Dual linkage refinement for YouTube video topic discovery

Understanding user generated videos have been an ever interesting research recently. While the amount of videos on video sharing websites, such as YouTube, becomes huge, the cost of visual content computation and the semantic gap make the text-based information to be the first choice for labeling work. However, text information is deficient and noisy on YouTube. In this paper, we propose the novel dual updating method for YouTube video topic discovery. We first enhance the document representation for each video with its related videos, then we extract meaningful topics via keyword cores, at last, the video response links and the correlations between keyword cores are used to refine the video soft clustering result. Experiments show that our method can give reliable topic descriptions and our document representation can help to increase the performance of common methods.