An Experiment using MNIST dataset.
| Classifier | Distance | Preprocessing | Correct Ratio | Error Ratio |
|---|---|---|---|---|
| KNN (kdtree, K=5) | L2 (Euclidean) | None | 96.90 % | 3.10 % |
| KNN (kdtree, K=5) | L3 | None | 97.28 % | 2.72 % |
| KNN (kdtree, K=5) | L1 | None | 96.26 % | 3.74 % |
| KNN (kdtree, K=1) | L2 (Euclidean) | None | 96.91 % | 3.09 % |
| KNN (kdtree, K=20) | L2 (Euclidean) | None | 96.27 % | 3.73 % |
| KNN (kdtree, K=5) | L2 (Euclidean) | Binaryzation (threshold=10) | 96.63 % | 3.37 % |
| KNN (kdtree, K=5) | L2 (Euclidean) | Downsample (factor=2) | 96.60 % | 3.40 % |
| KNN (kdtree, K=5) | L2 (Euclidean) | Downsample (factor=4) | 93.21 % | 6.79 % |
| KNN (kdtree, K=5) | L2 (Euclidean) | Blur (factor=2) | 97.58 % | 2.42 % |
| KNN (kdtree, K=5) | L2 (Euclidean) | Sum (image into scalar) | 18.33 % | 81.67 % |
Use a random-chosen subset of training data as sample to choose
the best parameter K for our KNN Classifier. The result is
- Choose Distance: L3
- Choose K=3
- Result: Correct Ratio = 97.58%, Error Ratio = 2.42%