Information Processing & Management最新文献

Zero-shot learning (ZSL) aims to identify new classes by transferring semantic knowledge from seen classes to unseen classes. However, existing models lack a differentiated understanding of different attributes and ignore the impact of global context information. Therefore, we propose a multi-granularity contrastive zero-shot learning model based on attribute decomposition. Specifically, as attributes are the carriers of semantic knowledge, we first classify attributes into key attributes and common attributes, i.e., attribute decomposition, and the importance of common attributes is increased by key attribute mask prediction. Then, inspired by Navon’s global–local paradigm, we work out the multi-granularity contrastive learning model, which is composed of the global learning module and the local one, to further enhance the interaction between the global and local information. Finally, zero-shot image classification is achieved by training a multi-granularity contrastive learning model. The method is experimented on three public ZSL benchmark datasets (i.e., AWA2, CUB, and SUN). Compared with the existing model, this model improves the accuracy by 2.2%/5.4% (AWA2/SUN) on conventional ZSL, 2.5%/1.6%/6.3% (AWA2/CUB/SUN) on generalized ZSL, further verifying the effectiveness of this model.

In most scenarios of graph-based tasks, graph neural networks (GNNs) are trained end-to-end with labeled samples. Labeling graph samples, a time-consuming and expert-dependent process, leads to huge costs. Graph data augmentations can provide a promising method to expand labeled samples cheaply. However, graph data augmentations will damage the capacity of GNNs to distinguish non-isomorphic graphs during the supervised graph representation learning process. How to utilize graph data augmentations to expand labeled samples while preserving the capacity of GNNs to distinguish non-isomorphic graphs is a challenging research problem. To address the above problem, we abstract a novel asymmetric augmented paradigm in this paper and theoretically prove that it offers a principled approach. The asymmetric augmented paradigm can preserve the capacity of GNNs to distinguish non-isomorphic graphs while utilizing augmented labeled samples to improve the generalization capacity of GNNs. To be specific, the asymmetric augmented paradigm will utilize similar yet distinct asymmetric weights to classify the real sample and augmented sample, respectively. To systemically explore the benefits of asymmetric augmented paradigm under different GNN architectures, rather than studying individual asymmetric augmented GNN (A²GNN) instance, we then develop an auto-search engine called Asymmetric Augmented Graph Neural Architecture Search (A²GNAS) to save human efforts. We empirically validate our asymmetric augmented paradigm on multiple graph classification benchmarks, and demonstrate that representative A²GNN instances automatically discovered by our A²GNAS method achieve state-of-the-art performance compared with competitive baselines. Our codes are available at: https://github.com/csubigdata-Organization/A2GNAS.

An emerging service model in online health communities (OHCs) is that of medical teams comprising multiple physicians who collaborate to offer diagnoses and recommendations to patients. Given its multiple information sources, this model has the potential to deliver high-quality services and enhance patient satisfaction. However, the effect of a wider range of information on patient satisfaction has yet to be empirically examined. Therefore, the current research aims to examine the effect of multiple sources of health-related information on the satisfaction of patients in OHCs. We construct a sample model and empirically test it using a dataset comprising 115,367 consultation records sourced from WeDoctor. The results show that responses from multiple physicians in OHC medical teams increase patient satisfaction. In addition, we explore the moderating effects of team composition and team replies. The results show that physicians with higher titles and affiliations with the same department and the same question's replies from multiple physicians all play a positive moderating role, while reply time plays a negative moderating role. This research enriches the existing literature by focusing on patient satisfaction in the context of OHCs and offers recommendations for research and practice.

