Extend Auto-Sklearn with PyOD algorithms

[johnantonn]

Modify autosklearn to accommodate PyOD models and verify their successful execution in the context of autosklearn.

[johnantonn]

Main points:

• Incorporate models by introducing new modules (.py files) named after the class of the model and implementing the required functions.
• Import the models as indicated by the documentation: https://automl.github.io/auto-sklearn/master/extending.html
• Modify the code so that the test_split includes all labeled points available and not the unlabeled and verify that the execution is correct, i.e. other features were not altered.
• Keep incorporating PyOD models along with their hyperparameter space and setup the final datasets for the experiments.

[johnantonn]

Changes:

• Updated auto-sklearn to most recent version 0.14.4
• Created submodule inside mai-thesis to link auto-sklearn own repo where the changes will take place
• Updated both repo's and now they're linked
• Proceeded with necessary changes for enabling anomaly detection PyOD algorithms to work inside auto-sklearn in an unsupervised learning manner but with supervised validation for estimating f(n) = auc in bayesian optimization

Note: maybe consider extending auto-sklearn 2.0 now that is available and proved to be more optimal? Currently the code for auto-sklearn 2.0 does not seem available.

[johnantonn]

Incorporated the below algorithms:

• Probabilistic
  • ABOD
  • COPOD
  • ECOD
  • MAD (can only be used with univariate data)
  • SOS
• Linear models
  • LMDD
  • MCD
  • OCSVM
  • PCA
• Proximity-based
  • COF
  • CBLOF
  • LOF
  • KNN
  • HBOS
  • LOCI
  • SOD
  • ROD
• Outlier ensembles
  • IForest

Notes:

• Included as many valid hyperparams as possible in order to be expressive and consistent.
• Did not include any NN algorithms since their HPO would be a NAS problem not covered by Auto-Sklearn.
• Did not include any other outlier ensembles (just IForest) since the rest require a list of simple algorithms as input.

[johnantonn]

Class diagram and execution flow of Auto-Sklearn for classification tasks:

[johnantonn]

In order to tackle the critical point of model evaluation, the execution flow was stuidies and mapped to the aforementioned diagarm. Based on that:

• A pre-defined index split needs to be created and pass to PredefinedSplit.
• The split can be considered a heuristic at this point, since the decision on what percentage of outliers to include in the training set and/or the percentage of inliers to include in the validation set can be modified.