1、SVD与PCA原理:
https://www.cnblogs.com/pinard/p/6251584.html
2、KNN 算法
KNN 算法其实简单的说就是“物以类聚”,也就是将新的没有被分类的点分类为周围的点中大多数属于的类。它采用测量不同特征值之间的距离方法进行分类,思想很简单:如果一个样本的特征空间中最为临近(欧式距离进行判断)的K个点大都属于某一个类,那么该样本就属于这个类。这就是物以类聚的思想。
当然,实际中,不同的K取值会影响到分类效果,并且在K个临近点的选择中,都不加意外的认为这K个点都是已经分类好的了,否则该算法也就失去了物以类聚的意义了。
KNN算法的不足点:
(1)、当样本不平衡时,比如一个类的样本容量很大,其他类的样本容量很小,输入一个样本的时候,K个临近值中大多数都是大样本容量的那个类,这时可能就会导致分类错误。改进方法是对K临近点进行加权,也就是距离近的点的权值大,距离远的点权值小。
(2)、计算量较大,每个待分类的样本都要计算它到全部点的距离,根据距离排序才能求得K个临近点,改进方法是:先对已知样本点进行剪辑,事先去除对分类作用不大的样本。
适用性:
适用于样本容量比较大的类域的自动分类,而样本容量较小的类域则容易误分
算法描述:
1、计算已知类别数据集合汇总的点与当前点的距离
2、按照距离递增次序排序
3、选取与当前点距离最近的K个点
4、确定距离最近的前K个点所在类别的出现频率
5、返回距离最近的前K个点中频率最高的类别作为当前点的预测分类
function y = knn(X, X_train, y_train, K)
%KNN k-Nearest Neighbors Algorithm.
%
% INPUT: X: testing sample features, P-by-N_test matrix.
% X_train: training sample features, P-by-N matrix.
% y_train: training sample labels, 1-by-N row vector.
% K: the k in k-Nearest Neighbors
%
% OUTPUT: y : predicted labels, 1-by-N_test row vector.
[~,N_test] = size(X);
predicted_label = zeros(1,N_test);
for i=1:N_test
[dists, neighbors] = top_K_neighbors(X_train,y_train,X(:,i),K);
% calculate the K nearest neighbors and the distances.
predicted_label(i) = recog(y_train(neighbors),max(y_train));
% recognize the label of the test vector.
end
y = predicted_label;
end
%---------------------------------------------------------------------
%最近K近邻的部分
function [dists,neighbors] = top_K_neighbors( X_train,y_train,X_test,K )
% Author: Ren Kan
% Input:
% X_test the test vector with P*1
% X_train and y_train are the train data set
% K is the K neighbor parameter
[~, N_train] = size(X_train);
test_mat = repmat(X_test,1,N_train);
dist_mat = (X_train-double(test_mat)) .^2;
% The distance is the Euclid Distance.
dist_array = sum(dist_mat);
[dists, neighbors] = sort(dist_array);
% The neighbors are the index of top K nearest points.
dists = dists(1:K);
neighbors = neighbors(1:K);
end
%---------------------------------------------------------------------
%利用概率求解测试集预测标签
function result = recog( K_labels,class_num )
%RECOG Summary of this function goes here
% Author: Ren Kan
[~,K] = size(K_labels);
class_count = zeros(1,class_num+1);
for i=1:K
class_index = K_labels(i)+1; % +1 is to avoid the 0 index reference.
class_count(class_index) = class_count(class_index) + 1;
end
[~,result] = max(class_count);
result = result - 1; % Do not forget -1 !!!
end
https://blog.csdn.net/rk2900/article/details/9080821
https://blog.csdn.net/llp1992/article/details/45040685