查看更多>>摘要：Conventional domain adaptation tries to leverage knowledge obtained from the single source domain to recognize the data in the target domain, where only one modality exists in the source domain. This neglects the scenario that source domain can be acquired from multi-modal data, such as RGB data and depth data. In addition, conventional domain adaptation approaches generally assume source and target domains have the identical number of categories, which is quite restrict for real-world applications. In practice, the number of categories in the target domain is often less than that in the source domain. In this work, we focus on a more practical and challenging task that recognizes RGB data by learning from RGB-D data under an unequal label scenario, which suffers from three challenges: i) the addition of depth information, ii) the domain mismatch problem and iii) the negative transfer caused by unequal label numbers. Our main contribution is a novel method, referred to as unequal Distribution Visual-Depth Adaption (uDVDA), which takes advantage of depth data and handles domain mismatch problem under label inequality, simultaneously. Experiments show that uDVDA outperforms state-of-the-art models on different datasets, especially under unequal label scenario.(c) 2022 Elsevier Inc. All rights reserved.
查看更多>>摘要：Shoeprints are important information collected at the crime scene and are of great value for forensic analysis. Shoeprints collected in real-world scenarios are normally unclear, abrasive, and lack contextual and other kinds of missing information. In this research, we apply a novel deep learning technique called restorable inpainting to repair shoeprint contours and missing parts. Existing inpainting methods aim to fill artificially occluded areas with plausible pixels, but these methods may not restore missing information for occlusions in shoeprint images. In addition, because no ground-truth shoeprints exist for training samples, inpainting occluded regions becomes challenging. In this paper, we propose DeepShoePaint, a novel deep learning approach to perform restorable inpainting by restoring synthetic information resembling desirable shoeprint images necessary for forensics. DeepShoePaint novelly adapts a probabilistic distribution borrowed from the variational autoencoder into a U-Net-like structure forming a unified architecture trained in an unsupervised fashion to restore occluded and masked regions to produce human-verifiable shoeprints. The experimental results reveal that DeepShoePaint achieves outstanding human inspection and statistical assessment results and outperforms conventional inpainting models. We believe that this study can provide valuable insights, not limited to inpainting, into restoring desirable shoeprints to automate and facilitate the forensic investigation and examination process instead of using handcrafted methods. (c) 2022 Elsevier Inc. All rights reserved.
查看更多>>摘要：The proliferation of a digital transformation area is inspiring researchers and practitioners in finance to embrace emerging innovative fintech development (i.e., finance + technology). In this study, we propose a field-aware attentive neural factorization machine (FAFM) model for large-scale data-driven company investment valuation. The proposed FAFM model utilizes the advantage of factorization machine (FM) to efficiently capture nonlinear feature interactions in a sparse dataset. We additionally consider field heterogeneity among features with fuzzy mutual information and develop an attention neural network to learn predictive strengths of pair-wise feature interactions. FAFM contributes to the literature by overcoming the limitation of FM that ignores field heterogeneity by factorizing pair-wise feature interactions with same weight. Further more, FAFM learns the prediction strengths in a stratified manner by using the attention deep learning mechanism, which demonstrates more structured control ability and allows for more leverage in tweaking the interactions in the feature-wise level. Experiments are conducted on a unique real dataset set consisting of 3,500 listed companies in the Chinese market with features from eight fields: demographics, annual reports, stock financial disclosure, land use, intellectual property, tax, bond financing, and certification. Results showed the superiority of FAFM on prediction accuracy and model interpretability over existing baselines. Our study provides a useful tool for company investment valuation that can not only generate accurate investment valuations but also provide interpretations of both individual features and their pair wise interactions effects, thereby allowing investors better investment decisions.(c) 2022 Elsevier Inc. All rights reserved.
