Efficient algorithms for fast integration on large data sets from multiple sources

BACKGROUND: Recent large scale deployments of health information technology have created opportunities for the integration of patient medical records with disparate public health, human service, and educational databases to provide comprehensive information related to health and development. Data integration techniques, which identify records belonging to the same individual that reside in multiple data sets, are essential to these efforts. Several algorithms have been proposed in the literatures that are adept in integrating records from two different datasets. Our algorithms are aimed at integrating multiple (in particular more than two) datasets efficiently. METHODS: Hierarchical clustering based solutions are used to integrate multiple (in particular more than two) datasets. Edit distance is used as the basic distance calculation, while distance calculation of common input errors is also studied. Several techniques have been applied to improve the algorithms in terms of both time and space: 1) Partial Construction of the Dendrogram (PCD) that ignores the level above the threshold; 2) Ignoring the Dendrogram Structure (IDS); 3) Faster Computation of the Edit Distance (FCED) that predicts the distance with the threshold by upper bounds on edit distance; and 4) A pre-processing blocking phase that limits dynamic computation within each block. RESULTS: We have experimentally validated our algorithms on large simulated as well as real data. Accuracy and completeness are defined stringently to show the performance of our algorithms. In addition, we employ a four-category analysis. Comparison with FEBRL shows the robustness of our approach. CONCLUSIONS: In the experiments we conducted, the accuracy we observed exceeded 90% for the simulated data in most cases. 97.7% and 98.1% accuracy were achieved for the constant and proportional threshold, respectively, in a real dataset of 1,083,878 records.