Presenting tiered recommendations in social activity streams
Modern social networking sites offer node-centralized streams that display recent updates from the other nodes in one's network. While such social activity streams are convenient features that help alleviate information overload, they can often become overwhelming themselves, especially high-throughput streams like Twitter’s home timelines. In these cases, recommender systems can help guide users toward the content they will find most important or interesting. However, current efforts to manipulate social activity streams involve hiding updates predicted to be less engaging or reordering them to place new or more engaging content first. These modifications can lead to decreased trust in the system and an inability to consume each update in its chronological context. Instead, I propose a three-tiered approach to displaying recommendations in social activity streams that hides nothing and preserves original context by highlighting updates predicted to be most important and de-emphasizing updates predicted to be least important. This presentation design allows users easily to consume different levels of recommended items chronologically, is able to persuade users to agree with its positive recommendations more than 25% more often than the baseline, and shows no significant loss of perceived accuracy or trust when compared with a filtered stream, possibly even performing better when extreme recommendation errors are intentionally introduced. Numerous directions for future research follow from this work that can shed light on how users react to different recommendation presentation designs and explain how study of an emphasis-based approach might help improve the state of the art.
Recommender systems, Visualization, Twitter, Social media, Human factors
Master of Science (M.Sc.)