Evaluation and optimization of machine learning models for edge processing of mobile network data

Abstract

The article addresses the evaluation and optimization of machine learning and deep learning models for execution on edge nodes when processing mobile network data. The relevance of the study is determined by the fact that models with similar predictive quality may differ substantially in inference latency, model artifact size, and computational requirements; therefore, selection based on accuracy alone does not guarantee suitability for edge deployment. The aim of the study is to build a consistent bank of deployment configurations by jointly evaluating predictive quality and execution characteristics before and after model optimization. The input is a fixed-dimensional integrated mobile network state vector containing statistical, temporal, event-based, resource, and data-quality features. XGBoost, MLP, GRU, TCN, and a Transformer-based model are compared for the same classification task. Optimized variants are produced using dynamic INT8 quantization, structured pruning, and knowledge distillation. Each configuration is characterized by macro-F1, balanced accuracy, recall of the critical class, median and p95 inference latency, and serialized model size. After a quality gate, Pareto screening removes configurations dominated simultaneously in predictive quality, latency, and size. The experimental study uses a synthetic sequence of 17,995 time windows with four mobile network state classes. The highest baseline macro-F1 is obtained by GRU and TCN, both close to 0.918. Structured TCN pruning reduces the model artifact by 23.0% and median latency by about 10% without degrading macro-F1. Distillation reduces the Transformer model size by 73.4% and almost halves its latency. In contrast, dynamic INT8 quantization greatly reduces the size of MLP and GRU models but increases latency on the reference CPU runtime because of quantization overhead. The results demonstrate that edge optimization should be assessed using a set of predictive and execution metrics rather than a single indicator. The resulting deployment bank can serve as a set of verified candidates for subsequent adaptive selection of a model and execution node in a distributed edge/cloud mobile network infrastructure.