How Deep Recommendation Models Are Improving Learner Engagement on an EdTech Platform

The Challenge

Recommendations on the platform came from editor-curated lists and simple "most popular in this category" rules, the same for every learner.

The existing approach made it difficult to:

• Suggest content that matched what a learner had already mastered or was struggling with.
• Bring new learners to relevant content before they had much history.
• Surface newly published courses that had no popularity data yet.
• Keep recommendations current as the catalog changed every week.
• Keep learners moving through a sequence instead of dropping off after one lesson.
• Measure whether a recommendation actually helped learning, not just clicks.

From a technical perspective, learner signals were spread across video events, quiz results, and enrollment records in separate systems. The catalog changed weekly, and recommendations had to load in the web and mobile apps within the page load budget. The system also had to respect learner privacy and data retention rules.

The Solution

InnovationM developed a two-stage deep recommendation system that first retrieves a short list of relevant content for each learner, then ranks it by likely learning value and engagement. The solution included:

• Unified learning event pipeline: Streaming video, quiz, and enrollment events into a single learner timeline used for training and real-time features, built through data engineering.
• Two-tower retrieval model: Training learner and content embeddings so the system can pull relevant candidates from the full catalog in milliseconds.
• Sequence-aware ranking: Using a transformer-based ranker that reads a learner's recent activity in order and predicts which next item they are most likely to start and complete.
• Skill-level signals: Estimating mastery by topic from quiz performance, so the ranker avoids content that is too easy or too advanced.
• Cold-start handling: Building content embeddings from lesson text, transcripts, and metadata so new courses and new learners get sensible suggestions from day one.
• Low-latency serving: Indexing embeddings in a vector store and serving the ranker behind a cached API, integrated into the web and mobile home screens.
• Experimentation framework: A/B testing against the existing rules and measuring completion and return visits alongside click-through.

The key focus was not simply increasing clicks on recommended content, but creating a steady path for each learner toward the goal they came to the platform for.

The Impact

The implementation helped the platform move from the same curated lists for every learner toward personalized next-step recommendations. Suggestions now adapt to each learner's progress and update as they study. The solution supported:

• Higher engagement with recommended lessons.
• Improved course completion.
• More learners returning in the weeks after sign-up.
• Faster discovery of newly published courses.
• Recommendations served within the app's page load budget.
• A foundation for testing every future model change with confidence.

By bringing together unified learner data, deep retrieval and ranking models, predictive AI, and disciplined experimentation, InnovationM showed how deep learning development services can turn a large catalog into a guided learning path for every learner.