TOWARDS OPTIMIZING ENERGY CONSUMPTION IN MOBILE PHONES OVER Wi-Fi: A DATA DRIVEN APPROACH

Abstract

Contemporary mobile devices are increasingly equipped with multiple network interfaces with diversified characteristics. Due to ubiquity and commendable amount of throughput bestowed by the Wi-Fi technology, mobile phones furnished with WLAN interface have grown exponentially in recent years. Vis-à-vis to other network interfaces, Wi-Fi is least power efficient in the idle state and causes highest energy overhead when scanning for new networks. Hence futile Wi-Fi network scanning, Wi-Fi idling create significant predicaments to modern day mobile users such as excessive battery dilution and privacy impairment. This data-driven approach focuses on exploring user perspective of Wi-Fi usage, where main objective is to model Wi-Fi usage of mobile users based on their past usage behaviour and thereby effectively inferring the active Wi-Fi usage requirements of mobile users to reduce probable inactive periods of Wi-Fi interface. Temporal, mobile application usage, mobile operational state and location contexts have been used as the feature categories to model the proposed prediction framework. Series of machine learning algorithms such as Naïve Bayes, C4.5, RandomForest and SVM have been experimented on the trace data collected from Rice-Livelab user study conducted by the Rice University,Texas. We propose a systematic feature engineering process to overcome challenges embedded in Wi-Fi usage data; such as size, noise and data imbalance. The study utilizes Sampling, Ensemble and Hybrid as techniques to overcome the limitations caused by the class imbalance phenomena. For the validation purposes, we use wide range of evaluation metrics to analyse the effectiveness of applying classification algorithms and class imbalance mitigation techniques on our dataset. Experimented results indicated that decision tree based classification algorithms; C4.5 and RandomForest perform well with class imbalance phenomena, while functional classification algorithm SVM perform poorly with it. The proposed solution is intended to predict active & inactive Wi-Fi usage, where Wi-Fi interface can be turned on whenever it’s required and switch it off when it’s not; reducing futile network scanning and idling periods of Wi-Fi interface.