Hot Event Tracking And Visual Summary Research And Implementation On Weibo

With the rapid development of the microblogging service platform, the number of Weibo users is growing significantly as well. People spread and obtain information about hot event via Weibo. However, the number of these microblogs is extremely large while the contents are highly similar. What’s more, the quality of the information is varied. Therefor, it urgent needs related technology to organize them. For a collection of microblogs which are related to a hot event and the embedded images, the aim of this thesis is to detect and track the sub-events and select representative images for sub-events. The problem of event tracking is converted into cluster tracking in this thesis, and track the clustering evolutionary mode over time.The microblog stream is modeled as a dynamic post network in this thesis. Then the problem of event tracking is converted into dynamic clustering tracking on the dynamic post network. The clustering algorithm CDBSCAN is proposed via improving DBSCAN algorithm, which is appropriate for incremental clustering. Some research is based on single update incremental clustering algorithm. In order to improve the efficiency, the batch update incremental clustering algorithm is adopted in this thesis. After a fixed time interval, the batch delete update incremental clustering and the batch add update incremental clustering is employed at each moments while the cluster evolutionary mode is recoded. According to the record, the cluster evolutionary mode between neighboring moments can be obtained. The results of experiment show that this algorithm can detect major sub-events efficiently and track the whole life of cluster evolution.There are a few research which simultaneously use text and images in the event tracking field. The author proposes a visual summary algorithm to select representative images for event. At first, the noise images are filtered according to their attributions. Then a model by SVM based on color histogram and edge direction histogram is learned to recognize noise images and filter them. As for the images in the same text cluster, after CDBSCAN clustering, the image of highest priority will be selected in every image cluster. Then these selected images are sorted by hot degree. The results of experiment show that the selected images are close to the event and can help users to understand the event. A hot event tracking and visual summary system is implemented at last. This system adopts incremental clustering to track clustering, and adopts visual summary to get representative images.