Search

• Temporal information is effectively integrated into the representation model. • A new prediction model is proposed to gain temporal information. • The memory capacity of the replay buffer is smaller than the existing methods. • The policy lag is proven to be decreased quickly in maximum-entropy reinforcement learning. Partially observable Markov decision process (POMDP) is a key challenging problem in the application of reinforcement learning since it comprehensively describes real agent-environment interactions. Recent works mainly utilize conventional reward signals to train a representation that converts POMDPs to MDPs. However, rewards alone are not enough for a good representation without temporal information. In this paper, we first introduce a novel Recurrent Prediction Model to integrate temporal information into the representation that solves POMDP problems by training three additional unsupervised prediction models, named transition model, reward recovery model, and observation recovery model. This paper secondly makes a modification of the data structure of vanilla replay buffer to reduce the memory usage and thirdly proposes an off-policy correction algorithm to decrease the policy lag in POMDPs. The experiments show that our model achieves better performance in partially observable environments on both stand-alone and distributed training systems. [ABSTRACT FROM AUTHOR]

This paper studies model predictive security control (MPSC) for networked switched systems under denial-of-service (DoS) attacks. Most of existing works only adjust the triggering scheme when being attacked. Different from them, this paper proposes a novel timing-response event-triggering scheme (TR-ETS) to reduce the impact of attacks on system performance, which can not only configure system resources adaptively, but also accurately detect attack information and compensate the attacked data. Specifically, the proposed scheme includes two event-based triggers, which can dynamically and jointly regulate the communication/calculation ability, generate virtual attack sequences and acquire the number of passive packet loss. Then, based on the triggered states, a class of model predictive controllers is designed to optimize the control action. Due to possible strong attacks, a security control framework including network and local loops be introduced and a permissable type-switching mechanism (PTM) is used. Under the permissable controllers (i.e., network and local controllers), sufficient conditions for the stability of closed-loop switched systems are derived. In addition, a set of model predictive optimization algorithm using linear matrix inequalities (LMIs) technique is addressed. Finally, the effectiveness of the proposed method is verified by illustrative examples. [ABSTRACT FROM AUTHOR]

This paper introduces a prediction model rooted in sparse representation and transfer learning, with the primary objective of predicting user behavior during rumor propagation. Users' behavior is dynamic, and rumor, rumor-refuting, and rumor-promoting messages interact dynamically, according to the model. Firstly, this paper proposes to compensate for the low performance of propagation prediction models due to data sparsity at the beginning of rumor topics' lifes. Transfer learning is used to compensate for the data sparsity problem at the beginning of the topic. To map the rumor topic space effectively, a low-rank dense vectorization algorithm based on sparse representation is proposed. Finally, to mine the potential impact of multiple types of information on users, a model based on three-party game theory is constructed. It considers the complex interaction between rumor, rumor-refuting, and rumor-promoting information in the propagation of rumor. Additionally, this paper develops a model of dynamic user behavior prediction using Rumor-Attention-Mechanism-Graph-Attention (RAM-GAT) to predict rumor propagation. Experiments demonstrate that our model can effectively mine the interaction influence between multiple types of information. It can accurately predict user behavior when initial data is insufficient. In addition, rumor propagation patterns and trends are revealed. • A user behavior prediction model is proposed. • Sparse representation can be used to express the rumor feature space. • The complex influences that influence user behavior are quantified. [ABSTRACT FROM AUTHOR]

This paper proposes an approach for the analysis and comparison of stochastic solvers based on the statistical distribution of their variables. The observed variables of the stochastic solvers are the runtime and number of function evaluations required to reach a (sub)-optimal solution. These variables were measured using a target approach. We extended the conventional approach, which usually predicts the average value of variables, by predicting their statistical distributions. If possible, we can predict, not only the average value, but also all the values of the observed variables according to the identified distribution and given probability. We can also predict the probability of reaching (sub)-optimal solutions according to the variable's values. The approach was empirically validated by comparing solvers for discrete and continuous problems. In our experiment, the differences between predicted and measured values of runtime, number of function evaluations and probability are 10% or less. This proves that the proposed approach is useful and can be used for the analysis and comparison of stochastic solvers. Although, the approach has some limitations, it can mitigate the issue of stochastic solvers being unable to provide (sub)-optimal solutions, and can be used to determine the stopping optimization criteria. • An approach to the comparison of stochastic solvers. • The predictive modeling based on optimal and sub optimal solutions. • The analysis of the distribution of runtime and number of function evaluations. • The predictive models based on distribution of solvers' variables are established. • The predictive models provide probability of reaching the optimal solutions. [ABSTRACT FROM AUTHOR]

