Investor sentiment significantly impacts investors’ decision-making behavior and stock market returns. Therefore, how to analyze and utilize investor sentiment for stock price prediction is an important research direction in finance. Many studies have explored the influence of investor sentiment on stock market price fluctuations, but these studies have not comprehensively considered factors such as overall market trends, sector rotation, and individual stock characteristics. Addressing this issue, this paper combines sentiment analysis with deep learning methods to construct a CNN-LSTM-Attention model incorporating sentiment indices, aiming to assist investors in predicting closing prices. We selected feature indicators that reflect sentiment in the overall market, sectors, and individual stock, constructed a formula for calculating sentiment indices, and calculated sentiment indices to further explore the correlation between sentiment indices and stock trends. To enhance prediction accuracy, we designed and constructed a CNN-LSTM-Attention model based on attention mechanisms, integrating convolutional neural networks and long short-term memory neural networks. This model comprehensively utilizes sentiment indices and stock data containing fundamental data, historical trading data, and technical indicators for closing price prediction. Using the stock of “iFlytek” as an example, experimental results demonstrate that compared to models such as RNN and LSTM, this model achieves the best results, indicating that incorporating sentiment indices can effectively improve prediction performance.
This paper introduces a novel approach to improving performance through the integration of cross-modal interaction and attention mechanisms, specifically channel and spatial attention for sentiment analysis tasks. By focusing on the synergistic relationship between textual and visual information, our proposed Cross-Modal Interaction Attention (CMIA) model leverages the strengths of both data types to achieve superior classification accuracy. In this paper we detail the methodology of channel and spatial attention mechanisms, demonstrating how they prioritize relevant features and spatial locations within the data. Through a series of experiments, including an ablation study on the MVSA dataset, we show that our model not only outperforms existing approaches either traditional or SOTA, but also highlights the importance of cross-modal interactions for a deeper understanding of sentiment. Our findings suggest significant potential for the application of attention mechanisms in various domains, pushing the boundaries of current neural network capabilities in sentiment analysis.
The intelligentization of flight control systems is an important research direction in the field of aeronautics and astronautics. This paper designs a hybrid intelligent control strategy based on deep reinforcement learning algorithms, achieving precise control of flight attitude. The performance of the control method is verified through simulation experiments, showing significant improvements in attitude stability, anti-interference ability, and control accuracy compared to traditional PID controllers. The experimental results demonstrate that the new method has a significant improvement in control accuracy and system response time, opening up a new direction for the research of intelligent control systems for aircraft.
Regarding security concerns in image storage and transmission, particularly for medical images, we propose a selective encryption method for quantum color images EQIRHSI, which is built on quantum LoG edge detection and quantum morphological processing. Leveraging HSI color space’s visual perception characteristics, the algorithm first extracts regions of interest (ROI) via quantum LoG edge detection, then refines ROI boundaries through quantum dilation and erosion. Subsequently, the color information of the ROI is encrypted using a double random phase encryption technique, while the intensity information of the ROI is encrypted by combining quantum walks, 2D-SCLMS chaotic system, XOR, and XNOR operations. Finally, quantum circuits are designed and simulation experiments are implemented, demonstrating that this method can efficiently encrypt and transmit the ROI of the image while ensuring image security. After decryption, the location and information of the ROIs are accurately restored, demonstrating high security and practicality.
Data valuation is an emerging area in machine learning that aims to assess the contribution of training data to a model’s performance on unseen data. Among existing approaches, Data-OOB leverages out-of-bag (OOB) estimate to enable efficient data valuation and has shown promising results in tasks such as noisy data detection. However, our analysis reveals two key limitations of Data-OOB: (1) the weak learners it employs are vulnerable to noisy data, and (2) it suffers from information loss during the exploitation of weak learners. To address these issues, we propose a novel data valuation framework called OOB-CM. This method assigns a dynamic threshold to each data point and uses it to define a personalized curriculum for every weak learner. The curriculum selectively filters training samples at each stage of training, shielding weak learners from the influence of noise. Furthermore, we introduce a multi-round voting strategy that aggregates the discriminative signals from weak learners by selecting the top- p predicted classes for each sample. This strategy significantly improves the utilization of weak learners’ predictive power. We conduct comprehensive evaluations across multiple scenarios, and the results consistently demonstrate that OOB-CM outperforms existing methods in terms of robustness and effectiveness.
