Multi-agent path finding (MAPF) is a safety-critical scenario where the goal is to secure collision-free trajectories from initial to desired locations. However, due to system complexity and uncertainty, integrating learning-based controllers with MAPF is challenging and cannot theoretically guarantee the safety of the learned controllers. In response, our study proposes a verified safe multi-agent neural control (VSMANC) approach for MAPF, focusing on the unified training of Decentralized Control Barrier Functions (DCBF) and controllers to enhence safety. VSMANC enables all agents to concurrently learn controllers and DCBFs using a unified loss function designed to maximize safety, adhere to standard control policies, and incorporate path-finding-related heuristics. We also propose a formal verification-guided retraining process to both verify the properties of the learned DCBFs and generate counterexamples for retraining, thereby providing a verified safety guarantee. We validate our approach through shape formation experiments and UAV simulations, demonstrating significant improvements in safety and effectiveness in complex multi-agent environments.

Dynamic networks are data structures that represent the interactions among various entities in real-world systems, with their topology and node properties evolving over time. However, prevailing approaches typically derive node embeddings through aggregating temporal neighbor nodes of adjacent several hops, thus failing to capture the long temporal dependencies in dynamic networks. Furthermore, existing research on dynamic networks focuses on node- and edge-level tasks, lacking the support of graph-level tasks. To address the limitations of current approaches, this paper proposes TP-GNN, a novel continuous dynamic graph neural network model intended for graph classification in dynamic networks, which offers two primary advantages: (1) TP-GNN captures the long temporal dependencies via a novel message-passing method based on the information flow among the nodes, and (2) it learns the network evolution process from edge order for accurate dynamic network analytics. We evaluate the performance of TP-GNN in five datasets, including a new dataset we created from a Java software project. The results show that our method outperforms state-of-the-art approaches in graph classification with an average improvement of 4.91% in terms of$F_{1}$Score11Codes and dataset are available at https://github.com/Jie-0828/TP-GNN..

System logs, recording critical information about system operations, serve as indispensable tools for system anomaly detection. Graph-based methods have demonstrated superior performance compared to other methods in capturing the interdependencies of log events. However, existing methods often neglect the complex substructure patterns of nodes within log graphs, making it challenging to capture the subtle alteration in event type, structure, and the location of exceptions that indicate node anomalies. To address this limitation, this paper proposes a novel framework called Substructure-aware Log Anomaly Detection at Code File Level (SLAD). It first introduces a Monte Carlo Tree Search strategy tailored specifically for log anomaly detection to discover representative substructures. Then, SLAD incorporates a substructure distillation way to enhance the efficiency of anomaly inference based on the representative substructures. After that, we introduce a soft pruning to obtain key substructure for nodes. Experimental results show SLAD outperforms all baselines. Particularly, SLAD demonstrates at least 15 times faster than substructure-based graph learning methods in anomaly inference.

In recent years, various distillation methods for semantic segmentation have been proposed. However, these methods typically train the student model to imitate the intermediate features or logits of the teacher model directly, thereby overlooking the high-discrepancy regions learned by both models, particularly the differences in instance edges. In this paper, we introduce a novel approach, called Difference-aware Distillation, to address this limitation. Our proposed method detects the discrepancies among the teacher model and the student model in the logit space through two masking mechanisms (i.e., masking by logit differences with respect to the ground truth labels and masking by differences in the predictive class probabilities), and guides the student model to restore the teacher's features with the focus on these highly-discrepant regions, resulting in improved segmentation performance. With the features jointly masked by these two mechanisms, the student model learns to preserve the teacher's features via a feature generation module, thus achieving better representation. Our experimental evaluation on three datasets, Cityscapes, Pascal2012, and ADE20 K, demonstrates our proposed approach outperforms several baselines considered. Further visualization analysis confirms that our method effectively directs the student model's attention to the discrepancies, such as the edges of small objects and the interiors of large objects.

Robust principal component analysis (RPCA) has gained popularity for handling high-dimensional data. The nuclear norm minimization (NNM) in RPCA is a classical method and has been widely investigated, which can recover low-rank and sparse matrices with high probability under certain conditions. However, NNM shrinks all singular values by the same threshold and over-penalizes larger singular values, resulting in this model being biased. Therefore, we propose a new method based on the truncated$l_{1-2}$norm to solve this problem in this paper, which is unbiased and flexible to capture the low-rank structure of the data matrix more accurately while separating the sparse noise. We also develop a robust and efficient algorithm to solve the proposed nonconvex optimization model, with the computational complexity and convergence discussed. Then the proposed scheme is applied to synthetic data as well as real-world data, including video background subtraction, facial shadow removal, and anomaly detection, for testing. These experimental results demonstrate that our proposed method is effect and superior in accuracy and robustness compared to other state-of-the-art methods.

Recent years have seen a growing interest in deep learning-based approaches to localize faults in software. However, existing methods have not reached a satisfying level of accuracy. The main reason is that the feature extraction of faulty code elements is insufficient. Namely, these deep learning-based methods will learn some features that are not relevant to fault localization, and thus ignore the features related to fault localization. We propose SupConFL, a new framework for statement-level fault localization. Our framework combines the statement-level abstract syntax tree with the statement sequence, and adopt controllable attention-based LSTM to locate the faulty elements. The training is done through contrastive learning between the faulty code and its fixed version. By comparing the faulty code with the fixed code, the model can learn richer features of the faulty code elements. Our experiments on Defects4j-1.2.0 dataset show that our method outperforms the current state-of-the-art. Specifically, SupConFL improves Top-1 score by 7.96% in comparison with the current state-of-the-art. In addition, our method has also achieved good results in cross-project experiments.

