The experiments performed by
Caruana and Niculescu-Mizil were very comprehensive and beyond the
scope of my project. In my experiments, I have not calibrated any of
the algorithms using Platt Scaling or Isotonic regression.
The
authors used different software packages for different learning
algorithms. For example the authors have used the Breiman-Cutler
implementation of random forests since it produced better results. Similarly they have used SVMLight for SVMs. While conducting my experiments, I have only used the algorithms provided by Weka.
The use of different implementations and lack of calibration may have
contributed to the different results obtained by me. Since I did not
perform any calibration, boosted trees did not perform as well as expected.
Data sets Used:
The following 7 data sets were used by me
Adult, Bacteria, California Housing, Cover_type, HS(Indian pine92), letter p1, letter p2
Algorithms Evaluated:
The following algorithms were evaluated
Naive Bayes, Random Forests, Decision trees, Boosted trees, Bagged trees, Boosted stumps, ANNs, SVMs, Logistic regression and KNN.
Performance Metrics Used:
Accuracy, F-score, ROC, RMSE, APR(average precision)
For
some metrics such as accuracy higher values are desirable while for
some measures such as RMSE, lower scores are desirable. For such scores
I have inverted the score( 1-RMSE) in order to maintain uniformity
among the metrics.
Continue to "results"
Back
Back to project main page
Back to Varun's home page