigm—thereby eliminating heuristic design and multi-stage pipelines. Notably, it is the first to extend latent distance modeling to single-event dynamic settings, enabling extreme node identification and quantitative assessment of influence dynamics. The key contributions are: (1) a hierarchical, interpretable structural-aware representation; (2) seamless unification of embedding learning across static and dynamic networks; and (3) state-of-the-art performance on community detection, anomaly detection, and temporal influence evaluation—significantly outperforming multi-stage baselines.

Existing graph deep learning approaches for multivariate time series forecasting predominantly focus on architectural innovations, lacking systematic methodology studies on problem formalization, model design principles, and evaluation paradigms. This paper introduces the first unified methodology framework for graph-enhanced time series forecasting. It formally defines the forecasting problem by modeling dynamic inter-variable dependencies via learnable graph structures. It establishes principled design guidelines integrating graph neural networks (GNNs), spatiotemporal modeling, and deep temporal models (e.g., TCN and Informer variants). Furthermore, it proposes a reproducible evaluation protocol with built-in mechanisms to ensure interpretability and scalability. The framework shifts forecasting model development from empiricism toward first-principles reasoning, providing standardized guidance for relation-aware global modeling and articulating key open challenges. (149 words)

This study addresses the problem of predicting both the timing and location of link formation in complex networks. It proposes a closed-form, non-Markovian model that integrates latent hyperbolic geometry with long-range memory of historical interactions, thereby unifying geometric structure and memory effects within a single framework for the first time. The resulting approach features few parameters and strong interpretability, offering a principled method for temporal link prediction. By modeling network dynamics through a non-Markovian process and deriving probabilistic predictions, the model achieves excellent agreement with empirical connection probabilities across multiple large-scale real-world networks. These results reveal that network evolution is fundamentally governed by the interplay between geometric constraints and memory-driven mechanisms.

This work addresses nonparametric modeling of temporal networks exhibiting dynamic features such as memory and periodicity, while preserving node exchangeability. To this end, the authors propose a unified framework based on dynamic decorated graphons, decomposing the modeling task into two stages: temporal process modeling and network structure estimation. This approach is the first to accommodate complex temporal dependencies under the constraint of exchangeability, offering rigorous nonparametric convergence rate guarantees. By integrating block models with a two-stage estimator under Hölder smoothness assumptions on the temporal dynamics, the method accommodates diverse edge evolution processes, including autoregressive and Markovian structures. Experiments on both synthetic data and a real-world hospital contact network successfully recover latent community structures and time-varying interaction patterns, demonstrating the method’s empirical effectiveness and theoretical soundness.

Existing dynamic graph modeling approaches are limited by short-term dependencies, static neighborhood semantics, and retrospective temporal assumptions, hindering their ability to capture transferable evolutionary patterns. This work proposes the Temporal Graph Pattern Machine (TGPM), a novel foundation model framework for temporal graphs centered on universal evolutionary patterns, thereby departing from conventional task-specific paradigms. TGPM employs time-biased random walks to generate interaction blocks and leverages a Transformer backbone to learn multi-scale structural semantics and long-range dependencies. It further introduces self-supervised objectives—including masked token modeling and next-time prediction—to uncover underlying evolutionary dynamics. The method achieves state-of-the-art performance on both transductive and inductive link prediction tasks and demonstrates significantly enhanced cross-domain transferability.

This work addresses the challenge that existing link prediction models struggle to disentangle users’ intrinsic preferences from the amplification effect of algorithmic feedback on network homophily. The authors propose the first dynamic graph analysis framework based on a multivariate Hawkes process, which explicitly decouples users’ inherent interaction tendencies from algorithmic feedback mechanisms. Central to this approach is a novel bias metric driven by instantaneous interaction intensity, capturing real-time reinforcement dynamics beyond conventional cumulative measures. The framework is theoretically grounded, with formal proofs establishing the stability and convergence of the induced dynamics. Experimental results demonstrate that the proposed bias metric effectively quantifies the strength of algorithmic feedback under diverse link prediction strategies, offering a reliable tool for understanding how algorithms shape network evolution.