Latent Skill Embedding for Personalized Lesson Sequence Recommendation
Abstract
Students in online courses generate large amounts of data that can be used to personalize the learning process and improve quality of education. In this paper, we present the Latent Skill Embedding (LSE), a probabilistic model of students and educational content that can be used to recommend personalized sequences of lessons with the goal of helping students prepare for specific assessments. Akin to collaborative filtering for recommender systems, the algorithm does not require students or content to be described by features, but it learns a representation using access traces. We formulate this problem as a regularized maximumlikelihood embedding of students, lessons, and assessments from historical studentcontent interactions. An empirical evaluation on largescale data from Knewton, an adaptive learning technology company, shows that this approach predicts assessment results competitively with benchmark models and is able to discriminate between lesson sequences that lead to mastery and failure.
Latent Skill Embedding for Personalized Lesson Sequence Recommendation
Siddharth Reddy 
Department of Computer Science 
Cornell University 
Ithaca, NY 
sgr45@cornell.edu 
Igor Labutov 
Department of Electrical and Computer Engineering 
Cornell University 
Ithaca, NY 
iil4@cornell.edu 
Thorsten Joachims 
Department of Computer Science 
Cornell University 
Ithaca, NY 
tj@cs.cornell.edu 
\@float
copyrightbox[b]
\end@float
•Mathematics of computing Probabilistic representations; •Computing methodologies Learning in probabilistic graphical models; •Applied computing Computerassisted instruction;

Probabilistic Embedding; Sequence Recommendation; Adaptive Learning
The popularity of online education platforms has soared in recent years. Companies like Coursera and EdX offer Massive Open Online Courses (MOOCs) that attract millions of students and highcalibre instructors. Khan Academy has become a hugely popular repository of videos and interactive materials on a wide range of subjects. Elearning products offered by universities and textbook publishers are also gaining traction. These platforms improve access to high quality educational content for anyone connected to the Internet. As a result, people who would otherwise lack the opportunity are able to consume materials like video lectures and problem sets from courses offered at top universities. However, in these online environments learners often lack the personalized instruction and coaching that can potentially lead to significant improvements in educational outcomes. Furthermore, the educational content may be contributed by many authors without a formal underlying structure. Intelligent systems that learn about the educational properties of the content, guide learners through custom lesson plans, and quickly adapt through feedback could help learners take advantage of large and heterogeneous collections of educational content to achieve their goals.
The extensive literature on intelligent tutoring systems (ITS) and computerassisted instruction (CAI) dates back to the 1960s. Early efforts focused on approximating the behavior of a human tutor through rulebased systems that taught students South American geography [?], electronics troubleshooting [?], and programming in Lisp [?]. Today’s online education platforms differ from early ITSes in their ability to gather data at scale, which facilitates the use of machine learning techniques to improve the educational experience. Relatively little academic work has been done to design systems that use the massive amounts of data generated by students in online courses to provide personalized learning tools. Learning and content analytics [?], instructional scaffolding in educational games [?], hint generation [?], and feedback propagation [?] are a few topics currently being explored in the personalized learning space.
Our aim is to build a domainagnostic framework for modeling students and content that can be used in many online learning products for personalized lesson sequence recommendation. A common data source available in products is a stream of interaction data, or access traces that log student interactions with modules of course content. These access traces have the form Student A completed Lesson B and Student C passed assessment D. Lessons are content modules that introduce or reinforce concepts; for example, an animation of cellular respiration or a paragraph of text on Newton’s first law of motion. Assessments are content modules with passfail results that test student skills; for example, a trueorfalse question halfway through a video lecture. By relying on a coarsegrained, binary assessment result, we are able to gracefully handle many types of assessments (e.g., free response and multiple choice) as long as a student response can be labelled as correct or incorrect.
We use access traces to embed students, lessons, and assessments together in a joint semantic space, yielding a representation that can be used to reason about the relationship between students and content (e.g., the likelihood of passing an assessment, or the skill gains achieved by completing a lesson). The model is evaluated on simple synthetic scenarios, as well as largescale real data from Knewton, an education technology company that offers personalized recommendations and activity analytics for online courses [?]. The data set consists of 2.18 million access traces from over 7,000 students, recorded in 1,939 classrooms over a combined period of 5 months.
