# Effective Feature Learning with Unsupervised Learning for Improving the Predictive Models in Massive Open Online Courses

**Authors:** Mucong Ding, Kai Yang, Dit-Yan Yeung, Ting-Chuen Pong  
**arXiv:** [1812.05044](https://arxiv.org/abs/1812.05044)  
**Date:** 2018-12-12  
**Summary page:** https://www.summarizepaper.com/en/paper/1812.05044v2/

## Key points

- Personalized intervention schemes can improve MOOCs' effectiveness
- Predictive models are key to achieving this improvement
- Handcrafted features are a challenge to design for this task
- The authors propose an unsupervised learning approach using a modified auto-encoder and LSTM network
- This approach captures underlying learning patterns in the content domain and temporal nature of clickstream data
- The new features obtained by the modified LSTM-AE are more parsimonious and discriminative than original features
- These learned features improve prediction accuracy by up to 17% compared with supervised neural networks and reduce overfitting to low-performing students.
- Unsupervised feature learning can enhance predictive modeling in MOOCs and other educational settings where large amounts of data are available.

## Summary

The effectiveness of learning in massive open online courses (MOOCs) can be significantly improved by introducing personalized intervention schemes that rely on building predictive models. To address the challenge of designing handcrafted features for this task, authors Mucong Ding, Kai Yang, Dit-Yan Yeung and Ting-Chuen Pong propose an unsupervised learning approach. They train a modified auto-encoder (AE) combined with the long short-term memory (LSTM) network to obtain a fixed-length embedding for each input sequence. This captures the underlying learning patterns in the content domain and temporal nature of clickstream data. The new features obtained by the modified LSTM-AE are more parsimonious and discriminative for predicting students' performance than original features. Using simple supervised learning models, these learned features improve prediction accuracy by up to 17% compared with supervised neural networks and reduce overfitting to low-performing students. This approach is generic and not restricted to specific supervised learning models or prediction tasks for MOOC learning analytics. The authors' findings suggest that unsupervised feature learning can enhance predictive modeling in MOOCs and other educational settings where large amounts of data are available.