A k-nearest neighbors based approach applied to more realistic activity recognition datasets

Abstract

Due to the latest technological advances, the current society has the possibility to store large volumes of data in the majority of the problems of the daily life. These data are useless if there is not a set of techniques available to analyze them with the objective of obtaining knowledge that facilitates the problem resolution. This paper focuses on the techniques provided by data mining as a tool for intelligent data analysis in the field of human activity recognition, specifically in the application of two techniques of data mining capable of carrying out the extraction of knowledge from data that are not as accurate and exact as desirable. This type of data reflects the true nature of the information collected on a day-to-day basis. The proposed techniques allow us to perform a preprocessing of the data by means of an instance selection that improves the computational requirements of the system response, obtaining satisfactory accuracy results. Several experiments are carried out on a real world dataset and various datasets obtained from the previous one in a synthetic way to simulate more realistic datasets that illustrate the potential of the techniques proposed.