Image tag recommendation based on novel tensor structures and their decompositions

In this paper, we address the problem of image tagging and we propose automatic methods for image tagging, using tensor decompositions. Tensors are a suitable way of mathematically representing multilink relations. Another, complementary structure that captures the aforementioned high-order relations is the hypergraph. More specifically, three different matrices are derived from the hypergraph, namely, the incidence, adjacency, and affinity matrices. The just mentioned matrices are used to create slices of a novel tensor structure, which combines users´ and images´ relations. Four methods are exploited to decompose the tensor, i.e., the Higher Order Singular Value Decomposition (HOSVD), the Canonical Decomposition/Parallel Factor Analysis (CANDECOMP/ PARAFAC, CP), the Non-negative Tensor Factor Analysis (NTF), and Tucker Decomposition (TD). Experiments conducted on a dataset retrieved from Flickr demonstrate the potential of the proposed approach.