Feature selection in single and ensemble learning‐based bankruptcy prediction models

Feature selection is an important data preprocessing step for the construction of an effective bankruptcy prediction model. The prediction performance can be affected by the employed feature selection and classification techniques. However, there have been very few studies of bankruptcy prediction that identify the best combination of feature selection and classification techniques. In this study, two types of feature selection methods, including filter‐ and wrapper‐based methods, are considered, and two types of classification techniques, including statistical and machine learning techniques, are employed in the development of the prediction methods. In addition, bagging and boosting ensemble classifiers are also constructed for comparison. The experimental results based on three related datasets that contain different numbers of input features show that the genetic algorithm as the wrapper‐based feature selection method performs better than the filter‐based one by information gain. It is also shown that the lowest prediction error rates for the three datasets are provided by combining the genetic algorithm with the naïve Bayes and support vector machine classifiers without bagging and boosting.     

SBFS：基于搜索的软件缺陷预测特征选择框架

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,有助于合理分配测试资源,并最终提高被测软件产品的质量。但在搜集缺陷预测数据集的时候,由于考虑了大量与代码复杂度或开发过程相关的度量元,造成数据集内存在维数灾难问题。借助基于搜索的软件工程思想,提出一种新颖的基于搜索的包裹式特征选择框架SBFS。该框架在实现时,首先借助SMOTE方法来缓解数据集内存在的类不平衡问题,随后借助基于遗传算法的特征选择方法,基于训练集选出最优特征子集。在实证研究中,以NASA数据集作为评测对象,以基于前向选择策略的包裹式特征选择方法FW、基于后向选择策略的包裹式特征选择BW、不进行特征选择的Origin作为基准方法。最终实证研究结果表明：SBFS方法在90%的情况下,不差于Origin法。在82.3%的情况下,不差于BW法。在69.3%的情况下,不差于FW法。除此之外,我们发现若基于决策树分类器,则应用SMOTE方法后,可以在71%的情况下,提高模型性能。而基于朴素贝叶斯和Logistic回归分类器,则应用SMOTE方法后,仅可以在47%和43%的情况下,提高模型的预测性能。     

面向软件缺陷个数预测的混合式特征选择方法*

针对软件缺陷数据集中不相关特征和冗余特征会降低软件缺陷个数预测模型的性能的问题,提出了一种面向软件缺陷个数预测的混合式特征选择方法-HFSNFP。首先,利用ReliefF算法计算每个特征与缺陷个数之间的相关性,选出相关性最高的m个特征;然后,基于特征之间的关联性利用谱聚类对这m个特征进行聚类;最后,利用基于包裹式特征选择思想从每个簇中依次挑选最相关的特征形成最终的特征子集。实验结果表明,相比于已有的五种过滤式特征选择方法,HFSNFP方法在提高预测率的同时降低了误报率,且G-measure与RMSE度量值更佳;相比于已有的两种包裹式特征选择方法,HFSNFP方法在保证了缺陷个数预测性能的同时可以显著降低特征选择的时间。     

克隆代码有害性预测中的特征选择模型

为解决克隆代码有害性预测过程中特征无关与特征冗余的问题,提出一种基于相关程度和影响程度的克隆代码有害性特征选择组合模型。首先,利用信息增益率对特征数据进行相关性的初步排序;然后,保留相关性排名较高的特征并去除其他无关特征,减小特征的搜索空间;接着,采用基于朴素贝叶斯等六种分类器分别与封装型序列浮动前向选择算法结合来确定最优特征子集。最后对不同的特征选择方法进行对比分析,将各种方法在不同选择准则上的优势加以利用,对特征数据进行分析、筛选和优化。实验结果表明,与未进行特征选择之前对比发现有害性预测准确率提高15.2~34个百分点以上;与其他特征选择方法比较,该方法在F1测度上提高1.1~10.1个百分点,在AUC指标上提升达到0.7~22.1个百分点,能极大地提高有害性预测模型的准确度。     

Statistics-based wrapper for feature selection: An implementation on financial distress identification with support vector machine

