AIEd Knowledge BaseAIEd KB
On this page

Concept

Learning Analytics

created 2026-08-09 · updated 2026-11-08

Learning analytics — the measurement, collection, analysis, and reporting of data about learners and their contexts for the purpose of understanding and optimizing learning. AI has transformed learning analytics from descriptive dashboards to predictive and prescriptive systems.

Questions to Consider

  • Most people assume collecting more learning data automatically improves education. The page argues analytics become meaningful only when they feed back into an intervention — otherwise they merely describe or flag without changing learning. Where have you seen data collected that never led to any action?
  • Imagine a dashboard tells you a student is 'at risk' — a prediction. What separates that from genuinely actionable guidance that a teacher or institution can actually carry out? The page suggests prediction alone is not enough.
  • The page notes AI has moved learning analytics from describing what happened to predicting what will happen and prescribing what to do next. Which of these three generations have you experienced, and what was missing in the others?
  • In one study, three different AI models produced sharply different support plans for the same student data, and the links between analytics indicators and recommended help were mostly weak. What does this suggest about trusting an AI's advice at face value?
  • Learning analytics sits in a privacy tension: the more granular the data, the more revealing — and the more powerful the intervention. Where would you draw the line on what is collected about you or your students, and who should decide?

Introduction

AI-enhanced analytics

  • Predictive analytics: Machine learning on learner interaction data predicts outcomes — from at-risk identification to knowledge state estimation.

  • Validation matters as much as prediction. Schuetze, Yan, and Carvalho (2025) show that predictive knowledge-state models (BKT, BKT-with-Forgetting, AFM) look accurate when fit retroactively to a full session history, yet under time-based cross-validation — predicting the next session from prior ones, how analytics are actually deployed — they overestimate learner performance, miss spacing/forgetting dynamics, and can mis-order practice conditions. The caution for analytics: retrospective fit can mask poor forward predictive validity on longitudinal data, so indicator and dashboard models should be validated walk-forward.

  • Validate the item-attribute mapping behind the estimate. The Q-matrix linking items to attributes is an assumption, not data: rejected entries for 14 of 918 analytical-thinking items, once corrected, cut a spurious knowledge-state pattern from 14.2% to 3.8% — so mastery analytics should audit their Q-matrix, not only their predictive fit (Feng & Huang, 2026).

  • Curriculum-anchored predictive analytics: Pradeesh et al. (2026) estimate knowledge states within Outcome-Based Education by tracing course outcomes directly from LMS interaction and attainment data, using OBE affinity mappings (course–program outcome relations) to structure concept links and a memory-augmented network to model cross-outcome impact — reaching 89.81% AUC and beating DKT, DKVMN, EKT, and SimpleKT on live university engineering data, while staying only competitive (not superior) on general-purpose ASSISTments data.

  • Interpretable progress prediction with an action window: Zhang, Jeffries & Koprinska (2025) predict module-level student progress in large-scale online programming courses from content-interaction log features, using glass-box decision trees that match black-box accuracy (85–91%) while flagging "No submission" dropout outcomes up to 7–8 days before module deadlines — an explicit, real-time window for intervention rather than a bare risk flag, and an exploratory typology of disengaged-at-risk, disengaged-but-successful, and engaged high-performer profiles.

  • Approval does not predict the engagement pattern. Trace data from a learner-facing analytics tool in a GAI-supported programming course split 27 students into Deep, Self-Oriented and Peer-Oriented Explorers plus a final-stage Limited Explorers profile, yet favorable perceptions of the assistant did not differ significantly across those profiles (Sun et al. (2026)).

  • Federated, explainable risk modeling across institutions (2026). Villegas-Ch et al. (2026) extend risk modeling beyond single-institution prediction by training a multitask (performance + dropout) model across simulated institutions via federated learning, so raw student data never leaves each institution. Under controlled heterogeneity (label skew, class imbalance, temporal drift, structural missingness) the model preserves ranking accuracy (OULAD AUC 0.918) and structurally stable feature-importance rankings, yet probabilistic calibration drifts — decoupling ranking performance from probability reliability. For early-warning systems this is a caution that threshold-based interventions may need per-institution calibration, and an argument for evaluating analytics along discrimination, calibration, robustness, and explainability at once.

