The purpose of this note is to provide an introduction to the Generalized Linear Model and a matricial formulation of statistical learning derived from the class of exponential distributions with dispersion parameter. We assume that the reader is comfortable with linear algebra and multivariable calculus, has an understanding of basic probability theory and is familiar with supervised learning concepts.
A number of regression and classification models commonly used in supervised learning settings turn out to be specific cases derived from the family of exponential distributions. This note is organized as follows:
- Section 2 describes the family of exponential distributions and their associated Generalized Linear Model. The family described in  counts a significant number of distributions including e.g., the univariate Gaussian, Bernoulli, Poisson, Geometric, and Multinomial cases. Other distributions such as the multivariate Gaussian lend themselves to a natural generalization of this model. In order to do so, we extend the family of exponential distributions with dispersion parameter  to include symmetric positive definite dispersion matrices.
- Section 3 derives the Generalized Linear Model Cost Function and its corresponding Gradient and Hessian all expressed in component form. We derive the expressions associated with the general case that includes a dispersion matrix. We also derive simplified versions for the specific case when the dispersion matrix is a positive scalar multiple of the identity matrix.
- In Section 4, we limit ourselves to distributions whose dispersion matrix is a positive scalar multiple of the identity matrix. These are precisely the ones described in . We express their associated Cost Function, Gradient and Hessian using concise matrix notation. We will separately analyze the case of the multivariate Gaussian distribution and derive its associated Cost Function and Gradient in matrix form in section 7.
- Section 5 provides a matricial formulation of three numerical algorithms that can be used to minimize the Cost Function. They include the Batch Gradient Descent (BGD), Stochastic Gradient Descent (SGD) and Newton Raphson (NR) methods.
- Section 6 applies the matricial formulation to a select set of exponential distributions whose dispersion matrix is a positive scalar multiple of the identity. In particular, we consider the following:
- Univariate Gaussian distribution (which yields linear regression).
- Bernoulli distribution (which yields logistic regression).
- Poisson distribution.
- Geometric distribution.
- Multinomial distribution (which yields softmax regression).
- Section 7 treats the case of the multivariate Gaussian distribution. It is an example of an exponential distribution with dispersion matrix that is not necessarily a positive scalar multiple of the identity matrix. In this case, the dispersion matrix turns out to be the precision matrix which is the inverse of the covariance matrix. We derive the corresponding Cost Function in component form and also express it using matrix notation. We then derive the Cost function’s Gradient, express it in matrix notation and show how to to minimize the Cost Function using BGD. We finally consider the specific case of a non-weighted Cost Function without regularization and derive a closed-form solution for the optimal values of its minimizing parameters.
- Section 8 provides a python script that implements the Generalized Linear Model Supervised Learning class using the matrix notation. We limit ourselves to cases where the dispersion matrix is a positive scalar multiple of the identity matrix. The code provided is meant for educational purposes and we recommend relying on existing and tested packages (e.g., scikit-learn) to run specific predictive models.