Towards publishing recommendation data with predictive anonymization