Topic networks deal with developing image representations using image features, for example, bag-of-words representations. However, the representation of social media images (or social images) and the web’s visual contents, called multimedia objects, are the issues of social media analysis. Social images accompany user tags, which are associated with images and are closely related to image content. This paper addresses the issue of building better image representations in social images.

The authors propose a new visual topic network (VTN) model, modeling content and relations of the images simultaneously. This method contributes to hierarchical image representations for social images, defines image relations, organizes image collection using correlation, and models multi-level image relations using a link probability function. The authors claim that the motivation for the proposed VTN model is the relational topic model used in document analysis.

They use two social media data sets crawled from Flickr (NUS-WIDE and MIRFlikr-25K), containing 269,648 and 25,000 images linked to 1,000 and 1,386 tags with 81 concepts and 23 labels defined in the data sets, respectively. For experiments, they compare term frequency-inverse document frequency, latent Dirichlet allocation (LDA), context-and-content-based multimedia retrieval (C2MR), and VTN, having variants such as supervised-LDA, supervised-C2MR, unsupervised-C2MR, and VTN (binary- and multi-) valued quantization.

In the future, the authors intend to enhance the scope of the proposed algorithm while modeling image relations to other social media information. They provide further insights on extending VTN models to semi-supervised VTNs, which makes this paper worth reading.