Residuals refer to the difference between the recorded value of a dependent variable and the predicted value of a dependent variable for every row in a data set. Plotting the residual values against the predicted values is a time-honoured model assessment technique and a great way to see all your modelling results in two dimensions.

Boosting models like XGBoost, LightBoost have been some of the go to model for tabular datasets.This is because these models are highly performant and are able to learn various non-linear relationships within the datasets.However sometimes in the real world,creating an high performant model can result in a non-intuitive model. There are also requirements to include domain knowledge into the ML models.

"How confident are you that the model performance (on test data) is not below 0.75 AUC ?".I came across this question in my work.This inspired me to write this blog.

Error analysis simply means looking the misclassified data points and trying to infer why is the model making error on this subpopulation. There are some great videos/blogs about how to do error analysis for text/image data. e.g video by Andrew Ng.

While browsing Twitter , I came across this tweet by Abhishek.

All practioners of ML know that feature selection is important and is helpful to reduce the unwanted signal in your dataset.

For most of the problems in real world, we have to optimise on two or more objective functions . Eg in the case of machine learning problems we may have to optimise on two or more objective function like maximise accuracy vs increase interpretability . Almost always these objective functions will be orthogonal to each other.

Hello there, this is my personal blog to note down my thoughts both personal and professional. I plan to write about technology, the books I have read, the side project I am working on etc. I also plan to share my ideas freely here, so that others can discuss it with me and contribute as well .