data-science-repo

Welcome to my data science repo.

Projects

1. heart disease - predict if a new patient will have heart disease
   Description: supervised machine learning algorithm - K nearest neighbor (KNN)
   Results: 82% accuracy

2. credit card customer segmentation - segment credit card customers based on spend and attributes
   Description: unsupervised machine learning algorithm - k means clustering
   Results: Segmented the data into clusters and provide commentary on similar characteristics of each cluster to deploy targetted marketing strategies.

3. predicting insurance costs - construct a predictive model for total medical insurance cost
   Description: predict medical cost to allow hospitals to predict revenue and plan procedures needed by patient population - linear regression model
   Results: regression coefficients of 'age', 'bmi', and 'is_smoker' and their efffects on charges

4. stochastic gradient descent - predict best time to go to the gym to have no wait time on machines
   Description: 60,000 observations gym dataset with timestamp, number of people and other determinent characteristics - stochastic gradient descent on linear regression
   Results: Best time to go: 5-8 am on Saturdays, Fridays, Tuesdays. Busy times are around January (New Years) and August (school semester begins). ~9am, ~5pm, ~11pm are busiest times.

5. logistic regression model in python - classify heart disease that uses logistic regression model
   Description: predict if a patient has heart disease by using a logistic regression classification model. Data: Cleveland Clinic Foundation
   Results: Test accuracy 75%, sensitivty was 79% and specificity was 72%.

6. decision tree and random forest model - predict productivity performance of working teams.
   Description: find features that help increase employee productivity. Data: UCI Machine Learning repository - garment production
   Results: Increasing incentives would help boost productivity as it plays a vital role and was found to be the root of every tree we produced.

7. optimizing ml models - machine learning optimization tasks to improve/optimize a model to predict the extent of fire damage to a forest
   Description: Walk through tasks for optimizing ML models (Feature engineering, model selection, cross validation, regularization). Data: UCI Machine Learning repository - forest fires
   Results: Among candidate models, the backward selection using two features performs best, with an average MSE of -2.17. However, it's on log-scale so this suggest that predictions are off by a magnitude of 2. On the surface, this suggests that the models are not good predictors. However This problem is known to be a difficult one. The extreme skew in the outcome hurts many of the assumptions needed by linear models. This demonstrates that machine learning is not a universal fix and that several problems in data science have chararateristics that makes prediction difficult.