Conversation
|
I'm not sure why, but there are some problems with the c++ solver on Mac. It works perfectly on linux. I'll take a look. |
|
It's fixed! I'm having a hard time pushing the changes, but I'll do it this weekend. |
|
@aldro61 the mmitboost and mmitboost.predict has most of the code same...as we cant find the prediction value without running it, and as its iterative process hence mmitboost cant be used for prediction So is there any way to improve this? |
|
There might be some code that can be reused, maybe packaged into one or many utility functions. But before that, we should make sure that the implementation is ok. I think that the Adaboost algorithm that you implemented is for classification. The algorithm is different for regression. Scikit-learn implements the following version (Adaboost.R2) https://www.researchgate.net/publication/2424244_Improving_Regressors_Using_Boosting_Techniques. Note that according to this algorithm, the training data needs to be resampled with replacement before training each tree. Also, there seems to be a learning rate hyperparameter that can be tuned by the user. |
|
ok....i'll check that and update |
Revert "Gsoc doc"
return result including trees, B value, predictions of each tree
Merge branch 'master' of https://github.com/aldro61/mmit Conflicts: Rpackage/R/mmif.R Rpackage/R/mmif.cv.R Rpackage/R/mmit.cv.R Rpackage/R/mmitboost.R Rpackage/R/mmitboost.predict.R Rpackage/man/mmit.Rd Rpackage/src/registerDynamicSymbol.c
| upper = prediction - target.mat[,2] + margin | ||
| upper[upper < 0] <- 0 | ||
| if(loss == 'hinge'){ | ||
| cost = lower + upper |
There was a problem hiding this comment.
@aldro61 How to make sure that cost is between 0 to 1?
| #' | ||
| #' @export | ||
| mmitboost.predict <- structure(function(target.mat, mmitboost_results) { | ||
| final_scores <- weighted.mean(mmitboost_results$pred, mmitboost_results$B) |
There was a problem hiding this comment.
@aldro61 How to implement weighted median?
2 participants
No description provided.