基于用户重购行为的产品推荐方法

Product Recommendation Method Based on User Repurchase Behavior

重复购买是消费者日常消费决策中的常见现象,考虑用户重购行为对于提升产品个性化推荐准确性至关重要.然而针对用户重购行为建模和预测的研究工作相对较少,还有很多问题有待解决.已有推荐技术主要通过深度挖掘产品、用户或时间某一层面信息来进行重购产品推荐,忽略了对多层次信息融合建模方法的研究,同时也忽略了重购推荐结果的可解释性需求.因此,融合多层次用户偏好信息,构建了具有双层注意力机制的可解释用户重复消费推荐方法.该方法融合注意力机制和指针生成网络,多层次提取并学习用户重购偏好,同时基于信息处理理论构建S型用户重购动态偏好函数,融合产品流行度信息进行重购产品和新颖产品的混合推荐,提高了模型可解释性和准确性.真实数据集上的实验结果表明,所提方法在多个性能指标上都优于对比方法,且学习出的参数具备较好的可解释性.此外,通过回归分析验证了S型重购动态偏好函数的可信性,进一步增强了理论的可解释性.

Repeat purchase is a common phenomenon in daily consumption decisions,and is crucial for improving the accuracy of personalized product recommendations.However,there is limited research on modeling and predicting user repurchase behaviors.Many issues need to be studied.It is notable that most of existing methods only optimize product,basket or time series level information for repurchase recommendation.They not only ignore the modeling of multi-level information fusion,but also ignore the interpretability requirement for repurchase recommendation results.To address the above issues,we propose a novel interpretable repurchase recommendation method based on dual attention mechanism.Our method combines attention mechanism and pointer generation network to extract multi-level user repurchase preference.We also propose an S-type user repurchase dynamic preference function based on information processing theory and integrate product popularity information for mixed recommendation of repurchase and new products,which improves the interpretability and recommendation accuracy of our method.The experimental results on two real world datasets show that our proposed method outperforms baselines in multiple performance indicators and the learned parameters have good interpretability.In addition,regression studies demonstrate the reliability of our S-type repurchase dynamic preference function,which further enhances the interpretability of our method from a theoretical perspective.