IBM Data WH Generalized Linear

Linear regression is a long-established statistical technique for classifying records based on the values of numeric input fields. Linear regression fits a straight line or surface that minimizes the discrepancies between predicted and actual output values. Linear models are useful in modeling a wide range of real-world phenomena owing to their simplicity in both training and model application. However, linear models assume a normal distribution in the dependent (target) variable and a linear impact of the independent (predictor) variables on the dependent variable.

There are many situations where a linear regression is useful but the above assumptions do not apply. For example, when modeling consumer choice between a discrete number of products, the dependent variable is likely to have a multinomial distribution. Equally, when modeling income against age, income typically increases as age increases, but the link between the two is unlikely to be as simple as a straight line.

For these situations, a generalized linear model can be used. Generalized linear models expand the linear regression model so that the dependent variable is related to the predictor variables by means of a specified link function, for which there is a choice of suitable functions. Moreover, the model allows for the dependent variable to have a non-normal distribution, such as Poisson.

The algorithm iteratively seeks the best-fitting model, up to a specified number of iterations. In calculating the best fit, the error is represented by the sum of squares of the differences between the predicted and actual value of the dependent variable.