|
 
|
[RFC13] A Weighted Multidimensional Scaling approach for representing the result of a search
Conférence Internationale avec comité de lecture : y-BIS 2013: Joint Meeting of Young Business and Industrial Statisticians, September 2013, pp.xx, Instanbul, Turkey,Mots clés: information retrieval, data visualization, multidimensional scaling.
Résumé:
Ranking by decreasing relevance is the dominant paradigm for presenting the result of search. Classical search procedures consist in sorting a set of objects by proximity values with respect to a reference object (or, more in general, a query) and replacing the values by their rank. However, there are many cases where sorting is not the best way to present the results of a search. Whatever the relevance measure employee (a measure of similarity, the structure of links, etc..), the ranking mixes several sources of information and it does not highlight significant structure in the data. We propose Multidimensional scaling (MDS) [1] to go beyond the ranking, by enriching the information carried by the ranking with the information about data structure. MDS represents measurements of similarity (or dissimilarity) among pairs of objects as distances between points into lower dimensional configuration. Since MDS models data as distances among points in a geometric space, we can visualize the result of the search on a 2D space, in order
to yield information easily and promptly interpretable by the user. Weighted MDS (WMDS) models [2,3] are defined by a configuration which minimizes a weighted sum of the squared differences between original proximities and the distances in the configuration. In this framework we developed a particular system of weights which privileges proximities among objects close to the reference. Moreover, we introduced in the weight system a tuning parameter allowing WMDS to go from the optimization of the information carried by the ranking with respect to the reference object (proximities between the reference and the other objects are perfectly reproduced) to the information carried by the MDS analysis (proximities among all the objects are reproduced as well as possible). This search procedure allows highlighting relevant structures in the data by controlling the information loss about the information carried by the ranking.
BibTeX
| @inproceedings { | |||
| RFC13, | |||
| title | = | "{A Weighted Multidimensional Scaling approach for representing the result of a search}", | |
| author | = | " G. Russolillo and M. Ferecatu and M. Crucianu ", | |
| booktitle | = | "{y-BIS 2013: Joint Meeting of Young Business and Industrial Statisticians}", | |
| year | = | 2013, | |
| month | = | "September", | |
| pages | = | "xx", | |
| address | = | "Instanbul, Turkey", | |
| } | |||
