Predicting student performance has become a vital objective in educational data mining, however existing models remain limited by limited feature sets, opaque decision processes, and the absence of real-time, context-aware interpretative intelligence. This study presents a novel hybrid Artificial Intelligence (AI) architecture that integrates cognitive, emotional, and general intelligence indicators with formative assessment/class performance data to construct a multidimensional demonstration of student performance. The system utilizes a mathematically formalized feature-weighting mechanism that allows domain experts to regulate the influence of key predictors prior to training, enabling a tunable and educator-configurable modeling process. Interaction between cognitive, emotional, and behavioral properties are explicitly engineered to capture complex effects often overlooked in prior research. An ensemble predictive model combining Random Forests, Gradient Boosting, and Linear Regression is used to predict student grades, with uncertainty measured through variance-based confidence intervals. The integrated system exhibited strong predictive capability, attaining high accuracy (R 2 = 0.9014, 95% CI [0.8882, 0.9121], 10,000 bootstrap resamples) with low prediction error, confirming the stability and reliability of the ensemble model. Understanding why a model produces a particular prediction is often as important as the prediction itself. To address the interpretability gap, the system incorporates SHapley Additive exPlanations (SHAP)-based global and local explanations, a counterfactual simulation engine that estimates the marginal impact of improving specific attributes, and a local generative AI layer deployed via Ollama that creates these outputs into structured, real-time narrative feedback. Running the Large Language Model (LLM) locally ensures privacy-preserving, institution-compliant interpretability while permitting interactive, student specific explanations. This integration of predictive modeling, explainability, simulation, and local generative reasoning creates a unified decision-support framework that exceeds traditional black-box approaches. Experimental assessment shows that the model not only achieves viable predictive accuracy but also provides actionable insights into feature influence, model stability, and potential intervention pathways.
Early and accurate detection of plant diseases is essential for improving crop productivity and supporting sustainable agricultural practices. Mango (Mangifera indica), one of the most economically important tropical fruit crops worldwide, is highly susceptible to several leaf diseases that can significantly reduce yield and fruit quality. This study presents an explainable deep learning framework for automated ten-class mango leaf disease classification using a publicly available dataset containing approximately 5000 images distributed across Anthracnose, Bacterial Canker, Cutting Weevil, Dag Disease, Die Back, Galls, Healthy, Powdery Mildew, Red Rust, and Sooty Mould classes. Unlike conventional AlexNet-based approaches, the proposed framework integrates Bayesian hyperparameter optimization, two-stage fine-tuning, test-time augmentation (TTA), and Grad-CAM-based explainability to improve both predictive performance and interpretability. Experimental results demonstrate that the proposed framework achieves a final test accuracy of 99.2% using eight-view test-time augmentation (TTA), with corresponding precision, recall, and F1-score of approximately 99.1%. The model also achieves micro-average and macro-average ROC–AUC scores of 1.00 on the curated test set, indicating strong discriminative capability across the considered disease categories. Comparative evaluations against MobileNetV2, DenseNet201, ConvNeXt, and LeafNet demonstrate competitive or superior classification performance while maintaining moderate computational complexity and architectural simplicity. Furthermore, Grad-CAM visualizations qualitatively highlight biologically relevant disease symptoms, including lesions, necrotic regions, rust spots, fungal growth, and tissue discoloration, thereby enhancing model interpretability. Although field-scale validation and quantitative explainability assessment remain subjects for future investigation, the proposed framework provides an accurate, computationally efficient, and interpretable approach for automated mango leaf disease diagnosis with strong potential for smart agriculture and precision farming applications.
