Lithology identification is a fundamental task in unconventional petroleum exploration, and acquiring accurate lithological information is vital for reservoir development. However, this task faces three interconnected challenges: limited labeled samples, severe class imbalance, and poor model interpretability, which collectively restrict classification accuracy and practical application. To address these challenges, this paper proposes a KAN-based Adaptive Self-training Semi-supervised Ensemble Learning (KASSEL) framework for lithology identification. The KASSEL framework leverages Kolmogorov-Arnold Networks (KAN) as base classifiers and introduces a novel two-stage ADASYN algorithm to handle class imbalance issues. By integrating adaptive sampling with self-training semi-supervised learning, the framework effectively utilizes both labeled and unlabeled data to build a robust ensemble system. Experimental results demonstrate that KASSEL achieves a macro F1-score of 0.89 on the test dataset, significantly outperforming conventional machine learning and deep learning methods. The framework substantially improves minority class recognition where conventional methods often fail, while maintaining enhanced interpretability through SHAP-based post-hoc analysis combined with KAN’s learnable activation functions. The framework has been successfully applied to the S tight reservoir, achieving high-precision lithology classification and providing engineers with an efficient workflow for rapid lithology determination. The KASSEL framework offers an efficient, accurate, and practically viable solution for lithology identification in unconventional reservoir exploration and development.