Recent research has made huge progress in sentiment analysis, however, challenges remain in balancing model complexity with performance, also how to train models to better fit into conversations is another problem. This research proposes a novel algorithm for sentiment analysis called BertmUCB. It leverages Reinforcement learning's capability to exploit historical feedback and language models' efficiency to gain better performance, and also considers the feasibility of deploying into reality. The author first fine-tunes a RoBERTa model with a two-layer Multi-Layer Perceptron (MLP) classifier on top of the simplyweibo_4_moods dataset, then, a prediction enhancement method optimized by a Proportion-Integration-Differentiation (PID) controller is applied to the MLP's output. PID adapts parameters in prediction enhancement to better fit different datasets. After prediction enhancement, the author uses a modified Upper Confidence Bound (UCB) algorithm to perform the arm selection. In the experiment, BertmUCB outperforms language models, especially on datasets where language models struggle to distinguish fine-grained sentiments, and an ablation study is conducted to quantify the contribution of each module.

Stroke risk prediction is critical for identifying at-risk populations in healthcare as well as for early diagnosis and resource optimization in stroke management. This study's main goal was to create a reliable stroke prediction model by analyzing a publically accessible dataset containing 4981 records and 11 variables using machine learning techniques. Initial data preprocessing consisted of converting categorical variables to numerical values and addressing class imbalance through methods such as class weighting and the Synthetic Minority Oversampling Technique (SMOTE). Three basic models - logistic regression, decision trees, and random forests - were implemented to establish base performance. To improve prediction accuracy and address imbalances, optimized ensemble models were built using stacking and hyperparameter tuning through grid search and cross-validation. This paper evaluates the performance of the model based on composite indicators. While the base model demonstrated high accuracy and recall, the optimized model had a superior performance with an AUC score of 0.94, showing that the capacity to distinguish between stroke and non-stroke cases has significantly improved. By offering a useful predictive tool that can precisely estimate the risk of stroke and direct creative ways to early intervention, this study advances the field of healthcare analytics.

With the development of artificial intelligence and related fields, technologies for image recognition and tracking objects have been continuously innovating. Starting from the initial single-object tracking, through numerous research improvements, multi-object tracking technology has been applied in various fields today, such as autonomous vehicles, security surveillance, and industrial automation production lines, all of which require the capability of multi-object tracking. Multi-object tracking technology can accomplish tasks that people cannot, significantly improving work efficiency and accuracy, greatly facilitating people's lives. However, it also faces many challenges, one of which is the target overlap problem during tracking, which largely affects the tracking effect. This paper will introduce several solutions and conduct a simple analysis and comparison. Existing methods are basically starting from the causes of target overlap, solving the problem at its source. It is believed that there will be more and better solutions to this problem in the future, ultimately making the multi-object tracking process smoother and more accurate.

This paper explores effective practices for transforming Master's programs into an online format, focusing on the Software Engineering and Cybernetics program at Siberian Federal University. The study's relevance stems from the rapid digitalization of education and growing interest in online formats, making Master's programs more accessible to working students. We present an analysis of successful practices for implementing online courses, including structural elements, design, content, multimedia, teaching methods, student interaction, assessment, and feedback. Special attention is given to the selection of an online learning platform, with Moodle being justified as the most suitable system for implementing the Master's program. The paper also presents a portrait of an online master's student, helping to identify needs and expectations, which aids in optimizing the educational process. The study emphasizes the importance of applying hybrid methods and complex systems approaches to create a high-quality educational experience and achieve program goals.

It is difficult to collect disease pictures of the plants due to the seasonal and regional features of diseases which can lead to poor effect of recognition models. Moreover, conventional enhancement techniques like image rotation and scaling often fail to produce images that accurately reflect real disease characteristics. Taking tomato late blight disease as an example, its typical pathological features demonstrate a complex textural structure and spatial distribution pattern. This study proposes a diffusion model with an attention mechanism. By building two models, basic and enhanced, this study explores the effect of class embedding, attention mechanism and pathology feature simulation on the quality of generation. The results show that the pictures generated by the enhanced model are significantly better than the basic model in terms of rationalization of pathological features. The FID and KID values of the healthy and late blight categories are 74, 68, 0.21, 0.18 respectively, showing that the synthetic data were able to simulate the visual characteristics of real diseases to some degree. This study provides a low-cost data augmentation scheme for disease recognition model training under small sample conditions, which helps to improve the robustness of classification models.

