CS Class Predicts Course Figures

The debate over preregistration last spring inspired an examination of how to best predict class size and allocate teaching fellows. Now, an algorithm developed by a computer science class last fall may have the solution.

After a semester’s work, the students of Computer Science 96 (CS 96), “System Design Projects,” found they could prevent the misallocation of 1,200 teaching fellows (TFs) and proposed their model of course enrollment predictions to University administrators.

Administrators said they may use the model to make TF assignments in the future.

“We may use some of their techniques, combined with human input on new courses, to estimate section needs next fall,” said Dean of the College Benedict H. Gross ’71.

Because students do not have to choose their courses until the end of the first week of the semester—after shopping period—the College must estimate in advance how many teaching fellows a course will require.

Many graduate students, who depend on these positions for teaching experience and money, often find themselves either in a course with too many or too few students to instruct.

To avoid last minute TF shuffling, Dean of the Faculty William C. Kirby proposed last fall a system of preregistration, under which students would turn in their study cards one semester in advance.

The proposal met with strong opposition from students and faculty; in response, Kirby tabled the plan.

According to its course description, CS 96 aimed to resolve the debate on preregistration, reporting its “target application is prediction of student enrollments based on historical data.”

Gordon McKay Professor of Computer Science Stuart M. Shieber ’81, who taught the course, said that the administration actively encouraged its creation.

Currently, the college estimates need for teaching fellows, textbooks, and classroom size, but these estimations are not standardized in a mathematical model.

After surveying 13 percent of Harvard undergraduate and graduate students, the CS 96 students decided that TF hiring efficiency could be improved with better modeling of course enrollment, according to Sheiber.

Shieber’s students canvassed over 10 years of data—historic enrollment, course type, department, CUE guide ratings, time slot, and other variables—to determine an algorithm, which Shieber called a “machine learning system” (MLS).

To test the program, the students in previous years against the program. On average, the system lowered human error in allocating TFs approximately three-fold. If the College were to implement the program, according to Shieber, it could cut down on its misplacement by 1,200 TFs.

“Personally, I thought that we did a pretty good job since what we were doing was highly theoretical,” said Christopher D. Bockman ’06, a student in CS 96. “We did the best we could with the given time that we had.”

Students also examined the deviation of section size from the ideal of 15 students to evaluate the misallocation of TFs. For example, larger sections indicate that not enough TFs were hired for a course.

When comparing expected section size deviation, the MLS reduced the deviation by about half, minimizing the misassignment of TFs.

Shieber warned, however, that the CS model itself is not foolproof. For example, it would not recognize that a course such as a sophomore tutorial is a required class for all concentrators and would most likely underestimate the number of students who will take the course.

And though it can correct TF misallocation, the model would not address classroom allocation, textbook shortages, and coursepack availability, Sheiber said.

But with human corrections, he maintains that MLS gives fairly good predictions for course enrollments.

The class recommended that the College implement a policy of “mandatory non-binding preregistration,” such as anonymous preregistration, to supplement the program.

Dean of the Graduate School of Arts and Sciences Peter Ellison said that although he had not seen a formal report from the class, he is very interested in the implications of course enrollment predictions.

“We will never be able to make all the [TF] appointments in advance, since no enrollment prediction methods are foolproof,” Ellison said. “[But] if the CS 96 project is helpful in this, I would love to use it.”

—Staff writer Risheng Xu can be reached at xu4@fas.harvard.edu.