2
$\begingroup$

I have a dataset of genes with features that describe the genes at different scales (epigenetic, protein, cell, drug data etc. all numeric data). I use this dataset in supervised ML with a xgboost regression model scoring the genes between 0 to 1 (with 1 being most likely to cause a disease and 0 being least likely).

However, for my data collection, I collect data such as the number of epigenetic sites per gene. Features like this are affected by the length of the gene and this will maybe skew prediction (e.g. genes which are larger will have more hits/sites potentially tricking the model into thinking they are more likely disease-causing genes only due to their bigger size).

To address this I give the model gene length as a feature, hoping it will control for any positive correlation with gene length. However, I'm having blackbox issues proving this is what's happening, I use SHAP and Friedman's H-statistic to show how gene length interacts with other features but these 2 have conflicting results.

I am looking to see if there is some other way I can use gene length to regulate the other features, whether as a pre-processing step before or still as a feature in the model itself. I have a biology background and I'm teaching myself stats/ML, so I'm not sure if this is possible - but is there a way to use a feature as a covariate or mediator of other features?

$\endgroup$

1 Answer 1

1
$\begingroup$

I believe normalizing your data could help your problem: if you scale all your dataset values to be between 0 and 1, the problem you described stops existing. If you don't know how to do this, you can for example read this article, but it's very very easy to apply and it will most likely solve your problem!

$\endgroup$
2
  • $\begingroup$ Thank you for this, with this in mind I wouldn't need to purposely add gene length as a feature at all right? $\endgroup$
    – DN1
    Commented Aug 26, 2020 at 10:54
  • $\begingroup$ @DN1 You wouldn't need to add it for "mediation" purposes; however, I believe it would probably be a valueable feature for your dataset (I'm from biological engineering). But, keeping this in mind, adding too many features might lead to overfitting, so it's all a matter of balancing that $\endgroup$
    – Johanna
    Commented Aug 26, 2020 at 11:05

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.