Anomaly detection is an important task that improves the maturity and stability of a software during its development. System logs record rich information about the running states of the software and reveal key insights of anomalous behaviors. This paper addresses anomaly detection using system log data and aims to resolve two challenges: First, different from most existing supervised learning-based anomaly detection methods that rely heavily on expensive, manually-curated labels, we aim to design an algorithm to make the most of scarce label information. Second, as a typical software system would contain very few anomalies, we aim to address the data imbalance issue which is often overlooked by existing studies. To address the challenges above, we propose LGLog, a semi-supervised anomaly detection framework that is based on system logs. First, LGLog transforms log sequences into graphs and employs an unsupervised graph learning model for pre-training. Then, LGLog mitigates the data imbalance issue by learning significant latent space representation of log events via reconstruction loss and node invariance loss, and further applies a weight balance method. Experiments indicate that LGLog outperforms compared approaches, and demonstrates the effectiveness of LGLog in the presence of scarce labels and imbalanced log data.

Hyperspectral image (HSI) denoising is a challenging task, not only because it is unavoidably contaminated by severe mixed noises, but also for its hard-to-recover spatial-spectral structure. Since it has been found that HSI has low-rank property, low-rank models have received extensive attention in dealing with the HSI denoising task. However, these models either use nuclear norm, which can only obtain sub-optimal solutions, or require some predefined information that is difficult to determine. To address these issues, in this paper we propose a new HSI denoising model based on non-convex fraction function, which has excellent performance in removing mixed noises. Specifically, the proposed model can capture the rank information of HSI automatically, which allows it to separate clean HSI from noises more accurately. Then, an iterative optimization algorithm is developed by exploiting the framework of the augmented Lagrange multiplier (ALM). Meanwhile, the subproblems at each iteration can be solved by the proximal operator with a closed-form solution. Besides, the convergence of the proposed algorithm is also provided theoretically. Extensive experiments implemented with simulated and real datasets demonstrate that our proposed model performs better than state-of-the-art models in HSI denoising. MATLAB code is available at https://github.com/wangzhi-swu/HSI-Denosing.

We parameterize schemata in Boyce-Codd Normal Form (BCNF) by the number n of minimal keys they exhibit. We show that n quantifies a trade-off between access variety and update complexity. Indeed, access variety refers to the number of different ways by which every entity over the schema is represented uniquely, while update complexity refers to the number of attribute sets for which uniqueness needs to be preserved during updates. As normalization aims at minimizing the level of effort required to preserve data consistency during updates, we establish an algorithm that returns a lossless, dependency-preserving 3NF decomposition where the subset of output schemata not in BCNF is minimized and redundant BCNF schemata are eliminated from the highest to the lowest n exhibited. In particular, if a lossless, dependency-preserving BCNF decomposition exists, our algorithm returns one where the maximum n across all output schemata is minimized. Experiments with synthetic and real-world data quantify the impact of n on the update and query performance over schemata in BCNF with n minimal keys, and show insight into the efficacy of our algorithm suite.

Low-rank minimization aims to recover a matrix of minimum rank subject to linear system constraint. It can be found in various data analysis and machine learning areas, such as recommender systems, video denoising, and signal processing. Nuclear norm minimization is a dominating approach to handle it. However, such a method ignores the difference among singular values of target matrix. To address this issue, nonconvex low-rank regularizers have been widely used. Unfortunately, existing methods suffer from different drawbacks, such as inefficiency and inaccuracy. To alleviate such problems, this article proposes a flexible model with a novel nonconvex regularizer. Such a model not only promotes low rankness but also can be solved much faster and more accurate. With it, the original low-rank problem can be equivalently transformed into the resulting optimization problem under the rank restricted isometry property (rank-RIP) condition. Subsequently, Nesterov's rule and inexact proximal strategies are adopted to achieve a novel algorithm highly efficient in solving this problem at a convergence rate of O(1/K) , with K being the iterate count. Besides, the asymptotic convergence rate is also analyzed rigorously by adopting the Kurdyka- ojasiewicz (KL) inequality. Furthermore, we apply the proposed optimization model to typical low-rank problems, including matrix completion, robust principal component analysis (RPCA), and tensor completion. Exhaustively empirical studies regarding data analysis tasks, i.e., synthetic data analysis, image recovery, personalized recommendation, and background subtraction, indicate that the proposed model outperforms state-of-the-art models in both accuracy and efficiency.

In existing ensemble learning algorithms (e.g., random forest), each base learner’s model needs the entire dataset for sampling and training. However, this may not be practical in many real-world applications, and it incurs additional computational costs. To achieve better efficiency, we propose a decentralized framework:Multi-Agent Ensemble.The framework leverages edge computing to facilitate ensemble learning techniques by focusing on the balancing of access restrictions (small sub-dataset) and accuracy enhancement. Specifically, network edge nodes (learners) are utilized to model classifications and predictions in our framework. Data is then distributed to multiple base learners who exchange data via an interaction mechanism to achieve improved prediction. The proposed approach relies on a training model rather than conventional centralized learning. Findings from the experimental evaluations using 20 real-world datasets suggest that Multi-Agent Ensemble outperforms other ensemble approaches in terms of accuracy even though the base learners require fewer samples (i.e., significant reduction in computation costs).