Energy Characterization and Optimization of Embedded Data Mining Algorithms: A Case Study

Abstract—Data mining has been flourishing in the information-based world. In data mining, the DTW-kNN framework is widely applied for classification in miscellaneous application domains. Most of the studies in the DTW-kNN framework focus on accuracy and speedup. However, with increasingly emphasis on applications of mobile and embedded systems, energy efficiency becomes an urgent consideration in data mining algorithm design. In this paper, we present our work on energy characterization and optimization of data mining algorithms. Through a case study of the DTW-kNN framework, we investigate multiple existing strategies to improve the energy efficiency without any loss of algorithm accuracy. To the best of our knowledge, this is the first work about energy characterization and optimization of data mining algorithms on embedded computing testbeds. All the experiments are implemented on a developed energy measurement testbed. The experimental results indicate that the distance matrix calculation is the bottleneck of the DTW-kNN framework, which accounts for 89.14 % on average of the total energy. With several optimized methods, the reduction of the total energy in the DTW-kNN framework can reach as much as 74.6%. Keywords—Dynamic Time Warping; k-Nearest Neighbors; en-ergy optimization; embedded computing testbed I