isomap mds 实现

Isomap 和 mds 实现

数据集：iris

数据集描述：一共五个字段，前四个字段表示花的四个属性，最后一个字段表示花的类别。共有150组数据，三种类别

Mds方法思路：首先计算任意两朵花的欧式距离结果放入disM2矩阵,接着计算B2，然后对B降维为m*2，得到result2矩阵。

Isomap方法思路：首先计算任意两朵花的欧式距离结果放入disM矩阵，选取k=10，使得前k个最短路径记为其欧式距离，其余记为-1，接着根据最短路径算法，计算任意两点的测地线距离。最后将元素测地线距离矩阵作为mds算法输入，计算降维后的result矩阵。

降维结果：

 

主要代码如下：

#计算欧式距离

for i inrange(m.shape[0]):
    disM[i][i]=0
    for j in range(m.shape[0]):
        if i!=j:
            t=0
            for n in range(m.shape[1]):
                t=t+(m[i][n]-m[j][n])**2
            disM[i][j]=t
            disM2[i][j]=t
            if len(myNeighbor[i])<k:
                #myNeighbor[i].append(j)
                orderInsert(disM[i],myNeighbor[i],j)
            elif disM[i][myNeighbor[i][k-1]]>t:
                test=myNeighbor[i][k-1]
                x=myNeighbor[i].pop()
                disM[i][x]=-1
                orderInsert(disM[i],myNeighbor[i], j)
            else:
                disM[i][j]=-1

#最短路径算法 计算测地线距离
for i in range(disM.shape[0]):
    mylist=np.arange(0,disM.shape[0],1)
    mylist=mylist.tolist()
    while len(mylist)!=0:
        index=minIndex(disM[i],mylist)
        minDis=disM[i][index]
        mylist.remove(index)#disM[i][index]此时是最小的.将index这个下标从mylist删去已经找到i到index最小距离
        if minDis==-1:
            continue
        for t,item in enumerate(myNeighbor[index]):
            alt=minDis+disM[index][item]
            if disM[i][item]==-1 or alt<disM[i][item] :
                disM[i][item]=alt

#计算mds和isomap的B DISI DISTJ DIST

for i in range(B.shape[0]):
    distI[i]=(1/m)*sum(disM[i]**2)
    distJ[i]=(1/m)*sum(disM[0:m][i]**2)
    distI2[i] = (1 / m) * sum(disM2[i] ** 2)
    distJ2[i] = (1 / m) * sum(disM2[0:m][i] ** 2)
for i in range(m):
    for j in range(m):
        B[i][j]=-0.5*(disM[i][j]**2-distI[i]-distJ[j]+dist)
        B2[i][j] = -0.5 * (disM2[i][j] ** 2 - distI2[i] - distJ2[j] + dist2)

#特征值分解
r,vec=np.linalg.eig(B)
r2,vec2=np.linalg.eig(B2)
d2=2
largest=heapq.nlargest(d2,range(len(r)),r.take)
largest2=heapq.nlargest(d2,range(len(r)),r2.take)
V=vec[0:m,largest]
V2=vec2[0:m,largest2]
r=r[largest]
r2=r2[largest2]
R=np.zeros(shape=(d2,d2),dtype=complex)
R2=np.zeros(shape=(d2,d2),dtype=complex)
for i in range(d2):
    R[i][i]=r[i]
    R2[i][i] = r2[i]
result=np.dot(V,R)
result2=np.dot(V2,R2)

#画图

f,ax=plt.subplots(2,1)
ax[0].scatter(x1,y1,color='r',marker='o',s=40,facecolors='none')
ax[0].scatter(x2,y2,color='b',marker='o',s=40,facecolors='none')
ax[0].scatter(x3,y3,color='g',marker='o',s=40,facecolors='none')
ax[0].set_title("isomap")

ax[1].scatter(x4,y4,color='r',marker='o',s=40,facecolors='none')
ax[1].scatter(x5,y5,color='b',marker='o',s=40,facecolors='none')
ax[1].scatter(x6,y6,color='g',marker='o',s=40,facecolors='none')
ax[1].set_title("mds")
plt.show()