Deep Contextual Novelty Detection with Context Prediction

Contextual novelty detection models detect novelties with respect to a given context. This is crucial in streaming scenarios where the definition of both normal and novel evolve over time. Such models however require contextual labels not only for training but also for detection during deployment. This creates an often unreasonable burden for additional contextual labels during the deployment of these models. In order to eliminate the need for these labels, we propose to predict this contextual information using an auxiliary prediction strategy which takes advantage of the rarity of novel examples, allowing these labels to instead be inferred. The inferred labels are then used as a conditioning criterion for deep autoencoders. We evaluate our approach on a large, public industrial machine sound dataset and show that our approach can successfully recognise context and use this to effectively condition novelty detection models, allowing them to outperform their unconditioned counterparts.