Learning Vocabularies to Embed Graphs in Multimodal Rank Aggregation Tasks

Abstract

This paper introduces Supervised Bag of Graphs (SBoG), a supervised vocabulary learning approach for multi-modal graph-based rank aggregation tasks. In our formulation, collection objects are represented based on complementary views provided by different ranks, defined in terms of multiple modalities. Ranks are encoded into a graph (fusion graph), which is later embedded into a vector representation (fusion vector), based on a vocabulary of graph words. SBoG explores different strategies for exploring collection labels to define suitable vocabularies that lead to effective representations. Experiments considered the use of SBoG-based representations in multimedia classification tasks. Obtained results demonstrate that SBoG leads to gains up to 28% when compared with state-of-the-art and traditional approaches.