Amazon.com uses recommender systems to customize their store for every user. “It’s as if you walked into a store and the shelves started rearranging themselves, with what you might want moving to the front, and what you’re unlikely to be interested in shuffling further away” (Two Decades of Recommender Systems at Amazon.com).
Could the vision of Amazon– a store that carries every product and is customized to every customer– be extended to schools? Under this vision, an online school would transform itself based on the needs and preferences of each student. Recommender systems that help students choose what and how to learn could bring us one step closer to “The Everything School.”
Khan Academy’s mission, “to provide a free, world-class education to anyone, anywhere,” aligns with this vision of The Everything School. However, Khan Academy, and the online education sector in general, has not utilized recommendation systems in ways that are as sophisticated as online retail or content streaming sites. A scenario analysis, followed by some recommendations, provide a sketch of some steps toward revolutionizing education as Amazon revolutionized retail.
Khan Academy’s target users are students, parents, and teachers, almost all of whom are already participating in traditional schooling. Before March 2020, Khan Academy was typically used as a tool for enrichment or remediation by school teachers. Pandemic-related school closures pushed Khan Academy toward the center of many virtual classrooms, providing the majority of direct instruction and student work materials.
A second group of Khan Academy users are adults in college. KA offers content from a variety of common undergraduate-level courses, especially in math and computing. These users may be supplementing their education or learning independently.
The key goals of these users are to gain sufficient skill and understanding so that they can succeed in their traditional coursework. Although traditional schooling structures have been weakened by the pandemic, they are still the norm for the vast majority of K-12 and undergraduate-age students in the United States.
Khan Academy helps these users accomplish their goals by providing videos with direct instruction content on a variety of topics commonly taught in K-12 schools and undergraduate programs. KA also provides a large library of practice problems and prompts. KA provides some basic tracking and diagnostic tools that function slightly differently for students, parents, and teachers. There is a gesture toward gamification, with badges and points distributed for sustained engagement and correct responses from students.
In this hypothetical future scenario, target users are students, parents, and teachers. A growing number of parents and students are participating in independent schooling, guided by recommender systems, state guidelines, and personal interests. An additional group of learners is composed of working professionals interested in strengthening a particular skill. While there are websites in existence that already target both these groups, Khan Academy might be able to attract a growing share of them with a wider variety of content, zero cost, and expertly deployed recommenders.
The key goals of these hypothetical future users would be to receive a more individualized education than traditional K-12 schooling typically provides, or to acquire professional skills outside of a degree program.
Khan Academy could help these users accomplish their goals by providing a wider variety of educational content. In addition to instructional video and practice exercises, Khan Academy could include moderated message boards, live expert guest speakers, a digital text library, etc. By gathering user data and developing recommender algorithms, Khan Academy could guide students through a large array of learning opportunities according to their preferences for learning.
Currently the recommender system on Khan Academy is very basic. For K-12 math courses, KA offers Course Challenges, which identify a student’s strengths and weaknesses across certain areas of content in a course, and then refers students back to items pertaining to weaker skills. There is no provision for recommending a course to a student, under the assumption that a student’s coursework is determined at their traditional school. When content on KA is recommended, it sometimes acquires a golden star icon, although the content on the page is not rearranged or replaced according to recommendations. The KA recommender system at present is more like a simple decision tree, moving students back and forth through course content along a single predetermined path.
Right now, Khan Academy resembles a high-quality tutoring center rather than a school. This is largely due to its small variety of types of content, as well as it’s one-size-fits-all model for how a student might proceed through a course. (Perhaps the number is slightly greater than one, since there are several curricula for the same K-12 math courses on the site.) There are several things that KA could do to improve its utility to users by implementing recommender systems.
First, Khan Academy needs a much wider variety of content within the same course. Just as Amazon has many brands of similar products and serves as a marketplace for many vendors, Khan Academy could build on the YouTube model of content acquisition to also include all the other elements of effective instruction. Of course, quality control would be a major concern– one that Amazon and YouTube are also still working on.
Once KA had a great variety of content, there would be opportunities to start building a more interesting recommender system. One idea comes from the paper “Recommender System for e-Learning Based on Personal Learning Style.” The authors argue that students could be directed to content in different modalities (video, text, exercises, etc) based on their binary comparisons of other content they’ve already seen. The authors call this an Active Pairwise Relation Learner model, and note that it has been successful in helping customers select used cars. The basic strategy is that material is coded as belonging to any of several classes based on content, performance level assessed, learning style, and main sense modality. As students indicate their pairwise preferences (A is preferred to B, B preferred to C, etc.), the algorithm generates "a set of rules consistent with all the positive examples given by the users, such as:
(Visual) betterthan (Auditory) (Auditory) betterthan (Intrapersonal and Deductive)"
I think this idea is an interesting jumping-off point, but I wonder whether students’ preferences are always for the material that actually helps them learn the most. In my experience, some students consistently prefer low-challenge work, and some prefer to work on topics that they’re not prepared for. Both these preferences would result in little learning. This is an important point of divergence between recommendations on Amazon or YouTube, and on an education site: Whereas on Amazon or YouTube, “liking” and consuming the content constitutes success for the recommender, on an educational site, the goal is not merely for content to be consumed– it’s for learning to happen. Measuring that, and using these measurements as input for a recommender, would be a very interesting project to work on.
Linden, Greg, Brent Smith and Jeremy York. “Amazon.com Recommendations: Item-to-Item Collabroative Filtering.” IEEE Internet Computing. Jan/Feb 2003.
Qomariyah, Nunung Nurul and Ahmad Nurul Fajar. “Recommender System for e-Learning based on Personal Learning Style.” 2019 International Seminar on Research of Information Technology and Intelligent Systems.