This paper studied a facial emotion recognition model based on the fusion technology of Alpha algorithm and Convolutional Neural Network. By combining the spatial feature extraction ability of CNN with the advantages of alpha algorithm in sequence modeling, the performance of emotion recognition model was improved. This paper introduces the theoretical basis of deep learning and reinforcement learning, and proposes a model combining CNN and Alpha-like algorithm. The experimental results show that the accuracy of the fusion model is improved by about 2.1 times in the emotion classification task, especially in the recognition of anger and disgust, which is significantly higher than the traditional algorithm. However, it is also pointed out that the misclassification problem of the model on some complex emotion categories such as surprise and neutral still exists, and the performance of the model needs to be further improved in the future.

This paper explores the integration of computer vision techniques within medical database management systems, focusing on their potential to enhance data accuracy and operational efficiency in healthcare. The study analyzes current methodologies and proposes enhancements based on recent technological advancements, particularly in graphics processor architectures. The research employs a combination of theoretical analysis and computational modeling, with a detailed examination of the G80 graphics processor architecture. The Viola-Jones method is utilized as a model algorithm for object detection in medical images, implemented using the OpenCV library. The study evaluates the algorithm's performance on 1024x1024 pixel images, varying parameters such as sliding window size, scaling factor, and number of classifiers. A novel method for locating key areas in medical images using Haar-like and contour features is proposed and evaluated on a dataset of MRI images. The results demonstrate superior performance compared to reference techniques, maintaining over 90% accuracy in locating key areas even with a database of 200 images. This high accuracy is attributed to expanded Haar-like feature templates, efficient computation using integral images, and a comprehensive contour feature extraction process. The paper concludes that the proposed method shows promise for enhancing medical image analysis and computer-aided diagnosis.

This paper presents the design and implementation of a Retrieval-Augmented Generation (RAG) question-answering system based on a Streamlit web application. The system allows users to upload one or more PDF files and ask questions in natural language. It retrieves and indexes PDF content using FAISS for semantic vector search and BM25 for keyword-based retrieval, combining both methods to improve accuracy. A cross-encoder is used to rerank the retrieved results, ensuring that the most relevant passages are used to generate the final answer. The OpenAI language model then produces short and clear responses grounded in the retrieved context. The system also supports notes generation, multilingual interfaces in English and Chinese, and memory for continuous dialogue. Experiments conducted on academic PDFs show that hybrid retrieval with reranking performs better than single retrieval methods, giving more precise and meaningful answers. The Streamlit interface and Cloudflare tunnel make the system easy to deploy and share. This project demonstrates how RAG technology can effectively extract information from complex documents and provide practical tools for education, research, and enterprise applications.

With the gradual development of sports data analysis, data-driven game prediction has gradually become an important tool for improving the decision-making efficiency of teams and coaches. Basketball, as a team sport, is influenced by multiple factors, and traditional statistical analysis methods make it difficult to intuitively identify the factors that have the greatest impact on the game. This study proposes a reinforcement learning model based on the Proximal Policy Optimization (PPO) algorithm for predicting the winning rate of NBA games. By collecting career statistics of NBA players and combining them with the team's current season winning rate, a feature vector containing individual player characteristics and team winning rate is constructed. The model is trained using the PPO algorithm to identify the most critical features that affect the winning rate of games. This study not only provides new ideas for sports event prediction, but also provides solutions for future decision-making and dataization in big data environments.

With increasing global attention on environmental protection and carbon neutrality goals, waste management has become a crucial component of achieving sustainability. Traditional waste disposal methods, such as manual sorting, identification, and crushing, often suffer from low efficiency and inconsistency. This study proposes a waste management system that integrates Artificial Intelligence (AI) to enhance the accuracy and efficiency of waste treatment through automation. The proposed system addresses waste classification, recognition, and fragmentation, utilizing advanced AI technologies like Graph Neural Networks (GNN), Convolutional Neural Networks (CNN), Visual Transformers (ViT), and Deep Reinforcement Learning (DRL). To optimize the waste management process, the study used the KTH-TIPS dataset, which contains 800 high-resolution images of various material types. Robust model training was ensured through a series of preprocessing and data augmentation techniques. The experimental results demonstrated that CNN and Transformer architectures, including MobileNet, ResNet, and ViT, achieved high accuracies of 100.00%, 100.00%, and 96.91%, respectively. While the GNN and VGG architectures had slightly lower accuracies of 82.88% and 84.57%, they still demonstrated competitive performance. The experimental results illustrate the variation in training and testing losses across different models over the training cycles, revealing the learning dynamics and efficiency of each model.