Online New Event Detection Of Social Networks For Trusted Users And Topic Communities

Micro-Blog, one of the most popular social networking platforms, plays an increasingly important role in information transmission and user communication. Users can follow and share information, access to various network resources, and express their own opinions on the Internet through computers or mobile terminals. Obviously, the micro-blog has become a platform for collecting, sharing, communicating and disseminating real-time information. Furthermore, the Internet produces hundreds of millions of pieces of micro-blog data, which are untapped resources with immense commercial value.While social networks are gaining popularity and maturity, their users are ever more involved and active. Most breaking news is first posted on social networks. As a result, detecting new events has become the focus of the academia and the government. The effect of events multiplies by users’ posting and reposting. Some core users play more significance in the process. So the study of core users has always been the hotspot topic of the academia. However, current new event detection methods on core users lack efficiency.On the basis of traditional algorithms of new event detection, to meet the requirements of reliability, diversity and timeliness in the event detection process, this thesis proposes a new topic detection framework based on trusted users and topic communities. In conclusion, the feasibility, effectiveness and diversity of the above detection method can be proved through experiments, which ensures its high efficiency and stability while maintaining the acuity while detecting new event.The main work and contributions of this thesis are as follows,1. This thesis describes key techniques and research findings of traditional new event detection methods, including new events based from news flow and social networks. Then we proposes a new topic detection framework based on trusted users and topic communities.2. HttpWatch 9.1 is used to capture and analyze Web data stream. Moreover, the massive data is collected by simulating browsers’ behaviors and cleaned with pattern rules in an automatic manner.3. Through analyzing the definition and characteristics of generalized untrusted users, we proposes an algorithm named TR-Score (Trust Relevance Score Propagation Algorithm) to calculate and measure users’ credibility. The TR-Score algorithm gives each user a TR-score value to measure the users’malice and remove untrusted users eventually.4. In order to increase the diversity of topics, this thesis introduces a community division algorithm based on topic information after removing all the untrusted users. The algorithm is evaluated in a public data set and has achieved good results. In addition, community division can reduce the time complexity of the whole process of NED. All topic detection processes can be parallelized.5. Based on traditional core-user analysis, combined with the characteristics of users and events, the thesis proposes the Event-based User Authority Ranking Model (EBAUR). The model is described in details and then evaluated.6. After conducting seed tweet selection based on Tweet Reliability and degree of diffusion, the thesis achieves the Event-Merging Model based on SVM. The model is evaluated in details.Finally, the thesis evaluates the effects of the two ranking models (User Authority Ranking and Tweet Reliability Ranking) in NED with three metrics, namely miss rate, false alarm rate and average detecting time. The results of the experiments indicate that we only have to monitor core users instead of scanning all users. By doing this we hardly increase any miss rate and false alarm rate but dramatically decrease the average detecting time.