Latent space conditioning for improved classification and anomaly detection

We propose a variational autoencoder to perform improved pre-processing forclustering and anomaly detection on data with a given label. Anomalies howeverare not known or labeled. We call our method conditioned variationalautonencoder since it separates the latent space by conditioning on informationwithin the data. The method fits one prior distribution to each class in thedataset, effectively expanding the prior distribution to include a Gaussianmixture model. Our approach is compared against the capabilities of a typicalvariational autoencoder by measuring their V-score during cluster formationwith respect to the k-means and EM algorithms. For anomaly detection, we use a new metric composed of the mass-volume andexcess-mass curves which can work in an unsupervised setting. We compare theresults between established methods such as as isolation forest, local outlierfactor and one-class support vector machine.