Robust and Ubiquitous Mobility Mode Estimation Using Limited Cellular Information

Recent mobility mode estimation systems propose using signals from the serving cell tower only to be deployable on all phones. However, all available solutions depend on statistical feature engineering, providing relatively low accuracy. Moreover, their performance is adversely affected by variations in phone placement, hindering their real-world use. To address these limitations, we present <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">AutoSense</i>, the first mobility mode detection system to provide high accuracy using signals from the serving tower only while being robust to phone placement variations. <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">AutoSense</i> proposes a novel domain-specific deep learning-based model to perform automatic feature extraction and temporal processing, improving mode estimation accuracy compared to available solutions. Furthermore, <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">AutoSense</i> integrates a denoising autoencoder into its model to learn salient features robust to changing phone placements. As part of its design, <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">AutoSense</i> addresses several challenges, including cellular data sparsity, absence of motion information, and information decay in long-term dependencies. We conduct extensive evaluations using a real-world public dataset comprising different phone placements. Our results show that <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">AutoSense</i> significantly outperforms state-of-the-art systems, achieving up to 20% and 19% higher average precision and recall, respectively. In addition, <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">AutoSense</i> exhibits remarkable resilience to unseen placements, giving up to 60% better robustness on average compared to the state-of-the-art.