Email spam detection in English has been studied extensively under full supervision, but for low-resource languages such as Turkish the largest public dataset is more than an order of magnitude smaller, and every existing Turkish spam detector assumes full supervision. This study asks whether spam can be detected from only a handful of labeled examples per class, and how the leading few-shot approaches compare, on the Enron-Spam and Turkish corpora across three encoder backbones and shot sizes from 2 to 64. SetFit, a contrastive few-shot method, attains a macro-F1 of 0.911 on Turkish and 0.947 on English at 64-shot, within 3.8 and 4.7 F1 points of the three-seed full fine-tuning reference. Against additional baselines (TF-IDF, FastFit, LoRA, and in-context prompting of Mistral-7B and Llama-3.1-8B), in-context LLM prompting is strongest at the lowest shot sizes but gains little from more demonstrations, so the fine-tuned methods match it on English by 64-shot and overtake it on Turkish from 32-shot. SetFit is therefore not uniformly best, but it reaches competitive accuracy with a far smaller model and no generative inference. Purely cross-lingual transfer fails, confounding language and domain shift, whereas bilingual training recovers most of the monolingual performance.
Selecting a suitable bioprinter remains a challenging task involving conflicting criteria that are difficult to weigh. Existing fuzzy multi-criteria decision-making (MCDM) methods fail to derive the fuzzy priorities of criteria while maintaining efficiency, and it is also difficult to consider the dependencies among the criteria. This study aims to solve these problems by proposing a flexible fuzzy reinforcement learning MCDM (FFRL-MCDM) approach. In the FFRL-MCDM, first, a nonlinear calibrated fuzzy reinforcement learning (ncFRL) method is devised to accurately and efficiently derive the fuzzy priorities of criteria. Subsequently, realistically feasible reference points are defined by considering dependencies between criteria to modify the fuzzy technique for order preference by similarity to the ideal solution (FTOPSIS) approach. The proposed FFRL-MCDM approach has been applied to a real case in the literature. Experimental results showed that the proposed FFRL-MCDM approach improved the accuracy of deriving fuzzy priorities with extremely high efficiency. Furthermore, the overall performances and ranks of the evaluated bioprinters did change by comparing to more realistic reference points. However, in this case, the scope of the dependency treatment was restricted to pairwise linear dependencies.
Pedestrian detection remains a critical component of autonomous driving and intelligent surveillance systems, providing essential perceptual information for their safe and reliable operation. However, false positives, where non-pedestrian objects or backgrounds are incorrectly classified as pedestrian objects, which can severely compromise normal system functionality. This problem is particularly challenging because pedestrians exhibit significant inter-class and intra-class appearance variations. Consequently, FPs persist as a fundamental and unsolved challenge. This paper addresses this problem by introducing a Full-stage Refined Proposal (FRP) algorithm aimed at eliminating FPs within a two-stage CNN-based pedestrian detection framework. Specifically, in the training stage, the Training mode FRP algorithm introduces a novel approach for validating pedestrian proposals to effectively guide the model training process, thereby constructing a model with strong capabilities for false positive suppression. During the inference stage, two innovative strategies are implemented: the Classifier-guided FRP algorithm integrates a pedestrian classifier into the proposal generation process to yield high-quality proposals through pedestrian feature evaluation, and the Split-proposal FRP algorithm vertically divides all proposals, sending both the original and the sub-region proposals to the subsequent subnetwork to evaluate their confidence scores, filtering out those with lower sub-region pedestrian confidence scores. As a result, the proposed algorithm enhances the model's ability to suppress pedestrian false positives across all stages. Experiments on benchmarks and the SY-Metro dataset demonstrate the FRP algorithm's effectiveness in suppressing FPs, with better improvements in detection accuracy. For instance, on CityPersons, the full-mode FRP lowers the log-average miss rate from 71.87% to 69.27% (a 2.6% absolute improvement) over the plain MetroNext, at the cost of only 5 ms/frame additional latency on GPU. Deployment experiments on embedded platforms, such as the Jetson Nano, highlight the FRP algorithm's potential for real-world applications, offering a competitive solution for pedestrian detection in resource-constrained environments. The source code for this paper is available at https://github.com/guoqiang0148666/fsrp-algorithm .
