When it comes to memorising factual knowledge, there are systematic differences in learner ability that mean some learners memorise facts more easily than others. Similarly, some facts are systematically easier to learn than others. An adaptive learning system that detects and adjusts to these individual differences in ability and difficulty can improve learning efficacy by enabling learners to study at an appropriately challenging level. Such a system faces the cold start problem whenever it has not yet had the opportunity to adapt to its user or to the content. Having more accurate initial predictions of ability and difficulty can alleviate the problem.

Using authentic learning data from 140 thousand students, we evaluate several methods for alleviating the cold start problem in the adaptive fact learning system MemoryLab. This system adaptively estimates a separate rate-of-forgetting parameter for each fact that a learner practices, by interpreting the accuracy and latency of the learner’s responses to retrieval prompts through a computational cognitive model. We show that data-driven prediction of the system’s rate-of-forgetting parameter leads to more accurate estimates of learning at the start of a practice session, particularly when that prediction is based on fact-specific difficulty information. The observed improvements are similar in magnitude to those found in an earlier lab study, where using the predicted rate-of-forgetting values as starting estimates in a MemoryLab practice session significantly increased retention on a posttest. Based on the current large-scale evaluation, we expect that comparable retention gains can be achieved in real-world educational practice.