OULAD: Open University Learning Analytics Data

• OULAD is a comprehensive, open dataset comprising course data, demographic records, and over 10 million VLE clicks, crucial for benchmarking educational analytics.
• It is widely reused in LAK, EDM, and AIED research, supporting diverse analyses from early-warning prediction to intervention policy evaluation.
• The dataset requires meticulous preprocessing and temporal filtering to prevent leakage, thereby enhancing reproducibility and reliable predictive modeling.

Searching arXiv for papers on OULAD and related learning analytics benchmarks. The Open University Learning Analytics Dataset (OULAD) is an open, anonymized educational dataset from the Open University that has become a recurrent empirical resource in learning analytics, educational data mining, and artificial intelligence in education. In recent survey work on open educational datasets, it appears as dataset ID 109, is described as “course data and demographics,” and is recorded as having been reused in six papers across LAK, EDM, and AIED during 2020–2024; separate methodological papers describe it as containing data for 32,593 students across 22 module presentations, with demographics, registration records, assessment data, and virtual learning environment (VLE) interaction logs comprising more than 10 million clicks (Švábenský et al., 19 Feb 2026, Howard, 14 Jan 2025, Ayady et al., 17 Jul 2025). Some recent studies abbreviate the same resource as “OULA” while explicitly naming it the “Open University Learning Analytics Dataset” (Muresan et al., 11 Jan 2026, Junejo et al., 2024).

1. Status within the open learning analytics ecosystem

Within the recent open-data survey of learning analytics, OULAD occupies a conspicuous position not because it is the only public dataset, but because it is among the comparatively small set of repeatedly reused ones. The survey manually examined 1,125 papers from LAK, EDM, and AIED over five years, identified 172 datasets used in 204 publications, and found that most datasets were used only once: 132 datasets appeared in a single paper, whereas only 20 datasets were used in 3+ papers. OULAD belongs to this smaller reuse-intensive group, with a recorded usage count of 6, placing it among the more frequently reused public datasets in the surveyed literature (Švábenský et al., 19 Feb 2026).

The same survey also reports that a prior review of 74 learning analytics papers on predictive analytics found that only three papers used public datasets, and all three used OULAD. This does not establish a universal benchmark status in a formal sense, but it strongly suggests that OULAD functions as a de facto public reference dataset in parts of predictive learning analytics where openly reusable data remain scarce (Švábenský et al., 19 Feb 2026).

Its prominence in this ecosystem is tied to characteristics repeatedly emphasized across the literature: openness, anonymization, course-linked educational context, demographic information, and sufficiently structured interaction and assessment traces to support predictive modeling, fairness-oriented analysis, survival modeling, and benchmarking. The survey’s broader interpretation is that datasets with these properties help reproducibility, collaboration, and comparative evaluation, and OULAD is presented as one of the clearest examples of that role in current practice (Švábenský et al., 19 Feb 2026).

2. Dataset scope, structure, and educational context

OULAD is described as a large, publicly available dataset released by the Open University, with anonymized data from 32,593 students across 22 module presentations drawn from 7 modules: AAA, BBB, CCC, DDD, EEE, FFF, and GGG (Howard, 14 Jan 2025). One paper groups these as three Social Sciences modules—AAA, BBB, and GGG—and four STEM modules—CCC, DDD, EEE, and FFF (Howard, 14 Jan 2025).

Aspect	Reported value
Students	32,593
Module presentations	22
Modules	AAA, BBB, CCC, DDD, EEE, FFF, GGG
Raw organization	Seven linked tables/files
Core data families	demographics, registration, assessment, VLE interactions
Interaction scale	more than 10 million clicks

The raw data are organized into seven linked tables or files under a database schema. Across the papers, the repeatedly named components are studentInfo, studentRegistration, assessments, studentAssessment, studentVLE, and vle (Howard, 14 Jan 2025, Le et al., 25 May 2026). The dataset is therefore relational rather than immediately analysis-ready, and several papers explicitly note that substantial preprocessing is usually required before modeling (Howard, 14 Jan 2025).

The VLE portion is particularly important in later methodological work. It is represented as daily summaries of student clicks in one account, and as fine-grained clickstream interaction logs in others, with more than 10 million clicks or clickstream entries overall (Howard, 14 Jan 2025, Tertulino, 23 Aug 2025, Ayady et al., 17 Jul 2025). These logs are frequently linked to activity-type metadata through studentVle.id_site = vle.id_site, allowing interaction counts to be re-expressed by resource or activity category rather than only by raw timestamped events (Le et al., 25 May 2026).

