Cybersecurity awareness and training programmes are widely implemented in higher education institutions, yet their association with employees’ behavioural risk remains insufficiently quantified. Using an explanatory sequential mixed-methods design, this study analysed survey responses from 167 faculty, staff and managers in United Arab Emirates higher education institutions and 22 semi-structured interviews. Behavioural risk was operationalised as a 0–100 score derived exclusively from five self-reported secure-behaviour items, so that no predictor contributed to the outcome. Respondents who reported attending institutional training had substantially lower behavioural risk than those who had not (13.4 vs. 40.8; Hedges’ g = 1.07), and training exposure explained 15.2% of additional variance after controlling for role. This association was transmitted through policy knowledge (indirect β = −0.52, 95% CI [−0.69, −0.35]) and was not independent of perceived cybersecurity culture and technical readiness; culture did not moderate it. In leakage-safe repeated cross-validation, simpler models predicted behavioural risk better (support vector regression, R2 = 0.73) than the proposed deep feedforward network (R2 = 0.28). Interviews described absent, optional or awareness-only training and a weak, compliance-oriented culture. Training appears to be associated with lower risk chiefly through policy knowledge, within a broader institutional security climate.
Semantic identifier (SID) generative recommenders cast retrieval as autoregressive decoding over short sequences of residual-quantized codewords. Most published SID recommenders decode all requests with one fixed beam width, which is optimal only if the marginal utility of compute is constant across requests. We find that it is not; additional search changes little for users with consistent histories, but improves accuracy substantially for cold start, interest-drifting, and long-tail users. The main methodological contribution of this paper is request-level allocation of decoding compute, in which a router predicts the accuracy gain of each configuration in a finite decoding menu over the cheapest configuration and selects the configuration with the largest gain net of a priced cost. The selection rule follows from standard Lagrangian arguments, which we use as tools; what is specific to this work is the use of this per-configuration gain as the routing target for SID decoding instead of a difficulty or uncertainty score. The resulting framework, called Bastion, combines the router with a hierarchical process verifier over the quantization levels, a listwise terminal verifier, codeword consistency voting, and budget-constrained group relative policy optimization. On five benchmarks built from Amazon Reviews 2023, MovieLens, and MicroLens, Bastion improves NDCG@10 by 5.7% to 8.7% over the strongest baseline. Its mean latency on Amazon-Beauty is 14.5 ms, against 14.6 ms for a beam-20 TIGER decoder, while its median and 99th-percentile latencies are higher by 0.9 ms and 7.0 ms. At every evaluated budget, its accuracy exceeds that of all baselines on the three datasets for which accuracy versus latency frontiers was measured.
Educational resource allocation requires the simultaneous consideration of regional demand heterogeneity, spatial adjacency, and nonlinear prediction relationships. This study presents an integrated framework that combines Spatially Constrained Entropy-Weighted K-Means (SCEW-KMeans) with Bayesian Optimization–Random Forest (BO-RF). SCEW-KMeans employs entropy weights to quantify the informational heterogeneity of indicators related to educational supply, population demand, and fiscal capacity, incorporates county-level spatial adjacency into the clustering objective function, and identifies stable regional partitions through Bootstrap consensus selection. During the demand forecasting stage, a nested spatial block cross-validation strategy is adopted, with Linear Regression (LR), Support Vector Regression (SVR), Global Random Forest (RF), Extreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), and Categorical Boosting (CatBoost) serving as unified competitive benchmark models. Experimental results based on annual observational data from 95 county-level administrative units in Jiangsu Province from 2021 to 2023 demonstrated that the SCEW-KMeans and BO-RF joint framework achieved a mean squared error (MSE) of 94.9 and a coefficient of determination (R2) of 0.879 in predicting county-level per-student fiscal expenditure on education. The clustering results attain an average silhouette coefficient of 0.65, a Bootstrap-Adjusted Rand Index (ARI) of 0.91, and a spatial boundary cut ratio of 0.18.
Developing objective and interpretable neurophysiological measures is important for characterizing Parkinson’s disease (PD). EEG microstates capture sub-second large-scale brain-state dynamics, while sequence-level measures provide complementary information on their temporal organization. We compared microstate duration, occurrence, and coverage between patients with PD and healthy controls (HCs), and assessed microstate sequence complexity using multiscale entropy (MSE). To address the arbitrary numerical coding of microstate labels, all 120 possible label recodings were tested, together with symbolic and surrogate analyses. Under the primary group-specific template representation, PD showed reduced duration and coverage of microstates A and B, shorter duration but increased occurrence of microstate C, increased occurrence of microstate D, and increased duration, occurrence, and coverage of microstate E. MSE was higher in PD across all 20 scales, with the direction preserved across all label recodings despite variation in effect magnitude. Full-sequence Lempel–Ziv complexity and conditional entropy were also higher in PD. These effects were markedly attenuated after run compression, and no significant residual group differences were detected under surrogate models preserving occupancy, run-length structure, or first-order transitions. In a common-template sensitivity analysis using the same diagnosis-independent spatial reference for all participants, the class-specific temporal effects and symbolic sequence effects were substantially attenuated, whereas MSE remained higher in PD across all 20 scales, with significant overall group and group × scale effects. These findings suggest that MSE provides a comparatively robust integrative description of PD-related microstate sequence organization, while the observed increase in sequence complexity appears to be closely related to lower-order temporal reorganization rather than representing an independent higher-order complexity abnormality.
Feature selection plays an important role in medical classification, particularly when datasets contain high-dimensional and redundant information. At the same time, the performance of a Random Weight Network (RWN) depends on its configuration, making the joint optimization of feature subsets and network parameters a relevant problem. This study proposes a metaheuristic-based framework that integrates feature selection with RWN configuration optimization. The framework simultaneously determines the selected features, active hidden neurons, regularization parameter, and activation function. Eight binary transfer functions, including four S-shaped and four V-shaped functions, are investigated to convert continuous search variables into binary decisions for feature and hidden-neuron selection. The framework is evaluated using four metaheuristic algorithms: Grey Wolf Optimizer (GWO), Particle Swarm Optimization (PSO), Genetic Algorithm (GA), and Differential Evolution (DE). Experiments are conducted on 20 benchmark medical datasets under consistent experimental conditions. The results show that the S2 transfer function provides the most favorable overall performance among the investigated transfer functions. Using S2, GWO achieves the best overall Friedman ranking among the four optimizers, although the overall differences are not statistically significant at the 5% level for some evaluation measures. The resulting GWO-S2-RWN configuration also demonstrates competitive performance when compared with established feature-selection and classification methods reported in the literature. Overall, the findings show that jointly optimizing feature selection and RWN configuration provides a flexible approach for balancing classification performance and model complexity in medical classification tasks.
