Zachary A. Pardos Worcester Polytechnic Institute United States
Neil Heffernan Worcester Polytechnic Institute
Provide a URL to a web page, technical memorandum, or a paper.
http://users.wpi.edu/~zpardos/
Provide a general summary with relevant background information: Where does the method come from? Is it novel? Name the prior art.
The HMM used is a novel Bayesian HMM developed by Pardos & Heffernan [1,2] that predicts the probability of knowledge for each student at each opportunity as well as a prediction of probability of correctness on each step. The model learns individualized student specific parameters (prior, learn rate, guess and slip) and then uses these parameters to train skill specific models. The resulting model that considers the composition of user and skill parameters outperforms models that only take into account parameters of the skill. Models that only consider skill parameters are the predominant methods for prediction in the intelligent tutoring systems field. The Bayesian model was used in ensemble selection [3] and also to generate extra features for the decision tree classifier. The bagged decision tree classifier [4,5] was the primary classifier used and was developed by Leo Breiman and Adele Cutler.
Summarize the algorithms you used in a way that those skilled in the art should understand what to do. Profile of your methods as follows:
Please describe your data understanding efforts, and interesting observations:
There was a large percentage (>20%) of steps that had no skill tagging in any KC model. These steps were all assigned a separate skill id. When running cross-fold validation on the training set, rows with this null skill id were predicted with an RMSE around 0.20. Another obstacle to a skill centric analysis was the lack of prior responses to skills for some users. Around 5-8% of the test rows involved making a prediction for a user on step tagged with a skill that the user had never encountered before. For rows of this type, the RMSE was dramatically higher; an RMSE of ~0.271 for rows with prior skill responses and ~0.323 for rows without prior skill responses.
Details on feature generation:
Several features sets of the user and skill were created as well as features of the step, problem, section and unit. Inferences on student skill knowledge and predicted probability of correct from the top user model were also used as features. Several test sets were created and features were generated for the rows of those test sets from their respective training sets. In addition to percent correct features, features indicative of users gaming the system [6] were also generated. These features include the standard deviation of the user's step duration for a given skill. I believe the creation of internal test and training sets was a key decision in being able to explore prediction methods with a manageable dataset size. This also allowed for feedback without using the leaderboard. The internal fit measures were often very close to the leaderboard submission result.
Details on feature selection:
No response.
Details on latent factor discovery (techniques used, useful student/step features, how were the factors used, etc.):
More details on preprocessing:
Steps often had multiple KCs associated with them. In order to convert the multi tagging to single tagging for easier processing, two alternative KC taggings were created for each of the KC models. One alternate chose the KC with the lowest percent correct. The second alternate assigned a new KC id to unique sets of KCs.
Details on classification:
The Bayesian models and bagged decision trees were run on a number of dataset and feature set variations. Some features required previous skill information for the user being predicted. If this information did not exist, this feature was not created for the particular test row. This resulted in feature sets that did not cover the entire test set. Because of this predictions were made that covered different portions of the test set. Ensemble selection was used to blend the various model predictions. Because of the varying number of predictions between models, a special ensemble initialization technique was created whereby the best model was chosen first and subsequent models were chosen based on the RMSE of the predicted rows excluding the rows already added to the initialized ensemble. This allowed for models to be used for the portions of the test set in which they excelled. For instance, the rows of the test set containing skills not yet seen by the user were best predicted by a model that was not a top predicting model overall.
Details on model selection:
Scores shown in the table below are Cup scores, not leaderboard scores. The difference between the two is described on the Evaluation page.
A reader should also know from reading the fact sheet what the strength of the method is.
Please comment about the following:
The training of the naive decision tree classifiers of the random forest method could be parallelized. This resulted in a lower training time than the neural networks approach that was tried. Random forests were preferable also because of their small parameter space; minleaf and number of trees were the only significant parameters and these parameters take non continuous values with reasonable ranges. Because of this, a large portion of the parameter space can be reasonably explored.
Neural networks with 1-3 hidden layers were tried but with predictive performance far bellow that of bagged decision trees. SVMs were also tried with both linear and non-linear kernels. The linear kernel SVM parameters were explored using a coarse grid search and then a higher resolution search around the areas of low RMSE found in the first search. This approach resulted in prediction accuracies comparable to the neural network predictions.
Details on the relevance of the KC models and latent factors:
Details on software implementation:
Perl was used to hash the dataset features into numeric values. The parallel computing toolbox from MathWorks was used with MATLAB.
Details on hardware implementation. Specify whether you provide a self contained-application or libraries.
A 30 node rocks cluster was used to train the ~1,500 Bayesian skill models for each dataset. One 16 core and one 8 core machine was used to run the bagged decision tree classification.
Provide a URL for the code (if available):
Combining user features with skill features was very powerful in both modeling and classification approaches. Prediction error was very low for rows that had sufficient data to compile a complete user and skill feature set however error was very high for rows were the user did not have sufficient skill data. In order to increase prediction accuracy for these rows, imputing missing features could be very beneficial. What to do with these rows is a worthy area of future study since, while small, they significantly impacted overall RMSE.
List references below.
[1] Pardos, Z. A., Heffernan, N. T. In Press (2010) Modeling Individualization in a Bayesian Networks Implementation of Knowledge Tracing. In Proceedings of the 18th International Conference on User Modeling, Adaptation and Personalization. Hawaii. [2] Pardos, Z. A., Heffernan, N. T. In Press (2010) Navigating the parameter space of Bayesian Knowledge Tracing models: Visualizations of the convergence of the Expectation Maximization algorithm. In Proceedings of the 3rd International Conference on Educational Data Mining. Pittsburg. [3] R. Caruana, A. Niculescu-Mizil, G. Crew, and A. Ksikes. Ensemble selection from libraries of models. In Proceedings of the International Conference on Machine Learning (ICML), 2004. [4] Breiman, L., Friedman, J., Olshen, R. & Stone, C. (1984), Classification and Regression Trees, Wadsworth International. [5] Breiman, L. Random forests. Machine Learning, 45(1): 5?32, 2001. [6] Baker, R.S.J.d., Corbett, A.T., Roll, I., Koedinger, K.R. (2008) Developing a Generalizable Detector of When Students Game the System. User Modeling and User-Adapted Interaction, 18, 3, 287-314.