Modeling Information-Seeking Behaviors During Learning From Hypermedia With nBrowser

• In Event: 44.081 - Data-Driven Techniques That Complement Theoretical Study of Cognition and Learning With Technology

Abstract

Objective. Information-seeking behaviors are fundamental to learning about complex topics in the context of open-ended learning environment, where learners have opportunities to be autonomous and set their own goals while navigating the web (Land, 2000; Segedy, Biswas, & Sulcer, 2014). A primary concern in modeling navigational patterns across hypermedia is how learners regulate their own learning while searching for useful information in achieving learning objectives. This indicates a need to understand how learners select new informational sources and monitor the adequacy of content relative to a goal or sub-goal. This study proposes a new methodology referred to as network-based learner models that leverage web content mining techniques to improve the ability of open-ended learning environments to track learner navigational profiles and recommend the most relevant online resources. We examine the proposed method using nBrowser, an intelligent web browser that implements a spreading activation algorithm to converge a network-based model towards an optimal arrangement based on learners’ self-reported usefulness ratings of online resources.
Methods. A total of 68 pre-service teachers were randomly assigned to either an adaptive or non-adaptive version of nBrowser. The pre-service teachers navigated the web and designed a lesson plan where technology is used to assist students in sharing their work with their peers in the classroom of Linda the math teacher. The most striking result to emerge from the data is that learners across both conditions seldom report the usefulness of online resources in designing a lesson plan. The teachers rated 57% of online resources crawled from the web and featured on the network, and 8% of all websites visited during a session. As such, we aggregated the log-file data obtained from nBrowser at the level of site visits performed by the learners while filtering for only the sites that were rated during a session across both conditions as a means to train a behavioral detection model.
Results. A rule-induction algorithm (i.e., RIPPER; Cohen, 1995) combined with a feature extraction technique (i.e., repeated discretization by binning) was used to predict whether learners submitted a positive or negative rating for an online resource based on several behaviors observed in the log-file data, including (1) elapsed time on a webpage, (2) nBrowser panel selections, (3) total amount of pages visited during the session, and (4) lesson plan edits made while viewing a page in nBrowser. The results obtained from a 10-fold cross-validation procedure where the dataset was repeatedly split in a training and testing set suggests that the learner behavior detection model accurately classifies approximately 70% of examples of positive and negative ratings reported by learners (i.e., Moderate agreement; Cohen’s Kappa = 0.398).
Significance. The learner behavior detection model built in this study stands to complement the self-reported measure that informs the nBrowser recommender system. The low submission rate of usefulness rankings for online resources hinders the ability of intelligent systems to efficiently update the network representation, warranting the use of latent indicators of the utility of online resources to guide system recommendations.

Authors

• Eric G. Poitras, University of Utah

• Negar Fazeli, University of Utah