Prospective Data Fusion for Batch Filtering

This paper explores the problem of “global” data fusion rules, or combination of evidence. Three rather different schemes for computing the similarity between documents and a set of judged documents are employed. The goal is to learn whether rules for score combination (sometimes called “fusion rules” can be learned from a random selection of filtering problems L and then applied to another collection of filtering problems T. Results are encouraging, and set the stage for further research into “when” (that is, under what conditions) it will be effective to use two or more systems in fusion.