We propose a method for finding similar objects in a large-scale database, based on a graph index. Each vertex corresponds to an object, and two similar vertices are likely to be connected. The graph index, constructed by a “Like likes like strategy,” shows small-world behavior: any two vertices can be connected by a small number of steps. Graph index search terminates quickly and is applicable to many media, including text, image, and audio. We introduce “Pitarie”, an application of the algorithm which allows searching for similar picture books by both text and image.
Innovative Communication Laboratory