SMOTE initialisation expects n

问题

I have already pre-cleaned the data, and below shows the format of the top 4 rows:

     [IN] df.head()

    [OUT]   Year    cleaned
         0  1909    acquaint hous receiv follow letter clerk crown...
         1  1909    ask secretari state war whether issu statement...
         2  1909    i beg present petit sign upward motor car driv...
         3  1909    i desir ask secretari state war second lieuten...
         4  1909    ask secretari state war whether would introduc...

I have called train_test_split() as follows:

     [IN] X_train, X_test, y_train, y_test = train_test_split(df['cleaned'], df['Year'], random_state=2)
   [Note*] `X_train` and `y_train` are now Pandas.core.series.Series of shape (1785,) and `X_test` and `y_test` are also Pandas.core.series.Series of shape (595,)

I have then vectorized the X training and testing data using the following TfidfVectorizer and fit/transform procedures:

     [IN] v = TfidfVectorizer(decode_error='replace', encoding='utf-8', stop_words='english', ngram_range=(1, 1), sublinear_tf=True)
          X_train = v.fit_transform(X_train)
          X_test = v.transform(X_test)

I'm now at the stage where I would normally apply a classifier, etc (if this were a balanced set of data). However, I initialize imblearn's SMOTE() class (to perform over-sampling)...

     [IN] smote_pipeline = make_pipeline_imb(SMOTE(), classifier(random_state=42))
          smote_model = smote_pipeline.fit(X_train, y_train)
          smote_prediction = smote_model.predict(X_test)

... but this results in:

     [OUT] ValueError: "Expected n_neighbors <= n_samples, but n_samples = 5, n_neighbors = 6.

I've attempted to whittle down the number of n_neighbors but to no avail, any tips or advice would be much appreciated. Thanks for reading.

------------------------------------------------------------------------------------------------------------------------------------

EDIT:

Full Traceback

The dataset/dataframe (df) contains 2380 rows across two columns, as shown in df.head() above. X_train contains 1785 of these rows in the format of a list of strings (df['cleaned']) and y_train also contains 1785 rows in the format of strings (df['Year']).

Post-vectorization using TfidfVectorizer(): X_train and X_test are converted from pandas.core.series.Series of shape '(1785,)' and '(595,)' respectively, to scipy.sparse.csr.csr_matrix of shape '(1785, 126459)' and '(595, 126459)' respectively.

As for the number of classes: using Counter(), I've calculated that there are 199 classes (Years), each instance of a class is attached to one element of aforementioned df['cleaned'] data which contains a list of strings extracted from a textual corpus.

The objective of this process is to automatically determine/guess the year, decade or century (any degree of classification will do!) of input textual data based on vocabularly present.

回答1:

Since there are approximately 200 classes and 1800 samples in the training set, you have on average 9 samples per class. The reason for the error message is that a) probably the data are not perfectly balanced and there are classes with less than 6 samples and b) the number of neighbors is 6. A few solutions for your problem:

Calculate the minimum number of samples (n_samples) among the 199 classes and select n_neighbors parameter of SMOTE class less or equal to n_samples.
Exclude from oversampling the classes with n_samples < n_neighbors using the ratio parameter of SMOTE class.
Use RandomOverSampler class which does not have a similar restriction.
Combine 3 and 4 solutions: Create a pipeline that is using SMOTE and RandomOversampler in a way that satisfies the condition n_neighbors <= n_samples for smoted classes and uses random oversampling when the condition is not satisfied.