时间约束的实体解析中记录对排序研究

实体解析是数据集成和数据清洗的重要组成部分,也是大数据分析与挖掘的必要预处理步骤.传统的批处理式实体解析的整体运行时间较长,无法满足当前(近似)实时的数据应用需求.因此,研究时间约束的实体解析,其核心问题是基于匹配可能性的记录对排序.通过对多路分块得到的块内信息与块间信息分别进行分析,提出两个基本的记录匹配可能性计算方法.在此基础上,提出一种基于二分图上相似性传播的记录匹配可能性计算方法.将记录对、块及其关联关系构建二分图;相似性沿着二分图不断地在记录对结点与块结点之间传播,直到收敛.收敛结果可以通过不动点计算得到.提出近似的收敛计算方法来降低计算代价,从而保证实体解析的实时召回率.最后,在两个数据集上进行实验评价,验证了所提出方法的有效性,并测试方法的各个方面.

Research on Record Pair Ranking for Entity Resolution with Time Constraint

Entity resolution (ER) is an important aspect of data integration and data cleaning,and is also a necessary pre-process step of big data analytics and mining. Traditional batch based ER''s overall runtime is costly, and cannot satisfy current (nearly) real-time data applications'' requirements. So we focus on time constraint entity resolution(TC-ER), whose core problem is record pair ranking according to match probability. We analyze both information inner blocks and information across blocks from multi-pass blocking respectively, and propose two basic records match probability methods. We improve the basic methods by proposing an advanced record match probability method based on similarity flowing over a biparitite graph. A bipartite graph is constructed according to record pairs, blocks and relations between them. Similarities iteratively flow between pair nodes and block nodes over the bipartite graph until convergence. The convergence result is computed with fixpoint iterations. An approximate convergence computation mehod is proposed to reduce cost, and it improves real-time recall in TC-ER. Finally, we evaluate the proposed methods on two datasets, show their effectiveness and also test different aspects of the proposed methods.