基于市场效应的用户行为与协同过滤推荐研究

Recommender system is one of the most important parts in a variety of Internet-based applications, including e-commerce and e-business. Based on the deep analysis into the problems existed in current collaborative filtering recommendation methods used in the systems, combined with the understanding of the requirements of information technology development, especially the requirements from big data applications, the project will carry out research on a sort of phenomenon appeared in the activities of users' product selection. The phenomenon is referred to as effect-based behavior in the project and will be examined from both theoretical and empirical perspective. The understanding of the phenomenon is beloved to be able to enlighten approaches for research on recommendation methods and this project will also conduct research on new recommendation methods from this view. The project will address research topics in three aspects including study of effect-based user behavior, research on recommendation methods and related application issues, experiments on and comparisons among the related recommendation methods. Base on the original concept of effect-based behavior proposed in this project, a distinctive approach is suggested to deal with the associated questions in recommender systems research area. Full consideration has been given to the related aspects of the proposed project and the scheme of the research is well demonstrated in the proposal. The proposed concept is believed to be helpful in extending our understanding about user behavior in product selection and evaluation. Recommendation methods generated from this new understanding about user behavior are believed to be of effectiveness in solving problems such as data sparsity, the trade-off between the efficiency and accuracy of a recommender system, and etc., which exist in traditional recommendation methods, of both value in theoretical research and applications in real practices, and of help in enhancing the adaptability of recommender systems in big data application environment. After all, the research is hoped to gain fruitful results which will enrich theories and methods in recommender systems area, and has its significance in improving shopping environment for online users in their related network activities, has implication in booming the development of e-commerce application.

推荐系统是当前包括电子商务在内的各种网络应用的一个重要组成部分。课题在深入分析当前协同过滤推荐技术存在的问题的基础上，结合信息技术应用发展、尤其是大数据背景下的应用需求，针对人们在产品选择中存在的一些行为特征，提出效应行为概念，拟将在对其进行深入的理论和实证研究基础上，展开相应的推荐算法研究。研究内容包括基于市场效应的用户行为研究、推荐算法及其应用相关问题研究、相关推荐技术比较分析等三个方面。 课题对研究方案进行了充分论证和精心设计，研究思路新颖独特。对效应行为的研究，可以拓展对人们选择行为的认识框架，以此为基础进行推荐算法的分析设计，对有效解决推荐系统应用中数据的稀疏性、推荐的精确性与系统效率之间的矛盾等方面的问题有较高的理论和实际应用价值，使推荐系统在大数据背景下具有较强的适应性。研究成果将丰富推荐系统领域现有的理论和方法，对完善我国网络购物环境、促进电子商务应用发展具有积极意义。

课题研究以人们的评分行为是一种理性有限的行为（简称“有限理性评分行为”）为前提，以预测评分计算中所依赖的参照集（最近邻项目集）为切入点，在对协同过滤技术的基本原理和方法、以及人们的实际评分行为特点进行深入分析的基础上，在理论、方法、技术和实验等四个方面取得了如下创新性成果：.在理论方面：（1）基于对人们评分行为的分析理解，识别出了两种课题称之为“市场效应”和“权变评分”的行为现象；（2）深入分析了人们的有限理性行为以及“权变评分”行为对实际评分及预测误差产生的影响；（3）将评分误差区分为理性误差、行为误差和方法误差，结合实验，讨论了它们对预测准确性的影响及改进思路。.在方法层面：（1）基于市场效应，提出了一种比传统IBCF方法更为简便有效的评分预测方法（MECF）；（2）提出了一种称之为峰值权变的评分改进方法（RAP）和一种称之为邻域叠加的邻域修复方法（NSM）。将二者结合起来，形成一种称之为R-N方法的系统化的方法体系。.在技术层面：（1）提出了一种应用于拟合评分计算中的综合权重因子；（2）应用信息熵，提出了一种分析邻域叠加对邻域修复效果的度量方式；（3）针对现有各种实验检验指标的不足，提出了一种称之为全序度（TRD）的检验指标。.在实验层面：（1）进一步假设人们的这种有限理性评分行为产生的评分误差为随机误差，通过检验几种假设情形下的模拟评分，来分析在理想情况下评分预测能够获得的效果，为理解和分析评分预测方法能够得到的评分预测效果提供参考；（2）通过实验，详细分析了在各个分值上的评分误差，为理解人们实际评分中的“权变行为”提供了依据；（3）提出了一种“异常评分”分析方法，以进一步展示各种评分预测方法达到的预测效果及其不足。