![Data Collection - Methods
Uniform Random (UR): Randomly predict grades from a uniform distribution over the range [0, 4].
Global Mean (GM): Predict grades using the mean of all previously observed grades.
Mean of Means (MoM): Predict grades using an average of the global mean, the per student mean, and
the per-course mean.](https://image.slidesharecdn.com/article2-171107192725/85/Prediction-Next-Term-Student-Performance-Prediction-A-Recommender-Systems-Approach-8-320.jpg)
Prediction-Next-Term Student Performance Prediction: A Recommender Systems Approach
- 1. CEIT 720: Learning Anaytics In Education Predictive Methods Beste Ulus
- 2. Next-Term Student Performance Prediction: A Recommender Systems Approach Sweeney, M., Rangwala, H., Lester, J., Johri, A. (2016)
- 3. Context of the Study The purpose of the study presented in this paper was to apply state-of-the-art recommender systems techniques to the task of student performance prediction.
- 4. Research Questions In the present study, they compare more models on a larger dataset and conduct a more detailed analysis of feature importance, performance errors, and implications for educational applications.
- 5. Research Method of the Study Predictive Analytical Method
- 6. Data Collection - Data Students from a public university Instructor (classification, rank, tenure status) Student info (transfer or not) Courses Disciplines Letter grades Demographics data, such as age, race, sex, high school CEEB code and GPA, zip code, and 1600-scale SAT scores.
- 7. Data Collection - Methods Simple baselines MF-based methods Common regression models.
- 8. Data Collection - Methods Uniform Random (UR): Randomly predict grades from a uniform distribution over the range [0, 4]. Global Mean (GM): Predict grades using the mean of all previously observed grades. Mean of Means (MoM): Predict grades using an average of the global mean, the per student mean, and the per-course mean.
- 9. Data Collection - Methods Three methods based on Matrix Factorization (MF): Singular Value Decomposition (SVD) SVD-kNN: SVD post-processed with kNN Factorization Machine (FM)
- 10. Data Collection - Methods Four different regression models: Random Forest (RF) Stochastic Gradient Descent (SGD) Regression k-Nearest Neighbors (kNN) Personalized Linear Multiple Regression (PLMR)
- 11. Data Collection - Methods Evaluations are performed in terms of two common regression metrics: Root Mean Squared Error (RMSE) Mean Absolute Error (MAE).
- 12. Data Collection - Tools No information on which tool they utilized to analyze data.
- 13. Results The FM model outperforms all the others by a wide margin, indicating 2-way feature interactions play an important role in predicting performance. FM-RF hybrid is an effective method of overcoming the cold-start limitations of FMs and is the most effective method in general. When the test distribution differs from the training distribution, the FM model is liable to learn overconfident 2-way interactions that reduce its ability to generalize.
- 14. Results
- 15. Results
- 16. Implications For students, we can incorporate this information into a degree planning system. For educators, knowledge of which students have the lowest expected grades could provide opportunities to increase detection of at-risk students. For advisors, any additional information that helps them personalize their advice to each student could potentially help thousands of students.
- 17. Thank You
Editor's Notes
- #4: Is it a part of project? If yes name of the project? Is the study descriptive, diagnostic, predictive or prescriptive ?
- #7: Cold-start dyads have either a new student, a new course, or both. Appears in the prediction phase but not in previous terms training phase
- #9: 1 randomly guessing, 2, overall central tendency, 3, row and column averages Coldstart hiç biri yoksa GM herhangi biri varsa MoM
- #10: each grade is simply predicted as the dot product of the latent student and course feature vectors. We call this the factorized 2-way interaction of the student with the course. k-nearest neighbors (kNN) yields improved predictive performance. Pure collaborative filtering (CF) methods such as SVD and SVD-kNN are unable to make predictions for cold-start records. the FM model is able to capture all of the information captured by the simple baselines as well as the information captured by SVD. In general, it captures the global central tendency, 1-way (linear) relationships between the predictors and the grade (bias terms), and 2-way factorized interactions between each predictor and the grade:
- #11: Once built, the tree can be used for regression of new data samples. The Random Forest then combines many of these trees in a weighted averaging approach to make decisions regarding unseen data. SGD is a gradient-based optimization technique that updates the model parameters incrementally, rather than on the entire training set at once (which is what normal gradient descent would do). This reduces overfitting and significantly improves training time. The k-Nearest Neighbors (kNN) algorithm is a classic method for clustering samples based on similarity.
- #12: RMSE is the metric we use to compare methods. MAE allows us to understand the range of grades we might actually be predicting.