Time series classification algorithms applied to embedded systems

Abstract

Integrating sequential and temporal data into the Data Mining process is of one of the most important challenges in Machine Learning. In this project, we are mostly interested in developing time series classification algorithms. The k-nearest neighbor algorithm is a common approach to time series classification. This algorithm has been known to perform well, especially when allied to distance measures that can deal with time lags, such as the Dynamic Time Warping. However, the classical k-nearest neighbor algorithm is computationally intensive. One may solve this problem by using indexes to increase the efficiency of similarity queries. This project proposes to investigate indexing algorithms that have the properties of anyspace algorithms. Anyspace algorithms are able to deal with different amounts of memory, in such a way that the algorithm performance depends directly on the amount of available memory. Such algorithms allow specifying the amount of memory based on the performance required by an embedded application. This project also deals with classification methods based on induction of classification rules. An approach to induce rules from time series data is the identification of motifs. Motifs are frequently occurring subsequences that usually represent a phenomenon of interest. A convenient aspect of rules is the ease one finds in writing a procedural program which implements the rule's logic with little memory and processing resources. The algorithms developed in this post-doctoral stage will be applied in insect control and monitoring using devices developed by ISCA Technologies. (AU)