Abstract In this paper an action recognition method that can adaptively handle the problems of variations in camera viewpoint is introduced. Our contribution is three-fold. First, a space-sampling algorithm based on affine transform in multiple scales is proposed to yield a series of different viewpoints from a single one. A histogram of dense optical flow is then extracted over each fixed-size patch for a given generated viewpoint as a local feature descriptor. Second, a dimension selection procedure is also proposed to retain only the dimensions that have distinctive information and discard the unnecessary ones in the feature vector space. Third, to adapt to a situation in which video data in multiple viewpoints are used for training; an extended method with a voting algorithm is also introduced to increase the recognition accuracy. By conducting experiments using both simulated and realistic datasets (http://www.aislab.org/index.php/en/mvar-datasets), the proposed method is validated. The method is found to be accurate and capable of maintaining its accuracy under a wide range of viewpoint changes. In addition, the method is less sensitive to variations in subject scale, subject position, action speed, partial occlusion, and background. The method is also validated by comparing with state-of-the-art view-invariant action recognition methods using well-known i3DPost and MuHAVi public datasets.

Abstract Inthis article, wefocus on the classification problem to semi-supervised learning in non-stationary environment. Semi-supervised learning is a learning task from both labeled and unlabeled data points. There are several approaches to semi-supervised learning in stationary environment which are not applicable directly for data streams. We propose a novel semi-supervised learning algorithm, named STDS. The proposed approach uses labeled and unlabeled data and employs an approach to handle the concept drift in data streams. The main challenge in semi-supervised self-training for data streams is to find a proper selection metric in order to find a set of high-confidence predictions and a proper underlying base learner. We therefore propose an ensemble approach to find a set of high-confidence predictions based on clustering algorithms and classifier predictions. We then employ the Kullback-Leibler (KL) divergence approach to measure the distribution differences between sequential chunks in order to detect the concept drift. When drift is detected, a new classifier is updated from the new set of labeled data in the current chunk; otherwise, a percentage of high-confidence newly labeled data in the current chunk is added to the labeled data in the next chunk for updating the incremental classifier based on the proposed selection metric. The results of our experiments on a number of classification benchmark datasets show that STDS outperforms the supervised and the most of other semi-supervised learning methods.

Abstract Rain streaks usually give rise to visual degradation and cause many computer vision algorithms to fail. So it is necessary to develop an effective deraining algorithm as preprocess of high-level vision tasks. In this paper, we propose a novel deep learning based deraining method. Specifically, the multi-scale kernels and feature maps are both important for single image deraining. However, the previous works ignore the two multi-scale information or only consider the multi-scale kernels information. Instead, our method learns multi-scale information both from the perspectives of kernels and feature maps, respectively, by designing spatial contextual information aggregation module and pyramid network module. The former module can capture the rain streaks with different sizes and the latter module can extract rain streaks from different scales further. Moreover, we also employ squeeze-and-excitation and skip connections to enhance the correlation between channels and transmit the information from low-level to high-level, respectively. The experimental results show that the proposed method achieves significant improvements over the recent state-of-the-art methods in Rain100H, Rain100L, Rain1200 and Rain1400 datasets.

Abstract User-based collaborative filtering is widely used in recommendation systems, which normally comprises three steps: (1) finding the nearest conceptual neighbors, (2) aggregating the neighbors’ ratings to predict the ratings of unrated items, and (3) generating recommendations based on the prediction. Existing algorithms mainly focus on steps 1 and 3 but neglect subtle treatments of aggregating neighbors’ suggestions in step 2. Based on the discovery of psychology that (i) users’ responses to positive and negative suggestions are different, and (ii) users may respond differently from one another, this paper proposes a Personal Asymmetry Response-based Suggestions Aggregation (PARSA) algorithm, which first uses a linear regression method to learn each user’s response to negative/positive suggestions from neighbors and then uses a gradient descent algorithm for optimizing them. In addition, this paper designs an Identical Asymmetry Response-based Suggestions Aggregation (IARSA) baseline algorithm, which assumes that all the users’ responses to suggestions are identical as references to verify the key contribution of the heuristics employed in our PARSA algorithm that user may responses differently to positive and negative suggestions. Three sets of experiments are designed and implemented over two real-life datasets (i.e., Eachmovie and Netflix) to evaluate the performance of our algorithms. Further, in order to eliminate the influence of different similarity measures, this paper selects three kinds of similarity measures to discover neighbors. Experimental results demonstrate that most people indeed pay more attention to negative suggestions and our algorithms achieve better prediction and recommendation performances than the compared algorithms under various similarity measures.

