Interpretability v/s Flexibility for different statistical learning methods

 

(Source: An Intoduction to Statistical Learning with Applications in R written by James G., Witten D., Hastie T. and Tibshirani R. and published by Springer Text in Statistics. This is a part of my machine learning specialization course from CU Boulder.)

    I learned from this topic that we can choose a method from multiple statistical learning topics based on our goal. This book shares an example saying that if we are interested in predicting the accurate values in the stock market and not much concerned about inference (how each predictor variable is associated with the output), we can use high-flexibility approaches like deep learning. However, if we are concerned about the inference, we can use high-interpretability models like Lasso or Least-squares. The book also states that sometimes the high-interpretability models give more accurate results.

Birefringence

    Aside from dispersion, materials usually have the same index of refraction for all orientations of the electric field in the light. However, some materials, usually nearly pure crystals, have two indices of refraction for the same wavelength of light. The difference is due only to the orientation of the light's electric field. Therefore, light with its electric field oscillating in the vertical direction experiences a different index of refraction from the index of refraction for an electric field oscillating in the horizontal direction. This is called birefringence. (Source: Optics for Dummies by Duree)

Variables in a Linear Regression model (machine learning)

There are two types of variables in a dataset in general:
1) Real valued (ex.: 1,2,3,4)
2) Categorical (ex.: Blood Group)
    Categorical variables have two sub-types:
    a) Ordinal (ex.: Age group 25-55, or Grades A-D)
    b) Non-ordinal (ex.: Male/Female, Plant/Animal)

    It is comparatively easy to build linear regression model around real-values and ordinal-categorical variables that can relate to regression. However, nonordinal-categorical variables pose a problem for a "regression" model. This is  because the regression equations,

y = a0 + a1X1 + a2X2 +....

cannot rely on their arbitrary nature to predict output values. Nonetheless, these nonordinal categorical variables can always be transformed into real-valued or binary forms to make use of them in a regression problem.