Our work builds on the existing literature in psychometric user modeling. The Rasch model estimates the probability of a student passing an assessment using latent concept proficiency and assessment difficulty parameters [?]. The twoparameter logistic item response theory (2PL IRT) model adds an assessment discriminability parameter to the result likelihood [?]. Both models assume that a map from assessments to a small number of underlying concepts is known a priori. We propose a datadriven method of learning content representation that does not require a priori knowledge of contenttoconcept mapping. Though this approach sacrifices the interpretability of expert ratings, it has two advantages: 1) it does not require laborintensive expert annotation of content and 2) it can evolve the representation over time as existing content is modified or new content is introduced.
Lan et al. propose a sparse factor analysis (SPARFA) approach to modeling graded learner responses that uses assessmentconcept associations, concept proficiencies, and assessment difficulty [?]. The algorithm does not rely on an expert concept map, but instead learns assessmentconcept associations from the data. Multidimensional item response theory [?] also learns these assocations from the data. We extend the ideas behind SPARFA and multidimensional item response theory to include a model of student learning from lesson modules, which is a key prerequisite for recommending personalized lesson sequences.
Bayesian Knowledge Tracing (BKT) uses a Hidden Markov Model to model the evolution of student knowledge (which is discretized into a finite number of states) over time [?]. Further work has modified the BKT framework to include the effects of lessons through an inputoutput Hidden Markov Model [?, ?, ?]. Similarly, SPARFA has been extended to model timevarying student knowledge and the effects of lesson modules [?]. Item response theory has also been extended to capture temporal changes in student knowledge [?, ?]. Recurrent neural networks have been used to trace student knowledge over time and model lesson effects [?]. Similar ideas for estimating temporal student knowledge from binaryvalued responses have appeared in the cognitive modeling literature [?, ?]. We extend this work in a multidimensional setting where student knowledge lies in a continuous state space and lesson prerequisites modulate knowledge gains from lesson modules.
Our model also builds on previous work that uses temporal embeddings to predict music playlists [?]. While Moore et al. focused on embedding objects (songs) in a metric space, we propose a nonmetric embedding where the distances between objects (students, assessments, and lessons) are not symmetric, capturing the natural progression in difficulty of assessments and the positive growth of student knowledge.
We now describe the Latent Skill Embedding, a probabilistic model that places students, lessons, and assessments in a joint semantic space that we call the latent skill space. Students have trajectories through the latent skill space, while assessments and lessons are placed at fixed locations. Formally, a student is represented as a set of latent skill levels ; a lesson module is represented as a vector of skill gains and a set of prerequisite skill requirements ; an assessment module is represented as a set of skill requirements .
Students interact with lessons and assessments in the following way. First, a student can be tested on an assessment module with a passfail result , where the likelihood of passing is high when a student has skill levels that exceed the assessment requirements and viceversa. Second, a student can work on lesson modules to improve skill levels over time. To fully realize the skill gains associated with completing a lesson module, a student must satisfy prerequisites (only partly fulfilling the prerequisites will result in relatively smaller gains, see Equation Latent Skill Embedding for Personalized Lesson Sequence Recommendation for details). Time is discretized such that at every timestep , a student completes a lesson and may complete zero or many assessments. The evolution of student knowledge can be formalized as the graphical model in Figure Latent Skill Embedding for Personalized Lesson Sequence Recommendation, and the following subsections elaborate on the details of this model.
For student , assessment , and result ,
(1) where is the logistic function and . and are constrained to be nonnegative (for details see the Parameter Estimation section). A pass result is indicated by , and a fail by . The term can be rewritten as , where is the angle between and ; it can be interpreted as “relevant skill”. The term can be interpreted as general (i.e. not conceptspecific) assessment difficulty. The expression is visualized in Figure Latent Skill Embedding for Personalized Lesson Sequence Recommendation. The bias term is a studentspecific term that captures a student’s general (assessmentinvariant and timeinvariant) ability to pass; it can be interpreted as a measure of how well the student guesses correct answers. The bias term is a modulespecific term that captures an assessment’s general (studentinvariant and timeinvariant) difficulty. differs from the difficulty term in that it is not bounded; see the Parameter Estimation section for details. These bias terms are analogous to the bias terms used for modeling song popularity in [?]. Our choice of differs from traditional multidimensional item response theory, which uses where and are not bounded (although in practice, suitable priors are imposed on these parameters).
