Applied Intelligence最新文献

Recommendation systems founded on graph neural networks (GNN) have been extensively employed because of their exceptional recommendation efficiency. Nevertheless, numerous recommendation biases also crop up, We have observed that delicate details such as gender and age are frequently implicitly apprehended by recommendation systems, culminating in unfair recommendations, and the associated algorithms of GNN will magnify this bias. To tackle these difficulties, this paper puts forth a method of introducing the notion of causal fairness into the issue of fairness in GNN-based recommendation systems, to accomplish counterfactual fairness of user-sensitive information and thereby attain unbiased recommendations. Specifically, given a GNN-based recommendation system model, which is implemented in our devised fairness framework, chiefly obtaining equitable effects through two facets: (1) attaining user embedding fairness through the counterfactual fairness technique; (2) mitigating the prejudiced impact caused by the GNN algorithm using the proposed central association subgraph method. The amalgamation of these two facets ultimately delivers unbiased recommendations. The effectiveness and sophistication of our proposed method for mitigating partiality problems in GNN recommendation systems from a causal perspective (MGRC) have been proven via experiments on four real-world datasets.

The massive amount of data generated by the proliferation of Internet of Things (IoT) devices has become one of the key factors driving the advancement of artificial intelligence (AI) technology. However, the lack of storage space and limited computational power of edge devices make it difficult to directly process large data volumes or run complex machine learning algorithms on these devices. At the same time, existing Federated Learning (FL) schemes still face a number of shortcomings, including a single point of failure, vulnerability to poisoning attacks, and a lack of incentives. To address the above issues, we propose DSFL, a blockchain-based framework for fair data sharing and FL. Specifically, we combine digital envelope technology and one-way accumulator with smart contracts to design fair, secure, and trustworthy data sharing protocols that facilitate edge devices to share data proactively, realize the value of data and reduce storage pressure. In addition, we propose blockchain extension schemes suitable for coupling with FL to improve training efficiency. Importantly, the node management mechanism and incentive algorithms are designed to effectively monitor and trace the behavior of nodes, and promote the virtuous cycle of model training and the motivation of participants. Experimental results show that DSFL is able to ensure fair data sharing and efficient model training without the involvement of trusted third parties. In particular, it is able to achieve model accuracy close to that of existing popular schemes even when 40% of the nodes are lazy, providing an excellent defense against malicious nodes.

With the advent of AI-based image synthesis tools and techniques, Deepfakes have become a serious problem as they pose a massive threat to one’s information security and personal privacy. Several architectures have been proposed to achieve robust Deep Fake detection. However, these methods suffer a drastic drop in performance if the images are visually degraded or have low resolution. To resolve these two issues, a novel FreqFaceNet model has been proposed that employs two novel attentions namely, Wavelet Attention and Fourier Attention, for extracting important frequency-based features from low-resolution images. The extraction of frequency-based features ensures minimal interference of noise due to image compression or low resolution. The proposed model excels on two public benchmark datasets—the DFDC and CelebDF. On the DFDC dataset, FreqFaceNet achieves 98.041% accuracy, an AUC value of 99.748, and a Mathews Correlation Coefficient (MCC) value of 93.857, while on the CelebDF dataset, it obtains an accuracy of 98.325%, an AUC value of 99.81, and an MCC value of 92.819. Qualitative analysis of the proposed model indicates strong classification capabilities. An ablation study has also been conducted to verify the complementary contributions of both Wavelet and Fourier Attention mechanisms.

Multi-task dense scene understanding is a fundamental research area in computer vision (CV). By predicting pixels, perceiving, and reasoning about multiple related tasks, it improves both accuracy and data efficiency. However, it faces the challenge that some tasks may require more independent feature representations, and excessive sharing can lead to interference between tasks. To address this issue, we propose a novel Pixel-Attention-Guided Unet (PAG-Unet). PAG-Unet incorporates a Pixel-Attention-Guided Fusion module (PAG Fusion) and a Multi-Task Self-Attention module (MTSA) to enhance task-specific feature extraction and reduce task interference. PAG Fusion leverages the relationship between shallow and deep features by using task-specific deep features to calibrate the distribution of shared shallow features. This suppresses background noise and enhances semantic features, thereby fully extracting task-specific features for different tasks and achieving feature enhancement. MTSA considers both global and local spatial interactions for each task during task interactions, capturing task-specific information and compensating for the loss of crucial details, thus improving prediction accuracy for each task. Our method achieves superior multi-task performance on the New York University Depth v2(NYUD-v2) and PASCAL Visual Object Classes Context(PASCAL-Context) datasets, with most metrics significantly outperforming previous state-of-the-art methods. The code is available at https://github.com/UPLI-123/Pag-Unet.

