Home

Welcome to the FTRLp wiki!

FTRLp is an online, Logistic Regression with Elastic Net regularization that is very efficient at obtaining sparsity. The algorithm is particularly well suited for massive datasets with very high dimensional categorical features and has proven to perform very well in massive Click-Through-Rate prediction tasks.

In this implementation you can find a data generation object to extract, transform (separate, change and hash features) and finally yield data to the classifier one row at a time.

This code follows the algorithm by H. B. McMahan et. al. It minimizes the LogLoss function iteratively with a combination of L2 and L1 (centralized at the current point) norms and adaptive, per coordinate learning rates.

References:

• Follow-the-Regularized-Leader and Mirror Descent: Equivalent Theorems and L1 Regularization, H. Brendan McMahan
• Ad Click Prediction: a View from the Trenches, H. Brendan McMahan et. al.