Prediction Probability Threshold???

[Vedant-R]

Hey @mrdbourke

I was going through the video of the course - 247. Visualising our model's most wrong predictions.

Is there a way where we can decide what threshold we should set for prediction probabilities, like for example in the course we have assumed 0.5, would there be some technique by which we would be able to decide whether it should be 0.5 or 0.7 or 0.3?

Thank you in advance.

[mrdbourke]

Hey @Vedant-R,

Yes, there is a way. The first I'd start with is experimental.

What level do you need your model to perform at?

I'd try with different thresholds and see whether or not your model is still feasible.

For example, try 0.5, then 0.3 then 0.7 and see which gets the best performance for your specific use case.

The second is looking into more automated methods.

I'd search for "how to set probability threshold for machine learning models" and see what you find.

But I'd start with the first way (experimenting) to get an idea of what happens when you change the threshold.

[Vedant-R]

Thanks @mrdbourke for your answer.