KNeighborsClassifier结合10折交叉验证始终预测相同值如何解决？

你在结合KNN与10折交叉验证对全量数据集测试时，出现所有折叠的预测结果全部为类别3的问题，真实标签包含1、2、3三类。

问题代码 #

import numpy as np
import pandas as pd # data processing, CSV file I/O (e.g. pd.read_csv)
import os
import scipy.io   
from sklearn.neighbors import KNeighborsClassifier
from sklearn import metrics
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import StandardScaler
from torch.utils.data import Dataset, DataLoader
from sklearn import preprocessing
import torch
import numpy as np
from sklearn.model_selection import KFold
from sklearn.neighbors import KNeighborsClassifier
from sklearn import metrics

def load_mat_data(path):
    mat = scipy.io.loadmat(DATA_PATH)
    x,y = mat['data'], mat['class']
    x = x.astype('float32')
    # stadardize values
    standardizer = preprocessing.StandardScaler()
    x = standardizer.fit_transform(x) 
    return x, standardizer, y

def numpyToTensor(x):
    x_train = torch.from_numpy(x)
    return x_train

class DataBuilder(Dataset):
    def __init__(self, path):
        self.x, self.standardizer, self.y = load_mat_data(DATA_PATH)
        self.x = numpyToTensor(self.x)
        self.len=self.x.shape[0]
        self.y = numpyToTensor(self.y)
    def __getitem__(self,index):      
        return (self.x[index], self.y[index])
    def __len__(self):
        return self.len

datasets = ['/home/katerina/Desktop/datasets/GSE75110.mat']

for DATA_PATH in datasets:

    print(DATA_PATH)
    data_set=DataBuilder(DATA_PATH)

    pred_rpknn = [0] * len(data_set.y)
    kf = KFold(n_splits=10, shuffle = True, random_state=7)

    for train_index, test_index in kf.split(data_set.x):
        #Create KNN Classifier
        knn = KNeighborsClassifier(n_neighbors=5)
        #print("TRAIN:", train_index, "TEST:", test_index)
        x_train, x_test = data_set.x[train_index], data_set.x[test_index]
        y_train, y_test = data_set.y[train_index], data_set.y[test_index]
        #Train the model using the training sets
        y1_train = y_train.ravel()
        knn.fit(x_train, y1_train)
        #Predict the response for test dataset
        y_pred = knn.predict(x_test)
        #print(y_pred)
        # Model Accuracy, how often is the classifier correct?
        print("Accuracy:",metrics.accuracy_score(y_test, y_pred))
        c = 0
        for idx in test_index:
            pred_rpknn[idx] = y_pred[c]
            c +=1
    print("Accuracy:",metrics.accuracy_score(data_set.y, pred_rpknn))
    print(pred_rpknn, data_set.y.reshape(1,-1))

输出结果 #

/home/katerina/Desktop/datasets/GSE75110.mat
Accuracy: 0.2857142857142857
Accuracy: 0.38095238095238093
Accuracy: 0.14285714285714285
Accuracy: 0.4
Accuracy: 0.3
Accuracy: 0.25
Accuracy: 0.3
Accuracy: 0.6
Accuracy: 0.25
Accuracy: 0.45
Accuracy: 0.33497536945812806
[3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3]

真实标签 #

tensor([[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 2,2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2,2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2,2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2,2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 3, 3, 3,3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3]])

解决方案 #

• 修正代码逻辑错误：load_mat_data函数定义时传入了形参path，但实际调用scipy.io.loadmat时用的是全局变量DATA_PATH，建议修改为mat = scipy.io.loadmat(path)，避免后续修改路径时出现逻辑错误。
• 修正数据预处理逻辑：你当前是对全量数据集做标准化，会引入数据泄露，应该在每个交叉验证折叠内，仅用当前训练集拟合StandardScaler，再分别对训练集和测试集做变换，不要提前处理全量数据。
• 调整K值：目前使用的k=5不一定适配你的数据集，建议从小到大尝试不同k值（如1、3、7、9、11），避免k值过大导致预测偏向多数类。
• 更换距离度量：高维特征下欧氏距离的区分度会大幅下降，可以尝试更换为曼哈顿距离、余弦距离，示例：KNeighborsClassifier(n_neighbors=5, metric='cosine')。
• 优化权重策略：可以把KNN的权重设置为按距离加权，让距离更近的邻居对预测结果影响更大，降低多数类的干扰，示例：KNeighborsClassifier(n_neighbors=5, weights='distance')。
• 过滤冗余特征：计算每个特征和标签的相关性，去掉和标签完全无关的噪声特征，提升距离计算的有效性。
• 处理类别不平衡：如果确认类别分布存在较大偏差，可以对少数类做过采样、对多数类做欠采样，平衡不同类别的样本量。

内容的提问来源于stack exchange，提问作者KateB