Abstract
Student attrition is a critical challenge for higher education, with roughly one-third of students in the US failing to complete their degrees. This paper introduces MSNF-MTCL, a novel neural network architecture that fuses structured academic data with unstructured advising notes to predict five distinct retention risks. By leveraging an 18-year longitudinal dataset and a cascaded multi-task learning approach, the model achieves over 98% accuracy in dropout prediction. Results demonstrate that integrating textual narratives from advisors significantly enhances predictive performance, particularly as more historical data becomes available.
Introduction
College dropout rates represent a global crisis with significant economic and social consequences. While universities collect vast amounts of data, most existing predictive models are "siloed," focusing either on static/temporal structured data (such as GPA and attendance) or unstructured text (such as advising notes). These models often fail to capture the spatiotemporal connection between a student's academic performance and the personal life events documented in text.
Furthermore, traditional models typically provide a binary prediction of whether a student will leave. This research expands the scope to five distinct risks: future dropout, imminent (next semester) dropout, the type of dropout (temporary vs. permanent), stay duration, and the specific cause of attrition (e.g., financial, health, or academic issues).
Methodology: MSNF-MTCL Architecture
The authors propose a two-part architecture designed to handle the complexity of multimodal educational data.
1. Multimodal Spatiotemporal Neural Fusion (MSNF)
This layer acts as a feature extractor that fuses three data modalities:
- Unstructured Text: Employs a BERT-based framework to embed advising notes, followed by a Bidirectional Long Short-Term Memory (BiLSTM) network to capture chronological sequences.
- Temporal Structured Data: Uses an LSTM to model trajectories of performance variables like GPA and credits over time.
- Static Demographics: Utilizes a 1D CNN to process fixed information such as gender and family income.
2. Multi-Task Cascade Learning (MTCL)
Rather than predicting risks in parallel, the model uses a cascaded hierarchy. Fundamental tasks (e.g., predicting if a dropout will occur) are placed at earlier layers. The features learned here flow upward to assist in solving more complex tasks, such as identifying the specific cause among 15 different categories.
Experimental Results
The study utilized a massive longitudinal dataset from a private university, spanning 18 years and including records for 36,445 students. The full MSNF-MTCL model (V3) achieved a peak 98.78% accuracy for predicting future dropouts, significantly outperforming baseline models that relied on simpler text embeddings or single data modalities.
The Power of Narrative History
A key finding of the study is that the model's accuracy improves significantly as the number of available advising notes increases. While GPA indicates that a student is struggling, advising notes provide the "why" (e.g., family bereavement or financial hardship). Fusing these provides a 360-degree view of the student's narrative over time.
Fairness and Ethics
Recognizing that educational data can harbor inherent biases, the authors integrated the AIF360 fairness toolkit. By employing "Prejudice Remover" techniques, the researchers ensured the model does not unfairly flag students based on protected attributes like gender or socioeconomic status.
Conclusion and Future Work
MSNF-MTCL demonstrates that the "hidden" information in unstructured advising notes is vital for accurate retention risk analysis. While the model showed exceptional performance in identifying dropouts due to financial or mental health issues, it faced challenges with unpredictable events like family deaths. Future work includes developing Educational BERT (EBERT), a model pre-trained specifically on academic advising notes to better capture the nuances of student-advisor communication.
Summary updated