BSENet: A Data-Driven Spatio-Temporal Representation Learning for Base Station Embedding

Base station (BS) plays a critical role in the wireless network. There has been some research on exploring the spatio-temporal information of BS in different fields. However, these works lack re-usability, which require the new researcher to re-do the work of representing the spatio-temporal information. To solve this problem, we propose a neural network model based on autoencoder and representation learning called BSENet to learn embedding of BS based on raw data. The embedding contains spatio-temporal information of BS. In this way, other fields that are related to BS can make use of spatio-temporal information with BSENet embeddings. Besides the spatial information, BSENet can maintain the independence of BS. Moreover, we use bi-directional LSTM to get temporal information and propose a time dropout method to improve the generalization ability. We propose a soft threshold to consider all spatial relations. In addition, we introduce weight to enhance compatibility. We treat the missing states as the inputs to deal with the missing values. The results of clustering show that BSENet embedding is better than other embeddings. In experiments of mobile traffic prediction, BSENet embedding helps a temporal model to improve the Mean Squared Error (MSE) from 0.01463 to 0.01334, which is similar to 0.01302 of a spatio-temporal model. At the same time, the training time increases a little but is still $5.0 times$ faster than spatio-temporal model.