Using Data-Driven Discovery Techniques for the Design and Improvement of Educational Systems John Stamper Pittsburgh Science of Learning Center Human-Computer Interaction Institute Carnegie Mellon University 4/8/2013 The Classroom of the Future Which picture represents the “Classroom of the Future”? 2 The Classroom of the Future The answer is both! Depends of how much money you have... … but maybe not what you think… 3 The Classroom of the Future Rich vs. Poor – Poor kids will be forced to rely on “cheap” technology – Rich kids will have access to “expensive” teachers We are seeing this today! – Waldorf school in Silicon Valley – no technology – NGLC Wave III Grants – MOOCs – Growth of adaptive technology companies – Online instruction – … and more… 4 What does this mean? My view is that we cannot stop this, I believe we must accept that economics will force this route. We should focus on improving learning technology • New ways to improve teacher-student access • Add more adaptive features to learning software Adaptive learning, at scale, using data! 5 Educational Data Mining • “Educational Data Mining is an emerging discipline, concerned with developing methods for exploring the unique types of data that come from educational settings, and using those methods to better understand students, and the settings which they learn in.” – www.educationaldatamining.org 6 Types of EDM methods (Baker & Yacef, 2009) • Prediction – Classification – Regression – Density estimation • Clustering • Relationship mining – – – – Association rule mining Correlation mining Sequential pattern mining Causal data mining • Distillation of data for human judgment • Discovery with models 7 Emerging Communities • Society for Learning Analytics Research – First conference: LAK2011 • International Educational Data Mining Society – First conference: EDM2008 – Publishing JEDM since 2009 • Plus an emerging number of great people working in this area who are (not yet) closely affiliated with either community Emerging Communities • Joint goal of exploring the “big data” now available on learners and learning • To promote – New scientific discoveries & to advance learning sciences – Better assessment of learners along multiple dimensions • Social, cognitive, emotional, meta-cognitive, etc. • Individual, group, institutional, etc. – Better real-time support for learners EDM Methods to discuss • Prediction – understand what the student knows • Discovery with models – improve understanding of the structure of knowledge 10 LearnLab Pittsburgh Science of Learning Center (PSLC) • Created to bridge the Chasm between science & practice – Low success rate (<10%) of randomized field trials • LearnLab = a socio-technical bridge between lab psychology & schools – E-science of learning & education – Social processes for research-practice engagement • Purpose: Leverage cognitive theory and computational modeling to identify the conditions that cause robust student learning 11 LearnLab: Data-driven improvement infrastructure Ed tech + wide use = Research in practice Algebra Cognitive Tutor + = Chemistry Virtual Lab English Grammar Tutor Educational Games • 2004-14, ~$50 million • Tech enhanced courses, assessment, & research • School cooperation • In vivo experiments Interaction data is surprisingly revealing Online interactions => state tests • Accurate assessment during learning • Detect student work ethic, engagement … R = .82 Learning Curve Analysis • Discover better models of what is hard to learn Flat curve => improvement opportunity DataShop • Central Repository – Secure place to store & access research data – Supports various kinds of research • Primary analysis of study data • Exploratory analysis of course data • Secondary analysis of any data set • Analysis & Reporting Tools – Focus on student-tutor interaction data – Data Export • Tab delimited tables you can open with your favorite spreadsheet program or statistical package • Web services for direct access 14 14 Repository • • • • • Allows for full data management Controlled access for collaboration File attachments Paper attachments Great for secondary analyses How big is DataShop? 15 How big is DataShop? Domain Files Language Papers Datasets Student Actions Students Student Hours 64 11 78 6,237,523 6,499 6,877 222 53 189 75,754,530 37,218 173,175 Science 92 19 93 13,849,756 16,939 45,465 Other 18 12 50 8,604,016 13,018 31,111 396 95 410 104,445,825 73,674 256,630 Math Total As of April 2013 16 What kinds of data? • By domain based on studies from the Learn Labs • Data from intelligent tutors • Data from online instruction • Data from games The data is fine grained at a transaction level! 17 Web Application Getting to DataShop • Explore data through the DataShop tools • Where is DataShop? – http://pslcdatashop.org – Linked from DataShop homepage and learnlab.org • http://pslcdatashop.web.cmu.edu/about/ • http://learnlab.org/technologies/datashop/index.php 19 19 DataShop Terminology • KC: Knowledge component – also known as a skill/concept/fact – a piece of information that can be used to accomplish tasks – tagged at the step level • KC Model: – also known as a cognitive model or skill model – a mapping between problem steps and knowledge components 20 Getting the KC Model Right! The KC model drives instruction in adaptive learning – Problem and topic sequence – Instructional messages – Tracking student knowledge 21 What makes a good KC Model? • A correct expert model is one that is consistent with student behavior. • Predicts task difficulty • Predicts transfer between instruction and test The model should fit the data! 22 Good KC Model => Good Learning Curve • An empirical basis for determining when a cognitive model is good • Accurate predictions of student task performance & learning transfer – Repeated practice on tasks involving the same skill should reduce the error rate on those tasks => A declining learning curve should emerge 23 A Good Learning Curve 24 How do we make KC Models? 25 Traditionally CTA has been used But Cognitive Task Analysis has some issues… – Extremely human driven – It is highly subjective – Leading to differing results from different analysts And these human discovered models are usually wrong! 26 If Human centered CTA is not the answer How should these models be designed? They shouldn’t! The models should be discovered not designed! 