面向群体推荐问题的多用户兴趣建模方法研究

Recommender system has become the most effective technology to solve information overload. Nowadays, social tools strengthen the connections with each other, more and more user behavior is playing out in the form of group activities, resulting huge demand for group recommendation. However, a group is a complex system produced by multiple users’ interaction instead of a direct aggregation of individual users, and its behavior reflects the changes-from independence to the self-organization, from closed to interaction and from single to diversity, which determine the difficulty of the problem. Disciplines such as sociology, psychology and social network have analyzed group behavior, and sketched the silhouette of group activities. Because of this, the project is to explore the decision-making mechanism of group activities based on the theory and method of group behavior analysis, and then guide machine learning techniques to study multi-user interest modeling method, including three focus tasks about social analysis of member relationship, dispersive extraction of local feature, diverse aggregation of local interest. These tasks can form a complete group activity learning framework including analysis, extraction and aggregation. The implementation of this project will deal effectively with group recommendation, enrich recommendation theory and method, extend application range of recommender system and make Internet service of our country more open, more social and more intelligent.

推荐系统已经成为解决信息过载问题最有效的技术。如今，社交工具加强了彼此联系，用户越来越多的行为都是以群体活动形式来展开，由此产生了面向群体推荐的极大需求。然而，群体不是个体的简单汇聚，而是多个用户相互作用形成的系统，其行为体现为从独立到自组织、从封闭到互动、从单一到多样，这些特点决定了该问题的困难性。社会学、心理学、社会网络等学科对群体行为有了初步分析，勾画出群体活动的整体轮廓。鉴于此，本项目在交叉学科有关群体行为分析的理论与方法基础之上，探究群体活动决策机制，以此为导向利用机器学习等技术研究多用户兴趣建模方法，重点开展社会化成员关系分析、分散化局部特征提取、多元化局部兴趣聚集三方面工作，这些工作可以形成完整的一套包含分析、提取、聚集的群体活动学习框架。本项目的实施将有效应对群体推荐问题，丰富推荐理论和方法，扩展推荐系统适用范围，推动我国互联网服务向更开放化、更社会化、更智能化方向发展。

互联网社交工具的发展加强了彼此联系，网络用户越来越多的活动都是以群体形式来展开，但是其活动决策繁琐耗时，由此产生了面向群体推荐的极大需求。群体不是个体的简单叠加，而是多个用户自组织形成的复杂系统，其行为体现为从独立到互动的变化。然而，当前主要解决方法是直接聚集个体兴趣产生预测结果，忽视了成员互动带来的影响，没有妥善解决该问题。社会学、心理学、复杂网络等学科对群体行为有了初步分析，大致勾画了群体活动的轮廓。鉴于此，本课题组参考交叉学科有关群体行为分析的理论与方法开展了相关研究，从成员关系及互动情况入手，在此基础上从互动角度深入探究了群体活动决策机制，以此为导向基于主题模型提出了用户兴趣建模方法，已经在关系分析、特征提取、兴趣聚集方面取得了一系列研究成果，形成了一套完整的群体活动学习框架，可以更准确地认知群体兴趣，提供更符合群体活动需求的信息，产业化后可以提高互联网社交平台的智能化服务水平。