Chandan Gautam

Data imputation via evolutionary computation, clustering and a neural network

In this paper, two novel hybrid imputation methods involving particle swarm optimization (PSO), evolving clustering method (ECM) and autoassociative extreme learning machine (AAELM) in tandem are proposed, which also preserve the covariance structure of the data. Further, we removed the randomness of AAELM by invoking ECM between input and hidden layers. Moreover, we selected the optimal value of Dthr using PSO, which simultaneously minimizes two error functions viz., (i) mean squared error between the covariance matrix of the set of complete records and that of the set of total records, including imputed ones and (ii) absolute difference between the determinants of the two covariance matrices. The proposed methods outperformed many existing imputation methods in majority of the datasets. Finally, we also performed a statistical significance testing to ensure the credibility of our obtained results. Superior performance of one of the hybrids is attributed to the power of hybrid of local learning, global optimization and global learning. Both methods resolved a nagging issue of the difficult choice of Dthr value and its dominant influence on the results in ECM based imputation. We conclude that the proposed models can be used as a viable alternative to the existing ones for the data imputation.

Evolving clustering based data imputation

Missing data is an inevitable problem in many disciplines. In this paper, we employed an Evolving Clustering Method (ECM) based imputation method and performed sensitivity analysis of the influence of threshold value (Dthr) on imputation results over 12 datasets. We experimented on a large range of Dthr values from 0.001 to 0.999, in steps of 0.001, in order to see which value of Dthr would perform better imputation compared to K-Means+MLP. Thereby, we provided an upper bound for the Dthr value in ECM algorithm. Further, we tested the effectiveness of the online clustering based imputation method on 12 datasets under 10-fold cross validation set up. ECM yielded better performance compared to K-Means + Multilayer perceptron hybrid algorithm, appearing in literature. It is due to strong local learning capability of ECM and selection of an optimal Dthr value.

Counter propagation auto-associative neural network based data imputation

In this paper, we propose two novel methods viz., counterpropagation auto-associative neural network (CPAANN) and grey system theory (GST) hybridised with CPAANN for data imputation. The effectiveness of these methods is demonstrated on 12 datasets and the results are compared with that of various extant methods. Wilcoxon signed rank test conducted at 1% level of significance, indicated that the proposed methods are statistically significant against all methods. The spectacular success of CPAANN can be attributed to the local learning, global approximation and auto-association that take place in tandem in a single architecture. Furthermore, significantly CPAANN turned out to be the best in the class of AANN architectures used for imputation. The reason could be the competitive learning that is intrinsic to the CPAANN architecture, but conspicuously absent in other auto-associative neural network architectures.

On the Construction of Extreme Learning Machine for One Class Classifier

One Class Classification (OCC) has been prime concern for researchers and effectively employed in various disciplines for outlier or novelty detection. But, traditional methods based one class classifier is very time consuming due to its iterative process for various parameters tuning. This paper presents four novel different OCC methods with their ten variants based on extreme Learning Machine (ELM). As we know, threshold decision is a crucial factor in case of OCC, so, three different threshold declining criteria have been employed so far. Our proposed classifiers mainly lie in two categories i.e. out of four proposed one class classifiers, two classifiers belong to reconstruction based and two belong to boundary based. In four proposed methods, two methods perform random feature mapping and two methods perform kernel feature mapping. These methods are tested on three benchmark datasets and exhibit better performance compared to eleven traditional one class classifiers.

On the construction of extreme learning machine for online and offline one-class classification—An expanded toolbox

Abstract One-class classification (OCC) has been prime concern for researchers and effectively employed in various disciplines. But, traditional methods based one-class classifiers are very time consuming due to its iterative process and various parameters tuning. In this paper, we present six OCC methods and their thirteen variants based on extreme learning machine (ELM) and online sequential ELM (OSELM). Our proposed classifiers mainly lie in two categories: reconstruction based and boundary based, where three proposed classifiers belong to reconstruction based and three belong to boundary based. We are presenting both types of learning viz., online and offline learning for OCC. Out of six methods, four are offline and remaining two are online methods. Out of four offline methods, two methods perform random feature mapping and two methods perform kernel feature mapping. We present a comprehensive discussion on these methods and their comparison to each other. Kernel feature mapping based approaches have been tested with RBF kernel and online version of one-class classifiers is tested with both types of nodes viz., additive and RBF. It is well known fact that threshold decision is a crucial factor in case of OCC, so, three different threshold deciding criteria have been employed so far and analyze the effectiveness of one threshold deciding criteria over another. Further, these methods are tested on two artificial datasets to check their boundary construction capability and on eight benchmark datasets from different discipline to evaluate the performance of the classifiers. Our proposed classifiers exhibit better performance compared to ten traditional one-class classifiers and ELM based two one-class classifiers. Through proposed one-class classifiers, we intend to expand the functionality of the most used toolbox for OCC i.e. DD toolbox. All of our methods are totally compatible with all the present features of the toolbox.

Construction of multi-class classifiers by Extreme Learning Machine based one-class classifiers

Construction of multi-class classifiers using homogeneous combination of Extreme Learning Machine (ELM) based one-class classifiers have been proposed in this paper. Each class has been trained using individual one-class classifier and any new sample will belong to that class, which will yield maximum value. Proposed methods can be used to detect unknown outliers using multi-class classifiers. Two recently proposed one-class classifiers viz., kernel and random feature mapping based one-class ELM, is extended for multi-class construction in this paper. Further, we construct one-class classifier based multi-class classifier in two ways: with rejection and without rejection of few samples during training. We also perform consistency based model selection for optimal parameters selection in one-class classifier. We have tested the generalization capability of the proposed classifiers on 6 synthetic datasets and two benchmark datasets.

A Fast Adaptive Classification Approach Using Kernel Ridge Regression and Clustering for Non-stationary Data Stream

Classification on non-stationary data requires faster evolving of the model while keeping the accuracy levels consistent. We present here a faster and reliable model to handle non-stationary data when a small number of labelled samples are available with the stream of unlabelled samples. An active learning model is proposed with the help of supervised model, i.e. Kernel Ridge Regression (KRR) with the combination of an unsupervised model, i.e. K-means clustering to handle the concept drift in the data efficiently. Proposed model consumes less time and at the same time yields similar or better accuracy compared to the existing clustering-based active learning methods.

Keystroke user recognition through extreme learning machine and evolving cluster method

User Identification and User Verification are the primary problems in the area of Keystroke Dynamics. In the last decade there has been massive research in User Verification, and lesser research in User Identification. Both approaches take a username and a passphrase as input. In this paper, we introduce this problem of replacing authentication systems with the passphrase alone. This is done by using neural network based approach i.e. Extreme Learning Machine. ELM is a fast Single hidden layer feed forward network (SLFN) with good generalization performance. However the hidden layer in ELM does not have to be tuned. As an evolutionary step, we use a clustering based Semi-supervised approach (ECM-ELM) to User Recognition to combat variance in the accuracy of traditional ELMs. This research aims not only to address User Recognition problem but also to remove the instability in the accuracy of ELM. As per our simulation, ECM-ELM achieved a stable accuracy of 87% with the CMU Keystroke Dataset, while ELM achieved an unstable average accuracy of 90%.