Deep Learning on Big, Sparse, Behavioral Data.

Area of Science:

• Machine Learning
• Behavioral Data Analysis
• Computational Science

Background:

• Deep learning (DL) shows promise beyond computer vision and natural language processing.
• Large, sparse behavioral datasets are increasingly common in big data research.
• Logistic regression (LR) has been a benchmark for sparse data classification.

Purpose of the Study:

• Investigate deep learning (DL) for classifying large, sparse behavioral data.
• Compare DL performance against regularized logistic regression (LR) and other shallow methods.
• Identify conditions and reasons for DL's effectiveness in behavioral data analysis.

Main Methods:

• Extensive search of DL architecture variants.
• Comparative analysis of DL versus regularized LR on 39 behavioral classification tasks.
• Evaluation of DL feature extraction and interpretability methods.

Main Results:

• Deep learning (DL) demonstrated significant performance improvements over logistic regression (LR) for sparse behavioral data classification.
• DL often outperformed other shallow techniques, though the gain might not always justify the computational cost.
• DL's effectiveness is linked to data characteristics, with better performance on higher signal-to-noise ratio datasets.

Conclusions:

• Deep learning (DL) offers a powerful approach for analyzing large, sparse behavioral datasets, surpassing traditional methods like logistic regression (LR).
• Understanding DL's feature learning and interpretability, particularly at the initial layers, is crucial for its application.
• Future research should explore advanced interpretability techniques like instance-level counterfactual explanations for DL models.