Granular computing, which simulates human cognition by partitioning objects into multiple granules, serves as a valuable approach for handling data with uncertainty and has been widely applied to multilabel feature selection. However, existing methods based on granular computing typically assume equal importance across all labels, which limits their ability to capture the relative recognition of features within the label space. In this article, we address this limitation by analyzing label space dependencies through the lens of rough set theory. In addition, a novel method for multilabel feature selection is introduced, which incorporates relative uncertainty information. First, we propose an object grid-based acceleration method to speed up the computation of relationships in the process of constructing neighborhood granularity, and provide a new definition of neighborhood granularity on this basis. Second, based on the constructed neighborhood granularity, we define uncertainty measures under multilabel data and analyze their corresponding theoretical properties. Finally, by analyzing the dependencies within the label space, we derive an importance matrix for the labels and, by combining it with the defined uncertainty measure, develop a multilabel feature selection method that incorporates relative uncertainty. The experimental results validate the effectiveness of the proposed method.

Utilizing various auxiliary optimization problems (AOPs) to help the optimization for constrained multiobjective problems (CMOPs) has recently drawn substantial attention. However, two key issues remain underexplored: the design of effective AOPs and the efficient allocation of iteration resources for these AOPs. Specifically, the design of AOPs directly affects the ability to identify high-quality solutions, while an effective allocation mechanism can reduce wasted iterations on less promising AOPs. In this study, we propose a novel algorithm, DRLAOP, to tackle these challenges. DRLAOP begins by analyzing the intrinsic optimization requirements of CMOPs and designs AOPs accordingly. Then, it employs a DRL-guided iteration resource allocation (DRL-IRA) mechanism to dynamically map the optimization landscape and allocate iteration resources to the most promising AOPs. Comparative experiments are carried out on 33 benchmark CMOP instances and nine real-world applications, with 19 state-of-the-art algorithms. The results demonstrate that DRLAOP consistently outperforms or matches the performance of its peers, validating that DRLAOP not only excels in discovering optimal solutions but also ensures efficient use of iteration resources.

Mengzhu Liu, Zhengqiu Zhu, Chuan Ai, Chen Gao, Xinghong Li, Lingnan He, Kaisheng Lai, Yingfeng Chen, Xin Lu, Yong Li, Quanjun Yin. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

Trajectory data is crucial for various applications but often suffers from incompleteness due to device limitations and diverse collection scenarios. Existing imputation methods rely on sparse trajectory or travel information, such as velocity, to infer missing points. However, these approaches assume that sparse trajectories retain essential behavioral patterns, which place significant demands on data acquisition and overlook the potential of large-scale human trajectory embeddings. To address this, we propose ProDiff, a trajectory imputation framework that uses only two endpoints as minimal information. It integrates prototype learning to embed human movement patterns and a denoising diffusion probabilistic model for robust spatiotemporal reconstruction. Joint training with a tailored loss function ensures effective imputation. ProDiff outperforms state-of-the-art methods, improving accuracy by 6.28\% on FourSquare and 2.52\% on WuXi. Further analysis shows a 0.927 correlation between generated and real trajectories, demonstrating the effectiveness of our approach.

Solving the multi-UAV task allocation problem with complex constraints (MTAPCc) by means of the constrained multi-objective evolutionary algorithms (cMOEAs) is novel research in the field of Operation Research. Its advantages mainly consist of two aspects. One is that it can find feasible solutions that satisfy the constraints within an acceptable time. The other is that the obtained Pareto solution set can offer more options for decision-makers. This paper presents a dualneighborhood search based dual-population coevolutionary algorithm (ccDNCA), which can specifically solve the constrained multi-objective combinatorial optimization problems (cMCOPs) based on permutation encoding, including the MTAPCc. The dual-population coevolutionary framework and the multistrategy collaborative constraint handling method of ccDNCA can effectively improve the efficiency of constraint handling and the ability of finding better solutions. The dual-neighborhood alternating local search (DN-ALS) framework can effectively increase the proportion of feasible solutions during the evolution and enhance the quality of the final solution set. The strategy pool integrated with multiple local search strategies can push the search towards regions with better objective values and constraint values, while enhancing the generalization ability of ccDNCA. In the experimental part, by comprehensively comparing the solution results of ccDNCA with those of other advanced algorithms, it is demonstrated that ccDNCA has significant superiority when dealing with cMCOPs based on permutation encoding, such as the MTAPCc and the Vehicle Routing Problem with Time Window constraints (VRPTW).

