为什么knn总是预测相同的数字？我该如何解决这个问题？数据集在这里链接。
代码：
import numpy as npimport pandas as pd # data processing, CSV file I/O (e.g. pd.read_csv)import osimport scipy.io from sklearn.neighbors import KNeighborsClassifierfrom sklearn import metricsfrom sklearn.model_selection import train_test_splitfrom sklearn.preprocessing import StandardScalerfrom torch.utils.data import Dataset, DataLoaderfrom sklearn import preprocessingimport torchimport numpy as npfrom sklearn.model_selection import KFoldfrom sklearn.neighbors import KNeighborsClassifierfrom sklearn import metricsdef 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, ydef numpyToTensor(x): x_train = torch.from_numpy(x) return x_trainclass 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.lendatasets = ['/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：
/home/katerina/Desktop/datasets/GSE75110.matAccuracy: 0.2857142857142857Accuracy: 0.38095238095238093Accuracy: 0.14285714285714285Accuracy: 0.4Accuracy: 0.3Accuracy: 0.25Accuracy: 0.3Accuracy: 0.6Accuracy: 0.25Accuracy: 0.45Accuracy: 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]我尝试将knn与k折交叉验证结合，以使用10个折叠测试整个数据集。问题是knn在每个折叠中总是预测3的数组。我想要预测的类别是这些：
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]]
回答：
TL;DR
这与StandardScaler有关，将其更改为简单的归一化处理。
例如：
from sklearn import preprocessing...x = preprocessing.normalize(x)解释：
您使用的标准化处理会执行以下操作：
样本`x`的标准分数计算如下： z = (x - u) / s其中`u`是训练样本的均值，如果`with_mean=False`则为零，`s`是训练样本的标准差，如果`with_std=False`则为一。
而您实际上希望这些特征帮助KNN决定哪个向量更接近。
在normalize中，归一化是分别对每个向量进行的，因此不会影响甚至有助于KNN区分向量
对于KNN，StandardScaler实际上可能损害您的预测。最好在其他形式的数据中使用它。
import scipy.iofrom torch.utils.data import Datasetfrom sklearn import preprocessingimport torchimport numpy as npfrom sklearn.model_selection import KFoldfrom sklearn.neighbors import KNeighborsClassifierfrom sklearn import metricsdef load_mat_data(path): mat = scipy.io.loadmat(DATA_PATH) x, y = mat['data'], mat['class'] x = x.astype('float32') # stadardize values x = preprocessing.normalize(x) return x, ydef numpyToTensor(x): x_train = torch.from_numpy(x) return x_trainclass DataBuilder(Dataset): def __init__(self, path): self.x, 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.lendatasets = ['/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.view(-1) 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 in loop:", 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))Accuracy in loop: 1.0Accuracy in loop: 0.8571428571428571Accuracy in loop: 0.8571428571428571Accuracy in loop: 1.0Accuracy in loop: 0.9Accuracy in loop: 0.9Accuracy in loop: 0.95Accuracy in loop: 1.0Accuracy in loop: 0.9Accuracy in loop: 1.0Accuracy: 0.9359605911330049