Domain generalizable person Re-identification methods (DG person ReID) often focus on aligning source-domain distributions to learn domain-invariant features. However, they commonly overlook the pivotal role of hard samples and subtle yet important appearance differences across domains in refining decision boundaries, leading to suboptimal performance when encountering visually similar pedestrians. To address these issues, we propose a framework that combines adaptive uncertainty-driven hard identity mining with a multi-level feature propagation strategy. Specifically, we first identify and prioritize challenging samples across domains through an uncertainty estimation mechanism coupled with dynamic cross-domain weight allocation, thereby adaptively enhancing the model’s discriminative capability for confusable identities. We then incorporate both cross-domain and intra-domain feature propagation to integrate fine-grained information across multiple source domains, strengthening the model’s adaptability to unseen target domains. Extensive experiments on nine real-world benchmarks demonstrate that our method consistently outperforms state-of-the-art DG person ReID approaches, achieving up to 5.3% mAP and 6.6% Rank-1 improvements over the leading baselines.

Single Domain Generalization (SDG) aims to develop models capable of generalizing to unseen target domains using only one source domain, a task complicated by substantial domain shifts and limited data diversity. Existing SDG approaches primarily rely on data augmentation techniques, which struggle to effectively adapt training dynamics to accommodate large domain shifts. To address this, we propose LEAwareSGD, a novel Lyapunov Exponent (LE)-guided optimization approach inspired by dynamical systems theory. By leveraging LE measurements to modulate the learning rate, LEAwareSGD encourages model training near the edge of chaos, a critical state that optimally balances stability and adaptability. This dynamic adjustment allows the model to explore a wider parameter space and capture more generalizable features, ultimately enhancing the model's generalization capability. Extensive experiments on PACS, OfficeHome, and DomainNet demonstrate that LEAwareSGD yields substantial generalization gains, achieving up to 9.47\% improvement on PACS in low-data regimes. These results underscore the effectiveness of training near the edge of chaos for enhancing model generalization capability in SDG tasks.

Few-shot object detection (FSOD) aims to detect objects of interest with limited annotated data, which is particularly challenging in remote sensing images (RSIs) due to significant multi-scale variations and diverse object orientations. We propose a novel FSOD framework designed explicitly for RSIs to address these challenges. Our approach incorporates a Multi-scale Enhancement Fusion (MEF) module and a Direction-aware Attention Region Proposal Network (DA-RPN). The MEF module improves multi-scale feature extraction, relieving the issue of semantic sparsity caused by direct channel compression in feature pyramid networks. Meanwhile, the DA-RPN introduces direction-aware spatial attention to better capture objects with diverse orientations, improving the model’s adaptability to varying object directions. We evaluate our model on the public RSI datasets NWPU VHR-10.v2 and DIOR. Experimental results demonstrate that our method exhibits superior detection accuracy and robustness across varying scales and orientations, showing competitive performance compared to existing FSOD approaches.

ABSTRACT Traditional text classification models face challenges in handling long texts and understanding topic transitions in dialogue scenarios, leading to suboptimal performance in automatic speech recognition (ASR)‐based multi‐round dialogue intent classification. In this article, we propose a few‐shot contrastive learning‐based multi‐round dialogue intent classification method. First, the ASR texts are partitioned, and role‐based features are extracted using a Transformer encoder. Second, refined sample pairs are forward‐propagated, adversarial samples are generated by perturbing word embedding matrices and contrastive loss is applied to positive sample pairs. Then, positive sample pairs are input into a multi‐round reasoning module to learn semantic clues from the entire scenario through multiple dialogues, obtain reasoning features, input them into a classifier to obtain classification results, and calculate multi‐task loss. Finally, a prototype update module (PUM) is introduced to rectify the biased prototypes by using gated recurrent unit (GRU) to update the prototypes stored in the memory bank and few‐shot learning (FSL) task. Experimental evaluations demonstrate that the proposed method outperforms state‐of‐the‐art methods on two public datasets (DailyDialog and CM) and a private real‐world dataset.

Few-shot learning(FSL) methods imitate the procedure by which humans learn knowledge from a few samples to identify a novel category. Some FSL methods emphasize obtaining an accurate prototype to precisely measure the similarity with query samples. However, due to the scarcity of data and the noise introduced by background information and other unrelated classes, the prototypes are usually biased and decrease the model’s effectiveness. In this paper, we propose a Knowledge Distillation and GRU-based Prototype Completion Network (KDG-ProComNet) to alleviate the above problem. In KDG-ProComNet, we propose a background discrimination pre-training module to identify the foreground object of samples and use it to obtain more accurate prototypes, thereby reducing the destructive impact of noise information on prototypes. Furthermore, we construct a memory bank to store the prototypes and update them via GRU to rectify biased prototypes and enhance the representativeness of prototypes in the feature space. Extensive experiments are conducted based on miniImageNet and tieredImageNet, and the results demonstrate the outstanding performance of the proposed method.

