MIMU: Mobile WiFi Usage Inference by Mining Diverse User Behaviors

Mobile WiFi is a newly emerging service in recent years, which provides convenience for users to access online resources and increases revenues for operators via services such as advertisements and application promotions. However, in practice, the prohibitively high system implementation and operational costs, especially the costs of perpetual data traffic, hinder the further deployment of mobile WiFi services. In this paper, we present MIMU, a usage inference system for data traffic saving suitable for ubiquitous mobile WiFi services. We demonstrate the performance of the system via an example from the real-world nationwide edge computing mobile WiFi infrastructure. To address the impact of diverse user behaviors, we investigate the WiFi network usage from the perspective of users and devices, focusing on two unique features of mobile WiFi: user mobility regularity and access irregularity. In particular, we first design a deep learning-based two-dimension usage predictor to infer the future mobile WiFi usage with 1) a user dimension model with temporal attention addressing dominant users with heavy bus WiFi usage, and 2) a device dimension model with spatial attention addressing diverse WiFi usage and connection. Based on the results of the predictor, an application of content caching is implemented in an iterative fashion to save the data traffic. We evaluate MIMU by real-world bus WiFi system data sets of three major cities with 6,643 bus WiFi devices and 150k daily active users in total. Our results show that MIMU outperforms state-of-the-art methods in terms of usage inference. Moreover, we summarize the lessons learned from our large-scale bus WiFi system investigation.