Topic Detection with Danmaku: A Time-Sync Joint NMF Approach

Topic detection on web videos can effectively help collecting users’ feedback and emotional tendency. With the features of relatively short, topic alignment and time synchronization, Danmaku comments can significantly extend the applications of topic detection. However, most of the current topic detection approaches fall short of considering the interior relation between adjacent time-steps which ignores the underlying temporal effects. To address this problem, we introduce a Joint Online Nonnegative Matrix Factorization model (JO-NMF) to detect latent topics with automatically exploiting Danmaku comments. Experimental results show great advantages of our proposed model on real-world Danmaku datasets. The results show our model outperforms baselines in topic detection with perplexity and RMSE for the noisy temporal data.