ParzenWindow-PR_Experiment

Implement the parzenwindow, which is the experiment 1 in Pattern Recognition.(Harbin Institute of Technology, CS, Master). I've read a few ParzenWindow samples which realized by Python, but none of them are correct, so I spend three days in understanding and programing it. Hope it'll help you:)

Formula

The implementation strictly follow the "Pattern Classification second edition". Specifically, the references are in page 138, it is as follows:

The flowchart of ParzenWindow

Code illustration

方法名	所属文件	功能
generateDate	parzen_main.py	生成样本
ShowData	parzen_main.py	绘制样本点
dist	parzen_main.py	计算L2距离
gaussCore	parzen_main.py	高斯核函数
ShowData	parzen_main.py	绘制样本点
PN	parzen_main.py	绘制样本点
ShowData	parzen_main.py	Pn公式计算
calculate_acc	parzen_main.py	准确率计算
calculate_acc	parzen_main.py	样本点预测
main	parzen_main.py	主函数
calc_statistics	initia_distribution.py	核对方差、均值
main	initia_distribution.py	绘制初始分布

Expriments and results

We let test samples' num stay 300 during the whole lab, and adjust the training num N and window sizeh.

initial distribution

the left is the test set, the other is the train set.

N=30 h=0.5 accuracy=0.5367

The tint dots are test samples' inference result, you can see most of them are around the corresponding train samples, and this phenomenon illustrates we are right.

N=30 h=1 accuracy=0.4233

N=30 h=5 accuracy=0.3333

N=90 h=0.5 accuracy=0.56

N=90 h=1 accuracy=0.4667

N=90 h=5 accuracy=0.3333

N=300 h=0.5 accuracy=0.6

N=300 h=1 accuracy=0.5367

N=300 h=5 accuracy=0.5367

N=1200 h=0.5 accuracy=0.6667

N=1200 h=1 accuracy=0.5667

N=1200 h=5 accuracy=0.3833

Set N=1200, the accuracy changes with h.