Time-aware online reputation analysis

Social media has become an integral part of society. Omnipresent mobile devices allow for immediate sharing of experiences. Experiences can be about brands and other entities. For social media analysts a collection of posts mentioning a brand can serve as a magnifying glass on the prevalent opinion towards a brand: The overall estimation of a its reputation is increasingly based on the aggregation of a brand's reputation polarity in social media posts. This polarity of reputation is currently annotated manually. However, with the dramatic increase of social media, this is no longer feasible.
This thesis aims to facilitate and automate parts of the process to estimate the reputation of a brand. We motivate this by performing user studies with expert social media analysts. We analyse three resulting datasets: a questionnaire, log data of a manual annotation interface, and videos of annotating experts following the think-aloud protocol.
Based on the indicators used for manual annotation, we proceed with the development of algorithms for the automatic estimation of reputation polarity. Unlike earlier, static evaluation scenarios, we follow a dynamic scenario, which mimics the daily workflow of social media analysts. Our algorithms are successful because we distinguish between reputation and sentiment.
The second part of this thesis is motivated by the analysts' desire for automation of retrieval and filtering of new media. For information retrieval, we present two improvements to existing algorithms.
We conclude that many aspects of the annotation of reputation can be automated - using in particular time series analysis, memory models, and low-impact help from expert social media analysts.