基于排序的个性化推荐系统关键问题研究

发布时间：2018-12-09 09:57

【摘要】：本文调研了目前推荐系统中常用的推荐算法,发现了在目前推荐系统中大多采用基于评分预测的方法。基于评分预测的方法过多地看重了准确预测评分的重要性。因而,本文提出了基于排序的学习方法来建立推荐模型。本文的研究内容主要包括以下几方面:1、将三种 LTR(Learning To Rank)算法,包括 Pointwise、Pairwise与List-wise,迁移至推荐任务中,并建立模型。并提出基于排序模型的几点改进。系统采取了基于“样本对排序”的学习来建立推荐模型,并加入了零样本采样以提高训练效果。2、非结构化文本的建模。提出了基于词向量并结合人工定义维度的方法来建立文本模型。3、评分模型与评论文本模型的融合。在分别建立用户评分、用户评论的模型之后,本文提出了融合评分与评论两部分模型的方法。结合文本建模与对排序模型,使得推荐的结果更加具有可解释性。
[Abstract]:This paper investigates the commonly used recommendation algorithms in the current recommendation system, and finds that most of the current recommendation systems use the method based on score prediction. The method based on score prediction values the importance of accurate prediction. Therefore, this paper proposes a learning method based on ranking to establish a recommendation model. The main research contents of this paper are as follows: 1. Three kinds of LTR (Learning To Rank) algorithms, including Pointwise,Pairwise and List-wise, are migrated to the recommended task, and the model is established. Some improvements based on the sort model are proposed. The system adopts the learning of "sample to sort" to establish the recommendation model, and adds zero sample sampling to improve the training effect. 2. Modeling of unstructured text. A method based on word vector and artificial definition dimension is proposed to build text model. 3. Fusion of scoring model and comment text model. After establishing the model of user rating and user comment, this paper proposes a method to combine the two models of rating and comment. Combined with text modeling and sorting model, the recommended results are more interpretable.
【学位授予单位】：北京邮电大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.3

【参考文献】