Online semi-supervised web video classification via heterogeneous attribute fusion

Learning large scale of web video data requires considering unlabeled data and heterogeneous information. A novel online semi-supervised learning method is proposed for the web video classification, which adopts graphs as base classifiers on each view of texts and videos, and propagates labels by linear neighborhood propagation algorithm. The unlabeled data are chosen online with co-training strategy on multiple graphs and base classifiers are incrementally updated. The proposed method increases the classification accuracy and is suitable for online semi-supervised learning of large scale of web video data. Experimental results show that the average accuracy of this method is approximately 8.3% higher than support vector machines, and the accuracy of learners after the online incremental learning increases by approximately 3%.