RSM412H1 Lecture Notes - Lecture 5: Logistic Regression, Linear Regression, Predictive Power
Document Summary
Classification models have qualitative target (categories: y is not normally distributed, useful for yes/no questions, some classification problems are linear. Can also use on more than 2 categories. Observations in different categories can be separated by a linear model. Need a curve to classify between different categories. Can use neural networks for this: classification problems are often non-linear. Logistic regression: if representing a binary variable using a linear relationship: We may see probabilities less than 0 or greater than 1. Linear regression does not accurately predict probability that y=1. Will fall as long as more predictors are added, so need other measures to adjust for number of predictors. Model comparison: need to look at, how complicated is trained candidate model after training, model performance, how well model does on training dataset, model complexity, pick model that minimizes aic or bic. Lower limit of 0 and upper limit of positive infinity. Range of negative infinity to positive infinity.