Abstract This paper is designed based on the mathematical models for bi-level programming in Stackelberg game under type-2 fuzzy environment. The parameters of the objective functions on both levels are considered as type-2 fuzzy numbers in the first case whereas the parameters of the objective functions and the constraints are chosen as type-2 fuzzy numbers in the second case. Critical value based reduction methods are applied to reduce type-2 fuzzy numbers to type-1 fuzzy numbers in the first case. After that, centroid method is used for completely defuzzifying type-2 fuzzy numbers. Besides this, the obtained results are compared with the help of LINGO iterative scheme and genetic algorithm. Coming to the second case, a chance constraint programming with the help of generalized credibility measure is utilized to convert the fuzzy problem to its equivalent crisp form. LINGO iterative scheme is used to solve the deterministic problem using fuzzy programming. The sensitivity analysis is shown to different credibility levels of right hand side of the constraints to find the value of objective function in each level. Finally, real-life based numerical problems are presented to show the performance of the proposed models and techniques. At last, conclusion about the findings and outlook are described.

Abstract With the explosion of unlabelled and high-dimensional data, unsupervised feature selection has become an critical and challenging problem in machine learning. Recently, data representation based model has been successfully deployed for unsupervised feature selection, which defines feature importance as the capability to represent original data via a reconstruction function. However, most existing algorithms conduct feature selection on original feature space, which will be affected by the noisy and redundant features of original feature space. In this paper, we investigate how to conduct feature selection on the dictionary basis space of the data, which can capture higher level and more abstract representation than original low-level representation. In addition, a similarity graph is learned simultaneously to preserve the local geometrical data structure which has been confirmed critical for unsupervised feature selection. In summary, we propose a model (referred to as DGL-UFS briefly) to integrate dictionary learning, similarity graph learning and feature selection into a uniform framework. Experiments on various types of real world datasets demonstrate the effectiveness of the proposed framework DGL-UFS.

Abstract Online learning represents an important family of efficient and scalable algorithms for large-scale classification problems. Many of them are linear with fast computational speed, but when faced with complex classification, they more likely have low accuracies. In order to improve accuracies, kernel trick is applied, however, it often brings high computational cost. In fact, discriminative information is vital in classification which is still not fully utilized in these algorithms. In this paper, we proposed a novel online linear method, called Sketch Discriminatively Regularized Online Gradient Descent Classification (SDROGD). In order to exploit inter-class separability and intra-class compactness, SDROGD utilizes a matrix to characterize the discriminative information and embeds it directly into a new regularization term. This matrix can be updated by the sketch technique in an online manner. After applying a simple but effective optimization, we show that SDROGD has a good time complexity bound, which is linear with the feature dimension or the number of samples. Experimental results on both toy and real-world datasets demonstrate that SDROGD has not only faster computational speed but also much better classification accuracies than some related kernelized algorithms.

Extreme learning machine (ELM) and its variants have been widely used in the field of object recognition and other complex classification tasks. Traditional deep learning architectures like Convolutional Neural Network (CNN) are capable of extracting high-level features, which are the key for the models to make right decisions. However, traditional deep architectures are confronted with solving a tough, non-convex optimization problem, which is a time-consuming process. In this paper, we propose two hierarchical models, i.e., Random Recursive Constrained ELM (R2CELM) and Random Recursive Local- Receptive-Fields-Based ELM (R2ELM-LRF), which are constructed by stacking with CELM or ELM-LRF, respectively. Besides, inspired by the stacking generalization philosophy, random projection and kernelization are incorporated as their constitutive elements. R2CELM and R2ELM-LRF not only fully inherit the merits of ELM, but also take advantage of the superiority of CELM and ELM-LRF in the field of image recognition, respectively. The essence of CELM is to constrain the weight vectors from the input layer to the hidden layer to be consistent with the directions from one class to another class, while ELM-LRF is adept at exploiting the local structures in images through many local receptive fields. In the empirical results, R2CELM and R2ELM-LRF demonstrate their better performance in testing accuracy on the six benchmark image recognition datasets, compared with their basic learners and other state-of-the-art algorithms. Moreover, the proposed two deep ELM models need less training time when compared with traditional Deep Neural Network (DNN) based models.

