In the NIPS 2012 conference, a workshop on "Personalizing education with machine learning" was organized by Michael C. Mozer, Javier R. Movellan, Robert V. Lindsey and Jacob Whitehill. The workshop consisted of twenty short oral presentations and was very well attended. The themes covered by the talks included, for instance, applying reinforcement learning models, stochastic optimal control theory, and Bayesian inference models on the diagnosis and decision making in educational contexts. In general, data and text mining techniques were used to model learning processes and to guide pedagogical processes. In the following, a small subset of the talks is described in some detail. Andrew Ng's talk "The Online Revolution: Education for Everyone" is discussed in another Cognitive Systems blog post.
Inferring Conceptual Knowledge From Unstructured Student Writing". Their approach was based on the idea that continuous, passive assessment can be used to elucidate conceptual knowledge. In other words, text mining was applied on students' writings to analyze their progress and to see if the text mining results can be used to predict course outcomes. This approach can be built on teachers' existing instruction and the wealth of unstructured data in an unintrusive manner. Ming and Ming showed that topic models of unstructured student writing can predict course outcomes. The work was motivated by the fact that in many cases the conceptual structure of the domain is not known well enough to facilitate detailed conceptual modeling. The study was based on texts written in online discussion forums during courses biology and economics lasting for five or six weeks. There were two or more mandatory discussion questions per week. It was found out that extra weeks of data improve the predictions. Ming and Ming also found out that using hierarchical topic modeling (based on hLDA) improves the results over traditional topic modeling (based on pLSA). In the methodological remarks, also cognitive components based on hidden-state conditional random fields were mentioned. Potential other text sources that can be used include online tutoring, informal learning environments, annotations on e-texts, and Wiki contributions.
Min Chi gave a talk on her work with Kurt VanLehn, Diane Litman and Pamela Jordan entitled "Empirically evaluating the application of reinforcement learning to the induction of effective and adaptive pedagogical tactics". The talk covered mathematics and science in which solving domain problems often consists of multiple domain principles and applying them in an appropriate way step-by-step. An instructor may choose to explain a step to the students ("tell") or ask the students to formulate the step themselves ("elicit"). A potential positive result of eliciting is the generation effect but also frustration may be caused if the task approves to be too difficult. Telling is accurate from the point of view of presentation but from the students' point of view it may lead into lack of attention and shallow processing. Chi presented a detailed account on how pedagogical tactics can be induced in the framework of Markov Decision Processes. The basic idea is to use reinforcement learning to determine what is the best action for the tutor to take in any learning context in order to maximize student learning. The actions in the model are elicit and tell, and the states are student performance and concept difficulty. The reward for the reinforcement learning consists of the student learning gains. In her approach, the transition probabilities are estimated from the training dataset. Min Chi listed also challenges like state representation not being clear and the state transitions being unknown. The experimental results were based on teaching students concepts in physics related to work and energy.
Vikram Ramanarayanan presented results on analyzing videos of educational situations. In his talk "A framework for unusual event detection in videos of informal class settings", he studied how engagement and disengagement could be recognized from videos in which 5 to 12 year old children are learning science. The results were promising even though formally the recognition measures were quite low. This was, however, to a large degree because the human annotators of the videos had often not labeled obvious cases of engagement or disengagement. Methodologically, linear dynamical systems modeling was used including a jump-Markov time-series model. The basic motivation of the work is to develop methods for real-time analysis of learning situations.