Continual Relation Extraction (CRE) has found widespread web applications (e.g., search engines) in recent times. One significant challenge in this task is the phenomenon of catastrophic forgetting, where models tend to forget earlier information. Existing approaches in this field predominantly rely on memory-based methods to alleviate catastrophic forgetting, which overlooks the inherent challenge posed by the varying memory requirements of different relations and the need for a suitable memory refreshing strategy. Drawing inspiration from the mechanisms of Dynamic Random Access Memory (DRAM), our study introduces a novel CRE architecture with an asynchronous refreshing strategy to tackle these challenges. We first design a DRAM-like architecture, comprising three key modules: perceptron, controller, and refresher. This architecture dynamically allocates memory, enabling the consolidation of well-remembered relations while allocating additional memory for revisiting poorly learned relations. Furthermore, we propose a compromising asynchronous refreshing strategy to find the pivot between over-memorization and overfitting, which focuses on the current learning task and mixed-memory data asynchronously. Additionally, we explain the existing refreshing strategies in CRE from the DRAM perspective. Our proposed method has experimented on two benchmarks and overall outperforms ConPL (the SOTA method) by an average of 1.50% on accuracy, which demonstrates the efficiency of the proposed architecture and refreshing strategy.

The growing usage of online crowdfunding platforms has fundamentally changed the traditional modes of fundraising and donation. Previous studies have mainly focused on the performance and ethical issues of online crowdfunding. In contrast, there is a dearth of information about the complexity of online donation behaviors. To explore the characteristics of fundraising and donation in online crowdfunding campaigns, we conduct a comprehensive analysis of fundraising and donation behaviors based on 151163 campaigns, with 188955849 donations created from 2016 to 2020 in one of the most popular medical crowdfunding (MCF) platforms called Easy Fundraising in China. We propose four indicators, namely, diversity, uncertainty, concentration, and consistency, to characterize the preferences of individual donors in choosing the donation amounts. Furthermore, we investigate the fundraising temporal dynamics and collective donation characteristics of crowdfunding campaigns using statistical methods. Results show that the first three days after the creation of a crowdfunding campaign is the most efficient fundraising period that largely determines the completion of the campaign. Donors who donate early are more generous than those who donate later. Individual donors prefer donation amounts in multiples of five, such as 5, 10, 20, and 50, and rarely change their donation amounts, which is irrelevant to the patients’ locations. The empirical results obtained in this study provide valuable insights to improve crowdfunding management, public welfare systems’ construction, and human donation behaviors’ understanding.

Since the outbreak of the coronavirus disease 2019 (COVID-19), the issue of how to maintain economic development while containing the epidemic has become a significant concern for decision-makers. Though lockdown measures are verified to be very effective in containing the epidemic, its economic costs and other influences have not been fully explored. As a result, decision-makers in many countries are still hesitant to include the lockdown measure in an intervention strategy in response to COVID-19. To address this issue, we propose a universal computational experiment approach for policy evaluation and adjustment based on the Artificial societies, Computational experiments, Parallel execution (ACP) concept. First, we innovatively construct a model via observable CO2 emissions, which is able to estimate the economic costs affected by nonpharmaceutical interventions. Furthermore, based on the population movement data, a risk source model is proposed to estimate the local transmission risk for any prefectures outside the epicenter. Finally, we integrate the data models in a high-resolution agent-based artificial society and carry out large-scale computational experiments supported by the Tianhe supercomputer. Policy adjustments and evaluations are carried out in four cities: Wenzhou, Guangzhou, Beijing, and Wuhan. Our research findings show important implications for policy-making: 1) the local transmission of a city can be almost contained if lockdowns are adopted immediately when the risk index is larger than 1.645, 1.960, or 2.576 at the 90%, 95%, or 99% confidence interval, respectively; 2) if lockdowns are required, in-advance lockdown measures facilitate mitigation efficacy and reduce economic loss; and 3) lockdowns lasting for 7–14 days in a prefecture would be effective in controlling the spread of the epidemic. The duration of the measure should be prolonged with the increment of the initial transmission risk.

Binary neural networks (BNNs) constrain weights and activations to +1 or -1 with limited storage and computational cost, which is hardware-friendly for portable devices. Recently, BNNs have achieved remarkable progress and been adopted into various fields. However, the performance of BNNs is sensitive to activation distribution. The existing BNNs utilized the Sign function with predefined or learned static thresholds to binarize activations. This process limits representation capacity of BNNs since different samples may adapt to unequal thresholds. To address this problem, we propose a dynamic BNN (DyBNN) incorporating dynamic learnable channel-wise thresholds of Sign function and shift parameters of PReLU. The method aggregates the global information into the hyper function and effectively increases the feature expression ability. The experimental results prove that our method is an effective and straightforward way to reduce information loss and enhance performance of BNNs. The DyBNN based on two backbones of ReActNet (MobileNetV1 and ResNet18) achieve 71.2% and 67.4% top1-accuracy on ImageNet dataset, outperforming baselines by a large margin (i.e., 1.8% and 1.5% respectively).