Abstract The recently proposed twin extreme learning machine (TELM) requires solving two quadratic programming problems (QPPs) in order to find two non-parallel hypersurfaces in the feature that brings in the additional requirement of external optimization toolbox such as MOSEK. In this paper, we propose implicit Lagrangian TELM for classification via unconstrained convex minimization problem (ULTELMC) and further suggest iterative convergent schemes which eliminates the requirement of external optimization toolbox generally required in solving the quadratic programming problems (QPPs) of TELM. The solutions to the dual variables of the proposed ULTELMC are obtained using iterative schemes containing ‘plus’ function which is not differentiable. To overcome this shortcoming, the generalized derivative approach and smooth approximation approaches are suggested. Further, to test the performance of the proposed approaches, classification performances are compared with support vector machine (SVM), twin support vector machine (TWSVM), extreme learning machine (ELM), twin extreme learning machine (TELM) and Lagrangian extreme learning machine (LELM). Moreover, non-requirement to solve QPPs makes the iterative schemes find the solution faster as compared to the reported methods that finds the solution in dual space. Computational times required in finding the solutions are also presented for comparison.

Abstract Sorting the swarm of robots is required when the robots are carrying different loads and they can not simply swap the loads. In 2016, Zhou et al. presented an interesting algorithm to sort a swarm of robots, wherein the authors made a main tree and a feedback tree to assign a topology to the robots, based on which the robots moved while arranging themselves in a sorted order. While the approach was very interesting and the results were critically analyzed by the authors, we see a critical problem that the approach did not account for collisions because of which the results can be very different. In this paper, we extend the work of Zhou et al. by enabling the robots to avoid collision by using a geometric approach called as “follow the gap” method. Together both the algorithms allow robot swarm to sort themselves in a straight line while avoiding collision simultaneously.

As a novel swarm intelligence optimization algorithm, brain storm optimization (BSO) has its own unique capabilities in solving optimization problems. However, the performance of traditional BSO strategy in balancing exploitation and exploration is inadequate, which reduces the convergence performance of BSO. To overcome these problems, a multi-strategy BSO with dynamic parameters adjustment (MSBSO) is presented in this paper. In MSBSO, four competitive strategies based on improved individual selection rules are designed to adapt to different search scopes, thus obtaining more diverse and effective individuals. In addition, a simple adaptive parameter that can dynamically regulate search scopes is designed as the basis for selecting strategies. The proposed MSBSO algorithm and other state-of-the-art algorithms are tested on CEC 2013 benchmark functions and CEC 2015 large scale global optimization (LSGO) benchmark functions, and the experimental results prove that the MSBSO algorithm is more competitive than other related algorithms.

Feature selection plays a critical role in many applications that are relevant to machine learning, image processing and gene expression analysis. Traditional feature selection methods intend to maximize feature dependency while minimizing feature redundancy. In previous information-theoretical-based feature selection methods, feature redundancy term is measured by the mutual information between a candidate feature and each already-selected feature or the interaction information among a candidate feature, each already-selected feature and the class. However, the larger values of the traditional feature redundancy term do not indicate the worse a candidate feature because a candidate feature can obtain large redundant information, meanwhile offering large new classification information. To address this issue, we design a new feature redundancy term that considers the relevancy between a candidate feature and the class given each already-selected feature, and a novel feature selection method named min-redundancy and max-dependency (MRMD) is proposed. To verify the effectiveness of our method, MRMD is compared to eight competitive methods on an artificial example and fifteen real-world data sets respectively. The experimental results show that our method achieves the best classification performance with respect to multiple evaluation criteria.

