Burst Topic Detection Oriented Large-Scale Microblogs Streams

In microblogs, emergent events spread quickly and produce tremendous influence. Burst of public opinion is widely concerned by government and enterprise. Existing burst topic detection methods only consider one type of entity, such as word or tag. However, Chinese microblogs contain not only new or colloquial words, but also contain some pictures and links, burst patters of which are difficult to detect. To tackle this problem, we propose a real-time burst topic detection framework for multi-type entites. Different from existing method, our method does not require Chinese word segmentation, but generates new words lastly. In this framework，the window size is adjusted based on the microblogs streams dynamically. In order to measure the burst weight of entity, the spread influence of entity is calculated. Moreover, the high order co-clustering algorithm based on non-negative matrix decompostition is used to cluster two types of entities, message and user simultaneously. While the detection of burst topic, we can also obtain the related messages and participating users, which can be used to analyze the cause of burst topic. Experimental on a large Sina Weibo dataset show that our algorithm has higher accuracy and earlier detection of the burst topic compared with the existing algorithms.