Recently proposed decoupled training methods emerge as a dominant paradigm for long-tailed object detection. But they require an extra fine-tuning stage, and the dis-jointed optimization of representation and classifier might lead to suboptimal results. However, end-to-end training methods, like equalization loss (EQL), still perform worse than decoupled training methods. In this paper, we re-veal the main issue in long-tailed object detection is the imbalanced gradients between positives and negatives, and find that EQL does not solve it well. To address the problem of imbalanced gradients, we introduce a new version of equalization loss, called equalization loss v2 (EQL v2), a novel gradient guided reweighing mechanism that re-balances the training process for each category independently and equally. Extensive experiments are performed on the challenging LVIS benchmark. EQL v2 outperforms origin EQL by about 4 points overall AP with 14 ∼ 18 points improvements on the rare categories. More importantly, it also surpasses decoupled training methods. With-out further tuning for the Open Images dataset, EQL v2 improves EQL by 7.3 points AP, showing strong generalization ability. Codes have been released at https://github.com/tztztztztz/eqlv2

The two-stage methods for instance segmentation, e.g. Mask R-CNN, have achieved excellent performance recently. However, the segmented masks are still very coarse due to the downsampling operations in both the feature pyramid and the instance-wise pooling process, especially for large objects. In this work, we propose a new method called RefineMask for high-quality instance segmentation of objects and scenes, which incorporates fine-grained features during the instance-wise segmenting process in a multi-stage manner. Through fusing more detailed information stage by stage, RefineMask is able to refine high-quality masks consistently. RefineMask succeeds in segmenting hard cases such as bent parts of objects that are oversmoothed by most previous methods and outputs accurate boundaries. Without bells and whistles, RefineMask yields significant gains of 2.6, 3.4, 3.8 AP over Mask R-CNN on COCO, LVIS, and Cityscapes benchmarks respectively at a small amount of additional computational cost. Furthermore, our single-model result outperforms the winner of the LVIS Challenge 2020 by 1.3 points on the LVIS test-dev set and establishes a new state-of-the-art. Code will be available at https://github.com/zhanggang001/RefineMask.

Changes in network state are a common source of instability in networks. An update event typically involves multiple flows that compete for network resources at the cost of rescheduling and migrating some existing flows. Previous network updating schemes tackle such flows independently, rather than as the entity of an update event. They only optimize the flow-level metrics for the flows involved in an update event. In this paper, we present an event-level abstraction of network update that groups flows of an update event and schedules them together to minimize the event completion time (ECT). We then study the scheduling problem of multiple update events for achieving high scheduling efficiency and preserving fairness. The designed least migration traffic first (LMTF) method schedules all update events in the FIFO order, but it avoids head-of-line blocking by randomly fine-tuning the queue order of some events. It can considerably reduce the update cost, the average, and tail ECTs of update events. In addition, we design a general parallel-LMTF (P-LMTF) method to guarantee fairness and further improve scheduling efficiency among update events. This improves the LMTF method by opportunistically updating multiple events simultaneously. The comprehensive evaluation results indicate that the average ECT of our approach is up to 10× faster than the flow-level scheduling method for network update events, and its tail ECT is up to 6× faster. Our P-LMTF method incurs a 75% reduction in the average ECT compared with FIFO when the network utilization exceeds 70%, and it achieves a 42% reduction in tail ECT.

This paper proposes a novel object detection framework named Grid R-CNN, which adopts a grid guided localization mechanism for accurate object detection. Different from the traditional regression based methods, the Grid R-CNN captures the spatial information explicitly and enjoys the position sensitive property of fully convolutional architecture. Instead of using only two independent points, we design a multi-point supervision formulation to encode more clues in order to reduce the impact of inaccurate prediction of specific points. To take the full advantage of the correlation of points in a grid, we propose a two-stage information fusion strategy to fuse feature maps of neighbor grid points. The grid guided localization approach is easy to be extended to different state-of-the-art detection frameworks. Grid R-CNN leads to high quality object localization, and experiments demonstrate that it achieves a 4.1% AP gain at IoU=0.8 and a 10.0% AP gain at IoU=0.9 on COCO benchmark compared to Faster R-CNN with Res50 backbone and FPN architecture.

The prediction of short-term volatile traffic becomes increasingly critical for efficient traffic engineering in intelligent transportation systems. Accurate forecast results can assist in traffic management and pedestrian route selection, which will help alleviate the huge congestion problem in the system. This paper presents a novel hybrid DTMGP model to accurately forecast the volume of passenger flows multi-step ahead with the comprehensive consideration of factors from temporal, origin-destination spatial, and frequency and self-similarity perspectives. We first apply discrete wavelet transform to decompose the traffic volume series into an appropriation component and several detailed components. Then we propose a more efficient tracking model to forecast the appropriation component and a novel Gaussian process model to forecast the detailed components. The forecasting performance is evaluated with real-time passenger flow data in Chongqing, China. Simulation results demonstrate that our hybrid model can achieve on average 20%-50% accuracy improvement, especially during rush hours.