In this paper we investigate the usefulness of eye tracking data for predicting emotions relevant to learning, specifically boredom and curiosity. The data was collected during a study with MetaTutor, an intelligent tutoring system (ITS) designed to teach concepts about the human circulatory system while promoting effective use of self-regulated learning strategies. We use a variety of machine learning algorithms to predict students' self-reported emotions from gaze data features, and find that we can predict both emotions reliably using gaze. We examine the optimal amount of interaction time needed to make predictions, as well as which features are most predictive of each emotion, providing insight into how to detect when students disengage from the MetaTutor learning environment.
This paper presents a user study that investigates the factors affecting student attention to user-adaptive hints during interaction with an educational computer game. The study focuses on Prime Climb, an educational game designed to provide individualized support for learning number factorization skills in the form of textual hints based on a model of student learning. We use eye-tracking data to capture user attention patterns on the game adaptive-hints and present results on how user performance, hint timing, and attitude toward getting help all affect the degree to which students attend to hints. We also show that improved attention to hints is a worthwhile goal, at least in Prime Climb, because when they are attended to hints can positively affect a student's performance with the game.
Classification of high-dimensional data presents many difficulties; classifiers tend to overfit the data, and many classification algorithms do not scale well as the number of features increases. This paper compares the results of common feature reduction techniques such as PCA against an efficient implementation of the Johnson-Lindenstrauss (JL) embedding known as the Fast JL Transform. We perform feature reduction on two types of high-dimensional data, bag-of-words representations of text data, and eye tracking data. We show that FJLT, a non-data-aware random projection, can offer performance that is highly competitive with existing feature reduction methods, and enable classification of high-dimensional data using computationally complex algorithms.
Random Forests (RF) and Dropout networks are currently two of the most effective machine learning algorithms available. However, so far a study directly comparing the accuracy of both on the same dataset has not been performed. We hope to fill this gap by testing the classification accuracy of both of these ensemble methods on a novel dataset of American Sign Language (ASL) hand signs collected using the Microsoft Kinect. Results show that dropout nets achieve a higher gesture classification accuracy, particularly as the number of classification labels increases. Further, a neural network trained with dropout outperforms the same net without dropout, demonstrating the effectiveness of the technique. Individual gesture recognition accuracy as well as computation times for both algorithms will be presented.
An emerging field in user-adaptive systems is affect adaptivity: modeling and responding to an estimation of the user's emotional state. This could be particularly useful in the context of Intelligent Tutoring Systems (ITS), where learning gains are sensitive to the user's affect. Previous research used an empirically validated Dynamic Bayesian Network (DBN) to create an ITS affect model, but could not plan affect-sensitive responses. This paper will extend this research by converting the model into a Partially Observable Markov Decision Process (POMDP) representation, in order to compute a plan of interventions for the ITS agent to take given an estimation of the user's mood and goals. Two different methods for solving the POMDP will be compared: Incremental Pruning (an exact method) and SARSOP (an approximate method). Factors affecting the tractability of POMDP representations will also be discussed.
We created a real-time, collaborative voting system that allows students in a university course to indicate if they do not understand the material, or if the lecture is moving too slowly. When enough students agree, an Arduino X-Bee sends a haptic notification to a device worn on the professor's wrist.
We developed an online system to track the progress through the courses needed for a degree. Clicking a requirement brings up a filtered list of suitable courses. Complex dependencies between prerequisites are captured in the database.
On my own initiative, I built a website to help track student attendance from handwritten lists obtained during tutorials. Since deciphering handwriting is difficult, the user is only required to type a fragment of the name before the website generates a list of matching students from the course. The site was subsequently adopted and hosted by the University of Regina.