TrAdaBoost can predict_proba or not  #111

Activity

1. antoinedemathelin commented on Aug 10, 2023

   

   antoinedemathelin

   on Aug 10, 2023

   Collaborator

   More actions

   Hi @jnuvenus,
   Thank you for your interest in the Adapt library!
   There is no predict_proba method implemented on the Tradaboost object but you can access the predict_proba method of each estimator in model.estimators_. Here is a little example to compute the weighted average probabilities among Tradaboost estimators (Note that the authors of Tradaboost suggests to use only the last half estimators for prediction):

   import numpy as np
   from sklearn.linear_model import LogisticRegression
   from adapt.utils import make_classification_da
   from adapt.instance_based import TrAdaBoost
   
   Xs, ys, Xt, yt = make_classification_da()
   model = TrAdaBoost(LogisticRegression(), n_estimators=10, Xt=Xt[:10], yt=yt[:10],
                      verbose=0, random_state=0)
   model.fit(Xs, ys);
   
   N = len(model.estimators_)
   probas = [m.predict_proba(Xt) for m in model.estimators_[int(N/2):]]
   probas = np.stack(probas, axis=-1)
   
   weights = np.array(model.estimator_weights_)
   weights = weights[int(N/2):]
   weights /= weights.sum()
   
   weighted_avg_probas = probas.dot(weights)
   weighted_avg_probas

   Thanks to your question, we will add the predict_proba feature for Tradaboost in the next release of Adapt