For student who worked on a lesson with skill gains and no prerequisites at time , the updated student state is
(2) where the covariance matrix is diagonal. For a lesson with prerequisites ,
(3) where is the logistic function and . The intuition behind this equation is that the skill gain from a lesson should be weighted according to how well a student satisfies the lesson prerequisites. A student can compensate for lack of prerequisites in one skill through excess strength in another skill, but the extent to which this tradeoff is possible depends on the relative weights of the prerequisites. The same principle applies to satisfying assessment skill requirements. With prerequisites, the vector field of skill gains is nonuniform (without prerequisites, it is uniform); for example, see Figure Latent Skill Embedding for Personalized Lesson Sequence Recommendation.
Our model differs from [?] in that we explicitly model the effects of prerequisite knowledge on gains from lessons. Lan et al. model gains from a lesson as an affine transformation of the student’s knowledge state.
We compute MAP estimates of model parameters by maximizing the following objective function:
(4) where is the set of assessment interactions, is the set of lesson interactions, is a regularization term that penalizes the norms of embedding parameters (not bias terms), and is a regularization parameter. Nonnegativity constraints on embedding parameters (not bias terms) are enforced.
regularization is used to penalize the size of embedding parameters to prevent overfitting. The bias terms are not bounded or regularized. This allows to be positive for assessment modules that are especially easy, and to be negative for students who fail especially often. We solve the optimization problem with box constraints using the LBFGSB [?] algorithm. We randomly initialize parameters and run the iterative optimization until the relative difference between consecutive objective function evaluations is less than . Averaging validation accuracy over multiple runs during crossvalidation reduces sensitivity to the random initializations (since the objective function is nonconvex).
To verify the correctness of our model and to illustrate the properties of the embedding geometry that the model captures, we conducted a series of experiments on small, syntheticallygenerated interaction histories. Each scenario is intended to demonstrate a different feature of the model (e.g., recovering student knowledge and assessment requirements in the absence of lessons, or recovering sensible skill gain vectors for different lessons). For the sake of simplicity, the embeddings do not use bias terms. The scenarios shown next are annotated versions of plots made by our embedding software.
Figure Latent Skill Embedding for Personalized Lesson Sequence Recommendation demonstrates an extremely simple embedding. The key observation here is that the model recovered positive skill gains for lesson L1, and “correctly” arranged Alice and assessment A1 in the latent space. Initially, Alice fails A1, so her skill level is behind the requirements of A1. After completing L1, Alice passes A1, indicating that her skill level has probably improved past the requirements of A1. Note that this scenario could have been explained with only one latent skill.
Figure Latent Skill Embedding for Personalized Lesson Sequence Recommendation depicts a twodimensional embedding, where an intransitivity in assessment results requires more than one latent skill to explain. The key observation here is that the assessments are embedded on two different axes, meaning they require two completely independent skills. This makes sense, since student results on A1 are uncorrelated with results on A2. Fogell fails both assessments, so his skill levels are behind the requirements for A1 and A2. McLovin passes both assessments, so his skill levels are beyond the requirements for A1 and A2. Evan and Seth are each able to pass one assessment but not the other. Since the assessments have independent requirements, this implies that Evan and Seth have independent skill sets (i.e. Evan has enough of skill 2 to pass A2 but not enough of skill 1 to pass A1, and Seth has enough of skill 1 to pass A1 but not enough of skill 2 to pass A2).
In Figure Latent Skill Embedding for Personalized Lesson Sequence Recommendation, we replicate the setting in Figure Latent Skill Embedding for Personalized Lesson Sequence Recommendation, then add two new students Slater and Michaels, and two new lesson modules L1 and L2. Slater is initially identical to Evan, while Michaels is initially identical to Seth. Slater reads lesson L1, then passes assessments A1 and A2. Michaels reads lesson L2, then passes assessments A1 and A2. The key observation here is that the skill gain vectors recovered for the two lesson modules are orthogonal, meaning they help students satisfy completely independent skill requirements. This makes sense, since initially Slater was lacking in Skill 1 while Michaels was lacking in Skill 2, but after completing their lessons they passed their assessments, showing that they gained from their respective lessons what they were lacking initially.
In Figure Latent Skill Embedding for Personalized Lesson Sequence Recommendation, we replicate the setting in Figure Latent Skill Embedding for Personalized Lesson Sequence Recommendation, then add a new assessment module A3 and a new lesson module L1. All students initially fail assessment A3, then read lesson L1, after which McLovin passes A3 while everyone else still fails A3. The key observation here is that McLovin is the only student who initially satisfies the prerequisites for L1, so he is the only student who realizes significant gains from taking L1.
