基于编码器增强的DeepFM推荐模型  

DeepFM Recommendation Model Based on Encoder Enhancement

作　　者：胡克富 王卫东 HU Kefu;WANG Weidong(School of Computing,Jiangsu University of Science and Technology,Zhenjiang 212100)

机构地区：[1]江苏科技大学计算机学院,镇江212100

出　　处：《计算机与数字工程》2025年第1期158-163,共6页Computer & Digital Engineering

摘　　要：点击率(CTR)预测用于预测用户点击推荐物品的概率,是推荐系统和在线广告的关键任务。CTR预测模型中缺乏对高效的特征交互以及对特征交互的可解释性。论文提出了一种具有编码器的EnDeepFM推荐模型(Deep Neural Networks with Encoder Enhanced Factorization Machine,EnDeepFM),通过Transformer编码器对嵌入特征进行编码,利用双线性函数生成不同特征对的不同特征相似度,从编码器生成的嵌入有利于进一步的特征交互。最后,在真实数据集Criteo和MovieLens上进行对比实验,实验结果表明所提出的算法比DeepFM模型具有更好的预测性能。Click-through rate(CTR)prediction is used to predict the probability of users clicking on recommended items,which is a key task for recommender systems and online advertising.Efficient feature interactions and interpretability of feature inter⁃actions are lacking in CTR prediction models.This paper proposes an EnDeepFM recommendation model with an encoder(Deep Neural Networks with Encoder Enhanced Factorization Machine,EnDeepFM),which encodes the embedded features through the Transformer encoder,and uses the bilinear function to generate different feature similarities of different feature pairs.Embeddings generated from the encoder facilitate further feature interactions.Finally,comparative experiments are conducted on the real datas⁃ets Criteo and MovieLens,and the experimental results show that the proposed algorithm has better predictive performance than the DeepFM model.

关 键 词：CTR预测 编码器 深度学习 DeepFM 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]