A user-embedded temporal attention neural network for IoT trajectories prediction.

Abstract

Over the past two decades, sequential recommendation systems have garnered significant research interest, driven by their potential applications in personalized product recommendations. In this article, we seek to explicitly model an algorithm based on Internet of Things (IoT) data to predict the next cell reached by the user equipment (UE). This algorithm exploits UE embedding and cell embedding combining the visit time interval information, and uses sliding window sampling to process more UE trajectory data. Furthermore, we use the attention mechanism, removed the query matrix operation and the attention mask, to obtain key information in data and reduce the number of parameters to speed up training. In the prediction layer, combining the positive and negative sampling and computing cross entropy loss also provides assistance to increase the precision and dependability of the entire model. We take the six adjacent cells of the current cell as candidates due to the limitation of the space problem, from which we predict the next destination cell of track movement. Extensive empirical study shows the recall of our algorithm reaches 0.5766, which infers the optimal result and high performance of our model.