Profile-Based Object Matching for Information Integration

an important role in many information management contexts, including information integration, data warehousing, information extraction, and text joins in databases. As the “Related Work” sidebar describes, researchers have proposed numerous object-matching solutions in both the AI and database communities. Virtually all these solutions assume that the target objects, or tuples, share the same set of attributes, and they match tuples by comparing attribute similarity. However, tuples often have nonoverlapping, or disjoint, attributes. This is frequently true in information integration, especially when merging tuples from different data sources. Because these data sources are typically developed independently, they often have overlapping, but different schemas. As an example, suppose a government agency is trying to merge two branch offices and their databases (S1 and S2), which both list information about people living in Champaign, Illinois. As Figure 1 shows, each tuple contains information about a single person. Because the databases cover the same geographical area, they contain many duplicate tuples. For example, S1’s tuple S. Riley, 105 Spring St, 61801, $95,000 refers to the same person as S2’s Sarah Riley, E. Spring St, 61801, 38; the shared attributes are name, street, and zip, while the disjoint attributes are income (S1) and age (S2). Existing systems can effectively match attributes, despite variations and errors: the name “Sarah Riley,” for example, might be abbreviated as “S. Riley” or even mistyped as “Sarah Rilye.” However, these methods don’t exploit disjoint attributes in the matching process. Our Profile-Based Object Matching solution—as embodied in the PROM system we’re developing—exploits disjoint attributes to maximize matching accuracy. The key to PROM is that disjoint attributes are often correlated, and we can leverage this correlation to perform a “sanity check” on object matching. So if we had the S1 tuple Mike Smith, E. White Street, 61820, $100,000 and the S2 tuple Mike Smith, E. White Street, 61820, 6, prior solutions would declare a match because the shared attributes match perfectly. However, the disjoint attributes (income in S1 and age in S2) indicate that, combined, the two tuples give us “Mike Smith,” a six-year-old with a $100,000 yearly income. This is possible but very unlikely. Thus, PROM would reject the match.