融合异构信息的低秩分解推荐模型研究

Aiming at the data sparsity and the un-interpretable challenges of traditional recommender systems, designing effective recommendation frameworks with the fusion of the enriched Web information, has become a promising research direction. Existing methods, however, suffer from the following three limitations. (1) conflicts on personalized biases of heterogeneous individual information; (2) conflicts on personalized biases of heterogeneous social information; and (3) time-consuming problem in fusing heterogeneous context information. In this project, we are going to investigate and improve the factorization-based recommendation approaches with the fusion of heterogeneous information. For the first limitation, a personalized weighted linear combination of latent feature vectors from heterogeneous latent spaces is proposed; for the second limitation, a generative process and a regularization function for fusing heterogeneous social information are designed, and graph-based algorithms are employed to quantify the relationship under homogeneous social network; and for the third limitation, the context latent feature vectors are linearly combined with personalized weights, after the linear transformations from heterogeneous latent spaces. Therefore, all the above limitations can be solved. At last, all the above three components are combined into a unified model. With the help of heterogeneous information, the data sparsity problem can be lightened; and with the learned personalized biases on various kinds of information, recommendation results are more interpretable. Under the information overload environment, the project will improve user experiences by finding their interested items, and will enlarge the revenue obtained from Web services by finding their potential users.

针对传统个性化推荐技术在稀疏数据可适性、推荐结果可解释性上面临的挑战，设计有效融合信息特征的推荐模型已成为推荐系统领域研究的热点。而现有方法存在异构个体信息偏好冲突、异构社交关系信息偏好冲突、异构上下文信息融合复杂度过高等问题。为此，本课题深入研究并改进融合异构信息的低秩分解推荐模型。通过建立独立的异构主题空间，并予以个性化加权融合，准确描述个性化异构个体信息偏好；通过设计融合异构社交关系信息的数据生成过程和正则化项，以及建立基于图结构的关联度量化方法，准确描述个性化异构社交关系信息偏好；通过将异构空间潜在向量映射到同构空间并加以线性融合的方式，降低融合异构上下文信息的计算复杂度。最后将上述改进整合成统一模型，通过对异构信息的有效融合，更准确地适刻画稀疏数据；同时通过对个性化异构信息偏好的准确描述，更完整的解释推荐结果。本课题将有力改善信息过载时代的用户体验及Web服务的收益。

课题以基于低秩分解的协同过滤作为关键技术，结合神经网络、主题分析、支持向量机等统计学习模型，研究个性化推荐技术中有效融入异构个体信息、关系信息及上下文信息的方法，提升推荐效果，解决推荐系统在数据稀疏、模型可解释性较弱等方面的不足。.. 研究内容包括：（1）融合异构个体信息的推荐模型，具体工作包括“低秩分解与异构主题分析的联合模型”、“基于指数族的混合回归模型”和“基于合成噪声的神经网络正则化模型”；（2）融合关系信息的推荐模型，具体工作包括“融合多特征的点击模型”、“流数据的关系分析预测模型”和“稀疏支持向量的优化模型”； （3）融合上下文信息的推荐模型，具体工作包括“面向特征向量的非线性迁移协同过滤模型”、“面向主题分布的非线性迁移协同过滤模型”和“面向缺失时序预测的多角度学习模型”。.. 研究结果表明，有效的融入个体、关系、上下文等异构信息，能够大幅度提升个性化推荐效果；将低秩分解技术与主题分析相结合，能够使推荐模型清晰的解释用户的个性化偏好及跨领域迁移特性。.. 研究成果发表于国际重要学术期刊及会议共9篇，其中包括IJCAI、AAAI等中国计算机学会（CCF）认定的A类论文3篇，B类论文1篇。