High-Order Factorization Machine Based on Cross Weights Network for Recommendation

Factorization machines are important methods for discovering complex and evolving relationships among data entities by crafting combinatorial features automatically. Because of good data representation ability, recently Deep Neural Networks (DNNs) have been applied to factorization machines to learn high-order interactions. However, the interactions with DNNs are generated implicitly. In this paper, we propose a novel factorization machine named High-order Cross Factorization Machine (HCFM), which is an efficient and explicit high-order feature interaction factorization machine. The main component that we design for HCFM is an explicit cross network: Cross Weights Network (CWN). CWN considers both the weights of different feature combinations and interaction orders, which can capture the inherent correlations of real-word data. In CWN, cross and compression layers can learn different weights of feature interactions efficiently. Different weights of interaction orders can be learned by the weights pooling layer. CWN not only retains the important interaction features but also avoids the complex interaction computations. Comprehensive experiments are conducted on three real-world recommendation datasets to verify the validity of the model. The experiment results have demonstrated its superiority over the other state-of-art algorithms, in terms of both accuracy and space complexity.

Abstract licence: CC BY-NC-ND

A person counts once a day per paper. We keep no cookie and no address: a reader is a hash that changes every day. Any month under five reads shows no number.