Multi-label learning (MLL) copes with applications such as text classification and image recognition where each instance is associated with multiple labels. Currently, some label-specific feature conversion approaches such as LIFT serve each individual label well, however do not take advantage of label correlation. Some label correlation exploitation approaches such as GLOCAL have independent model prediction and label correlation stages, thus limiting their learning ability. This paper introduces LSTC, a new algorithm that generates label-specific features and exploits third-order label correlation. On the one hand, for each label, we choose the same number of positive and negative representative instances inspired by density peak. These instances are used to generate a new data matrix with label-specific features. On the other hand, we train a paired output prediction network for each label based on these matrices of its own and that of two auxiliary labels. In this way, third-order label correlations are implicitly exploited. Specifically, the two auxiliary labels are the most similar and least similar, respectively, to avoid getting stuck in local optima. Experiments are conducted on seventeen benchmark datasets in comparison with ten popular algorithms. Results on eight measures demonstrate the superiority of LSTC on data from various domains except the text-domain. The source code is available at github.com/fansmale/lstc. • We adopt density peaks under Gaussian kernel to extract label-specific features. • We train the model for each label using the data of two auxiliary labels. • We use the softmax function to obtain numerical predictions from paired outputs. • Our algorithm shows superiority in ranking-based measures of non-text data. [ABSTRACT FROM AUTHOR]

In this paper, we investigate the impact of optimal hyper-parameter configuration in relational topic models. The main goal is to validate the hypothesis that single-objective Bayesian Optimization (BO) can discover a hyper-parameter setting that leads a set of relational topic models to simultaneously ensure good prediction capabilities and significant topics from a qualitative perspective. Our research, as a result of a comparative analysis performed on 7 state-of-the-art models, 5 performance measures and 3 datasets, has highlighted three main findings: (1) the majority of relational topic models are not able to offer a good trade-off between classification capabilities and topic interpretability; (2) single-objective optimization of hyper-parameters, targeted on maximizing the F1-Measure, is able to create topics that are also optimal with respect to the Kullback Leibler divergence measure; (3) the Pareto frontiers across several performance metrics reveals that the most promising trade-off between the performance metrics can be obtained by Constrained Relational Topic Models. [ABSTRACT FROM AUTHOR]

Machine learning models, especially deep neural network (DNN) models, have widespread and valuable applications in business activities. Training a deep learning model for commercial use requires plenty of private data, expert knowledge, and computing resources. The huge commercial value of such trained models has attracted the attention of attackers. Attackers can construct a dataset by repeatedly querying the target model for the output of the requested samples and then train a substitute model on this dataset that functions similarly to the target model. In this paper, we propose a defense scheme based on physical unclonable function (PUF) against such black-box model extraction attacks. We deploy a PUF on the user side and the corresponding PUF model on the service provider side to ensure that only legitimate users can obtain the correct model predictions. Our experimental results show that by choosing a suitable fuzzy extractor threshold d , legitimate users can recover more than 99.5% of the prediction results with a little additional computational overhead to the service provider. We perform a model extraction attack in the most favorable case for the attacker, and the prediction accuracy of the obtained substitute model is only about 10%, which demonstrates the effectiveness of our proposed scheme. Compared to existing defenses, our scheme not only effectively prevents black-box model extraction attacks but also ensures that the accuracy of the prediction service for legitimate users is not affected. • Defense based on PUF obfuscation against black-box model extraction attacks. • A PUF and the corresponding PUF model ensure that legitimate users can obtain correct model predictions. • Legitimate users can correctly recover more than 99.5% of the predictions with a little additional computational overhead. • The prediction accuracy of the model extracted by the attacker is only about 10%. [ABSTRACT FROM AUTHOR]

