0
$\begingroup$

Currently I am working with a biological dataset with a range of 0-to-1 to do a multi-task regression with Deep Learning. However, this dataset has an empty gap in the range 0 to 0.2 (however there are several datapoints which are 0s) which can't be interpolated due to biological reason. In addition, there is also high tendency of a skewed data around 0.3. Until now I have only tried to train my model with standard HuberLoss and MSELoss and tried several type of models. However, I still don't get "good" prediction result with pearson correlation of only (~0.65). I came across an article about weighted MSE but most of them talking about classification problem rather than regression. In addition, I read an article about imbalanced dataset in regression but seems like they did interpolation for the data.

My question: is there other method to improve my model further? If so how could I approach it?

Many thanks in advance!

Edit: I had tried simple weighting for my data based on binning of the values' frequency but it didn't help much.

$\endgroup$
4
  • $\begingroup$ I don't know which correlation you refer to, but using this to quantify your model's performance is most likely a bad method. $\endgroup$ Commented Apr 14 at 15:50
  • $\begingroup$ Thanks for your response. I am using pearson correlation, and my expectation is that my predicted value will be linearly related into the true value. I know that there are other metrics like MAE, MSE(used for loss_fn), etc. but since the value range from 0-to-1 it will "look-like" small. Hence, I decided to use perason correlation with a graph for checking my result. $\endgroup$ Commented Apr 14 at 19:25
  • $\begingroup$ Folks will be better able to offer advice if they better understand the Business Domain. You were rather vague about what X and y represent, and whether 0.1 corresponds to “infeasibly high energy” or “toxic” or something else. The more we understand about your experimental situation, the more likely we can help. $\endgroup$
    – J_H
    Commented Apr 14 at 19:57
  • $\begingroup$ Really thanks for the advice @J_H. Adding the context, I am dealing with "RNA modification" where the modification occurs "co-transcriptionally". My hypothesis is that "RNA sequence only" is enough to say that "hey (protein X)" you need to modify this site as much as y%. Back to my data, 0 means no modification, 1 shows that the site is fully modified. However, to prevent false-positive from the pipeline, any site with <0.1 will be treated as 0s. $\endgroup$ Commented Apr 14 at 20:10

0

Your Answer

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