ModelFlow

Develop predictive models with one configurable workflow. The workflow configuration should be understandable to anyone interested.

Entrypoints

Pass raw data through feature-transforms pipeline: python -m src.features.main

With transformed data, estimate models, using modules in src.models

Configuration Files

Configure the data sample. Data are real, while models are intangible abstractions.

• From what (storage) source do we extract data?
• What filter subsets the training data?
• What filter subsets new test data?
• What outcome do we seek to predict? How should its missing values be interpreted?
• What data columns constitute subject attributes -- column-concatenated to predictions, for subject identifiers and exploratory analysis?

Configure the feature transforms pipeline, which yields model-ready inputs from raw values. After declaring preset transformers, proceed by feature, declaring each one's transforms. With feature-then-transforms flow, intent is analyst-friendly thought process. (Under the hood, details convert to transform-then-features structure, according to transformer APIs.)

• Uniquely name the pipeline. Arbitrarily many could transform a dataset.
• Name transform functions (transformers) with default arguments.
• Declare model features. For each,
  • What data type?
  • Which transforms apply?

Configure subdirectories which house feature transform pipeline artifacts. Expect several models will co-exist. The subdirectory structure must partition:

• Two pipelines predicting different outcomes
• Two pipelines predicting same outcome, but trained from different samples
• Two pipelines predicting same outcome & trained from same samples, but following different transform procedures