[Paper Review] Multi-view anomaly detection via robust probabilistic latent variable models

Tomoharu Iwata, Makoto Yamada|arXiv (Cornell University)|Dec 5, 2016

Anomaly Detection Techniques and Applications24 references30 citations

TL;DR

This paper proposes a robust probabilistic latent variable model for multi-view anomaly detection, where normal instances are generated from a single shared latent vector across views, while anomalous instances are modeled with multiple view-specific latent vectors. Using Dirichlet process priors to infer the number of latent vectors per instance, the method enables Bayesian inference via a stochastic EM algorithm, achieving improved detection performance on noisy multi-view data.

ABSTRACT

We propose probabilistic latent variable models for multi-view anomaly detection, which is the task of finding instances that have inconsistent views given multi-view data. With the proposed model, all views of a non-anomalous instance are assumed to be generated from a single latent vector. On the other hand, an anomalous instance is assumed to have multiple latent vectors, and its different views are generated from different latent vectors. By inferring the number of latent vectors used for each instance with Dirichlet process priors, we obtain multi-view anomaly scores. The proposed model can be seen as a robust extension of probabilistic canonical correlation analysis for noisy multi-view data. We present Bayesian inference procedures for the proposed model based on a stochastic EM algorithm. The effectiveness of the proposed model is demonstrated in terms of performance when detecting multi-view anomalies.

Motivation & Objective

• To address the challenge of detecting anomalies in multi-view data where inconsistencies across views indicate abnormality.
• To model normal instances as sharing a single latent representation across views, while allowing anomalous instances to have multiple, view-specific latent vectors.
• To enable automatic inference of the number of latent vectors per instance using non-parametric Bayesian priors.
• To improve robustness to noise in multi-view data compared to existing methods like probabilistic CCA.
• To develop a scalable Bayesian inference procedure for the proposed model using a stochastic EM algorithm.

Proposed method

• Model normal instances as generated from a single shared latent vector across all views, ensuring consistency.
• Model anomalous instances as generated from multiple, view-specific latent vectors, reflecting inconsistency.
• Use a Dirichlet process prior over the number of latent vectors per instance to enable non-parametric inference.
• Apply a stochastic expectation-maximization (SEM) algorithm for Bayesian inference of latent variables and model parameters.
• Formulate the likelihood of each view given its latent vector, assuming conditional independence given the latent structure.
• Incorporate robustness to noise by allowing flexible modeling of view-specific variations through the latent vector structure.

Experimental results

Research questions

• RQ1 How can we effectively detect anomalies in multi-view data by identifying inconsistent views?
• RQ2 To what extent can a probabilistic latent variable model with non-parametric priors improve anomaly detection in noisy multi-view settings?
• RQ3 Can the number of latent vectors per instance be reliably inferred without prior knowledge?
• RQ4 How does the proposed model compare in performance to probabilistic canonical correlation analysis under noisy conditions?
• RQ5 What is the impact of using a stochastic EM algorithm on scalability and inference accuracy?

Key findings

• The proposed model achieves superior performance in detecting multi-view anomalies compared to baseline methods on noisy data.
• The use of Dirichlet process priors enables automatic and data-driven inference of the number of latent vectors per instance.
• The model effectively captures view inconsistencies, which are indicative of anomalous instances.
• Bayesian inference via stochastic EM provides stable and scalable parameter estimation for the latent variable model.
• The method demonstrates robustness to noise, outperforming standard probabilistic CCA in multi-view anomaly detection tasks.
• Empirical results confirm that the model’s anomaly scores are more discriminative for anomalous instances than those from existing approaches.