AI Course Recommender
My Roles and Responsibilities
Primary UX Designer, Prototyping, Wireframing, Qualitative Research
Team
Brighten Jelke, Quentin Griffin
DePaul University is one of the largest private universities in the world, serving nearly 21,000 students and offers over 130 majors. With this large amount of variety, comes a great deal of information across both undergraduate and graduate levels.
The degree requirements for graduation consist of combinations of core-courses in each students’ major, as well as elective courses to provide a curated education geared towards students’ interests and career goals.
Students can, and are encouraged to, work alongside their academic advisors to help craft their course schedules to meet these requirements, tangential to their own academic and career goals.
Problem Space
Students currently suffer from a lack of centralized information regarding future course offerings and the academic content of courses pertaining to their interests. Much of the time spent interacting with advisors is dedicated to finding this information rather than discussing the student’s options.
By implementing a Machine Learning/Artificial Intelligence model for course recommendations, we hope to address these aforementioned issues in the following ways:
Allowing students to input their preferences through keywords, time/modality, and course characteristics
Considering the above information as well as the student’s history of coursework, their prerequisites needed, and the college’s course offering history
Dynamically recommending a course schedule by quarter from the current term to the end of their academic career at DePaul with multiple variations
User and Business Needs
Our main user base was identified as current DePaul students and advisors. However, we acknowledged that there were varying levels of distinction between undergraduate (rigidity based on requirements) and graduate (personalization and tailoring of education) students when creating their course schedules. Advisors were identified as facilitators for decision-making and discussion around course schedule crafting, as noted in our initial discussion with subject matter experts (students and faculty at varying levels).
When defining our problem, we utilized the Lean Canvas method to help understand where the current issues are in the course scheduling process, what we should research, and where effectve data-driven design can be implemented.
Our research objectives were to:
Document current course registration/search process, systems, and sources of relevant information
Identify student pain points in course scheduling process
Validate assumption that an ML/AI course recommender would be an improvement
Research Methods
We planned to conduct user and expert interviews, do a walkthrough of the current course registration system as a team, and synthesize results with affinity diagramming.
We began with conducting initial interviews with the following:
[8] Target Users Interviews (DePaul Undergraduate or Graduate students)
We took into consideration whether the student was at the undergraduate or graduate level, and how the course scheduling experience may have differed
[2] Expert Interview (DePaul Academic Advisor / CampusConnect Expert)
We took into consideration whether the advisor was at the undergraduate or graduate level to inform our understanding of advisor roles in course scheduling
Initial Interviews
From our 8 initial qualitative interviews, we found:
Some students don't go to their advisor for course recommendations at all; could be due to anxiety, lack of trust or deep relationship
Time spent on course scheduling varies by student bandwidth, regardless of motivation level
Advisors have different approaches to advising
Faculty advisors have much less control over course logistics (course substitutions, timing of offerings) than academic advisors
Graduate students are more concerned with finding specific topics rather than logistical issues
Undergraduate students have more difficulty deciding when to take courses
Some students choose courses based on their interests, while others choose them based off of the instructor or fellow peers
Here are some notable quotes from our conversations:
“My (decision-making) process is very much tied to what other people are doing.” - Graduate Student
“My advising with students usually involves a lot of back-and-forth. I tell them, ‘Let me send you the info when I find it so that we don’t waste time’” - Faculty Advisor
“I feel nervous to meet with my advisor, and I only ever contact them via email if I have a specific question.” - Undergraduate Student
“Each advisor has their own varying perception of their role, and it exists on a spectrum” - Faculty Advisor
Key Metaphors and Design Principles
From our insightful conversations, we synthesized all of the information into key metaphors to understand how to bridge the existing gap between student and advisors. Our three pillars guiding us forward were: centralization, personalization, and facilitation. We identified that students were oftentimes too overwhelmed to find the information meaningful to them, but still wanted to personalize and tailor their class schedule to their needs. Our team also did not want to cut ties from the advisor’s role in a student’s academic career, so we prioritized this in our work.
Data Journey Map
Creating a data journey map was imperative in designing the AI/ML model, as we had to understand where current pieces of data were located, transferred, and transformed. Additionally, we wanted to understand where our opportunities to automate the process were. We found that the main areas for data-driven design were during the stages where students manually check their course history, credits needed, and course availability.
Data Types and Algorithm Selection
Outlining the available data types and aggregate data needed informed our choice for a recommendation algorithm using content-based filtering (recommends similar content) in the initial stages and collaborative filtering (recommends content from similar users) once more user data is available.
However, we had some considerations regarding sharing information about other students, course evaluation details (such as student information), as well as prioritizing informed decision making through transparency on data collection. For us, all of the data we will be using will either: (1) be already available to the university or (2) be input by the user.
Prototype
From our data identification, we created a functioning prototype of how this AI model would recommend both an undergraduate and graduate student courses until graduation. Some of the main features include:
Onboarding with tooltip regarding data collection information and consent
Preferences input for users with keywords and course evaluation metrics
Course schedule recommendation with sharing to advisors / peers
Design Decisions, Next Steps, and Reflections
When testing our first iteration of the prototype, we made three key decisions based on feedback received from usability test participants:
Removed ranking system of upvoting/downvoting recommendations due to ethical concerns of preferential treatment (classes with specific professors being downvoted more frequently simply because of the professor)
Added prompt to meet with advisor at the end to facilitate student-advisor relations and conversation around course scheduling, as well as provide clarity on the expectation of the Course Recommender (only recommends, does not register students for classes)
Clarified “Consider Past Student Evaluations” and “Keywords” sections with tooltips, which now includes explanation of student evaluation metrics (1-5 scale) and types of keywords (professor name, course name, subject matter)
Since our time was limited to ten concentrated weeks, our next steps in this project would be to
Work with data scientists to refine data collection points
Conversion of course evaluations into measurable and aggregate metrics
Run a higher-fidelity test with actual data to see if recommendations align
Implement recommendation feature based on what others within one’s department are taking
This project was one of the unique cases where I was able to implement my skills in UX design in a machine learning context through both research and digital design. Working directly with data encouraged me to further my understanding of data analysis and model building to better understand how to not only work with data, but ensure it is used ethically as a designer. Less than a year after this project, I earned an Advanced Data Analytics certificate from Google, which taught me how to build such models using real data; I hope to work as closely with data throughout my career.