Quantity forecast of mobile subscribers with Time-Dilated Attention

Abstract

The quantity forecast of mobile subscribers requires accurate and reliable results for obtaining insights into user trends and facilitating effective business management. Due to the complexity inherent in mobile subscriber data, influenced by subscriber tendencies and device popularity, capturing its underlying regularities poses a challenge. In this research, a novel Time-Dilated Attention (TDA) model is proposed, complemented by a feature extraction method characterized by high interpretability and distinguishability. Its efficacy and implications are explored on a real-world mobile subscriber dataset. TDA facilitates the acquisition of more informative representations, while our feature extraction method enhances the ability to discern dissimilar samples, thereby improving the stability of mobile subscriber trend analysis. The approach is validated on three additional datasets to assess its robustness. Experimental findings on the target mobile subscriber dataset demonstrate that the proposed approach achieves reductions in MAE, RMSE, and Theil’s U by 1.45%, 5.28%, and 5.12%, respectively, compared to the strongest baseline methods. Additionally, it attains the second-best performance in terms of MedAE. Furthermore, this model consistently ranks within the top two positions for nine out of twelve metrics on the additional datasets, underscoring its generalizability.