  • Synthetic-data validation is a prerequisite for privacy-preserving analytics. A structural check on synthetic versions of four annual study-habit cohorts (117-120 learners over eighteen weeks) found partition descriptors agreeing closely in two of four cohorts while week-to-week variation ran 2.6 to 4.9 times lower without exception (coefficient of variation 0.086-0.152 against 0.399-0.539 for the real cohorts), and only 36% of findings reached on synthetic pilot data were confirmed on the real data across twenty-five validation requests (Inoue & Yasutake (2026)).

  • Engagement analytics: Engagement measurement and intensity modeling quantify how students interact with AI systems.

  • Completion rates can hide an engagement collapse. In a group-randomized university trial, a course-integrated AI tutor left homework submission unchanged while platform participation fell sharply — page views by 0.37 standard deviations, active days by 0.51 — so the analytics signal that mattered was the activity the tool displaced, not the completion metric that stayed flat (Liu et al. (2026)).

  • Feedback analytics: Feedback classification and quality assessment analyze the feedback students receive.

  • Network analysis: Co-Occurrence Network Analysis and epistemic network analysis reveal interaction patterns.

  • Privacy tensions: Privacy concerns grow as analytics become more granular and AI-driven.

  • A governance review framework for student data. LEAGUE proposes six pillars — Lawfulness, Equity, Agency, Governance, Utility, and Ethics by Design — treating FERPA and GDPR compliance as the floor and recommending scheduled reassessment as models drift (Varadaraju & Vijayakumar (2026)).

Sensor and gaze analytics is a second instrument family feeding the same learner models, drawing on eye tracking, EEG, electrodermal activity and heart rate to estimate attention, engagement and cognitive load (Sensor and Gaze Analytics). Its validity and consent problems are steeper than trace data's, because the measurements are biometric and the step from body to mental state is an interpretation.

Model course concurrency, not just sequence. TRACE gives every course in a semester the same positional encoding, and jointly predicting courses and grades cut grade-prediction error to 0.1339 MAE — a 46.4% reduction over a grades-only transformer — on ten years of institutional data (Savala (2026)).

Interaction data is a representation problem before it is a modeling one. Taufer et al. (2026) treat each human-AI exchange as one structured record carrying the computation that motivated the request, the request, the response, the user's feedback and the execution that followed, rather than as an isolated prompt. Across 140 student-assignment pairs in a 35-student deployment, interaction count alone correlated with successfully executed cells at r = 0.12, and the same call could be an error recovery, a next step, a concept check or a validation depending on the state around it. Thirty-seven pairs with access produced no interactions at all, so absence of AI use in a log cannot be read as absence of need.

The learning analytics cycle

Learning analytics is canonically framed as a cycle that begins with learner activity producing data, which is processed into measures and indicators that are then translated into interventions — and the intervention feeds back into learner activity to close the loop. The intervention step is what distinguishes analytics from mere monitoring or prediction: without it, analytics describe and flag but never change learning. This cycle is the organizing frame for understanding where AI tools (dashboards, feedback generators, prescriptive recommenders) sit in the pipeline and which step they automate.

A learner-facing dashboard is only as good as its reach: in a Hong Kong Grade 9 writing study a prompt classifier reached a macro F1 of 0.757, about a third of the 46 students opened the three GenAI-use dashboards and the group differences in copying and learning-oriented prompting were not reliable (Fong et al. (2026)).

A frequency encoding of qualitative data is an entry point, not a verdict: in a ten-participant study, some instructors read frequency encodings as a productive way in while others warned they obscure rare but critical responses, so the design keeps every aggregate view linked back to verbatim student text (Nguyen et al. (2026)).

From description to intervention

  • The evidence base is not yet intervention evidence. Synthesizing four studies across student, disciplinary, educator and institutional levels, Rienties, Divjak and Law (2026) find the field dominated by cross-sectional, self-report, discourse-based and expert-judgment designs, and call for intervention research testing causal mechanisms and longer-term, distributional outcomes.

