Outliers In Training Data Set Research Articles

A robust multilayer extreme learning machine using kernel risk-sensitive loss criterion

More recently, extreme learning machine (ELM) has emerged as a novel computing paradigm that enables the neural network (NN) based learning to be achieved with fast training speed and good generalization performance. However, the single hidden layer NN using ELM may be not effective in addressing some large-scale problems with more computational efforts. To avoid such limitation, we utilize the multilayer ELM architecture in this article to reduce the computational complexity, without the physical memory limitation. Meanwhile, it is known to us all that there are a lot of noises in the practical applications, and the traditional ELM may not perform well in this instance. Considering the existence of noises or outliers in training dataset, we develop a more practical approach by incorporating the kernel risk-sensitive loss (KRSL) criterion into ELM, on the basis of the efficient performance surface of KRSL with high accuracy while still maintaining the robustness to outliers. A robust multilayer ELM, i.e., the stacked ELM using the minimum KRSL criterion (SELM-MKRSL), is accordingly proposed in this article to enhance the outlier robustness on large-scale and complicated dataset. The simulation results on some synthetic datasets indicate that the proposed approach SELM-MKRSL can achieve higher classification accuracy and is more robust to the noises compared with other state-of-the-art algorithms related to multilayer ELM.

Power transformer fault diagnosis under measurement originated uncertainties

This paper addresses the problem of diagnosing the fault symptoms of power transformers with measurement originated uncertainties, which arise from the imprecision of samples (i.e. due to noises and outliers) and the effect of class imbalance (i.e. samples are unequally distributed between different fault types) in a training dataset used to identify different fault types. Two fuzzy support vector machine (FSVM) algorithms namely fuzzy c-means clustering-based FSVM (FCM-FSVM) and kernel fuzzy c-means clustering-based FSVM (KFCM-FSVM) have been applied in this paper to deal with any noises and outliers in training dataset. In order to reduce the effect of class imbalance in training dataset, two approaches including between-class weighting and random oversampling have been adopted and integrated with FCM-FSVM and KFCM-FSVM. The case studies show that KFCM-FSVM algorithm and its variants have consistent tendency to attain satisfied classification accuracy in transformer fault diagnosis using dissolved gas analysis (DGA) measurements.

Robust location algorithm based on weighted least-squares support vector machine (WLS-SVM) for non-line-of-sight environments

One of the main problems facing accurate location in wireless communication systems is non-line-of-sight (NLOS) propagation. Though learning location methods perform well in NLOS environments, learning location methods may be improved further since these methods do not consider outliers in the training data set. In this paper, we extend weighted least squares support vector machine (WLS-SVM) algorithm to mobile location problem. The proposed method can effectively suppress outliers with different weights. In simulation, we analyze the effects of the number of training points, the percentage of outliers in training data set, the standard deviation and mean of outliers, and the standard deviation of measurement error. Simulation results show that the proposed algorithm clearly outperforms three other algorithms (LS method, kernel method and LS-SVM based method). Key words: Weighted least-squares support vector machine (WLS-SVM), non-line-of-sight (NLOS).

Genetic algorithm based neural network approaches for predicting churn in cellular wireless network services

Marketing research suggests that it is more expensive to recruit a new customer than to retain an existing customer. In order to retain existing customers, academics and practitioners have developed churn prediction models to effectively manage customer churn. In this paper, we propose two genetic-algorithm (GA) based neural network (NN) models to predict customer churn in subscription of wireless services. Our first GA based NN model uses a cross entropy based criterion to predict customer churn, and our second GA based NN model attempts to directly maximize the prediction accuracy of customer churn. Using real-world cellular wireless services dataset and three different sizes of NNs, we compare the two GA based NN models with a statistical z-score model using several model evaluation criteria, which include prediction accuracy, top 10% decile lift and area under receiver operating characteristics (ROC) curve. The results of our experiments indicate that both GA based NN models outperform the statistical z-score model on all performance criteria. Further, we observe that medium sized NNs perform best and the cross entropy based criterion may be more resistant to overfitting outliers in training dataset.