Implementation and Performance Analysis of Data-Mining Classification Algorithms on Smartphones

Abstract

Smartphones are increasingly being used to capture data and perform complex tasks, however, this rarely extends to the local training of data models. This study investigates the implementation of data mining classification algorithms on smartphones, using 20 classifiers and nine mixed-design datasets on three devices. Their accuracy and processing speed are further compared against a laptop computer using our cross-platform ‘DataBench’ software. Results show that smartphones not only deliver classification accuracy equal to that of more powerful computers when using the same algorithms, as expected, but also, as many as 75% of the 180 algorithm/dataset learning tasks tested were completed on smartphone hardware within three seconds. However, tests further show that the increased complexity of newer algorithms searching for ever greater classification accuracy is resulting in model-build times growing at an exponential rate. Additional testing identified that while a single algorithm execution can have negligible effect on battery life, power efficiency is affected by algorithm complexity, data size and attribute type. The increased processing demand of local model-learning on smartphones also results in increased power dissipation. Yet, even on a continuous-loop execution basis, mobile temperature gains over a 15-min period did not exceed 7 °C. Our conclusion is that smartphones are ready to form self-reliant mobile data-mining solutions able to efficiently execute a wide range of classification algorithms. This offers numerous advantages, including data security and privacy improvements, removal of reliance on network connectivity and delivery of personalised learning.