Bridging Complexity and Interpretability: A Two-Phase Clustering Framework

Unsupervised learning has consistently been a chal-lenging aspect of machine learning, particularly in robotic tasks such as autonomous navigation, object grasping, and anomaly detection. These tasks require the system to identify patterns in data without any labeled examples, making it difficult to achieve reliable and accurate outcomes. In medical applications, such as tumor detection and segmentation, the difficulties of unsupervised learning become even more noticeable. Without labeled data to guide the models, discovering meaningful patterns often leads to poor interpretability, limiting the effectiveness of these systems in critical medical decision-making processes. In these contexts, transparency and understanding are essential, especially for safety-critical tasks. Recognizing this need, we introduce a novel two-phase clustering approach that combines latent space transformation using an autoencoder with weighted feature augmentation and post-clustering in the original feature space, thereby improving explainability by making the clusters more interpretable and highlighting key features for a better understanding of the results. Building on this, our method, applied to a real-world dataset characterized by complex human behaviors developed by our group, demonstrates superior performance with improved silhouette scores and comprehensive analysis, effectively capturing complex customer behaviors while enhancing interpretability and providing key insights for data-driven decisions.