8.3.1 Predictive Modeling

Predictive modeling in learning analytics uses historical and real-time behavioral data to estimate future states of learner knowledge, performance, or engagement. The defining characteristic of predictive models is that their outputs are forward-looking probability estimates — not summaries of what has already happened. This prospective orientation is what makes predictive modeling so directly useful to learning engineers: it enables proactive rather than reactive interventions.

Bayesian Knowledge Tracing: The Foundational Model

The prototype for predictive modeling in learning analytics is Bayesian Knowledge Tracing (BKT), developed by Albert Corbett and John Anderson at Carnegie Mellon and formally published in 1994.[LE-LS-AP-011] BKT treats each knowledge component (KC) as a latent binary variable: either a learner has acquired it or has not. The model uses four parameters to update the probability of mastery with each practice opportunity:

• P(L₀) — the prior probability that a learner already knows the KC before instruction
• P(T) — the probability of transitioning from unlearned to learned after each opportunity
• P(S) — the probability of a learner who knows the KC making a "slip" (answering incorrectly)
• P(G) — the probability of a learner who does not know the KC "guessing" correctly

Using Bayes' theorem, the model updates the posterior probability of mastery after each response. When the posterior exceeds a mastery threshold, the system moves the learner to the next KC. This mechanism was central to the Cognitive Tutor systems deployed in thousands of U.S. schools throughout the 1990s and 2000s.[LE-LS-AP-014]

BKT was co-developed by Albert Corbett, who also collaborated on the broader Intelligent Tutoring System architecture documented by Corbett, Koedinger, and colleagues.[LE-LS-PP-015][LE-LS-AP-002] The model remains the most widely deployed student model in production learning systems worldwide, precisely because its parameters are interpretable and its computational requirements are modest.

Performance Factors Analysis

Performance Factors Analysis (PFA) extends BKT's logic by modeling learning as a function of cumulative prior successes and failures on each KC, rather than using the four-parameter Bayesian update. PFA emerged from analyses of DataShop log data and has shown competitive predictive accuracy with BKT on a range of educational datasets. Kenneth Koedinger's DataShop initiative at CMU provided the large-scale data infrastructure that made PFA comparisons possible.[LE-LS-PP-005]

The practical difference between BKT and PFA for learning engineering is in how they represent forgetting and spacing effects. PFA's linear structure makes it easier to incorporate cumulative practice counts as a predictor, while extensions of BKT have been developed to model temporal forgetting through decay parameters.

Student Success Prediction and Early Alert Systems

Beyond KC-level mastery prediction, predictive modeling at the course or program level addresses the question: which learners are at risk of poor outcomes or withdrawal? These models typically aggregate a broader set of features — submission timing, forum participation, video engagement, grade trajectory — into a course-level risk score.

The Learning Engineering Toolkit describes this form of predictive analytics as a key component of the LE practitioner's toolkit for understanding learner trajectories and designing timely support.[LE-LS-GL-007] The High-Leverage Opportunities report identified predictive advising — flagging at-risk learners and routing them to institutional support — as one of the highest-leverage LE infrastructure investments available to institutions.[LE-LS-GL-003]

Predictive models for early alert systems face a distinctive challenge: they must be accurate enough early in the term to allow intervention, but early in the term there is little data to train on. Feature engineering — selecting and constructing the right input variables from log data — is therefore a critical skill. Work published in the IEEE Transactions on Learning Technologies has documented feature sets that yield useful prediction from week-one data in large online courses.[LE-LS-JO-003]

Regression Models for Outcome Forecasting

Logistic and linear regression models, despite their simplicity relative to deep learning approaches, remain workhorses of applied learning analytics predictive modeling. They are interpretable — coefficients directly express the marginal predictive contribution of each feature — and they generalize reasonably well across institutional contexts when trained on sufficient data. The EDM Conference regularly publishes comparative studies showing that regression baselines outperform more complex models when training data is limited, which is the typical situation for a single course or institution.[LE-LS-CE-003]

Evaluation and Deployment Considerations

Predictive models must be evaluated not just on aggregate accuracy (AUC, RMSE) but on their behavior across subpopulations. A model that performs well on average may systematically underestimate risk for historically underserved students, producing the appearance of equity while perpetuating unequal support allocation. The learning engineer deploying a predictive model has an obligation to examine performance disaggregated by demographic group before deployment — a principle that is discussed further in section 8.5.

Further Reading

Source: wrgr/lecommons — curated by the learning engineering community. Confidence: medium — lecommons-curated; not yet independently expert-validated in this context.

Landmark Academic Papers

• Knowledge Tracing: Modeling the Acquisition of Procedural Knowledge — Albert T. Corbett, John R. Anderson (1994). User Modeling and User-Adapted Interaction · doi:10.1007/BF01099821 · ~2,500 citations · tier: foundational

  Source: lecommons/landscape/data/papers.json · ID: LE-LS-AP-011 · confidence: medium · expert-validated: false

• Cognitive Tutors: Lessons Learned — John R. Anderson, Albert T. Corbett, Kenneth R. Koedinger et al. (1995). Journal of the Learning Sciences · doi:10.1207/s15327809jls0402_2 · ~1,800 citations · tier: foundational

  Source: lecommons/landscape/data/papers.json · ID: LE-LS-AP-014 · confidence: medium · expert-validated: false

Policy, Reports & Grey Literature

• High-Leverage Opportunities for Learning Engineering — Ryan S. Baker, Ulrich Boser, Allison Shelley (2021). University of Pennsylvania Center for Learning Analytics · link

  Source: lecommons/landscape/data/grey_literature.json · ID: LE-LS-GL-003 · confidence: low · expert-validated: false

• Learning Engineering Toolkit — Jim Goodell, Janet Kolodner (2023). Taylor & Francis / Routledge · link

  Source: lecommons/landscape/data/grey_literature.json · ID: LE-LS-GL-007 · confidence: low · expert-validated: false

Key People

• Albert T. Corbett (profile), Carnegie Mellon University — Co-developed BKT; co-designed Algebra Cognitive Tutor

  Source: lecommons/landscape/data/people.json · ID: LE-LS-PP-015 · confidence: medium · expert-validated: false

• Kenneth R. Koedinger (profile), Carnegie Mellon University — DataShop founder; Cognitive Tutor pioneer

  Source: lecommons/landscape/data/people.json · ID: LE-LS-PP-005 · confidence: medium · expert-validated: false

Lecommons enrichment applied 2026-04-17. All items pending expert validation.

Community Recommended Reading

Designing a Student Progress Panel for Formative Practice: A Learning Engineering Process — Rachel Van Campenhout, Michelle Selinger, Bill Jerome (2023, ICLS 2023)

Chapter 5: LE Toolkit — Learning Engineering is a Process — Aaron Kessler et al. (2022)

Context enrichment applied 2026-04-17. See wrgr/lecommons for source data.