baigang/spark_multiboost

An implementation of the multi-class/multi-label classifier, of which the training is carried out using AdaBoost.MH on Apache Spark.

spark_multiboost's Issues

Binary classifier in mllib.classification as base learners

As describe in the paper, \phi(x) can be any binary classification model. So we may use logistic regression and support vector machines implemented in mllib.classification for this component. It is straightforward to conver the label vector into a single 0/1 label by checking the sigh of the weighted label sum.

Update to Spark 1.5.1

Current version depends to old lib.
It is an issue when trying to use with a project depending of last version of Spark

Better documentations.

We need a compelling description of the project.

Use 0/1 label instead of -1/+1

By specifying negative samples as 0 instead of -1, we can better leverage the sparse representation of Vector. After we parse the data points, we can simply convert it to -1/+1 for coherence in the algorithm.

Documentation

Is there some examples existing outside of the tests? With prediction of multilabel data?

Hamming Tree implementation

Hamming tree is a slightly ``stronger'' weak learner reusing some mechanics from Decision Stump.