一种基于自训练的众包标记噪声纠正算法

针对众包标记经过标记集成后仍然存在噪声的问题,提出了一种基于自训练的众包标记噪声纠正算法(Selftraining-based label noise correction, STLNC). STLNC整体分为3个阶段:第1阶段利用过滤器将带集成标记的众包数据集分为噪声集和干净集.第2阶段利用加权密度峰值聚类算法构建数据集中低密度实例指向高密度实例的空间结构关系.第3阶段首先根据发现的空间结构关系设计噪声实例选择策略;然后利用在干净集上训练的集成分类器对选择的噪声实例按照设计的实例纠正策略进行纠正,并将纠正后的实例加入到干净集,再重新训练集成分类器;重复实例选择与纠正过程直到噪声集中所有的实例被纠正;最后用最后一轮训练得到的集成分类器对所有实例进行纠正.在仿真标准数据集和真实众包数据集上的实验结果表明STLNC比其他5种最先进的噪声纠正算法在噪声比和模型质量两个度量指标上表现更优.     

基于Box-Cox转换的集成跨项目软件缺陷预测方法

软件缺陷预测通过预先识别出被测项目内的潜在缺陷程序模块,可以优化测试资源的分配并提高软件产品的质量。论文对跨项目缺陷预测问题展开了深入研究,在源项目实例选择时,考虑了三种不同的实例相似度计算方法,并发现这些方法的缺陷预测结果存在多样性,因此提出了一种基于Box-Cox转换的集成跨项目软件缺陷预测方法BCEL,具体来说,首先基于不同的实例相似度计算方法,从候选集中选出不同的训练集,随后针对这些数据集,进行针对性的Box-Cox转化,并借助特定分类方法构造出不同的基分类器,最后将这三个基分类器进行有效集成。基于实际项目的数据集,验证了BCEL方法的有效性,并深入分析了BCEL方法内的影响因素对缺陷预测性能的影响。     

Using neural network ensembles for bankruptcy prediction and credit scoring

Bankruptcy prediction and credit scoring have long been regarded as critical topics and have been studied extensively in the accounting and finance literature. Artificial intelligence and machine learning techniques have been used to solve these financial decision-making problems. The multilayer perceptron (MLP) network trained by the back-propagation learning algorithm is the mostly used technique for financial decision-making problems. In addition, it is usually superior to other traditional statistical models. Recent studies suggest combining multiple classifiers (or classifier ensembles) should be better than single classifiers. However, the performance of multiple classifiers in bankruptcy prediction and credit scoring is not fully understood. In this paper, we investigate the performance of a single classifier as the baseline classifier to compare with multiple classifiers and diversified multiple classifiers by using neural networks based on three datasets. By comparing with the single classifier as the benchmark in terms of average prediction accuracy, the multiple classifiers only perform better in one of the three datasets. The diversified multiple classifiers trained by not only different classifier parameters but also different sets of training data perform worse in all datasets. However, for the Type I and Type II errors, there is no exact winner. We suggest that it is better to consider these three classifier architectures to make the optimal financial decision.     

Benchmarking Least Squares Support Vector Machine Classifiers

In Support Vector Machines (SVMs), the solution of the classification problem is characterized by a (convex) quadratic programming (QP) problem. In a modified version of SVMs, called Least Squares SVM classifiers (LS-SVMs), a least squares cost function is proposed so as to obtain a linear set of equations in the dual space. While the SVM classifier has a large margin interpretation, the LS-SVM formulation is related in this paper to a ridge regression approach for classification with binary targets and to Fisher's linear discriminant analysis in the feature space. Multiclass categorization problems are represented by a set of binary classifiers using different output coding schemes. While regularization is used to control the effective number of parameters of the LS-SVM classifier, the sparseness property of SVMs is lost due to the choice of the 2-norm. Sparseness can be imposed in a second stage by gradually pruning the support value spectrum and optimizing the hyperparameters during the sparse approximation procedure. In this paper, twenty public domain benchmark datasets are used to evaluate the test set performance of LS-SVM classifiers with linear, polynomial and radial basis function (RBF) kernels. Both the SVM and LS-SVM classifier with RBF kernel in combination with standard cross-validation procedures for hyperparameter selection achieve comparable test set performances. These SVM and LS-SVM performances are consistently very good when compared to a variety of methods described in the literature including decision tree based algorithms, statistical algorithms and instance based learning methods. We show on ten UCI datasets that the LS-SVM sparse approximation procedure can be successfully applied.     

