Precision Education: Using Prior Students' Data to Personalize Instruction for Future Students

• In Event: 34.075 - Symposium on Intelligent Tutoring Systems, BIG DATA–Learning Analytics, and Automated Humanlike Tutoring: Similarities and Differences

Abstract

As students use technology-based instructional systems, they generate a wealth of electronic data. These large data sets offer tremendous new opportunities to create personalized instructional experiences for individual learners by mining the data from prior users. Machine learning algorithms can be used to explore the data set, and recognize and automatically recommend different pathways through instructional space, some of which will be better than others for particular students.

Much research points to the benefits of personalized tutoring (e.g., Bloom, 1984). For example, human tutors present problems that help diagnose the student’s sources of difficulty, choose problems within a student’s “Zone of Proximal Development” (meaning problems that can be solved with expert assistance), scaffold the student to a successful solution, and then attribute the success to student effort and enhanced understanding (Brown, Ellery, & Campione, 1994; John-Steiner & Mahn, 1996; Lepper, Woolverton, Mumme, & Gurtner, 1993). Expert tutors balance the goals of providing challenge to ensure cognitive conflict, resolution and improved understanding, with the goal of sustaining the learner’s motivation and persistence (Merrill, Reiser, Ranney, & Trafton, 1992).

The model of the “intelligent tutoring system” (ITS) is based on the idea that a machine can create the experience of working with a human tutor. These highly-structured systems involve adaptive instruction based on a model of what the student is estimated to know relative to a defined set of skills (Woolf, 2009). In an ITS, students work “at their own pace” but, in effect, they all follow the predefined instructional pathways that have been determined by domain experts. One limitation of the ITS approach is the cost, estimated to be 200 hours of knowledge engineering to generate an hour of adaptive instruction (Woolf & Cunningham, 1987). Also, the hand-crafted approach represented in the ITS model is rapidly being overtaken by a user-driven model in which students themselves locate the learning objects they need (e.g., Khan Academy). And although benefits have been associated with intelligent tutoring, there are substantial inconsistencies in findings and effect sizes across studies (Kulik & Fletcher, 2016; Steenbergen-Hu & Cooper, 2013).
The presentation will outline an emerging alternative to the ITS model in which instruction is personalized via a data-driven approach. Students are guided through the material based on estimates of their states and, importantly, their similarity to prior students who were successful (Authors, 2009; Koedinger, D’Mello, McLaughlin, Pardos, & Rose, 2015). Examples from project data involving an instructional system used by over 200,000 students in more than 60 school districts will be included to illustrate the feasibility and potential of the precision education approach.

Author

• Dexter Fletcher, Institute for Defense Analyses