Bin Yang

Hypertension is a leading contributor to cardiovascular morbidity worldwide. When blood pressure and routine clinical variables are available during clinic visits or health checks, guideline thresholds support binary classification but do not necessarily provide reliable calibrated probabilities for decision support. Machine learning studies in hypertension have concentrated on predicting future disease incidence rather than binary classification at the time of measurement, and calibration and validation design are reported inconsistently across this literature. This study examines whether machine learning, trained under a stratified protocol with held‐out calibration and test subsets, can improve probabilistic reliability in post‐measurement classification while remaining comparable to a clinical blood pressure rule. We analysed a public hypertension dataset of 4240 records with a positive‐class prevalence of 31.1% using a scikit‐learn pipeline that integrates preprocessing, recursive feature elimination, Optuna‐based hyperparameter tuning, and isotonic calibration on a held‐out calibration set. A tuned decision tree was compared against a clinical rule based on systolic blood pressure of at least 140 mmHg or diastolic blood pressure of at least 90 mmHg across 30 stratified random splits. Accuracy and recall were comparable to the clinical rule, with no statistically significant Wilcoxon differences, whereas Brier score and expected calibration error were significantly lower for the machine learning model, with mean values of 0.081 versus 0.090 and 0.027 versus 0.084. Interpretability analysis was consistent with blood‐pressure‐related predictors dominating model attributions. These findings suggest that machine learning classification after blood pressure measurement aligns with guideline‐based decisions while offering better calibrated probabilities for clinical decision support.

Along with the development of Internet of Things systems, numerous edge intelligent devices can obtain a large amount of user data, and the analysis of this user data can be applied to business scenarios, such as user click-through rate (CTR) prediction. In the recommendation, advertising and other scenarios, the users at the edge have high response requirements for CTR prediction model training and inference. In the current edge scenario of CTR prediction, there are problems of overly complex model structure and highly sparse original features, which makes it difficult to deploy CTR prediction models at the edge. Therefore, we propose KD-based graph attention FI model (KD-GAFIM), a lightweight recommendation algorithm that combines graph neural networks (GNNs) with knowledge distillation (KD). The approach uses graph attention networks to flexibly capture feature dependencies in a way that maintains a small model size while augmenting the feature vector with feature dependencies. And by sharing the embedding layer of the teacher model, KD-GAFIM improves the efficiency of user CTR prediction. On top of that, we also propose a feature compression strategy guided by model interpretability, which identifies high-contributing features for inference and model refinement based on their performance in model interpretability. This strategy improves efficiency, making KD-GAFIM suitable for training and inference on edge devices. We conducted extensive experiments on multiple datasets. The experimental results show that KD-GAFIM outperforms various state-of-the-art CTR prediction models, demonstrating that GNN-based KD models can improve model performance while reducing model size and feature dimensionality, and have significant potential for application at the edge.

In recent years, collaborative classification of multimodal data, e.g., hyperspectral image (HSI) and light detection and ranging (LiDAR), has been widely used to improve remote sensing image classification accuracy. However, existing fusion approaches for HSI and LiDAR suffer from limitations. Fusing the heterogeneous features of HSI and LiDAR proved to be challenging, leading to incomplete utilization of information for category representation. Additionally, during the extraction of spatial features from HSI, the spectral and spatial information are often disjointed. It leads to the difficulty of fully exploiting the rich spectral information in hyperspectral data. To address these issues, we proposed a multimodal data fusion framework specifically designed for HSI and LiDAR fusion classification, called Modality Fusion Vision Transformer (MFViT). We have designed a stackable modality fusion block (MFB) as the core of our model. Specifically, these blocks mainly consist of multimodal cross-attention modules and spectral self-attention modules. The proposed novel multimodal cross-attention module for feature fusion addresses the issue of insufficient fusion of heterogeneous features from HSI and LiDAR for category representation. Compared to other cross-attention methods, it reduces the alignment requirements between modal feature spaces during cross-modal fusion. The spectral self-attention module can preserve spatial features while exploiting the rich spectral information and participating in the process of extracting spatial features from HSI. Ultimately, we achieve overall classification accuracies of 99.91%, 99.59%, and 96.98% on three benchmark datasets respectively, surpassing all state-of-the-art methods, demonstrating the stability and effectiveness of our model.

