Robust and interpretable tabular data classification via multi-channel image conversion and channel-wise gating

Deep learning for tabular data remains challenging because heterogeneous feature types, missing values, and limited sample sizes complicate learning representations that are both expressive and reliable. Existing table-to-image approaches enable convolutional modeling, but many either require costly spatial layout design or map all features into a single image plane, which can obscure feature identity, introduce artificial spatial bias, and limit feature-level interpretability. To address this limitation, we propose TabGCNet, a CNN-based framework that consists of a parameter-free Tabular-to-Multi-channel Image (Tab2MI) module with a lightweight Gated shallow Convolutional Network (GCNet). Tab2MI renders each feature value, including numerical, categorical, and missing entries, into a dedicated image channel, thereby preserving explicit feature-to-channel correspondence without auxiliary encoding or layout search. Built on this representation, GCNet introduces a learnable channel-wise gating layer before shallow convolutions, allowing channel reweighting while producing intrinsic column-level importance scores during forward inference. Together, these components enable efficient prediction with intrinsic feature-level interpretability. Experiments on ten public tabular datasets show that TabGCNet achieves competitive overall classification performance, remains robust under label-preserving corruption, and yields feature rankings that are broadly consistent with predictive utility, while retaining the efficiency advantages of compact convolutional architectures. These results indicate that TabGCNet is a practical and auditable framework for heterogeneous tabular classification, particularly in low-latency or safety-critical scenarios.