{"title":"Link prediction for static knowledge graphs based on attention mechanism and deep graph propagation","authors":"Rongyuan Qin, Jianzhong Chen, Ying Pan, Xiong Li","doi":"10.1007/s40747-026-02555-2","DOIUrl":"https://doi.org/10.1007/s40747-026-02555-2","url":null,"abstract":"The ongoing expansion of application scenarios for knowledge graphs enhances the research significance of link prediction. Current graph neural network methodologies utilized for link prediction on static knowledge graphs frequently suffer from node feature over-smoothing and information degradation in deeper layers, hindering the comprehensive utilisation of structural patterns and semantic relationships within the graphs. This study aims to overcome these limitations by focusing on the static knowledge graph completion problem and proposing a link prediction method that incorporates attention mechanisms and deep graph propagation. This method enhances the efficacy of link prediction by optimising semantic graph representations and analysing entity relationships. This paper’s primary contribution is a sophisticated graph propagation model called Multi-Head Attention and Dynamic Pruning Graph (MADPGraph), designed for static knowledge graph link prediction, integrating multi-head attention relation-aware modules and dynamic pruning techniques. The approach initially creates query-oriented subgraphs to encapsulate multi-hop neighbourhood contextual information focused on head entities. A relation-aware multi-head attention technique aggregates query relations, edge relations, and node features, enabling deep information propagation across multiple attention heads. An inter-layer adaptive dynamic pruning module is introduced after generating contextualised node representations for each query. This module dynamically identifies semantically pertinent nodes during layer-wise propagation to limit subgraph size and is designed to reduce the risk of feature homogenization caused by irrelevant deep propagation. It retains semantic information along essential pathways and improves model generalisation by ensuring uniform representations across various trimmed paths. Experimental findings on typical static knowledge graph datasets and inductive benchmark datasets indicate that MADPGraph achieves competitive performance relative to existing mainstream approaches, thereby supporting the effectiveness and generalisation ability of the proposed strategy in static contexts.","PeriodicalId":10524,"journal":{"name":"Complex & Intelligent Systems","volume":"14 1","pages":""},"PeriodicalIF":5.8,"publicationDate":"2026-10-30","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149287758","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Linh Do Nhat, Dzi Lam Tran Tuan, Vinh Truong Hoang
{"title":"The architectural mirage: capacity ratios fail as a safeguard against harmful knowledge distillation in cross-domain pill retrieval","authors":"Linh Do Nhat, Dzi Lam Tran Tuan, Vinh Truong Hoang","doi":"10.1007/s40747-026-02531-w","DOIUrl":"https://doi.org/10.1007/s40747-026-02531-w","url":null,"abstract":"Compact models on edge devices need knowledge distillation (KD) to close their cross-domain accuracy gap, but the choice of teacher decides whether distillation helps or harms. We characterise that failure at a scale that tests selection rules rather than illustrates them: 180 teacher–student pairings on ePillID over three training seeds, and 180 more on an independently derived CURE benchmark. Harmful transfer is common (21.8% and 27.3% of pairings) and concentrates in compact transformer students on both benchmarks (45.0% versus 8.6% for CNN students on ePillID; 55.0% versus 11.4% on CURE; single-seed labels). The capacity ratio, the safeguard much of the literature relies on and this study originally proposed, fails three ways: it discriminates weakly (screening AUC 0.66); its fitted threshold does not transfer between benchmarks (17.1 and 24.0, neither of them the asserted 12); and at 12 it admits no teacher for the two smallest students, forbidding distillation where it is most needed. Of the replacements we test, one result is modestly positive and the rest are negative. No training-free signal we evaluate, including four similarity metrics, their fusions, and five external transferability estimators, ranks teachers better than taking the strongest teacher by its own benchmark accuracy, by more than the measured noise floor. For the easier question of whether a pairing