1
$\begingroup$

I've been playing with a problem around matching product names. I've trained a model based on a variety of different features (numbers only, Levenshtein distance, pulling out package sizes, brands, etc.). The model is spitting out predictions of match vs. no match (1 or 0). Ultimately, I understand that this is giving me what I want... it either matches or doesn't match.

What is the conventional wisdom around getting a match score? Is it simply adding up all of the feature scores and dividing, essentially an average? At the end of the day, I want to send in a product to an API and receive a list of the highest probable matches if the algorithm really can't get me to a "match".

$\endgroup$

1 Answer 1

1
$\begingroup$

If the model is a soft classifier (i.e. it predicts a probability before converting it to a class), then the simple option is to use the underlying probability (e.g. with a predict_proba function).

Another option is to directly train a regression model: in the training data, any instance with class "match" is represented as 1 and any instance with class "no match" is represented as 0. This way the model directly predicts a numeric value which can be used as a score.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.