基于用户行为特征的深度混合推荐算法

A Deep Hybrid Recommendation Algorithm Based on User Behavior Characteristics

现有大多数推荐方法学习的是每个特征的固定表示,然而用户行为偏好随上下文特征发生变化,特征在上下文中具有不同的重要性,因此,特征的固定表示造成模型给出的推荐结果不准确。为解决此问题,本文提出基于深度因式分解机并融合信息提取单元和交叉网络结构的混合推荐模型(deep and cross factorization machine information extraction unit,IEU-DeepCFM)。首先,自注意力机制和上下文信息提取器组成的信息提取单元模块对不同上下文中的每个特征学习上下文感知特征表示;然后,利用深度交叉因式分解机在提取用户低、高阶特征的同时来挖掘用户更多的显式交叉信息;最终实现对用户行为特征的点击率预测。在MovieLens电影数据集和Avazu广告点击率数据集上进行消融和对比实验,结果表明,本文所提出的模型在AUC和LogLoss指标上均得到提高和改善,证明了该模型的合理性。

The hybrid recommendation model,named IEU-DeepCFM(deep and cross factorization machine information extraction unit),is proposed in this paper,which is based on the deep factorization machine and integrates the information extraction unit and cross network structure.In the proposed model,a fixed representation of each feature is learned by most existing recommendation methods.However,it is recognized that user behavioral preferences change with contextual features,and features have different importance in different contexts.Therefore,inaccurate recommendation results may be caused by the fixed representation of features given by the model.To address this issue,the information extraction unit module is introduced,consisting of a selfattention mechanism and a contextual information extractor.This module learns context-aware feature representations for each feature in various contexts.Subsequently,a deep cross factorization machine is employed to mine low-and high-order features of the user.This enables users to receive more explicit cross-information,ultimately leading to click-through rate predictions based on user behavioral characteristics.The results of ablation and comparison experiments conducted on the MovieLens movie dataset and the Avazu advertising clickthrough rate dataset demonstrate the improvement in both AUC and LogLoss indicators achieved by the proposed model.This confirms the rationality of the model.