CHS: Small: Promoting Unexpected Information Discovery: An Interactive Framework for Computational Serendipity

University Of North Carolina At Charlotte, Charlotte NC

Investigators

Abstract

Serendipity is a concept associated with unexpected discoveries that are valuable. However, most existing recommendation and information retrieval algorithms, including those used by common search tools and social media, do not include serendipity as an element they use to choose which content to show. Instead, the criteria used to build and evaluate these algorithms rewards choices that reinforce what people already know or believe, rather than promoting unexpected, serendipitous discoveries. This project's goal is to make progress on how to predict and foster serendipity when people interact with information online. One key challenge is how to measure potential serendipity; to do this, the team will create measures that capture important subcomponents of serendipity such as surprise, value, and curiosity. Another is how to develop algorithms that promote serendipity while still addressing people's needs for relevant information; toward this, the team will create algorithms that balance relevance and serendipity over time and that develop better models of the values and curiosity of the people who use them. Doing this promises to make contributions to the study of recommender systems, information retrieval, and human-computer interaction, along with developing real tools to support serendipity and knowledge discovery in libraries. Specifically, the investigators will first develop an interactive framework for computational serendipity that is independent of application domain. The framework consists of four components: Surprise, Value, Curiosity, and Sequence Composition. For the Surprise component, we hypothesize that users will be surprised when presented with items that violate their expectations as predicted by our computational model of them. We will study two computational models of surprise based on recent advances in text mining and deep learning techniques. The Surprise component will be balanced by a Value component, implemented with a traditional collaborative filtering (CF) recommender approach to ensure that the surprising item is liked by the user. The Curiosity component reasons about personalized levels and patterns of surprise that stimulate users' curiosity and sustain their interest to explore. The Sequence Composition component synthesizes the outputs from the other three components, identifying how much surprise is just surprising enough to be approachable, but not so much to generate anxious feelings. The output is a sequence of suggestions that guides the user toward increasingly surprising concepts, to engage their curiosity over time. User feedback will be incorporated to update the Surprise, Value, and Curiosity components for discovering new sets of serendipity sequences. As a use case, the investigators will implement and evaluate the framework as a serendipity recommender, SerenCat, designed to promote and study students' discovery in learning materials in a university library environment. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.