Support vector machine (SVM) is an effective tool for financial distress identification (FDI). However, a potential issue that keeps SVM from being efficiently applied in identifying financial distress is how to select features in SVM-based FDI. Although filters are commonly employed, yet this type of approach does not consider predictive capability of SVM itself when selecting features. This research devotes to constructing a statistics-based wrapper for SVM-based FDI by using statistical indices of ranking-order information from predictive performances on various parameters. This wrapper consists of four levels, i.e., data level, model level based on SVM, feature ranking-order level, and the index level of feature selection. When data is ready, predictive accuracies of a type of SVM model, i.e., linear SVM (LSVM), polynomial SVM (PSVM), Gaussian SVM (GSVM), or sigmoid SVM (SSVM), on various pairs of parameters are firstly calculated. Then, performances of SVM models on each candidate feature are transferred to be ranking-order indices. After this step, the two statistical indices of mean and standard deviation values are calculated from ranking-order information on each feature. Finally, the feature selection indices of SVM are produced by a combination of statistical indices. Each feature with its feature selection index being smaller than half of the average index is selected to compose the optimal feature set. With a dataset collected for Chinese FDI prior to 3 years, we statistically verified the performance of this statistics-based wrapper against a non-statistics-based wrapper, two filters, and non-feature selection for SVM-based FDI. Results from unseen dataset indicate that GSVM with the statistics-based wrapper significantly outperformed the other SVM models on the other feature selection methods and two wrapper-based classical statistical models.     

Using support vector machine with a hybrid feature selection method to the stock trend prediction

In this paper, we developed a prediction model based on support vector machine (SVM) with a hybrid feature selection method to predict the trend of stock markets. This proposed hybrid feature selection method, named F-score and Supported Sequential Forward Search (F_SSFS), combines the advantages of filter methods and wrapper methods to select the optimal feature subset from original feature set. To evaluate the prediction accuracy of this SVM-based model combined with F_SSFS, we compare its performance with back-propagation neural network (BPNN) along with three commonly used feature selection methods including Information gain, Symmetrical uncertainty, and Correlation-based feature selection via paired t-test. The grid-search technique using 5-fold cross-validation is used to find out the best parameter value of kernel function of SVM. In this study, we show that SVM outperforms BPN to the problem of stock trend prediction. In addition, our experimental results show that the proposed SVM-based model combined with F_SSFS has the highest level of accuracies and generalization performance in comparison with the other three feature selection methods. With these results, we claim that SVM combined with F_SSFS can serve as a promising addition to the existing stock trend prediction methods.     

A novel ensemble-based wrapper method for feature selection using extreme learning machine and genetic algorithm

This paper presents a novel wrapper feature selection algorithm for classification problems, namely hybrid genetic algorithm (GA)- and extreme learning machine (ELM)-based feature selection algorithm (HGEFS). It utilizes GA to wrap ELM to search for the optimum subsets in the huge feature space, and then, a set of subsets are selected to make ensemble to improve the final prediction accuracy. To prevent GA from being trapped in the local optimum, we propose a novel and efficient mechanism specifically designed for feature selection problems to maintain GA’s diversity. To measure each subset’s quality fairly and efficiently, we adopt a modified ELM called error-minimized extreme learning machine (EM-ELM) which automatically determines an appropriate network architecture for each feature subsets. Moreover, EM-ELM has good generalization ability and extreme learning speed which allows us to perform wrapper feature selection processes in an affordable time. In other words, we simultaneously optimize feature subset and classifiers’ parameters. After finishing the search process of GA, to further promote the prediction accuracy and get a stable result, we select a set of EM-ELMs from the obtained population to make the final ensemble according to a specific ranking and selecting strategy. To verify the performance of HGEFS, empirical comparisons are carried out on different feature selection methods and HGEFS with benchmark datasets. The results reveal that HGEFS is a useful method for feature selection problems and always outperforms other algorithms in comparison.     

Constructing response model using ensemble based on feature subset selection

In building a response model, determining the inputs to the model has been an important issue because of the complexities of the marketing problem and limitations of mental models for decision-making. It is common that the customers' historical purchase data contains many irrelevant or redundant features thus result in bad model performance. Furthermore, single complex models based on feature subset selection may not always report improved performance largely because of overfitting and instability. Ensemble is a widely adopted mechanism for alleviating such problems. In this paper, we propose an ensemble creation method based on GA based wrapper feature subset selection mechanism. Through experimental studies on DMEF4 data set, we found that the proposed method has, at least, two distinct advantages over other models: first, the ability to account for the important inputs to the response model; second, improved prediction accuracy and stability.     

