Vertical Federated Learning for Higher-Order Factorization Machines

Abstract

In the real world, multiple parties sometimes have different data of common instances, e.g., a customer of a supermarket can be a patient of a hospital. In other words, datasets are sometimes vertically partitioned into multiple parties. In such a situation, it is natural for those parties to collaborate to obtain more accurate prediction models; however, sharing their raw data should be prohibitive from the point of view of privacy-preservation. Federated learning has recently attracted the attention of machine learning researchers as a framework for efficiently collaborative learning of predictive models among multiple parties with privacy-preservation. In this paper, we propose a lossless vertical federated learning (VFL) method for higher-order factorization machines (HOFMs). HOFMs take into feature combinations efficiently and effectively and have succeeded in many tasks, especially recommender systems, link predictions, and natural language processing. Although it is intuitively difficult to evaluate and learn HOFMs without sharing raw feature vectors, our generalized recursion of ANOVA kernels enables us to do it. We also propose a more efficient and robust VFL method for HOFMs based on anonymization by clustering. Experimental results on three real-world datasets show the effectiveness of the proposed method .