Machine Learning Done Wrong

Logistic Regression

Logistic regression, often used in machine learning, presents challenges in interpreting coefficients, as Cheng-tao Chu explains. He argues that interpreting these coefficients can lead to misleading insights, especially when dealing with collinear features like temperature and humidity 1. This collinearity can cause coefficients to shift unpredictably, making them unreliable for interpretation.

Your transaction amount could be very important and if you change your scale, it could be insignificant without actually adding more information or reducing any information.

--- Cheng-tao Chu

Kyle Polich agrees, noting the importance of clear communication when sharing model results with non-technical audiences 1.

   

Nonlinear Models

In addressing nonlinear data, Cheng-tao Chu highlights the advantages of tree-based algorithms over linear models. He uses fraud detection as an example, where features like billing and shipping addresses require more complex modeling than linear combinations can offer 2. Tree-based models can handle such complexity by creating decision rules that better capture the nuances of the data.

As for what algorithms are more robust for those cases, I would say some high variance model like tree based algorithm or non parametric models would be a better fit in those cases.

--- Cheng-tao Chu

Kyle Polich adds that dealing with outliers is crucial, as they can distort model performance if not properly managed 2.