For unsupervised learning, we propose a new clustering method which incorporates belief peaks into a linear label propagation strategy. The proposed method aims to reveal the data structure by finding out the exact number of clusters and deriving a fuzzy partition. Firstly, the cluster centers and outliers can be identified by the improved belief metric, which makes use of the whole data distribution information so as to correctly highlight the cluster centers without the limitation of massive neighbor points. Secondly, an informative initial fuzzy cluster assignment for each remaining point is created by considering the distances between its neighbors and each cluster center, then the fuzzy label of each point will be iteratively updated by absorbing its neighbors’ label information until the fuzzy partition is stable. The label propagation assignment strategy provides a valuable alternative technique with explicit convergence and linear complexity in the field of belief-peaks clustering. The effectiveness of the proposed method is tested on seven commonly used real-world datasets from the UCI Machine Learning Repository, and seven synthetic datasets in the domain of data clustering. Comparing with several state-of-the-art clustering methods, the experiments reveal that the proposed method enhanced the clustering results in terms of the exact numbers of clusters and the Adjusted Rand Index. Further, the parameter analysis experiments validate the robustness to the two tunable parameters in the proposed method.

Arising from the need of all time for optimization of irrigation systems, distribution network and cable network, Clustered Shortest-Path Tree Problem (CluSPT) has been attracting a lot of attention and interest from the research community. On the other hand, the Multifactorial Evolutionary Algorithm (MFEA) is one of the most recently exploited realms of Evolutionary Algorithms (EAs) and its performance in solving optimization problems has been very promising. Considering these characteristics, this paper describes a new approach using the MFEA for solving the CluSPT. The MFEA has two tasks: the goal of the first task is to determine the best tree (w.r.t. cost minimization) which envelops all vertices of the CluSPT while the goal of the second task is to find the fittest solution possible for the problem. The purpose of the second task is to find good materials for implicit genetic transfer process in MFEA to improve the quality of CluSPT. To apply this new algorithm, a decoding scheme for deriving individual solutions from the unified representation in the MFEA is also introduced in this paper. Furthermore, evolutionary operators such as population initialization, crossover and mutation operators are also proposed. These operators are applicable for constructing valid solution from both sparse and complete graph. Although the proposed algorithm is slightly complicated for implementation, it can enhance ability to explore and exploit the Unified Search Space (USS). To prove this increment in performance i.e, to assess the effectiveness of the proposed algorithm and methods, the authors implemented them on both Euclidean and Non-Euclidean instances. Experiment results show that the proposed MFEA outperformed existing heuristic algorithms in most of the test cases. The impact of the proposed MFEA was analyzed and a possible influential factor that may be useful for further study was also pointed out.

The dimensionality reduction of the unbalanced semi-supervised problem is difficult because there are too few labeled samples. In this paper, we propose a new dimensionality reduction method for the unbalanced semi-supervised problem, called sparse locality preserving projection (SLPP for short). In the past work of solving the semi-supervised dimensionality reduction problems, they either abandon some unlabeled samples or do not utilize the implicit discriminant information of unlabeled samples. While, SLPP learns the optimal projection matrix with the full use of the discriminant information and the geometric structure of the unlabeled samples. Here, we preserve the geometric structure of the rest unlabeled samples and their k-nearest neighbors after increasing the number of labeled samples by label propagation. The optimization problem of SLPP can be easily solved by a generalized eigenvalue problem. Results on various data sets from UCI machine learning repository and two hyperspectral data sets demonstrate that SLPP is superior to other conventional reduction methods.

Ontology matching aims at identifying the correspondences between instances and data properties of different ontologies. The use of data mining approach in matching ontology problem is reviewed in this article. We propose DMOM (Data Mining for Ontology Matching based instances) framework to select data properties of instances efficiently. The framework exploits data mining techniques to select the most appropriate features to match ontologies. Moreover, three strategies have been investigated to select the relevant features for the matching process. The first one called exhaustive, explores the enumerate search tree randomly by generating at each iteration a subset of feature attributes, where each node is evaluated by running the matching process on its selected attributes. The second approach called statistical, it uses some statistical values to select the most relevant properties. The third one called FIM (Frequent Itemsets Mining), it explores the correlation between different properties and selects the most frequent properties describing the overall instances of the given ontology. To demonstrate the usefulness of DMOM framework, several experiments have been carried out on OAEI (Ontology Alignment Evaluation Initiative) and DBpedia ontology databases. The results show that the third strategy, FIM, outperforms the two other strategies (Exhaustive, and Statistical). The results also reveal that DMOM outperforms the state-of-the-art ontology matching approaches in terms of execution time and the quality of the matching process.

