EdNet: A Large-Scale Hierarchical Dataset in Education

Source

Core Claim

EdNet provides large-scale sequences of student activities from an AI tutoring/self-study platform.

Action-Time-Series Notes

  • The time-series unit is student activity over time across question solving, lecture consumption, and purchases.
  • Actions can be interpreted as student/tutor platform events rather than clean control inputs.
  • It is useful for student-state dynamics and learning-path recommendation.

Foundation TSFM Relevance

Agenda slotVerdictEvidenceMissing pieces
Latent-state predictionpartially closesThe paper frames EdNet around knowledge tracing over 131M interactions from 784K students.The latent state is education-specific and not a general numeric system-state benchmark.
Time representation and irregular event streamspartially closesEdNet-KT1 through KT4 add timestamps, elapsed time, question interactions, lectures, purchases, and other ordered platform actions.Needs a canonical event schema for TSFM ingestion and missingness handling.
Causal structure, counterfactuals, and controladjacentThe paper discusses RL-based learning-path recommendation and student simulators using action/state notation.Logged events are not enough for policy-quality counterfactuals without intervention design.