INFFC: An iterative class noise filter based on the fusion of classifiers with noise sensitivity control

In classification, noise may deteriorate the system performance and increase the complexity of the models built. In order to mitigate its consequences, several approaches have been proposed in the literature. Among them, noise filtering, which removes noisy examples from the training data, is one of the most used techniques. This paper proposes a new noise filtering method that combines several filtering strategies in order to increase the accuracy of the classification algorithms used after the filtering process. The filtering is based on the fusion of the predictions of several classifiers used to detect the presence of noise. We translate the idea behind multiple classifier systems, where the information gathered from different models is combined, to noise filtering. In this way, we consider the combination of classifiers instead of using only one to detect noise. Additionally, the proposed method follows an iterative noise filtering scheme that allows us to avoid the usage of detected noisy examples in each new iteration of the filtering process. Finally, we introduce a noisy score to control the filtering sensitivity, in such a way that the amount of noisy examples removed in each iteration can be adapted to the necessities of the practitioner. The first two strategies (use of multiple classifiers and iterative filtering) are used to improve the filtering accuracy, whereas the last one (the noisy score) controls the level of conservation of the filter removing potentially noisy examples. The validity of the proposed method is studied in an exhaustive experimental study. We compare the new filtering method against several state-of-the-art methods to deal with datasets with class noise and study their efficacy in three classifiers with different sensitivity to noise.     

Democratic instance selection: A linear complexity instance selection algorithm based on classifier ensemble concepts

Instance selection is becoming increasingly relevant due to the huge amount of data that is constantly being produced in many fields of research. Although current algorithms are useful for fairly large datasets, scaling problems are found when the number of instances is in the hundreds of thousands or millions. When we face huge problems, scalability becomes an issue, and most algorithms are not applicable.Thus, paradoxically, instance selection algorithms are for the most part impracticable for the same problems that would benefit most from their use. This paper presents a way of avoiding this difficulty using several rounds of instance selection on subsets of the original dataset. These rounds are combined using a voting scheme to allow good performance in terms of testing error and storage reduction, while the execution time of the process is significantly reduced. The method is particularly efficient when we use instance selection algorithms that are high in computational cost. The proposed approach shares the philosophy underlying the construction of ensembles of classifiers. In an ensemble, several weak learners are combined to form a strong classifier; in our method several weak (in the sense that they are applied to subsets of the data) instance selection algorithms are combined to produce a strong and fast instance selection method.An extensive comparison of 30 medium and large datasets from the UCI Machine Learning Repository using 3 different classifiers shows the usefulness of our method. Additionally, the method is applied to 5 huge datasets (from three hundred thousand to more than a million instances) with good results and fast execution time.     

A multi-objective optimisation approach for class imbalance learning

Class imbalance limits the performance of most learning algorithms since they cannot cope with large differences between the number of samples in each class, resulting in a low predictive accuracy over the minority class. In this respect, several papers proposed algorithms aiming at achieving more balanced performance. However, balancing the recognition accuracies for each class very often harms the global accuracy. Indeed, in these cases the accuracy over the minority class increases while the accuracy over the majority one decreases. This paper proposes an approach to overcome this limitation: for each classification act, it chooses between the output of a classifier trained on the original skewed distribution and the output of a classifier trained according to a learning method addressing the course of imbalanced data. This choice is driven by a parameter whose value maximizes, on a validation set, two objective functions, i.e. the global accuracy and the accuracies for each class. A series of experiments on ten public datasets with different proportions between the majority and minority classes show that the proposed approach provides more balanced recognition accuracies than classifiers trained according to traditional learning methods for imbalanced data as well as larger global accuracy than classifiers trained on the original skewed distribution.     

