Optimizing the Feature Selection Process for Better Accuracy in Datasets with a Large Number of Features (Student Abstract)

With the rapid development of artificial intelligence in recent years, the research on image processing, text mining, and genome informatics has gradually deepened, and the mining of large-scale databases has begun to receive more and more attention. The objects of data mining have also become more complex, and the data dimensions of mining objects have become higher and higher. Compared with the ultra-high data dimensions, the number of samples available for analysis is too small, resulting in the production of high-dimensional small sample data. High-dimensional small sample data will bring serious dimensional disasters to the mining process. Through feature selection, redundancy and noise features in high-dimensional small sample data can be effectively eliminated, avoiding dimensional disasters and improving the actual efficiency of mining algorithms. However, the existing feature selection methods emphasize the classification or clustering performance of the feature selection results and ignore the stability of the feature selection results, which will lead to unstable feature selection results, and it is difficult to obtain real and understandable features. Based on the traditional feature selection method, this paper proposes an ensemble feature selection method, Random Bits Forest Recursive Clustering Eliminate (RBF-RCE) feature selection method, combined with multiple sets of basic classifiers to carry out parallel learning and screen out the best feature classification results, optimizes the classification performance of traditional feature selection methods, and can also improve the stability of feature selection. Then, this paper analyzes the reasons for the instability of feature selection and introduces a feature selection stability measurement method, the Intersection Measurement (IM), to evaluate whether the feature selection process is stable. The effectiveness of the proposed method is verified by experiments on several groups of high-dimensional small sample data sets.

Software defect prediction models are used for predicting high risk software components. Feature selection has significant impact on the prediction performance of the software defect prediction models since redundant and unimportant features make the prediction model more difficult to learn. Ensemble feature selection has recently emerged as a new methodology for enhancing feature selection performance. This paper proposes a new multi-criteria-decision-making (MCDM) based ensemble feature selection (EFS) method. This new method is termed as MCDM-EFS. The proposed method, MCDM-EFS, first generates the decision matrix signifying the feature’s importance score with respect to various existing feature selection methods. Next, the decision matrix is used as the input to well-known MCDM method TOPSIS for assigning a final rank to each feature. The proposed approach is validated by an experimental study for predicting software defects using two classifiers K-nearest neighbor (KNN) and naïve bayes (NB) over five open-source datasets. The predictive performance of the proposed approach is compared with existing feature selection algorithms. Two evaluation metrics – nMCC and G-measure are used to compare predictive performance. The experimental results show that the MCDM-EFS significantly improves the predictive performance of software defect prediction models against other feature selection methods in terms of nMCC as well as G-measure.

Cancer is a world-famous deadly disease. According to the World Health Organization (WHO), cancer is the second leading cause of death globally and is responsible for an estimated 9.6 million deaths in 2018. One well-known technique for cancer detection is the DNA microarray technique. DNA microarray technology provides an opportunity for researchers to analyze thousands of gene expression profiles at the same time to determine whether a person has cancer or not. However, one of the problems in DNA microarray data is the large number of features that require feature selection. In overcoming these problems, this study will use the feature selection Support Vector Machine-Recursive Feature Elimination (SVM-RFE) and Chi-Square and use the Naive Bayes classification method. The accuracy results from using feature selection with those that are not will be compared. The accuracy between using the two feature selection methods will also be compared to find which feature selection method is better when combined with the Naive Bayes classification method. To get an overall picture of the performance comparison, this study also considers precision, recall, and F1-score. The best accuracy results obtained were 100% lung cancer data with SVM-RFE and Chi-Square, 99.6% ovarian cancer with SVM-RFE, 93.7% breast cancer with SVM-RFE, and 90% colon cancer with SVM- RFE.

Computer-aided diagnosis has the potential of increasing diagnostic accuracy by providing a second reading to radiologists. In many computerized schemes, numerous features can be extracted to describe suspect image regions. A subset of these features is then employed in a data classifier to determine whether the suspect region is abnormal or normal. Different subsets of features will, in general, result in different classification performances. A feature selection method is often used to determine an "optimal" subset of features to use with a particular classifier. A classifier performance measure (such as the area under the receiver operating characteristic curve) must be incorporated into this feature selection process. With limited datasets, however, there is a distribution in the classifier performance measure for a given classifier and subset of features. In this paper, we investigate the variation in the selected subset of "optimal" features as compared with the true optimal subset of features caused by this distribution of classifier performance. We consider examples in which the probability that the optimal subset of features is selected can be analytically computed. We show the dependence of this probability on the dataset sample size, the total number of features from which to select, the number of features selected, and the performance of the true optimal subset. Once a subset of features has been selected, the parameters of the data classifier must be determined. We show that, with limited datasets and/or a large number of features from which to choose, bias is introduced if the classifier parameters are determined using the same data that were employed to select the "optimal" subset of features.

