8.3.3 Data Mining in Learning Analytics

Educational Data Mining (EDM) is the application of data mining techniques — originally developed for commercial and scientific data analysis — to educational datasets, with the goal of understanding learners and learning environments. The field was formally codified by Baker and Yacef in their 2009 state-of-the-field paper, which established both a disciplinary identity and a taxonomy of methods that continues to organize EDM research.[LE-LS-AP-005] The International Educational Data Mining Society and the annual EDM Conference are the primary institutional homes for this work.[LE-LS-CO-006][LE-LS-CE-003]

The Baker-Yacef Taxonomy

Baker and Yacef identified five primary categories of EDM methods, each targeting a different kind of discovery:

1. Prediction — Building models that use known variables to predict an unknown or future variable. Classification (predicting categorical outcomes such as pass/fail) and regression (predicting continuous outcomes such as final exam score) are the primary techniques. Prediction in EDM includes both the KC-level mastery prediction covered in 8.3.1 and the course-level early warning predictions used in institutional analytics.

2. Clustering — Grouping learners or learning events into categories based on similarity, without pre-specified category labels. K-means, hierarchical clustering, and model-based clustering have all been applied to segment learner populations by behavioral profile, identify distinct problem-solving strategies, or discover types of assessment items that share underlying structure.

3. Relationship mining — Discovering associations, correlations, or causal structures among variables in educational data. Association rule mining (finding rules of the form "if a learner exhibits X behavior, they also tend to exhibit Y behavior") and correlation analysis are common techniques. Sequential pattern mining — a specialized form of relationship mining — discovers frequent orderings of events in interaction logs.

4. Knowledge discovery — Distilling verified models of cognitive or instructional processes from data, often in collaboration with human experts. This category bridges empirical data analysis and cognitive science: the goal is not just accurate prediction but interpretable models that advance understanding of how learning works.

5. Distillation for human judgment — Transforming complex analytical outputs into visualizations and summaries that support human decision-making. This category recognizes that the endpoint of EDM is often not automated action but improved human insight.

Ryan Baker's research program at the Penn Center for Learning Analytics has consistently spanned all five categories, with particular emphasis on prediction models for affect and engagement states and their distillation into actionable signals for instructors and adaptive systems.[LE-LS-PP-008]

EDM vs. Learning Analytics: Goals and Emphasis

EDM and Learning Analytics overlap significantly in methods but have historically emphasized different goals. EDM has emphasized algorithmic development, methodological rigor, and building models that generalize across contexts. Learning Analytics, as practiced through the LAK Conference and the Journal of Learning Analytics, has emphasized deployment in institutional contexts, equity concerns, stakeholder engagement, and the translation of analytical outputs into educational decisions.

The Journal of Educational Data Mining publishes detailed algorithmic work — new model architectures, comparative evaluations, benchmark datasets — while the Journal of Learning Analytics tends to publish implementation studies, ethical analyses, and institutional case studies.[LE-LS-JO-002] Learning engineers working in applied contexts need both: the methodological rigor that EDM brings and the deployment wisdom that LA practice has accumulated.

Sequence Mining and Process Mining

Among the EDM techniques with the most direct relevance to learning engineering, sequence mining and process mining stand out. Sequence mining discovers frequent ordered patterns in interaction logs — for example, the sequence of hint requests, errors, and correct responses that characterizes different problem-solving strategies. Process mining reconstructs the process model (the set of decision points and paths) that generated an observed set of event sequences.

These techniques are particularly valuable for understanding how learners navigate non-linear content structures, where they get stuck in adaptive systems, and what distinguishes efficient from inefficient learning paths. Kenneth Koedinger's DataShop platform was designed specifically to support this kind of sequence-level analysis, storing interaction logs in a format that enables KC-level, step-level, and session-level queries across millions of learning events.[LE-LS-PP-005] The OpenSimon Toolkit — including DataShop and LearnSphere — provides the open infrastructure through which the EDM community accesses and analyzes these large-scale educational datasets.[LE-LS-GL-010]

Practical Application in the LE Cycle

The Learning Engineering Toolkit positions data mining as the most appropriate analytical approach during the "challenge identification" and "data investigation" phases of the iterative LE process — when the learning engineer is still forming hypotheses rather than testing them.[LE-LS-GL-007] A clustering analysis that reveals three distinct learner behavioral profiles in a new course context, for example, generates hypotheses about what those profiles mean and what interventions each might need. Those hypotheses then motivate the building of inference models (to detect which profile a new learner belongs to in real time) and predictive models (to forecast where each profile is heading).

This sequential use of method families — EDM for exploration, inference modeling for real-time characterization, predictive modeling for proactive intervention — reflects mature learning analytics practice.

Community Recommended Reading

Data-Informed Course Improvement: The Application of Learning Engineering in the Classroom — Anne Fensie (2023, ICLS 2023)

ID: lecommons-rl-data-informed-course-improvement-the-application-of-learning-engineeri · type: reading list item · confidence: medium · expert-validated: false

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

Source: goodell-ch5-le-is-a-process · confidence: medium · expert-validated: false

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

8.3.3 Data Mining in Learning Analytics

The Baker-Yacef Taxonomy

EDM vs. Learning Analytics: Goals and Emphasis

Sequence Mining and Process Mining

Practical Application in the LE Cycle

Further Reading

Landmark Academic Papers

Policy, Reports & Grey Literature

Key People

Organizations & Conferences

Community Recommended Reading