A Machine Learning Based Early Detection and Intervention System for At-Risk Students
Open Access Depositedhowever, even the most modern of studies have not provided educators with insights on how to proceed once students are identified by initial classification models. This praxis presents a hybrid system that classifies students at-risk of dropping out and uses student meta-data to recommend intervention plans for positively classified at-risk students. This approach aims to save educators time by offering evidence based strategies for intervention plans. The hybrid system is composed of two models a classification model, and a language model with context provided by a Retrieval Augmented Generation (RAG) architecture. There were 5 classification models compared for use within this study, Logistic Regression (LR), Decision Trees (DT), Stacking Ensemble (SE), XGBoost (XGB) and Random Forest (RF). Models were trained on 4 years of student data from a single district in the State of Michigan. The model that performed the best consistently was the Logistic Regression model trained on the ADASYN sampled dataset, achieving an accuracy of 92%. Features that contributed the most towards the prediction of a dropout student were also analyzed leveraging SHAP these features included the current grade, gender and whether they are a Section 31A student. The collection of resources for the RAG architecture was composed of documents from four sources including strategies from the Michigan Strategy Bank (MSB), What Works Clearinghouse (WWC), the National Dropout Prevention Center (NDPC) and the Educational Resources Information Center (ERIC). Chunking strategies for the RAG were also compared to determining which strategy yielded the most relevant results. This was measured by calculating the similarity between the documents retrieved, the prompt passed to the LLM and the full intervention model response. Documents were chunked by token length and by paragraph. The chunking strategy that yielded the most relevant results was chunking by token length.
Students who dropout out of high school face lifelong consequences that affect both themselves and society. To reduce the number of dropout students, students need to be identified early and then interventions must be implemented in a timely manner. There has been significant advancement using machine learning to classify at-risk students over the last decade
- All rights reserved
Notice to Authors
If you are the author of this work and you have any questions about the information on this page, please use the Contact form to get in touch with us.