GBDT-FM

An implementation of GBDT+FM

The implementation of FM (factorization Machines) comes from Zhengruifeng (https://github.com/zhengruifeng/spark-libFM)

In this implementation, GBDT is used for feature transformation and FM optimize those 0-1 features for CTR prediction with SGD.

args for input

args（0）： HDFS file

args（1）： repeat time of algorithm e.g. 5

args（2）： number of trees of GBDT e.g. 30

args（3）： number of iteration of FM e.g. 20

args（4）： step size of FM e.g. 0.1

other parameters in GBDT e.g. depth of tree and that of in FM e.g. bias, regularization can be set mannually.

two read file method is provided

1. LibSVM Format, spark RDD labelled point

2. text format, split by "," with first column of label and rest the data