<p>Classification problems often have a large number of features, but not all of them are useful for classification. Irrelevant and redundant features may even reduce the classification accuracy. Feature selection is a process of selecting a subset of relevant features, which can decrease the dimensionality, shorten the running time, and/or improve the classification accuracy. There are two types of feature selection approaches, i.e. wrapper and filter approaches. Their main difference is that wrappers use a classification algorithm to evaluate the goodness of the features during the feature selection process while filters are independent of any classification algorithm. Feature selection is a difficult task because of feature interactions and the large search space. Existing feature selection methods suffer from different problems, such as stagnation in local optima and high computational cost. Evolutionary computation (EC) techniques are well-known global search algorithms. Particle swarm optimisation (PSO) is an EC technique that is computationally less expensive and can converge faster than other methods. PSO has been successfully applied to many areas, but its potential for feature selection has not been fully investigated. The overall goal of this thesis is to investigate and improve the capability of PSO for feature selection to select a smaller number of features and achieve similar or better classification performance than using all features. This thesis investigates the use of PSO for both wrapper and filter, and for both single objective and multi-objective feature selection, and also investigates the differences between wrappers and filters. This thesis proposes a new PSO based wrapper, single objective feature selection approach by developing new initialisation and updating mechanisms. The results show that by considering the number of features in the initialisation and updating procedures, the new algorithm can improve the classification performance, reduce the number of features and decrease computational time. This thesis develops the first PSO based wrapper multi-objective feature selection approach, which aims to maximise the classification accuracy and simultaneously minimise the number of features. The results show that the proposed multi-objective algorithm can obtain more and better feature subsets than single objective algorithms, and outperform other well-known EC based multi-objective feature selection algorithms. This thesis develops a filter, single objective feature selection approach based on PSO and information theory. Two measures are proposed to evaluate the relevance of the selected features based on each pair of features and a group of features, respectively. The results show that PSO and information based algorithms can successfully address feature selection tasks. The group based method achieves higher classification accuracies, but the pair based method is faster and selects smaller feature subsets. This thesis proposes the first PSO based multi-objective filter feature selection approach using information based measures. This work is also the first work using other two well-known multi-objective EC algorithms in filter feature selection, which are also used to compare the performance of the PSO based approach. The results show that the PSO based multiobjective filter approach can successfully address feature selection problems, outperform single objective filter algorithms and achieve better classification performance than other multi-objective algorithms. This thesis investigates the difference between wrapper and filter approaches in terms of the classification performance and computational time, and also examines the generality of wrappers. The results show that wrappers generally achieve better or similar classification performance than filters, but do not always need longer computational time than filters. The results also show that wrappers built with simple classification algorithms can be general to other classification algorithms.</p>

In many applications, like function approximation, pattern recognition, time series prediction, and data mining, one has to build a model relating some features describing the data to some response value. Often, the features that are relevant for building the model are not known in advance. Feature selection methods allow removing irrelevant and/or redundant features to only keep the feature subset that are most useful to build a prediction model. The model is simpler and easier to interpret, reducing the risks of overfitting, non-convergence, etc. By contrast with other dimensionality reduction techniques such as principal component analysis or more recent nonlinear projection techniques (Lee & Verleysen 2007), which build a new, smaller set of features, the features that are selected by feature selection methods preserve their initial meaning, potentially bringing extra information about the process being modeled (Guyon 2006). Recently, the advent of high-dimensional data has raised new challenges for feature selection methods, both from the algorithmic point of view and the conceptual point of view (Liu & Motoda 2007). The problem of feature selection is exponential in nature, and many approximate algorithms are cubic with respect to the initial number of features, which may be intractable when the dimensionality of the data is large. Furthermore, high-dimensional data are often highly redundant, and two distinct subsets of features may have very similar predictive power, which can make it difficult to identify the best subset.

We present an integrated algorithm for simultaneous feature selection (FS) and designing of diverse classifiers using a steady state multiobjective genetic programming (GP), which minimizes three objectives: 1) false positives (FPs); 2) false negatives (FNs); and 3) the number of leaf nodes in the tree. Our method divides a c -class problem into c binary classification problems. It evolves c sets of genetic programs to create c ensembles. During mutation operation, our method exploits the fitness as well as unfitness of features, which dynamically change with generations with a view to using a set of highly relevant features with low redundancy. The classifiers of i th class determine the net belongingness of an unknown data point to the i th class using a weighted voting scheme, which makes use of the FP and FN mistakes made on the training data. We test our method on eight microarray and 11 text data sets with diverse number of classes (from 2 to 44), large number of features (from 2000 to 49 151), and high feature-to-sample ratio (from 1.03 to 273.1). We compare our method with a bi-objective GP scheme that does not use any FS and rule size reduction strategy. It depicts the effectiveness of the proposed FS and rule size reduction schemes. Furthermore, we compare our method with four classification methods in conjunction with six features selection algorithms and full feature set. Our scheme performs the best for 380 out of 474 combinations of data sets, algorithm and FS method.

