Movement Attentive Graph Embedding for Improving Next Point of Access Prediction

Next Point-of-Attachment (PoA) prediction is a deep learning-based approach to provide seamless connectivity in the mobile networks from heterogenous and dense user access. Existing recurrent neural models utilize time series characteristics to extract and learn the sequential association changes occurs between all Base Stations (BSs). Since the user moves from certain BS to its adjacent BS, the deployments of these BSs are a crucial feature for improving prediction accuracy. To this end, we propose a Graph Embedding - Long Short-Term Memory (GE-LSTM) model using Recurrent Neural Network (RNN) and Graph Neural Network (GNN) jointly. The time-series pattern of the user movement is extracted by LSTM, and the movement between adjacent BSs is emphasized as GNN through GE. The joint property of the proposed model is implemented through mixture-of-experts approach from GNN and LSTM. The performance is evaluated along with the RNN families and Transformer models, furthermore, it explores the significance of the graph-based mobility embeddings. In our experiment, the proposed model outperforms other SOTA models by achieving the maximum accuracy 94%.