基于GPU数据库的新型推荐引擎研究

Recommendation system is an effective way to overcome the problem of “information overload”. Recently, with the development of E-commerce, recommender systems are being demanded by an ever-increasing number of applications. Existing research has obtained great achievements in user obvious or implicit behavior analysis, user preference modeling, recommendation engine algorithms design and recommendation applications. However, challenging research problems have been identified, such as data sparsity, and cold start, which will yield poor recommendation accuracy. Another challenge is to provide “persuasive” context-aware explanation to recommendation results. Moreover, scalability and efficiency often become the bottleneck for computation intensive recommendation methods to provide online recommendations. The large-scale, dynamical, and heterogeneous social media, E-commerce and mobile data make the research of these problems more difficult. In this project, we will conduct the research from aspects of theory and method, low-level database support, and high-level application verification. We will focus on the problems of accuracy, efficiency and explanation of recommendation system and propose a suit of methods on user preference modeling, recommendation algorithm efficiently executing, and recommendation explanation. We will develop new user modeling methods based on Bayes deep learning, parallel primitive operations and recommendation algorithms based on GPU, and unified storage and index methods for heterogeneous data, to improve the accuracy, performance, and explanation ability of recommendation system. Moreover, we will verify the proposed methods in the field applications.

推荐系统是解决“信息过载”问题的有效方式之一。近年来，随着电子商务的蓬勃发展，推荐系统在各领域的应用需求日益凸显。目前的研究在用户显性和隐性行为分析、用户偏好建模、推荐引擎算法设计和推荐应用等方面，已经取得了许多有价值的研究成果，然而在用户偏好建模和推荐的准确性、推荐引擎的高性能、推荐结果的可解释性等方面还存在着诸多不足。社会媒体、电子商务和移动应用的大规模、动态性及异构性，使得新型推荐系统在这几方面的研究面临巨大挑战。本项目拟围绕推荐系统的三个核心要素（用户偏好、推荐算法、异构数据）开展研究，探讨用户偏好建模、推荐引擎算法、异构数据存储和访问的基础理论与方法，拟提出基于贝叶斯深度学习的用户偏好建模、基于GPU 的并行原语操作和推荐算法、支持异构数据的统一数据存储和索引方法等，解决了新型推荐系统在大数据时代所面临的准确性、性能和可解释性等方面的难题，并通过面向领域的应用验证研究结果的效果。

本项目围绕推荐系统的三个核心要素（用户偏好、推荐算法、异构数据）展开了全面和系统的研究工作,并按照计划完成了全部研究工作。在推荐系统中的用户偏好建模研究、基于上下文的推荐引擎设计、数据稀疏和冷启动问题中推荐算法研究、因子分解机推荐模型研究、基于模糊粗糙集的特征选取算法研究、GPU排序算法研究、GPU连接算法研究、数据存储分区算法研究、高维学习型索引研究等方面取得进展。共发表和录用论文26篇。其中1篇发表在顶级国际期刊TKDE上，1篇发表在顶级国际会议AAAI上，4篇发表在著名国际会议CIKM、WSDM、ECML/PKDD和DASFAA上，4篇发表在著名国际期刊Information Science和ACM Transactions on the Web上。出版了两本教材。结合本项目申请了2项发明专利，培养了7名博士生和11名硕士生。