Enhancing Classification Accuracy Through Feature Extraction: a Comparative Study of Discretization and Clustering Approaches on Sensor-Based Datasets

Abstract

Accuracy in a classification problem is directly related to the ability of features to adequately represent the differences between classes. In sensor-based datasets, measurements taken from the sensor form feature vectors. Measuring a given physical signal with different sensors enables it to be expressed with various feature vectors. For this reason, using sensor fusion is preferred in data acquisition. However, each new sensor added to the system brings problems such as complex sensory and supply circuit structures, extra energy consumption, signal sampling complexity, and time-consumption. On the other hand, in cases where sensor fusion cannot be applied, the ability of data from one sensor to represent classes may be insufficient. To avoid these problems, discretization and clustering approaches are suitable to derive more features from fewer sensors. The aim is to improve the accuracy of classifiers by deriving new feature vectors that can represent sensor data. This research reveals the contributions of clustering and discretization approaches as feature extraction methods to improve classification accuracy. In this study, three widely used machine learning techniques are investigated on Perfume, Wine, Seeds, and Gas datasets from the UCI repository. This comprehensive empirical study indicates that the accuracy of classifiers improves by up to 20% on datasets obtained from some sensors by using both discretization and clustering as feature-extracting methods.