The dataset’s educational context is formal higher education delivered through the Open University. Several studies exploit the fact that module and presentation identifiers are explicit, so the same dataset supports analyses at the level of the student, the student–module–presentation enrollment, the course presentation, or temporally sliced within-course trajectories (Silva et al., 10 Apr 2026, Le et al., 25 May 2026, Muresan et al., 11 Jan 2026).

3. Units of analysis and target definitions

A distinctive feature of OULAD research is that the same underlying data support multiple units of analysis. In leakage-controlled early outcome prediction, each instance is one learner in one course run indexed by (code_module, code_presentation, id_student) (Le et al., 25 May 2026). In temporal dropout modeling, the unit is the enrollment, defined as the student–module–presentation combination (Silva et al., 10 Apr 2026). In heterogeneous graph modeling, a registration node represents a student’s enrollment in a specific course/presentation instance (Muresan et al., 11 Jan 2026). In federated learning, the course module itself can be treated as the federation unit, with AAA through GGG mapped to separate simulated institutions (Tertulino, 23 Aug 2025).

Outcome definitions vary materially across studies. OULAD’s original final-result labels are reported as Distinction, Pass, Fail, and Withdrawn (Muresan et al., 11 Jan 2026, Ayady et al., 17 Jul 2025, Junejo et al., 2024). From these, different tasks are constructed:

• Binary success prediction: Pass + Distinction versus Fail + Withdrawn (Muresan et al., 11 Jan 2026, Le et al., 25 May 2026, Ayady et al., 17 Jul 2025).
• At-risk failure prediction: at_risk = 1 if final result is Fail, at_risk = 0 if final result is Pass or Distinction, with Withdrawn excluded (Tertulino, 23 Aug 2025).
• Four-class forecasting: Distinction, Fail, Pass, and Withdrawn treated as separate labels, with output-layer label encoding 0,1,2,3 respectively in one 1D-CNN study (Junejo et al., 2024).
• Temporal dropout prediction: withdrawal is treated as an event only when the enrollment has administrative status Withdrawn together with a valid unregistration date; all others are right-censored at end of observation (Silva et al., 10 Apr 2026).
• Decision-policy learning: the target is not a course outcome but an oracle intervention action chosen from TUTOR_CALL, CONTENT_PUSH, REMINDER, REDUCE_LOAD, and NO_ACTION (Atkinson, 28 Jun 2026).

These varying operationalizations are methodologically consequential. Performance figures reported on OULAD are therefore not interchangeable unless the target definition, exclusion rules, and temporal cutoff design are also aligned. This is one reason later benchmark papers emphasize harmonized evaluation protocols rather than single summary scores (Silva et al., 10 Apr 2026, Le et al., 25 May 2026).

4. Feature engineering and temporal representations

OULAD’s most heavily used predictive signals combine assessment evidence with behavioral traces from the VLE. In one federated early-warning study, an early assessment window is defined as the first 90 days of a module, producing average_early_score and early_assessments_count; the same study derives engagement-volume features such as total_clicks and distinct_days_active, and activity-specific features including clicks_on_oucontent, clicks_on_quiz, clicks_on_forumng, and clicks_on_homepage, with missing values imputed with zero (Tertulino, 23 Aug 2025). That paper’s correlation analysis reports average_early_score as the strongest negative correlate of at_risk, with engagement measures also negatively associated with risk (Tertulino, 23 Aug 2025).

Other work makes temporality even more explicit. A heterogeneous-graph study engineers a dynamic Partial Grade feature at 13 time points—7%, 15%, 23%, 30%, 38%, 46%, 54%, 60%, 70%, 77%, 85%, 93%, and 100% of the semester—corresponding to 20 through 260 days since semester start (Muresan et al., 11 Jan 2026). A multivariate time-series study converts VLE logs into weekly sequences of shape $(n\_samples,\ n\_weeks,\ n\_activities)$, with observation windows ranging from 5 to 40 weeks and channels corresponding to activity types such as forum, quiz, and resource (Ayady et al., 17 Jul 2025).

Leakage-controlled work on OULAD has made representation order a first-class methodological issue. Under LEAP, early prediction at cutoff $t$ is defined through cutoff-first truncation,

$$\mathcal{R}_i^{(\le t)} = \{r \in \mathcal{R}_i : \tau(r)\le t\}, \qquad x_i^{(t)} = \phi(\mathcal{R}_i^{(\le t)}),$$

so that joins and aggregations must be applied only after temporal filtering (Le et al., 25 May 2026). The same protocol audits feature provenance by checking that the maximum timestamp touched in each feature group satisfies $T_{i,g}^{(t)} \le t$ (Le et al., 25 May 2026). This is a direct response to the literature’s concern that apparently “early” prediction can be inflated if future assessment information leaks into feature construction.