We use data processed by Knewton, an adaptive learning technology company. Knewton’s infrastructure uses studentcontent access traces to generate personalized recommendations and activity analytics for partner organizations with online learning products. The data describes interactions between college students and two science textbooks. The Book A data set was collected from 869 classrooms from January 1, 2014 through June 1, 2014. It contains 834,811 interactions, 3,471 students, 3,374 lessons, 3,480 assessments, and an average assessment pass rate of 0.712. The paths that students take are biased by direction from instructors, a recommender system, and the sequence of chapters in the textbook. The Book B data set was collected from 1,070 classrooms from January 1, 2014 through June 1, 2014. It contains 1,349,541 interactions, 3,563 students, 3,843 lessons, 3,807 assessments, and an average assessment pass rate of 0.693.
Both data sets are filtered to eliminate students with fewer than five lesson interactions and content modules with fewer than five student interactions. To avoid spam interactions and focus on the outcomes of initial student attempts, we only consider the first interaction between a student and an assessment (subsequent interactions between student and assessment are ignored).
Model Book A Book B Test Validation Test Validation 1 N N N 0.673 0.614 2 N N Y 0.818 0.788 3 Y N N 0.692 0.630 4 Y N Y 0.798 0.775 5 Y Y N 0.724 0.629 6 Y Y Y 0.811 0.785 7 1PL IRT 0.812 0.778 8 2PL IRT 0.780 0.686 9 2D MIRT 0.817 0.776 Table \thetable: Test AUC, validation AUC, and standard error of validation AUC for variations of the embedding model and benchmark IRT models. We evaluate the embedding model on the task of predicting results of heldout assessment interactions, and compare it to three benchmark models: the one and twoparameter logistic item response theory models, and a twodimensional item response theory model. The 1PL IRT model, also known as the Rasch model, has the following assessment pass likelihood: for student and item , where is student proficiency and is item difficulty, and is the logistic link function [?]. The 2PL model extends the likelihood as follows: , where is the item discriminability [?]. The 2D MIRT model, which is a multidimensional generalization of 2PL, has the following pass likelihood: , where are the student factors, are the item factors, and is the item offset [?]. Note that we have not explicitly included Bayesian Knowledge Tracing as a benchmark model since it requires content modules to be annotated with concept tags, while the Latent Skill Embedding does not.
We use tenfold crossvalidation to select the regularization parameter and learning update variance for the embedding model (see Figure Latent Skill Embedding for Personalized Lesson Sequence Recommendation for the exploration on Book A), as well as regularization parameters for the benchmark IRT models. On each fold, we train on the full histories of 90% of students and the truncated histories of 10% of students, and validate on the assessment interactions immediately following the truncated histories. Truncations are made just before the last assessment interactions for each student (maximizing the size of the training set). We have also examined the effect of randomizing the truncation in student histories, and find no substantial changes to our results.
After selecting model hyperparameters using crossvalidation, we evaluate the models on a heldout test set of students (20% of the students in the complete data set) that was not visible during the earlier parameter selection phase. The same truncation method is used for evaluation on the test set.
Our performance metric is area under the ROC curve (AUC), which measures the discriminative ability of a binary classifier that assigns probabilities to class membership.
To gain insight into which components of the embedding model contribute most to its predictive power, we conduct a lesion analysis. For the sake of simplicity, we restrict ourselves to using a twodimensional embedding (later, we describe the effect of varying the embedding dimension ). We start with an embedding model that ignores lesson interactions and does not use bias terms. We then gradually add components to the embedding model to examine their effects on prediction AUC. Specifically, we evaluate embeddings with and without lesson parameters , prerequisite parameters for lessons, and bias terms . Each variant of the model corresponds to a row in Table Latent Skill Embedding for Personalized Lesson Sequence Recommendation.
From these results, we observe the following: including bias terms in the assessment result likelihood (Equation Latent Skill Embedding for Personalized Lesson Sequence Recommendation) gives a large and statistically significant performance gain ( for the standard ttest comparing validation AUCs of row 5 vs. 6 on Book A); an embedding with lesson prerequisites and bias terms performs comparably to the best benchmark IRT model.
One issue that may have affected the findings is the biased nature of student paths, which has been discussed by [?]. In the data, we observe that student paths are heavily directed along common routes through modules. We conjecture that this bias dulls the effect of modeling lesson prerequisites in the embedding, and causes the inclusion of bias terms to give a large performance boost. Most students attack a module with the same background knowledge, so an embedding that captures the variation in students who work on the same module is not as valuable.