Ordinal-based and frequency-based integration of feature selection methods for sentiment analysis

Feature subset selection with the aim of reducing dependency of feature selection techniques and obtaining a high-quality minimal feature subset from a real-world domain is the main task of this research. For this end, firstly, two types of feature representation are presented for feature sets, namely unigram-based and part-of-speech based feature sets. Secondly, five methods of feature ranking are employed for creating feature vectors. Finally, we propose two methods for the integration feature vectors and feature subsets. An ordinal-based integration of different feature vectors (OIFV) is proposed in order to obtain a new feature vector. The new feature vector depends on the order of features in the old vectors. A frequency-based integration of different feature subsets (FIFS) with most effective features, which are obtained from a hybrid filter and wrapper methods in the feature selection task, is then proposed. In addition, four well-known text classification algorithms are employed as classifiers in the wrapper method for the selection of the feature subsets. A wide range of comparative experiments on five widely-used datasets in sentiment analysis were carried out. The experiments demonstrate that proposed methods can effectively improve the performance of sentiment classification. These results also show that proposed part-of-speech patterns are more effective in their classification accuracy compared to unigram-based features.     

Application of wrapper approach and composite classifier to the stock trend prediction

The research on the stock market prediction has been more popular in recent years. Numerous researchers tried to predict the immediate future stock prices or indices based on technical indices with various mathematical models and machine learning techniques such as artificial neural networks (ANN), support vector machines (SVM) and ARIMA models. Although some researches in the literature exhibit satisfactory prediction achievement when the average percentage error and root mean square error are used as the performance metrics, the prediction accuracy of whether stock market goes or down is seldom analyzed. This paper employs wrapper approach to select the optimal feature subset from original feature set composed of 23 technical indices and then uses voting scheme that combines different classification algorithms to predict the trend in Korea and Taiwan stock markets. Experimental result shows that wrapper approach can achieve better performance than the commonly used feature filters, such as χ²-Statistic, Information gain, ReliefF, Symmetrical uncertainty and CFS. Moreover, the proposed voting scheme outperforms single classifier such as SVM, kth nearest neighbor, back-propagation neural network, decision tree, and logistic regression.     

Hybrids of support vector machine wrapper and filter based framework for malware detection

Malware replicates itself and produces offspring with the same characteristics but different signatures by using code obfuscation techniques. Current generation Anti-Virus (AV) engines employ a signature-template type detection approach where malware can easily evade existing signatures in the database. This reduces the capability of current AV engines in detecting malware. In this paper we propose a hybrid framework for malware detection by using the hybrids of Support Vector Machines Wrapper, Maximum-Relevance–Minimum-Redundancy Filter heuristics where Application Program Interface (API) call statistics are used as a malware features. The novelty of our hybrid framework is that it injects the filter’s ranking score in the wrapper selection process and combines the properties of both wrapper and filters and API call statistics which can detect malware based on the nature of infectious actions instead of signature. To the best of our knowledge, this kind of hybrid approach has not been explored yet in the literature in the context of feature selection and malware detection. Knowledge about the intrinsic characteristics of malicious activities is determined by the API call statistics which is injected as a filter score into the wrapper’s backward elimination process in order to find the most significant APIs. While using the most significant APIs in the wrapper classification on both obfuscated and benign types malware datasets, the results show that the proposed hybrid framework clearly surpasses the existing models including the independent filters and wrappers using only a very compact set of significant APIs. The performances of the proposed and existing models have further been compared using binary logistic regression. Various goodness of fit comparison criteria such as Chi Square, Akaike’s Information Criterion (AIC) and Receiver Operating Characteristic Curve ROC are deployed to identify the best performing models. Experimental outcomes based on the above criteria also show that the proposed hybrid framework outperforms other existing models of signature types including independent wrapper and filter approaches to identify malware.     