• A review and comparison of surrogate models used in single-objective SAEAs. • Review surrogate models in terms of absolute and relative fitness models. • The construction and prediction time of models, and model's accuracy are compared. • SAEAs with different surrogate models are compared. Surrogate-assisted evolutionary algorithms (SAEAs), which use efficient surrogate models or meta-models to approximate the fitness function in evolutionary algorithms (EAs), are effective and popular methods for solving computationally expensive optimization problems. During the past decades, a number of SAEAs have been proposed by combining different surrogate models and EAs. This paper dedicates to providing a more systematical review and comprehensive empirical study of surrogate models used in single-objective SAEAs. A new taxonomy of surrogate models in SAEAs for single-objective optimization is introduced in this paper. Surrogate models are classified into two major categories: absolute fitness models, which directly approximate the fitness function values of candidate solutions, and relative fitness models, which estimates the relative rank or preference of candidates rather than their fitness values. Then, the characteristics of different models are analyzed and compared by conducting a series of experiments in terms of time complexity (execution time), model accuracy, parameter influence, and the overall performance when used in EAs. The empirical results are helpful for researchers to select suitable surrogate models when designing SAEAs. Open research questions and future work are discussed at the end of the paper. [ABSTRACT FROM AUTHOR]

• A novel dynamic social interaction modeling mechanism based on structure entropy minimization. • A separate scene node to model the motion patterns and behavioral preferences. • A multi-time-scale motion tendency modeling module. • State-of-the-art results on the ETH, UCY and SDD datasets. Trajectory is an important basis for reflecting the behavior of moving agents and can be used for various applications. Autonomous systems navigating in complex scenes should have the ability to predict the future locations of surrounding agents and avoid collisions. Agents in a scene interact with each other constantly, which is the greatest challenge of trajectory prediction. However, existing methods have problems with insufficient, superfluous or inaccurate interactions, whether they are distance-, attention- or dense graph-based interactive mechanisms. Moreover, the dynamic complexity and validity of social interactions cannot be quantified. In this paper, we propose a dynamic social interaction modeling mechanism based on structure entropy minimization. Specifically, structure entropy minimization provides a principle for detecting and quantifying the natural or true interactions between agents. Additionally, considering that most agents under the same scenario often have similar behavioral preferences, we introduce scene features to model the interscene variance and intrascene consistency of motion patterns. Finally, through multi-time-scale motion tendency modeling, we can simultaneously model short-term and long-term intentions to alleviate accumulated errors for pedestrians avoiding collisions. We evaluate our proposed method on three trajectory datasets, and the experiments and comparisons demonstrate that our method outperforms comparative state-of-the-art methods while demonstrating the capacity to model social interaction and motion patterns. [ABSTRACT FROM AUTHOR]

• A novel hierarchical graph representation learning model for drug-target binding affinity prediction. • Represents the drug-target binding affinity data as a hierarchical graph. • Hierarchically integrates coarse- and fine-level information in a coarse-to-fine manner. Computationally predicting drug-target binding affinity (DTA) has attracted increasing attention due to its benefit for accelerating drug discovery. Currently, numerous deep learning-based prediction models have been proposed, often with a biencoder architecture that commonly focuses on how to extract expressive representations for drugs and targets but overlooks modeling explicit drug-target interactions. However, known DTA can provide underlying knowledge about how the drugs interact with targets that is beneficial for predictive accuracy. In this paper, we propose a novel hierarchical graph representation learning model for DTA prediction, named HGRL-DTA. The main contribution of our model is to establish a hierarchical graph learning architecture to integrate the coarse- and fine-level information from an affinity graph and drug/target molecule graphs, respectively, in a well-designed coarse-to-fine manner. In addition, we design a similarity-based representation inference method to infer coarse-level information when it is unavailable for new drugs or targets under the cold start scenario. Comprehensive experimental results under four scenarios across two benchmark datasets indicate that HGRL-DTA outperforms the state-of-the-art models in almost all cases. [ABSTRACT FROM AUTHOR]