will harm at all, a correspondence-aware manifold divergence leads on ePillID; its AUC margin over the capacity ratio is unresolved at eleven student clusters, but at a threshold fitted out of sample, and on the nine students where both rules give a decision, it admits 3 harmful pairings against that rule’s 9 while discarding 24 beneficial ones against 39, it still leads when the rule is given its own fitted threshold, and it is defined for every student. The screen does not transfer visibly to CURE, where no signal we test separates harm from safety, in a design with power only against strong screens.","PeriodicalId":10524,"journal":{"name":"Complex & Intelligent Systems","volume":"76 1","pages":""},"PeriodicalIF":5.8,"publicationDate":"2026-10-30","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149293033","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
{"title":"Exploiting representational redundancy in transformer hidden states through token representation reuse","authors":"Xinpeng Xu","doi":"10.1007/s40747-026-02564-1","DOIUrl":"https://doi.org/10.1007/s40747-026-02564-1","url":null,"abstract":"Transformer-based large language models construct rich contextualized token representations across multiple layers, yet redundancy within their intermediate hidden-state spaces remains insufficiently explored as an inference-time representation property. This paper introduces token representation reuse (TRR), an inference-time framework that identifies highly similar contextual token representations and replaces selected hidden states with preceding anchor representations under similarity and reuse-budget constraints. Unlike token pruning or merging, TRR preserves the original sequence length and positional structure and requires neither model retraining nor parameter modification. We further provide an end-to-end complexity analysis of representation discovery and characterize the local conditions under which identical hidden states can, in principle, permit projection reuse, while explicitly showing that such equality does not generally persist across subsequent transformer layers. Experiments on BoolQ, AGNews, SST-2, and TruthfulQA using Qwen2.5-1.5B-Instruct and Llama-3-8B demonstrate representation reuse ratios of up to 12.40% while largely preserving downstream predictive performance. Under the frozen default configuration, paired statistical analysis shows no significant accuracy change on AGNews or SST-2, while TruthfulQA exhibits a statistically significant improvement of <jats:inline-formula> <jats:alternatives> <jats:tex-math>$$+2.92$$</jats:tex-math> <mml:math xmlns:mml=\"http://www.w3.org/1998/Math/MathML\"> <mml:mrow> <mml:mo>+</mml:mo> <mml:mn>2.92</mml:mn> </mml:mrow> </mml:math> </jats:alternatives> </jats:inline-formula> percentage points on Qwen2.5-1.5B-Instruct (95% CI <jats:inline-formula> <jats:alternatives> <jats:tex-math>$$[+1.15,+4.69],$$</jats:tex-math> <mml:math xmlns:mml=\"http://www.w3.org/1998/Math/MathML\"> <mml:mrow> <mml:mo>[</mml:mo> <mml:mo>+</mml:mo> <mml:mn>1.15</mml:mn> <mml:mo>,</mml:mo> <mml:mo>+</mml:mo> <mml:mn>4.69</mml:mn> <mml:mo>]</mml:mo> <mml:mo>,</mml:mo> </mml:mrow> </mml:math> </jats:alternatives> </jats:inline-formula> p = 0.001); BoolQ shows a significant decrease of <jats:inline-formula> <jats:alternatives> <jats:tex-math>$$-1.50$$</jats:tex-math> <mml:math xmlns:mml=\"http://www.w3.org/1998/Math/MathML\"> <mml:mrow> <mml:mo>-</mml:mo> <mml:mn>1.50</mml:mn> </mml:mrow> </mml:math> </jats:alternatives> </jats:inline-formula> percentage points. End-to-end runtime profiling of the same TRR algorithmic procedure shows that the current dense implementation incurs additional latency because the <jats:inline-formula> <jats:alternatives> <jats:tex-math>$$ O(n^2d) $$</jats:tex-math> <mml:math xmlns:mml=\"http://www.w3.org/1998/Math/MathML\"> <mml:mrow> <mml:mi>O</mml:mi> <mml:mo>(</mml:mo> <mml:msup> <mml:mi>n</mml:mi> <mml:mn>2</mml:mn> </mml:msup> <mml:mi>d</mml:mi> <mml:mo>)</mml:mo> </mml:mrow> </mml:math> </jats:alternatives> </jats:inline-formula> similarity-discovery cost is not offset by hardware-level","PeriodicalId":10524,"journal":{"name":"Complex & Intelligent Systems","volume":"20 1","pages":""},"PeriodicalIF":5.8,"publicationDate":"2026-10-30","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149293031","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
