Evolving Memory-Based Intelligent Systems For Long-Term Knowledge Retention In Personalized Education Applications

Abstract

The intelligent tutoring systems of today make use of knowledge tracing models that aim at estimating the understanding and competency levels of a student and suggest personalized learning activities accordingly; but majority of the knowledge tracing models have adopted a single static knowledge model that updates the knowledge based on every interaction, without any decision-making process on what knowledge to retain, consolidate, or forget for long periods of time. Consequently, the prediction made on how much a student will be able to recall his knowledge after several days or weeks becomes weak. This paper presents an Evolving Memory Knowledge Tracing (EMKT) paradigm which adds an Ebbinghaus inspired forgetting-decay component and a retention agent to a dynamic memory network, which helps the system in retaining the important knowledge slots and discarding or replacing irrelevant or low utility knowledge. The model is benchmarked using a dataset simulated based on the ASSISTments tutoring interaction logs for a 90-day period, where retention is specifically measured at 1-day, 7-day, 30-day, and 90-day intervals post-concept usage. In comparison to a Deep Knowledge Tracing-based model and a static Dynamic Key-Value Memory Network-based model, the novel EMKT model retains an AUC score of 0.80 for retention prediction at 90 days, as compared to 0.59 and 0.65 for the two models, respectively, while performing similar at the 1-day mark. This clearly shows that evolution of memory as opposed to mere adaptation of memory helps in better retention prediction in personalized education systems.