基于排序蒸馏的序列化推荐算法

Sequential recommendation algorithm based on sequencing distillation

为解决当前基于知识蒸馏的推荐算法排名有效性和效率低,以及现有知识蒸馏模型更强调的是静态和单一知识迁移的问题,提出一种基于排序蒸馏的序列化推荐算法。训练一个性能优越、规模大的教师模型,训练一个符合移动终端设备的小模型即学生模型,使学生模型在教师模型的指导下学习排序。学生模型实现了与教师模型相似的排名性能,且学生模型规模较小提高了在线推荐效率。通过在数据集MovieLens和Gowalla上的实验,验证了该模型增强了学生模型的学习效果,缓解了学生模型学习不充分导致排名不佳的问题。模型可以自然地运用于序列化推荐的模型中,具有很好的通用性。

To solve the problems of low ranking effectiveness and efficiency of current recommendation algorithms based on knowledge distillation and the existing knowledge distillation models that emphasize static and single knowledge transfer,a seria-lized recommendation algorithm based on sequencing distillation was presented.A teacher model with superior performance and large scale was trained,and a small model that conformed to the mobile terminal equipment,namely the student model,was trained,so that the student model could learn and sort under the guidance of the teacher model.The student model achieves the ranking performance similar to the teacher model,and the smaller size of the student model improves the online recommendation efficiency.Through experiments on the dataset MovieLens and Gowalla,it is verified that the model enhances the learning effect of the student model,alleviates the problem of poor ranking caused by insufficient learning of the student model,and the model can be naturally applied to the model of serialized recommendation,with good universality.