基于多维辅助反馈的数据生成和物品推荐机理研究

Recommender systems aim to provide users with personalized services, and have become an essential component of Web intelligent systems. However, recommender systems suffer from the data sparsity problem, which cause difficulty in learning user preference and greatly deteriorate recommendation accuracy. To resolve this issue, traditional approaches mainly take into account additional data of users and items, but ignore the users’ behaviors and feedback towards items, which expresses their preference more directly. Hence, this proposal intends to handle the concerned issue from the perspective of user-item interactions. Specifically, we will generate more high-quality target feedback based on multi-dimensional auxiliary feedback, whereby more powerful recommendation models can be proposed. The research is composed of three parts: (1) correlation analysis for multi-dimensional implicit feedback, including the linear and non-linear associations between multi-type auxiliary feedback, and especially those between auxiliary feedback and target feedback; (2) learning approaches to generate new data, including the methods based on users’ direct and indirect actions towards a target item; (3) recommendation models with multi-type feedback, including the impact of auxiliary feedback on existing approaches, the learning-to-rank approaches based on multi-type auxiliary feedback, and the confidence-aware deep collaborative filtering approaches. The related research outcomes will provide solutions to recommender systems based on multiple-type side information, resolving the data sparsity problem and improving the accuracy of item recommendations. Hence, our research has an important impact on both theoretical and practical aspects in the field of recommender systems.

推荐系统旨在为用户提供个性化的推荐服务，已成为Web智能系统的重要组成部分。然而，数据稀疏问题制约了推荐系统对用户偏好的学习，严重影响了推荐的准确性。针对该问题，现有算法主要是集成额外的用户或物品信息，却忽略了能直接反映用户偏好的行为反馈。本项目拟从用户-物品的交互角度，以多维度的辅助反馈为基础，生成更多的高质量目标反馈，提出多维联合推荐模型。重点研究：（1）多维隐式反馈的关联分析，包括多维辅助反馈之间、特别是与目标反馈之间的线性或非线性相关性等；（2）预期反馈的生成方法，包括基于用户对目标物品的直接或间接行为的目标反馈生成方法等；（3）融合多维反馈的推荐模型，包括预期反馈对现有算法的影响、基于多维反馈的排序学习算法、置信度感知的深度协同过滤算法等。研究成果将为融合多类型辅助信息的推荐系统提供解决方案，对进一步解决数据稀疏问题，提高物品推荐的准确性，具有重要的理论意义与实际应用价值。

数据稀疏是推荐系统领域的重要基础问题，它严重制约着对用户偏好的学习及物品推荐。本项目旨在通过使用多类型的辅助反馈（信息），例如时间、评论、图片、用户行为（点击和浏览等）等，更好地学习用户偏好，缓解数据稀疏的影响。具体来说，本项目从三个方面开展了针对性的研究。一是提出新型的融合多类型辅助反馈的推荐模型。通过分析不同类型的辅助反馈与目标反馈之间的相关性，例如时间对偏好的影响，更好地表达用户或物品的特征表示，提升物品推荐的准确性。二是提出推荐模型的学习和训练效率。融合多类型的辅助信息，也会相应地增加模型的学习难度和训练效率。通过提出动态采样策略、设计对抗生成网络等方法，寻找更有训练价值的负样本，从而提升物品推荐的效果。同时，利用向量的二值化加速计算，叠加模型优化时的二值化约束，提升模型的生成效率。三是进行前期的相关扩展研究。由于多类型的辅助反馈之间通常存在一定的序列关系，针对用户行为序列开展相关的外延研究，是非常有必要的。主要的研究成果包括：发表论文28篇，其中SCI期刊论文16篇，EI会议论文12篇，出版了1部学术专著。