This paper addresses the problem of anomaly detection in multidimensional time series data, where labeled samples are often scarce and data patterns are complex. We propose a semi-supervised learning algorithm that combines labeled and unlabeled data to enhance detection performance. To mitigate the risk of classifier degradation caused by mislabeled data, we introduce a fuzzy clustering-based selection mechanism. The algorithm selects high-confidence samples from the unlabeled data using fuzzy C-means clustering and iteratively adds them to the training set through a self-training process. Experiments on transformer oil chromatography data show that the proposed method achieves better performance than traditional semi-supervised and supervised approaches. The results demonstrate the effectiveness of integrating clustering techniques into semi-supervised learning for robust anomaly detection under limited label conditions.
Recently, Deep neural networks (DNNs) have attained remarkable achievements across numerous visual recognition tasks. Nevertheless, the existing deep neural network models are characterized by high computational costs and substantial memory usage, which pose significant barriers to their deployment in devices with limited memory resources or applications with strict latency requirements. Consequently, model compression and acceleration for deep networks without causing notable degradation in model performance is in urgent need. This paper provides a comprehensive review of the recent techniques employed for compacting and accelerating DNN models. From the perspective of model architecture, the main approaches are compact structure design and neural architecture search. From an algorithmic dimension, methods are briefly categorized into static compression methods and dynamic acceleration methods, specifically covering implementation strategies such as model pruning, parameter quantization, low-rank factorization, and knowledge distillation. For each category, we demonstrate the development of mainstream methods as well as the characteristics and advantages of each method. We also provide insightful analysis about the integration of multiple methods, their advantages and drawbacks.
Graph neural networks (GNNs) have been widely applied in graph-based learning tasks. However, the feature collection stage is increasingly becoming the bottleneck of the existing GNN systems when training large-scale graphs. To solve this problem, this paper proposes FCGraph, a novel efficient GNN training strategy by combining a feature cache policy of multi-hop (k-hop) neighbors and a hierarchical GPU-centric data access method. First, a k-hop neighbors based cache policy is proposed by exploring vertex data access characteristics in k-layer GNN models, which focuses on reducing CPU-GPU data transfer overhead. Second, based on the cache policy, node features are partitioned according to node access frequency, and a novel GPU-centric data access method is presented for feature collection. Further, FCGraph is scaled to multi-GPU systems equipped with NVLink, in which the memory access hierarchies are explored. The evaluation of some representative datasets shows that FCGraph can improve end-to-end training performance over the advanced GNN training systems of DGL and PyTorch-Direct by 3.72 × and 1.46 × on average, respectively.
Time series data, widely used in industrial and scientific domains, often contain noise such as outliers and missing values, which degrade rule mining performance. To address this, we propose a robust regression rule mining algorithm tailored for time series data. We introduce Approximate Conditional Regression Rules (ACRR), which relax strict rule constraints to uncover approximate attribute relationships. The algorithm combines predicate-based condition generation with linear regression validation, and incorporates modules for outlier detection and iterative missing value imputation. Experiments on real-world datasets demonstrate that our method effectively discovers meaningful rules and maintains robustness in noisy environments.
In the Internet of Vehicles (IoV) environment, the frequent changes in network topology are caused by the high-speed movement of vehicles, the diversity of road layouts, and the limited transmission range of vehicles. Therefore, selecting appropriate routing to ensure robust communication between vehicles is one of the main challenges in IoV. This paper focuses on the research of routing protocols in urban IoV scenarios. By integrating the IoV and software-defined network (SDN) architectures, a software-defined vehicular network architecture is constructed to enhance the reliability and efficiency of information transmission between vehicles. Utilizing the global view of the central controller, this study applies more vehicular information, including link stability, node congestion degree, and packet congestion degree, in route selection to improve path performance. Additionally, to address various complex scenarios in IoV, this paper proposes different deployment strategies, including proactive strategy, proactive-optimization strategy, and proactive-restriction strategy, to flexibly cope with complex scenarios and enhance the quality of communication services. Simulation results indicate that compared to the original passive strategy, the proactive-restriction strategy reduces end-to-end delay by 19
With the continuous advancement of automation technology, six-degree-of-freedom robotic arms are becoming more and more widely used in manufacturing, medical care, logistics, and services. Thanks to their flexibility and efficiency, these robots are gradually replacing traditional manual operations, improving production efficiency and reducing labour costs. However, how to achieve efficient path planning and joint angle optimisation for robots in complex and ever-changing environments remains a major challenge. To address this problem, this paper proposes a comprehensive application solution that combines the A algorithm with the simulated annealing algorithm. The A algorithm uses a heuristic search method that can quickly find the optimal path in a large state space, making it particularly suitable for multi-objective path planning in dynamic environments. The algorithm improves the speed and accuracy of path planning by comprehensively considering actual and heuristic costs. It is highly adaptable and can be adjusted in real time to meet different task requirements. The simulated annealing algorithm, as a global optimization tool, can effectively avoid local optima and help achieve an optimal configuration of joint angles. The algorithm performs well in high-dimensional optimization problems, improves the movement accuracy and efficiency of the robotic arm, and reduces energy consumption. Experimental results show that the path planning and joint angle optimisation strategy combining A* and simulated annealing algorithms can effectively cope with the challenges in complex environments and improve the overall performance of the robotic arm. The research results of this paper provide theoretical support and practical guidance for the application of intelligent robots in complex environments, and have important academic and practical value. Future research can explore the combined application of other optimisation algorithms to further enhance the adaptability of robotic arms in a wider range of scenarios, so as to promote the sustainable development and innovation of robotics.
