Limits and Chances of Social Information Retrieval
Christoph Fuchs · mediaTUM – the media and publications repository of the Technical University Munich (Technical University Munich) · 2016
The prevailing approaches for web search are mainly driven by content similarities and disregard social relationships between the information seeker and the information provider. Furthermore, only explicitly published information is considered. Although several social search approaches exist, only a small subset interprets social search as querying other people's information spaces. Following concepts like homophily from the social sciences, the objective of this thesis is to assess the potential of social information retrieval approaches to satisfy information needs. Therefore, a specific, but also highly customizable social information retrieval concept is developed, prototypically implemented, and evaluated in various usage scenarios. The results allow to identify limits of and success factors for social information retrieval systems. By conducting a survey with 112 participants, we show that using one's social network is a valid method to satisfy information needs, but privacy is considered as a potential threat for information seekers (an additional survey also confirmed the results for information providers, n=608). The analysis of two large social networking datasets from Twitter and Facebook indicate that content from socially close people is perceived as more important by the information seeker than content from other people, affirming social information retrieval as promising method to satisfy information needs. As part of the thesis, a social information retrieval concept is developed that is specific and specific enough to be implemented prototypically, but also sufficiently flexible and parameterizable to cover a broad range of social information retrieval scenarios. The distributed character of the system leads to smaller document collections which allow to apply semantically richer modeling approaches like latent topic models or explicit concept representations. Using these prototypes, various aspects of the social information retrieval workflow are evaluated using (1) datasets covering socially relevant information (scientific abstracts as expertise profiles, social question & answer platforms) and (2) data obtained from a real-world social information retrieval experiment using the developed prototypes with 121 participants in the course of three weeks. The social information retrieval experiment consists of a manual mode relying on human intelligence to route questions and reply to answers (considered as the hypothetical upper bound w.r.t. quality), an automatic mode (routing and content identification done by the system), and a specific use case (social product search). The results confirm that an adjusted interaction pattern successfully mitigates the participants' reluctance to share information. The findings indicate that social closeness is positively correlated with the reply's degree of relevance. Based on the collected data, serendipitous effects can not be linked to social closeness, but appear to co-occur with high degrees of content knowledge similarity. The outcome of the social product search experiment suggests that socially close people are interested in the same products with a higher probability than socially distant people. This could be interpreted as confirmation that social networks can support buying decisions. Overall, the results indicate that social information retrieval is a promising enhancement of existing tools for information gathering, especially for information needs that benefit from personal judgment.