Adaptive State-Space Neural Expectation-Maximization Approach for Mobile Network Traffic Forecasting

The paper develops an end-to-end probability-informed approach for short-term mobile network traffic forecasting. In continuation of previous studies on probability informing via finite normal mixtures, a Neural Expectation-Maximization (Neural EM) block is proposed with three new enhancements: a neural-network initialization of the mixture parameters followed by E- and M-steps; canonical ordering of mixture components combined with a state-space processing of the resulting parameter vector; and a self-tuning loss in which the weight of a probabilistic anchor based on the mixture's log-likelihood is adjusted by a reinforcement learning agent. The block is end-to-end trainable and can be attached to arbitrary machine learning algorithms and neural network architectures without changes to their basic structure. On the mobile traffic data, attaching the block consistently improves the predictive accuracy of two host architectures: an LSTM is improved by up to 9.5% and a Transformer encoder by up to 19.6%. The applicability is also verified on geophysical data, where the proposed approach additionally stabilizes the training across random runs.