Financial distress prediction using support vector machines: Ensemble vs. individual

Financial distress prediction (FDP) is of great importance to both inner and outside parts of companies. Though lots of literatures have given comprehensive analysis on single classifier FDP method, ensemble method for FDP just emerged in recent years and needs to be further studied. Support vector machine (SVM) shows promising performance in FDP when compared with other single classifier methods. The contribution of this paper is to propose a new FDP method based on SVM ensemble, whose candidate single classifiers are trained by SVM algorithms with different kernel functions on different feature subsets of one initial dataset. SVM kernels such as linear, polynomial, RBF and sigmoid, and the filter feature selection/extraction methods of stepwise multi discriminant analysis (MDA), stepwise logistic regression (logit), and principal component analysis (PCA) are applied. The algorithm for selecting SVM ensemble's base classifiers from candidate ones is designed by considering both individual performance and diversity analysis. Weighted majority voting based on base classifiers’ cross validation accuracy on training dataset is used as the combination mechanism. Experimental results indicate that SVM ensemble is significantly superior to individual SVM classifier when the number of base classifiers in SVM ensemble is properly set. Besides, it also shows that RBF SVM based on features selected by stepwise MDA is a good choice for FDP when individual SVM classifier is applied.     

基于量子进化算法的包装式特征选择方法

针对监督分类中的特征选择问题, 提出一种基于量子进化算法的包装式特征选择方法. 首先分析了现有子集评价方法存在过度偏好分类精度的缺点, 进而提出基于固定阈值和统计检验的两种子集评价方法. 然后改进了量子进化算法的进化策略, 即将整个进化过程分为两个阶段, 分别选用个体极值和全局极值作为种群的进化目标. 在此基础上, 按照包装式特征选择遵循的一般框架设计了特征选择算法. 最后, 通过15个UCI数据集分别验证了子集评价方法和进化策略的有效性, 以及新方法相较于其它6种特征选择方法的优越性. 结果表明, 新方法在80%以上的数据集上取得相似甚至更好的分类精度, 在86.67%的数据集上选择了特征个数更小的子集.     

融合异常检测与随机森林的微博转发行为预测方法

针对目前微博转发行为预测具有的特征选择任意性、准确率不高的问题,提出了融合异常检测与随机森林的微博转发行为预测方法。首先,提取用户基本特征、博文基本特征、博文内容主题特征,并基于相对熵计算用户活跃度、博文影响力;其次,通过结合过滤式与封装式特征选择方法筛选出关键特征组;最后,融合异常检测与随机森林算法,依据筛选后的关键特征组进行微博转发行为预测,并利用袋外数据误差估计设置随机森林中的决策树和特征数。在真实新浪微博数据集上与基于逻辑回归、决策树、朴素贝叶斯、随机森林等算法的微博转发行为预测方法进行实验对比,结果表明所提方法的预测准确率(90.5％) 高于基准方法中最优的随机森林方法的预测准确率,同时验证了特征筛选方法的有效性。     

Bankruptcy prediction using SVM models with a new approach to combine features selection and parameter optimisation

Due to the economic significance of bankruptcy prediction of companies for financial institutions, investors and governments, many quantitative methods have been used to develop effective prediction models. Support vector machine (SVM), a powerful classification method, has been used for this task; however, the performance of SVM is sensitive to model form, parameter setting and features selection. In this study, a new approach based on direct search and features ranking technology is proposed to optimise features selection and parameter setting for 1-norm and least-squares SVM models for bankruptcy prediction. This approach is also compared to the SVM models with parameter optimisation and features selection by the popular genetic algorithm technique. The experimental results on a data set with 2010 instances show that the proposed models are good alternatives for bankruptcy prediction.     

Fast wrapper feature subset selection in high-dimensional datasets by means of filter re-ranking

This paper deals with the problem of supervised wrapper-based feature subset selection in datasets with a very large number of attributes. Recently the literature has contained numerous references to the use of hybrid selection algorithms: based on a filter ranking, they perform an incremental wrapper selection over that ranking. Though working fine, these methods still have their problems: (1) depending on the complexity of the wrapper search method, the number of wrapper evaluations can still be too large; and (2) they rely on a univariate ranking that does not take into account interaction between the variables already included in the selected subset and the remaining ones.Here we propose a new approach whose main goal is to drastically reduce the number of wrapper evaluations while maintaining good performance (e.g. accuracy and size of the obtained subset). To do this we propose an algorithm that iteratively alternates between filter ranking construction and wrapper feature subset selection (FSS). Thus, the FSS only uses the first block of ranked attributes and the ranking method uses the current selected subset in order to build a new ranking where this knowledge is considered. The algorithm terminates when no new attribute is selected in the last call to the FSS algorithm. The main advantage of this approach is that only a few blocks of variables are analyzed, and so the number of wrapper evaluations decreases drastically.The proposed method is tested over eleven high-dimensional datasets (2400-46,000 variables) using different classifiers. The results show an impressive reduction in the number of wrapper evaluations without degrading the quality of the obtained subset.     

