Phase 1 Analysis: local computation, replicated, no aes encryption
Our linear regression, logistic regression, and tree ensembles (both classifiers and regressors) seem to be robust to a wide range of inputs and varying parameters.
Our's and Sklearn outputs match up to 4 or more decimal points for linear regression, up to 2 decimal points for logistic regression, and up to 4 or more decimals for tree ensembles
Ensemble trees outputs have the closest match to Sklearn, but are much slower than linear or logistic regression. One ensemble with only 10 trees (Sklearn's default is 100 trees) and 10 features runs about 10 minutes locally without aes encryption.
The lower precision of logistic regression's output probabilities could be driven by sigmoid.
The goal is to test if Linear Regression, Logistic Regression, and Tree-based models maintain acceptable arithmetic precision.