Semi-supervised partial label learning is an emerging weakly supervised learning paradigm dealing with partially labeled data and unlabeled data simultaneously. The supervision information is weak and even missing for training instances. Most existing researches simply propagate labels to unlabeled data and focus on inducing a weighted classifier model, ignoring the importance of acquiring reliable label confidences for training instances. In this paper, by label set assignment, the valid label supervision information can be preserved and propagated to unlabeled instances to a greater extent. Then, the training instances are projected into a lower-dimensional feature space. Reliable label confidence can be obtained by maximizing the dependence between the feature description and the associated label information. After that, the predictive model can be induced by preserving the intrinsic manifold structure in both feature space and label space. Accordingly, a novel semi-supervised partial label learning approach via dependence-maximized label set assignment (Dlsa) is proposed, where an accurate and robust model can be induced based on reliable label confidences obtained by label set assignment and dependence-maximized dimensionality reduction. Extensive experiments on five real-world data sets validate the effectiveness and superiority of the proposed method. [ABSTRACT FROM AUTHOR]

• A hierarchical clustering-based manifold transfer learning (CMTL) method is developed for dynamic multi-objective optimization problems. • A multi-source transfer learning (MSTL) method is designed to fully exploit all historical knowledge, which can excavate valuable knowledge from the historical environment to predict a high-quality initial population for the new environment. • By incorporating the proposed CMTL and MSTL transfer learning methods, a new DMOEA, called MSTL-DMOEA, is proposed for dynamic optimization. • The superiority of the proposed MSTL-DMOEA over existing prediction-based and transfer-based methods is demonstrated. Recently, dynamic multiobjective evolutionary algorithms (DMOEAs) with transfer learning have become popular for solving dynamic multiobjective optimization problems (DMOPs), as the used transfer learning methods in DMOEAs can effectively generate a good initial population for the new environment. However, most of them only transfer non-dominated solutions from the previous one or two environments, which cannot fully exploit all historical information and may easily induce negative transfer as only limited knowledge is available. To address this problem, this paper presents a multiple source transfer learning method for DMOEA, called MSTL-DMOEA, which runs two transfer learning procedures to fully exploit the historical information from all previous environments. First, to select some representative solutions for knowledge transfer, one clustering-based manifold transfer learning is run to cluster non-dominated solutions of the last environment to obtain their centroids, which are then fed into the manifold transfer learning model to predict the corresponding centroids for the new environment. After that, multiple source transfer learning is further run by using multisource TrAdaboost, which can fully exploit information from the above centroids in new environment and old centroids from all previous environments, aiming to construct a more accurate prediction model. This way, MSTL-DMOEA can predict an initial population with better quality for the new environment. The experimental results also validate the superiority of MSTL-DMOEA over several competitive state-of-the-art DMOEAs in solving various kinds of DMOPs. [ABSTRACT FROM AUTHOR]

As the key step of anomaly detection for multivariate time-series (MTS) data, learning the relations among different variables has been explored by many approaches. However, most existing approaches overlook the heterogeneity among variables, that is, different types of variables (continuous numerical variables, discrete categorical variables or hybrid variables) may have different edge distributions. In this paper, we propose a novel semi-supervised anomaly detection framework based on a heterogeneous feature network (HFN) for MTS. Specifically, we first combine the embedding similarity subgraph generated by sensor embedding and the feature value similarity subgraph generated by sensor values to construct a time-series heterogeneous graph, which fully utilizes the rich heterogeneous mutual information among variables. Then, a prediction model containing nodes and channel attentions is jointly optimized to obtain better time-series representations. This approach fuses the state-of-the-art technologies of heterogeneous graph structure learning (HGSL) and representation learning. Experimental results on four sensor datasets from real-world applications demonstrate that our approach achieves more accurate anomaly detection compared to baseline methods, laying a foundation for the rapid positioning of anomalies. [ABSTRACT FROM AUTHOR]

