Entity resolution framework using rough set blocking for heterogeneous web of data

摘要

Entity Resolution (ER) is the method of resolving two similar entities used in the process of data cleaning and data integration. However, existing ER Framework lead to exhaustive pairwise comparisons. The most efficient ER method is blocking, inherently uses exponential pair-wise comparisons for the large databases, leading to poor efficiency in resolving the entities. The real world data can either be homogeneous or heterogeneous, generally of two forms, clean-clean ER which does not have any duplicates or dirty-ER which have duplicates within the dataset. Entity Resolution framework is associated with two phases namely the block building phase which construct the blocks where the similar entities are grouped into a single block for effective indexing, while the aim of block processing phase is to reduce the number of redundant pair-wise comparisons. Another perspective is handling of the entity associated with heterogeneous data, in the proposed work the block building phase aims to gather related entities with different representations into a single block with an approximation space. For this purpose semantic-dominance rough set has been used to cluster the attributes of related entities having a varied schema. The similarity between the entities associated with the clustered attributes is determined using a rough-Jaccard similarity measure, grouped to form blocks of varied, but limited size. The pair-wise comparisons between the blocks of entities are carried out only when the lower approximation of the blocks are same, determined by the proposed multi-criteria Pareto optimality, else the entities are not compared, which signifies, the overall number of pair-wise comparisons is reduced. A performance analysis of the proposed technique has been tested on four real-world, highly heterogeneous datasets, and the validation of these algorithms has yielded 99.98% effectiveness and 98.3% efficiency in block comparison when compared to token blocking and attribute clustering methods.