A randomized-encouragement design supplies the causal anchor. In a randomized-encouragement experiment on ASSISTments, Lim, Heffernan & Sales (2025) randomly prompted teachers new to the platform to view the assignment report after creating an assignment. Among the 330-teacher analytic sample (156 control, 174 encouragement; 281 for value-added), viewing the first-assignment report significantly raised next-assignment completion (Anderson-Rubin F = 4.092, p = 0.044) - at least 2 percentage points overall. In classes under 44 students the gain was at least 24.5 percentage points (AR F = 7.475, p = 0.007), while only 54.0% of 121,890 assignments had their reports viewed and starting and value-added outcomes did not move. The same platform data carries the field's largest closed-loop rating case: Johnson et al. (2025) analyzed 5,214,211 student-question sessions in which thumbs-down (2.00 per 1,000) tracked generator-controlled features and student-level variance dominated question-level variance (BIC 101,572.9 vs 123,342.0).

Learning analytics has evolved through three generations in the knowledge base: descriptive (what happened?), predictive (what will happen?), and prescriptive (what should we do?). AI enables the prescriptive layer — analytics that directly trigger instructional interventions. A key frontier is the actionability gap: SC2R (Le, Abel & Laforge 2026) shows that prediction alone is insufficient for decision support, and that counterfactual recourse becomes operationally meaningful only when recommendations are semantically feasible and machine-checkable — constrained by timing, budget, immutability, and availability via SHACL validation, rather than merely model-valid. This moves the field beyond risk scores toward recommendations that institutions can actually enact, with human oversight preserved.

A direct empirical test of the prescriptive layer comes from López-Pernas et al. (2026), who asked three LLMs to recommend support plans for 4,500 synthetic student vignettes.

A complementary, teacher-centered test of how analytics reach the classroom comes from Mejia-Domenzain et al. (2026), who designed a learning analytics dashboard (DashED) to communicate ML-derived Self-Regulated Learning profiles to teachers in two blended contexts. Their 100-teacher study shows that the presentation of analytics is itself a barrier to action: teachers systematically preferred simpler, more traditional charts (bar plots, pie charts) even when more complex designs (e.g., heatmaps) yielded richer insights, and higher visualization literacy predicted deeper, more detailed interpretation (e.g., more teachers identifying trends in time-series data). For group comparison, teachers favored superposition over juxtaposition and full-information plots over explicit difference encoding. The actions teachers proposed were shaped by the content represented and their Teaching level rather than the plot type — university teachers favored weekly tests and course-level adaptation, while vocational teachers proposed direct, individualized coaching — so the prescriptive step depends as much on how analytics are visualized as on the model. Their finding is cautionary: correlations between LA indicators and recommended support were mostly weak, cross-model recommendations diverged sharply for the same student, and support was frequently allocated regardless of who needed it most. The authors conclude that current LLMs are not yet reliable as prescriptive models for student support at scale, reinforcing that the prescriptive step still requires validation, fine-tuning, and human oversight rather than off-the-shelf automation.

The step after the score — who gets support, and whether it was allocated where it was needed — belongs to student support and success, which carries the outreach, referral, and allocation evidence this page's prescriptive layer feeds.

An integrity outcome enters the same predictive layer. Akçapınar (2026) predicts AI-assisted cheating risk from the first eight weeks of Moodle and video-player traces, reaching an AUC of 0.763 with logistic regression, and argues the signal is usable precisely because it arrives before the exam rather than during it. A complementary line targets the interpretation of analytics itself. Marques et al. (2026)'s FACTRIA framework organizes the biasing factors behind an indicator - pipeline, institutional, course, and demographic - and a reflection-prompting chatbot raised the sub-factors 11 stakeholders weighed per case from 1.41 to 3.45 (d = 2.15), shifting them from direct readings to conditional ones.