Building bagging on critical instances

The ensemble method is a powerful data mining paradigm, which builds a classification model by integrating multiple diversified component learners. Bagging is one of the most successful ensemble methods. It is made of bootstrap-inspired classifiers and uses these classifiers to get an aggregated classifier. However, in bagging, bootstrapped training sets become more and more similar as redundancy is increasing. Besides redundancy, any training set is usually subject to noise. Moreover, the training set might be imbalanced. Thus, each training instance has a different impact on the learning process. This paper explores some properties of the ensemble margin and its use in improving the performance of bagging. We introduce a new approach to measure the importance of training data in learning, based on the margin theory. Then, a new bagging method concentrating on critical instances is proposed. This method is more accurate than bagging and more robust than boosting. Compared to bagging, it reduces the bias while generally keeping the same variance. Our findings suggest that (a) examples with low margins tend to be more critical for the classifier performance; (b) examples with higher margins tend to be more redundant; (c) misclassified examples with high margins tend to be noisy examples. Our experimental results on 15 various data sets show that the generalization error of bagging can be reduced up to 2.5% and its resilience to noise strengthened by iteratively removing both typical and noisy training instances, reducing the training set size by up to 75%.     

Belief Combination of Classifiers for Incomplete Data

Data with missing values,or incomplete information,brings some challenges to the development of classification,as the incompleteness may significantly affect the performance of classifiers.In this paper,we handle missing values in both training and test sets with uncertainty and imprecision reasoning by proposing a new belief combination of classifier(BCC)method based on the evidence theory.The proposed BCC method aims to improve the classification performance of incomplete data by characterizing the uncertainty and imprecision brought by incompleteness.In BCC,different attributes are regarded as independent sources,and the collection of each attribute is considered as a subset.Then,multiple classifiers are trained with each subset independently and allow each observed attribute to provide a sub-classification result for the query pattern.Finally,these sub-classification results with different weights(discounting factors)are used to provide supplementary information to jointly determine the final classes of query patterns.The weights consist of two aspects:global and local.The global weight calculated by an optimization function is employed to represent the reliability of each classifier,and the local weight obtained by mining attribute distribution characteristics is used to quantify the importance of observed attributes to the pattern classification.Abundant comparative experiments including seven methods on twelve datasets are executed,demonstrating the out-performance of BCC over all baseline methods in terms of accuracy,precision,recall,F1 measure,with pertinent computational costs.     

Hand gesture recognition using two-level speed normalization,feature selection and classifier fusion

Hand gesture recognition provides an alternative way to many devices for human computer interaction. In this work, we have developed a classifier fusion based dynamic free-air hand gesture recognition system to identify the isolated gestures. Different users gesticulate at different speed for the same gesture. Hence, when comparing different samples of the same gesture, variations due to difference in gesturing speed should not contribute to the dissimilarity score. Thus, we have introduced a two-level speed normalization procedure using DTW and Euclidean distance-based techniques. Three features such as ‘orientation between consecutive points’, ‘speed’ and ‘orientation between first and every trajectory points’ were used for the speed normalization. Moreover, in feature extraction stage, 44 features were selected from the existing literatures. Use of total feature set could lead to overfitting, information redundancy and may increase the computational complexity due to higher dimension. Thus, we have tried to overcome this difficulty by selecting optimal set of features using analysis of variance and incremental feature selection techniques. The performance of the system was evaluated using this optimal set of features for different individual classifiers such as ANN, SVM, k-NN and Naïve Bayes. Finally, the decisions of the individual classifiers were combined using classifier fusion model. Based on the experimental results it may be concluded that classifier fusion provides satisfactory results compared to other individual classifiers. An accuracy of 94.78 % was achieved using the classifier fusion technique as compared to baseline CRF (85.07 %) and HCRF (89.91 %) models.     

DF-SVM: a decision forest constructed on artificially enlarged feature space by support vector machine

Enlarging the feature space of the base tree classifiers in a decision forest by means of informative features extracted from an additional predictive model is advantageous for classification tasks. In this paper, we have empirically examined the performance of this type of decision forest with three different base tree classifier models including; (1) the full decision tree, (2) eight-node decision tree and (3) two-node decision tree (or decision stump). The hybrid decision forest with these base classifiers are trained in nine different sized resampled training sets. We have examined the performance of all these ensembles from different point of views; we have studied the bias-variance decomposition of the misclassification error of the ensembles, then we have investigated the amount of dependence and degree of uncertainty among the base classifiers of these ensembles using information theoretic measures. The experiment was designed to find out: (1) optimal training set size for each base classifier and (2) which base classifier is optimal for this kind of decision forest. In the final comparison, we have checked whether the subsampled version of the decision forest outperform the bootstrapped version. All the experiments have been conducted with 20 benchmark datasets from UCI machine learning repository. The overall results clearly point out that with careful selection of the base classifier and training sample size, the hybrid decision forest can be an efficient tool for real world classification tasks.     

