Towards Temporality-Sensitive Recurrent Neural Networks through Enriched Traces

Thomas Sergent (1, 2, 3) , François Bouchet (2, 1) , Thibault Carron (1, 2)

1 MOCAH - Modèles et Outils en ingénierie des Connaissances pour l'Apprentissage Humain
2 SU - Sorbonne Université
3 Lalilo, Paris

Educational traces are distinctive compared to the usual data a recurrent neural network encounters: there is a difference between two consecutive educational traces generated by a same learner if they are separated by 2 minutes or 2 months. Indeed, in the latter case, the learner who generated the trace may have forgotten the associated skill, which is less likely in the former case. Recurrent Neural Networks have seen a surge of popularity in the recent few years thanks to Deep Knowledge Tracing. While the focus has mostly been on the network architecture, we propose here a novel framework where traces are enriched with information relative to the temporality before they are used to train the network, and assess the performance on two datasets (Lalilo and ASSISTments 2012), which is not improved by this approach .

Mots clés: Educational Data Mining; Neural Networks; Trace enrich- ment; Temporality

Domaines: Environnements Informatiques pour l'Apprentissage Humain

Thomas Sergent, François Bouchet, Thibault Carron. Towards Temporality-Sensitive Recurrent Neural Networks through Enriched Traces. International Conference on Educational Data Mining (EDM 2020), IEDMS, Jul 2020, Ifrane, Morocco. pp.658-661. ⟨hal-03005331⟩