After we trained a deep learning model for regression problem, how to refine it? What are the possible reasons for its poor performance?
If you perform a linear regression model and it doesn't fit it probably means the relationship is not linear. You can add additional variables or addition terms (2nd or third or higher order polynomials or other functions) and you will usually get a pretty good fit that should work within the range of data it was developed from but should not be extrapolated very far outside of that range.
What are you trying to regress? Is a kinetic reaction rate form appropriate for this system? Is a polynomial good enough?
If you actually know the kinetics/equilibrium etc, that defines the problem you are trying to solve you should use the correct relationships with as few extra "fudge factors" as you can get away with to accurately match the observed data.
Could you specify the problem in more detail and unambiguously. In its current form, one can only guess. Do you mean neural networks?
Dear Anran Ran,
Get a plot of the residuals of the regression and look at the shape of the cloud of residuals compared with the line of the regression. The residuals should be regularly spread out along the line. If they have, for example, a curved shape in relationship to the line, replace your variable x by x squared and run the regression again. Look at the shape of the cloud of residuals again. Depending on the shape of the cloud of residuals there are diferent types of transformation of the original variable x to improve the fit.
A good reference on the matter is: Draper, N.R. and H. Smith, Applied Regression Analysis, John Wiley and Sons, Inc, New York, 1966.
- Badges
- Science topic
- Similar topics
- Organizational Psychology
- Training