Researchers have developed a progressive representation recovery framework, known as RGSM-STKG, designed to improve temporal reasoning in sparse temporal knowledge graphs (STKGs). Temporal knowledge graphs, which extend standard graph structures by incorporating timestamps for facts, are essential for modeling evolving real-world information. However, they frequently suffer from data incompleteness, where missing relations or unrecorded events create structural and temporal sparsity. This sparsity does not merely result in the absence of individual facts but actively disrupts the evolution trajectories of entities and relations by degrading their representations. To address these challenges, the team formulated sparse TKG reasoning as a progressive representation-recovery process, ensuring that structural enhancement precedes temporal reconstruction and multi-scale inference. Experimental evaluations on eight sparse subsets across three major benchmarks—ICEWS14, ICEWS05-15, and WIKI—showed that the model consistently outperformed existing baselines, achieving a 43.06% improvement in Mean Reciprocal Rank (MRR) on the highly sparse ICEWS14_2% subset.
The fundamental challenge in STKG reasoning arises from the coupling of structural and temporal sparsity. When insufficient neighboring entities and relational observations are present at a given timestamp, entity representations become incomplete or unstable. If these degraded observations are irregularly distributed across time, the continuity of entity-relation evolution is interrupted, leading to missing latent information. Conventional models often treat these as isolated issues, whereas the RGSM-STKG framework integrates three successive stages to recover the underlying knowledge. First, a dual-structure perception module alleviates representation degradation within individual snapshots. Second, a Mamba-based dynamic evolution module captures long-range dependencies to reconstruct coherent state sequences. Third, a wavelet-based multi-scale inference module explicitly models heterogeneous temporal variations before final entity prediction, providing a robust solution for environments characterized by severe structural and temporal incompleteness.
In the first stage, the dual-structure perception module employs a progressive refinement process to address snapshot-level degradation. This stage begins with a temporally relevant neighborhood perception mechanism. By applying a temporal decay function to historical interactions, the model prioritizes neighborhood evidence closer to the current historical timestamp, effectively down-weighting distant, less relevant interactions. The sampled neighborhood structural evidence is then aggregated using a GraphSAGE operator to generate neighborhood-enhanced entity representations. This output is further refined through a relation-aware structural refinement process modeled after Relation Graph Convolutional Networks (RGCN). By incorporating relation-specific transformations, the model ensures that the distinct semantic roles an entity may play across different relations are preserved. This dual-approach produces structurally enhanced historical entity and relation states, which serve as the foundation for subsequent temporal modeling.
The second stage, the dynamic evolution module, addresses the temporal continuity of these enhanced states by utilizing Mamba, a selective Structured State-Space Model (SSM). Given that historical states remain irregularly distributed, Mamba models temporal dependencies across the sequence using a selective state-space mechanism that adapts to input data. The model processes the historical entity and relation state sequences through Mamba-based evolution blocks, effectively reconstructing latent entity-relation trajectories even when historical observations are discontinuous. By maintaining the evolved representation at each historical position, the framework produces temporally coherent state sequences. To estimate the state at an unobserved target timestamp, the model performs a residual temporal transition over these reconstructed sequences, providing reliable anchor representations for query-conditioned prediction.
The final stage focuses on multi-scale temporal inference, introducing a wavelet-based decomposition technique along the temporal axis of the query-conditioned sequences. Rather than focusing on the occurrence frequency of entities, this approach analyzes the variation of latent representations across different temporal scales. The model utilizes a single-level Daubechies 1 (Db1) wavelet transform to decompose the temporal trajectory into low- and high-frequency components. The low-frequency component represents locally smoothed temporal information, while the high-frequency component captures complementary local variations between neighboring temporal states. These components are processed independently through learnable transformations before being reconstructed via an inverse wavelet transform. This process allows the model to differentiate between persistent temporal patterns and rapid local fluctuations, which are then aggregated by multi-head attention to generate a query-aware representation for target prediction.
The researchers validated the RGSM-STKG framework through extensive experiments on multiple sparse subsets of the ICEWS14, ICEWS05-15, and WIKI datasets. The sparse subsets were constructed using Bernoulli sampling with predefined retention probabilities ranging from 0.3% to 5%, ensuring consistent testing under varied sparsity levels. On the ICEWS14 benchmark, which is known for its sparse timestamp distribution, the model showed significant gains over the STKGR-PR baseline, with MRR improvements of 43.06%, 49.41% on Hits@3, and 29.24% on Hits@10 for the 2% retention scenario. These results indicate that even when temporal information is severely restricted, the progressive recovery of structural and temporal states effectively bolsters prediction accuracy compared to methods that do not explicitly model these coupled effects.
