Entity Resolution via Graph Clustering and LSH Blocking

The patent segments the entity resolution process into blocking and matching phases. The blocking phase groups documents into buckets based on similarity, and only documents within the same bucket are compared in the matching phase. This segmentation reduces the number of comparisons needed when updating entity resolution results, as new documents only need to be blocked and matched against relevant buckets rather than the entire dataset.

The patent performs blocking as a preliminary action before matching. By pre-grouping documents into buckets based on similarity metrics (such as Jaccard similarity, cosine similarity, or edit distance), the system prepares the data structure in advance so that when new documents arrive or existing documents change, only the affected buckets need to be reprocessed rather than performing full pairwise comparisons across the entire dataset.

The patent divides the dataset into multiple buckets based on document similarity using blocking techniques. This segmentation ensures that computational resources are focused only on comparing documents within the same bucket, rather than performing exhaustive pairwise comparisons across the entire dataset. The segmentation maintains resolution completeness by ensuring that similar documents are grouped together in the same bucket.

The patent applies partial action by performing entity resolution only on subsets of documents (those within the same bucket) rather than all possible document pairs. This selective approach reduces computational resource consumption while still achieving complete entity resolution for the relevant subsets. The blocking mechanism ensures that no potentially matching documents are missed by organizing them into appropriate buckets before matching.

The patent segments the blocking mechanism into multiple independent similarity metrics (e.g., Jaccard similarity, cosine similarity, edit distance, inter-document references). Each metric can be computed independently and combined to form the final blocking decision. This modular segmentation improves efficiency by allowing parallel computation of different metrics while managing complexity through clear separation of concerns.

The patent creates a universal blocking mechanism that can handle multiple types of similarity metrics and data formats through a unified framework. The blocking system is designed to accommodate different similarity measures (textual, structural, reference-based) and can be configured based on specific application requirements, making it versatile without requiring separate blocking mechanisms for each metric type.

The patent performs blocking as a preliminary action that can be reused across multiple update cycles. Once documents are blocked into buckets, this blocking structure can be maintained and updated incrementally as new documents arrive or existing documents change. This preliminary blocking action prevents the need to recompute blocking structures frequently, reducing resource wastage while maintaining data accuracy through incremental updates.

The patent enables selective reprocessing by identifying and discarding only the portions of the dataset that have changed or may be affected by changes. When updating entity resolution results, the system recovers and reprocesses only the affected buckets rather than the entire dataset. This selective approach maintains data accuracy for changed portions while avoiding unnecessary reprocessing of unchanged data, thereby reducing computational resource wastage.