Deep FM-Based Predictive Model for Student Dropout in Online Classes

Abstract

The student's high dropout rate is a severe issue in online learning courses, and as a result, it is creating concerns for academics and administrators in the field of education. A practical method of preventing dropouts is predicting students' likelihood of dropping out. This study aims to predict students' dropouts with two datasets, namely HarvardX Person-Course Academic Year 2013 De-Identified and MOOC datasets, using an explainable factorization machine and deep-learning approach. With the solvable approach, the aim is to enable the interpretation of the predictive models to produce actionable insights for related online educational interventions. This approach creates a DeepFM-based prediction model for student dropout, which entails multiple processes, including data preparation, feature engineering, model construction, training, assessment, and deployment. Moreover, the DeepFM design combines a factorization machine with DNN models to forecast student dropouts. It examines performance metrics, including recall, F1 score, accuracy, precision, and AUC-ROC. After ten iterations and 64 batches, the DeepFM model accurately predicted student dropout from online courses with a 99% accuracy rate on validation data. It also outperformed other techniques because of its capacity to capture complicated non-linear connections between features, combine dense and sparse information, and consider the unique properties of online learning. This study illustrated using an explainable factorization machine learning and DNN approach called DeepFM to interpret the underlying reasons for predicting students' dropout from online classes. Moreover, this approach has the potential to be extended to additional Massively open online courses (MOOC) datasets to assist educators and institutions in identifying at-risk students and providing targeted interventions to enhance their learning results.

A Study of Online Academic Risk Prediction Based on Multivariate Time Series Features

“…[16] uses three RNN-based models to handle time series features: basic RNN, GRU, and LSTM. Ultimately, the basic RNN achieves the best performance; [5] creates a DeepFM-based prediction model that can effectively simulate low-and high-order feature interactions without explicit feature engineering; [8] uses a two-channel CNN to automatically extract local feature information, then uses an attention mechanism to focus on important features and suppress unnecessary features, and finally uses a temporal convolutional network (TCN) to capture temporal features on different time scales.…”

“…from a large amount of personal information (e.g., place of origin, major, test scores). [5] selects 18 features that have strong relationships with dropping out based on prior knowledge and then uses these features to predict students in danger of dropping out. [6] performs nine rounds of feature extraction and feature generation to obtain new features, combines some of the features to form new features, eliminates features with high correlation to avoid redundancy, and finally combines all the generated new features as the prediction features.…”

Predicting Factors that Affect East Asian Students’ Reading Proficiency in PISA

“…It outperforms CNN when it is applied to KDD Cup 2015 dataset. DeepFM [25] is DNN and factorization machine hybrid, achieving 99% in validation data. In work by [26] multiple linear regression (MLR), multilayer perceptron (MLP) and classification and regression tree (CART) are compared with MLP and CART performing better than MLR.…”