Image tagging using tensor decomposition
Michail Panagopoulos, Constantine L. Kotropoulos · 2015
Social media growing has resulted into a huge amount of information. Meta-data, accompanying the raw data, can assist data manipulation and processing, e.g. the tags assigned to social images. Many systems for automatic image tagging are based on pair-wise relations. Recent approaches focus on relations among multiple vertices (i.e. items), which are respresented as hyperedges in a hypergraph. In this work, an image tagging methodology is proposed that exploits a factorization of a tensor, capturing high-order relations among multiple Flickr images. The proposed approach uses an extended graph, where the vertices represent images as well as users. By analysing the data communities and their similarities, image annotation is proposed.