baura / machine_learning Goto Github PK

#This is my project for Machine Learning course from the Data Science specialization of Coursera.

##I proceeded as follows:

1. Download the csv files and call them training and testing, respectively;
2. Correct the values like #DIV0!, turning them into NA's;
3. Divide the training set into train_train, for 70% of its observations, and test_train;
4. Check how many columns have more than 90 % of observations missing and kick them out. I made sure that most of these NA's values are such for most of these columns;
5. Looked at the distribution of each numeric variable of those remained (55), subdivided by their classe value;
6. Performed principal component analysis, as many variables showed high collinearity;
7. Played around with classification methods on the preprocessed train_train subset and looked at the accuracy on predicting the outcome classe on test_train (preprocessed with the same rotation as in train_train);
8. All those methods with an accuracy higher than 0.5 were then used in a combined predictor that worked as a "majority vote". Recall that, having 5 levels of classe, a random predictor would have 0.2 accuracy;
9. Use this combined predictor to predict classe in testing.