Stock market prediction is a very noisy problem and the use of any additional information to increase accuracy is necessary. In this paper, for the stock daily return prediction problem, the set of features is expanded to include indicators not only for the stock to be predicted itself but also a set of other stocks and currencies. Afterwards, different feature selection and classification methods are utilized for prediction. The daily close returns of the 3 most traded stocks (GARAN, THYAO, and ISCTR) in Borsa Istanbul (BIST) are predicted using indicators computed on those stocks, indicators for all the other stocks listed in the BIST100 index, and indicators on the dollar-gold prices. Twenty-five different indicators on daily stock prices are computed to form feature vectors for each trading day. These feature vectors are assigned class labels according to the daily close returns. Expanding the feature space with BIST100 stocks features results in a high dimensional feature space, with possibly noisy or irrelevant features. Therefore, feature selection methods are utilized to select the most informative features. In order to determine relevance scores of features, fast filter-based methods, gain ratio and relief, are used. Experiments are performed based on individual stock features, dollar-gold features (DG), BIST100 stock features (BIST100), and a combination of BIST100 and DG with and without feature selection. Using the gain ratio feature selection with a gradient boosting machine (GBM), the movements of GARAN stock were predicted with an accuracy of 0.599 and an F-measure of 0.614. For THYAO, the relief feature selection with the GBM gave an accuracy of 0.558, and for ISCTR, the gain ratio feature selection with logistic regression achieved an accuracy of 0.581. It was found that using BIST100 stock features boosts classification results for all stocks in terms of accuracy.

While the focus in Chap. 5 was on descriptors that were made up of homogeneous features, the focus in this chapter is on descriptors that are composed of features of different types. Heterogeneous descriptors are created from a combination of various types of features as described in Chap. 4. The features are selected in such a way that the combined feature set delivers the most information about an image or sub-image. Also discussed in this chapter are various feature-selection methods that choose the optimal or suboptimal feature set from a large number of features. Both filter and wrapper feature selection algorithms are discussed, including max-min, sequential forward selection, sequential backward selection, plus l take away r, and branch and bound algorithms.

Selecting relevant features is a common task in most OMICs data analysis, where the aim is to identify a small set of key features to be used as biomarkers. To this end, two alternative but equally valid methods are mainly available, namely the univariate (filter) or the multivariate (wrapper) approach. The stability of the selected lists of features is an often neglected but very important requirement. If the same features are selected in multiple independent iterations, they more likely are reliable biomarkers. In this study, we developed and evaluated the performance of a novel method for feature selection and prioritization, aiming at generating robust and stable sets of features with high predictive power. The proposed method uses the fuzzy logic for a first unbiased feature selection and a Random Forest built from conditional inference trees to prioritize the candidate discriminant features. Analyzing several multi-class gene expression microarray data sets, we demonstrate that our technique provides equal or better classification performance and a greater stability as compared to other Random Forest-based feature selection methods.

Feature selection is used for finding a feature subset that has the most discriminative information from the original feature set. Large number of features often includes many garbage features. We propose a novel feature selection method on the basis of the estimation of Bayes discrimination boundary. The experimental results on heart single proton emission computed tomography (SPECT) data shows the fundamental effectiveness of the proposed method compared to the conventional forward feature selection methods.

