cancel
Showing results for 
Search instead for 
Did you mean: 

What is Mono models with monotonicity constraints?

What is Mono models with monotonicity constraints?

Interpretability Basics

  • Screenshot 2021-07-08 102049.jpg, if domain knowledge suggest the effects should be monotonic. 

 

1 Solution

Accepted Solutions
Inactive
Not applicable

Monotonicity is when a line always slopes upwards or always slopes downwards.
It is important that models behave according to business rules. So, sometimes monotonic behavior is highly desirable. For example:

  • for credit scoring, an applicant with a higher income should always have a lower probability of a loan going bad, everything else being equal
  • for pricing, a price should always be higher when the quantity of product is higher, or when the quality of product is higher

But sometimes, due to chance, the historical data causes a feature effects line that isn't monotonic when it should be. 

colinpriest_0-1625798430198.png

In the Lending Club example above, you can see occasional upward blips in the feature effects and an upward slope at the end. This is not desirable behavior. 

colinpriest_2-1625798975936.png

So, DataRobot gives you the choice to set constraints on the slope of feature effects. When you first set up a project, before you click the start button, use advanced settings to set which features must slope upwards versus the target, and which must slope downwards.

colinpriest_1-1625798682018.png

In the screenshot above, you can see the feature effects on a model where I have applied these constraints. No more blips. No more upward slope at the end.

Colin

 

 

View solution in original post

2 Replies
Inactive
Not applicable

Monotonicity is when a line always slopes upwards or always slopes downwards.
It is important that models behave according to business rules. So, sometimes monotonic behavior is highly desirable. For example:

  • for credit scoring, an applicant with a higher income should always have a lower probability of a loan going bad, everything else being equal
  • for pricing, a price should always be higher when the quantity of product is higher, or when the quality of product is higher

But sometimes, due to chance, the historical data causes a feature effects line that isn't monotonic when it should be. 

colinpriest_0-1625798430198.png

In the Lending Club example above, you can see occasional upward blips in the feature effects and an upward slope at the end. This is not desirable behavior. 

colinpriest_2-1625798975936.png

So, DataRobot gives you the choice to set constraints on the slope of feature effects. When you first set up a project, before you click the start button, use advanced settings to set which features must slope upwards versus the target, and which must slope downwards.

colinpriest_1-1625798682018.png

In the screenshot above, you can see the feature effects on a model where I have applied these constraints. No more blips. No more upward slope at the end.

Colin

 

 

Thank you so so much! I am glad that I asked.