Conventional Text-to-SQL research tackles the problem of solving user questions in natural language by generating the corresponding SQL queries. Most of the recent works are dedicated to improving model’s robustness and generalizability in cross-domain settings. However, model’s capability in solving geography-related questions remains unexploited. In this paper we propose GS-SQL, a new framework that jointly model the schema item alignment and geospatial semantics in the question. The proposed framework consists of an improved abstract syntax tree for representing spatial queries, a novel spatial entity tagging module for locating entities in the question, and a spatial semantics extraction module for determining the spatial relationship between the entities. Then we propose GeoSpatialSpider, a dataset that introduces geospatial queries, requiring model to yield spatial functions and nested SQL inside functions. Finally we evaluate the proposed method on our dataset Experimental results show the effectiveness of our abstract syntax tree and GS-SQL in parsing geospatial semantics while preserving traditional Text-to-SQL capabilities.

Abstract Accurate traffic flow prediction is valuable for satisfying citizens’ travel needs and alleviating urban traffic pressure. However, it is highly challenging due to the complexity of the urban geospatial structure and the highly nonlinear temporal and spatial dependence on human mobility. Most existing works proposed to rely on strict periods (e.g. daily and weekly) and separate the extraction of temporal and spatial features. Besides, most Recurrent Neural Network (RNN)-based models either fail to capture variations of spatial–temporal features in adjacent timestamps or ignore details of closeness. In this paper, we propose a Multi-attention based Hybrid-convolution Spatial-temporal Recurrent Network (MHSRN) for region-based traffic flow prediction. In MHSRN, we leverage a hybrid-convolution module to capture both shifting features and rich information at the nearest timestamps, and we apply the downsampling procedure to reduce the computation of RNN-based model. Furthermore, we propose to adopt a space-aware multi-attention module to re-perceive global and local spatial–temporal features. We conduct extensive experiments based on three real-world datasets. The results show that the MHSRN outperforms other challenging baselines by approximately 0.2–8.1% in mean absolute error on all datasets. On datasets other than TaxiBJ, the MHSRN reduces the root mean square error by at least 2.8% compared with the RNN-based model.

Learning the next Point-of-Interest (POI) is a highly context-dependent human movement behavior prediction task, which has gained increasing attention with the consideration of massive spatial-temporal trajectories data or check-in data. The spatial dependency, temporal dependency, sequential dependency and social network dependency are widely considered pivotal to predict the users’ next location in the near future. However, most existing models fail to consider the influence of other users’ movement patterns and the correlation with the POIs the user has visited. Therefore, we propose a Global User Graph Enhanced Network (GUGEN) for the next POI recommendation from a global and a user perspectives. First, a trajectory learning network is designed to model the users’ short-term preference. Second, a geographical learning module is designed to model the global and user context information. From the global perspective, two graphs are designed to represent the global POI features and the geographical relationships of all POIs. From the user perspective, a user graph is constructed to describe each users’ historical POI information. We evaluated the proposed model on three real-world datasets. The experimental evaluations demonstrate that the proposed GUGEN method outperforms the state-of-the-art approaches for the next POI recommendation.

Crowd flow forecasting is of great significance for urban traffic management and personal travel planning. Due to the complexity of the urban geographic structure and the highly nonlinear temporal and spatial dependence on crowd flow, accurate forecasting becomes very challenging. Recent research works usually divided cities into regions of the same size and coded as heat-maps, for cities with complex terrain, heat-maps contain many invalid data, which have a negative effect on the acquisition of spatial dependence. In order to decrease the effect, we encode the crowd flow into graphs and propose a multi-graph neural network based model to solve the crowd flow forecasting problem. We first construct two K-NN graphs by the Euclidean distance and the Pearson correlation coefficient respectively and the spatial dependence is captured through the spatial block composed of Graph Attention Networks(GAT) and ChebNet, then another ChebNet is deployed to fuse the spatial dependency of the two graphs. Afterward, we adapted a LSTM to capture the temporal dependence of all regions separately and use self-attention mechanism and fully connected layer to (a) the receptive filed of CNN's filget prediction results. Extensive experiment results based on two real-world datasets demonstrate that our model achieves an important performance on other baselines.

Recently, Domain Adaptation has received great attention in addressing domain shift problem, where source domain and target domain follow different distribution. Existing methods often seek to minimize the marginal distribution and the conditional distribution across domains however treat them equally, which is not reasonable in real applications. Moreover, it is easy to misclassify on target data with large within-class scatter. In this paper, we propose a Compact Cluster-based Balanced Distribution Adaptation (CC-BDA) method for cross-domain classification. CC-BDA exploits the balanced domain adaptation by increasing the balance factor dynamically in each iteration until convergence. In addition, we construct compact clusters to reduce within-class scatter alongside the domain adaptation, which endows samples with the same label much closer from their means. Finally, we present analysis and performance evaluation over seven state-of-the-art baseline methods. The results show that CC-BDA significantly outperforms other adaptation methods in accuracy and execution time cost.