Automated diabetic retinopathy (DR) severity classification faces persistent challenges from limited annotated datasets, severe class imbalance across the five-level International Clinical Diabetic Retinopathy (ICDR) scale, and domain shift between imaging environments. This paper presents an integrated deep learning framework that extends our earlier Fused Ensemble Convolutional Inception Network (93.2% accuracy) through three complementary innovations. First, a class-conditional latent denoising diffusion probabilistic model (DDPM) generates anatomically realistic synthetic fundus images for underrepresented DR severity levels, addressing data scarcity and class imbalance simultaneously. Second, a SimCLR contrastive self-supervised pre-training protocol, applied to the combined real-and-synthetic corpus, learns discriminative retinal representations prior to supervised fine-tuning. Third, a five-model ensemble (EfficientNet-B3, ResNet-50, VGG16, VGG19, and DenseNet-121) fused via a learned Pattern-Net achieves improved generalization through architectural diversity. Evaluated on the APTOS 2019 dataset under 5-fold stratified cross-validation, the integrated framework achieves 95.1% accuracy, 0.928 quadratic weighted kappa (QWK), 0.91 F1-score, and 0.97 AUC. Class-wise improvements over the EfficientNet-B3 baseline reach +23.8 percentage points for Severe DR (61.8 → 85.6%) and +17.7 for Mild DR (71.2 → 88.9%). Independent zero-shot validation on EyePACS yields 78.4% accuracy, 0.809 QWK, 0.738 macro F1, and 0.887 macro AUC, quantifying the expected performance reduction under domain shift. Fold-level variability and paired out-of-fold statistical analyses are additionally reported, and Grad-CAM visualization is included for qualitative model interpretability. These results demonstrate the synergistic benefit of generative augmentation with self-supervised representation learning for imbalanced medical image classification, with implications for scalable automated DR screening in resource-limited settings.
Teleoperated Robot Avatars enable operators to provide customer service from remote locations. Effective persuasive interaction requires more than information presentation; customers' attention and engagement must be elicited before their attitudes and behavior can be influenced. Social presence may support this type of engagement. Based on prior research, we conceptualize a Robot Avatar's social presence as the extent to which customers perceive it, i.e., not as a machine or information medium, but as a social entity customers can attend to and interact with. Therefore, we examined a design approach intended to enhance a Robot Avatar's social presence by forming a social context around it. We developed GayaRobo, comprising one teleoperated Robot Avatar and two supporting Gaya robots. Rather than approaching users as independent agents, the Gaya robots respond to the Robot Avatar's utterances via speech and movement, thereby forming a social context centered on the avatar. We conducted a field experiment at a commercial complex in Osaka, Japan, to compare observable early-stage engagement behaviors of passersby toward the Robot Avatar across three experimental conditions. The same three-robot configuration was used under three conditions: the Gaya robots responded to the Robot Avatar, only the Robot Avatar spoke, or only the Gaya robots spoke. Two annotators coded four behaviors among 1,379 passersby in 270 min of video recordings: looking at the Robot Avatar, approaching it, talking to it, and picking up a flyer. After correction for multiple testing, only approaching the Robot Avatar was significantly associated with experimental condition and occurred more frequently than expected under the inter-robot interaction condition. These findings provide behavioral evidence that the inter-robot interaction condition was associated with approaching the Robot Avatar, an observable indicator of early-stage engagement. Because social presence was not directly measured, the findings did not demonstrate that GayaRobo enhanced the Robot Avatar's social presence. Nevertheless, they highlight approach behavior as a relevant observable outcome for future evaluations of designs that consider the social context surrounding a Robot Avatar in customer service settings.
Accurate financial time-series forecasting is essential for quantitative investment. However, existing methods often perform inadequately when handling the complex dynamics of financial data, including non-stationarity, volatility clustering, and market regime transitions. These characteristics can adversely affect both forecasting accuracy and the quality of investment decisions. To address these challenges, this study develops AHAT-Fin, an improved Transformer-based forecasting framework that integrates adaptive hybrid attention with market state perception. AHAT-Fin integrates dual-path adaptive hybrid attention, generalized autoregressive conditional heteroskedasticity (GARCH)-based dynamic volatility adjustment, and hierarchical market-state perception based on multi-window long short-term memory (LSTM) networks. The model is evaluated using daily trading data from seven representative U.S. stocks and selected constituents of the CSI 300 Index. The results show that AHAT-Fin generally outperforms the benchmark models, achieving an average reduction of 20.2% in mean squared error, an 8.4% improvement in directional accuracy, and a 63.4% increase in the Sharpe ratio. Module-level evaluation further confirms that all three components contribute to the overall predictive performance, with the market state perception module playing a particularly important role in improving risk-adjusted investment performance. Visual analyses of attention weights and market state identification further demonstrate the model’s ability to extract multidimensional financial features and capture transitions across market states. These findings demonstrate the potential of combining volatility-aware modeling, adaptive attention, and market-state perception for financial forecasting and quantitative strategy design.
