Outlier detection in reference data using Breiman's outlier score
Detecting incorrectly labeled data in training and validation data sets is important. One of the approaches consists in computing an outlier score for every sample using the proximity measured in the trees of a random forest model after training. For every sample, one counts how many times it falls in the same leaf as other samples of the same class. The average across all trees in the forest is the outlier score.
La méthode décrite ici, proposée par Breiman, n’a pas fait l’objet d’une publication avant son dé- cès. Le détail est cependant disponible dans un rapport technique : www.stat.berkeley.edu/~breiman/ Using_random_forests_v4.0.pdf. Par ailleurs, le code Fortran développé par Breiman est aussi dispo- nible à www.stat.berkeley.edu/~breiman/RandomForests/cc_software.htm
The outlier score would be computed after training, then the outliers would be removed and a new training would be performed with the cleansed data. A specific application for computing the outlier score and adding it to the sample file as an additional attribute would be nice. Then filtering with sqlite or ogr2ogr would allow to perform the removal.