Evaluating a Visual Presentation of Retrieved Documents

摘要

In this paper we study a visualization system for the purpose of helping the user to locate interesting material in the retrieved data. The system works by placing the documents into 1-, 2-, and 3-dimensional space and positioning them according to the inter-document similarity. We compute the quality of the visualization by simulating a user searching for the relevant material and calculating average precision of that search. We compare the numbers for the visualization with the same measure taken for traditional ranked list. We show that the visualization performs - on average - significantly better than ranked list. We show a significant advantage of multidimensional visualizations over the 1 dimensional one. We also show that the difference between 2 and 3 dimensional visualizations is very small.