基于联合潜在因子模型的跨领域信息推荐系统研究

随着互联网信息资源的迅猛增长，传统的信息推荐系统已很难适应用户规模、推荐项目数量的快速提升，特别是仅仅只能针对单一目标领域内的信息对象提供推荐，而无法扩展到跨领域的信息推荐。因此，本项目提出了跨领域多源异构信息推荐这一新问题以及相应的原创性理论模型：联合潜在因子模型。该模型对跨领域的非重合用户群体、信息对象内容以及各领域中用户与信息对象的交互模式进行联合建模，通过获取跨领域中用户聚类对信息对象评价模式的共性特征和个性特征来构建高效的跨领域信息推荐算法，提高跨领域信息推荐结果的准确性和多样性。本项目的研究内容主要包括：（1）联合潜在因子模型及其优化算法的理论研究；（2）跨领域中基于聚类技术的用户兴趣模型构建研究；（3）跨领域中基于迁移学习的多源异构信息对象模型构建研究；（4）基于联合潜在因子模型的跨领域信息推荐算法研究；（5）跨领域大数据信息推荐系统演示平台研究。

With the rapid growth of Internet resources, most recommender systems aim to provide recommendations or rating predictions of an active user on a set of items belonging to only a single domain (e.g., movies or books) based on the historical user-item preference records, which usually suffer the data sparsity problem. However, there exists a considerable number of publicly available user-item rating datasets from multiple domains, which can have dependencies and correlations among the domains. Thus, this project addresses a new problem of Cross-domain Recommender System research, which is based on a new learning framework called Collective Latent Factor Model (CLFM) to model the users, heterogeneous items and the rating data simultaneously and capture the common cluster-level user-item rating pattern shared across domains as well as the domain-specific cluster-level rating pattern from each domain, which contains the discriminative information propitious to improve across recommendation accuracy. In this project, major research targets include: (1) to create and perfect the basic theory of Collective Latent Factor Model, proposing two kinds of model constructions in computational framework and probabilistic framework respectively, and developing the efficient optimization algorithms for these model choices; (2) by employing the framework of Collective Latent Factor Model, proposing the User Interest Model by deriving the latent user-cluster factor across multiple domains based on clustering algorithms, and integrating the social network information of users into the model; (3) based on the Collective Latent Factor Model, proposing the Item Topic Model by deriving the latent item-cluster factor from the available rating data on items, and combining the across-domain heterogeneous items to construct the topic model with the aid of transfer learning. (4) by using joint nonnegative matrix tri-factorization based Collective Latent Factor Model, proposing the across-domain recommendation algorithms to learn shared common rating pattern across multiple domains and also learn the domain-specific cluster-level rating pattern from each domain to improve across recommendation accuracy. (5) designing and implementing the Map and Reduce functions using data and task parallelism from the across-domain information, and proposing a parallel collaborative filtering algorithm for large scale data based on CLFM.