Is there any way to set up some rules from features in a classification model?

Assume that we want to classify an employee as someone who will be terminated or not. We found that average hourly pay and working hours are the two most significant features for the model. Is there any way that we can say how changes in these two features impact the chance?

For example, employees whose average pay is less than $20 per hour have a higher chance of quitting or those who work more than 40 hours per week have a higher chance?

Is there anyway to come up with these thresholds?

  • $\begingroup$ What type of analysis have you done? Interpretation of results depends almost entirely on what processes you applied to get those results. And what sort of outputs do you have? Was this a logit, multivariate regression, etc.? What sorts of covariates did you deem not significant enough, and how did you make that decision? $\endgroup$
    – Upper_Case
    May 6, 2019 at 16:45
  • $\begingroup$ I used pearson correlation to find significant features and then trained the model with boosted tree. I just have the probabilities for termination and considering a 0.5 threshold, I have a label of either 0 or 1. I know that looking at trees may give me an idea of the break points for features but wanted to see if there are any statistical analysis that we can use. $\endgroup$
    – Fatima
    May 6, 2019 at 17:12
  • 1
    $\begingroup$ You can have a look at partial dependence plots or ceteris paribus profiles. Or to linear regression. Your model strategy will require a very good validation strategy as it easily overfits. $\endgroup$
    – Michael M
    May 7, 2019 at 3:53


Your Answer

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