Description

Abstract

Markovian Semantic Indexing, a new method for mining user queries by defining keyword

relevance as a connectivity measure between Markovian states modeled after the user queries. The

proposed system is dynamically trained by the queries of the same users that will be served by the

system. Consequently, the targeting is more accurate, compared to other systems that use external

means of non dynamic or non adaptive nature to define keyword relevance. A stochastic distance,

in the form of a generalized euclidean distance, was constructed by means of an Aggregate

Markovian Chain and proved to be optimal with respect to certain Markovian connectivity

measures that were defined for this purpose. A comparison to Latent Semantic Indexing and

probabilistic Latent semantic Indexing revealed certain theoretical advantages of the proposed

method (MSI). Experiments have shown that MSI achieves better retrieval results in sparsely

annotated image data sets.