Learning Probabilistic Latent Structure for Outlier Detection from Multi-view Data

Mining anomalous objects from multi-view data is a challenging issue as data collected from diverse sources have more complicated distributions and exhibit inconsistently heterogeneous properties. Existing multi-view outlier detection approaches mainly focus on transduction, which becomes very costly when new data points are introduced by an input stream. Besides, the existing detection methods use either the pairwise multiplication of cross-view data vectors to quantify outlier scores or the predicted joint probability to measure anomalousness, which are less extensible to support more sources. To resolve these challenges, we propose in this paper a Bayesian probabilistic model for finding multi-view outliers in an inductive learning manner. Specifically, we first follow the probabilistic projection method of latent variable for exploring the structural correlation and consistency of different views. Then, we seek for a promising Bayesian treatment for the generative model to approach the issue of selecting the optimal dimensionality. Next, we explore a variational approximation approach to estimate model parameters and achieve model selection. The outlier score for every sample can then be predicted by analyzing mutual distances of its representations across different views in the latent subspace. Finally, we benchmark the efficacy of the proposed method by conducting comprehensive experiments.