We often get into the state of blaming advanced data science models not working up to the mark without understanding the nuances of its working which surely requires a decent understanding of the mathematics behind these models. Let’s take an example of the following questions:

– Should I build a separate model for each city, or a single model for all the cities?
– What is the right metric for this problem: “finding the most influential people in human history based on Wikipedia data”?
– Why my gradient-boosting model is unable to beat a basic rule-based model?
– I have installed a library that implements a recommendation-model which optimizes for showing most relevant items to a user. How do I tweak it so that it maximizes my revenue instead?

The big theme behind it is that the black box approach often leads to poor choices at various steps involving data collection, data processing & wrangling, feature engineering, model selection, model evaluation and lot more. Knowing mathematics behind a model is no longer a luxury but indeed a necessity now. Kindly refer to an article by Pulkit Bansal for an interesting perspective about this issue.

Go to link →