A novel hybrid credit scoring model based on ensemble feature selection and multilayer ensemble classification

Credit scoring focuses on the development of empirical models to support the financial decision‐making processes of financial institutions and credit industries. It makes use of applicants' historical data and statistical or machine learning techniques to assess the risk associated with an applicant. However, the historical data may consist of redundant and noisy features that affect the performance of credit scoring models. The main focus of this paper is to develop a hybrid model, combining feature selection and a multilayer ensemble classifier framework, to improve the predictive performance of credit scoring. The proposed hybrid credit scoring model is modeled in three phases. The initial phase constitutes preprocessing and assigns ranks and weights to classifiers. In the next phase, the ensemble feature selection approach is applied to the preprocessed dataset. Finally, in the last phase, the dataset with the selected features is used in a multilayer ensemble classifier framework. In addition, a classifier placement algorithm based on the Choquet integral value is designed, as the classifier placement affects the predictive performance of the ensemble framework. The proposed hybrid credit scoring model is validated on real‐world credit scoring datasets, namely, Australian, Japanese, German‐categorical, and German‐numerical datasets.     

Classification in the presence of class noise using a probabilistic Kernel Fisher method

In machine learning, class noise occurs frequently and deteriorates the classifier derived from the noisy data set. This paper presents two promising classifiers for this problem based on a probabilistic model proposed by Lawrence and Schölkopf (2001). The proposed algorithms are able to tolerate class noise, and extend the earlier work of Lawrence and Schölkopf in two ways. First, we present a novel incorporation of their probabilistic noise model in the Kernel Fisher discriminant; second, the distribution assumption previously made is relaxed in our work. The methods were investigated on simulated noisy data sets and a real world comparative genomic hybridization (CGH) data set. The results show that the proposed approaches substantially improve standard classifiers in noisy data sets, and achieve larger performance gain in non-Gaussian data sets and small size data sets.     

基于集成一致性的多源跨领域情感分类模型

现有的跨领域情感分类方法大多只利用了单个源领域到目标域的迁移特征,没有充分考虑目标域实例与不同源域之间的联系。针对此问题,本文提出一种无监督的多源跨领域情感分类模型。首先利用单个源域到目标域的迁移特征训练基分类器,并对不同的基分类器加权;然后将不同基分类器对目标域实例预测的集成一致性作为目标函数,优化该目标函数,得到不同基分类器的权重;最后利用加权后的基分类器得到目标域的情感分类结果。该模型在Amazon数据集上进行了多源域情感迁移实验,取得了较好的实验结果,相对其他基线模型,在4组实验中平均提升了0.75%。     

A sentiment classification model based on multiple classifiers

With the widespread usage of social networks, forums and blogs, customer reviews emerged as a critical factor for the customers’ purchase decisions. Since the beginning of 2000s, researchers started to focus on these reviews to automatically categorize them into polarity levels such as positive, negative, and neutral. This research problem is known as sentiment classification. The objective of this study is to investigate the potential benefit of multiple classifier systems concept on Turkish sentiment classification problem and propose a novel classification technique. Vote algorithm has been used in conjunction with three classifiers, namely Naive Bayes, Support Vector Machine (SVM), and Bagging. Parameters of the SVM have been optimized when it was used as an individual classifier. Experimental results showed that multiple classifier systems increase the performance of individual classifiers on Turkish sentiment classification datasets and meta classifiers contribute to the power of these multiple classifier systems. The proposed approach achieved better performance than Naive Bayes, which was reported the best individual classifier for these datasets, and Support Vector Machines. Multiple classifier systems (MCS) is a good approach for sentiment classification, and parameter optimization of individual classifiers must be taken into account while developing MCS-based prediction systems.     

Improvements in image categorization using codebook ensembles

The problem of object category classification by committees or ensembles of classifiers, each of which is based on one diverse codebook, is addressed in this paper. Two methods of constructing visual codebook ensembles are proposed in this study. The first technique introduces diverse individual visual codebooks using different clustering algorithms. The second uses various visual codebooks of different sizes for constructing an ensemble with high diversity. Codebook ensembles are trained to capture and convey image properties from different aspects. Based on these codebook ensembles, different types of image representations can be acquired. A classifier ensemble can be trained based on different expression datasets from the same training image set. The use of a classifier ensemble to categorize new images can lead to improved performance. Detailed experimental analysis on a Pascal VOC challenge dataset reveals that the present ensemble approach performs well, consistently improves the performance of visual object classifiers, and results in state-of-the-art performance in categorization.     

