Finding out which values lead Random tree to a decision

Question

I have a dataset of machines that produce plastic parts. A camera evaluates whether a plastic part was produced correctly or not (binary classification). I'm trying to figure out which factors influence a part being wrongly produced. E.g. I have different temperature values of the machine parts during the production.

I'm using a Random Forest to classify the data. The test dataset is being recognized quite well. The next step is to figure out which values lead to a wrongly produced part (e.g. when temperature > 150K: Part is broken). I've searched the internet but I couldn't find any information about this.

At the moment I'm trying a brute force method where I simply generate a test dataset where I go through different value ranges. But so far everything is classified as wrongly produced part.

Are there other methods I can use to get the values?

Thank you!

Answer 1

If your dataset was separable into a series of neat decisions, then a classification and regression tree (CART), would give you the type of solution you're looking for. The price you pay for random forest is that by reducing the variance through generating many random trees, you're also reducing the interpretability of the model significantly. Solutions that can give you local solutions are LIME or you can calculate SHAP values or feature importances, but that gives you importances in the context of the model and may not be useful for the type of decision you're looking to make.