DGFFM: Generalized Field-aware Factorization Machine based on DenseNet

• Main Authors: Zhang, Qing-Long, Rao, Lu, Yang, Yubin
• Format: Conference Proceeding
• Language: English
• Subjects: Computational modeling; Feature extraction; Frequency modulation; Indexes; Motion pictures; Neural networks; Predictive models
• ISSN: 2161-4407
• DOI: 10.1109/IJCNN.2019.8851933

In this paper, we build a generalized field-aware factorization machine (GFFM) based on FFM, which stores different feature embeddings in multiple files separately instead of one single file. By making use of the corresponding location relationship between the field index and feature index, GFFM can significantly reduce the computation time. Also, features in GFFM are subdivided into dynamic ones and static ones. When modeling, the temporal variation of dynamic features such as fashion trends and user behavior preferences are considered to establish an accurate dynamic model based on the time window. We further propose DGFFM which uses the wide & deep framework to jointly train GFFM and DenseNet, aiming to combine the benefits of traditional machine learning methods including their faster learning speed on low-rank features and ability to extract high dimensional features. DGFFM add extra first-order identity features into the FM/FFM related models for the first time by adopting the multi-storage pattern designed in this paper. Experimental results on four real-world datasets demonstrate that GFFM can obtain higher accuracy and computational efficiency compared with the state-of-the-art methods used in recommendation and CTR tasks, and DGFFM can achieve an even higher prediction accuracy than recent DNN models.