Anomaly detection is a fundamental and important problem in real-world applications, such as clinic diagnosis, fault detection, fraud detection, cybersecurity, etc. This thesis considers anomaly detection with multi-view data. We propose to develop multi-view anomaly detection models that are semi-supervised, inductive and robust. Our core assumption is there exists a latent space, where different views of a normal instance are mapped to a single point, whereas views of an abnormal instance are mapped to different points. Two models are proposed: one is a probabilistic latent variable model(LVM) with Gaussian distribution assumptions on the shared latent variable and noise term(or projection error); the other is a Bayesian model with Robust (Student-t) assumptions on the `common' latent variable and projection noise term (in other words, the latter model can be considered as having a robust likelihood). Following the proposed model, we define the outlier score measurement which evaluates the likeliness of an instance being the outlier. We then employ the Expectation-Maximization (EM) algorithm and Variational Inference (VI) to estimate the parameters of those models respectively. If the score is greater than a pre-setting threshold, we identify it as a outlier. We experiment the proposed models on nine real-world data sets, and the results show our methods consistently outperform the state-of-the-art competitors. We prove the consistency of our Bayesian model according to Evidence Lower Bound(ELBO) criterion.