基于聚类技术的推荐算法研究

【作者】 张亮；

【导师】 尚明生；

【作者基本信息】 电子科技大学 ， 计算机技术， 2012， 硕士

【摘要】 推荐系统是帮助用户评估他没有发现的内容，从而克服信息超载的一种有效工具。对于推荐系统的研究，既有重大的社会意义，又有重大的经济价值。自上世纪九十年代推荐系统作为一个独立的研究问题以来，所进行的相关研究涵盖了数据挖掘、人工智能、人机交互和用户行为学等多个学科领域。针对实际应用中的一些问题，数据挖掘方面研究者们从推荐算法角度给出了众多解决和改进办法。这些工作使得推荐系统已经成功应用于各个商业系统中。虽然个性化推荐技术已经在互联网领域内得到广泛应用，但是仍普遍存在数据稀疏性、算法可扩展性和“冷启动”等亟待解决的问题。利用聚类分析和复杂网络等方法，本文致力于理解推荐系统各要素，在一定程度上解决推荐系统的现有问题。主要工作如下：1.提出了一种基于项目聚类的推荐算法。算法的核心目的在于在保留了其它已有基于聚类的推荐算法在解决数据稀疏性优势的基础上，使用聚类分析技术对原始信息进行处理，并通过引入归属度的概念，将真实用户的行为模型转化为兴趣模型从而进行了更精准的推荐。在多个不同类型数据集上进行了多次实验并使用离线仿真的方式对算法的精确性进行评估，结果表明算法在预测精确性上也对原始算法有着较大程度上的提高。2.提出了一种基于跨电商行为的交叉推荐算法。在分析了现有推荐系统在电子商务应用中的主要任务和面临问题基础上，针对新用户的“冷启动”问题使用用户在多个不同类别电子商务网站访问的交叉行为信息提进行推荐。经过离线仿真，算法可以提供具有相当精确性和个性化的推荐。即使只利用一个站外电商数据进行交叉推荐，其精确度可以远远超过随机推荐，与此同时保持和随机推荐相近的多样性和新颖性。这些都保证算法有着良好的用户体验。随着我们对用户已知信息的不断引入，算法的精确性获得进一步提高。这部分工作为推荐系统的商业应用提供了全新的可供分析的数据对象和全新的研究视角。更多还原

【Abstract】 Recommendation System is a tool for helping users to evaluate the content he didnot discover; it is also a tool to overcome the information overload. The research onRecommendation System is both of great social significance and significant economicvalue. Since the nineties of the last century, Recommendation System being proposed asan independent subject, related research covers data mining, artificial intelligence,human-computer interaction, user behavior and other disciplines. Targeted on someproblems in the practical applications, data mining researchers give out a number ofsolutions and improvement strategies from the perspective of algorithms. These worksmake that Recommendation System has been successfully used in various commercialsystems.Although personalized recommendation technology has been widely used in theInternet, there are still some problems which should be solved, such as data sparsenessproblem, scalability of algorithms and "cold start" problem. The thesis aims tounderstand the various elements of the recommendation systems and solve existingproblems using approaches such as cluster analysis and complex networks, The mainresults are as follows:1. The thesis proposes a recommention algorithm based on item-based clustering.The core of the algorithm aims to deal with the original information andprovide a more accurate recommendation for real users by transformingbehavior model into interest model through the introduction of users’belongingcoefficient, while retain the advantages of other cluster-based recommendationalgorithms in solving data sparsity problem. Many experiments with couple ofdifferent kinds of data sets are carried out and we also use offline simulation toevaluate the accuracy of the algorithm, results show that the predictionaccuracy of this algorithm has a greater enhancement compared with theorigional algorithm.2. The thesis proposes a multi-B2C crossing ranking recommendation algorithm.The algorithm makes use of users’ accessing information in a number of different categories of E-commerce sites to recommend, after we analysis themain tasks and problems of existing recommendation systems in E-commerceapplications, the recommendation is of accuracy and personality by offlinesimulation. Even using only the data of one site to recommend, its accuracycan be much better than random recommendation, while maintaining randomrecommendation’s diversity and novelty. These are to ensure that the algorithmhas a good user experience. As more known information is taken into account,the accuracy of this algorithm has a greater enhancement. This part of workprovides a brand new object of data and perspective for commercial applicationof recommendation systems.更多还原