Bridging the Gap: From Statistical Distributions to Machine Learning Loss Functions

When I consulted researchers on which statistical analysis to use for their data, a common first step was to think about the distribution of the target variable: Is it a count, like the number of emails received within an hour? Poisson distribution it is. Is the target variable a binary category, like whether or not a customer bought a product? Use the Bernoulli distribution. There’s a long list of such distributions and too many ways you can modify them. Like truncated versions, mixtures, and so on. It’s a bit like a game where you can customize your character: An easy way to get stuck before you even get started. Thanks for reading Mindful Modeler! Subscribe for free to receive new posts and support my work. Anyways, based on the presumed distribution the next step was to fit a GLM or GAM, which ultimately meant maximizing the likelihood of the chosen distribution. In contrast to classic statistical modeling, supervised machine learning seems pretty dumb. Sophisticated discussion of

When I consulted researchers on which statistical analysis to use for their data, a common first step was to think about the distribution of the target variable: Is it a count, like the number of emails received within an hour? Poisson distribution it is. Is the target variable a binary category, like whether or not a customer bought a product? Use the Bernoulli distribution. There’s a long list of such distributions and too many ways you can modify them. Like truncated versions, mixtures, and so on. It’s a bit like a game where you can customize your character: An easy way to get stuck before