Rethinking Long-Term Evolution Throughput Prediction from Passive Drive Tests: Target Semantics and Dependence-Aware Validation

As seen in this research, passive phone bitrate logs cannot be taken as a sign of actual Long-Term Evolution (LTE) network capacity. The study compares conventional random validation with a combination of target-semantics analysis, dependence-aware validation, spatial holdout testing, and simple baseline comparisons, and demonstrates the superiority of the combined approach to avoid over-estimating the performance of machine-learning models in drive-test data. The main contribution is a more reliable framework for evaluation of cellular throughput prediction models based on passive mobile measurements.