【域适应】基于迁移成分分析(TCA) 的典型二分类问题(python)

关于

传统迁移成分分析，区别于深度迁移学习，是一种常规的对特征数据进行对齐的一种技术。

广泛应用于，两个域（源域，目标域）数据特征分布不一致，需要实现目标域数据分析任务中。

文章参考：https://home.cse.ust.hk/~jamesk/papers/ijcai09.pdf

工具

Python

方法实现

通过循环遍历TCA隐藏空间维度和超参数，来获得基于TCA方法的源域到目标域的迁移技术。

主要函数和流程：

import pandas as pd
import numpy as np
from sklearn.svm import SVC
from TCA import TCA# source domain training data
Xs_y = pd.read_csv('Xs_y.csv')
Xs = np.array(Xs_y.iloc[:,:-1])
ys = np.array(Xs_y.iloc[:,-1])
# target domain training data
Xt_y = pd.read_csv('Xt_y.csv')
Xt = np.array(Xt_y.iloc[:,:-1])
yt = np.array(Xt_y.iloc[:,-1])
# test data
test = pd.read_csv('Xtest.csv')
Xtest = np.array(test.iloc[:,:-1])
ytest = np.array(test.iloc[:,-1])reg = SVC()def acc(y_pre, y):right = abs(y_pre - y).sum()/2return (len(y_pre) - right)/len(y_pre)
################################################################
# case one 
# training on target domina data (Xt_y.csv) and test on testing dataset (Xtest.csv)
pre_label = reg.fit(Xt, yt,).predict(Xtest)
rate = acc(pre_label,ytest)
print('Without transfer, we have acc rate %f' % rate )# case two
# training on mapped space (4-d) with source (Xs_y.csv) and target domina data (Xt_y.csv), test on testing dataset (Xtest.csv)
acc_list = [rate]
numbers = np.linspace(0.01, 1, 10)
for j in range(len(numbers)):model = TCA(dim=1, lamda=numbers[j], gamma=0.5)Train_X = np.vstack((Xs,Xt))Xs_new, Xt_new = model.fit(Train_X,Xtest)Train_y = np.hstack((ys,yt))transfer_pre_label = reg.fit(Xs_new, Train_y).predict(Xt_new)transfer_rate = acc(transfer_pre_label,ytest)acc_list.append(transfer_rate)print('lamda = {}'.format(numbers[j]))print('With transfer, we have acc rate %f' % transfer_rate ,'\n')# plot the result
import matplotlib.pyplot as plt
from matplotlib.ticker import MaxNLocatorfig, ax = plt.subplots()
x = numbersplt.plot(x,  acc_list[1:],marker='o',linestyle='-', color='blue',label='TCA transfer with different lamda')
ax.axhline(y=acc_list[0], color='green', linestyle='--', label='without TCA transfer')
#ax.xaxis.set_major_locator(MaxNLocator(integer=True))plt.xlabel('lamda')
plt.ylabel('classification accuracy on test data')ax.legend()
plt.savefig('iteration number.png',bbox_inches = 'tight',dpi=600)
plt.savefig('iteration number.svg',bbox_inches = 'tight',dpi=600)
plt.show()