Methods and network analysis

Network methods are core to learning analytics: transition network analysis (TNA) models temporal sequences of learner actions (e.g., the revision and chat loops in chatbot-scaffolded writing), and epistemic network analysis (ENA) maps how codes/constructs co-occur across activity — together revealing the process of learning and learner-AI interaction rather than only its product.(Penny: Transition Network Analysis of Learner-Chatbot Interactions in Scaffolded EFL Writing)(Tracing GenAI Literacy: Student-AI Interaction Patterns in Academic Writing)

  • Clustering clickstreams in self-directed model building. An, Hammock & Goel (2025) combined activity sequence analysis, hierarchical clustering, and Markov chain models on the clickstreams of 315 online learners who built 822 ecological models in VERA, distilling nine fine-grained, transition-based behavioral clusters into three broader patterns (Observation, Construction, Exploration). Their work demonstrates that combining sequence analysis with Markov chain modeling can uncover meaningful behavior in unstructured, self-directed tasks even in the complete absence of demographic or contextual data.

  • LA and GenAI shape learning design differently (2026). Claassen et al. (2026) used ENA on 11 instructor focus groups to compare how learning analytics versus Generative AI inform Learning Design decision-making. LA discussions centered on contextual information, course-level design, and creative Problem Solving (LA for diagnosing engagement and targeting support), while GenAI discussions centered on assessment design and designing for student self-determination (GenAI for ideation and assessment development). Context and Creativity were central across both — a reminder that analytics inform design only within pedagogical context and instructor autonomy.

  • Talk beats fusion when the construct is social, and the best source is per construct. Rather than maximizing modalities, a LAK26 comparison of five feature configurations on collaborative modeling segments found text-only embeddings strongest overall (mean ROC AUC 0.650, 95% CI 0.615 to 0.685) against text plus logs (0.613) and text plus logs with a context window (0.623). Logs measured 0.526 alone, near chance for a social construct, yet they were the only source that improved detection of reflection (0.5862), while planning depended on the surrounding talk rather than on environment traces. Feature choice is a claim about the construct, not a budget decision (Zhang et al., 2026).

  • Behavioral traces are not measures of cognitive state — especially for special-needs learners. Vasileiou et al. (2026) warn that interaction traces (time on task, repeated errors, Help-Seeking, irregular progress) can help teachers locate where a learner needs support, but are not transparent measures of internal states: slow responding may reflect careful thought, language processing, distraction, inaccessible design, or technical difficulty. "Algorithms should not infer an executive deficit or diagnostic status from behavioral data alone" — analytics are defensible as one source of information within professional assessment and dialogue with the learner.

  • Design analytics: mining planned activity sequences rather than traces (2026). Divjak, Svetec and Horvat (2026) applied Markov chains and sequential pattern mining to the designed sequence of 29,064 teaching and learning activities across 554 courses planned in an open learning-design tool. The transition matrix peaked at Assessment to Discussion (0.332), self-transitions dominated Practice (0.317) and Acquisition (0.292), and the strongest consecutive rule was Acquisition to Assessment to Practice to Practice (confidence 0.743, lift 1.449), while the most frequent four-step path was Acquisition, Practice, Practice, Assessment (120 occurrences). Discussion and Assessment were the most reachable types and Production the most distant and sporadic. The study is a reminder that learning analytics need not begin with LMS traces: design-time data can expose the pedagogical grammar of a course before any student arrives, though the authors stress that resemblance to flipped, inquiry-based or project-based designs is not evidence of intent.

  • Self-explaining distilled LLMs (2026): A two-stage pipeline distills a black-box learning-analytics estimator and its post-hoc interpretation into a small, open-weight Large Language Models (LLMs) that returns both an individual-level estimate and a natural-language explanation. A faithfulness-first audit evaluates whether narrations match the attributions they describe; Simulation shows near-lossless recovery (r > .90) with an oracle mentor, offering a more transparent, deployable path for analytics (Distilling Black-Box Machine Learning into a Small, Self-Explaining Language Model for Learning Analytics).