In a regime where students who work on a module come from a variety of skill backgrounds, our model that includes lesson prerequisites may further improve results. Preliminary evidence for this is presented in Figure Latent Skill Embedding for Personalized Lesson Sequence Recommendation, where we recreate our analysis of the two Knewton data sets on several public data sets of student interactions and a private data set from an online language learning game [?, ?, ?, ?]. After conducting a lesion analysis for each data set, we compute the relative difference between the validation AUC for a twodimensional embedding with lesson prerequisites and the validation AUC for a twodimensional embedding without lesson prerequisites. We also measure the “entropy” of student paths by assuming that student paths can be modeled as Markov chains and computing the entropy of the transition probability matrix for each data set. A relationship exists between student path entropy and the relative AUC gain from using lesson prerequisites, providing some evidence to support the hypothesis that modeling lesson prerequisites is more beneficial in a setting with larger variation in student paths.
In other experiments, we explored the parameter space of the embedding model by varying the regularization constant and embedding dimension . Not explicitly shown are the results for changing . In summary, we find that increasing embedding dimension substantially improved performance for embedding models without bias terms, but that it has little effect on performance for embeddings with bias terms. The former is expected, since the embedding itself must be used to model general student passing ability and general assessment difficulty.
We perform several sensitivity analyses on Book A and observe the following: prediction AUC is most affected by a student’s recent history (see Figure Latent Skill Embedding for Personalized Lesson Sequence Recommendation); the number of full student histories in the training set has a strong effect on prediction AUC, via the quality of module embeddings (see Figure Latent Skill Embedding for Personalized Lesson Sequence Recommendation); prediction AUC decays when assessment results in the training set are noisy (see Figure Latent Skill Embedding for Personalized Lesson Sequence Recommendation); the length of a student’s history is weakly related to prediction AUC (see Figure Latent Skill Embedding for Personalized Lesson Sequence Recommendation). These findings lead to two key qualitative insights regarding model performance: (1) for a course offered regularly (e.g., over several semesters), the model will improve steadily as log data is collected from students who complete the course, and (2) the model performs best when assessments are crafted to test specific skills and minimize noise in outcomes.
The ability to predict future performance of students on assessments, while a useful metric for evaluating the learned embedding, does not address the more important task of adaptive tutoring via customized lesson sequence recommendation. We introduce a surrogate task for evaluating the sequence recommendation performance of the model based entirely on the observational data of student interactions, by assessing the model’s ability to recommend “productive” paths amongst several alternatives.
The size of the data set creates a unique opportunity to leverage the variability in learning paths to simulate the setting of a controlled experiment. For this evaluation, we use a larger version of the Book A data set, containing 14,707 students and 14,327 content modules. We find that the data contains many instances of student paths that share the same lesson module at the beginning and the same assessment module at the end, but contain different lessons along the way. We call these instances bubbles, for example see Figure Latent Skill Embedding for Personalized Lesson Sequence Recommendation, which present themselves as a sort of experimental evidence on the relative merits of two different learning progressions. We can thus use these bubbles to evaluate the ability of an embedding to recommend a learning sequence that leads to success, as measured by the relative performance of students who take the recommended vs. the notrecommended path to the assessment module at the end of the bubble.
We use the full histories of 70% of students to embed lesson and assessment modules, then train on the histories of heldout students up to the beginning of a bubble. The lesson sequence for a student is then simulated over the initial student embedding, using the learning update (Equation Latent Skill Embedding for Personalized Lesson Sequence Recommendation) to compute an expected student embedding at the end of the bubble (which can be used to predict the passing likelihood for the final assessment using Equation Latent Skill Embedding for Personalized Lesson Sequence Recommendation). The path that leads the student to a higher pass likelihood on the final assessment is the “recommended” path. Our performance measure is , where is the outcome at the end of the recommended path and is the outcome at the end of the other path ( is failing and is passing). This measure can be interpreted as “expected gain” (averaged over many bubbles) from taking recommended paths, or how “successful” the paths recommended by the model are when compared to the alternative.