<p>Classification aims to identify a class label of an instance according to the information from its characteristics or features. Unfortunately, many classification problems have a large feature set containing irrelevant and redundant features, which reduce the classification performance. In order to address the above problem, feature selection is proposed to select a small subset of relevant features. There are three main types of feature selection methods, i.e. wrapper, embedded and filter approaches. Wrappers use a classification algorithm to evaluate candidate feature subsets. In embedded approaches, the selection process is embedded in the training process of a classification algorithm. Different from the other two approaches, filters do not involve any classification algorithm during the selection process. Feature selection is an important process but it is not an easy task due to its large search space and complex feature interactions. Because of the potential global search ability, Evolutionary Computation (EC), especially Particle Swarm Optimization (PSO), has been widely and successfully applied to feature selection. However, there is potential to improve the effectiveness and efficiency of EC-based feature selection. The overall goal of this thesis is to investigate and improve the capability of EC for feature selection to select small feature subsets while maintaining or even improving the classification performance compared to using all features. Different aspects of feature selection are considered in this thesis such as the number of objectives (single-objective/multi-objective), the fitness function (filter/wrapper), and the searching mechanism. This thesis introduces a new fitness function based on mutual information which is calculated by an estimation approach instead of the traditional counting approach. Results show that the estimation approach works well on both continuous and discrete data. More importantly, mutual information calculated by the estimation approach can capture feature interactions better than the traditional counting approach. This thesis develops a novel binary PSO algorithm, which is the first work to redefine some core concepts of PSO such as velocity and momentum to suit the characteristics of binary search spaces. Experimental results show that the proposed binary PSO algorithm evolve better solutions than other binary EC algorithms when the search spaces are large and complex. Specifically, on feature selection, the proposed binary PSO algorithm can select smaller feature subsets with similar or better classification accuracies, especially when there are a large number of features. This thesis proposes surrogate models for wrapper-based feature selection. The surrogate models use surrogate training sets which are subsets of informative instances selected from the training set. Experimental results show that the proposed surrogate models assist PSO to reduce the computational cost while maintaining or even improving the classification performance compared to using only the original training set. The thesis develops the first wrapper-based multi-objective feature selection algorithm using MOEA/D. A new decomposition strategy using multiple reference points for MOEA/D is designed, which can deal with different characteristics of multi-objective feature selection such as highly discontinuous Pareto fronts and complex relationships between objectives. The experimental results show that the proposed algorithm can evolve more diverse non-dominated sets than other multi-objective algorithms. This thesis introduces the first PSO-based feature selection algorithm for transfer learning. In the proposed algorithm, the fitness function uses classification performance to reduce the differences between domains while maintaining the discriminative ability on the target domain. The experimental results show that the proposed algorithm can select feature subsets which achieve better classification performance than four state-of-the-art feature-based transfer learning algorithms.</p>

Feature selection is widely used as the first stage of classification task to reduce the dimension of problem, decrease noise, improve speed and relieve memory constraints by the elimination of irrelevant or redundant features. One approach in the feature selection area is employing population-based optimization algorithms such as particle swarm optimization (PSO)-based method and ant colony optimization (ACO)-based method. Ant colony optimization algorithm is inspired by observation on real ants in their search for the shortest paths to food sources. Protein function prediction is an important problem in functional genomics. Typically, protein sequences are represented by feature vectors. A major problem of protein datasets that increase the complexity of classification models is their large number of features. This paper empowers the ant colony optimization algorithm by enabling the ACO to select features for a Bayesian classification method. The naive Bayesian classifier is a straightforward and frequently used method for supervised learning. It provides a flexible way for dealing with any number of features or classes, and is based on probability theory. This paper then compares the performance of the proposed ACO algorithm against the performance of a standard binary particle swarm optimization algorithm on the task of selecting features on Postsynaptic dataset. The criteria used for this comparison are maximizing predictive accuracy and finding the smallest subset of features. Simulation results on Postsynaptic dataset show that proposed method simplifies features effectively and obtains a higher classification accuracy compared to other feature selection methods.

A common situation in classification tasks is to deal with unbalanced datasets, an issue that appears when the majority class(es) has a large number of samples compared to the minority class(es). This problem is even more significant when the datasets have a large number of features but only a few samples, as is the case with microarray datasets. Traditionally, an approach to alleviate this problem has been the application of sampling methods to obtain more balanced classes, increasing the number of samples in the minority class (replicating samples or generating new synthetic samples), or decreasing the number of samples in the majority class. In this study, we have compared different balancing methods, including a novel method that applies sampling in both the minority and majority classes. The interest in applying feature selection in combination with balancing methods has also been explored. In view of the results, a recommendation of sampling method, feature selection, and classifier is proposed to improve the classification results according to the type of dataset.

Feature selection is a key step for activity classification applications. Feature selection selects the most relevant features and considers how to use each of the selected features in the most suitable format. This paper proposes an efficient feature selection method that organizes multiple subsets of features in a multilayer, rather than utilizing all selected features together as one large feature set. The proposed method was evaluated by 13 subjects (aged from 23 to 50) in a lab environment. The experimental results illustrate that the large number of features (3 vs. 7 features) are not associated with high classification accuracy using a single Support Vector Machine (SVM) model (61.3% vs. 44.7%). However, the accuracy was improved significantly (83.1% vs. 44.7%), when the selected 7 features were organized as 3 subsets and used to classify 10 postures (9 motionless with 1 motion) in 3 layers via a hierarchical algorithm, which combined a rule-based algorithm with 3 independent SVM models.