Recently, click-through rate (CTR) prediction is a challenge problem in the aspect of online advertising. Some researchers have proposed deep learning-based models that follow a similar embedding and MLP paradigm. However, the corresponding approaches generally ignore the importance of capturing the latent user interest behind user behaviour data. In this paper, we present a novel attentive deep interest-based network model called ADIN. Specifically, we capture the interest sequence in the interest extractor layer, and the auxiliary losses are employed to produce the interest state with the deep supervision. First, we model the dependency between behaviours by using a bidirectional gated recurrent unit (Bi-GRU). Next, we extract the interest evolving process that is related to the target and propose an interest evolving layer. At the same time, attention mechanism is embedded into the sequential structure. Then, the model learns highly non-linear interactions of features based on stack autoencoders. An experiment has been done using four real-world datasets, the proposed model achieves superior performance than the existing state-of-the-art models.

This paper proposes a novel hybrid multi-objective algorithm named Multi-objective Spotted Hyena and Emperor Penguin Optimizer (MOSHEPO) for solving both convex and non-convex economic dispatch and micro grid power dispatch problems. The proposed algorithm combines two newly developed bio-inspired optimization algorithms namely Multi-objective Spotted Hyena Optimizer (MOSHO) and Emperor Penguin Optimizer (EPO). MOSHEPO contemplates many non-linear characteristics of power generators such as transmission losses, multiple fuels, valve-point loading, and prohibited operating zones along with their operational constraints, for practical operation. To evaluate the effectiveness of MOSHEPO, the proposed algorithm has been tested on various benchmark test systems and its performance is compared with other well-known approaches. The experimental results demonstrate that the proposed algorithm outperforms other algorithms with low computational efforts while solving economic and micro grid power dispatch problems.

Abstract In the field of machine learning and data mining, feature interaction is a ubiquitous issue that cannot be ignored and has attracted more attention in recent years. In this paper, we proposed the Symmetrical Complementary Coefficient which can quantify feature interactions very well. Based on it, we improved the Sequential Forward Selection (SFS) algorithm and proposed a new feature subset searching algorithm called SCom-SFS which only needs to consider the feature interactions between adjacent features on a given sequence instead of all of them. Moreover, discovered feature interactions can speed up the process of searching for the optimal feature subset. In addition, we have improved the ReliefF algorithm by screening out representative samples from the original data set, and need not to sample the samples. The improved ReliefF algorithm has been proved to be more efficient and reliable. An effective and complete feature selection algorithm RRSS is obtained through the combination of the two modified algorithms. According to the experimental results, the proposed algorithm RRSS outperformed five classic and two latest feature selection algorithms in terms of size of resulting feature subset, Accuracy, Kappa coefficient, and adjusted Mean-Square Error (MSE).

Abstract Many state-of-the-art shape features have been proposed for the shape recognition task. In this paper, to explore whether a shape feature influences object segmentation, we propose a specific shape feature, Fisher shape (a form of bag of contour fragments), and we combine this with the appearance feature with multiple kernel learning to create a pipeline of object segmentation system. The experimental results on benchmark datasets clearly demonstrate that the pipeline of object segmentation is effective and that the Fisher shape can improve object segmentation with only the appearance feature.

Abstract The no-wait flow shop scheduling problem (NWFSP), as a typical NP-hard problem, has important ramifications in the modern industry. In this paper, a jigsaw puzzle inspired heuristic (JPA) is proposed for solving NWFSP with the objective of minimizing makespan. The core idea behind JPA is to find the best match for each job until all the jobs are scheduled in the set of process. In JPA, a waiting time matrix is constructed to measure the gap between two jobs. Then, a matching matrix based on the waiting time matrix is obtained. Finally, the optimal scheduling sequence is built by using the matching matrix. Experimental results on large-scale benchmark instances show that JPA is superior to the state-of-the-art heuristics.

In this paper, we propose a classification system which is able to decide whether a company is innovative or not, based only on its public website available on the internet. As innovativeness plays a crucial role in the development of myriad branches of the modern economy, an increasing number of entities are expending effort to be innovative. Thus, a new issue has appeared: how can we recognise them? Not only is grasping the idea of innovativeness challenging for humans, but also impossible for any known machine learning algorithm. Therefore, we propose a new indirect technique: a diversified stacked generalisation method, which is based on a combination of a multi-view approach and a genetic algorithm. The proposed approach achieves better performance than all other classification methods which include: (i) models trained on single datasets; or (ii) a simple voting method on these models. Furthermore, in this study, we check if unaligned feature space improves classification results. The proposed solution has been extensively evaluated on real data collected from companies’ websites. The experimental results verify that the proposed method improves the classification quality of websites which might represent innovative companies.

