BeSci + ML

An ambitious project of combining machine learning and behavioral science --- Behavioral Machine Learning (not sure what it is, we're trying to invent it). Currently this repo consists of several disjoint pieces of code split into the following folders:

1. besci_loss compares the conventional ML to the behavioral ML on NHANES DBQ-dataset. This whole project kinda failed because conventional ML performs better and I'm not sure what to do about it.

2. coin_flips simulates the data of flipping coins described by dummy features and analyzes the model trained on that data. In particular, it shows a problem of using shap on one-hot encoded data.

3. nhanes_dbq_synthesis uses NHANES DBQ-dataset and builds a simulator using additional behavioral rules. In this formlulation the input consist of 6 demographic features: age, gender, race, income, education, marital status. The output is a 5-dimensional probability vector of food preferences. The resulting simulator is non-deterministic and synthesizes data that alignes with the underlying behavioral rules.

4. nhanes_dbq_explanation uses 12 years of NHANES DBQ-surveys to learn people's food preferences from 6 demographic features and then analyze importance of these demographic features on the food choice preferences.

5. brfss contains the data from BRFSS dataset, which has 450,000 rows and 358 columns. No idea what to do with all this data but it's here for future inspiration.