Machine learning has become an important part of modern malware detection. However, many high-performing models act like black boxes. This makes it difficult to use them effectively in security operations. This research presents an interpretable framework for classifying malware. It uses gradient-boosted decision trees that are trained on static data from Windows PE files. The proposed model is trained on a dataset from the EMBER benchmark. It delivers strong detection results, with an accuracy of 90.72%, a recall of 93.51%, and an ROC-AUC of 0.9691. To improve transparency, SHAP (Shapley Additive Explanations) is added to the pipeline. This helps identify which structural and statistical file attributes most influence model predictions. Experimental evaluations show that SHAP successfully identifies important features like entropy patterns and header anomalies. This provides analysts with useful insights that boost confidence and help with incident response. While creating detailed explanations can be computationally intensive, especially at scale, the findings suggest that combining machine learning with explainability leads to a more reliable and useful malware detection system. Future improvements could involve real-time explanation techniques, integration with threat intelligence frameworks like MITRE ATT&CK, and the addition of dynamic behavioral attributes.