查看更多>>摘要：Existing solutions focus on improving overall accuracy for imbalanced and complex loan datasets, resulting in a lower precise recall for default samples. To embrace these challenges, based on peer-to-peer loan application information, we proposed a multi-view representation learning with Kolmogorov-Smirnov (KS) to effectively organize these complex data and predict default. Firstly, the features were automatically represented as multi views based on their discreteness and correlation difference. Then, a corresponding multi-view deep neural network (MV-DNN) was developed to obtain knowledge in a multi-view way. Here, we firstly designed different view learning layers to obtain knowledge in corresponding views. Subsequently, to interact with the knowledge in different views, an information fusion layer was developed to fuse the acquired information. To face the challenge from imbalanced data distribution, the KS was set as evaluation metric to assist in training MV-DNN to improve the distinguishing ability for two classes of samples. The experimental results show compared with the MV-DNNs based on random and k means multi-view strategies, and other advanced models, our method could provide optimal comprehensive performance and the most stable multi-view organizing results. Furthermore, we also verified the KS is the key component to assist the model in dealing with the imbalanced dataset.(c) 2022 Elsevier Inc. All rights reserved.
查看更多>>摘要：Expert finding is an important research field in community question answering (CQA). Traditional expert finding methods mainly exploit topic analysis and authority calculation methods to identify high-quality experts in certain fields. To avoid recommending questions to those experts who do not display the willingness or ability to provide high quality answers, user interest drift and user quality should be considered. This study proposes a novel method named high-quality domain expert finding in CQA based on multi granularity semantic analysis and interest drift (HQExpert). Firstly, HQExpert considers different semantic granularities by employing two models, a coarse-grained topic model LCLDA and a fine-grained model (BERT), to capture the domain information of questions and users more accurately. Secondly, to address the diverse interests of the users, a user interest drift model in HQExpert is developed to dynamically represent the changes in the interests of the users at different periods. In addition, a user quality model is developed to further optimize the professional level of the user, finding experts who can provide high-quality answers and are interested in the current question. Finally, extensive experiments on two datasets from different domains demonstrate that the proposed HQExpert model can significantly improve the accuracy of finding high-quality experts. (c) 2022 Elsevier Inc. All rights reserved.
查看更多>>摘要：The density peak clustering (DPC) algorithm is a famous density-based method for exploring, analyzing, and deriving information from data. In the case of application, it encounters the following bottlenecks. Firstly, its Euclidean distance based similarity measurement is prone to misclassification of neighbors. Secondly, its clustering results are significantly influenced by human factors (controlling the cut-off distance parameter dc and selecting the clustering center manually). Finally, the local density q is affected by parameter dc and cannot reflect the sparseness of the data distribution. To overcome these deficiencies, a novel density clustering algorithm, called NAPC, is proposed. First, a Divergence distance is defined to evaluate the similarity of points under the refined Euclidean distance space. Second, the specific value of dc is obtained based on the theory of Adjusted Boxplot. Then, the local density q is reinterpreted based on the above Divergence distance and newly assigned dc. Finally, a judgement index is given to determine which points can be regarded as the center points in the Divergence distance space. The performance of our proposal is evaluated by comparing with several existing clustering methods on various datasets. Experimental results prove that NAPC performs much better than several comparison algorithms. It can identify clusters of various shapes and spatial dimensions with minimal human intervention.
查看更多>>摘要：Intensity inhomogeneity and high noise are two common but challenging issues in image segmentation and is particularly pronounced when the two issues simultaneously appear in one image. As a result, most existing level set methods yield poor performance when applied to these images. To address this issue, this paper proposes a robust variational level set method (RVLSM) based on adaptive diffusion mechanism and local cluster criterion, which can not only correct the severe inhomogeneous intensity but also denoise in segmentation. Specifically, we first define an adaptive-scale representation term using the proposed adaptive diffusion mechanism to transform the image data towards diffusion induced space, which successfully restrains different types/levels of noise while enhancing image details. Then, a new bias field correction term is constructed via estimating the bias in transformed domain to better correct the severe inhomogeneous intensity while segmentation. Finally, an enhanced fourth-order piecewise polynomial penalty term is designed to eradicate numerical calculation instability and tedious re-initialization during the evolution of the level set. The experimental results on synthetic and real images with severe intensity inhomogeneity and high noise demonstrate the superiority of the proposed method over most existing methods in both accuracy and robustness.(c) 2022 Elsevier Inc. All rights reserved.
查看更多>>摘要：Emojis have been widely used in social media as a new way to express various emotions and personalities. However, most previous research only focused on limited features from textual information while neglecting rich emoji information in user-generated content. This study presents two novel attention-based Bi-LSTM architectures to incorporate emoji and textual information at different semantic levels, and investigate how the emoji information contributes to the performance of personality recognition tasks. Specifically, we first extract emoji information from online user-generated content, and concatenate word embedding and emoji embedding based on word and sentence perspectives. We then obtain the document representations of all users from the word and sentence levels during the training process and feed them into the attention-based Bi-LSTM architecture to predict the Big Five personality traits. Experimental results show that the proposed methods achieve state-of-the-art performance over the baseline models on the real dataset, demonstrating the usefulness and contribution of emoji information in personality recognition tasks. The findings could help researchers and practitioners better understand the rich semantics of emoji information and provide a new way to introduce emoji information into personality recognition tasks.(c) 2022 Elsevier Inc. All rights reserved.