This observational study is potentially confounded by many hidden variables. For example, it may be that one group of students systematically takes recommended paths while another group of students does not, leading to results at the end of a bubble that are mostly dictated by the teachers directing the groups, or other studentspecific hidden factors, rather than path quality. To best approximate the settings of a randomized controlled trial in our observational study, we use the standard propensity score matching approach for debiasing observational data [?, ?]. The key idea behind propensity score matching is to subset the observed data in a way that balances the distribution of the features (“hidden variables”) describing subjects in the two conditions, as it would be expected in a randomized experiment. The validity of any conclusion drawn from the observational data debiased in this way hinges on the assumption that all confounding variables that determine selfselection have been accounted for in the features prior to matching. In this study, we hypothesize that the set of all lesson modules and assessment modules (with outcomes) that the learner attempted throughout his or her duration in the online system is sufficient to compensate for any selfselection in the taken learning paths. Formally, we represent learners in a feature space such that , where if student passed module (lessons are always “passed”), if student has not completed module , and if student failed module .
We use PCA to map to a lowdimensional feature space where students are described by 1,000 features, which capture 80% of the variance in the original 14,327 features. A logistic regression model with regularization is used to estimate the probability of a student following the recommended branch of a bubble, i.e. the propensity score, given the student features (the regularization constant is selected using crossvalidation to maximize average loglikelihood on heldout students). Within each bubble, students who took their recommended branch are matched with their nearest neighbors (by absolute difference in propensity scores) from the group of students who did not take their recommended branch. Matching is done with replacement (so the same student can be selected as a nearest neighbor multiple times) to improve matching quality, trading off bias for variance. Multiple nearest neighbors can be matched (we examine the effect of varying ), trading off variance for bias.
Figure Latent Skill Embedding for Personalized Lesson Sequence Recommendation shows the results of the experiment, showing by how much students gain by following the path recommended by our embedding. We use the same embedding configuration as in row 6 of Table Latent Skill Embedding for Personalized Lesson Sequence Recommendation, which uses prerequisites and bias terms in a twodimensional embedding model with lesson. Naturally, our evaluation metric of gain in the pass rate from following a recommended path would depend strongly on the relative merits of the recommended and alternative paths. We therefore plot the gain that the recommended path achieves in relation to the difference in path quality, as measured by the absolute difference in pass rates between the two paths. Figure Latent Skill Embedding for Personalized Lesson Sequence Recommendation shows that the model generally able to recommend more successful paths, and this finding is robust to the choice of nearest neighbors used during propensity matching. As expected, the effect of the system recommendation is larger when there is a significant difference between the quality of the two paths.
We presented a general model that learns a representation of student knowledge and educational content that can be used for personalized instruction. The key idea lies in using a multidimensional embedding to capture the dynamics of learning and testing. Using a largescale data set collected in realworld classrooms, we (1) demonstrate the ability of the model to successfully predict learning outcomes and (2) introduce an offline methodology as a proxy for assessing the ability of the model to recommend personalized learning paths. We show that our model is able to successfully discriminate between personalized learning paths that lead to mastery and failure.
An implementation of the Latent Skill Embedding and the IPython notebooks used to conduct experiments are available online at http://siddharth.io/lentil.
We would like to thank members of Knewton for valuable feedback on this work. This research was funded in part by the National Science Foundation under Awards IIS1247637, IIS1217686, and IIS1513692, a grant from the John Templeton Foundation provided through the Metaknowledge Network at the University of Chicago and through the Cornell Presidential Research Scholars program.
 [1] E. Andersen. Automatic Scaffolding for Procedural Learning. PhD thesis, 2014.
 [2] M. Caliendo and S. Kopeinig. Some practical guidance for the implementation of propensity score matching. Journal of economic surveys, 22(1):31–72, 2008.
 [3] J. R. Carbonell. Ai in cai: An artificialintelligence approach to computerassisted instruction. ManMachine Systems, IEEE Transactions on, 11(4):190–202, 1970.
 [4] S. Chen, J. L. Moore, D. Turnbull, and T. Joachims. Playlist prediction via metric embedding. In Proceedings of the 18th ACM SIGKDD international conference on Knowledge discovery and data mining, pages 714–722. ACM, 2012.
 [5] A. T. Corbett and J. R. Anderson. Knowledge tracing: Modeling the acquisition of procedural knowledge. User Modeling and UserAdapted Interaction, 4(4):253–278, 1994.
 [6] C. Ekanadham and Y. Karklin. Tskirt: Online estimation of student proficiency in an adaptive learning system. Machine Learning for Education Workshop at ICML, 2015.