Collaborative robots operating in modern manufacturing environments require trajectory planning strategies that simultaneously support efficiency, adaptability, and operational safety. This study proposes a Multi-Level Deep Deterministic Policy Gradient (ML-DDPG) framework for safety-constrained trajectory optimization. The framework combines a strategic waypoint level, a continuous DDPG motion-control policy, and a reactive feasibility projection that corrects unsafe actions before execution. The trajectory-planning problem is formulated as a Markov Decision Process with reward terms for goal progress, obstacle clearance, smoothness, energy, execution time, and safety. Descriptive experiments in a custom PyTorch-based three-dimensional environment compare ML-DDPG with conventional planners and the DDPG, TD3, and SAC baselines. In the retained single training run, ML-DDPG produced observed values of 98.7% success, 1.08 mm trajectory error, a 0.98 smoothness index, and no recorded collision. These values are single-run observations and are not means over independent seeds; variability and statistical significance cannot be estimated from the available record. The exported traces show faster reward growth within the 200-episode budget, and the single-run ablation record associates the largest observed gains with the combined multi-level, reward, and projection components. One UR5e demonstration yielded 0.018 m mean tracking error, 0.042 m maximum tracking error, and no collision in that trial. A separate ABB welding sequence is included only as qualitative feasibility evidence because synchronized trajectory, force, and weld-quality logs were unavailable. The findings motivate further evaluation, while multi-seed training, repeated hardware trials, complete reporting of evaluation procedures, and cross-platform validation remain necessary before broader generalization.
The escalating sophistication of cyber threats against organizational infrastructure demands detection frameworks that simultaneously achieve high classification performance, statistical robustness, and operational transparency. Most published intrusion detection frameworks optimize for accuracy alone, omitting leakage-free evaluation protocols, formal significance testing, and the auditable explanations required by contemporary AI governance frameworks. This paper presents CyberShield Analytics, a SHAP-guided two-level stacking ensemble framework for network intrusion detection and organizational risk resilience. The study addresses the binary intrusion-detection task (Attack vs. Normal); the 10 underlying traffic categories of the dataset are used only for exploratory characterization and are not predicted as separate classes. Evaluated on the benchmark UNSW-NB15 dataset comprising 257,673 network flow records, the framework integrates nine heterogeneous base learners-Logistic Regression, Decision Tree, Random Forest, Extra Trees, Gradient Boosting, XGBoost, LightGBM, Multilayer Perceptron, and K-Nearest Neighbors-with an XGBoost meta-learner operating on leakage-free out-of-fold probability estimates. A systematic feature engineering pipeline expands the original 45-feature space to 67 dimensions through 22 domain-informed interaction and transformation features. The ensemble achieves Accuracy = 0.9535, F 1 = 0.9535, and AUC-ROC = 0.9929 on a fully held-out test set of 51,535 instances, outperforming all individual base learners. Statistical validity is confirmed via 5-Fold Stratified Cross-Validation (mean F 1 = 0.9499 ± 0.0013), with paired t -tests showing significant superiority over XGBoost (t = 7.165, p = 0.002), LightGBM (t = 25.108, p < 0.0001), and Extra Trees (t = 3.178, p = 0.034). Error evaluation confirms the lowest MAE (0.04653) and RMSE (0.21571) among all evaluated models. A tri-method Explainable AI (XAI) layer comprising SHAP, DeepLIFT, and Integrated Gradients applied to representative base learners consistently identifies sttl, ct_dst_sport_ltm, and byte_ratio as the dominant threat-discriminating features. The CyberShield framework bridges the critical gap between predictive performance and operational explainability. Improvements should be read as gains in robustness, calibration, and statistical reliability at comparable accuracy rather than a headline accuracy increase. The framework offers actionable threat intelligence for Security Operations Centers (SOCs) and cyber risk governance in enterprise environments.
