如何在 PySpark 中使用 TfidfVectorizer

Question

I am very new at using Pyspark and have some issues with Pyspark Dataframe.

I'm trying to implement the TF-IDF algorithm. I did it with pandas dataframe once. However, I started using Pyspark and now everything changed:( I can't use Pyspark Dataframe like dataframe['ColumnName'] . When I write and run the code, it says dataframe is not iterable. This is a massive problem for me and has not been solved yet. The current problem below:

With Pandas:

tfidf = TfidfVectorizer(vocabulary=vocabulary, dtype=np.float32)
tfidf.fit(pandasDF['name'])
tfidf_tran = tfidf.transform(pandasDF['name'])

With PySpark:

tfidf = TfidfVectorizer(vocabulary=vocabulary, dtype=np.float32)
tfidf.fit(SparkDF['name'])
tfidf_tran = tfidf.transform(SparkDF['name'])

---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
~\AppData\Local\Temp/ipykernel_19992/3734911517.py in <module>
     13 vocabulary = list(vocabulary)
     14 tfidf = TfidfVectorizer(vocabulary=vocabulary, dtype=np.float32)
---> 15 tfidf.fit(dataframe['name'])
     16 tfidf_tran = tfidf.transform(dataframe['name'])
     17 

E:\Anaconda\lib\site-packages\sklearn\feature_extraction\text.py in fit(self, raw_documents, y)
   1821         self._check_params()
   1822         self._warn_for_unused_params()
-> 1823         X = super().fit_transform(raw_documents)
   1824         self._tfidf.fit(X)
   1825         return self

E:\Anaconda\lib\site-packages\sklearn\feature_extraction\text.py in fit_transform(self, raw_documents, y)
   1200         max_features = self.max_features
   1201 
-> 1202         vocabulary, X = self._count_vocab(raw_documents,
   1203                                           self.fixed_vocabulary_)
   1204 

E:\Anaconda\lib\site-packages\sklearn\feature_extraction\text.py in _count_vocab(self, raw_documents, fixed_vocab)
   1110         values = _make_int_array()
   1111         indptr.append(0)
-> 1112         for doc in raw_documents:
   1113             feature_counter = {}
   1114             for feature in analyze(doc):

E:\Anaconda\lib\site-packages\pyspark\sql\column.py in __iter__(self)
    458 
    459     def __iter__(self):
--> 460         raise TypeError("Column is not iterable")
    461 
    462     # string methods

TypeError: Column is not iterable

Answer 1

Tf-idf is the term frequency multiplied by the inverse document frequency. There isn't an explicit tf-idf vectorizer within the MlLib for dataframes in the Pyspark library, but they have 2 useful models that will help get you to the tf-idf. Using the HashingTF , you'd get the term frequencies. Using the IDF , you'd have the inverse document frequencies. Multiply the two together, and you should have an output matrix matching what you would be expecting from the TfidfVectorizer you specified originally.