The fast growth of digital content in streaming systems has made the problem of too much information more serious. It also brings big challenges to traditional recommendation methods that use experience-based similarity or simple neural network models. We put forward an improved graph convolution framework for personalized movie recommendations to solve three key problems: poor expandability, sparse data, and difficulties in modeling high-level interactions. First, we build a bipartite graph of users and items based on a big dataset of movie ratings. Then we use a simple multi-layer graph convolution method to get high-level collaborative information through standardized neighborhood spread. Different from standard LightGCN models that use inner-product calculation for scoring, our method combines an MLP prediction module with Batch Normalization, non-linear activation functions and Dropout regularization. This design lets us model the interactions between users and items more clearly and keeps the system structure efficient at the same time. The test results from big interaction data sets show the model has steady convergence and good generalization ability. It also gets competitive results in top-K recommendation tests, and there is no obvious overfitting during the training process. We find that mixing simple graph spread with non-linear prediction can improve both the ability to show data features and recommendation precision in big and sparse data environments. This research provides a framework that can expand well for recommendation systems with better structure. It also lays a good base for the future combination of graph learning and semantic model building.

Remaining Useful Life (RUL) prediction is crucial for predictive maintenance in complex engineering systems. In recent years, deep learning methods have become the dominant approach for RUL prediction due to their ability to capture complex temporal dependencies. Long Short-Term Memory (LSTM) networks, originally designed for sequence modeling, have been widely applied in time-series prediction tasks. The Transformer architecture, known for its powerful attention mechanism, has achieved remarkable success in various sequential data analysis domains. However, these methods typically assume a single global degradation pattern, which may limit their performance under varying operating conditions. To address this issue, this paper presents a two-fold investigation: first, a comparative performance analysis of three prominent architectures—LSTM, Transformer, and Mixture of Experts (MoE). Second, we focus on the optimization of the MoE framework by proposing a Regime-Aware MoE (RA-MoE). This model integrates regime identification techniques (K-Means, HMM, and VAE) to optimize the gating mechanism. Experimental results show that while LSTM remains the most robust performer among the candidate architectures, the proposed RA-MoE significantly enhances the performance of the standard MoE architecture, demonstrating the effectiveness of regime-aware optimization in complex scenarios.

Risks in supply chains and financial networks propagate through graph structure; early warning and targeted intervention are critical for mitigating cascading failures. Existing methods suffer from the prediction–intervention decoupling: prediction models are trained without awareness of downstream intervention decisions, and intervention policies rely on fixed heuristics (e.g., hand-tuned mixing weights) that do not adapt to data. We propose a decision-aware framework that couples GNN-based risk prediction with a data-adaptive intervention policy—the mixing weight between predicted risk and structural centrality is learned from validation performance rather than being fixed a priori. Epidemic dynamics (SIR, LTM) provide features; a GCN backbone predicts node risk; the intervention policy is selected by maximizing intervention benefitΔon a validation set. Experiments on Email-Enron, Facebook, and Wiki-Vote show that the adaptive policy achieves AUC 0.78 andΔ22.4%, outperforming fixed-heuristic baselines (centrality-only15.8%, prediction-only19.6%) and retaining advantage under edge noise and limited observation. We argue that learning to intervene, which closes the loop between prediction and intervention via validation-driven policy selection, is a principled step toward decision-aware risk management.

Adolescent mental health has emerged as a critical public health challenge, with traditional screening methods often hindered by time lags and limited reach. As social media becomes a primary channel for emotional expression among youth, automated sentiment analysis offers a promising pathway for real-time monitoring. This study constructs an automated recognition system using a large-scale corpus of 52,573 labeled social media entries across seven psychological dimensions, including depression, anxiety, suicidal ideation, and stress. By employing the TF-IDF algorithm for multi-dimensional feature extraction and a Logistic Regression model for multi-class classification, the proposed scheme achieves an overall recognition accuracy of 74.8%. Experimental results reveal significant linguistic patterns across mental states: the "normal" category exhibited the highest discriminability (F1-score = 0.896), while the "stress" category proved the most challenging to identify (F1-score = 0.553) due to its semantic overlap with daily emotional fluctuations. Feature analysis further confirms that specific "psychological fingerprints"—such as the high frequency of first-person pronouns in the depression group and uncertainty-related queries in the anxiety group—can serve as reliable predictors. This research validates the feasibility of large-scale, non-invasive psychological screening and provides a data-driven framework for early campus crisis intervention and precise psychological support.