Abstract This research presents a novel aggregating method for constructing an aggregated topic model that is composed of the topics with greater coherence than individual models. When generating a topic model, a number of parameters have to be specified. The resulting topics can be very general or very specific, which depend on the chosen parameters. In this study we investigate the process of aggregating multiple topic models generated using different parameters with a focus on whether combining the general and specific topics is able to increase topic coherence. We employ cosine similarity and Jensen-Shannon divergence to compute the similarity among topics and combine them into an aggregated model when their similarity scores exceed a predefined threshold. The model is evaluated against the standard topics models generated by the latent Dirichlet allocation and Non-negative Matrix Factorisation. Specifically we use the coherence of topics to compare the individual models that create aggregated models against those of the aggregated model and models generated by Non-negative Matrix Factorisation, respectively. The results demonstrate that the aggregated model outperforms those topic models at a statistically significant level in terms of topic coherence over an external corpus. We also make use of the aggregated topic model on social media data to validate the method in a realistic scenario and find that again it outperforms individual topic models.

Dynamic uncertain causality graph (DUCG), which is based on probability theory, is used for uncertain knowledge representation and reasoning. However, the traditional DUCG has difficulty expressing the causality of the events with crisp numbers. Therefore, an intuitionistic fuzzy set based dynamic uncertain causality graph (IFDUCG) model is proposed in this paper. The model focuses on describing the uncertain event in the form of intuitionistic fuzzy sets, which can handle with the problem of describing vagueness and uncertainty of an event in the traditional model. Then the technique for order preference by similarity to an ideal solution (TOPSIS) method is combined with IFDUCG for knowledge representation and reasoning so as to integrate more abundant experienced knowledge into the model to make the model more reliable. Then some examples are used to validate the proposed method. The experimental results prove that the proposed method is effective and flexible in dealing with the difficulty of the fuzzy event of knowledge representation and reasoning. Furthermore, we make a practical application to root cause analysis of aluminum electrolysis and the results show that the proposed method is available for workers to make decisions.

Intelligent manufacturing poses a challenge for fault diagnosis of rotor systems to meet the three tasks: whether exists faults, faults location and quantitative diagnosis. Traditional methods hardly meet all the three tasks in online fault diagnosis. This paper proposes a modified classification and regression tree (CART) algorithm named D-CART algorithm to provide much faster fault classification by reducing the iteration times in computation while still ensuring accuracy. Experiments are carried on to achieve a comprehensive online fault diagnosis for rotor systems such as faults location, faults types and quantitative analysis of unbalanced mass in this paper. In comparison with the other 4 novel CART-based algorithms, the experimental results indicate that the speed of D-CART algorithm is improved by a factor of 23.92 compared to the fastest improved algorithm (Adaboost-CART) and a model accuracy of up to 96.77%. Thus demonstrating the speed superiority of D-CART algorithm in both diagnosing locations of different faults types and determining the loading masses of unbalanced faults. The proposed method has the potential to realize high-accuracy online fault diagnosis for rotor systems.

Multiple kernel learning (MKL) for feature selection utilizes kernels to explore complex properties of features, which has been shown to be among the most effective for feature selection. To perform feature selection, a natural way is to use the l0-norm to get sparse solutions. However, the optimization problem involving l0-norm is NP-hard. Therefore, previous MKL methods typically utilize a l1-norm to get sparse kernel combinations. However, the l1-norm, as a convex approximation of l0-norm, sometimes cannot attain the desired solution of the l0-norm regularizer problem and may lead to prediction accuracy loss. In contrast, various non-convex approximations of l0-norm have been proposed and perform better in many linear feature selection methods. In this paper, we propose a novel l0-norm based MKL method (l0-MKL) for feature selection with non-convex approximations constraint on kernel combination coefficients to select features automatically. Considering the better empirical performance of indefinite kernels than positive kernels, our l0-MKL is built on the primal form of multiple indefinite kernel learning for feature selection. The non-convex optimization problem of l0-MKL is further refumated as a difference of convex functions (DC) programming and solved by DC algorithm (DCA). Experiments on real-world datasets demonstrate that l0-MKL is superior to some related state-of-the-art methods in both feature selection and classification performance.

