Index-Supported Similarity Search Using Multiple Representations

Similarity search in complex databases is of utmost interest in a wide range of application domains. Often, complex objects are described by several representations. The combination of these different representations usually contains more information compared to only one representation. In our work, we introduce the use of an index structure in combination with a negotiation-theory-based approach for deriving a suitable subset of representations for a given query object. This most promising subset of representations is determined in an unsupervised way at query time. We experimentally show how this approach significantly increases the efficiency of the query processing step. At the same time the effectiveness, i.e. the quality of the search results, is equal or even higher compared to standard combination methods.