Rotation Forest Regressor - Githubissues

abduskhazi / PL-Binding-Affinity-Prediction-using-ML

This repository is maintained for the documentation and coding of the MSc project @ Bioinformatics Lab Uni-Freiburg.

MIT License

1 stars 0 forks source link

Rotation Forest Regressor #5

Open abduskhazi opened 3 years ago

abduskhazi commented 3 years ago

The following tasks were completed for this issue -

Read about rotation forest models. Refered the papers
- https://arxiv.org/abs/1809.06705
- https://www.researchgate.net/publication/6806976_Rotation_Forest_A_New_Classifier_Ensemble_Method
Added the given (forked) code as a submodule.
Adapted the code for regression purpose. The given code was only written for classification problem. Please refer - https://github.com/abduskhazi/RotationForest/pull/1/files
Ran the model with the following parameters RotationForest(n_trees=100, n_features=15, sample_prop=0.5, bootstrap=True) # features = partitions here

Observations (Execution ID = 2837580558; commit = a01287573da06664d69943648a23bf023f04e4b8)

Fitting the model is costly. All features were used for training. (No multithreading/multiprocessing support in the model).
- Time taken for fitting ~25m 29s
- Validation r2 score = 0.776733213218239
- Training r2 score = 0.9672355525699042
- r2 score (shuffled protein columns) = 0.6013427042718706
- r2 score (shuffled ligand columns) = -0.014683489295441499
The accuracy is approximately same as regular random forest regression.

abduskhazi commented 3 years ago

With 1000 trees and 25 partitions:
- Validation r2 score = 0.7727133682818564
- Training r2 score = 0.9699304387683184
With 1000 trees and 15 partitions:
- Validation r2 score = 0.7775481476893422
- Training r2 score = 0.9693611573975324
With 1000 trees and 5 partitions:
- Validation r2 score = 0.7399648622707944
- Training r2 score = 0.9645953975367217