查看更多>>摘要：DeepFakes are widespread on social networks, and they result in severe information concerns. Although various detection methods have been proposed, there are still practical limitations. Previous specific artifact-based methods were insufficient to capture finegrained features, which limited their effectiveness against advanced DeepFakes. Current DNN-based detectors tend to trade high costs for performance improvement, and are not efficient enough, given that DeepFakes can be created easily by mobile apps, and DNNbased models require expensive computational resources. Furthermore, most methods lack generalizability under the cross-dataset scenario. In this work, we instead mine the more subtle and generalized defects of DeepFakes and propose the fused facial region_feature descriptor (FFR_FD), which is only a vector of the discriminative feature description, for effective and fast DeepFake detection. We show that DeepFake faces have fewer feature points than real ones, especially in facial regions. FFR_FD capitalizes on such key observations, and thus has strong generalizability. We train a random forest classifier with FFR_FD to achieve efficient detection. Extensive experiments on six large-scale DeepFake datasets demonstrate the effectiveness of our lightweight method. Our model generalizes well on the challenging Celeb-DF (v2) dataset, with 0.706 AUC, which is superior to most stateof-the-art methods. (c) 2022 Elsevier Inc. All rights reserved.
查看更多>>摘要：Discrete and delayed laboratory analyses of product quality restrict the operational opti-mization of industrial processes. However, it is challenging to build an accurate online esti-mation model for product quality because of complex process dynamics, multiple working conditions, and multi-rate characteristics. Therefore, a multimode mechanism-guided pro-duct quality variable estimation model is proposed in this study. First, representative fea-tures are extracted from high-dimensional and redundant process variables via both feature engineering and deep learning to describe the internal reaction state. Then, the rep-resentative features are used to partition the data samples which are used to train the multi-mode long short-term memory (LSTM) network to increase the adaptability of the estimation model. Finally, the LSTM units are cascaded to learn the variation in the quality variable against time to handle the multi-rate problem. The mechanism models are placed in parallel with the LSTM units to guide the learning process. The estimation model utilizes production data, mechanism knowledge and working condition information, which increases model interpretability and adaptability. A zinc fluidized bed roaster is used to illustrate the proposed estimation approach. The simulation results demonstrate the feasi-bility and effectiveness of the proposed multi-rate estimation approach.(c) 2022 Elsevier Inc. All rights reserved.
查看更多>>摘要：In this paper, we divide the population into three groups: susceptible individuals (S), infec-tious individuals (I) and removed individuals (R), and propose a rumor propagation dynamic model with Allee effect and cross-diffusion. Next, we have analyzed a general form of cross-diffusion model with time delay, and drawn a general conclusion of linear stability analysis of Turing bifurcation. However, Turing bifurcation analysis cannot give the specific shapes of the patterns under certain conditions. With the help of the "Multiple Scale Analysis" method, we derive the expression of the amplitude equation for the general form of weakly nonlinear models. Finally, we apply the above theorems to the analysis of our previously proposed model, and derive the appearance condition of the Turing bifurcation and the expression of the amplitude equation respectively. Through the numerical simulations, we have verified the correctness of the above theoret-ical analysis.(c) 2022 Elsevier Inc. All rights reserved.
查看更多>>摘要：Cyber-attacks can tamper with measurement data from physical systems via communication networks of smart grids, which could potentially lead circuit breakers to trip creating a false fault in the absence of any faulty section. Accordingly, a fault diagnosis method should first determine whether a fault is actually present; however, current diagnosis methods of power systems struggle to achieve this goal. This paper proposes a novel method for fault diagnosis based on memory spiking neural P systems, which can distinguish false faults caused by measurement tampering attacks. The proposed method consists of three modules with the functions of suspicious fault section detection, measurement tamper attack identification and fault diagnosis, respectively. The suspicious fault section detection module is used to find candidate sections to reduce the fault diagnosis scope. The attack identification module is designed to identify whether a possibly faulty section is under the measurement tampering attack or not. The fault diagnosis module is devised to diagnose true faults, detecting both the fault sections and their corresponding fault types. To achieve the above goals, inspired by the memory recall mechanism of human brains, a memory spiking neural P system and a corresponding general matrix reasoning algorithm are proposed, which can synthetically utilize the remote measurements and remote signals via a new modeling mechanism. Finally, case studies based on the IEEE 14 and IEEE 118 bus systems verify the feasibility and effectiveness of the proposed method.(c) 2022 Elsevier Inc. All rights reserved.
查看更多>>摘要：This paper investigates the bipartite time-varying output formation tracking issue for general linear multi-agent systems with multiple leaders and matched uncertainties. The outputs of followers are required to form two antagonistic prescribed time-varying subformations, one of which tracks the convex combination of the outputs of multiple leaders and the other tracks the symmetric convex combination. An independent asynchronous fully distributed event-triggered control protocol is formulated by using relative information between neighboring agents. It is shown that, with the formulated protocol, the bipartite time-varying output formation tracking can be achieved if the feasible formation condition is satisfied. The Zeno behavior is proved to be excluded. Then, we further construct a novel self-triggered control protocol to avoid continuous monitoring of estimate error. Since the protocol incorporates both event-triggered control and adaptive control, it efficiently avoids continuous communication among agents and can be implemented in a fully distributed manner. Moreover, it is noteworthy that in the case where relative information is available while absolute information is not, the protocol is applicable. Finally, the feasibility of the constructed protocols is verified by a numerical example.(c) 2022 Elsevier Inc. All rights reserved.
NSTL
Elsevier