Justify role of Similarity Diffusion Process in cross-media topic ranking: an empirical evaluation

Recently top performing cross-media topic detection employs Similarity Diffusion Process (SDP) to rank the interestingness of topics from a large number of candidates. SDP models the polysemous phenomenon from short and less-constrained user-generated data by assuming the similarities between two multi-media data should be divided into intersected topics. The noise in SDP plays an important role to explain the generation of the similarity. However, it is unclear what kind of noise is more appropriate for different modalities in cross media: SDP under different noises should has the lower false positives when topics are successfully detected. In this paper, we provide an in depth analysis of two types of noises (Poisson and Gaussian) for this task. In the evaluation, we observe that the combination of Poisson noise and topic sizes performs best while Gaussian noise has a faster optimization speed than that of Poisson one.