Random forest risk and reason to churn/terminate

[andrasszabo]

Hello,
I’m relatively a newbie in KNIME and ML, however, I have successfully built several workflows now, and understand the basics.
My problem might be a bit more advanced for my level, however, based on what amazing stuff you can do with KNIME, I thought this should be possible.

My workflow uses random forest to predict which employees will leave the company. It works as it should with a solid level of accuracy, but it would be even better to include the percentage of how likely an employee will resign. I only have the prediction and the confidence for now and I didn’t find any post or workflow which does this.
Also, I was able to extract the importance of each feature, but somehow still missing why each line has that prediction, so why it’s 1 and why 0.

I would really appreciate if you could help me with this.
Best,
Andras

[mlauber71]

You can get a 0-1 score typically by activating an additional setting in most predictor nodes.

If you are interested in interpreting model performances beyond the true/false prediction you can find an example here:

Looking for options to evaluate a decision tree

One ‘brutal’ way could be to convert the decision tree rules into SQL and then use the SQL code. Or use the ruleset to make some kind of If then rules. Not very elegant I have to admit. [27] [35] kn_example_decision_tree.knwf (510.8 KB)

Random Forest Model Not Working

I attached a working example of a Random Forest with H2O.ai nodes and a 0/1 Target from a Kaggle competition (please note the example only uses the first 1.000 lines, you should remove this restriction or use your own data for your own purpouses). In order to get an idea about the quality of you model you might want to read this article: I always advise not just to use a Scorer but rather a metric like Gini/AUC to see the relative quality of the model. In the end the ‘score’ is not a magic…

[andrasszabo]

Thank you very much for this, exactly what I needed!

Now somehow if I load my new set of data to the learner node there are only 1 positive prediction and about 99.5 % is 0.

[mlauber71]

“prediction” is a relative business. The model basically gives you a rank and out of convenience the group with more % gets the ‘prediction’ especially if your targets are highy imbalanced this might not be a good approach.

You might have to decide yourself where the threshold is. There are measures and statistics that can help you. You can calculate Cohen’s Kappa and F1 score for various score groups and decide which gives the best balance of predictions.

This also depends on your business case. What happens if your model is wrong. Some misplaced marketing calls or someone getting hurt because of the wrong treatment.

Also you might balance your target with SMOTE algorithm.

Also this article might give you further insights.

Imbalanced Classification Problems in R