{"title":"Domain-adaptive epidemic forecasting with hybrid SEIR-N-BEATS framework for robust cross-regional prediction under distribution shifts","authors":"Xihe Qiu, Ying Xie, Sulan Wei, Xiaoxiao Ge, Haoyu Wang, Xiaoyu Tan, Liang Liu, Xin Xu, Wen Bao, Fenghao Sun, Lu Gan","doi":"10.1007/s40747-026-02510-1","DOIUrl":"https://doi.org/10.1007/s40747-026-02510-1","url":null,"abstract":"Epidemic forecasting across heterogeneous regions faces significant challenges due to domain shifts in healthcare infrastructure, population demographics, and intervention policies. We present SEIR-N-BEATS, a hybrid transfer learning framework integrating epidemiological theory with neural pattern recognition for robust cross-domain prediction. Our approach combines a modified susceptible-exposed-infectious-removed (SEIR) model with neural basis expansion analysis for time series (N-BEATS) to address domain adaptation challenges. The SEIR component incorporates time-varying transmission rates and policy-aware mechanisms to capture distributional shifts, while N-BEATS learns domain-specific residual patterns through hierarchical blocks with interpretable decomposition. We evaluated cross-domain generalization using epidemic surveillance data from multiple regions with varying healthcare systems (December 2019–May 2022), comparing performance via MAE, RMSE, MAPE, and Pearson correlation under both in-distribution and out-of-distribution scenarios. SEIR-N-BEATS achieved 1.03% MAPE with 19.1% error reduction compared to mechanistic SEIR baselines, demonstrating superior transfer capability across heterogeneous domains (MAE range 1.32–576.39). The framework exhibited exceptional out-of-distribution robustness, maintaining near-zero MAPE variation ( <jats:inline-formula> <jats:alternatives> <jats:tex-math>$$0.00% pm 0.01%$$</jats:tex-math> <mml:math xmlns:mml=\"http://www.w3.org/1998/Math/MathML\"> <mml:mrow> <mml:mn>0.00</mml:mn> <mml:mo>%</mml:mo> <mml:mo>±</mml:mo> <mml:mn>0.01</mml:mn> <mml:mo>%</mml:mo> </mml:mrow> </mml:math> </jats:alternatives> </jats:inline-formula> ) under noise perturbations, compared to substantial MAPE degradation in pure machine-learning baselines (LSTM <jats:inline-formula> <jats:alternatives> <jats:tex-math>$$39.13% pm 2.8%$$</jats:tex-math> <mml:math xmlns:mml=\"http://www.w3.org/1998/Math/MathML\"> <mml:mrow> <mml:mn>39.13</mml:mn> <mml:mo>%</mml:mo> <mml:mo>±</mml:mo> <mml:mn>2.8</mml:mn> <mml:mo>%</mml:mo> </mml:mrow> </mml:math> </jats:alternatives> </jats:inline-formula> , N-BEATS <jats:inline-formula> <jats:alternatives> <jats:tex-math>$$243.23% pm 15.2%$$</jats:tex-math> <mml:math xmlns:mml=\"http://www.w3.org/1998/Math/MathML\"> <mml:mrow> <mml:mn>243.23</mml:mn> <mml:mo>%</mml:mo> <mml:mo>±</mml:mo> <mml:mn>15.2</mml:mn> <mml:mo>%</mml:mo> </mml:mrow> </mml:math> </jats:alternatives> </jats:inline-formula> ). Cross-regional validation confirmed effective domain adaptation with 23.4% average MAE improvement. This framework bridges epidemiological knowledge with data-driven learning, enabling interpretable cross-domain forecasting robust to distribution shifts, offering practical solutions for resource allocation and intervention planning across diverse healthcare settings.","PeriodicalId":10524,"journal":{"name":"Complex & Intelligent Systems","volume":"21 1","pages":""},"PeriodicalIF":5.8,"publicationDate":"2026-10-28","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149287762","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
{"title":"Integration of solar power and isolated industrial micro-grids based on LIF-CoCoSo models with Sugeno–Weber approaches","authors":"Zeeshan Ali, Nadeem Javaid, Miin-Shen Yang","doi":"10.1007/s40747-026-02560-5","DOIUrl":"https://doi.org/10.1007/s40747-026-02560-5","url":null,"abstract":"","PeriodicalId":10524,"journal":{"name":"Complex & Intelligent Systems","volume":"14 1","pages":""},"PeriodicalIF":5.8,"publicationDate":"2026-10-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149287765","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
{"title":"Advances in fuzzy utility theory for risk management: convergence, risk premium bounds, and aggregation of risk aversion","authors":"Mehdi Shams","doi":"10.1007/s40747-026-02522-x","DOIUrl":"https://doi.org/10.1007/s40747-026-02522-x","url":null,"abstract":"","PeriodicalId":10524,"journal":{"name":"Complex & Intelligent Systems","volume":"19 1","pages":""},"PeriodicalIF":5.8,"publicationDate":"2026-10-26","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149287768","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}