Knowledge-Based Systems最新文献

The N-body problem is a remarkable research topic in physics. We propose a new feature extraction model inspired by the N-body trajectory and test its feature extraction capability. In the first part of the research, an open-access electroencephalogram (EEG) dataset is used to test the proposed method. This dataset has three classes, namely (i) Alzheimer's Disorder (AD), (ii) frontal dementia (FD), and (iii) control groups. In the second step of the study, the EEG signals were divided into segments of 15 s in length, which resulted in 4,661 EEG signals. In the third part of the study, the proposed new self-organized feature engineering (SOFE) model is used to classify the EEG signals automatically. For this SOFE, two novel methods were presented: (i) a dynamic feature extraction function using a graph of the N-Body orbital, termed N-BodyPat, and (ii) an attention pooling function. A multileveled and combinational feature extraction method was proposed by deploying both methods. A feature selection function using ReliefF and Neighborhood Component Analysis (RFNCA) was used to choose the most informative features. An ensemble k-nearest neighbors (EkNN) classifier was employed in the classification phase. Our proposed N-BodyPat generates seven feature vectors for each channel, and the utilized EEG signal dataset contains 19 channels. In this aspect,133 (=19 × 7) EkNN-based outcomes were created. To attain higher classification performance by employing these 133 EkNN-based outcomes, an iterative majority voting (IMV)-based information fusion method was applied, and the most accurate outcomes were selected automatically. The recommended N-BodyPat-based SOFE achieved a classification accuracy of 99.64 %.

Despite the remarkable progress made in visual reinforcement learning (RL) in recent years, sample inefficiency remains a major challenge. Many existing approaches attempt to address this by extracting better representations from raw images using techniques like data augmentation or introducing some auxiliary tasks. However, these methods overlook the environmental dynamic information embedded in the collected transitions, which can be crucial for efficient control. In this paper, we present STAR: State-Action Representation Learning, a simple yet effective approach for visual continuous control. STAR learns a joint state–action representation by modeling the dynamics of the environment in the latent space. By incorporating the learned joint state–action representation into the critic, STAR enhances the value estimation with latent dynamics information. We theoretically show that the value function can still converge to the optima when involving additional representation inputs. On various challenging visual continuous control tasks from DeepMind Control Suite, STAR achieves significant improvements in sample efficiency compared to strong baseline algorithms.

Graph Contrastive Learning (GCL) effectively mitigates label dependency, defining positive and negative pairs for node embeddings. Nevertheless, most GCL methods, including those considering communities, overlooking the simultaneous influence of community and node—a crucial factor for accurate embeddings. In this paper, we propose Dual influenced Community Strength-boosted Multi-Scale Graph Contrastive Learning (DCMSL), concurrently considering community and node influence for comprehensive contrastive learning. Firstly, we define dual influenced community strength which can be adaptable to diverse datasets. Based on it, we define node cruciality to differentiate node importance. Secondly, two graph data augmentation methods, NCNAM and NCED, respectively, are put forward based on node cruciality, guiding graph augmentation to preserve more influential semantic information. Thirdly, a joint multi-scale graph contrastive scheme is raised to guide the graph encoder to learn data semantic information at two scales: (1) Propulsive force node-level graph contrastive learning—a node-level graph contrastive loss defining the force to push negative pairs in GCL farther away. (2) Community-level graph contrastive learning—enabling the graph encoder to learn from data on the community level, improving model performance. DCMSL achieves state-of-the-art results, demonstrating its effectiveness and versatility in two node-level tasks: node classification and node clustering and one edge-level task: link prediction. Our code is available at: https://github.com/HanChen-HUST/DCMSL.

Keyphrase Extraction (KE) aims to identify a concise set of words or phrases that effectively summarizes the core ideas of a document. Recent embedding-based models have achieved state-of-the-art performance by jointly modeling local and global contexts in Unsupervised Keyphrase Extraction (UKE). However, these models often ignore either sentence- or document-level contexts, leading directly to weak or incorrect global significance. Furthermore, they rely heavily on local significance, making them vulnerable to noisy data, particularly in long documents, resulting in unstable and suboptimal performance. Intuitively, hierarchical contexts enable a more accurate understanding of the candidates, thereby enhancing their global relevance. Inspired by this, we propose a novel Hierarchical Context-aware Unsupervised Keyphrase Extraction method called HCUKE. Specifically, HCUKE comprises three core modules: (i) a hierarchical context-based global significance measure module that incrementally learns global semantic information from a three-level hierarchical structure; (ii) a phrase-level local significance measure module that captures local semantic information by modeling the context interaction among candidates; and (iii) a candidate ranking module that integrates the measure scores with positional weights to compute a final ranking score. Extensive experiments on three benchmark datasets demonstrate that the proposed method significantly outperforms state-of-the-art baselines.

