Dependency Clustering Across Measurement Scales

摘要

How to automatically spot the major trends in large amounts of heterogeneous data? Clustering can help. However, most existing techniques suffer from one or more of the following drawbacks: 1) Many techniques support only one particular data type, most commonly numerical attributes. 2) Other techniques do not support attribute dependencies which are prevalent in real data. 3) Some approaches require input parameters which are difficult to estimate. 4) Most clustering approaches lack in interpretability. To address these challenges, we present the algorithm Scenic for dependency clustering across measurement scales. Our approach seamlessly integrates het-erogenous data types measured at different scales, most importantly continuous numerical and discrete categorical data. Scenic clusters by arranging objects and attributes in a cluster-specific low-dimensional space. The embedding serves as a compact cluster model allowing to reconstruct the original het-erogenous attributes with high accuracy. Thereby embedding reveals the major cluster-specific mixed-type attribute dependencies. Following the Minimum Description Length (MDL) principle, the cluster-specific embedding serves as a codebook for effective data compression. This compression-based view automatically balances goodness-of-fit and model complexity, making input parameters redundant. Finally, the embedding serves as a visualization enhancing the interpretability of the clustering result. Extensive experiments demonstrate the benefits of Scenic.