| Merging has been investigated in several application contexts, which range from political elections to metasearch engines. Methods of merging for retrieval results which present special characteristics such as certain documents being in some but not in all lists, retrieval results, and in which only rank positions in each list are known, are not yet fully investigated. We focus on merging retrieval results from distributed information systems in order to gain coverage and to improve retrieval.; Recent research in the area of fusion has shown that in very special cases improvement is achieved. These cases are related to the amount of overlapping among the sources to be merged.; By selectively choosing the rankings, the problems relating to too much overlap can be overcome. Formally stated, given a finite set of rankings, we provide mechanisms to analyze rankings and to merge them. A rigorous analysis of rankings based on metrics [24] allows us to perform query specific search engine classification and recommendation of rankings for further merging. Two methods to merge, which take into consideration absence of local similarities, missing items among the search engines/lists, etc. are developed and evaluated. Experimental results on several TREC collections were conducted. Results favored our proposed methods, OWA and Reasonable, over known similar methods. |
