Using collaborative filtering to weave an information tapestry

Reviewer: Thomas C. Lowe

Filtering processes are the same as, or descended from, selective dissemination of information (or SDI, although that acronym has been commandeered). The general concepts apply to traditional document collections, where an SDI system brings selected new accessions to appropriate patrons' attention, or automatically copies and routes news briefs, memoranda, directives, or information. The notions apply equally to productivity tools for environments increasingly deluged with electronic mail. Email is the motivation for Tapestry, a system that features, in addition to content-based filtering, human contribution to the filtering, called collaborative filtering . Those who read a document in Tapestry can annotate that document by recording their responses to each document in an append-only annotation store. Documents are indexed only once; documents are never re-indexed or purged in the ideal case, so the indexing and annotation information are in append-only storage. Therefore, new indexing criteria cannot be introduced, and endorsements cannot be expunged. A Tapestry requirement for “continuous semantics” results in the restriction that, for any fixed query, the number of documents returned by that query must be a weakly monotonic increasing function of time. So, as the authors illustrate, a valid query is “all bug reports that [were] not answered within two weeks,” and an invalid query is “all bug reports that are over two weeks old and have not been answered.” This restriction may not be inconsistent with a document-streaming SDI processor's typical application. The paper describes the advantages of Tapestry's collaborative nature, its implementation under Sybase with translation from its peculiar language to SQL, and existing and planned electronic mail connections. Eventually, Tapestry is intended to be a replacement for current email systems.