Effective Amazon Machine Learning
上QQ阅读APP看书,第一时间看更新

Assumptions for multiclass linear models

For a linear model to offer reliable predictions, predictors must satisfy a certain number of conditions. These conditions are known as the Assumptions of Multiple Linear Regression (http://www.statisticssolutions.com/assumptions-of-multiple-linear-regression/):

  • Linear relationship: The predictors should have some level of linear relationship with the outcome
  • Multivariate normality: The predictors should follow a Gaussian distribution
  • No or little multicollinearity: The predictors should not be correlated to one another
  • Homoscedasticity: The variance of each predictor should remain more or less constant across the whole range of values

Of course, these assumptions are seldom verified. But there are ways to transform the data to approach these optimal conditions.