Deep graph clustering(DGC), which seeks to partition nodes into disconnected clusters through clustering algorithms, remains a challenging research task. Recent advancements have drawn significant attention to reliable clustering algorithms due to their promising performance. However, due to the limitation of clustering algorithms, existing methods for DGC cannot effectively mine hard pseudo label as high-confidence pseudo label, thereby neglecting the crucial samples. To solve the above problem, this paper proposes a Distance-guided Pseudo Label Graph Clustering Network(DPLC) by introducing a comprehensive hard node measure criterion to improve the clustering performance. Specifically, DPLC first divides the nodes into disconnected clusters based on clustering algorithm. Then, this network uses carefully selected clustering nodes as high-confidence pseudo label, aiming the network for the feature distribution alignment can be more accurate. Additionally, different from most DGCs only use contrastive learning, we optimize high similarities between disconnected nodes which may have higher similarity than the connected nodes in the graph representation, preventing irrational message-passing and then preserving structural information. Extensive experiments on five public datasets demonstrate the superiority of DPLC against competitors.
The granular computing can mimic the way humans think when dealing with large-scale complex problems and has been effectively utilized in the area of anomaly detection. However, most anomaly detection algorithms based on granular computing are incapable of managing incomplete data. Throughout this paper, we propose the fuzzy information entropy tailored for incomplete data, along with its application in anomaly detection algorithms for incomplete data. First, fuzzy set and fuzzy similarity relation with respect to incomplete data are defined. Second, the fuzzy entropy and its related metrics are proposed, which include incomplete fuzzy information entropy, incomplete fuzzy joint entropy, incomplete fuzzy conditional entropy, incomplete fuzzy mutual information, and incomplete fuzzy complement entropy. Additionally, two indicators, the incomplete entropy ratio factor and incomplete potential difference, are proposed. Subsequently, an anomaly detection model is established using the proposed theoretical approach. Initially, a relation matrix is formed using the fuzzy similarity relations of the incomplete information system. Next, the incomplete entropy ratio factor and incomplete potential difference are calculated for specific attribute sets of attributes. Then, anomaly factors for objects within specific attribute sets are formulated, and anomaly scores are calculated to evaluate the extent of anomaly in these objects. Finally, the related anomaly detection algorithm, ADIFIE (Anomaly Detection for Incomplete data based on Fuzzy Information Entropy), is proposed. Comparative experiments with mainstream anomaly detection algorithms are carried out on public UCI datasets and synthetic datasets. Experimental results demonstrate that the ADIFIE algorithm attains the best performance in the conducted experiments, excelling in three-quarters of the datasets.
We have researched generating distinctive implication expressions from table datasets, which we call rules. We have realized a rule generation system that we call the DIS-Apriori rule generator. This paper aims to utilize the implication expressions obtained as rules. In the first half, we focus on the dependency between attributes and the positive region and expand the positive region when contradictions are recognized. In the second half, we use the obtained rules to calculate the dependency between attributes based on the expanded positive region. The proposed method is expected to become a new method for dependency calculation.
In this work, we propose a method to explain the performance of a neural network-based brain of a machine learning agent trained with a Proximal Policy Optimization (PPO) algorithm. By explaining, we mean to discover probable criteria on which the PPOtrained agent system is making certain decisions. An additional complication is that the agent retrieves information about the environment using a camera. Our proposed explanation method consists of Gradient-weighted Class Activation Mapping (GradCAM) and Learning from Examples using Rough Sets (LEM2) algorithm. We also show how using a threshold on rules support and Laplace, we can reduce the number of rules and measure how reducing the number of rules affects the accuracy of approximation of the explained model. To the best of our knowledge, this is the first work proposing a PPO-generated network explanation for visual-based agents using the rough sets-based method. Our method has achieved a high efficiency of explanation of neural network performance on the test set, reaching F1 score = 0.93 and an Average per-class accuracy of 0.94. In some cases, we have also reduced the number of rules to a few or a dozen while maintaining a high level of explainability efficiency. We publish all source codes so our experiments can be reproduced.
Clustering algorithms are crucial in data mining and pattern recognition. However, most existing clustering methods operate based on a single granularity, which requires frequent distance calculations between sample points. These methods often exhibit unstable performance when handling high-dimensional data and clusters with complex shapes. To address this issue, this paper proposes a multi-objective driven granular ball clustering (MOGBC) algorithm, where granular balls represent a relatively coarse granularity with respect to the data. In the granule ball generation phase, principal component analysis (PCA) is used to determine the splitting direction of the granular ball. In the optimization phase, the multi-objective alignment loss function is minimized by balancing intra-cluster compactness (ICC) and inter-cluster separation (ICS). A neighborhood voting mechanism is also employed to determine the assignment of fuzzy boundary points. Experimental results demonstrate that MOGBC achieves strong clustering performance across 10 datasets, with significant advantages in clustering evaluation metrics such as ACC, NMI, and ARI, thereby validating the algorithm's effectiveness and robustness.
