R2 dips after tuning different ways in which the feature can bucketed

[anjineyulutv]

We can plot the loss curve against different empirical bin size and interpolate in the loss curve which is the optimal bin size @likespeanuts .Let me know if you need more clarifications

[anjineyulutv]

This approach involves the following steps:

Define a range of bin sizes to evaluate. For each bin size: a. Bin the feature(s) using the specified bin size. b. Train a machine learning model (e.g., regression or classification) using the binned features. c. Compute and store the loss (e.g., mean squared error for regression, log loss for classification) on a validation set. Plot the loss curve against the bin sizes. Interpolate the loss curve to find the bin size that corresponds to the minimum loss.

[likespeanuts]

okay @anjineyulutv i will look into it il let you know if you need more clarifications..

[likespeanuts]

@anjineyulutv I think im done

[likespeanuts]

i maximized r^2 and minimized MSE

[anjineyulutv]

So does it improve R2 @likespeanuts

[likespeanuts]

somehow it doesn't i used Decision Tree Regressor fort this cause clade recommended it il try and switch up models maybe that might help..

[anjineyulutv]

Please try to not change the algorithm for now. Try getting the optimal bin size for linear regression only

[likespeanuts]

@anjineyulutv i have updated it, r^2 value hasn't improved still.. i need to review the code and will update further tomorrow..

[anjineyulutv]

@likespeanuts I want you to think on the lines of, given R2 value by business ,get a bin size ,via sort of backpropagation. Think about it. Please take time :) You can also think about variable size binning. You can think like cut the feature into sections based on the curvature say into N curvatures and bin each curvature into M bins and assign global bins from 1 to M*N .Let's see how it goes