27 Solution – We have lots of log data from tutors and other systems – We can harness this data to validate and improve existing student models 28 Human-Machine Student Model Discovery DataShop provides easy interface to add and modify KC models and ranks the models using AFM 29 29 Human-Machine Student Model Discovery 3 strategies for discovering improvements to the student model – Smooth learning curves – No apparent learning – Problems with unexpected error rates 30 A good cognitive model produces a learning curve Without decomposition, using just a single “Geometry” skill, no smooth learning curve. But with decomposition, 12 skills for area, a smooth learning curve. Is this the correct or “best” cognitive model? (Rise in error rate because poorer students get assigned more problems) Inspect curves for individual knowledge components (KCs) Many curves show a reasonable decline Some do not => Opportunity to improve model! 32 No apparent Learning 33 Problems with Unexpected Error Rates 34 Inspect problems to hypothesize new KC labels • Here scaffolding is originally absent, but other problems have fixed scaffolding – They start with columns for square & area These strategies suggest an improvement – Hypothesized there were additional skills involved in some of the compose by addition problems – A new student model (better BIC value) suggests the splitting the skill. 36 Redesign based on Discovered Model Our discovery suggested changes needed to be made to the tutor – Resequencing – put problems requiring fewer skills first – Knowledge Tracing – adding new skills – Creating new tasks – new problems – Changing instructional messages, feedback or hints 37 Study : Current tutor is control • Current fielded tutor only uses scaffolded problems Study: Treatment • Scaffolded, given areas, plan-only, & unscaffolded • Isolate practice on problem decomposition Study Results • Much more efficient & better learning on targeted decomposition skills Instructional time (minutes) by step type 30 20 Post-test % correct by item type 1 Area and other steps 0.95 Composition steps 0.9 0.85 10 0.8 0.75 0 Control: Original tutor Treatment: Modelbased redesign Composition Area 0.7 Control: Original Treatment: Modeltutor based redesign Translational Research Feedback Loop Design Discover Deploy Data Can a data-driven process be automated & brought to scale? Yes! • Combine Cognitive Science, Psychometrics, Machine Learning … • Collect a rich body of data • Develop new model discovery algorithms, visualizations, & on-line collaboration support 42 DataShop’s “leaderboard” ranks discovered cognitive models 100s of datasets coming from ed tech in math, science, & language Some models are machine generated (based on human-generated learning factors) Some models are human generated 43 Metrics for model prediction • AIC & BIC penalize for more parameters, fast & consistent • 10 fold cross validation • Minimize root mean squared error (RMSE) on unseen data 44 Automated search for better models Learning Factors Analysis (LFA) (Cen, Koedinger, & Junker, 2006) • Method for discovering & evaluating cognitive models • Finds model “Q matrix” that best predicts student learning data • Inputs Data: Student success on tasks over time Factors hypothesized to explain learning • Outputs Rank order of most predictive Q matrix Parameter estimates for each Simple search process example: modifying Q matrix by input factor to get new Q’ matrix • Q matrix factor Sub split by factor Neg-result • • Produces new Q matrix Two new KCs (Sub-Pos & Sub-Neg) replace old KC (Sub) • Redo opportunity counts LFA: Best First Search Process • Original Model BIC = 4328 Split by Embed 4301 4320 4322 Split by Backward 4322 4313 • Split by Initial 4312 4322 4325 50+ Search algorithm guided by a heuristic: AIC Start with single skill cog model (Q matrix) 4320 4324 15 expansions later 4248 Cen, H., Koedinger, K., Junker, B. (2006). Learning Factors Analysis: A general method for cognitive model evaluation and improvement. 8th International Conference on Intelligent Tutoring Systems. Scientist “crowd”sourcing: Feature input comes “for free” Union of all hypothesized KCs in human generated models Scientist generated models 48 Validating Learning Factors Analysis • Discovers better cognitive models in 11 of 11 datasets … Koedinger, McLaughlin, & Stamper (2012). Automated student model improvement. In Proceedings of the Fifth International Conference on Educational Data Mining. [Conference best paper.] Data from a variety of educational technologies & domains Statistics Online Course English Article Tutor Algebra Cognitive Tutor Numberline Game 50 Applying LFA across domains Variety of domains & technologies 11 of 11 improved models 9 of 11 equal or greater learning Can we go even bigger? 52 Competitions? KDD Cup Competition Knowledge Discovery and Data Mining (KDD) is the most prestigious conference in the data mining and machine learning fields KDD Cup is the premier data mining challenge 2010 KDD Cup called “Educational Data Mining Challenge” Ran from April 2010 through June 2010 54 KDD Cup Competition Competition goal is to predict student responses given tutor data provided by Carnegie Learning Dataset Students Steps File size Algebra I 2008-2009 3,310 9,426,966 3 GB Bridge to Algebra 2008-2009 6,043 20,768,884 5.43 GB 55 KDD Cup Competition 655 registered participants 130 participants who submitted predictions 3,400 submissions KDD Cup Competition Advances in prediction, cognitive modeling, new methods applied to EDM Spawned a number of workshops and papers The datasets are now in the “wild” and showing up in non KDD conferences New competitions to continue momentum 57 Marigames.org • Two stage competition with $100,000 in prizes – $50,000 Game Development – $50,000 Educational Data Mining • Goal is to go beyond individual datasets • This requires common data formats 58 Take aways • The amount of data coming from educational technology is growing exponentially • Huge potential for EDM to improve educational systems • Optimal instructional design requires discoveries (The student is not like me) • These methods require common forms of data for analysis (standards!) 59 Opportunities • New Learning Science and Engineering professional masters degree at Carnegie Mellon University • New concentration in Learning Analytics, MA in Cognitive Studies in Education at Teachers College, Columbia University • Other programs in the works 60 Thank you Special Thanks to: Ken Koedinger, Director LearnLab Ryan Baker, President IEDMS Steve Ritter, Carnegie Learning 61 http://pslcdatashop.org Questions? john@stamper.org http://dev.stamper.org 62