Automatic traffic sign detection has great potential for intelligent vehicles. The ability to detect small traffic signs in large traffic scenes enhances the safety of intelligent devices. However, small object detection is a challenging problem in computer vision; the main problem involved in accurate traffic sign detection is the small size of the signs. In this paper, we present a deconvolution region-based convolutional neural network (DR-CNN) to cope with this problem. This method first adds a deconvolution layer and a normalization layer to the output of the convolution layer. It concatenates the features of the different layers into a fused feature map to provide sufficient information for small traffic sign detection. To improve training effectiveness and distinguish hard negative samples from easy positive ones, we propose a two-stage adaptive classification loss function for region proposal networks (RPN) and fully connected neural networks within DR-CNN. Finally, we evaluate our proposed method on the new and challenging Tsinghua-Tencent 100K dataset. We further conduct ablation experiments and analyse the effectiveness of the fused feature map and the two-stage classification loss function. The final experimental results demonstrate the superiority of the proposed method for detecting small traffic signs.

Abstract In data analysis, incomplete data commonly occurs and can have significant effects on the conclusions that can be drawn from the data. Incomplete data cause another problem, so-called uncertainty which leads to producing unreliable results. Hence, developing effective techniques to impute these missing values is crucial. Missing or incomplete data and noise are two common sources of uncertainty. In this paper, an effective method for imputing missing values is introduced which is robust to uncertainties that are arising from incompleteness and noise. A kernel-based method for removing the noise is designed. Using the belief function theory, the class of incomplete data is determined. Finally, every missing dimension is imputed considering the mean value of the same dimension of the members belonging to the determined class. The performance has been evaluated on real-world data sets from UCI repository. The results of the experiments have been compared with state-of-the-art methods, which show the superiority of the proposed method regarding classification accuracy.

Hesitant fuzzy linguistic preference relation (HFLPR) as a new preference relation is introduced to express the decision makers’ (DMs’) hesitant preference information for each pairwise comparison between different alternatives or criteria. In this paper, the priority vector and consistency of HFLPR are discussed based on a two-stage optimization and multiplicative consistency. Based on the original hesitant preference information, the multiplicative consistency index of an HFLPR is defined to measure the consistency level of the HFLPR. For an unacceptable multiplicative consistent HFLPR, a goal programming model, which is an integer optimization model, is developed to derive an acceptable, multiplicative, consistent HFLPR. According to probability sampling, a linguistic preference relation (LPR) with the best consistency level and an LPR with the worst consistency level with regard to an HFLPR are defined. Combining the two LPRs, a two-stage optimization framework is constructed to obtain the HFLPR’s priority vector, which considers the DM’s risk preference. A multi-stage optimization approach is proposed to solve decision-making problems by integrating the goal programming model and the two-stage optimization framework. Two real life problems are analyzed to show the feasibility of the proposed approach.

Abstract The capacity of a robot to automatically adapt to new environments is crucial, especially in social robotics. Often, when these robots are deployed in home or office environments, they tend to fail because they lack the ability to adapt to new and continuously changing scenarios. In order to accomplish this task, robots must obtain new information from the environment, and then add it to their already learned knowledge. Deep learning techniques are often used to tackle this problem successfully. However, these approaches, complete retraining of the models, which is highly time-consuming. In this work, several strategies are tested to find the best way to include new knowledge in an already learned model in a deep learning pipeline, putting the spotlight on the time spent for this training. We tackle the localization problem in the long term with a deep learning approach and testing several retraining strategies. The results of the experiments are discussed and, finally, the best approach is deployed on a Pepper robot.