Massive high-resolution medical images have become an important medium for medical record storage, disease diagnosis, online consultation and remote diagnosis and treatment to ensure people’s health. Traditional full-image encryption technology is an important means of safeguarding patient privacy. However, for massive high-resolution medical images, traditional encryption technology cannot meet the requirements for high processing speed. Due to the significant separation between organ regions and background regions in medical images, local encryption technology for regions of interest (ROI) in medical images is expected to meet the high processing rate requirements for large volumes of medical images. Meantime, the non-destructiveness of decrypted medical images is necessary to ensure the precision medicine. This paper primarily studies the efficient, secure, and lossless encryption and decryption technology for the local ROI in medical images. The binary edge image of the medical image is obtained using U2-Net, thus the original image is divided into ROI and the region of the background (ROB) according to the positional relationship between the original image and the ROI in the binary image. The ROI bit plane is scrambled using the Chen chaotic system, and the key matrix is generated by the Henon chaotic map to diffuse the scrambled bit plane image. The encrypted ROI and the unencrypted ROB are combined into a ciphertext image. Run-length encoding is used to compress and encode the ROI binary image, then the encoding results are randomly embedded into the ciphertext image, which can achieve lossless decryption while improving security.
In response to the suboptimal obstacle avoidance performance of Autonomous Underwater Vehicle clusters (AUV clusters) and the issues of slow convergence speed and poor network performance associated with MATD3 networks, we propose a novel AUV cluster obstacle avoidance algorithm based on DHMATD3. By computing the sliding window average of the iterative TD error during the MATD3 network training process, we dynamically adjust the delayed update frequency of the MATD3 network. Subsequently, we utilize a GRU network to process historical sample data and incorporate it into the experience pool, thereby altering the sampling method of the MATD3 network experience pool to expedite the learning of excellent data. Given that certain outputs of the GRU network may contain irrelevant information, we introduce an attention mechanism to ensure that the GRU network focuses solely on useful information, thereby optimizing the structure of the MATD3 network. We apply this approach to address AUV cluster obstacle avoidance problems. Experimental results demonstrate that this algorithm effectively enhances the performance and convergence speed of the MATD3 network, leading to more optimal AUV cluster obstacle avoidance decisions and increased stability.
Current software development has moved away from the traditional manual workshop model and emphasizes improving software product quality. To address the issue of repetitive work, software reuse techniques can be adopted to continually enhance the quality and efficiency of software development. Software reuse primarily involves reutilizing existing software knowledge during the software development process, effectively reducing maintenance costs incurred during development and controlling the overall software development expenses. Software components are an effective form of supporting software product reuse and serve as the core technology for enabling software reuse. Component-based software engineering techniques emphasize the use of reusable software “components” to design and construct programs, aiming to assemble these components within a software architecture to achieve software reuse and improve the quality and productivity of software products. However, selecting the most suitable components from the results of component retrieval requires understanding the different usages of each component in the retrieval results. The existing methods suffer from excessive reliance on manual approaches and errors caused by inter-component relationships. Therefore, this paper proposes a component selection technique called UEBCS (Usage Example-Based Component Selection). This technique leverages steps such as clustering analysis and hierarchical classification to achieve optimal component selection. UEBCS has shown excellent results in terms of both efficiency and accuracy in selecting components. This method provides technical support for software developers in the software development process and holds significant practical significance for enhancing software development quality and efficiency, as well as promoting the deepening development of the software industry.
