Melvyl Recommender Project Findings
The Melvyl Recommender Project http://www.cdlib.org/inside/projects/melvyl_recommender/, which explored next-generation services for library catalogs, has reached its conclusion. The project was funded by the Andrew W. Mellon Foundation.
Popular commercial services such as Google, eBay, Amazon, and Netflix have evolved quickly over the last decade to help people find what they want, developing information retrieval strategies such as usefully ranked results, spelling correction, and recommendations. Library catalogs, in contrast, have changed little and are not well equipped to meet changing needs and expectations. The Melvyl Recommender Project explored methods and feasibility of closing this gap. An additional extension project to the Melvyl Recommender Project carried out deeper explorations into the most interesting and promising questions raised during the original project, and to add obvious missing pieces of functionality. The principal area of investigation was the impact of adding full-text objects to what had previously been a metadata-only index.
Overall findings from both portions of the project include:
- The text-based discovery application, the eXtensible TextFramework (XTF) that was the backbone of the project’s system (known as Relvyl) proved capable of scaling to millions of records and hundreds of concurrent users, indicating that this is an approach worth pursuing for providing ranking, recommendation and other types of functionality with an online catalog.
- Use of an index based single word spelling correction algorithm addressed 90 percent of misspelled single words.
- Initial examination of faceted browsing and FRBR-like document groups indicated that each of these features could substantially improve the patron’s experience of working with large result sets.
- User assessment confirmed that users prefer relevance ranked results over unranked results, although more investigation is required to determine whether content-based ranking with or without different types of weights (based on circulation or holdings) is more effective.
- Two types of recommendation strategies were explored: circulation-based (patrons who checked this out also checked,out) and text-similarity (More like this…). User assessment was conducted against the first type and showed that users like getting recommendations, which are useful for performing academic tasks, and they can also serve a unique query expansion function.
- Adjustments to keyword searching strategies, document scoring and the index-based spelling correction dictionary allowed for an effective combination of full-text and metadata only records into one system, in which neither type of record was privileged.
Much of the functionality explored in both phases of the project can be found in the Relvyl prototype (http://rec-proto.cdlib.org/xtf/search?style=melrec&brand=melrec)More information about the entire project can be found on the CDL website http://www.cdlib.org/inside/projects/melvyl_recommender/.