聚类算法2——LVQ学习向量量化(算法步骤及matlab代码)

看了西关书的聚类算法，算法原理很容易明白，接下来就是整理成自己的理解思路，然后一步一步来实现算法，那么就来做吧。

学习向量量化（LVQ）算法步骤

输入：样本集（带标记），原型向量个数，各原型向量预设的类别标记，学习率（0，1）

输出：最终的原型向量

Step1:初始化函数 ，init_lvq()

Step1.1: 载入样本集，初始化学习率，初始化原型向量

输入: txt样本数据，原型向量个数q

输出： 样本集，原型向量

Repeat

Step2:聚类： lvq_cluster()

输入：原型向量、样本

输出：最近原型向量

Step2.1: 计算每个样本与原型向量集的距离，并找到与xi最近的原型向量

Step2.2: 根据标签，保存到样本容器

Step3: 更新原型向量并，保存原型向量对应的样本容器: update_vector();

若标签相同，p’ = p+yiTa *(x-p)

否则：p’ = p-yiTa *(x-p)

Until

Setp4: 中止迭代条件： is_stop();

若P与p’的欧式距离<0.01 ,停止迭代

LVQ代码及样本下载链接

ok啦，接下来不废话上代码（Matlab发布形式）

%Algorithm: learning vector Quantization
%input: 带标记的样本集
%output：原型向量
function Main()
clc;clear;close
%1.0
melon_data = load('melon4.0_lables.txt');
q = 5;
learn_rate = 0.1;
[learn_vector,sample_data]= init_lvq(melon_data,q);
[sample_rows,~] = size(sample_data);
sample_calss_box =zeros(sample_rows,2*q); % store class
show_now(learn_vector,sample_data,sample_calss_box);
max_loop = 100;
for loop = 1:max_loop
    temp_vector = learn_vector;
    for i = 1:sample_rows
       close_LV_ind  = lvq_cluster(sample_data(i,1:2),learn_vector);
       [learn_vector, sample_calss_box]= update_vector(i,close_LV_ind,sample_data,learn_vector,learn_rate,sample_calss_box);
    end
    if is_stop(learn_vector,temp_vector)
        show_now(learn_vector,sample_data,sample_calss_box);
        fprintf('迭代次数为：%d\n',loop);
        break;
    end
    show_now(learn_vector,sample_data,sample_calss_box);
    sample_calss_box =zeros(sample_rows,2*q); %clear
end
end

subfunction

%1.0
function [learn_vector,sample_data]= init_lvq(melon_data,q)
sample_data = melon_data(:,2:end);
learn_vector =[sample_data(5,:);
    sample_data(12,:);
    sample_data(18,:);
    sample_data(23,:);
    sample_data(29,:)];
end
%2.0
function close_LV_ind  = lvq_cluster(sample_data,learn_vector)

dist = pdist2(learn_vector(:,1:2),sample_data);
[~,close_LV_ind] = min(dist);
end

function [cur_vector,sample_box]= update_vector(i,close_LV_ind,sample_data,learn_vector,learn_rate,sample_calss_box)
    if sample_data(i,end)==learn_vector(close_LV_ind,end)
       learn_vector(close_LV_ind,1:2) = learn_vector(close_LV_ind,1:2) +...
           learn_rate*(sample_data(i,1:2)-learn_vector(close_LV_ind,1:2));
    else
        learn_vector(close_LV_ind,1:2) = learn_vector(close_LV_ind,1:2) -...
           learn_rate*(sample_data(i,1:2)-learn_vector(close_LV_ind,1:2));
    end
    same_label_ind = find(learn_vector(:,end)==sample_data(i,end));
    if ~isempty(same_label_ind)
        [~,min_BDI ]= min( pdist2(learn_vector(same_label_ind,1:2), sample_data(i,1:2)));%min_box_dist_ind
        zero_ind = find(sample_calss_box(:,(same_label_ind(min_BDI)*2))==0);
        sample_calss_box(zero_ind(1),(same_label_ind(min_BDI)*2-1):(same_label_ind(min_BDI)*2)) = sample_data(i,1:2);
    end
    cur_vector = learn_vector;
    sample_box = sample_calss_box;
end

function  isStop = is_stop(learn_vector,temp_vector)
isStop = 0;
cons_ind = find(pdist(learn_vector-temp_vector)>0.00000001);
if isempty(cons_ind)
    isStop =1;
end

end

function show_now(learn_vector,sample_data,sample_calss_box)
close
label1 = [sample_data(1:8,:);sample_data(22:end,:)];
label2 = sample_data(9:21,:);
plot(label1(:,1),label1(:,2),'+b');hold on
plot(label2(:,1),label2(:,2),'+r'); %2
%vector
plot(learn_vector(1,1),learn_vector(1,2),'pb');
plot(learn_vector(4,1),learn_vector(4,2),'pb');
plot(learn_vector(5,1),learn_vector(5,2),'pb');
plot(learn_vector(2,1),learn_vector(2,2),'pr');%2
plot(learn_vector(3,1),learn_vector(3,2),'pr');
%sample box
plot(sample_calss_box(:,1),sample_calss_box(:,2),'ro');
plot(sample_calss_box(:,3),sample_calss_box(:,4),'bo');
plot(sample_calss_box(:,5),sample_calss_box(:,6),'go');
plot(sample_calss_box(:,7),sample_calss_box(:,8),'rs');
plot(sample_calss_box(:,9),sample_calss_box(:,10),'bs');
xlabel('density');ylabel('sugar rate');
end

迭代次数为：41