Extraction of Characteristic Frequent Visual Patterns by Distributed Representation

Abstract

Frequent pattern mining is one of the mostimportant and fundamental tasks in data mining. While alarge number of sophisticated techniques on frequent patternmining are proposed, two essential drawbacks on frequentpattern mining, i.e. the explosion of discovered patterns andless comprehensibility of patterns, are still remained unsolved. In this paper, we propose an application of distributed representationto the area of frequent pattern mining in order toalleviate the drawbacks and to derive characteristic patternswith high understandability. More precisely, given a set ofvisual patterns derived from databases on tagged thumbnailimages in social media, we attempt to identify characteristicvisual patterns by the cluster analysis in the vector spaceobtained by distributed representation. In addition, to helpunderstanding of the meanings of visual patterns, we associateplural tag patterns having similar vector representations witheach visual pattern. A series of experiments are conductedto assess the effectiveness of the proposed framework usingreal tagged thumbnail images in Nicovideo(nicovideo.jp). Theresults confirm that the proposed framework can provide bettervector representation of visual patterns compared with otherdimensionality reduction techniques to identify characteristicpatterns with high understandability.