Accessing fitted underlying scikit-learn models after fitting PyOD models

[AhmetZamanis]

Hi, I've been experimenting with some of the PyOD anomaly detection algorithms. I have a question regarding some PyOD models which are based on scikit-learn implementations. In general, I wonder if there is a way to access the fitted underlying scikit-learn models and retrieve some attributes not available through the PyOD wrapper. I'll share two examples:

The Gaussian Mixture Model gmm.GMM is a wrapper around scikit-learn's GaussianMixture. According to the documentation, gmm.GMM predicts anomaly scores & binary anomaly labels, and stores various other attributes. But it does not predict the component labels as GaussianMixture does, or store the labels for the training data as an attribute. I wonder if it's possible to somehow retrieve the component labels so they can be easily used for "cluster plots". In comparison, it's possible to do this with cblof.CBLOF, through the cluster_labels_ attribute.

The second one is the PCA-based outlier detector, pca.PCA. Again, this method predicts anomaly scores & labels, and stores some PCA attributes like PC loadings and variances explained. But it's not possible to retrieve the actual principal component values for the datapoints. In contrast, the scikit-learn implementation of PCA transforms the data into principal component values & returns these values with .transform(). I wonder if there is a way to retrieve the PC values either directly, or by accessing the fitted underlying PCA object. These could be useful to plot against the binary anomaly labels, to see how the PC values differ between anomalies and non-anomalies.

Thanks in advance for any answers.

[yzhao062]

PCA is also a wrapper -> https://github.com/yzhao062/pyod/blob/66010dd4cbf614b51d34bab9827151b409ae7595/pyod/models/pca.py#L230C12-L230C12

and the attributes are available at:

[AhmetZamanis]

Thank you, I did not realize .detector_ provides what I needed.