故障特征组合选择方法

特征选择方法主要包括过滤方法和绕封方法。为了利用过滤方法计算简单和绕封方法精度高的优点,提出一种组合过滤和绕封方法的特征选择新方法。该方法首先利用基于互信息准则的过滤方法得到满足一定精度要求的子集后,再采用绕封方法找到最后的优化特征子集。由于遗传算法在组合优化问题上的成功应用,对特征子集寻优采用了遗传算法。在数值仿真和轴承故障特征选择中,采用新方法在保证诊断精度的同时,可以节省大量选择时间。组合特征选择方法有较好的寻优特征子集的能力,能够节省选择时间,具有高效、高精度的双重优点。     

ORESP：基于有序回归的软件缺陷严重程度预测方法

为提高软件缺陷严重程度的预测性能,通过充分考虑软件缺陷严重程度标签间的次序性,提出一种基于有序回归的软件缺陷严重程度预测方法ORESP.该方法首先使用基于Spearman的特征选择方法来识别并移除数据集内的冗余特征,随后使用基于比例优势模型的神经网络来构建预测模型.通过与五种经典分类方法的比较,所提的ORESP方法在四种不同类型的度量下均可取得更高的预测性能,其中基于平均0-1误差(MZE)评测指标,预测模型性能最大可提升10.3％;基于平均绝对误差(MAE)评测指标,预测模型性能最大可提升12.3％.除此之外,发现使用基于Spearman的特征选择方法可以有效提升ORESP方法的预测性能.     

Wrapper–Filter Feature Selection Algorithm Using a Memetic Framework

This correspondence presents a novel hybrid wrapper and filter feature selection algorithm for a classification problem using a memetic framework. It incorporates a filter ranking method in the traditional genetic algorithm to improve classification performance and accelerate the search in identifying the core feature subsets. Particularly, the method adds or deletes a feature from a candidate feature subset based on the univariate feature ranking information. This empirical study on commonly used data sets from the University of California, Irvine repository and microarray data sets shows that the proposed method outperforms existing methods in terms of classification accuracy, number of selected features, and computational efficiency. Furthermore, we investigate several major issues of memetic algorithm (MA) to identify a good balance between local search and genetic search so as to maximize search quality and efficiency in the hybrid filter and wrapper MA     

Selection of Relevant Visual Feature Sets for Enhanced Depression Detection using Incremental Linear Discriminant Analysis

Presently, while automated depression diagnosis has made great progress, most of the recent works have focused on combining multiple modalities rather than strengthening a single one. In this research work, we present a unimodal framework for depression detection based on facial expressions and facial motion analysis. We investigate a wide set of visual features extracted from different facial regions. Due to high dimensionality of the obtained feature sets, identification of informative and discriminative features is a challenge. This paper suggests a hybrid dimensionality reduction approach which leverages the advantages of the filter and wrapper methods. First, we use a univariate filter method, Fisher Discriminant Ratio, to initially reduce the size of each feature set. Subsequently, we propose an Incremental Linear Discriminant Analysis (ILDA) approach to find an optimal combination of complementary and relevant feature sets. We compare the performance of the proposed ILDA with the batch-mode LDA and also the Composite Kernel based Support Vector Machine (CKSVM) method. The experiments conducted on the Distress Analysis Interview Corpus Wizard-of-Oz (DAIC-WOZ) dataset demonstrate that the best depression classification performance is obtained by using different feature extraction methods in combination rather than individually. ILDA generates better depression classification results in comparison to the CKSVM. Moreover, ILDA based wrapper feature selection incurs lower computational cost in comparison to the CKSVM and the batch-mode LDA methods. The proposed framework significantly improves the depression classification performance, with an F1 Score of 0.805, which is better than all the video based depression detection models suggested in literature, for the DAIC-WOZ dataset. Salient facial regions and well performing visual feature extraction methods are also identified.