Fine-grained biometric image recognition aims to achieve classification of subclasses by processing detailed features, which is still a critical problem to be solved in computing due to the small differences between subclasses. In recent years, Transformer model, which was originally used in natural language processing, has been applied to computer vision. The transformer model splits the image into patches and calculates the weights between different parts to obtain a better feature representation. In this paper, we propose a model of transformer for fine-grained biometric image recognition. Specifically, in the process of patch coding by the model, our model generates corresponding weights for all patches, and saves corresponding attention scores. To verify the effectiveness of our method, we conducted experiments on the CUB-200-2011 and Stanford Dog datasets.

Aspect-based sentiment analysis (ABSA) plays a significant role in the field of big data and aims to distinguish the sentiment polarity of specific aspects in given sentences; however, the previous works on ABSA had two limitations. They mainly considered semantic features, rather than syntactic dependency features, and paid too much attention to the context words, while ignoring the high-level interaction of multiple representations of aspects themselves. To cope with these limitations, we propose a new method based on the graph convolutional network (GCN) and the multi-head attention mechanism, called the attention interactive encoder network (AIEN). The GCN was used to obtain the syntactic information that has the greatest syntactic impact on the aspect based on the syntax dependency tree. The multi-head attention mechanism can not only obtain the context-aware information of the given aspects, but also the interaction information between multiple representations of the aspect itself. The high-level information generated by the interaction of multi-dimensional features can produce a stronger representation ability for the aspect. Our experiments with the proposed model on five benchmark datasets showed that our model outperformed other works significantly. The experimental results further demonstrated the feasibility and applicability of our proposed model in the ABSA task.

With the rapid development of satellite and internet of things (IoT) technology, it becomes more and more convenient to acquire high-resolution satellite images from the ground. Extraction of urban vegetation from high-resolution satellite images can provide valuable suggestions for the decision-making of urban management. At present, deep-learning semantic segmentation has become an important method for vegetation extraction. However, due to the poor representation of context and spatial information, the effect of segmentation is not accurate. Thus, vegetation extraction based on Deep Attention Model (VEDAM) is proposed to enhance the context and spatial information representation ability in the scenario of vegetation extraction from satellite images. Specifically, continuous convolutions are used for feature extraction, and atrous convolutions are introduced to obtain more multi-scale context information. Then the extracted features are enhanced by the Spatial Attention Module (SAM) and the atrous spatial pyramid convolution functions. In addition, image-level feature obtained by image pooling encoding global context further improves the overall performance. Experiments are conducted on real datasets Gaofen Image Dataset (GID). From the comparative experimental results, it is concluded that VEDAM achieves the best mIoU (mIoU = 0.9136) of vegetation semantic segmentation.

As the typical application of computational intelligence in software engineering, cross-project defect prediction (CPDP) uses labeled data from other projects (source projects) for building models to predict the defects in the current projects (target projects), helping testers quickly locate the defective modules. But class imbalance and different data distribution among projects make CPDP a challenging topic. To address the above two problems, we propose a two-phase feature importance amplification (TFIA) CPDP model in this paper which can solve these two problems from domain adaptation phase and classification phase. In the domain adaptation phase, the differences in data distribution among projects are reduced by filtering both source and target projects, and the correlation-based feature selection with greedy best-first search amplifies the importance of features with strong feature-class correlation. In the classification phase, Random Forest works as the classifier to further amplify the importance of highly correlated features and establish a model which is sensitive to highly correlated features. We conducted both ablation experiments and comparison experiments on the widely used AEEEM database. Experimental results show that TFIA can yield significant improvement on CPDP. And the performance of TFIA CPDP model in all experiments is stable and efficient, which lays a solid foundation for its further application in practical engineering.

The visual cues from multiple support regions of different sizes and resolutions are complementary in classifying a candidate box in object detection. Effective integration of local and contextual visual cues from these regions has become a fundamental problem in object detection. In this paper, we propose a gated bi-directional CNN (GBD-Net) to pass messages among features from different support regions during both feature learning and feature extraction. Such message passing can be implemented through convolution between neighboring support regions in two directions and can be conducted in various layers. Therefore, local and contextual visual patterns can validate the existence of each other by learning their nonlinear relationships and their close interactions are modeled in a more complex way. It is also shown that message passing is not always helpful but dependent on individual samples. Gated functions are therefore needed to control message transmission, whose on-or-offs are controlled by extra visual evidence from the input sample. The effectiveness of GBD-Net is shown through experiments on three object detection datasets, ImageNet, Pascal VOC2007 and Microsoft COCO. This paper also shows the details of our approach in wining the ImageNet object detection challenge of 2016, with source code provided on \url{https://github.com/craftGBD/craftGBD}.