0

I am trying to do a prediction with Lasso regression. When I plotted the predicted Vs. Actual, I got a very low correlation between them (shown in the figure). However, I feel that with some kind of transformation, I should be able to make the predicted values to lie on the (predicted = Actual) line.

Is this possible? Or are there any parameters in Lasso or any other regression technique, that I can tune to get better predicted results for this case?

Predicted vs. Actual with Lasso

Gihan
  • 17
  • 5
  • can you elaborate on how you fit using the lasso. theres already a runing parameter – user795305 May 11 '17 at 19:17
  • Why do you use so much space horizontally? If you cut out the strips of white space on the left & right, you are essentially left with a structureless point cloud. "I feel that with some kind of transformation, I should be able to make the predicted values to lie on the (predicted = Actual) line." - why do you feel that way? Right now, it seems like your data & model has a lot of unexplained residual variation. If you can't get a handle on that, transforming the lasso model likely won't help. And even if you do include every relevant predictor, you won't get perfect predictions... – Stephan Kolassa May 11 '17 at 19:32
  • ... because of the bias-variance tradeoff. [And some ML problems simply are hopeless.](https://stats.stackexchange.com/q/222179/1352) – Stephan Kolassa May 11 '17 at 19:32
  • Ben, I use LassoCV as implemented in sk-learn. These are my parameters of the model, 'alphas': [1, 0.1, 0.05, 0.001, 0.0005], 'copy_X': True, 'cv': None, 'eps': 0.001, 'fit_intercept': True, 'max_iter': 10000, 'n_alphas': 100, 'n_jobs': 1, 'normalize': False, 'positive': False, 'precompute': 'auto', 'random_state': None, 'selection': 'cyclic', 'tol': 0.0001, 'verbose': False – Gihan May 11 '17 at 20:02
  • It's hard to work with what you've given us. Visualization of the raw data would be helpful. – user795305 May 13 '17 at 20:17

0 Answers0