异质信息网络中元路径感知的评分协同过滤

Meta Path-Aware Rating Collaborative Filtering in Heterogeneous Information Network

基于邻域的协同过滤(Neighborhood-Based Collaboration Filtering,NBCF)具有简单、可解释等优点一直备受关注且被广泛使用.然而,仅利用用户-项目的历史交互信息使得NBCF并不能获得很好的推荐性能.随着网络的快速发展,信息网络中包含了大量不同类型的对象和关系,越来越多丰富的语义信息可以被进一步挖掘和利用,自然构成了异质信息网络(Heterogeneous Information Network,HIN).基于HIN的推荐模型受到了研究者们的高度关注.相比于传统的推荐模型,基于HIN的推荐模型不仅能有效提高推荐性能,还能缓解冷启动和数据稀疏等问题.然而,现有基于HIN的推荐模型在保证模型有效性的同时往往需要学习较多的参数,参数的设定对模型性能有重要的影响.因此,本文提出一种无参数的HIN中元路径感知的评分协同过滤推荐模型HRCF,其主要思想是通过HIN中的元路径寻找评分的邻居,然后对邻居进行加权来估计该评分.首先,在不同的元路径上产生用户(项目)之间的交换矩阵;其次,将不同元路径上的交换矩阵进行整合计算用户(项目)之间的相似矩阵;最后将用户之间的相似矩阵、用户-项目的历史评分矩阵、项目之间的相似矩阵依次相乘并归一化从而一次性估计所有的评分.为验证HRCF模型的有效性,本文在公开的Douban Book和Yelp数据集上进行了实验.实验结果表明,HRCF模型的推荐精度优于目前存在的方法,且能很好地克服冷启动问题.

As an indispensable part of online shopping and e-commerce,recommendation system(RS)plays a vital role in helping users find products they are interested in.In many recommendation tasks,rating prediction is one of the most common and important ones.Moreover,among many recommendation models for rating prediction,Neighborhood-Based Collaborative Filtering(NBCF)has received a lot of attention and is widely used because of its advantages such as simplicity and interpretability.The main idea of the NBCF model is that similar users have similar preferences,which is relatively intuitive.However,NBCF is not able to obtain good recommendation performance only with historical user-item interaction information.The main reason is that the historical interaction information is often very sparse and the additional information is not enough,which makes the similarity measurement between users(items)less accurate.With the rapid development of networks,the information network contains many different types of objects and relationships,and more rich semantic information can be further tapped and utilized,which naturally constitutes Heterogeneous Information Network(HIN).So,HIN-based recommendation model has received extensive attention from researchers.Compared with traditional recommendation model,HIN-based recommendation models can not only effectively improve the recommendation performance,but also alleviate the problems of cold-start and data sparsity.However,most existing HIN-based recommendation models have many parameters which need to be learned while ensuring the validity,and the parameter setting has an important impact on the model’s performance.Particularity,some of the parameters in these models need to be adjusted manually,which results in a large amount of training time.Therefore,this paper proposes a parameterless meta path-aware rating collaborative filtering recommendation model in HIN(HRCF).Moreover,HRCF can be seen as an extension of NBCF in HIN,which makes HRCF inherit many advantages of NBCF.But the main difference between them is that NBCF utilizes similarity between users(items)to make recommendations,while HRCF directly measures similarity between ratings in HIN to make rating predictions.The main idea of HRCF is that ratings of similar users on similar items are close,and our statistical results in the Yelp dataset confirm this idea.In general,HRCF is to find the rating’s neighborhood through the meta-path in HIN and then weight the neighbors to estimate the rating.In addition,HRCF does not need to adjust or learn any parameters,and it can be converted to more convenient and fast matrix operations to perform.Specifically,we first generate commuting matrices between users(items)on different meta-paths.Then they are integrated to calculate the similarity matrix between users(items).Finally,the similarity matrix between users,the historical rating matrix of users-items,and the similarity matrix between items are sequentially multiplied and normalized to estimate all ratings at once.In order to verify the validity of the HRCF model,this paper conducted experiments on published Douban Book and Yelp datasets.The experimental results show that the recommendation accuracy of the HRCF model is better than the existing methods,and can well overcome the cold-start problem.Furthermore,we have also shown that a given set of meta-paths on those two datasets has a positive effect on the performance of the HRCF model.