Zero-shot learning (ZSL) aims to recognize unseen classes during training. Transductive methods have advanced in ZSL, however, often rely on pseudo labels based on confidence scores, leading to semantic misalignment between unseen-class image features and corresponding class semantic descriptions due to noisy pseudo labels. In this paper, we introduce a novel Consistency-Guided Pseudo-Labeling (CGPL) to generate high-quality pseudo labels, achieving robust mapping from visual to semantic space for unseen classes. CGPL incorporates a large-scale vision-language model as a collaborator with the ZSL model to generate high-quality pseudo-labels. Then, pseudo-labeled samples with consistent prediction of two models are added to the training set, to learn the visual-to-semantic mapping for unseen classes. Furthermore, we design a quasi-classification loss based on reconstructed unseen prototypes to learn accurate visual-semantic mapping. Consequently, CGPL is further encouraged to obtain higher-quality pseudo labels, and progressively learn the precise visual-semantic mapping for unseen classes throughout the iterative process. Our extensive experimental results across four benchmark datasets highlight the superior performance of CGPL in both CZSL and GZSL settings. [ABSTRACT FROM AUTHOR]

In view of the high propagation space and the complex networking of rumors, this paper proposes a group behavior prediction model based on sparse representation and interaction of complex messages. Firstly, to solve the difficulty in model training caused by the high dimension and complexity of rumor space, sparse representation is considered as the theoretical basis to construct sparse vectors for user node features, and to construct the node feature prediction submodel. Secondly, aiming at the dynamic interactive behavior among complex messages in the rumor space, the driving force of complex messages is quantified with the evolutionary game, the dynamic rumor propagation network is reconstructed, and the structure attribute prediction submodel is constructed. Finally, considering the advantages of model fusion in improving the generalization ability of the single model, the node feature prediction submodel—Submodel Based on SRC and the structure attribute prediction submodel—Submodel Based on Node2Vec are fused. Meanwhile, a dynamic group behavior prediction model under the influence of complex messages is constructed for the time-sensitive nature of rumor propagation. The experimental results show that the model not only effectively explores the interaction between complex messages but also accurately predicts the group behavior and depicts the rules of rumor propagation. [ABSTRACT FROM AUTHOR]

• Model a fusion user embedding with considering local and global interest semantics. • Explain social relations by learning neighbors' bootable and transitive influences. • Propose a hybrid personalized and neighbor attention model for link prediction. Friend link prediction is an important research problem in recommender systems. Existing network embedding and knowledge embedding methods mainly consider the structural relationships of entities, ignoring the explanatory role of text contents. In this paper, we present an explainable friend link recommendation method that leverages direct and indirect similarity of user pairs via fusion embedding of heterogeneous context information. In social networks, while considering user content interests, first, a fusion user embedding method was developed by incorporating external knowledge semantics. Second, for a user pair, we calculate their direct similar relationship using fusion user embeddings. Additionally, based on intermediate neighbors, we compute their indirect similar relationship by using an attention mechanism, which explains neighbors' bootable and transitive influences for learning the social relationship of user pairs. Then, a hybrid personalized and neighbor attention model for friend link prediction was proposed by considering direct and indirect factors. Finally, experiments were conducted on the Sina Weibo datasets, which indicates that the proposed method effectively predicts the friends of users and provides a good interpretation for link prediction recommendation results. [ABSTRACT FROM AUTHOR]

• Model a fusion user embedding with considering local and global interest semantics. • Explain social relations by learning neighbors' bootable and transitive influences. • Propose a hybrid personalized and neighbor attention model for link prediction. Friend link prediction is an important research problem in recommender systems. Existing network embedding and knowledge embedding methods mainly consider the structural relationships of entities, ignoring the explanatory role of text contents. In this paper, we present an explainable friend link recommendation method that leverages direct and indirect similarity of user pairs via fusion embedding of heterogeneous context information. In social networks, while considering user content interests, first, a fusion user embedding method was developed by incorporating external knowledge semantics. Second, for a user pair, we calculate their direct similar relationship using fusion user embeddings. Additionally, based on intermediate neighbors, we compute their indirect similar relationship by using an attention mechanism, which explains neighbors' bootable and transitive influences for learning the social relationship of user pairs. Then, a hybrid personalized and neighbor attention model for friend link prediction was proposed by considering direct and indirect factors. Finally, experiments were conducted on the Sina Weibo datasets, which indicates that the proposed method effectively predicts the friends of users and provides a good interpretation for link prediction recommendation results. [ABSTRACT FROM AUTHOR]

