Gas explosion has always been an important factor restricting coal mine production safety. The application of machine learning techniques in coal mine gas concentration prediction and early warning can effectively prevent gas explosion accidents. Nearly all traditional prediction models use a regression technique to predict gas concentration. Considering there exist very few instances of high gas concentration, the instance distribution of gas concentration would be extremely imbalanced. Therefore, such regression models generally perform poorly in predicting high gas concentration instances. In this study, we consider early warning of gas concentration as a binary-class problem, and divide gas concentration data into warning class and non-warning class according to the concentration threshold. We proposed the probability density machine (PDM) algorithm with excellent adaptability to imbalanced data distribution. In this study, we use the original gas concentration data collected from several monitoring points in a coal mine in Datong city, Shanxi Province, China, to train the PDM model and to compare the model with several class imbalance learning algorithms. The results show that the PDM algorithm is superior to the traditional and state-of-the-art class imbalance learning algorithms, and can produce more accurate early warning results for gas explosion.
Personalized tag recommendation algorithms generate personalized tag lists for users by learning the tagging preferences of users. Traditional personalized tag recommendation systems are limited by the problem of data sparsity, making the personalized tag recommendation models unable to accurately learn the embeddings of users, items, and tags. To address this issue, we propose a contrastive learning-based personalized tag recommendation algorithm, namely CLPTR. Specifically, CLPTR generates augmented views of user–tag and item–tag interaction graphs by injecting noises into implicit feature representations rather than dropping nodes and edges. Hence, CLPTR is able to greatly preserve the underlying semantics of the original user–tag or the item–tag interaction graphs and avoid destroying their structural information. In addition, we integrate the contrastive learning module into a graph neural network-based personalized tag recommendation model, which enables the model to extract self-supervised signals from user–tag and item–tag interaction graphs. We conduct extensive experiments on real-world datasets, and the experimental results demonstrate the state-of-the-art performance of our proposed CLPTR compared with traditional personalized tag recommendation models.
Neural networks have been widely employed in the field of object detection. Transformers enable effective object detection through global context awareness, modular design, scalability, and adaptability to diverse target scales. However, small object detection requires careful consideration due to its comprehensive computations, data requirements, and real-time performance challenges. To address these issues, we present SwinVision , an innovative framework for small object detection in low-light environments. This research shows a balanced approach between computational efficiency and detection accuracy for advancing object detection in low-light scenarios. Firstly, a Swin Transformer-based computing network is introduced and optimized for object detection in large-scale areas. The framework balances computational power and resource efficiency, surpassing conventional transformers. Secondly, we present the STLE module, which enhances the features of low-light images for beneficial object detection. The last building block is a specialized Swin-based detection block for accurate detection of small, detailed objects in resource-constrained scenarios. Experiments conducted on the VisDrone dataset significantly ameliorated existing methods such as YOLOv8x, with a 6.31% increase in mAP and 12.55% in AP50. SwinVision’s effectiveness in low-light environments, especially with small objects, establishes a foundation for robust detection systems adapting to various environmental challenges.
4 July 2024 in IEEE Transactions on Computational Social Systems
The goal of sequential recommendation algorithms is to predict personalized sequential behaviors of users (i.e., next-item recommendation). Learning representations of entities (i.e., users and items) from sparse interaction behaviors and capturing the relationships between entities are the main challenges for sequential recommendation. However, most sequential recommendation algorithms model relationships among entities in Euclidean space, where it is difficult to capture hierarchical relationships among entities. Moreover, most of them utilize independent components to model the user preferences and the sequential behaviors, ignoring the correlation between them. To simultaneously capture the hierarchical structure relationships and model the user preferences and the sequential behaviors in a unified framework, we propose a general hyperbolic translation-based sequential recommendation framework, namely HTSR. Specifically, we first measure the distance between entities in hyperbolic space. Then, we utilize personalized hyperbolic translation operations to model the third-order relationships among a user, his/her latest visited item, and the next item to consume. In addition, we instantiate two hyperbolic translation-based sequential recommendation models, namely Poincaré translation-based sequential recommendation (PoTSR) and Lorentzian translation-based sequential recommendation (LoTSR). PoTSR and LoTSR utilize the Poincaré distance and Lorentzian distance to measure similarities between entities, respectively. Moreover, we utilize the tangent space optimization method to determine optimal model parameters. Experimental results on five real-world datasets show that our proposed hyperbolic translation-based sequential recommendation methods outperform the state-of-the-art sequential recommendation algorithms.
Synthetic Aperture Radar (SAR) plays a vital role in ship safety monitoring and marine environmental protection. However, ship targets in SAR images are often small, have blurred edges, and suffer from strong background interference, making it difficult for traditional detection algorithms to balance accuracy and real-time performance. To address these challenges, this paper proposes LSD-Det, a lightweight SAR ship detection model improved from YOLOv8n. First, the YOLOv8n architecture is streamlined to reduce redundancy, enhancing suitability for SAR detection. Then, a Grouped Split Enhanced Channel Attention (GSECA) module is introduced in the backbone, combining average and max pooling with channel shuffle to improve recognition of small targets and suppress background noise. Additionally, a Global-Enhanced Dilated Wavelet Transform (GEDWT) module is embedded in the neck’s C2f structure to enhance multi-scale feature representation with minimal computational overhead. Furthermore, the original CIoU loss is replaced with PIoUv2, which accelerates convergence and improves bounding box regression accuracy. Experiments conducted on two publicly available SAR ship detection datasets, SSDD and HRSID, demonstrate that LSD-Det significantly reduces computational cost while maintaining high detection performance. Specifically, parameters are reduced by 65.7%, GFLOPs by 20.7%, and inference speed is notably improved. LSD-Det also improves mAP on SSDD by 1.2% (mAP@0.5) and 2.2% (mAP@0.5:0.95), with corresponding gains of 0.8% and 1.3% on HRSID. The experimental outcomes demonstrate that the proposed approach maintains a desirable trade-off between precision and computational efficiency, making it suitable for real-time SAR ship detection under limited-resource conditions.