Gaussian mixture distance for information retrieval

In most Information Retrieval (IR) applications, Euclidean distance is used for similarity measurement. It is adequate in many cases but this distance metric is not very accurate when there exist some different local data distributions in the database. We propose a Gaussian mixture distance for performing accurate nearest-neighbor search for Information Retrieval (IR). Under an established Gaussian finite mixture model for the distribution of the data in the database, the Gaussian mixture distance is formulated based on minimizing the Kullback-Leibler (KL) divergence between the distribution of the retrieval data and the data in database. We compared the performance of the Gaussian mixture distance with the well-known Euclidean and Mahalanobis distance based on a precision performance measurement. Experimental results demonstrate that the Gaussian mixture distance function is superior in the others for different types of testing data.