III-COR-Small: Multi-Relational Data Clustering with Probabilistic Mixture Models

随着将数据挖掘方法应用于现实生活问题的广泛尝试，人们越来越认识到现实生活数据往往是多关系的，涉及通过一组关系连接多个实体的观测。在涉及多关系数据的几个应用中，一个中心问题是同时发现跨相关实体的对象簇，例如电子商务中的客户簇和相关产品簇，推荐系统中的电影簇和相关用户簇，社交网络中的社区和共享内容等。关键的新颖方面是，一个实体中对象的聚类依赖于它与其他实体中的对象的关系，例如，如果用户喜欢相似的电影，则用户相似，反之亦然。本项目的主要目标是开发一种统一的统计方法来解决多关系聚类和多关系数据分析中的相关问题。为此，该项目研究了一类新的统计多关系混合模型，重点是多关系聚类的加法和乘法模型。由于设计的模块化，加法和乘法模型都可以结合领域特定的语义，以及使用适当的贝叶斯先验进行自动模型选择。此外，该项目还研究了适用于发现潜在多关系聚类的高效变分推理方法。该项目显著增强了数据挖掘的知识发现组件。理解几个学科中观察到的数据的关键线索，包括社会、生物和信息科学，通常分布在多个相关的观察中。该项目实现了一种统计方法来检测这种多关系数据中的潜在结构，这是从多个相关数据来源进行知识发现的重要一步。该项目在发展更紧密的跨学科合作和扩大对计算机科学的参与方面发挥了重要作用。基于人们对多种关系数据无处不在的日益认识，该项目有助于为下一代劳动力开发适当的教育材料。有关该项目的更多信息，请访问项目网站site:http://www.cs.umn.edu/~banerjee/multi-relational.

With widespread attempts to apply data mining methods to real lifeproblems, there is an increasing realization that real life data isoften multi-relational, involving observations connecting multipleentities through a set of relations. A central problem in severalapplications involving multi-relational data is to simultaneously findclusters of objects across related entities, e.g., customer clustersand related product clusters in e-commerce, movie clusters and relateduser clusters in recommendation systems, communities and sharedcontent in social networks, etc. The key novel aspect is that theclustering of objects in an entity, such as the set of movies orusers, depends on its relationships with objects in other entities,e.g., users are similar if they like similar movies, and vice versa.The primary goal of of this project is to develop a unifiedstatistical approach to multi-relational clustering and relatedproblems in multi-relational data analysis. Towards this end, theproject investigates a family of novel statistical multi-relationalmixture models, with focus on additive and multiplicative models formulti-relational clustering. Due to modularity of design, bothadditive and multiplicative models can incorporate domain specificsemantics as well as automatic model selection using appropriateBayesian priors. Further, the project investigates efficientvariational inference methods appropriate for discovering latentmulti-relational clusters.The project significantly empowers the knowledge discovery componentof data mining. Crucial clues to understanding observed data inseveral disciplines, including social, biological, and informationsciences, are often spread across multiple related observations. Theproject enables a statistical approach to detecting latent structurein such multi-relational data, which is an important step towardsknowledge discovery from multiple related data sources. The projectplays an important role in developing closer collaboration acrossdisciplines and broaden participation in computer science. Buildingon the increasing awareness regarding the ubiquity of multi-relationaldata, the project contributes to the development of appropriateeducational material for the next generation work-force. Furtherinformation on the project may be found at the project web site:http://www.cs.umn.edu/~banerjee/multi-relational.