Behind the paper: EduBridge Analytica

In June 2026, a study I co-authored with Henry Chukwudi John was published in the Babcock University Journal of Education: “EduBridge Analytica: Bridging Theory and Practice in Educational Data Mining for Action Research.” The headline is easy to say out loud: with our system-supported workflow, educators finished feature selection in 45 seconds on average instead of 342 — about 7.6× faster — while choosing 85–90% of the same features they had picked by hand. This page is the long version: what we built, how we tested it, what the numbers mean, and — just as important — what they don't.

The gap the paper aims at

Educational data mining (EDM) develops computational techniques for discovering patterns in educational data; learning analytics (LA) focuses on interpreting those patterns in context to improve teaching. Both fields are mature, and learning platforms generate more behavioural data than anyone can read. Yet the group with the most direct questions — educators running action research on their own classrooms — barely benefits. Recent reviews of the field (Kalita et al., 2025; Kurday & Vladova, 2025) keep finding the same imbalance: increasingly sophisticated algorithms on one side, and very little evidence about how educators actually use these tools on the other.

The pain concentrates in the unglamorous early stages. Before any dashboard or model, someone has to prepare the data and decide which of the many available variables matter — feature selection. That step assumes exactly the technical background most educators don't have, and the literature flags it as a persistent bottleneck for non-technical users of educational data (Fadillah et al., 2025). Fully automated selection exists, but it trades away transparency — and an action researcher who can't see why a variable was chosen can't defend the conclusion built on it.

The research question: how does a system-supported EDM–learning-analytics workflow, implemented in the EduBridge Analytica prototype, support educators to prepare data, select relevant features, and interpret analytic outputs to inform action research?

First a framework, then a prototype

The project started in 2023 as a literature review at African Leadership University and became a framework before it became software. The framework does one simple thing: it organises the analytical features an educator might use along two axes — the perspective they describe (student, course, content, instructor) and the source they come from (profiles, performance, interactions). It sounds almost too modest to publish, but that grid is exactly what the fuzzy early stage lacks: a structure you can navigate instead of a blank page you must invent.

EduBridge Analytica is that framework operationalised. It began as low- and mid-fidelity Figma wireframes and became a web application that walks an educator through the whole path: sign-in and a profile dashboard; dataset upload with a quality preview — column types, unique values and missing values, visible before any analysis; mapping the uploaded columns onto the standardised EduBridge schema; supported feature selection; and dashboards for performance, engagement and course-content insight. One design rule governed every screen: the system suggests, previews and accelerates — it never decides.

The study: what we actually did

The evaluation was mixed-methods and deliberately small. Two educators, purposively selected for different levels of experience with technology-enhanced learning, each answered the same three action-research questions twice — first manually, then with the prototype. Because no institutional dataset was available to us, we built a synthetic post-secondary dataset, informed by data from partner universities and designed to mirror the dimensionality of the real thing. It spans six categories:

Three measures carried the analysis: completion time (efficiency, by direct comparison), feature overlap between the manual and assisted selections (consistency), and the interview themes (experience).

Results, precisely

Manual selection took 342 seconds on average; the system-supported workflow took 45 — an 87% reduction. Overlap between what participants chose by hand and what they chose with the system stayed high on every question: 90% on the first, 85% on the second, 88% on the third. That pairing is the finding. Speed alone would be easy — any tool that picks for you is fast. Speed with high overlap says the system compressed the mechanics while leaving the judgement where it was.

The interviews pointed the same way. Both educators independently described the assisted workflow as easier and less mentally demanding than the manual process, and both said the dashboards helped them connect selected features back to their research questions. Both also asked for more flexibility when moving between different kinds of questions — which the paper reports as a limitation, not a footnote.

Why this matters — and what's actually new

Strip the terminology away and the situation is this: teachers sit on piles of data about their students — clicks, grades, attendance, submissions. Before any of it can teach them anything, they have to answer one unglamorous question: out of all these columns, which ones actually matter for what I'm asking? That step is slow, intimidating, and it's where most non-technical educators give up. Everything downstream — the dashboards, the models, the insight — never happens, because the first step never finishes.