Automating the generation of user interface (UI) code from design images has gained significant attention due to its potential to streamline application development. However, the effectiveness of deep learning models in this domain is often hindered by mismatches between UI images and their corresponding layout code, a common issue in image-text datasets. In this paper, we introduce a framework that locates and removes these mismatches, thereby improving the accuracy of UI code generation models. Our approach leverages a convolutional neural network to predict the alignment between UI components and layout code nodes, coupled with a tree-based heuristic algorithm to localize mismatches. Through extensive evaluation, we demonstrate that our method enhances the accuracy of UI code generation by approximately 15%, while significantly reducing the need for costly manual annotations. The proposed framework not only advances the state of automated UI code generation but also lays the foundation for creating high-quality, large-scale UI datasets, essential for future research and development in this field.

Deep matrix factorization (deep MF) is an increasingly popular unsupervised data-mining technique that operates as a deep decomposition rooted in traditional nonnegative matrix factorization (NMF). Compared with standard NMF, deep MF has shown excellent performance in the extraction of hierarchical information from complex datasets. For cases in which the data matrices corresponding to the dataset are symmetric—such as the adjacency matrix of an undirected graph in network analysis—this paper proposes a deep MF variant called deep non-smooth nonnegative symmetric matrix factorization (DNSSNMF). The aim of this work is to enhance the extraction of complex hierarchical structures in high-dimensional datasets and achieve the clustering of structures inherent in graphical representations by improving the goodness-of-fit of the factor matrix product. Accordingly, we successfully applied DNSSNMF to post-traumatic-stress-disorder (PTSD) datasets and synthetic datasets to extract several hierarchical communities. In particular, we extracted non-disjoint communities in the partial correlation network of psychiatric symptoms in PTSD, revealing correlations between different symptoms and leading to meaningful clinical interpretations. The results of our numerical experiments indicated promising applications of DNSSNMF in fields including network analysis and medicine.

Self-supervised heterogeneous graph neural networks have shown remarkable effectiveness in addressing the challenge of limited labeled data. However, current contrastive learning methods face limitations in leveraging neighborhood information for each node. Some approaches utilize the local information of the target node, ignoring useful signals from deeper neighborhoods. On the other hand, simply stacking convolutional layers to expand the neighborhood inevitably leads to over-smoothing. To address the problems, we propose HGNN-DB, a Self-supervised Heterogeneous Graph Neural Network Based on Deep and Broad Neighborhood Encoding to tackle the over-smoothing problem within heterogeneous graphs. Specifically, HGNN-DB aims to learn informative node representations by incorporating both deep and broad neighborhoods. We introduce a deep neighborhood encoder with a distance-weighted strategy to capture deep features of target nodes. Additionally, a single-layer graph convolutional network is employed for the broad neighborhood encoder to aggregate broad features of target nodes. Furthermore, we introduce a collaborative contrastive mechanism to learn the complementarity and potential invariance between the two views of neighborhood information. Experimental results on four real-world datasets and seven baselines demonstrate that our model significantly outperforms the current state-of-the-art techniques on multiple downstream tasks. The codes and datasets for this work are available at https://github.com/SSQiana/HGNN-DB.

Reinforcement learning (RL)-based multi-hop knowledge graph reasoning has achieved remarkable success in real-world applications and can effectively handle knowledge graph completion tasks. This approach involves a policy-based agent navigating the graph environment to extend reasoning paths and identify the target entity. However, most existing multi-hop reasoning models are typically constrained to stepwise inference, which inherently disrupts the global information of multi-hop paths. To overcome this limitation, we introduce discriminative features between valid and invalid paths as global information. Here, we propose a multi-hop path encoder specifically designed to extract these discriminative features. Firstly, a multi-hop path encoding module is employed to derive each path’s hidden features, using cross-attention mechanisms to strengthen the interaction between triple context and path features. Secondly, a discriminative feature extraction module is used to capture the differences between valid and invalid paths. Thirdly, an attention-enhanced gated fusion mechanism is implemented to integrate these discriminative features into the multi-hop inference decoder. We further evaluate our method on five standard datasets. Our method outperforms the baseline models, demonstrating the effectiveness of discriminative features in improving prediction performance, learning speed, and path interpretability.

Radar-based human motion recognition (HMR) technology has gained substantial importance across diverse domains such as security surveillance, post-disaster search and rescue operations, and the development of smart home environments. The intricate nature of human movements generates radar echo signals with pronounced non-stationary attributes, which encapsulate a wealth of target feature data. However, striking a balance between the precision of motion recognition and the requirement for real-time processing, especially in the context of extracting meaningful features from radar signals, remains a formidable challenge. This research paper introduces a novel approach to tackle this challenge. Firstly,we apply the multi-order fractional Fourier transform (m-FRFT) to radar echo signals, facilitating the extraction of micro-Doppler (m-D) frequency information. Secondly, we have developed an optimized feature selection model named MPG, which stands for m-D parameter screening based on genetic algorithm (GA) and adaptive weight particle swarm optimization (AWPSO). Thirdly, we apply the MPG model to the recursive feature elimination (RFE) algorithm to refine the representation of m-D frequency information, allowing for adaptive parameter adjustment and effective feature dimensionality reduction. The proposed method has been tested using human motion echo data collected from a Doppler radar prototype. The experimental outcomes demonstrate that our approach outperforms traditional feature extraction methods in terms of reducing feature dimensionality, computational efficiency, and classification accuracy.