Abstract For mutual benefit, data is shared among business organizations. However, this may result in privacy and security threats. To address this issue, privacy-preserving data mining is presented to sanitize the original database to hide all sensitive knowledge. Privacy-preserving utility mining is an extension of privacy-preserving data mining, the objective of which is to hide all sensitive high-utility itemsets and minimize the side effects on non-sensitive knowledge caused by the sanitization process. In this paper, three heuristic algorithms for privacy-preserving utility mining are proposed, namely, Selecting Maximum Utility item first (SMAU), Selecting Minimum Utility item first (SMIU) and Selecting Minimum Side Effects item first (SMSE). The quality of the database is well maintained because all of the proposed algorithms consider the side effects on the non-sensitive itemsets. Furthermore, to avoid performing multiple database scans, two table structures, T-table and HUI-table, are adopted to accelerate the hiding process by only scanning the database twice. The experimental results show that the proposed approaches successfully conceal all sensitive itemsets with fewer distortions of non-sensitive knowledge. Moreover, the influence of the database density on the proposed approaches is observed.

In order to protect data privacy, a new transfer group probability Naive Bayes algorithm TrGNB is proposed. TrGNB is applied to scenarios in which the source domain contains a large amount of labeled data and only a small amount of unlabeled data group probability information in the target domain. TrGNB integrates the ideology of transfer learning and group probability information into the Naive Bayes model, which not only improves the classification effect of the learning task in the target domain but also protects the data privacy. The TrGNB was verified on the 20-Newsgroups, Reuters-21578 and Email spam datasets. The experimental results show that TrGNB significantly improves the classification accuracy compared with the benchmark algorithms.

Abstract A multi-objective particle swarm optimization based on cooperative hybrid strategy (CHSPSO) is presented in this paper to solve complex multi-objective problems. Most algorithms usually contain only one strategy, which makes them unable to trade off the convergence and diversity when solving the complex multi-objective problems. The proposed cooperative hybrid strategy can effectively guarantee the convergence and the diversity of the algorithm. The multi-population strategy and the dynamic clustering strategy are employed to improve the convergence and the diversity. At the same time, the life strategy and lottery probability selection strategy are used to further ensure the diversity of the population. A series of test functions are used to verify the effectiveness of CHSPSO. The performance of the proposed algorithm is compared with other evolutionary algorithms. The results show that CHSPSO can obtain a better convergence and diversity for the complex multi-objective problems.

Abstract Fuzzy risk analysis plays an important role in mitigating the levels of harm of a risk. In real world scenarios, it is a big challenge for risk analysts to make a proper and comprehensive decision when coping with risks that are incomplete, vague and fuzzy. Many established fuzzy risk analysis approaches do not have the flexibility to deal with knowledge in the form of preferences elicitation which lead to incorrect risk decision. The inefficiency is reflected when they consider only risk analyst preferences elicitation that is partially known. Nonetheless, the preferences elicited by the risk analyst are often non-homogeneous in nature such that they can be completely known, completely unknown, partially known and partially unknown. In this case, established fuzzy risk analysis methods are considered as inefficient in handling risk, hence an appropriate fuzzy risk analysis method that can deal with the non-homogeneous nature of risk analyst’s preferences elicitation is worth developing. Therefore, this paper proposes a novel fuzzy risk analysis method that is capable to deal with the non-homogeneous risk analyst’s preferences elicitation based on grey numbers. The proposed method aims at resolving the uncertain interactions between homogeneous and non-homogeneous natures of risk analyst’s preferences elicitation by using a novel consensus reaching approach that involves transformation of grey numbers into grey parametric fuzzy numbers. Later on, a novel fuzzy risk assessment score approach is presented to correctly evaluate and distinguish the levels of harm of the risks faced, such that these evaluations are consistent with preferences elicitation of the risk analyst. A real world risk analysis problem in fiber industry is then carried out to demonstrate the novelty, validity and feasibility of the proposed method.

Abstract In this paper we propose a novel method for learning how algorithms perform. Classically, algorithms are compared on a finite number of existing (or newly simulated) benchmark datasets based on some fixed metrics. The algorithm(s) with the smallest value of this metric are chosen to be the ‘best performing’. We offer a new approach to flip this paradigm. We instead aim to gain a richer picture of the performance of an algorithm by generating artificial data through genetic evolution, the purpose of which is to create populations of datasets for which a particular algorithm performs well on a given metric. These datasets can be studied so as to learn what attributes lead to a particular progression of a given algorithm. Following a detailed description of the algorithm as well as a brief description of an open source implementation, a case study in clustering is presented. This case study demonstrates the performance and nuances of the method which we call Evolutionary Dataset Optimisation. In this study, a number of known properties about preferable datasets for the clustering algorithms known as k-means and DBSCAN are realised in the generated datasets.