Machine learning approaches have demonstrated impressive performance in Android malware detection. Yet, most—if not all—of these approaches face trade-offs among accuracy, interpretability, and scalability. Approaches based on simple features are interpretable but often fail to capture complex behaviors, while approaches that model holistic application patterns tend to obscure the specific code responsible for malicious activity. In this paper, we outline our vision for an accurate, scalable, and interpretable method-level malware detection framework. The core idea behind our approach is to filter out non-discriminative parts of applications before analyzing the remaining, application-specific behaviors at a finer level of granularity. We further discuss key challenges that must be addressed to effectively realize this approach and provide suggestions for future research directions.