MNIST Dataset Classification via Multiview Topological Data Analysis

Multiview learning promises richer representations by combining complementary perspectives. However, generating views that are both diverse and informative remains challenging. To explore this, we study digit classification using a multiview stacking model that integrates four spatial quadrants of each MNIST image plus a topological view derived from persistent homology (TDA). We use a random forest as both the base and meta learner in a stacking ensemble. By systematically combining subsets of spatial and topological views, we generate 32 distinct multiview variants. Evaluating these on 44,800 training and 11, 200 test images¿sampled from 80% of the 70,000-image MNIST dataset across 30 random splits¿we find that adding the topological view increases mean accuracy from 0.9668±0.001 to 0.9848±0.001, a 1.66 percentage point gain with very large effect size. Notably, in an extreme low-data setting with just 70 training and 30 test samples, the topological view continued to improve accuracy by 0.14-0.24 percentage points across 140 runs, while additional spatial views contribute little. These findings suggest that persistent homology captures complementary global structure missing from pixel-level partitions, and that incorporating TDA into a late-fusion framework offers a simple, architecture-independent path to robust classification under data scarcity. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2026.

MNIST Dataset Classification via Multiview Topological Data Analysis Chapter in Scopus