A Novel 2D-CNN Framework for Tabular Data Classification Based on Multi-Channel Fuzzy Clustering
Converting non-visual numerical matrices into artificial images has emerged as a major paradigm shift, enabling deep networks to leverage spatial inductive biases. However, single-layer transformations often fail to capture complex non-linear transitions, localized spatial associations, and multi-layered data structures. To overcome these challenges, this study introduces a novel, order-invariant tabular-to-image mapping framework that systematically constructs information-dense 4D artificial image tensors. The proposed architecture dynamically integrates normalized real features with cell-based Fuzzy C-Means memberships (m=2.0), Gustafson-Kessel elliptic variance partitions (m=1.5), and column-wise Possibilistic Fuzzy C-Means typicality indicators. Evaluated over ten diversified UCI benchmark datasets using a rigorous 10-fold stratified cross-validation pipeline, the framework successfully achieves a global Mean Rank of 3.50. Strikingly, despite being executed over a highly restricted, lightweight hyperparameter search grid via Random Search, no statistically significant difference was detected between the proposed model and top-tier, extensively optimized baselines like Support Vector Machines and Multi-Layer Perceptrons. Friedman test (χ^2=42.961, P
Authors
- Erdinç Karakullukçu (ORCID: https://orcid.org/0000-0001-8805-1694)
Institutions
- Karadeniz Technical University (TR)
Publication Details
- Journal
- Black Sea Journal of Engineering and Science
- Published
- 2026-11-15
- DOI
- https://doi.org/10.34248/bsengineering.1995246
- Primary Topic
- Advanced Clustering Algorithms Research
- Type
- article
- Field-Weighted Citation Impact
- 0.00