This paper addresses the challenge of detecting anomalies in Blockchain transactions, particularly Bitcoin transactions, using machine learning classifiers and explainability analysis. The study integrates Explainable Artificial Intelligence (XAI) techniques, specifically the Shapley Additive explanation (SHAP) method, to enhance the interpretability of model predictions. The authors propose an under-sampling algorithm called XGBCLUS, which balances the dataset by selecting a subset of instances from the majority class, and compare it with other under-sampling and over-sampling techniques. The effectiveness of these techniques is evaluated through various evaluation metrics such as accuracy, True Positive Rate (TPR), False Positive Rate (FPR), and ROC-AUC score. The results show that XGBCLUS outperforms other under-sampling methods in terms of TPR and FPR, while the ADASYN over-sampling technique performs better in reducing FPR. Additionally, the study compares the performance of single tree-based classifiers with ensemble classifiers, finding that ensemble models, particularly the voting classifier, achieve higher accuracy and FPR values. The paper also presents a set of rules derived from the tree-based model to interpret anomalous transactions. Overall, the study demonstrates the effectiveness of the proposed methods in detecting anomalies in Blockchain transactions and improving the interpretability of the models.This paper addresses the challenge of detecting anomalies in Blockchain transactions, particularly Bitcoin transactions, using machine learning classifiers and explainability analysis. The study integrates Explainable Artificial Intelligence (XAI) techniques, specifically the Shapley Additive explanation (SHAP) method, to enhance the interpretability of model predictions. The authors propose an under-sampling algorithm called XGBCLUS, which balances the dataset by selecting a subset of instances from the majority class, and compare it with other under-sampling and over-sampling techniques. The effectiveness of these techniques is evaluated through various evaluation metrics such as accuracy, True Positive Rate (TPR), False Positive Rate (FPR), and ROC-AUC score. The results show that XGBCLUS outperforms other under-sampling methods in terms of TPR and FPR, while the ADASYN over-sampling technique performs better in reducing FPR. Additionally, the study compares the performance of single tree-based classifiers with ensemble classifiers, finding that ensemble models, particularly the voting classifier, achieve higher accuracy and FPR values. The paper also presents a set of rules derived from the tree-based model to interpret anomalous transactions. Overall, the study demonstrates the effectiveness of the proposed methods in detecting anomalies in Blockchain transactions and improving the interpretability of the models.