Distance Metric Tracking

Recent work in distance metric learning focused on learning transformations of data that best align with provided sets of pairwise similarity and dissimilarity constraints. The learned transformations lead to improved retrieval, classification, and clustering algorithms due to the more accurate distance or similarity measures. Here, we introduce the problem of learning these transformations when the underlying constraint generation process is dynamic. These dynamics can be due to changes in either the ground-truth labels used to generate constraints or changes to the feature subspaces in which the class structure is apparent. We propose and evaluate an adaptive, online algorithm for learning and tracking metrics as they change over time. We demonstrate the proposed algorithm on both real and synthetic data sets and show significant performance improvements relative to previously proposed batch and online distance metric learning algorithms.