Feature importance and interaction are among the main issues in explainable artificial intelligence or interpretable machine learning. To measure feature importance and interaction, several methods, such as H-statistic and partial dependency, have been proposed. However, it is difficult to understand the practical implications of importance and interaction. In this paper, a new method for measuring feature importance and interaction is proposed. For the classification model, we observed correctly predicted cases in a predictive model and grouped them according to the characteristics of the cases. We derived a method for feature importance and interaction from group information. For the regression model, we grouped cases according to the change in the size of the prediction error. The proposed method supports the same rationale for feature importance and interaction. It also supports the decomposition of feature importance to feature power and feature interactions. To implement the proposed method, three visualization tools, including a feature interaction graph, are implemented. Through the proposed work, we can better understand the working mechanism of a predictive model. [ABSTRACT FROM AUTHOR]

Ordinal regression (OR) is a paradigm which learns a prediction model on the data with ordered classes. Despite much progress in OR, the existing OR works learn the classifier from only one view and the multi-view learning in OR has not been considered. What is more, there may exist uncertain information in the multi-view OR data. In this paper, we put forward a novel approach, called multi-view support vector ordinal regression with uncertain data (MORU), which can improve the OR classifier by incorporating the multi-view information and handling the data uncertainty. In our method, a series of parallel hyperplanes are applied to separate the multi-view ordered data, and the uncertain information is considered in the input data. Then, we adopt a heuristic framework to solve the OR learning problem. Experimental results have illustrated that our method obtains superior performance to the existing OR techniques. [ABSTRACT FROM AUTHOR]

• Propose a feasible paradigm for using multidomain data in big data analysis. • Apply multigranularity analysis to small-scale low-dimensional data sets. • Transform prediction into a classification problem using deep forest computation. When addressing analysis and prediction problems in a specific domain based on big data processing, the following problems often arise: only relationships between features in the domain itself are considered, and existing methods are not effective for training models on small sample data sets. The traditional approach usually obtains the relationships between single-domain features. Analysis and forecasting in the problem domain alone can quickly achieve good accuracy, but due to the limitations of the analysis domain, it becomes increasingly difficult to further improve the prediction accuracy. This paper proposes a novel data analysis approach compatible with small sample sets called multidomain data depth analysis (MODE). In contrast to traditional approaches, MODE emphasizes multidomain data and considers the relationships among feature domains in the original data. The features in each domain are orthogonally extracted, and feature dimensions are expanded in accordance with the characteristics of small data sets. A better prediction model can be obtained by using the expanded and strengthened features. We apply this approach to real big data from the field of sociology to predict annual income based on census data in experiments. The experimental results show that MODE offers a better prediction effect based on small multidomain samples. [ABSTRACT FROM AUTHOR]

[Display omitted] Data selection, which includes feature and instance selection, is often an important step in building prediction systems. Genetic algorithms (GA) frequently allow finding better solutions than classical methods in many areas. This is also true for the instance selection task. The main difficulties and challenges in GA-based instance selection are high computational complexity and decreasing performance with the dataset size growth. This is caused by the fact that each instance is encoded in one chromosome position. Hence bigger datasets result in longer chromosomes. The main contribution of this paper addresses the above problems in a three-step process. In the first step the dataset is divided into several consistent regions by fuzzy clustering. Then GA-based instance selection is performed independently within each cluster. Finally ensemble voting provides seamless aggregation of the partial results from the overlapping clusters. This improves dataset exploitation by more localized search and also takes the advantage of ensemble methods. This method significantly improves the predictive model performance and data reduction in comparison to instance selection performed on the whole training dataset. [ABSTRACT FROM AUTHOR]