  • Enablers of LA-based educational interventions (2026). Svetec, Divjak & Kadoić (2026) identify and prioritize seven enablers of trustworthy LA-based educational interventions via Delphi + AHP + SNAP: institutional strategic orientation, pedagogical & other research foundations, available resources, pedagogical support, ethics & data governance, stakeholder engagement, and quality assurance. Institutional strategic orientation ranked highest (and most influential on other enablers), with available resources second. Trustworthiness (ethical compliance, transparent/unbiased algorithms, pedagogical validity) is framed as the prerequisite without which LA-based interventions are not meaningful.

  • LLM interaction depth predicts task quality but not recall (2026). Tsiligkiris (2026) links turn-level LLM conversational telemetry (Depth/Volume/Pacing) to learning outcomes: explanation-seeking "depth" predicted independently marked task quality (β = 6.27) but not immediate recall — a dissociation between elaboration-driven comprehension and retrieval-driven consolidation that has implications for how Large Language Models (LLMs) interaction is measured and evaluated in LA.

  • A shared interaction unit is missing. Across 46 categorizations from 33 studies there is no shared meta-characteristic — similar labels name different phenomena — so the review proposes the interaction episode, a goal-directed bounded exchange, as the unit for correlating dialogue with skill acquisition (Borchers, Jansen & Weidlich (2026)).

  • Simulating collaborative discourse for learning analytics. Fang (2026) uses fine-tuned participant-specific LLM agents to reproduce collaborative problem solving dialogues, validated with Epistemic Network Analysis (ENA distance 0.17, permutation p = 0.65). The approach offers learning-analytics researchers a scalable way to generate authentic collaborative discourse for studying interaction dynamics, turn-taking, and thematic code trajectories without collecting new human data.

  • Open, reproducible data and trace-ready analytics. ASTRA releases a synthetic Benchmark with a trace-ready schema (N=540; 360 sessions; 1,440 episodes) for analyzing interaction and participation balance in collaborative programming. Log and trace data are the natural counterweight to self-report in this literature: the same construct is often measured twice, once by asking and once by observing, and the two do not always agree. Separately, an exploratory ML framework with SHAP analysis identified the learning-related constructs most associated with intended academic ChatGPT use among university students, prioritizing interpretability (An Exploratory Machine Learning Approach to Understanding Determinants of Future ChatGPT Use in Higher Education).

  • Reusable analytics infrastructure, validated only in-house (2026). Bai et al. (2026) reproduced published findings from three Georgia Tech AI assistants by changing configuration values rather than code, though the claim rests on reanalysis of existing single-institution datasets and the one capability extension was done by insiders.

Connections

Learning analytics connects to Knowledge Tracing (the core analytic), Formative Assessment (analytics-driven assessment), Learner Modeling and Adaptive Instruction (the learner representation analytics populate), Privacy (the ethical constraint), and Edtech Platform (where analytics are deployed). Because prescriptive analytics are increasingly evaluated on simulated learners — where synthetic student cohorts substitute for real cohorts in controlled tests — learning analytics also connects to student simulation.

  • What dashboards make visible decides what teachers act on. Köroğlu et al. (2026) reviewed 27 empirical studies (2016–2025) and built a socio-technical taxonomy of AI-supported lecturer decision-making across eight decision types: instructional, curriculum, assessment, feedback, learning-environment, emotional, administrative and ethical. Learning Analytics Dashboards were the most frequently reported system, and the coding shows support concentrated in the instructional, feedback and assessment decisions that behavioral text and log data can inform, while emotional, ethical, curriculum and learning-environment decisions were rarely supported. The authors read this as an attention effect rather than a capability limit: because the systems rendered behavioral student data visible and actionable, motivation, Metacognition, emotion and environment concerns fell outside what the data invited lecturers to consider.

  • Turning AI analyses into classroom talk is itself a teacher skill. In four secondary mathematics lessons, one teacher's 47 whole-class orchestrations converted dashboard output into collective argumentation by selecting and sequencing students' warrants, and the study names four recurring types (Lee, Park and Lee (2026)).

