Exploitation of information propagation patterns in social sensing

Social sensing;Summarization service;Fact-finder;Social dependency;Polarization;Information propagation patterns;Correlated error;Expectation maximization;Maximum likelihood;Matrix factorization;Hierarchical clustering;Tweet clustering

Amin, Md Tanvir Al

关键词: Social sensing;    Summarization service;    Fact-finder;    Social dependency;    Polarization;    Information propagation patterns;    Correlated error;    Expectation maximization;    Maximum likelihood;    Matrix factorization;    Hierarchical clustering;    Tweet clustering;   

Others  :  https://www.ideals.illinois.edu/bitstream/handle/2142/97270/AMIN-DISSERTATION-2017.pdf?sequence=1&isAllowed=y

美国|英语

来源: The Illinois Digital Environment for Access to Learning and Scholarship

【 摘 要 】

Online social media presents new opportunity for sensing the physical world. The sensors are essentially human, who share information in the broadcast social media. Such human sensors impose challenges like influence, bias, polarization, and data overload, unseen in the traditional sensor network. This dissertation addresses the aforementioned challenges by exploiting the propagation or prefential attachment patterns of the human sensors to distill a factual view of the events transpiring in the physical world.Our first contribution explores the correlated errors caused by the dependent sources. When people follow others, they are prone to broadcast information with unknown provenance. We show that using admission control mechanism to select an independent set of sensors improves the quality of reconstruction. The next contribution explores a different kind of correlated error caused by polarization and bias. During events related to conflict or disagreement, people take sides, and take a selective or preferential approach when broadcasting information. For example, a source might be less credible when it shares information conforming to its own bias. We present a maximum-likelihood estimation model to reconstruct the factual information in such cases, given the individual bias of the sources are already known. Our next two contributions relate to modeling polarization and unveiling polarization using maximum-likelihood and matrix factorization based mechanisms. These mechanisms allow us to automate the process of separating polarized content, and obtain a more faithful view of the events being sensed.Finally, we design and implement `SocialTrove', a summarization service that continuously execute in the cloud, as a platform to compute the reconstructions at scale. Our contributions have been integrated with `Apollo Social Sensing Toolkit', which builds a pipeline to collect, summarize, and analyze information from Twitter, and serves more than 40 users.