Reliable facial expression learning (FEL) involves the effective learning ofdistinctive facial expression characteristics for more reliable, unbiased andaccurate predictions in real-life settings. However, current systems strugglewith FEL tasks because of the variance in people's facial expressions due totheir unique facial structures, movements, tones, and demographics. Biased andimbalanced datasets compound this challenge, leading to wrong and biasedprediction labels. To tackle these, we introduce GReFEL, leveraging VisionTransformers and a facial geometry-aware anchor-based reliability balancingmodule to combat imbalanced data distributions, bias, and uncertainty in facialexpression learning. Integrating local and global data with anchors that learndifferent facial data points and structural features, our approach adjustsbiased and mislabeled emotions caused by intra-class disparity, inter-classsimilarity, and scale sensitivity, resulting in comprehensive, accurate, andreliable facial expression predictions. Our model outperforms currentstate-of-the-art methodologies, as demonstrated by extensive experiments onvarious datasets.