Background: Understanding how Regional Innovation Systems (RISs) drive innovation outputs remains a central question in innovation studies. Most existing empirical research relies on linear or single-indicator models, which fail to capture nonlinear interactions among the key RIS dimensions—Firms, Knowledge, Government, and Economy. Methodology: This study proposes an integrated analytical framework that combines Confirmatory Factor Analysis (CFA), CatBoost machine learning, and SHAP-based explainability to bridge theory-driven modeling with data-driven prediction. Using provincial panel data from China spanning 2011–2023, CFA is first employed to construct and validate four latent RIS dimensions. These latent constructs are then used as inputs in a CatBoost model to predict regional patent outputs, followed by SHAP analysis to quantify the marginal and interactive contributions of each dimension. Results: The CFA results confirm the reliability and validity of the four latent dimensions, establishing a robust structural foundation for the RIS. The CatBoost model achieves high predictive accuracy (log-transformed $R^2$ = 0.975, RMSE = 0.206), substantially outperforming traditional linear benchmarks. SHAP analysis indicates that the Firm dimension is the primary driver of innovation output, while Knowledge, Government, and Economy dimensions exhibit context-dependent moderating effects characterized by diminishing returns, threshold effects, and nonlinear synergies. Conclusions: By integrating latent-variable modeling with interpretable machine learning, this study develops a “CFA-CatBoost-SHAP” closed-loop paradigm for transparent and high-precision analysis of innovation mechanisms. This approach advances RIS theory by empirically validating its multidimensional structure, enriches the methodological toolkit for innovation research, and provides actionable insights for the design of targeted R&D and innovation policies.

Coreference resolution is a key task in Natural Language Processing. It is difficult to evaluate the similarity of long-span texts, which makes text-level encoding somewhat challenging. This paper first compares the impact of commonly used methods to improve the global information collection ability of the model on the BERT encoding performance. Based on this, a multi-scale context information module is designed to improve the applicability of the BERT encoding model under different text spans. In addition, improving linear separability through dimension expansion. Finally, cross-entropy loss is used as the loss function. After adding BERT and span BERT to the module designed in this article, F1 increased by 0.5% and 0.2%, respectively.

Background/Objectives: Knowledge Graphs (KGs) are often incomplete, which can significantly impact the performance of downstream applications. Manual completion of KGs is time-consuming and costly, emphasizing the importance of developing automated methods for KGC. Link prediction serves as a fundamental task in this domain. The semantic correlation among entity features plays a crucial role in determining the effectiveness of link-prediction models. Notably, the human brain can often infer information using a limited set of salient features. Methods: Inspired by this cognitive principle, this paper proposes a lightweight Bi-level routing attention mechanism specifically designed for link-prediction tasks. This proposed module explores a theoretically grounded and lightweight structural design aimed at enhancing the semantic recognition capability of language models without altering their core parameters. The proposed module enhances the model’s ability to attend to feature regions with high semantic relevance. With only a marginal increase of approximately one million parameters, the mechanism effectively captures the most semantically informative features. Result: It replaces the original feature-extraction module within the KGML framework and is evaluated on the publicly available WN18RR and FB15K-237 dataset. Conclusions: Experimental results demonstrate consistent improvements in standard evaluation metrics, including Mean Rank (MR), Mean Reciprocal Rank (MRR), and Hits@10, thereby confirming the effectiveness of the proposed approach.

Wearable neurotechnology depends critically on continuous movement monitoring to characterize motor impairment and recovery in real-world settings. While joint torque serves as a clinically essential kinetic marker, estimating it directly on-device from inertial signals remains challenging due to stringent computational, memory, and energy constraints. Lightweight pipelines typically omit computationally expensive time–frequency processing; however, this omission degrades the observability of dynamics encoded in 1D IMU signals and diminishes the effectiveness of standard knowledge distillation strategies. To enable reliable on-device torque inference, we propose a Physically Guided Dual-Consistency Knowledge Distillation (PDC-KD) framework that explicitly integrates biomechanical priors into the learning process through two collaborative pathways: parameter-manifold alignment and physics-guided compensation. The student network receives guidance through Fisher-information-weighted parameter transfer, ensuring robust knowledge distillation despite significant model capacity mismatch. Furthermore, the framework incorporates a physics-guided regularization term that enforces dynamically consistent torque trajectories via a numerically stable Cholesky-parameterized constraint. Experiments demonstrate that the student model preserves teacher-level predictive accuracy while operating within the stringent resource constraints of edge devices (achieving a 98% parameter reduction, ∼2× faster inference, and ∼1 ms latency). Moreover, the proposed method yields torque estimates with enhanced dynamical consistency, providing an efficient biosignal-processing solution for wearable neurotechnology platforms demanding real-time movement analytics.