4
$\begingroup$

Have you ever heard about Child-Pugh cirrhosis score? There are five features, each feature is discretized in three intervals. For each interval you get a point, eg 1 for the first, 2 for the second etc. After computing all points you sum them up and then you get a cirrhosis stages ranked from A to C. Each stage is correlated to a particular probability to survive. http://en.wikipedia.org/wiki/Child-Pugh_score

I have a DB with a many features and a label, the label is binary: ill and good. I have chosen probability estimation trees to get a white box model that estimates probabilities to be 'ill'. The model can show almost clearly the classification process, but I think that a Child-Pugh like score could be clearer than this for a medical doctor.

Is there something similar in literature?

$\endgroup$

1 Answer 1

1
$\begingroup$

One option is to take a single tree and make it small (e.g. 3-5 features). This will show a classification process that is easy to interpret.

If you want something closer to a "score", then try using a linear classifier (such as Naive Bayes or SVM). Again, use few (3-5) features and discretize them into levels if you want. They will compute something like $$Score = w_1 \cdot f_1 + w_2 \cdot f_2 + ... ,$$ where $f_i$ are feature values (1, 2, or 3 in your example) and $w_i$ is the weight for that feature (i.e. how many points you get for that feature).

The disadvantage of all this is that you need to use few features and small classifiers for them to be easy to interpret. But they may not work as well as your big classifier.

$\endgroup$

Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.