fCombMNZ: An Improved Data Fusion Algorithm
Mohammad Othman Nassar, Ghassan Kanaan · 2009
The data-fusion techniques have been investigated by many researchers and have been used in implementing several information retrieval systems. Fusion allows leveraging of the component systems in several ways by exploiting a number of effects, amongst them an effect called "chorus effect"; this effect suggest that for a particular document if it is retrieved by two systems; it will be "better" than another document retrieved by only one system, and if it is retrieved by three systems it will be "better" than another document retrieved by one or two systems, and so on. "Better" means the document has higher probability to be relevant. We suggest including a new simple but effective rule called "fairness rule" within well known data-fusion algorithm called CombMNZ to allow this algorithm from leveraging the component systems in an effective way, this rule is the first attempt in the literature to manipulate the normalized score values for documents under certain circumstances to maximize the CombMNZ ability to benefit from the "chorus effect" without sacrificing its ability to benefit from another effect called "skimming effect", skimming effect; suggests that relevant documents are most likely to occur on the top of the retrieved list for any retrieval system. Our results show an improvement in the mean average precision (MAP) values over the traditional CombMNZ when we used our "fairness rule", the improvement for TREC3 groups ranged from 0.66% to 2.88%, and for TREC5 ranged from 0.2% to 3.7%.