General Modeling Starting Point

fork this repo for new predictive modeling projects

Basic Idea

Data Classes

Data classes are the building blocks

Each data class should be some logical/semantic group of data:

• IP addresses
• credit attributes
• sales data
• etc...

Each data class should have:

• init where you set the value of some common key to be used in query filtering
• gather_data_submethod where data is gathered through DB queries, file reads, etc
• process_data_submethod where any data cleaning/transforming is done

Class Dictionary

A class dict specifies how the data should be joined together

my_class_dict = {
  DataClass1: ['key1', 'key3']
  DataClass2: ['key2']
}

Modeling Data

ModelingData takes a class dict and a list of ids and will build the modeling dataset

my_data = ModelingData(class_dict=my_class_dict, ids=[8234, 9314, 2828])
my_data.build()

Model Wrapper

The ModelWrapper implements the actual predictive modeling components

• tuning hyperparameters
• fitting a model
• saving a model

ModelWrapper takes a ModelingData object which is the data the model will learn from and a model_type which needs to correspond to a value in ModelLookup

my_model = ModelWrapper(data=my_data, model_type='xgboost', model_name='my_xgboost_model')
my_model.tune_hyperparams(n_calls = 20)
my_model.fit()
my_model.save(dir='~/saved_models')

Model Lookup

ModelLookup is basically a library of models that can be used seamlessly (most of the time) in the modeling framework

Scorer

Scorer takes a trained ModelWrapper object (or path to a pickled ModelWrapper object) and some new data and allows you to make predictions

new_data = ModelingData(class_dict=my_class_dict, ids=[1084, 3939, 9423])
new_data.build()
my_scorer = Scorer(model='~/saved_models/my_xgboost_model', data=new_data)
my_scorer.score()

Workflow

• Define some data classes
• Create a class_dict where the key is the class and the val is a list of keys to join on
• Build some modeling data
• Create a model wrapper and train a model
• Get some new data
• Pass a trained model and new data to a Scorer
• Make predictions
• ????
• PROFIT

TODO

• incorporate more robust joins in modeling_data.build() to allow for different foreign keys across data_classes

Additional Notes

xgboost package needs extra love on MacOS https://medium.com/@lenguyenthedat/installing-xgboost-on-os-x-1f63c1ed042