Accuracy Improvement of User Move Embedding with Latitude and Longitude Information

Abstract

Recently, with the rise of the popularity of wearable devices like smartphones, it is becoming possible to use user’s moving records as a part of big data. In previous research, user movements were modeled by Bi-directional LSTM trained with time series of Mesh-ID. However, since Mesh-ID was assigned artificially, its physical distance and relative position were not considered properly. Therefore, it might be difficult to train the model effectively by using only Mesh-ID. In this research, to solve this problem, we used additional latitude and longitude information in the new model. As a result, we confirmed the accuracy improvement for Mesh-ID prediction and the difference between the output of clustering with user embeddings for each model.