Machine Learning for Online Education Platforms

Completion is the number that matters

An online training provider selling professional courses to 20,000 learners a year typically sees a familiar curve: strong activity in week one, a sharp fall by week three, and a completion rate that nobody puts in the marketing. For subscription platforms, those learners cancel. For employer-funded training, the employer notices at renewal time.

Machine learning for online learning platforms is most useful here, because the platform already records exactly the behaviour that predicts dropping out: logins, video watch time, quiz attempts, forum activity and the gap since the last session. A model can spot the drift early enough for a nudge, a tutor message or a simpler route through the material to make a difference.

Where machine learning fits in edtech

• Dropout and churn prediction. Scoring learners weekly by their risk of disengaging, so tutors and automated nudges reach the right people.
• Course and content recommendations. Suggesting the next module or course based on what similar learners took and completed, rather than what they clicked.
• Adaptive practice. Choosing questions at the right difficulty from a learner's recent answers, which keeps practice useful without being demoralising.
• Marking assistance. Grouping similar short answers so tutors can mark them together, or drafting feedback for a tutor to edit.
• Content quality signals. Finding lessons where learners consistently rewind, fail the following quiz or abandon the course.
• Demand forecasting. Predicting enrolments by course and intake, which helps plan tutor capacity.

Dropout prediction: what to do with the score

A risk score on its own changes nothing. The design question is what happens next, and the answer has to be something the provider can sustain. A tutor calling every at-risk learner works for a cohort of 200; it does not work for 20,000.

1. Score learners weekly on recent engagement compared with learners who completed
2. Send the lowest-risk tier nothing, the middle tier an automated but specific nudge
3. Route the highest-risk tier to a tutor with a one-line reason, such as missed two consecutive assessments
4. Measure completion against a held-back group that received no intervention
5. Retire interventions that do not move completion, however sensible they seemed

Step four is the one most platforms skip. Without a comparison group you cannot tell whether the nudges helped or whether the learners who responded would have finished anyway.

The EU AI Act draws clear lines in education

Education is one of the areas the EU AI Act singles out. Systems used to decide admission, evaluate learning outcomes in ways that steer a learner's education, assess the appropriate level of education for someone, or monitor learners for prohibited behaviour during tests are classed as high-risk. Those carry obligations on risk management, data quality, documentation, human oversight and logging, which are phasing in. Emotion recognition in educational settings is prohibited outright, with narrow exceptions.

If you serve learners in the EU, that matters for design decisions today. A recommendation engine suggesting the next optional course is a very different proposition from a model that decides whether someone passes.

Why automated proctoring is usually a bad idea

Webcam-based cheating detection has a poor track record. It flags learners with unreliable internet, different lighting, disabilities or simply unusual behaviour, and it creates stress and complaints out of proportion to the cheating it catches. We generally advise against building it.

Better options are assessment design that is hard to cheat (open-book, applied, varied question banks), plagiarism checks with human review, and viva-style follow-ups for high-stakes results. They are less impressive in a sales deck and far more defensible.

If a model's mistake could fail someone's qualification, a person needs to make the decision and be able to explain it.

Recommendations that improve learning instead of clicks

Generic recommendation engines optimise for engagement, which in education can mean steering learners towards short, entertaining content rather than the module they need. Train the model on completion and progression, not clicks, and add simple rules for prerequisites so nobody is recommended an advanced course before the foundation one.

Our post on building a content recommendation engine covers the mechanics, and our wider piece on AI for education and edtech covers tutoring assistants and content generation.

Measure learning, not only engagement

It is easy for an edtech team to end up optimising time on platform because it is the number the analytics tool shows first. Time on platform is a poor proxy for learning. A learner who is stuck on the same lesson for forty minutes looks highly engaged.

Better targets are completion, assessment improvement over time, progression to the next course and, for professional training, whether the learner passed the external exam or reported using the skill at work. These outcomes arrive slowly and are noisier, which is exactly why they need deliberate collection from the start.

• Record assessment attempts and scores with timestamps, not just the final result
• Ask a short outcome question a few months after completion
• Link employer-funded learners to renewal outcomes where the contract allows
• Track whether recommended content was completed, since opening it proves little

How we would start with an edtech platform

At SpiderHunts we first check the event data. Many learning platforms record page views but not quiz attempts, or store progress as a single percentage that overwrites history. Fixing the tracking is often the first month of the project, and it pays off even if no model is ever built.

Then we build a simple dropout model, run it silently for a term, compare its warnings with who actually left and only then connect it to nudges. The modelling and product integration sit within our machine learning development work.

Frequently asked questions