Definition
Learning analytics (LA) is the measurement, collection, analysis, and reporting of data about learners and their contexts, for the purpose of understanding and optimizing learning and the environments in which it occurs. This definition, widely adopted following the first LAK Conference in 2011, has three operative components: the data sources (learner interactions, assessment responses, behavioral signals), the analytical processes (statistical modeling, machine learning, visualization), and the purpose (improving learning). [LE-LS-CO-003]
Within learning engineering, the "collection" dimension of this definition is largely handled by the data instrumentation knowledge area — covering logging infrastructure, xAPI statement design, and telemetry pipelines. [LE-LS-SG-002] The learning analytics knowledge area focuses on what happens once data is collected: how it is modeled, interpreted, and converted into decisions or design changes. [LE-LS-GL-007]
It is important to note that in LE practice, "learning analytics" is used as an umbrella term that encompasses methods from Educational Data Mining (EDM), statistics, and machine learning. EDM is formally a distinct discipline — discovery-oriented, algorithmically focused, and community-organized through the International EDM Society — but LE teams routinely apply EDM methods in service of LA goals. [LE-LS-AP-005]
What Is Measured
The scope of what LA measures spans several categories:
Performance data — correct/incorrect responses, partial credit scores, time-to-completion, error patterns, and retry rates on specific knowledge components. Performance data is the primary input to student models such as Bayesian Knowledge Tracing. [LE-LS-AP-011] In the Cognitive Tutor lineage developed at CMU, each student interaction is logged at the knowledge-component level, producing fine-grained mastery estimates that drive adaptive problem selection. [LE-LS-AP-014]
Behavioral and engagement data — time-on-task, hint request frequency, navigation patterns, pause durations, content revisit rates, help-avoidance, and "gaming the system" behaviors (deliberate exploitation of hint sequences without genuine problem-solving). Ryan Baker's work at the Penn Center for Learning Analytics has been foundational in developing detectors for gaming, off-task behavior, and boredom from such behavioral traces. [LE-LS-PP-008] [LE-LS-CO-004]
Affective and motivational signals — derived from behavioral proxies or, in some systems, physiological or facial expression sensors. These are more experimental than performance-based measures but are an active research area. [LE-LS-AP-005]
Outcome data — pre/post test scores, course grades, retention and transfer measures, and longer-term outcomes such as course completion and downstream performance. The doer effect research by Van Campenhout et al. used outcome data across seven courses to isolate the causal impact of active practice on learning gains, demonstrating how large-scale LA infrastructure enables causal inference at scale. [LE-LS-AP-012]
Who Uses It and For What Decisions
LA serves four primary stakeholder groups with distinct decision-making needs:
Learners use LA-generated feedback (progress indicators, mastery estimates, recommendations) to self-regulate — deciding when to seek more practice, when to seek help, and how to allocate study time. Adaptive systems such as ASSISTments (serving 500,000+ students) provide this feedback automatically. [LE-LS-PP-008]
Instructors use dashboards and early-warning flags to identify struggling students, evaluate whether instructional activities are working, and adjust pacing or content emphasis. The "High-Leverage Opportunities" report identified teacher-facing dashboards as a top-priority LA application. [LE-LS-GL-003]
Learning engineers and instructional designers use analytics to evaluate the effectiveness of designed interventions, identify knowledge components where learners systematically struggle, and generate hypotheses for iterative design improvements. [LE-LS-AP-002]
Institutional administrators and policymakers use aggregate analytics for program evaluation, equity monitoring (detecting differential outcomes by demographic group), and resource allocation. [LE-LS-GL-003]
Grain Size and Scope
LA can be applied at multiple levels of scope: the individual learner (inferring mastery of a single KC), the learning session (detecting disengagement), the cohort (identifying items with anomalously high error rates), the course (evaluating instructional sequence effectiveness), the institution (monitoring equity metrics), and across institutions (discovering generalizable learning phenomena). [LE-LS-AP-005]
DataShop / LearnSphere, developed at CMU under Kenneth Koedinger's leadership, is the world's largest open repository of educational interaction data and enables cross-study analytics at the multi-institution level. [LE-LS-PP-005] This scale of data is what enables statistical power sufficient to detect small effect sizes and to discover patterns that would be invisible in any single study.
The scope of LA also includes the learning environment itself — not just what learners do, but how the system responds, what content was presented, and what interventions were triggered. Full-loop analytics, in which the system's own actions are logged alongside learner responses, are necessary for evaluating the effectiveness of adaptive features. [LE-LS-AP-002]