Further experiments on the ICEWS05-15 dataset, which features denser temporal information and a longer historical span, provided additional confirmation of the model’s robustness. Across all tested sparsity levels in ICEWS05-15, the model consistently outperformed both established embedding-based reasoning approaches and more recent temporal modeling techniques. The performance gap highlights the model’s ability to leverage richer historical observations for more precise temporal evolution modeling. In comparisons where embedding-based models traditionally rely on static feature consistency, the results demonstrated that explicitly recovering dynamic temporal states yields a distinct advantage, particularly in settings where the knowledge graph exhibits complex, time-dependent evolution patterns.
To investigate the individual contributions of the progressive stages, the research team conducted ablation studies. Systematically removing the dual-structure perception module resulted in consistent performance degradation across both ICEWS14 and ICEWS05-15 datasets, confirming its necessity for mitigating snapshot-level representation degradation. Similarly, ablating the dynamic evolution module and the wavelet-based inference module caused measurable decreases in Mean Reciprocal Rank. These results emphasize that structural enhancement, temporal state reconstruction, and multi-scale frequency analysis are not redundant; rather, they provide complementary information that progressively refines the model’s understanding of the latent graph state. The consistent degradation across diverse sparsity levels underscores the model’s architectural stability.
A comparative analysis of temporal modeling architectures further supported the selection of Mamba over recurrent units like LSTM or self-attention mechanisms like the Transformer. While LSTM and Transformer-based variants remained competitive, the Mamba-based implementation achieved consistently higher scores across all evaluation metrics. This advantage was particularly pronounced under severe sparsity settings, suggesting that Mamba’s input-dependent state propagation is superior for capturing long-range dependencies in the presence of irregular and incomplete data. Furthermore, sensitivity analysis regarding wavelet configurations confirmed that the Db1 basis and a single-level decomposition provided the best balance between preserving coarse-scale information and local temporal nuances, while deeper decomposition levels inadvertently weakened the representation of critical local variations.
The study also utilized a controlled temporal masking experiment and a structural sparsification analysis to provide representation-level evidence of the framework’s recovery capabilities. By removing increasing proportions of historical facts, the team measured the stability of entity representations and the ability of the dynamic evolution module to recover disrupted information. The results showed that as historical incompleteness increased, the dynamic evolution module’s effectiveness in narrowing the gap between masked and reference states became more pronounced, as quantified by the Temporal Information Recovery Score (TIRS). Similarly, the structural sparsification analysis demonstrated that the dual-structure perception module maintains higher representation stability compared to ablated versions, directly validating the claim that the model preserves more robust and informative states despite limited structural evidence.
Computational efficiency analysis indicated that although RGSM-STKG contains more trainable parameters than the reinforcement-learning-based STKGR-PR baseline, it exhibits lower average training time and inference latency per query. This efficiency gain is attributed to the framework’s use of batched neural operations over learned representations rather than sequential action selection, which is often a bottleneck in path-based reasoning methods. Despite the slightly higher peak GPU memory requirement, the performance gains achieved by RGSM-STKG present a favorable trade-off for high-sparsity reasoning tasks. While the framework provides a robust solution, the researchers noted that its performance may decrease in scenarios with extreme scarcity or noisy observations, marking future directions for uncertainty-aware and inductive reasoning on larger-scale datasets.
Subject of Research: Sparse temporal knowledge graph reasoning
Article Title: RGSM-STKG: structure-aware representation enhancement and dynamic-evolution completion for sparse temporal knowledge graph reasoning
Article References: Dong, R., Peng, J., Wei, S., Hong, Q., & Yin, J. (2026). RGSM-STKG: structure-aware representation enhancement and dynamic-evolution completion for sparse temporal knowledge graph reasoning. Journal of King Saud University Computer and Information Sciences, 38(9), Article 1008. https://doi.org/10.1007/s44443-026-01423-w
Image Credits: AI Generated
DOI: 10.1007/s44443-026-01423-w
Keywords: Temporal Knowledge Graphs, Representation Learning, Mamba Architecture, Wavelet Decomposition, Data Sparsity, Structural Refinement, Dynamic Evolution, Graph Neural Networks, Multi-scale Analysis, Temporal Reasoning, Information Recovery, Entity Representations, ICEWS14 Benchmark, Neural-symbolic Reasoning, State-space Models
Cite Scienmag News
Denise Maddox. (November 20, 2026). New Framework Improves Reasoning Efficiency in Sparse Temporal Knowledge Graphs. Scienmag. https://scienmag.com/new-framework-improves-reasoning-efficiency-in-sparse-temporal-knowledge-graphs/
Denise Maddox. "New Framework Improves Reasoning Efficiency in Sparse Temporal Knowledge Graphs." Scienmag, 20 November 2026, https://scienmag.com/new-framework-improves-reasoning-efficiency-in-sparse-temporal-knowledge-graphs/. Accessed 20 November 2026.
Denise Maddox. "New Framework Improves Reasoning Efficiency in Sparse Temporal Knowledge Graphs." Scienmag. November 20, 2026. https://scienmag.com/new-framework-improves-reasoning-efficiency-in-sparse-temporal-knowledge-graphs/