Data streams are often subject to concept drift, which can gradually reduce the reliability of learning models over time in data stream mining. To maintain model accuracy and enhance its robustness, it is crucial to detect concept drift and update the learning model accordingly. The majority of drift detection methods rely on the assumption that true labels are immediately available, which is challenging to implement in real-world scenarios. Therefore, it is more practicable to detect concept drift in an unsupervised manner. This paper proposes an unsupervised Drift Detection method based on Stacked Autoencoder and Page-Hinckley test (DD-SAPH). DD-SAPH employs the stacked autoencoder as a medium to represent the distribution of historical data, which extracts hidden features from the reference window. To measure the difference between distributions of historical data and new data, the reconstruction error of the stacked autoencoder on the current window is employed. The Page-Hinckley test dynamically calculates thresholds to warn and alarm concept drift. Experimental results indicate that DD-SAPH outperforms the compared unsupervised algorithms when addressing concept drift on both synthetic and real datasets.
Federated Learning (FL) has emerged as a promising approach for preserving data privacy in recommendation systems by training models locally. Recently, Graph Neural Networks (GNN) have gained popularity in recommendation tasks due to their ability to capture high-order interactions between users and items. However, privacy concerns prevent the global sharing of the entire user-item graph. To address this limitation, some methods create pseudo-interacted items or users in the graph to compensate for missing information for each client. Unfortunately, these methods introduce random noise and raise privacy concerns. In this paper, we propose FedRKG, a novel federated recommendation system, where a global knowledge graph (KG) is constructed and maintained on the server using publicly available item information, enabling higher-order user-item interactions. On the client side, a relation-aware GNN model leverages diverse KG relationships. To protect local interaction items and obscure gradients, we employ pseudo-labeling and Local Differential Privacy (LDP). Extensive experiments conducted on three real-world datasets demonstrate the competitive performance of our approach compared to centralized algorithms while ensuring privacy preservation. Moreover, FedRKG achieves an average accuracy improvement of 4
In the field of intelligent digital healthcare, Continuous-flow microfluidic biochip (CFMB) has become a research direction of widespread concern. CFMB integrates a large number of microvalves and large-scale microchannel networks into a single chip, enabling efficient execution of various biochemical protocols. However, as the scale of the chip increases, the routing task for CFMB becomes increasingly complex, and traditional manual routing is no longer sufficient to meet the requirements. Therefore, this paper proposes an automatic routing framework for CFMB based on Genetic algorithm (GA) and A* algorithms. Specifically, we adopt a two-stage A* algorithm to design the routing between modules, using the routing results obtained from the A* algorithm as the basis for evaluating the quality of solutions in the GA algorithm. Then, the GA algorithm is used to search for the optimal approximate solution in the solution space. Experimental results show that this method can reduce routing length and minimize routing crossings, thereby improving the parallel transmission speed of reagents on CFMB. This approach provides a feasible solution for large-scale automated routing of CFMB in the field of intelligent digital healthcare.
De-rainy has become a pre-processing task for most computer vision systems. Combining recursive ideas to De-rainy models is currently popular. In this paper, the EAPRN model is proposed by introducing the elemental attention mechanism in the progressive residual network model. The elemental attention mainly consists of spatial attention and channel attention, which feature-weight the feature image in both spatial and channel dimensions and combine as elemental attention features. The introduction of elemental attention can help the model improve its fitness for the rain removal task, filter out important network layers and help the network process the rainy image. Experiments show that the EAPRN model has better visual results on different datasets and the quality of the De-rainy image is further improved.
With the development of science and technology, function fitting has penetrated into various fields of scientific research, scientific and technological innovation. For the function of fitting analysis of a given function image, there is no public software on the market that uses the idea of genetic algorithm to solve the problem of function fitting. In order to make up for the insufficiency of the existing software and seize the opportunity of function fitting in various industries, a function fitting system based on genetic algorithm was proposed. It is a low-threshold software with a wide range of applications. When designing the function of fitting and analyzing a given function image, the genetic algorithm is used. With the application of ray detection in Unity, the closest expression of function fitting is obtained. At the same time, the line graph of fitness reduction during genetic algorithm iteration is given.
In situations where ordinary people and the hearing-impaired person need to communicate, it is possible that the average person does not know sign language and thus communication may be impaired, which means that a technology or device to assist communication is needed. Therefore, this study develops a new cloud sign language translation system on smart device based on the Browser/Server architecture, so that when the hearing-impaired person makes a sign language movement in front of the user who using the system with a smart device (e.g., smart glasses), the screen of the smart device will display the subtitle of the sign languages. We use MediaPipe to recognize and collect sign language action data from WLASL dataset and provide it to TensorFlow’s 1D-CNN deep learning model for training, so as to realize the sign language translation function. In the test phase, we invited five experimenters to test the sign language translation system ten times for each person, and the final average accuracy rate was 72
