Long-tail distribution is widespread in many practical applications, where most categories contain only a small number of samples. This imbalance poses significant challenges for recognizing underrepresented classes, since they can be easily biased towards dominant classes and perform poorly on tail classes. To address these issues, we propose LTBoost, a two-phase approach: (1) confusion-matrix-guided mixup augmentation to improve rare-class representation and (2) logit adjustment via meta-data optimization to reduce majority-class bias. LTBoost enhances recall uniformity, illustrated by balanced accuracy and geometric mean metrics. We also introduce the coefficient of variation to better assess the uniformity of the recall distribution across all classes when the geometric mean can be zero. Experiments confirm LTBoost’s effectiveness in improving minority-class recognition and achieving balanced performance in real-world long-tailed datasets.