|Termin||Fr 29. März 2019|
|Kurzfassung||Outlier detection targets the discovery of abnormal data patterns. Typical scenarios, such as are fraud detection and predictive maintenance are particularly challenging, since the data is available as an infinite and ever evolving stream. In this thesis, we propose Adaptive Variational Autoencoders (AVA), a novel approach for unsupervised outlier detection in data streams.
Our contribution is two-fold: (1) we introduce a general streaming framework for training arbitrary generative models on data streams. Here, generative models are useful to capture the history of the stream. (2) We instantiate this framework with a Variational Autoencoder, which adapts its network architecture to the dimensionality of incoming data.
Our experiments against several benchmark outlier data sets show that AVA outperforms the state of the art and successfully adapts to streams with concept drift.