Fast Differentially Private Matrix Factorization

PDF

Differentially private collaborative filtering is a challenging task, both in terms of accuracy and speed. We present a simple algorithm that is provably differentially private, while offering good performance, using a novel connection of differential privacy to Bayesian posterior sampling via Stochastic Gradient Langevin Dynamics. Due to its simplicity the algorithm lends itself to efficient implementation. By careful systems design and by exploiting the power law behavior of the data to maximize CPU cache bandwidth we are able to generate 1024 dimensional models at a rate of 8.5 million recommendations per second on a single PC.
Submitted 6 May 2015 to Learning [cs.LG]
Published 7 May 2015
Updated 7 May 2015
Subjects: cs.LG cs.AI
http://arxiv.org/abs/1505.01419
http://arxiv.org/pdf/1505.01419.pdf

0 comments