Time-series data are widespread in real-world industrial scenarios. To recover and infer missing information in real-world applications, the problem of time-series prediction has been widely studied as a classical research topic in data mining. Deep learning architectures have been viewed as next-generation time-series prediction models. However, recent studies have shown that deep learning models are vulnerable to adversarial attacks. In this study, we prospectively examine the problem of time-series prediction adversarial attacks and propose an attack strategy for generating an adversarial time series by adding malicious perturbations to the original time series to deteriorate the performance of time-series prediction models. Specifically, a perturbation-based adversarial example generation algorithm is proposed using the gradient information of the prediction model. In practice, unlike the imperceptibility to humans in the field of image processing, time-series data are more sensitive to abnormal perturbations and there are more stringent requirements regarding the amount of perturbations. To address this challenge, we craft an adversarial time series based on the importance measurement to slightly perturb the original data. Based on comprehensive experiments conducted on real-world time-series datasets, we verify that the proposed adversarial attack methods not only effectively fool the target time-series prediction model LSTNet, they also attack state-of-the-art CNN-, RNN-, and MHANET-based models. Meanwhile, the results show that the proposed methods achieve a good transferability. That is, the adversarial examples generated for a specific prediction model can significantly affect the performance of the other methods. Moreover, through a comparison with existing adversarial attack approaches, we can see that much smaller perturbations are sufficient for the proposed importance-measurement based adversarial attack method. The methods described in this paper are significant in understanding the impact of adversarial attacks on a time-series prediction and promoting the robustness of such prediction technologies. [ABSTRACT FROM AUTHOR]

• A novel semi-supervised self-organizing fuzzy inference system (S3OFIS+) is introduced for data stream classification. • A novel parameter-free pseudo-labelling strategy is proposed for S3OFIS+ to effectively self-train from unlabelled data. • The proposed S3OFIS+ can self-learn from unlabelled data streams chunk-by-chunk in a single-pass manner with minimum human input. • The proposed S3OFIS+ can provide users high interpretability and explainability thanks to its prototype-based nature. As a powerful tool for data streams processing, the vast majority of existing evolving intelligent systems (EISs) learn prediction models from data in a supervised manner. However, high-quality labelled data can be difficult to obtain in many real-world classification applications concerning data streams, though unlabelled data is plentiful. To overcome the labelling bottleneck and construct a stronger classification model, a novel semi-supervised EIS is proposed in this paper. After being primed with a small amount of labelled data, the proposed method is capable of continuously self-developing its system structure and self-updating the meta-parameters from unlabelled data streams chunk-by-chunk in a non-iterative, exploratory manner by exploiting a novel pseudo-labelling strategy. Thanks to its transparent prototype-based structure and human-understandable reasoning process, the proposed method can provide users high explainability and interpretability while achieving great classification precision. Experimental investigation demonstrates the superior performance of the proposed method. [ABSTRACT FROM AUTHOR]

With the development of online social networks, user behavior prediction based on the data collected from these social networks has attracted increasing attention. In heterogeneous social networks, a node usually has several heterogeneous attributes to describe itself from different angles. However, most existing methods only utilize an attribute of each node and neglect other heterogeneous attributes. Therefore, this paper proposes a new user heterogeneous information embedding method, called user heterogeneous information embedding (UHIE). This method utilizes the attention mechanism to aggregate the heterogeneous attribute information of each neighbor to obtain their low-dimension representation. Then, the graph neural network is employed to aggregate the multi-relational information from neighbors to obtain the low-dimension representation of nodes. Furthermore, a new soft thresholding method is proposed to eliminate the unimportant information, called multi-head self-attention soft thresholding (MSST), which employs the multi-head self-attention mechanism to calculate an importance threshold for each feature. Based on UHIE and MSST, a new user behavior prediction model is proposed, called Heterogeneous Residual Self-Attention Shrinkage Network (HRSN). This model utilizes UHIE to aggregate heterogeneous information including all heterogeneous attribute information of nodes, and employs MSST to eliminate unimportant information. The experimental results on three real-world datasets show the superiority of the proposed model. [ABSTRACT FROM AUTHOR]

