RSM412H1 Lecture Notes - Lecture 3: Generalized Linear Model, Central Limit Theorem, Polynomial Regression
Document Summary
There is significant overlap between statistical learning and machine learning. Machine learning can be grouped into supervised and unsupervised. For discrete label variables: robustness tests if predictor is significant. Is =0: accuracy looks at the mse of the test data. Scatterplots are not able to capture the relationship between qualitative predictors and dependent variables. For qualitative predictors with only two levels, we can use a dummy variable. For qualitative predictors with more than two levels, need additional dummy variable for each additional level. Can accommodate non-linear relationship through various transformations to variables and to regression model. As degree increases, model becomes overly flexible and oddly shaped. Can relax normality assumption when n is large. Errors do not need to follow normal distribution because of central limit theorem. Clt ensures sampling distribution of estimates will converge toward normal distribution as n approaches infinity when: If there is non-normality in error: a. b. Relationship between x and y may not be linear.
