Show simple item record

dc.contributor.authorFarasat, Alireza
dc.date.accessioned2018-05-23T20:18:17Z
dc.date.available2018-05-23T20:18:17Z
dc.date.issued2017
dc.identifier.isbn9781369593358
dc.identifier.other1877972813
dc.identifier.urihttp://hdl.handle.net/10477/77284
dc.description.abstractThis dissertation introduces the socio-technical pedagogical paradigm called “crowdlearning'', aimed at STEM students in physical and virtual classrooms into collaborative problem-posing and problem-solving activities. This research also addresses the challenges of developing an Intelligent Tutoring Systems (ITS) such as scalability, modeling of student conceptual learning, personalizing optimal teaching policy by incorporating advances in machine learning and optimization techniques to explain how students learn and how an instructor can promote their conceptual understanding of the course contents. A probabilistic framework for formulating and predicting the dynamics of student learning using new tensor factorization models is proposed and aligned with a model predictive control approach to personalize the optimal teaching policy for each student. The first part of this study develops a web-based platform to engage STEM students and instructors in creative problem posing and solving activities, wherein the students experience deeper learning by gaining new perspectives on the use of subject-matter concepts and by learning with and from each other. The vision of Crowdlearning is that of a self-sustaining problem-posing and problem-solving environment, where the students of a given subject intermittently take on the roles as (A) the creators of subject-focused problems (problem statements/formulations with answer alternatives, hints, correct answers with explanations, etc.) and (B) problem solvers all on the online platform. The second part of this research introduces two new synthetic data generating models so that enables one to explore students conceptual learning in more depth. The proposed models benefit from the idea of Bayesian Knowledge Tracing and Latent Factor Analysis, the state-of-the-art models in modeling student learning and predicting their performance. The first model consider student learning as a discrete random variable and uses the idea of Factorial Hidden Markov Model to connect different knowledge components and observations overtime. The second model postulates that student learning dynamics is a continuous quantity captured by a first order difference equation. The student learning models developed throughout this dissertation are tested with this synthetic data. The third part improves current ITS state of the art model in terms of computational performance and proposes a Parallel Coordinate Descent algorithm in shared memory systems that significantly makes the recently proposed SPARse Factor Analysis (SPARFA) scalable. This new implementation enables student modeling analyses at scale, by presenting a parallel algorithm for performing SPARse Factor Analysis (SPARFA). SPARFA is accepted as a state-of-the-art machine learning approach for estimating student knowledge levels and predicting their performance in not-yet-taken educational tasks. However, with the Likelihood Maximization (LM) formulation behind it resulting in a non-convex optimization problem with many local minima, the scalability has been a challenge for SPARFA. While the original method employs an alternating optimization approach to approximately solve the SPARFA LM problem, this work employs a Parallel Coordinate Descent (PCD) algorithm parallelized in the shared memory setting. The performance of PCD is evaluated on varied problem instances with synthetic data and is found to successfully solve instance with up to 7.5 Million variables, which makes it suitable to make inference from the data of any real-world Massive Open Online Course. Moreover, this research enhances the current Latent Factor Analysis models and offers new methods for dealing with student performance prediction problem, framed as a Probabilistic Tensor Factorization problem. The proposed framework introduces new constrained tensor factorization models that enable one to discover meaningful patterns from highly sparse data. The key challenge of modeling and predicting students performance lies in the estimation of their conceptual understanding while recognizing both the temporal dynamics of knowledge acquisition and the differences in individual learning abilities. Responding to this challenge, new constrained tensor factorization models in probabilistic formulation are presented. The first model, called Homogeneous Student Model (HSM), is a probabilistic tensor factorization-based model that specifies student knowledge gain as a set of constraints in the model. The second model is based on a new tensor factorization model that parameterizes personalized student learning process and it is called Personalized Student Model (PSM). Finally, this study extends a new framework in student modeling that not only takes into account the effect of each question on student learning but it distinguishes also between the gains obtained form correctly and incorrectly answered questions (i.e. correctly answered questions lead to more knowledge gain). The proposed framework that is founded on top of constrained tensor factorization techniques, models the interconnection between different knowledge components as well. The proposed framework called Learning-By-Solving (LBS), is employed as a predictive model to optimize teaching policy by finding the optimal sequence of questions in a personalized way. Finding optimal teaching policy is still an open question in ITSs. (Abstract shortened by ProQuest.)
dc.languageEnglish
dc.sourceDissertations & Theses @ SUNY Buffalo,ProQuest Dissertations & Theses Global
dc.subjectApplied sciences
dc.subjectIntelligent tutoring systems
dc.subjectModel protective control
dc.subjectPersonalized learning systems
dc.subjectProbabilistic models
dc.subjectSparse tensor factorization
dc.titlePersonalization and Learning Curve Optimization in Intelligent Tutoring Systems
dc.typeDissertation/Thesis


Files in this item

Thumbnail

This item appears in the following Collection(s)

Show simple item record