In recent years, decentralized social networks have gained increasing attention, where each client maintains a local view of a social graph. To provide services based on graph learning in such networks, the server commonly needs to collect the local views of the graph structure, which raises privacy issues. In this paper, we focus on learning graph neural networks (GNNs) on decentralized social graphs while satisfying local differential privacy (LDP). Most existing methods collect high-dimensional local views under LDP through Randomized Response, which introduces a large amount of noise and significantly decreases the usability of the collected graph structure for training GNNs. To address this problem, we present Structure Learning-based Locally Private Graph Learning (SL-LPGL). Its main idea is to first collect low-dimensional encoded structural information called cluster degree vectors to reduce the amount of LDP noise, then learn a high-dimensional graph structure from the cluster degree vectors via graph structure learning (GSL) to train GNNs. In SL-LPGL, we propose a Homophily-aware Graph StructurE Initialization (HAGEI) method to provide a low-noise initial graph structure as learning guidance for GSL. We then introduce an Estimated Average Degree Vector Enhanced Graph Structure Learning (EADEGSL) method to further mitigate the negative impact of LDP noise in GSL. We conduct experiments on four real-world graph datasets. The experimental results demonstrate that SL-LPGL outperforms the baselines.

Mainstream person search models aim to jointly optimize person detection and re-identification (ReID) in a one-step manner. Despite notable progress, existing one-step person search models still face three major challenges in extracting discriminative features: 1) incomplete feature extraction and fusion hinder the effective utilization of multiscale information, 2) the models struggle to capture critical features in complex occlusion scenarios, and 3) the optimization objectives of person detection and ReID are in conflict in the shared feature space. To address these issues, this study proposes a novel adaptive shift and task decoupling (ASTD) method that aims to enhance the accuracy and robustness of extracting discriminative features within the region of interest. In particular, we introduce a scale-aware transformer to handle scale/pose variations and occlusions. This transformer incorporates scale-aware modulation to enhance the utilization of multiscale information and adaptive shift augmentation to learn adaptation to occlusions dynamically. In addition, we design a task decoupling mechanism to hierarchically learn independent task representations using orthogonal loss to decouple two subtasks during training. Experimental results show that ASTD achieves state-of-the-art performance on the CUHK-SYSU and PRW datasets. Our code is accessible at https://github.com/zqx951102/ASTD.

With the recent progress in graph-federated learning (GFL), it has demonstrated a promising performance in effectively addressing challenges associated with heterogeneous clients. Although the majority of advances in GFL have been focused on techniques for elucidating the intricate relationships among clients, existing GFL methods have two limitations. First, current methods comprising the use of low-dimensional graphs fail to accurately depict the associations between clients, thereby compromising the performance of GFL. Second, these methods may disclose additional information when sharing client-side hidden representations. This paper presents a structural GFL (SGFL) framework and a suite of novel optimization methods. SGFL addresses the limitations of existing GFL approaches with three original contributions. Firstly, our approach advocates the dynamic construction of federated learning (FL) graphs by leveraging the high-dimensional information inherent among clients, while enabling the discovery of hierarchical communities within clients. Secondly, we present SG-FedX, a novel federated stochastic gradient optimization algorithm that mitigates the effects of heterogeneity by intelligently using a global representation. Furthermore, SG-FedX introduces a strict sharing mechanism that protects client privacy more effectively by refraining from sharing client information beyond the model parameters. Our comparative evaluations, conducted against ten representative FL algorithms under challenging non-independently-and-identically-distributed settings, demonstrated the superior performance of SG-FedX. It was noted that, in the cross-dataset scenarios, SG-FedX outperformed the second-best baseline by 8.12% and 7.91% in personalization and generalization performance, respectively.

Multi-objective reinforcement learning (MORL) addresses the challenge of optimizing policies in environments with multiple conflicting objectives. Traditional approaches often rely on scalar utility functions, which require predefined preference weights, limiting their adaptability and efficiency. To overcome this, we propose the Dynamic Preference Inference Network (DPIN), a novel method designed to enhance sample efficiency by dynamically estimating the trajectory decision preference of the agent. DPIN leverages a neural network to predict the most favorable preference distribution for each trajectory, enabling more effective policy updates and improving overall performance in complex MORL tasks. Extensive experiments in various benchmark environments demonstrate that DPIN significantly outperforms existing state-of-the-art methods, achieving higher scalarized returns and hypervolume. Our findings highlight DPIN’s ability to adapt to varying preferences, reduce sample complexity, and provide robust solutions in multi-objective settings.

Federated learning is a popular machine learning paradigm that achieves decentralized model training on distributed devices, ensuring data decentralization, privacy protection, and enhanced overall learning effectiveness. However, the non-independence and identically distributed (i.e., non-IID) nature of medical data across different institutes has remained a significant challenge in federated learning. Current research has mainly focused on addressing label distribution skew and classification scenarios, overlooking the feature distribution skew settings and more challenging semantic segmentation scenarios. In this paper, we present communication-efficient Adversarial Contrastive Federated Learning (ACFL) for the prevalent feature distribution skew scenarios in medical semantic segmentation. The core idea of the approach is to enhance model generalization by learning each client’s domain-invariant features through adversarial training. Specifically, we introduce a global discriminator that, through contrastive learning in the server, trains to differentiate feature representations from various clients. Meanwhile, the clients learn common domain-invariant features through prototype contrastive learning and global discriminator training. Furthermore, by utilizing Gaussian mixture models for virtual feature sampling on the server, compared to transmitting raw features, the ACFL method possesses the additional advantages of efficient communication and privacy protection. Extensive experiments on two medical semantic segmentation datasets and extension on three classification datasets validated the superiority of the proposed method.