Related Experiment Videos
Robust and interpretable tabular data classification via multi-channel image conversion and channel-wise gating.
Zengshuai Wang1, Minhua Zheng2, Peter Xiaoping Liu3
1School of Mechanical, Electronic and Control Engineering, Beijing Jiaotong University, Beijing, 100044, China.
Summary
Deep learning on tabular data is difficult due to varied features and missing values. TabGCNet, a CNN framework, converts tabular data into multi-channel images for efficient, interpretable classification.
Area of Science:
- Computer Science
- Machine Learning
- Artificial Intelligence
Background:
- Deep learning on tabular data faces challenges with heterogeneous features, missing values, and small sample sizes.
- Existing table-to-image methods can obscure feature identity and introduce bias due to single-plane mapping or complex layout designs.
Purpose of the Study:
- To propose TabGCNet, a novel CNN-based framework for heterogeneous tabular classification.
- To enhance feature representation and interpretability in tabular data using a multi-channel image approach.
Main Methods:
- TabGCNet utilizes a parameter-free Tabular-to-Multi-channel Image (Tab2MI) module to render each feature into a dedicated image channel.
- A lightweight Gated shallow Convolutional Network (GCNet) with channel-wise gating is employed for efficient prediction and feature-level importance scoring.
Main Results:
- TabGCNet achieved competitive classification performance across ten public tabular datasets.
- The framework demonstrated robustness under label-preserving corruption and provided feature rankings consistent with predictive utility.
- Experiments confirmed the efficiency advantages of the compact convolutional architecture.
Conclusions:
- TabGCNet offers a practical and auditable solution for heterogeneous tabular classification.
- The framework is particularly suitable for low-latency or safety-critical applications requiring efficient and interpretable deep learning models.
