Olin Course Prediction

A prediction tool for Olin College course enrollment based on preregistration survey data and Olin's course history.

Usage

python olin_course_prediction.py

Preconditions

Installation of scikit learn and its dependencies: http://scikit-learn.org/stable/install.html

Also, Make sure to specify the locations of the Olin Course History data and Pre-Registration survey data in the main function of olin_course_prediction.py. Those files should be in the 'csv' format. Additionally, all of the pre-registration survey csv-files should be speicific to one semester and contain the semester information in the filename (eg "FA10").

Output

After running olin_course_prediction.py, there will be 2 files in the /results directory:

• test_data.xls: the course enrollment predictions for the 5 models (discussed below) we're considering to predict course enrollment with one sheet per course

• model_comparison.xls: one sheet with the cumulative error of each model compared with the acutal past enrollment of those courses (for the semesters with pre-registration data).

5 Models

To gauge how effective our predictions are, we run through the predictionson 20 different coures offered consistantly at Olin for 5 different models , each having different amounts of feature input data:

• Baseline Predicted Enrollment: No features, trains on past enrollment trends

• Gender Feature Enrollment: the baseline predicted enrollment model with the addition of the gender of the sudent added to the feature data

• Prereg Predicted Enrollment: the gender feature enrollment model with the addition of pre-registration data to the feature data

• Course History Predicted Enrollment: the gender feature enrollment model with the addition of feature data values (boolean) for every course at Olin as well as the major of the student

• Prereg + Course History Predicted Enrollment: the combination of all of our predictive strategies; this model should be the most effective at predicting course enrollment

Additionally, we compare each of these models' predicted enrollment with the actual enrollment and the pre-registration survey predicted enrollment.

Data Structures (Models)

There are several objects used to store the data from Olin's Course History in an effective way:

• Student: stores information like gender, starting/ending semesters, list of course offerings taken, and major

• Course: represents one course, like Software Design

• Course Offering: represents an offering of a single course, like the Fall 2014 offering of Software Design

• Major: and concentration

• Professor

• Graduating Class: this was initially designed to include information about the requirements for specific classes, but we didn't end up using it much

Predictive Model

We used scikit learn's Logistic Regression model:

http://scikit-learn.org/stable/modules/generated/sklearn.linear_model.LogisticRegression.html

Using the linear regression model, we can predict whether or not a student is likely to take a course in a specific semester, given their grade level.

So, to predict the enrollment of Software Design in the 1415FA semester, we make 4 separate linear regression models, one for each grade of students. These models consider the course histories of all students, noting those who took Software Design in the fall semester of the specified grade as well as the preregistration data for how many students within that grade said that they would take Software Design. The model then predicts for each student at Olin that semester the probability that the student would take Software Design. These probabilities are summed to determine the predicted total enrollment of the course.

We have found that a C value (input to the Logisitc Regression model) of 0.1 achieves reasonable results, although more fine-tuning could be done to find the optimal C value.

Our Results

The graph below shows the cumulative error of each model compared with the acutal enrollment for the following semesters: SP11, SP12, FA12, SP13, and FA13.

The light blue bar represents the Prereg + Course History Predicted Enrollment model, which outperforms the other models, but still does not do as good a job at course prediction as the pre-registration survey as a whole.

Suggestions for Improvement and Further Work

Improve the signal between the pre-registration data and a student's choice of course

Currently, the pre-registration survey provides data on how many people from each grade are expected to take a particular course. On the other hand, our model predicts what a single student will do, given their history. Even though the pre-registration data is the clearest signal of data we have to predict course enrollment theoretically, the weight of the pre-registration survey feature is not comparatively high in the simulations.

We suggest either attempting to predict the enrollment for a semester, not for individual students, given the pre-registration data as it currently stand. For the future it would be great if we could link the responses to the pre-registration survey with the students included in Olin's Course History, so we could have the pre-registration survey results for each student.