Work

Predicting student return intervals in online English learning using a CNN LSTM attention model

Year: 2026
Type: article
Abstract: Predicting when students resume activity can support learning-engagement analysis in large-scale online English education. This study evaluates DropoutShield, a convolutional neural network (CNN), long short-term memory (LSTM), and temporal attention framework, for predicting early versus later observed return to question-answering activity in the EdNet dataset. The input is restricted to the first 14 days after each learner’s first observed KT1 question; the outcome is the first subsequent recorded KT1 return on days 15–44 versus days 45–194. Of 784,301 processed KT1 learners, 134,435 met this observed-return definition and were divided into student-disjoint training, validation, and test partitions of 94,104, 20,165, and 20,166 learners. The joint KT1/KT4 sequence model used 13 daily channels, while the KT1-only comparison used five channels. Three stride-two ceiling-mode pooling operations reduced the sequence from 14 to two time positions before the LSTM and attention modules. Across five random seeds, the joint model attained ROC-AUC $$0.783528\pm 0.002143$$ , compared with $$0.782243\pm 0.002321$$ for KT1-only. On seed 42, the paired ROC-AUC difference was $$ - 0.001856$$ (95% confidence interval $$[ - 0.005094,0.001131]$$ ). Thus, the additional KT4 channels did not establish a reliable advantage in that fitted-model comparison. Since learners without a classified recorded return were excluded and individual follow-up completion was not established, these results assess return-interval classification among observed returners, not verified dropout risk. The limited temporal resolution after pooling also restricts interpretation of attention as day-specific importance.
Language: English
Open Access status: gold
Estimated APC paid: $1,590
OpenAlex ID: W7235404917
FWCI:
Cites: 39
Cited by:
Related to: 10