LSTM With Bayesian Loss for Ordinal and Imbalanced Channel Quality Prediction

Accurate prediction of the channel quality indicator (CQI) is crucial for enabling adaptive modulation and coding in next-generation wireless networks. However, CQI prediction remains challenging due to its ordinal nature and the frequent class imbalance observed in real-world datasets. This work proposes a novel long short-term memory (LSTM) model with Bayesian loss to address both the ordinal structure of CQI and the imbalanced distribution of CQI levels. The Bayesian component ensures well-calibrated probabilistic outputs, while the LSTM captures temporal dependencies in channel dynamics. Unlike previous models that rely on synthetic or idealized data, our model is trained and evaluated on real indoor millimeter wave (mmWave) scenarios, in environments with pedestrian mobility. Experimental results show that the proposed approach significantly improves accuracy and robustness over baseline models. These findings have important practical implications for dynamic wireless systems, where real-time, reliable CQI prediction is essential for improving network efficiency, throughput, and overall performance in challenging environments.