MNIST Classification — Binary, Multiclass, Multilabel & Multioutput

A complete tour of classification approaches on the MNIST handwritten digit dataset (70,000 images, 28×28 pixels).

Overview

Covers four distinct classification problem types on the same dataset, comparing models honestly with cross-validation rather than relying on raw accuracy (which is misleading on imbalanced classes).

Approach

Binary classification ("is it an 8?") — SGD Classifier baseline, evaluated against a DummyClassifier to prove the model actually learned something beyond guessing the majority class.
Threshold analysis — precision-recall and ROC curves to show the tradeoff between precision and recall as the decision threshold moves, instead of reporting one fixed operating point.
Model comparison — SGD (linear) vs Random Forest (ensemble) on the same binary task.
Multiclass classification — all 10 digits at once, with a 10×10 confusion matrix (normalized by row) to see which digits actually get confused with which.
Multilabel classification — predicting multiple independent yes/no labels per sample simultaneously, including ClassifierChain to model dependence between labels.
Multioutput classification — predicting all 784 pixel values at once (image denoising as a classification problem).

Results

Model / Task	Metric	Score
SGD (binary: "is it an 8?")	F1	0.58
SGD (binary)	ROC AUC	0.915
Random Forest (binary)	F1	0.90
Random Forest (binary)	ROC AUC	0.997
Multilabel (KNN)	Macro F1	0.978

The SGD → Random Forest jump illustrates why a linear decision boundary struggles on raw pixel data compared to a tree ensemble that can carve non-linear regions.

Tech stack

Python, scikit-learn, matplotlib

Notebooks

classification.ipynb — original notebook, my own work
classification_notebook_polished.ipynb — same analysis with AI-assisted documentation and formatting added for readability

2.1 KiB Raw Blame History Unescape Escape