Skip to the content ForoHackers
Log inJoin

Our "machine learning feature" is a hand-written if/else tree

Machine learning
#1059 · anonymous · · Report

Not naming the employer. Just venting.

be me, joined a team that "ships ML" ask to see the model it is a file with 140 nested if statements each one has a threshold thresholds were tuned by a guy by looking at a spreadsheet he left accuracy is "good enough" according to a slide

The strange part: I benchmarked a plain logistic regression on the same data and it was only slightly better. The tree is not stupid. It is just undocumented, untested and unversioned.

Do I tell them it is a decision tree, or do I let them keep the slide?

#1604 · anonymous · · Report

Tell them, gently. Hand-written rules are a perfectly respectable baseline. What is missing is a test set, a metric and a way to see when it drifts. Put the thresholds in a config file, add 20 labelled examples as regression tests, and you have already improved it more than any new model would.

#1605 · anonymous · · Report

logistic regression only slightly better than 140 ifs

That is not a failure of the tree, that is a sign the problem is easy. Keep the simpler thing you can explain.

Sign as a returning source (a key kept only in this browser)

Sending takes a few seconds here: your browser first works out a proof of work against spam.

Reply