  • Analytics that lightens a teacher's load is not analytics that lightens a learner's. In a review of 23 studies of AI in embodied learning, 62.5% of 40 role tags were sensing roles that appraise a movement, trajectory or artifact — work that otherwise fell to a teacher — and the authors warn that a reduction in teacher workload should not be read as a reduction in learner load. Their reason is a limit on what this kind of coding can show at all: recording the function a system performs says nothing about its effect on the learner, so role frequencies cannot settle whether the system offloaded cognition or deepened it (Kim and Han (2026)).

Process-level instrumentation is the descriptive layer's next step down. Prol et al. (2026) extended an Open Source Parsons-problem platform to record every block placement, removal and submission as a chronological trace, pair each submission with per-block correctness coloring and an attempt history, and optionally pass the traces through an Large Language Models (LLMs) pipeline that labels recurring difficulty patterns for instructor review. Across 68 students in an upper-division Java software-design course and 36 in an introductory Python course the analysis surfaced the same three difficulties — choosing the wrong exception type, substituting return for throw, and incorrect control-flow ordering — which correctness-and-attempt counts cannot expose, because they answer whether the arrangement was right rather than what the process was. The AI acts as an interpreter for the instructor rather than a scorer of the student, and its output is framed as a reviewable hypothesis about a class's difficulties rather than a grade.

  • The learning system wall: log data doesn't generalize across platforms. MATHia, i-Ready, and IXL each define even basic constructs — an "attempt," a "hint" — differently, so models can't transfer between them. A work-in-progress response attacks the problem from the video side: a Gemini pipeline distills tutoring screen recordings into a single screenplay transcript aligned to native MATHia logs as ground truth. What video recovers that no single stream does: whiteboard Scaffolding, silent calculator work in external tools, and off-screen context that native logs reduce to "idleness." Open challenges remain in temporal anchoring, clock drift, and separating platform events from other learning processes. (Towards breaking the learning system wall using multimodal tutoring transcriptions)
  • Interaction logs can lose to a constant baseline. Semester-level features from 16,851 LLM-tutor exchanges beat a median predictor of assignment scores by 0.114 points, an interval crossing zero, with 62.1% of error in the lowest quartile (Dong et al. (2026)).

Dashboards and early-alert systems concentrate the risk of Automation Bias: a flag asserts a judgment that a person may accept because it exists rather than because it was checked.

Holding the multimodal indicators constant and adding only a narrative explanation raised pre-service teachers' interpretation (η² = 0.166) and meaning construction (η² = 0.220), while actionable transformation did not move — the interpretation layer, not the data, was the bottleneck (Liu et al., 2026).

Common Misconceptions

  • "A knowledge-state estimate inherits its validity from the data." The Q-matrix linking items to attributes is an assumption, not data: rejected entries for 14 of 918 analyzed items means the mapping itself required adjudication. Validating the mapping is a step, not a formality.
  • LLMs reading clickstreams against named learning strategies. Radmehr et al. (2026) stage an LLM pipeline that locates evidence for a strategy in a student's trace, judges adherence, and cites the actions behind the judgment, reaching overall quality of 0.735 in a harder simulation against 0.704 for a flat baseline.
  • "Analytics need clickstream traces." Knowledge states can be estimated within Outcome-Based Education by tracing course outcomes directly from LMS interaction and attainment data, using OBE affinity mappings — the institutional record supplies the signal.
  • "Prediction without an action window is useful." Progress prediction is most usable when it is interpretable and paired with an action window: Zhang, Jeffries and Koprinska (2025) predict module-level outcomes, which is a form an advisor can act on, unlike a raw score.

Connected Concepts

Connected Articles

Connected FAQs

Metadata

Learning and instructionStudent Engagement
Technologies and techniquesKnowledge Tracing, Learner Modeling and Adaptive Instruction, Edtech Platform
Assessment and measurementFeedback, Formative Assessment
Research methods and evaluationAI Ed Evaluation
Equity, ethics, and responsible usePrivacy
Page kindEvaluation
SourceHistoryEditChat

Embed this page

Copy the code below to embed this page in a learning management system or other website. The embedded version shows only the page content — it hides the site header, navigation, and footer.