Images of similarity: A visual exploration of optimal similarity metrics and scaling properties of TREC topic-document sets

Multiple similarity measures for five TREC topic-document sets from the LDC TREC Collection Disk 1 are derived from the full text of documents. Each m easure on each set is scaled using SAS MDS under ordinal, interval, and MLE assumptions. The resulting 75 permutations are ploted. It is suggested tha t cosine-vector and overlap measures for similarity appear to recover optim al data relationships among the documents of the five sets. MLE assumptions appear to be required to model the data adequately.