Proposals

My ideas for my final project are...

1. Use Python data analysis packages such as Pandas, Numpy, and Statsmodels to measure price elasticity of grocery items (this is from my work)

2. Create a pull request to sklearn to enable individual attribute level regularization coefficients for generalized linear models. Currently, a scalar value is provided so that all attributes have the same regularization parameter. The change would enable the Sklearn API to either take a scalar value or an array of numbers of equal length to number of attributes in the model.

3. Build a flask app to output visualization of the beta distribution and output statistics for Bayesian AB testing

Execution plan

Proposal 2 was chosen to be implemented.

By the end of the following weeks I expect to have the following finished/implemented...

Week 3 - Finished execution plan and opened issue in issue tracker in github scikit-learn/scikit-learn#22350

Week 4 - Finish writing functionality code as well as code documentation

Week 5 - Finish testing feature and writing unit tests

Week 6 - Aggregate all testing material and submit for review on Sklearn github

Week 7 - Create sample data analysis and simple webpage to showcase changes API

Week 8+ - Finish webpage API and create youTube video

Actual project work

1. Sklearn pull request

Pull request was reviewed by three scikit-learn core devs and was approved but delayed due to argument over API implementation

Pull request: scikit-learn/scikit-learn#22485

Remote branch: https://github.com/xiaowei1234/scikit-learn/tree/glm_alpha

Sklearn issue: scikit-learn/scikit-learn#22350

2. Modeling API

Creates an API that will run models according to range of regularization penalty values on one of the variables using the RAND Health Insurance Experiment Data https://www.statsmodels.org/devel/datasets/generated/randhie.html

Able to run either synchronously or using parallel processing. Will store coefficients of a variable that has a constant penalty term of 1 and coefficients of another variable with different penalties. Will also store the run time of running the sequence of models with a different penalty for each.

Code

Sklearn pull request proofs

proof.py - comparing results of sklearn regression versus statsmodels regression using the same regularization penalties and output to log file of comparisons

files/proof_output.log - the output of proof.py

proof2.py - more proof of correctness using manual calculations as suggested by sklearn devs

API

main.py - the app which has get, two puts and a delete

modelresults.py - ModelResults class which stores model results from one series of penalties

database.py - Database class that stores all the ModelResults in a dictionary for the app

run_models.py - synchronously run models as well as get data and timeit functions

mp_models.py - use multiprocessing to run the models

tests/test_main.py - pytest unit tests for the app

plot

files/MP:_False....png saved plot from plotting. I couldn't get plotting to work with FastAPI and was running short on time. This is the plot produced by a unit test.