Personalizing Education for 50,000+ StudentsPersonalizing Education for 50,000+ Students

Pearson's online learning platform serves 50,000+ K-12 students across 120 school districts. Despite having a library of 15,000 learning modules (videos, interactive exercises, quizzes, and reading materials), every student in a given grade received the same content sequence. Teachers could manually adjust assignments for individual students, but with class sizes of 25-35, meaningful personalization was impossible. The result was predictable: advanced students were bored, struggling students fell further behind, and engagement metrics showed a steady decline - average session time had dropped 18% year-over-year, and course completion rates sat at 61%.

Pearson had previously tried two approaches to personalization. The first was a simple branching logic system where quiz scores above or below certain thresholds would route students to harder or easier content. This was too coarse - it treated all wrong answers the same and couldn't distinguish between a student who didn't understand the concept and one who made a careless arithmetic mistake. The second attempt was an off-the-shelf adaptive learning plugin from an ed-tech vendor, but it required content to be tagged in a proprietary format, couldn't integrate with Pearson's existing content management system, and only supported multiple-choice assessments. After 6 months of integration work, it was used by fewer than 5% of teachers and was quietly deprecated.

We started by working with Pearson's curriculum team and a panel of 12 teachers to build a knowledge graph mapping the prerequisite relationships between 2,200 discrete learning objectives across math, science, and ELA for grades 3-8. This graph became the backbone of the adaptive engine - rather than just adjusting difficulty, the system could identify specific knowledge gaps and route students to the exact prerequisite content they needed.

For the student modeling layer, we evaluated several approaches: Bayesian Knowledge Tracing (BKT), Deep Knowledge Tracing (DKT) using LSTMs, and a custom model we called Attentive Knowledge Tracing (AKT) that used transformer-based attention over a student's interaction history. After testing on 18 months of historical student interaction data (approximately 42 million learning events), the AKT model outperformed BKT by 23% and DKT by 11% on predicting next-question correctness (AUC 0.847). The key advantage was its ability to model forgetting curves and capture long-range dependencies - for example, recognizing that a student who struggled with fractions in October might need a brief refresher before tackling ratios in February.

The recommendation engine uses a multi-armed bandit approach (Thompson Sampling) to balance exploitation (serving content predicted to be at the student's optimal difficulty) with exploration (occasionally testing harder or different-format content to refine the student model). We tuned the exploration rate through A/B testing with a cohort of 3,000 students over 4 weeks, finding that a 15% exploration rate maximized both learning gains and engagement.

The adaptive engine is deployed as a set of microservices on Google Cloud Platform. The knowledge graph is stored in Neo4j and queried via a GraphQL API. The AKT student model runs on Cloud Run with GPU support, updating each student's mastery estimates in real-time as they interact with content. The recommendation service (Python, FastAPI) pulls from the student model and knowledge graph to generate a ranked list of next-best learning activities, which is served to the React frontend via a REST API. A teacher dashboard built with Next.js provides class-level and individual student views showing mastery heatmaps across the knowledge graph, predicted areas of difficulty, and suggested intervention points. All student data is encrypted and the system is FERPA and COPPA compliant. The content tagging system uses a semi-automated pipeline: a fine-tuned BERT model tags new content with learning objectives at 91% accuracy, with curriculum specialists reviewing and correcting tags before content enters the adaptive pool.