A variational autoencoder (VAE) is a generative model that compresses an input into a probabilistic latent representation and reconstructs it. Unlike a standard autoencoder, which gives an input a fixed encoding, a VAE represents each latent value with a mean and variance. Training seeks both accurate reconstruction and a well-formed latent space that can support generation of new content. section 16.2.2
For outlier detection, the generative ability is usually not the goal. Instead, reconstruction error estimates how unusual an input is: anomalous inputs tend to reconstruct less accurately. A VAE and a standard autoencoder may perform differently on a given dataset, so neither should be assumed superior. section 16.2.2
Listing 16.3 Example using the PyOD VAE detector
from pyod.models.vae import VAE np.random.seed(0) det = VAE(epochs=5, random_state=0) det.fit(X) pred = det.predict_proba(X)[:, 1]
A short PyOD example fits a VAE to X for five epochs, sets a random seed, and obtains outlier predictions with predict_proba(X)[:, 1]. It assumes X has already been collected. section 16.2.2
Check whether the VAE helps before relying on its flags. Compare it with an autoencoder and other detectors using synthetic test data or a pseudo-ground truth, and tune parameters against that evidence. A well-tuned VAE can also add a different perspective to an ensemble. On tabular data, deep-learning detectors are often slower and harder to tune than established alternatives without being more accurate, although large datasets or complex relationships among many features may favor them. section 16.2.2 section 16.0
