Correlation consistency constrained probabilistic matrix factorization for social tag refinement

With the permeation of Web 2.0, large-scale user contributed images with tags are easily available on social websites. However, the noisy or incomplete correspondence between images and tags prohibit us from precise image retrieval and effective management. To tackle this, we propose a social tag refinement method, named as Correlation Consistency constrained Probabilistic Matrix Factorization (CCPMF), to jointly model the inter- and intra-correlations among images and tags, and further to precisely reconstruct the image–tag correlation as a result. For CCPMF, we attempt to derive two low-rank factors by conducting a joint factorization upon the image–tag correlation matrix. Besides, two types of correlation consistency, i.e., the image–bias correlation consistency (from image similarity to tag relevance) and the tag–bias correlation consistency (from tag relevance to image similarity), are formulated as constraints in the factorization process. Finally, each untagged or noisily tagged image can be retagged according to the reconstructed image–tag correlations with the both derived latent factors. Experimental results on the NUS-WIDE dataset show the encouraging performance of our proposed algorithm over the state-of-the-arts.