Bayesian instance selection for the nearest neighbor rule

The nearest neighbors rules are commonly used in pattern recognition and statistics. The performance of these methods relies on three crucial choices: a distance metric, a set of prototypes and a classification scheme. In this paper, we focus on the second, challenging issue: instance selection. We apply a maximum a posteriori criterion to the evaluation of sets of instances and we propose a new optimization algorithm. This gives birth to Eva, a new instance selection method. We benchmark this method on real datasets and perform a multi-criteria analysis: we evaluate the compression rate, the predictive accuracy, the reliability and the computational time. We also carry out experiments on synthetic datasets in order to discriminate the respective contributions of the criterion and the algorithm, and to illustrate the advantages of Eva over the state-of-the-art algorithms. The study shows that Eva outputs smaller and more reliable sets of instances, in a competitive time, while preserving the predictive accuracy of the related classifier.     

Recognition of human actions using CNN-GWO: a novel modeling of CNN for enhancement of classification performance

Recognizing human actions from unconstrained videos turns to be a major challenging task in computer visualization approaches due to decreased accuracy in the feature classification performance. Therefore to improve the classification performance it is essential to minimize the ‘classification’ errors. Here, in this work, we propose a hybrid CNN-GWO approach for the recognition of human actions from the unconstrained videos. The weight initializations for the proposed deep Convolutional Neural Network (CNN) classifiers highly depend on the generated solutions of GWO (Grey Wolf Optimization) algorithm, which in turn minimizes the ‘classification’ errors. The action bank and local spatio-temporal features are generated for a video and fed into the ‘CNN’ classifiers. The ‘CNN’ classifiers are trained by a gradient descent algorithm to detect a ‘local minimum’ during the fitness computation of GWO ‘search agents’. The GWO algorithms ‘global search’ capability as well as the gradient descent algorithms ‘local search’ capabilities are subjected for the identification of a solution which is nearer to the global optimum. Finally, the classification performance can be further enhanced by fusing the classifiers evidences produced by the GWO algorithm. The proposed classification frameworks efficiency for the recognition of human actions is evaluated with the help of four achievable action recognition datasets namely HMDB51, UCF50, Olympic Sports and Virat Release 2.0. The experimental validation of our proposed approach shows better achievable results on the recognition of human actions with 99.9% recognition accuracy.     

SVOIS: Support Vector Oriented Instance Selection for text classification

Automatic text classification is usually based on models constructed through learning from training examples. However, as the size of text document repositories grows rapidly, the storage requirements and computational cost of model learning is becoming ever higher. Instance selection is one solution to overcoming this limitation. The aim is to reduce the amount of data by filtering out noisy data from a given training dataset. A number of instance selection algorithms have been proposed in the literature, such as ENN, IB3, ICF, and DROP3. However, all of these methods have been developed for the k-nearest neighbor (k-NN) classifier. In addition, their performance has not been examined over the text classification domain where the dimensionality of the dataset is usually very high. The support vector machines (SVM) are core text classification techniques. In this study, a novel instance selection method, called Support Vector Oriented Instance Selection (SVOIS), is proposed. First of all, a regression plane in the original feature space is identified by utilizing a threshold distance between the given training instances and their class centers. Then, another threshold distance, between the identified data (forming the regression plane) and the regression plane, is used to decide on the support vectors for the selected instances. The experimental results based on the TechTC-100 dataset show the superior performance of SVOIS over other state-of-the-art algorithms. In particular, using SVOIS to select text documents allows the k-NN and SVM classifiers perform better than without instance selection.     

A new learning schema based on support vector for multi-classification

A novel learning schema SVCMR based on support vector is proposed in this paper to address M-class classification issue. It creates a tree-shaped decision frame where M/2 nodes are constructed with the three-separation model as the basic classifier. A class selection rule is defined to ensure basic classifiers be trained in turn on pair of classes with maximum feature distance. Class contours are extracted as data representatives to reduce training set size. Another point is that parameters involved in SVCMR are learned from data neighborhood, which brings adaptation to various datasets and avoids pricy cost spent on searching parameter spaces. Experiments on real datasets demonstrate the performance of SVCMR can be competitive to those state-of-the-art classifiers but with the higher effectiveness than them.