Scipy hstack results in "TypeError: no supported conversion for types: (dtype('float64'), dtype('O'))"

Question

I am trying to run hstack to join a column of integer values to a list of columns created by a TF-IDF (so I can eventually use all of these columns/features in a classifier).

I'm reading in the column using pandas, checking for any NA values and converting them to the largest value in the dataframe like so :

  OtherColumn = p.read_csv('file.csv', delimiter=";", na_values=['?'])[["OtherColumn"]]
  OtherColumn = OtherColumn.fillna(OtherColumn.max())
  OtherColumn = OtherColumn.convert_objects(convert_numeric=True)

Then I read in my text column and run TF-IDF to create loads of features :

  X = list(np.array(p.read_csv('file.csv', delimiter=";"))[:,2])

  tfv = TfidfVectorizer(min_df=3,  max_features=None, strip_accents='unicode',  
        analyzer='word',token_pattern=r'\w{1,}',ngram_range=(1, 2), use_idf=1,smooth_idf=1,sublinear_tf=1)
  tfv.fit(X)

Finally, I want to join them all together, and this is where our error occurs and the program cannot run, and also I am unsure whether I am using the StandardScaler appropriately here :

  X =  sp.sparse.hstack((X, OtherColumn.values)) #error here
  sc = preprocessing.StandardScaler().fit(X)
  X = sc.transform(X)
  X_test = sc.transform(X_test)

Full error message:

---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-13-79d1e70bc1bc> in <module>()
---> 47 X =  sp.sparse.hstack((X, OtherColumn.values))
     48 sc = preprocessing.StandardScaler().fit(X)
     49 X = sc.transform(X)

C:\Users\Simon\Anaconda\lib\site-packages\scipy\sparse\construct.pyc in hstack(blocks, format, dtype)
    421 
    422     """
--> 423     return bmat([blocks], format=format, dtype=dtype)
    424 
    425 

C:\Users\Simon\Anaconda\lib\site-packages\scipy\sparse\construct.pyc in bmat(blocks, format, dtype)
    537     nnz = sum([A.nnz for A in blocks[block_mask]])
    538     if dtype is None:
--> 539         dtype = upcast(*tuple([A.dtype for A in blocks[block_mask]]))
    540 
    541     row_offsets = np.concatenate(([0], np.cumsum(brow_lengths)))

C:\Users\Simon\Anaconda\lib\site-packages\scipy\sparse\sputils.pyc in upcast(*args)
     58             return t
     59 
---> 60     raise TypeError('no supported conversion for types: %r' % (args,))
     61 
     62 

TypeError: no supported conversion for types: (dtype('float64'), dtype('O'))

Answer 1

As discussed in Numpy hstack - "ValueError: all the input arrays must have same number of dimensions" - but they do you many need to explicitly cast the inputs to sparse.hstack. The sparse code is not as robust as the core numpy code.

If X is a sparse array with dtype=float, and A is dense with dtype=object, several options are possible.

sparse.hstack(X, A) # error
sparse.hstack(X.astype(object), A) # cast X to object; return object
sparse.hstack(X, A.astype(float)) # cast A to float; return float
hstack(X.A, A) # make X dense, result will be type object

A.astype(float) will work if A contains some NaN. See http://pandas.pydata.org/pandas-docs/stable/gotchas.html regarding NaN. If A is object for some other reason (e.g. ragged lists), then we'll have to revisit the issue.

Another possibility is to use Pandas's concat. http://pandas.pydata.org/pandas-docs/stable/merging.html. I assume Pandas has paid more attention to these issues than the sparse coders.