Evaluation of the Impact of Initial Positions obtained by Clustering Algorithms on the Straight Line Segments Classifier

Abstract

Supervised learning is an important component of several applications, such as speech recognition, handwritten symbol recognition, data mining, among others. Supervised classification algorithms aim at producing a learning model from a labeled training set. Different methods and approaches have been proposed to overcome the two-class classification problem. Among the existing techniques in literature, the classifier based on Straight Line Segments (SLS Classifier) is worthy of note. This technique is based on distances between points and two sets of straight line segments, whose initial positions are obtained by applying the K-Means algorithm. Then, the gradient descent method finds its optimal positions that minimize the Mean Squared Error. This paper aims to study the impact of the initial positions on the classifier accuracy. For this purpose, we performed two experiments to demonstrate the stability of the classifier performance when the initial positions are not optimal (close to the samples): (i) random initial positions and; (ii) k-means positions displaced by adding Gaussian and uniform noises. In addition, we perform a comparison with positions obtained using different clustering algorithms. As expected, the results suggest that with an increased noise level, the classification rate decreases, however, such reduction was not significant as compared when using the random initial positions. It is worth mentioning that in most of the experiments, the classification rate of the SLS and the Bayes classifier are comparable.