Survival-analysis work extends the temporal perspective further by modeling discrete-time weekly hazards and early-window summaries. In the dynamic weekly arm, OULAD enrollments are expanded into person-period panels with covariates such as total_clicks_week, n_vle_rows_week, active_this_week, n_distinct_sites_week, cum_clicks_until_t, recency, and streak; in the comparable arm, the first four weeks are collapsed into enrollment-level summaries such as clicks_first_4_weeks, active_weeks_first_4, and mean_clicks_first_4_weeks (Silva et al., 10 Apr 2026). Across these representations, the literature repeatedly converges on the value of temporal-behavioral variables.

5. Benchmark tasks and empirical findings

Empirical work using OULAD spans centralized and federated classification, heterogeneous graphs, multivariate time series, multiclass neural forecasting, survival analysis, leakage-controlled early warning, and intervention-policy learning. Reported scores are not directly comparable because labels, cutoffs, and cohort constructions differ, but they collectively show the dataset’s unusual breadth as a benchmark substrate.

Study focus	Modeling frame	Reported result
Federated at-risk prediction (Tertulino, 23 Aug 2025)	Fail vs Pass/Distinction, Withdrawn excluded	Centralized LR ROC AUC $\approx 0.8645$; Federated LR + SMOTE $\approx 0.8543$; Federated DNN $\approx 0.8430$
Heterogeneous graph success prediction (Muresan et al., 11 Jan 2026)	Pass/Distinction vs Fail/Withdrawn	HGT Case 4/5 reaches 68.6% validation F1 at 20 days and 89.5% near semester end
Weekly multivariate time series (Ayady et al., 17 Jul 2025)	Binary and multiclass failure-risk prediction on BBB, DDD, FFF	FCN reaches up to 94% F1 after 40 weeks on FFF; BBB remains much lower
Four-class early forecasting (Junejo et al., 2024)	Distinction / Fail / Pass / Withdrawn	92% accuracy at 20% of course completion; 98% accuracy with all features
Leakage-controlled early prediction (Le et al., 25 May 2026)	Weekly cutoff-based binary prediction	Best ROC-AUC rises from 0.7151 at day 7 to 0.8602 at day 56; RF is best earliest, GBDT thereafter
Temporal dropout survival benchmark (Silva et al., 10 Apr 2026)	Withdrawal event with censoring	RSF leads the comparable arm with IBS = 0.1219 and TD concordance = 0.6023; Poisson Piecewise-Exponential narrowly leads the dynamic arm on IBS at 0.1396
High-stakes intervention policy (Atkinson, 28 Jun 2026)	Five-action oracle policy learning	ONNX Decision Transformer reaches macro-F1 0.79, macro-recall 0.85, and 0% action flip rate

Across these studies, a recurrent substantive pattern is that temporal and behavioral variables often dominate static background variables. The survival benchmark states this most explicitly, concluding from ablation and explainability that the dominant predictive signal is temporal-behavioral rather than demographic or structural (Silva et al., 10 Apr 2026). The multivariate time-series study makes a closely related observation at the course level: richer and more diverse interaction logs, especially in the selected STEM courses, are associated with much better failure-risk prediction than sparser interaction structures in the selected SHS course (Ayady et al., 17 Jul 2025).

At the same time, the later decision-policy paper shows that OULAD can support tasks beyond outcome prediction. Using a 12-dimensional prefix-only Entity-Attribute-Value state, it evaluates whether AI systems can reproduce a conservative oracle intervention boundary and reports that zero-shot GPT-4o over-prescribes intervention by 43.6 percentage points at day 56, whereas supervised policy learners trained on oracle-labeled trajectories achieve near-zero calibration error (Atkinson, 28 Jun 2026). This extends OULAD’s role from predictive benchmarking to decision-quality evaluation.

6. Reproducibility, tooling, and instructional use

Because the raw dataset is multi-file and relational, a substantial line of work focuses on preprocessing standardization. The ouladFormat R package was introduced specifically to load, clean, and format OULAD into a single flat file or tibble with one row per student (Howard, 14 Jan 2025). It contains ten functions, including dataset_demographics(), dataset_registration(), dataset_assessment(), dataset_VLE_time(), dataset_VLE_activity(), convert_VLE(), VLE_learning_classification(), and the main combined_dataset() workflow (Howard, 14 Jan 2025). The package supports click transformations such as binary, standardised by variable, standardised globally, and logarithmic, and maps activity types to pedagogical classifications including FSLM, FSLSM, VARK, and OLS (Howard, 14 Jan 2025). Its stated purpose is to improve reproducibility, comparability, accessibility, and teaching utility.

