面向消费行为大数据的用户建模与个性化推荐研究

With the rapid increase of Web based consumption service like e-commerce and O2O, the information gap between diverse personalized user needs and large scale unstructured comsumption goods has become larger and larger. It has become an critical task to solve the gap and increase consumption capability of the society by analyzing users' consumption behaviors, identifying users' preferences and interests, and further providing personalized recommendation to help stimulate users' potention consumption needs. The big consumption data provide the opportunity to solve this problem. However, the complicated relations among user consumption behaviors, the heterogeneity and polymorphism of user's preferences, and the dynamic and interactive characteristics of user data all bring new challenges to the conventional user modeling and recommender systems. To address these chanllenges, we propose to focus on the following three scientific problems, i.e. user interest representation based on complicated linked data, personalized recommendation modeling based on heterogeneous preferences, and online learning theory and evaluation mechanism for recommendation based on interactions. We study from the following four aspects, namely user interest profiling, collaborative ranking based recommendation, online recommendation techniques and online demonstration. Through our study, we aim to build the systematic theory foundation and critical techniques for user modeling and personalized recommendation based on big consumption data. In this way, we can make better application of user modeling and recommender systems in Web based consumption services, and furhter improve the consumption capability of the society.

随着以电子商务、O2O为代表的互联网消费服务的高速发展，用户个性化、多样化的消费需求与庞大、信息无结构的消费品数据之间存在的信息鸿沟也愈发明显，如何深入分析用户的消费行为，识别用户的偏好兴趣，进而提供个性化推荐来刺激用户的潜在消费需求，是弥合信息鸿沟、提高社会消费能力的关键难题。用户消费行为大数据为解决该问题提供了契机，然而用户消费行为复杂关联、用户偏好异构多态、用户数据实时交互的特性，都给传统的用户建模与推荐技术提出了全新的挑战。针对上述挑战，本课题围绕复杂关联数据下的用户兴趣表征、面向偏好异构的个性化推荐问题建模、面向交互的在线推荐学习理论和评估机制三个科学问题，从用户兴趣画像、协同排序推荐模型、在线学习推荐技术和在线验证四个层面展开研究，旨在建立面向消费行为大数据的用户建模与个性化推荐成体系的基础理论和关键技术，推动用户建模与推荐在互联网消费服务行业的应用，进一步促进我国的社会消费。

深入分析用户的消费行为，识别用户的偏好兴趣，并进一步提供个性化推荐来刺激用户的潜在消费需求，是工业界和学术界广泛关注的研究问题。本课题针对用户消费行为复杂关联、用户偏好异构多态、用户数据实时交互的挑战，以用户行为大数据为基本资源，以推荐为核心应用场景，围绕复杂关联行为数据下的用户兴趣表征、面向偏好异构的个性化推荐问题建模以及面向交互的在线推荐学习理论和评估机制三个科学问题，在多模态深层用户兴趣画像、建模相对偏好的协同排序推荐、混合反馈下在线学习推荐与评估以及在线数据验证四个方面取得了突破，形成了包括结构化多层次表征学习、深度匹配模型、强化在线学习等一系列原创性的理论方法与模型，在国内外重要会议和期刊发表(录用)学术论文51篇，获得国际学术奖励1次，申请专利3项。通过本课题的研究，能有效解决利用大规模用户行为数据进行推荐所面临的核心问题，进一步提升推荐技术，推动推荐系统在互联网中更为广泛有效的使用。