 [7] M. Feng, N. Heffernan, and K. Koedinger. Addressing the assessment challenge with an online system that tutors as it assesses. User Modeling and UserAdapted Interaction, 19(3):243–266, 2009.
 [8] J. GonzálezBrenes, Y. Huang, and P. Brusilovsky. General features in knowledge tracing: Applications to multiple subskills, temporal item response theory, and expert knowledge. In Proceedings of the 7th International Conference on Educational Data Mining (accepted, 2014), 2014.
 [9] Grockit. What do you know?, 2011.
 [10] Knewton. The knewton platform: A generalpurpose adaptive learning infrastructure. Technical report, 2015.
 [11] A. S. Lan, C. Studer, and R. G. Baraniuk. Timevarying learning and content analytics via sparse factor analysis. In Proceedings of the 20th ACM SIGKDD international conference on Knowledge discovery and data mining, pages 452–461. ACM, 2014.
 [12] A. S. Lan, A. E. Waters, C. Studer, and R. G. Baraniuk. Sparse factor analysis for learning and content analytics. J. Mach. Learn. Res., 15(1):1959–2008, Jan. 2014.
 [13] A. Lesgold et al. Sherlock: A coached practice environment for an electronics troubleshooting job. 1988.
 [14] W. Linden and R. K. Hambleton. Handbook of modern item response theory. New York, 1997.
 [15] J. Moore, S. Chen, D. Turnbull, and T. Joachims. Taste over time: The temporal dynamics of user preferences. In Conference of the International Society for Music Information Retrieval Conference (ISMIR), pages 401–406, 2013.
 [16] E. O’Rourke, E. Andersen, S. Gulwani, and Z. Popovic. A framework for automatically generating interactive instructional scaffolding. 2015.
 [17] Z. Pardos, Y. Bergner, D. Seaton, and D. Pritchard. Adapting bayesian knowledge tracing to a massive open online course in edx. In Educational Data Mining 2013, 2013.
 [18] Z. A. Pardos and N. T. Heffernan. Tutor modeling vs. student modeling. In Proceedings of the TwentyFifth International Florida Artificial Intelligence Research Society Conference, pages 420–425, 2012.
 [19] C. Piech, J. Huang, A. Nguyen, M. Phulsuksombati, M. Sahami, and L. Guibas. Learning program embeddings to propagate feedback on student code. arXiv preprint arXiv:1505.05969, 2015.
 [20] C. Piech, M. Sahami, J. Huang, and L. Guibas. Autonomously generating hints by inferring problem solving policies. 2015.
 [21] C. Piech, J. Spencer, J. Huang, S. Ganguli, M. Sahami, L. J. Guibas, and J. SohlDickstein. Deep knowledge tracing. CoRR, abs/1506.05908, 2015.
 [22] M. J. Prerau, A. C. Smith, U. T. Eden, M. Yanike, W. A. Suzuki, and E. N. Brown. A mixed filter algorithm for cognitive state estimation from simultaneously recorded continuous and binary measures of performance. Biological cybernetics, 99(1):1–14, 2008.
 [23] G. Rasch. Probabilistic models for some intelligence and attainment tests. ERIC, 1993.
 [24] M. D. Reckase. Multidimensional Item Response Theory. Springer Publishing Company, Incorporated, first edition, 2009.
 [25] P. R. Rosenbaum and D. B. Rubin. The central role of the propensity score in observational studies for causal effects. Biometrika, 70(1):41–55, 1983.
 [26] A. C. Smith, L. M. Frank, S. Wirth, M. Yanike, D. Hu, Y. Kubota, A. M. Graybiel, W. A. Suzuki, and E. N. Brown. Dynamic analysis of learning in behavioral experiments. The journal of neuroscience, 24(2):447–461, 2004.
 [27] J. SohlDickenstein. Temporal multidimensional item response theory, 2014.
 [28] J. Stamper, A. NiculescuMizil, S. Ritter, G. Gordon, and K. Koedinger. Algebra i 20052006, algebra i 20062007, bridge to algebra 20062007. Challenge data set from KDD Cup 2010 Educational Data Mining Challenge, 2010.
 [29] C. Zhu, R. H. Byrd, P. Lu, and J. Nocedal. Algorithm 778: Lbfgsb: Fortran subroutines for largescale boundconstrained optimization. ACM Transactions on Mathematical Software (TOMS), 23(4):550–560, 1997.
