面向互联网大数据的用户兴趣挖掘及预测研究

Nowadays, the applications in the Internet have accumulated huge and rich data, including the attributes of users and items, the social links between users, and the interactions between users and items. For example, users would like to read news, buy products, or check in tourism attractions, and then write comments to share their experiences. This project aims to exploit these data to mine and predict the interests of users to items, which can bring benefits to business, advertising, and the public. This project will study the following contents. 1) We will propose a new topic-sentiment model based on the dependencies on communities, regions, time, and the types of interactions in order to mine the fine-grained interests of users on items. 2) We will develop a new hybrid rating prediction model by combining matrix factorization and kernel-based nonlinear regression in order to predict the overall rating of users on items. 3) We will build the correlation between the topic-sentiment model and the hybrid rating prediction model. That is, the topic-sentiment model is used to estimate the historical user-item rating matrix from textual comments which is input into the hybrid rating prediction model. Thus, we can solve the problem on the lack of historical user-item rating matrices in the applications of the Internet and make the hybrid rating prediction model more useable. 4) We will evaluate the effectiveness of the proposed models using the real data from the Internet and user studies. The techniques on mining and predicting user interests can be applied in recommender systems so as to boost business profits and enrich people's quality of life.

互联网应用积累了海量的丰富数据，包括人的属性、物的属性、人与人的社交关系、人对物的互动关系。例如，用户浏览新闻、购买商品、签到旅游景点等，并且通过文本评论分享自己的经历。本项目拟利用这些数据，挖掘和预测用户的兴趣，这具有重要的商业、广告和人文价值。本项目将研究以下内容：1)建立基于社区、地区、时间和互动类型的主题情感模型，以挖掘用户对事物的细粒度兴趣。2)通过融合矩阵分解和基于核函数的多属性非线性回归，提出新的混合评分模型，以预测用户对事物的总体评分。3)建立主题情感挖掘模型和混合评分预测模型的联系，即通过主题情感模型估计历史用户-事物评分矩阵，作为混合评分预测模型的输入，以解决互联网应用中历史评分矩阵缺乏问题，从而提高混合评分预测模型的可用性。4)利用互联网大数据和用户实验数据，评价本项目提出的模型的效果。本项目开发的用户兴趣挖掘及预测技术，可应用于推荐系统，为大众和商家带来便利和利益。

互联网应用积累了海量的丰富数据，包括人的属性、物的属性、人与人的社交关系、人对物的互动关系。本项目利用这些数据，挖掘和预测用户的兴趣。本项目主要研究内容包括：（1）基于社区、地区、时间、互动类型的影响，提出了相应的情感偏好模型，从而更好挖掘用户对事物的兴趣。（2）评价和测试了情感偏好模型在真实数据中的表现。本项目的重要结果包括：（a）通过建模社区、地区、时间、互动类型对用户的情感偏好的影响，并且提出多标准决策方法，建模各因素在不同情况下的影响力，进而更好确定用户对事物的最终偏好程度。（b）基于深度学习技术，挖掘用户评论中语义信息，以及用户偏好随着时间迁移的变化规律，从而增强本项目用户偏好模型的准确性。（c）利用多个真实互联网大数据集，评价提出的情感偏好模型，实验结果验证了：与现有前沿技术相比，本项目模型的优越性。进一步，本项目偏好模型已经应用于商业推荐系统中。基于以上研究成果，本项目发表国际期刊SCI论文6篇。