On-device Deep Learning Location Category Inference Model

Abstract

We define Location Category Inference (LCI) as a task of predicting the category of a visited venue, such as bar, restaurant or university, given user location GPS coordinates and a set of venue candidates. LCI is an essential part of the hyper-personalization systems as its output provides deep insights into user lifestyle (has children, owns a dog) and behavioral patterns (regularly exercises, visits museums). Due to such factors as signal obstruction, especially in urban canyons, the GPS positioning is inaccurate. The noise in the GPS signal makes the problem of LCI challenging and requires researchers to explore models that incorporate additional information such as the time of day, duration of stay or user lifestyle in order to overcome the noise-induced errors. In this paper we propose an embeddable on-device LCI model which fuses spatial and temporal features. We discuss how initial clustering of locations helps limiting the GPS noise. Then, we propose a multi-modal architecture that incorporates socio-cultural information on when and for how long people typically visit venues of different categories. Finally, we compare our model with one nearest neighbor, a simple fully connected neural network and a random forest model and show that the multi-modal neural network achieves f1 score of 73.2% which is 6.6% better than the best of benchmark models. Our model outperforms benchmark models while being almost 180 times smaller in size at around 1.9Mb.