Amid the rapid advancement of edge machine learning, real-time, low-latency image recognition on resource-constrained mobile devices is increasingly in demand for intelligent daily scenarios yet existing research lacks focus on lightweight models for common small daily objects. This study addresses the research gap by exploring efficient recognition of 6 daily objects (Airpods, bottles, lipsticks, etc.) based on edge computing. A 220-image dataset (80% training, 20% testing) was built with preprocessing and augmentation. Lightweight CNN models (MobileNetV1/V2, EfficientNet) were trained via transfer learning on Edge Impulse with a control variable method. Results show that EfficientNet achieved 95.8% test accuracy without overfitting, while MobileNetV2 (160×160) achieved an optimal balance of accuracy (87.5%), speed, and size. Transfer learning resolves small-sample issues, and data augmentation boosts generalization. The optimized model was successfully deployed for real-time mobile recognition.

Recent advances in large language models (LLMs) have substantially improved their ability to perform complex reasoning tasks, with some models reaching or exceeding human expert performance in specific domains. Despite this progress, a systematic and comprehensive synthesis of existing research on LLMs' scientific reasoning capabilities remains limited. Existing studies either focus on narrow subdomains or single aspects of the field, lacking integration of benchmarks, models, and evaluation methodologies. To address this gap, this survey provides a structured review of three core aspects of scientific reasoning in LLMs: (1) benchmark datasets used to evaluate scientific reasoning performance, (2) representative LLMs exhibiting differing reasoning capabilities, and (3) evaluation methodologies designed to assess both reasoning outcomes and reasoning processes. In addition, this work analyses key challenges that constrain current progress, including hallucinations, domain-specific data scarcity, and inefficient overthinking in multi-step reasoning. This survey aims to provide a coherent reference framework for researchers and practitioners engaged in the development, evaluation, and application of LLMs for scientific reasoning.

Early diagnosis of thyroid diseases has challenges due to limitations of single-modal data. Benign and malignant thyroid nodules overlap in ultrasonographic features, and clinical data is crucial for imaging interpretation and practice. This study proposes a multimodal fusion strategy that integrates thyroid ultrasound images with clinical data features to improve nodule prediction accuracy. An EfficientNet - b4+BERT fusion model with multi - head self - attention mechanisms is developed for dynamic weighted feature vector fusion, which enhances image - text feature alignment efficiency by leveraging inter - modal complementarity. Comparative experiments show the model outperforms single models in accuracy and stability. The proposed technology performs well in high - risk thyroid tumor prediction and has significant clinical application value.

Optimization plays a fundamental role in machine learning, as most learning tasks can be formulated as the minimization of a loss function. From classical gradient descent to modern convex optimization theory, optimization algorithms have continuously evolved to meet the demands of large-scale data and high-dimensional models. This paper reviews the development of optimization methods in machine learning, focusing on gradient descent and its variants, stochastic optimization, and convex optimization theory. Through literature analysis, this study examines the theoretical foundations, convergence properties, and practical applications of these methods. Specifically, the research addresses three key questions: how gradient-based methods have evolved, what advantages convex optimization provides, and what challenges arise in non-convex optimization. The paper concludes that convex optimization offers strong theoretical guarantees, while gradient-based algorithms dominate practical large-scale machine learning tasks, especially in deep learning.

With the continuous expansion of the application of UAV swarms in tasks such as post-disaster search and rescue, environmental inspection, agriculture and forestry monitoring, and low-altitude security, the path planning problem has evolved from shortest-path search for a single UAV in static and known environments into a cooperative decision-making problem for multiple UAVs in dynamic and unknown environments. Ant colony optimization has the advantages of distributed search, strong global optimization capability, and easy integration into path cost functions, but it suffers from slow convergence, proneness to local optima, and insufficient adaptability to continuous spaces in complex environments. Deep reinforcement learning can achieve online decision-making and adaptive obstacle avoidance through interaction with the environment, making it more suitable for handling dynamic obstacles, partial observations, and multi-UAV cooperative tasks; however, it also faces challenges such as low sample efficiency, complex reward design, and limited generalization ability. Focusing on cooperative path planning for UAV swarms, this paper systematically reviews the research progress of ant colony optimization, deep reinforcement learning, and their hybrid methods, with emphasis on comparing the differences between the two types of methods in terms of environment modeling, coordination mechanisms, path quality, real-time response, and engineering deployability. On this basis, it further summarizes the key bottlenecks in current research, including unified modeling of complex constraints, global-local coordination interfaces, communication and information sharing, and simulation-to-reality transfer. The review concludes that constructing a hierarchical hybrid framework in which ant colony optimization is responsible for global candidate path generation and deep reinforcement learning is responsible for local online correction is an important development direction for improving the performance and engineering applicability of UAV swarm path planning.