Encoding Local Correspondence in Topic Models

Exploiting label correlations is a challenging and crucial problem especially in multi-label learning context. Labels correlations are not necessarily shared by all instances and have generally a local definition. This paper introduces LOC-LDA, which is a latent variable model that adresses the problem of modeling annotated data by locally exploiting correlations between annotations. In particular, we represent explicitly local dependencies to define the correspondence between specific objects, i.e. regions of images and their annotations. We conducted experiments on a collection of pictures provided by the Wikipedia "Picture of the day" website, and evaluated our model on the task of "automatic image annotation". The results validate the effectiveness of our approach.