Evolving Metric Learning for Incremental and Decremental Features

Online metric learning has been widely exploited for large-scale data classification due to the low computational cost. However, amongst online practical scenarios where features are evolving (e.g., some vanished and new augmented), most models cannot be successfully applied these scenarios, although they can tackle instances efficiently. To address challenge, we develop a Evolving Metric Learning (EML) model incremental decremental features, which handle instance feature evolutions simultaneously by incorporating with smoothed Wasserstein distance. Specifically, our contains two essential stages: Transforming stage (T-stage) Inheriting (I-stage). For T-stage, propose extract important information from while neglecting non-informative knowledge, forward it into survived transforming them low-rank discriminative space. It further explores intrinsic structure of heterogeneous samples reduce computation memory burden especially highly-dimensional data. I-stage, inherit performance T-stage then expand include augmented features. Moreover, distance is utilized characterize similarity relationships among complex samples, since not strictly aligned in different stages. In addition tackling challenges one-shot case, also extend multi-shot scenario. After deriving an efficient optimization strategy both extensive experiments on several datasets verify superior EML model.