3
$\begingroup$

I'm new to data science/ml and working on using the sklearn libraries to classify data. I'm currently using the KNeighborsClassifier with 5 fold cross validation whilst tweaking the k value but its producing a graph that looks quite strange.

I have my training data and test data in 2 different CSV files and load them in like this:

trainData = pd.read_csv('train.csv',header='infer')
testData = pd.read_csv('test.csv',header='infer')

I then separate the classifiers (Y is name of the column in my dataset that's the classification):

trainY = trainData['Y']
trainX = trainData.drop(['Y'],axis=1)

testY = testData['Y']
testX = testData.drop(['Y'],axis=1)

I use sklearn KNeighborsClassifier with 5 fold cross validation whilst tweaking the k value from 2 to 20:

trainAcc = []
testAcc = []

for i in range(2,20):
    clf = KNeighborsClassifier(n_neighbors=i, metric='minkowski', p=2)
    trainScores = cross_val_score(estimator=clf, X=trainX, y=trainY, cv=5, n_jobs=4)
    testScores= cross_val_score(estimator=clf, X=testX, y=testY, cv=5, n_jobs=4)
    trainAcc.append((i, trainScores.mean()))
    testAcc.append((i, testScores.mean()))

I then print the graph:

plt.plot([x[0] for x in trainAcc],[x[1] for x in trainAcc], 'ro-', [x[0] for x in testAcc],[x[1] for x in testAcc], 'bv--')

But I get something weird like this:

enter image description here

Can anyone explain where I went wrong and why my graph looks the way it does.

Thanks.

EDIT: It is indeed weird because when I run it without doing the cross-validation, I get a more normal graph like this:

clf.fit(X=trainX, y=trainY)
predTrainY = clf.predict(trainX)
predTestY = clf.predict(testX)
trainAcc.append(accuracy_score(trainY, predTrainY))
testAcc.append(accuracy_score(testY, predTestY))

enter image description here

$\endgroup$
6
  • $\begingroup$ This is certainly interesting. There's one obvious possible explanation for parity issues though: the neighbors' votes can result in a tie when k is even. $\endgroup$
    – Ben Reiniger
    Mar 4, 2020 at 19:51
  • $\begingroup$ @BenReiniger, What's strange about it is when I run without cross validation, e.g clf.fit(X=trainX, y=trainY) predTrainY = clf.predict(trainX) predTestY = clf.predict(testX) I get much more of a normal graph, so I'm not sure if I've applied the cross-validation correctly. Please see my edit. $\endgroup$
    – Tahmid
    Mar 4, 2020 at 20:14
  • 2
    $\begingroup$ Ah, yes, you are applying cross_val_score incorrectly. That function splits into folds, trains the model and predicts for each train/test split. Running it on your (trainX, trainY) gives the test-fold average score, and running it on your (testX, testY) again gives a test-fold average, with models trained on subsets of testX now. Seeing the two scores track closely is expected! Do you still see the parity effect on the lower plot, if you include odd indices? $\endgroup$
    – Ben Reiniger
    Mar 4, 2020 at 20:59
  • $\begingroup$ @BenReiniger, Thanks for the reply. When including odd and even k values from 1 - 17 on the bottom graph, the zigzag effect is present but the general shape of the graph still remains the same (both lines still slowly come towards each other). I think I understand what you mean, how can I get the train-folder average so that my cross-validation graph prints correctly (in similar fashion to the bottom graph)? $\endgroup$
    – Tahmid
    Mar 4, 2020 at 21:23
  • $\begingroup$ Ahh, I think I got it! I can just plot the training line as is (like I did in the bottom graph) because the cross-val is only for the test set/test folds? e.g trainAcc.append((i, accuracy_score(trainY, predTrainY))) testAcc.append((i, testscores.mean())) $\endgroup$
    – Tahmid
    Mar 4, 2020 at 21:43

3 Answers 3

2
$\begingroup$

To summarize from the comment thread: there are two "weird" things going on here.

1. The zig-zag.

As I addressed in the comments, and @BrianSpiering in an answer, this is probably a parity effect, arising from tied votes among the nearest neighbors when $k$ is even.

2. Training accuracy not decreasing (toward test accuracy) with increasing $k$.

This was caused by an incorrect usage of cross_val_score. You ran that function twice, separately on the training and test sets. But that means your results were the out-of-fold score on the training set (not the actual training score!), and scores coming from models fitted on (folds of) the test set (not scores on the test set from models trained on the training set).

Probably the cleanest way to approach this is to use GridSearchCV to search over the different values of $k$. You can then pull the results from cv_results_ for plotting. Otherwise, if you don't need them for anything else, you can remove either the cross-validation (just score on the train and test set) or the test set (just do cross-validation and get scores on the train and test folds).

$\endgroup$
0
$\begingroup$

One interpretation is the model has high accuracy when k is even number. An even number of groups in KNeighborsClassifier can result in a high number of ties (i.e., the model predicts a data point is equally likely to belong to multiple groups). The model has reduced accuracy when k is odd, ties are less likely to happen when k is odd.

It might be helpful to look at performance on the raw data. In particular, when the model is right and when the model is wrong.

$\endgroup$
1
  • $\begingroup$ What's strange about it is when I run without cross validation, e.g clf.fit(X=trainX, y=trainY) predTrainY = clf.predict(trainX) predTestY = clf.predict(testX) I get much more of a normal graph, so I'm not sure if I've applied the cross-validation correctly. Am I required to refit the model each time in my loop? I assumed cross_val_score handles that for me. Please see my edit. $\endgroup$
    – Tahmid
    Mar 4, 2020 at 20:19
0
$\begingroup$

You could try to use the same k, like k=8 and run your cross-validation model 100 times, maybe with some little shift on sampling each run, and plot the outcome of each run, to see if there would be some zigzag between the runs. Maybe your data has a lot of cases where for example the weight of the distance function is the same for two candidates(or even more), but you have to choose only one because of the limit of your k. You could try to find better distance functions for such scenarios, like allowing to have clusters of different size $k = x +${$1,-1,0$}

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.