Knowledge graph completion aims to infer the missing links of new elements, however, the missing links often lie in sparse regions of the graph. Primary subgraph-based methods rely heavily on structural information, which makes it difficult for them to play an essential role in sparse graph completion. To address this challenge, we propose a learning framework for feature-enhanced hierarchical reconstruction (FEHR). The proposed FEHR explores relational semantics at the global and local levels, minimizing the limitations of sparse structures. First, entity graphs are converted into relation graphs, and overreliance on the entity structure is reduced by obtaining prior knowledge on similar global graphs. Second, the relational features are further refined at the local level. Finally, an improved performer model expresses the degree of preference between the predicted behaviors and relations. Extensive inductive experiments showed that FEHR performs better than state-of-the-art baselines, achieving improvements in area under the prediction–recall curve (AUC-PR) and Hits@n metrics, ranging from 0.32% to 11.73%.

This study introduces the Abnormality Converging Scene Analysis Method (ACSAM) to detect abnormal group behavior using monitored videos or CCTV images in crowded scenarios. Abnormal behavior recognition involves classifying activities and gestures in continuous scenes, which traditionally presents significant computational challenges, particularly in complex crowd scenes, leading to reduced recognition accuracy. To address these issues, ACSAM employs a convolutional neural network (CNN) enhanced with Abnormality and Crowd Behavior Training layers to accurately detect and classify abnormal activities, regardless of crowd density. The method involves extracting frames from the input scene and using CNN to perform conditional validation of abnormality factors, comparing current values with previous high values to maximize detection accuracy. As the abnormality factor increases, the identification rate improves with higher training iterations. The system was tested on 26 video samples and trained on 34 samples, demonstrating superior performance to other approaches like DeepROD, MSI-CNN, and PT-2DCNN. Specifically, ACSAM achieved a 12.55% improvement in accuracy, a 12.97% increase in recall, and a 10.23% enhancement in convergence rate. These results suggest that ACSAM effectively overcomes the computational challenges inherent in crowd scene detection, offering a robust solution for real-time abnormal behavior recognition in crowded environments.

Forecasting time-to-risk poses a challenge in the information processing and risk management within financial markets. While previous studies have focused on centralized survival analysis, how to forecast the time to risk using multi-party data in a decentralized and privacy-preserving setting with the requirements of explainability and mitigating information redundancy is still challenging. To this end, we propose an explainable, privacy-preserving, decentralized survival analysis method. Specifically, we transform time-to-risk forecasting into a multi-label learning problem by independently modeling for multiple time horizons. For each forecasting time horizon, we use Taylor expansion and homomorphic encryption to securely build a decentralized logistic regression model. Considering the information redundancy among multiple parties, we design and add decentralized regularizations to each model. We also propose a decentralized proximal gradient descent method to estimate the decentralized coefficients. Empirical evaluation shows that the proposed method yields competitive forecasting performance and explainable results as compared to benchmarked methods.

The prevalence of fake news on social media poses devastating and wide-ranging threats to political beliefs, economic activities, and public health. Due to the continuous emergence of news events on social media, the corresponding data distribution keeps changing, which places high demands on the generalizabilities of automatic detection methods. Current cross-event fake news detection methods often enhance generalization by training models on a broad range of events. However, they require storing historical training data and retraining the model from scratch when new events occur, resulting in substantial storage and computational costs. This limitation makes it challenging to meet the requirements of continual fake news detection on social media. Inspired by human abilities to consolidate learning from earlier tasks and transfer knowledge to new tasks, in this paper, we propose a fake news detection method based on parameter-level historical event knowledge transfer, namely EvolveDetector, which does not require storing historical event data to retrain the model from scratch. Specifically, we design the hard attention-based knowledge storing mechanism to efficiently store the knowledge of learned events, which mainly consists of a knowledge memory and corresponding event masks. Whenever a new event needs to be detected for fake news, EvolveDetector retrieves the neuron parameters of all similar historical events from the knowledge memory to guide the learning in the new event. Afterward, the multi-head self-attention is used to integrate the feature outputs corresponding to these similar events to train a classifier for the new event. Experiments on public datasets collected from Twitter and Sina Weibo demonstrate that our EvolveDetector outperforms state-of-the-art baselines, which can be utilized for cross-event fake news detection.