The study shows two things at once, and the pairing is the point. First, the guided workflow made that step fast: what took nearly six minutes of staring and second-guessing took 45 seconds. Second, it didn't change the answers: 85–90% of what the educators chose with the system is what they would have chosen on their own, just slower.

That combination is also where the novelty sits. The empirical literature keeps measuring the glamorous end of the pipeline — model accuracy, prediction, visualisation — while the efficiency of data preparation and feature selection, the stage that actually decides whether an educator gets anywhere, is seldom measured at all. And the studies that do automate it tend to find that opaque automation lowers educator confidence rather than raising it. Our contribution is a direct, timed measurement of that neglected early stage, with real educators in the loop, showing you can compress it by an order of magnitude without taking the decision away from the person who has to stand behind it.

The practical payoff is easy to picture. A teacher with no data-science background wonders whether early engagement predicts who will struggle in her course. Yesterday that question meant a spreadsheet with forty columns and an afternoon she doesn't have. With this kind of workflow it means a few guided minutes — and every choice along the way is still hers to see, question, and defend.

What the paper doesn't claim

Future work, as the paper frames it: more participants, varied levels of system guidance, and a closer look at the flexibility educators asked for across different question types.

Where it sits in the literature

The result lines up with a pattern in recent empirical work. In a randomised experiment with more than 8,000 MOOC participants, Borrella and Ponce-Cueto (2025) found that dashboards supporting interpretation improved learning outcomes, while purely descriptive visualisation added cognitive load for negligible gain. Wang et al. (2025) reached a similar place with teacher-facing self-service analytics: structured workflows reduce technical difficulty, but unaligned schemas and unguided feature relevance still confuse users. And Lampropoulos and Evangelidis (2025) report that institutional adoption tracks whether a system helps people make sense of data in their own instructional context.

The through-line — and the paper's overarching argument — is that practical impact depends less on analytical sophistication than on how well the analytics are embedded in usable, interpretable workflows. Our study is one small, concrete data point for that claim at the feature-selection stage.

About the journal

The Babcock University Journal of Education (BUJED) is published by the Department of Education and Humanities at Babcock University, Ilishan-Remo, Ogun State, Nigeria. It publishes two issues a year — May/June and November/December — runs double-blind peer review, and is fully open access: every article is free to read under a Creative Commons Attribution 4.0 licence, receives a DOI minted through Zenodo (which is why ours resolves there), and is indexed in Google Scholar, OpenAIRE, Zenodo, Dimensions and Garuda.

Our paper sits in Volume 11, Number 1 (May/June 2026), pages 223–239, published on 1 June 2026. The full text is free to read wherever suits you: via the DOI at doi.org/10.5281/zenodo.20712838, on the journal's article page, or on ResearchGate.

Who did what

Author contributions, as declared in the paper: I handled the conceptual design, the integrated framework, the prototype's logic and platform, and the quantitative analysis. Henry Chukwudi John — my co-author at African Leadership University and the paper's corresponding author — led the study design, the literature review, the methodological structuring, the qualitative analysis, and the drafting and revision of the manuscript. Isaac Museveni assisted with the critical evaluation of the text. We're indebted to African Leadership University for supporting the research, to the partner universities whose data informed the synthetic dataset, and to the two educators who gave us their time and candour.

The part no one shows you is the distance between “it works” and “it's published.” The first reading list dates to 2023; the issue is dated May/June 2026. Review cycles, revisions, waiting — Henry carried more of that distance than anyone.

What stays with me isn't mainly the number — it's the stance. Most systems I build now, in research or industry, follow the same rule the prototype did: let the machine do the tedious part, make its work inspectable, and keep the human in charge of the judgement. The prototype and the numbers in context are also on the work page; if you want the primary source, the paper is open access and seventeen pages.