FINAL write up : CS 174 Spring 2012
- k-Nearest Neighbors
-
k-NN method is simple but strong approach to classification.
We implemented k-NN classifier for testing. We built several k-NN for several k. We used both HIK(Histogram Kernel Intersection) applied and not applied raw dataset for comparing results. As we can see the graph below, we got the best result (9.01%) when we set up k = 1. Right graph shows the maximum accuracy that was recorded for specific breed. However, we realized that k-NN method has a limitation of accuracy for multi class classification.
k-NN test result
- Multi-class SVM
-
The graph below is the result of multi-class SVM(one versus one). The multi-class SVM got 19% average accuracy. Multi-class SVM showed much better performance compare to k-NN method, but it is still poor accuracy. Moreover, this method required a lot of time and memory. We could barely run this method with Dartmouth research computing system.
SVM result