Field-Aware Factorization Machine Implemented by Java, with An Experiment Using Criteo Dataset.

Get The DataSet

refer to: https://github.com/guestwalk/kaggle-2014-criteo/blob/master/README, 'Get The Dataset' part.

Run main_script.sh

feature_engineer --> split train/validation set --> subsample --> change to libffm format. then run libffm:

java -Xmx65g -jar ffm.jar <eta> <lambda> <iter> <factor> <norm> <rand> <trset> <vaset>

eta: used for learning rate lambda: used for L2 regularization iter: max iterations factor: latent factor num norm: instance wise normalization rand: use random instance order when training trset: train set vaset: validation set

Experiment Results:

norm and rand only affect training speed. best eta is about 0.1, bigger eta hurt validation logloss, smaller eta get slow convergence.

when eta=0.1, factor=4, iter=10: lambda 0.00000 0.00001 0.00010 0.00100 0.01000 0.10000 best_logloss 0.45061 0.44930 0.44951 0.46919 0.58700 0.69321 convergence iter 3 5 10 10 10 10(very slow) convergence iter 10 means not oberserve convergence.

when lambda=0.0001, eta=0.1, iter=30: k=4,8,12 doesn't affect best_logloss and convergence iter.