Impact of Platt Scaling on Calibration in ML-Based Wireless Resource Allocation

In this paper, we study the calibration performance of a machine learning (ML)-based outage predictor applied to a single-user, multi-resource allocation system. Our approach models the wireless channel using Rayleigh fading with temporal characteristics that are consistent with Clarke’s 3D model. This allows us to account for the mobility of the receiver, introducing correlation between successive channel samples. Building upon this, we study the calibration performance of the outage predictor when Platt scaling is applied. Its calibration performance is assessed by generating histogram-based reliability plots with logarithmic binning. This outage predictor is trained using a customized outage loss function (OLF) as well as the commonly employed binary cross entropy. Using the negative log likelihood as an indicator, our results show that Platt scaling has a clear impact upon calibration performance, with greater enhancement observed at higher confidence levels and less improvement at lower ones. Furthermore, we observe that Platt scaling is particularly effective in minimizing overconfidence for predictors trained with OLF at lower classification thresholds and as the signal-to-noise ratio increases.