functional_definitions.md

Functional definitions

While being designed with genericity in mind, active_flow has to make assumptions on the objects it manipulates

When one calls

python aflow.[active|interactive] create_model {external_module.model_class} ...

active_flow will create the model from the create_model arguments, given the dependencies in the current python environment.

This instance - the model - has to pass certain checks. If not, active_flow will fail elegantly, and report actions that can be triggered for the model to be accepted by the CLI.

Active Machine Learning

Estimator to be sumbitted to the aflow.active CLI must implement a fit method, as defined in the scikit-learn API, or a partial_fit method, to increment the model once targets have been queried from the user for a given batch.

Interactive Data Mining

The API for data mining is currently under progress. For now the main design would require a miner to provide:

• a prefit method, to be called on the input data before the main -interactive - active_flow loop
• a generate_candidates method that takes no argument, returns an iterator (generator included)
• an update method to udpate the model given a candidate generated by generate_candidates, and validated by the oracle

Notes

1. methods like evaluate would not be required, as evaluation is done by the oracle in our typical workflow
2. "any_time" methods like skmine.itemsets.SLIM generate a new candidate given the - freshly - updated model, thus resulting in some sort of an "online" interactive pattern mining workflow