Scene images generally show the characteristics of large intra-class variety and high inter-class similarity because of complicated appearances, subtle differences, and ambiguous categorization. Hence, it is difficult to achieve a satisfactory accuracy by using a single representation. For solving this issue, we present a comprehensive representation for scene recognition by fusing deep features extracted from three discriminative views including the information of object semantics, global appearance and contextual appearance. These views show diversity and complementarity of features. The object semantics representation of the scene image, denoted by Spatial-layout-maintained Object Semantics Features (SOSF), is extracted from the output of a deep-learning-based multi-classes detector by using Spatial Fisher Vectors (SFV), which can simultaneously encode the category and layout information of objects. A multi-direction Long Short-Term Memory (LSTM)-based model is built to represent contextual information of the scene image, and, the activation of the fully connected layer of a Convolutional Neural Network (CNN) is used to represent the global appearance of scene image. These three kinds of deep features are then fused to draw a final conclusion for scene recognition. Extensive experiments are conducted to evaluate the proposed comprehensive representation on three benchmarks scene image database. The results show that the three deep features complement to each other strongly and are effective in improving recognition performance after fusion. The proposed method can achieve scene recognition accuracy of 89.51% on MIT67 database, 78.93% on SUN397 database, and 57.27% on Places365 databases, respectively, which are better percentages than the accuracies obtained by the latest reported deep-learning-based scene recognition methods.

Accurate and timely forecasting of traffic on local road networks is crucial for deploying effective dynamic traffic control, advanced route planning, and navigation services. This task is particularly challenging due to complex spatio-temporal dependencies arising from non-Euclidean spatial relations in road networks and non-linear temporal dynamics influenced by changing road conditions. This paper introduces the spatio-temporal network embedding (STNE) model, a novel deep learning framework tailored for learning and forecasting graph-structured traffic data over extended input sequences. Unlike traditional convolutional neural networks (CNNs), the model employs graph convolutional networks (GCNs) to capture the spatial characteristics of local road network topologies. Moreover, the segmentation of very long input traffic data into multiple sub-sequences, based on significant temporal properties such as closeness, periodicity, and trend, is performed. Multi-dimensional long short-term memory neural networks (MDLSTM) are utilized to flexibly access multi-dimensional context. Experimental results demonstrate that the STNE model surpasses state-of-the-art traffic forecasting benchmarks on two large-scale real-world traffic datasets.

A novel, multi-dimensional, spatiotemporal prediction framework is proposed to enhance air traffic flow prediction in increasingly complex aviation networks. This framework incorporates graph convolutional networks (GCNs) with multi-dimensional Long Short-Term Memory (LSTM) networks and multi-scale, temporal convolution, employing an attention mechanism to effectively capture spatiotemporal dependencies. By addressing irregular topologies and dynamic temporal trends, the framework models local air traffic patterns with improved accuracy. The experimental results demonstrate significant predictive accuracy improvements over traditional methods, particularly in accounting for the complex nature of air traffic flows. The model’s scalability and adaptability extend its application to various aviation networks, encompassing all airspace units within three local networks, rather than focusing solely on airport traffic. These findings contribute to the development of more intelligent, accurate, and adaptive air traffic management systems, ultimately enhancing both operational efficiency and safety.

Multimodal sentiment analysis (MSA) has exhibited superior performance compared to unimodal sentiment analysis. However, in real-world scenarios, multimodal data are prone to being incomplete due to various interferences. To address this issue, this paper combines cross-modal adaptive knowledge information with adversarial reconstruction of complete-incomplete multimodal representations, proposing a novel neural network model called Dynamic Information Transfer-Adversarial Learning (DIT-AL) for robust MSA of unaligned modal sequences with missing data. First, the model employs a symmetric structurebased attention feature fusion to facilitate local and global feature interactions across modalities, learning multimodal representations for complete-incomplete data pairs. Then, two mechanisms are utilized simultaneously to recover the semantic information of the missing data. The first mechanism dynamically learns the relative relationships between complete and incomplete modal representations through an adaptive information distillation module, bridging the semantic gaps caused by missing modal data and enhancing the representation capacity of incomplete multimodal representations. The second mechanism introduces adversarial learning and feature reconstruction based on paired multimodal representations to generate semantic information for incomplete data, thereby achieving consistency in the feature distribution between complete and incomplete data. Experimental results on three public benchmarks (CMU-MOSI, CMU-MOSEI, and CHSIMS) demonstrate that the proposed method can effectively improve the robustness of MSA under random data missing conditions by using adaptive information distillation combined with adversarial reconstruction. Compared with state-of-the-art methods, it achieves higher recognition accuracy for robust MSA

Semantic matching is essential for understanding natural language, but traditional models like BERT face challenges with random masking strategies, limiting their ability to capture key information. Additionally, BERT’s sentence vectors may “collapse,” making it difficult to distinguish between different sentences. This paper introduces a domain-finetuned semantic matching framework that uses dynamic masking and contrastive learning techniques to address these issues. The dynamic masking strategy enhances the model’s ability to retain critical information, while contrastive learning improves sentence vector representations using a small amount of unlabeled text. This approach helps the model better align with the needs of various downstream tasks. Experimental results show that after private domain training, the model improves semantic similarity between entities by 16.9%, outperforming existing models. It also demonstrates an 8.0% average improvement in semantic matching for diverse text. Performance metrics such as A@1, A@3, and A@5 are at least 26.1% higher than those of competing models. For newly added entities, the model achieves a 44.3% average improvement, consistently surpassing other models by at least 30%. These results collectively validate the effectiveness and superiority of the proposed framework in domain-specific semantic matching tasks.