Due to small-quantity and often imbalance of labeled samples, it is challenging to establish a robust and accurate prediction model through data-driven methods. To deal with the small dataset problem, new virtual samples may be generated via virtual sample generation (VSG) methods based on the trend of the original small raw dataset, thereby improving modeling performance. Effective VSG is desirable, but also challenging. Conventional VSG usually assumes that the raw sample set contains only a single operating mode. Taking multi-mode into account will improve the VSG based modeling performance since actual processes are often multi-mode. To this end, an information expansion function considering sample density and amount (IEDA) is first developed to expand the domain range of the attributes in this paper. Then, virtual samples under the multiple operating mode condition are generated by proposing a Gaussian mixture model based virtual sample generation (GMMVSG) method. Applications of GMMVSG on Tennessee Eastman benchmark process and an industrial hydrocracking process show significant improvement of modeling and predictions over other conventional VSG methods. [ABSTRACT FROM AUTHOR]

[Display omitted] • A time-varying consensus formation scheme for mobile robots with switching topology is built. • The robots' input can be simultaneously got by the NDMPC-based protocol. • The constraints of the formation system can be considered. In this paper, for controlling the time-varying consensus formation (TVCF) of multiply nonholonomic robots under switching information transforming topology, a generalized consensus formation system, which is controlled by a neural-dynamic-optimized distributed model predictive control (NDMPC) based consensus protocol strategy, is developed. The system consists of an auxiliary consensus maneuvering subsystem and a formation tracking subsystem. Through simultaneously stabilizing the generalized errors consisted of these two subsystems under the switching topologies, the consensus object is achieved. Within each sampling time, the NDMPC method can formulate and solve a constrained quadratic programming (QP) problem with time-varying desired formation pattern and get the optimal inputs for each robot in a distributed manner. The constraints of the system, as well as the switching structure brought by the changing topologies, can be tackled by utilizing the proposed method. In the end, numerical examples verify the effectiveness of the proposed formation control method. [ABSTRACT FROM AUTHOR]

High-precision short-term inbound passenger flow prediction is of great significance to the daily crowd management and line rescheduling in urban rail systems. Although current models have been applied to prediction, most methods need optimization to meet refined passenger flow management demand. In order to better predict the passenger flow, a novel Wave-LSTM model, based on long short-term memory network (LSTM) and wavelet, is introduced in this paper. In an empirical study with practical passenger flow data of Dongzhimen Station in the Beijing Subway system, the hybrid model exhibited more effective performance in terms of prediction accuracy than the existing algorithms, e.g., autoregressive integrated moving average (ARIMA), nonlinear regression (NAR), and traditional LSTM model. The study illustrates that our newly adopted model is a promising approach for predicting high-precision short-term inbound passenger flow. [ABSTRACT FROM AUTHOR]

Dynamic multi-objective optimization problems (DMOPs) are mainly reflected in objective changes with changes in the environment. To solve DMOPs, a transfer learning (TL) approach is used, which can continuously adapt to environmental changes and reuse valuable knowledge from the past. However, if all individuals are transferred, they may experience negative transfers. Therefore, this paper proposes a novel knowledge transfer method for the dynamic multi-objective evolutionary algorithm (T-DMOEA) to solve DMOPs, which consists of a multi-time prediction model (MTPM) and a manifold TL algorithm. First, according to the movement trend of historical knee points, the MTPM model uses a weighted method to effectively track knee points after environmental changes. Then, the knowledge of the suboptimal solution is reused in the non-knee point set using the manifold TL technique, which yields more high-quality individuals and speeds up the convergence. In the dynamic evolutionary process, the knee points and high-quality solutions are combined to guide the generation of the initial population in the next environment, ensuring the diversity of the population while reducing the computational cost. The experimental results show that the proposed T-DMOEA algorithm can converge rapidly in solving DMOPs while obtaining better-quality solutions. [ABSTRACT FROM AUTHOR]