查看更多>>摘要：There exists a key escrow issue in ciphertext-policy attribute-based encryption (CP-ABE). The key generator center issues all users' secret keys and can decrypt each ciphertext by calculating the corresponding secret key. Besides, applying CP-ABE in data sharing environment also brings another challenging problem which is users' revocation. To resolve the above issues, we provide a key escrow-free CP-ABE scheme with the user revocation, which withstands collusion attack between malicious users and revoked users. In our scheme, a secret key is calculated utilizing a secure key issuing protocol between key authority (KA) and data user (DU). KA is unable to obtain DU's secret value and generate the complete secret key independently, which solves the key escrow issue. When a user revokes from the system, the secret keys of the unrevoked users require to be updated. We introduce a group manager (GM) to update the unrevoked users' group secret keys and generate a re-encryption key. The re-encryption technology is applied to prevent the revoked users from decrypting ciphertexts. Moreover, the decryption cloud server provider (D-CSP) executes most of decryption operations to decrease computation costs. The performance analysis indicates that our scheme is practical and efficient. The security of the presented scheme is reduced to divisible computable Diffie-Hellman (DCDH) assumption.(c) 2022 Elsevier Inc. All rights reserved.
查看更多>>摘要：Aspect-based sentiment analysis is a classic fine-grained approach that aims to distinguish sentiment polarities towards a particular aspect target. The majority of research on this topic has been devoted to constructing syntax-based graph convolutional networks (GCNs) for context feature vectors. These approaches perform poorly in terms of node representation and capturing long-distance dependency. In this paper, we focus on the ability of graph convolution and propose an aggregated graph convolutional network (AGCN) to enhance the representation ability of target nodes. To exploit the node feature information, we introduce two aggregator functions to iteratively update the representation of each node from its local neighborhood. To extract more associated node information, we also apply the subdependency of nodes to aggregate the node features, and then employ the attention mechanism to capture the sentiment dependencies between different node feature information. The proposed AGCN is evaluated on large Chinese and English datasets to prove the effect of our model in aspect-based sentiment analysis. The experimental results show that our model is valid compared with other GCN-based methods.
Al-Andoli, Mohammed NasserTan, Shing ChiangCheah, Wooi Ping
24页
查看更多>>摘要：In this paper, a parallel deep learning-based community detection method in large complex networks (CNs) is proposed. First, a CN partitioning method is employed to divide the CN into multiple chunks to improve the efficiency in terms of space and time complexities. Next, the method is integrated with two optimization algorithms: (1) backpropagation (BP), which optimizes deep learning locally within each local chunk of the CN; (2) particle swarm optimization (PSO), which is used to improve the BP optimization involving all CN chunks. PSO utilizes a multi-objective function to improve the effectiveness of the proposed method. In addition, a distributed environment is set up to conduct parallel optimization of the proposed method so that multi-local optimizations could be performed simultaneously. A set of 16 real-world CNs in a range from small to large size are used to verify the effectiveness and efficiency of the method in a benchmark study. The proposed method is implemented in multi-machines with central processing unit (CPU) and graphics processing unit (GPU) devices. The results reveal the effective role of the proposed deep learning with hybrid BP-PSO optimization in detecting communities in large CNs, which requires minimum execution time on both CPU and GPU devices.(c) 2022 Elsevier Inc. All rights reserved.
查看更多>>摘要：Recent years have witnessed unprecedented success achieved by deep learning models in the field of computer vision. However, their vulnerability towards carefully-crafted adversarial examples has also attracted the increasing attention of researchers. Motivated by the observation that adversarial examples are due to the non-robust feature learned from the original dataset by models, we propose the concepts of salient feature (SF) and trivial feature (TF). The former represents the class-related feature, while the latter is usually adopted to mislead the model. We extract these two features with coupled generative adversarial network model and put forward a novel detection and defense method named salient feature extractor (SFE) to defend against adversarial attacks. Concretely, detection is realized by separating and comparing the difference between SF and TF of the input. At the same time, correct labels are obtained by re-identifying SF. Extensive experiments are carried out on MNIST, CIFAR-10, and ImageNet datasets where SFE shows superior results in effectiveness and efficiency compared with state-of-the-art baselines. Furthermore, we provide an interpretable understanding of the defense and detection process. The code of SFE could be downloaded from ( https://github.com/haibinzheng/SFE). (c) 2022 Elsevier Inc. All rights reserved.
