Accelerating approximate similarity queries using genetic algorithms

Searching for the exact answer to a similarity query is an expensive process considering computational resources, such as memory and processing time requirements. Moreover, comparison operations over multimedia data is even more expensive than over traditional data such as numbers and small character strings. Therefore, when comparing multimedia data, the comparison computations usually consider some properties extracted from the data elements. In this way, exact queries involving this kind of data return data that is exact regarding the properties compared, but not necessarily exact regarding the multimedia data itself. For example, searching for similar images regarding their colors return images whose color histogram are the most similar, but the retrieved images can be very different regarding, for instance, the shape the objects pictured. Therefore, for applications dealing with complex data types, trading exact answering with query time response can be worthwhile. In this paper we propose to use techniques based on genetic algorithms to allow retrieving data indexed in a metric access methods within a limited, user-defined, amount of time. We show that these techniques lead to much faster execution, without reducing the quality of the answer. We also present experimental evaluation using real datasets, showing that suitable results can be obtained in a fraction of the time required to obtain the exact answer.