OULAD is also used directly in educational data-science instruction. In one study, 72 postgraduate students organized into 12 groups used OULAD in a four-week group innovation project to formulate questions, preprocess data, analyze patterns, and interpret findings in learning analytics (Zahran et al., 22 Jan 2025). The paper reports that all groups followed a structured data-science methodology aligned with CRISP-DM and that their questions clustered into seven themes: Assessment and Attainment; Engagement and Attainment; Subject, Discipline, Performance, Engagement; Student Characteristics and Outcomes; Engagement Metrics; Temporal Factors; and Retention and Withdrawal (Zahran et al., 22 Jan 2025). This use case is notable because OULAD serves not merely as a benchmark corpus but as a shared design object for articulating what students themselves regard as useful learning analytics.

A related but non-dataset-specific framework is the LAVA model, “Learning Analytics and Visual Analytics,” which is explicitly described as not being about OULAD itself but as being applicable to OULAD-style course-level learner data (Chatti et al., 2023). LAVA expands the standard learning analytics cycle into eight stages—Learning Activities, Data Collection, Data Storage and Pre-processing, Analysis, Visualization, Perception and Knowledge, Exploration, and Action—and operationalizes indicator design through a Goal–Question–Indicator workflow (Chatti et al., 2023). A plausible implication is that OULAD’s repeated use in fixed predictive pipelines could be complemented by more interactive, user-centered indicator design and visual analytics workflows.

7. Methodological limits, controversies, and interpretive cautions

Despite its prominence, OULAD should not be treated as a fully self-explanatory benchmark. The large-scale survey that inventories it as dataset ID 109 does not provide a deep technical description of its internal schema, file-level structure, or exact variable list (Švábenský et al., 19 Feb 2026). That absence has practical consequences: papers often make distinct filtering, joining, and representation choices, and those differences can materially change the analytic sample and the reported performance (Howard, 14 Jan 2025).

Temporal leakage is the most explicit controversy surrounding early-warning use of OULAD. Under strict LEAP, the best day-7 ROC-AUC is about 0.7151 for Random Forest, but under intentionally leaky variants it rises dramatically: Random Forest reaches 0.9669 under Leaky-All and 0.9616 under Leaky-Assessment, while GBDT reaches 0.9704 under Leaky-All (Le et al., 25 May 2026). This is direct evidence that “early” predictive performance can be massively overstated if cutoff-first truncation is not enforced, especially through assessment information.

Generalizability is another recurrent boundary. The federated-learning study explicitly notes that it uses one dataset from one university and that its FL environment is simulated rather than production-deployed (Tertulino, 23 Aug 2025). The multivariate time-series study analyzes only three courses in detail and states that conclusions about STEM versus SHS are suggestive rather than definitive (Ayady et al., 17 Jul 2025). The survival benchmark similarly refuses to produce a single cross-arm ranking because its two harmonized arms are not directly comparable in representation, and its bootstrap intervals reflect sampling variability only, not refitting uncertainty (Silva et al., 10 Apr 2026).

Interpretation is also context-dependent even when the same dataset is shared. In the postgraduate co-design study, broad agreement existed on the direction of many relationships—such as higher engagement tending to associate with better performance—but groups still diverged because they chose different subsets, preprocessing decisions, and interpretive frames (Zahran et al., 22 Jan 2025). The paper explicitly characterizes many of these interpretations as contextual and speculative rather than deterministic (Zahran et al., 22 Jan 2025). This suggests that OULAD supports robust comparative work, but not context-free causal conclusions.

A further methodological caution concerns evaluation beyond predictive discrimination. In the high-stakes advisory study, LLM-as-judge scoring is reported as blind to intervention bias, rewarding fluent over-prescription rather than decision quality (Atkinson, 28 Jun 2026). In that setting, OULAD is used to show that a system can appear persuasive while being operationally miscalibrated. This widens the dataset’s significance: it is not only a benchmark for outcome prediction, but also a testbed for questions about calibration, policy fidelity, auditability, and the governance of educational AI.

Taken together, these strands of work indicate that OULAD’s importance lies less in any single canonical task than in a rare combination of openness, anonymization, structured course context, and reusable temporal interaction traces. The literature therefore supports viewing OULAD as one of the central public infrastructures through which contemporary learning analytics studies compare methods, test protocol design, and examine the boundary between predictive performance and deployable educational decision support (Švábenský et al., 19 Feb 2026).