查看更多>>摘要：With thousands of new questions posted every day on popular Q&A websites, there is a need for automated and accurate software solutions to replace manual moderation. In this paper, we address the critical drawbacks of crowdsourcing moderation actions in Q&A communities and demonstrate the ability to automate moderation using the latest machine learning models. From a technical point, we propose a multi-view approach that generates three distinct feature groups that examine a question from three different per-spectives: 1) question-related features extracted using a BERT-based regression model; 2) context-related features extracted using a named-entity-recognition model; and 3) gen-eral lexical features derived using statistical and analytical methods. As a last step, we train a gradient boosting classifier to predict a moderation action. For evaluation purposes, we created a new dataset consisting of 60,000 Stack Overflow questions classified into three choices of moderation actions. Based on cross-validation on the novel dataset, our approach reaches 95.6% accuracy as a multiclass task and outperforms all state-of-the -art and previously-published models. Our results clearly demonstrate the high influence of our feature generation components on the overall success of the classifier.(c) 2022 Elsevier Inc. All rights reserved.
查看更多>>摘要：Clustering is one of the most important tasks in the field known as 'Exploratory Data Analysis' (EDA). It explores the dependencies hidden in individual data attributes, dividing them from one set into smaller subsets. In this paper, a Parallel Complete Gradient Clustering Algorithm (PCGCA) is proposed. The Complete Gradient Clustering Algorithm (CGCA) provides a natural interpretation combined with no need for assumptions regarding the number of clusters, making it an appealing choice. Moreover, in CGCA, internal optimization procedures point out the parameters influencing the size of clusters. Algorithms based on kernel density estimation can, therefore, be applied for diverse practical scenarios. Another very useful usage is outlier detection - which is especially important in the currently fast-growing data industry. The described algorithm has been validated in terms of both the speed of calculation and the quality of the obtained solution. The quality of the solution was evaluated with the use of eleven clustering indexes calculated on six data sets. In addition, the obtained result was compared with several classical well-known methods of clustering.(c) 2022 Elsevier Inc. All rights reserved.
查看更多>>摘要：In most previous opinion dynamics research, the opinion evolution is usually defaulted to evolving over a single topic; however, in more general scenarios, agents often discuss several topics at the same time. This paper adopts a multidimensional opinion dynamics model in which the opinions are represented by vectors to reveal the multi-topic opinion pattern. To be more in line with reality, this paper employs a multidimensional version of the Friedkin-Johnsen (FJ) model, where each stubborn agent has a time-evolving stubbornness level. The theoretical analysis finds that the proposed model is convergent if the interpersonal influence matrix is stochastic indecomposable and aperiodic. To achieve unified consensus, opinion leaders are introduced, after which the consensus conditions for the proposed model are given. The theoretical conclusions suggest that in the proposed model, the stubborn agents are as prominent as the opinion leaders in the opinion formation process. Simulations are then conducted in two types of artificial networks, from which it is found that compared with the scalar version of the proposed model and standard multidimensional FJ models in which agents' stubbornness levels remain unchanged, the proposed model produce a more compact final opinion space. The results show that to generate smaller opinion distance and higher opinion correctness degree, it is necessary to ensure a lower proportion of stubborn agents and higher network connectivity. This is consistent with people's intuition. The population size is found to have little effect on the results, but larger number of topics result in larger opinion distances. These findings are enlightening and helpful to opinion managers.(c) 2022 Elsevier Inc. All rights reserved.
NSTL
Elsevier