I'm trying to use scikit-learn's DBSCAN implementation to clusterize a bunch of documents. First I create the TF-IDF matrix using scikit-learn's TfidfVectorizer (it's a 163405x13029 sparse matrix of type numpy.float64). Then I try to clusterize specific subsets of this matrix. Things work fine when the subset is small (say, up to a few thousand rows). But with large subsets (with tens of thousands of rows) I get ValueError: could not convert integer scalar.
Here's the full traceback (idxs is a list of indices):
ValueError Traceback (most recent call last)
<ipython-input-1-73ee366d8de5> in <module>()
193 # use descriptions to clusterize items
194 ncm_clusterizer = DBSCAN()
--> 195 ncm_clusterizer.fit_predict(tfidf[idxs])
196 idxs_clusters = list(zip(idxs, ncm_clusterizer.labels_))
197 for e in idxs_clusters:
/usr/local/lib/python3.4/site-packages/sklearn/cluster/dbscan_.py in fit_predict(self, X, y, sample_weight)
294 cluster labels
295 """
--> 296 self.fit(X, sample_weight=sample_weight)
297 return self.labels_
/usr/local/lib/python3.4/site-packages/sklearn/cluster/dbscan_.py in fit(self, X, y, sample_weight)
264 X = check_array(X, accept_sparse='csr')
265 clust = dbscan(X, sample_weight=sample_weight,
--> 266 **self.get_params())
267 self.core_sample_indices_, self.labels_ = clust
268 if len(self.core_sample_indices_):
/usr/local/lib/python3.4/site-packages/sklearn/cluster/dbscan_.py in dbscan(X, eps, min_samples, metric, algorithm, leaf_size, p, sample_weight, n_jobs)
136 # This has worst case O(n^2) memory complexity
137 neighborhoods = neighbors_model.radius_neighbors(X, eps,
--> 138 return_distance=False)
139
140 if sample_weight is None:
/usr/local/lib/python3.4/site-packages/sklearn/neighbors/base.py in radius_neighbors(self, X, radius, return_distance)
584 if self.effective_metric_ == 'euclidean':
585 dist = pairwise_distances(X, self._fit_X, 'euclidean',
--> 586 n_jobs=self.n_jobs, squared=True)
587 radius *= radius
588 else:
/usr/local/lib/python3.4/site-packages/sklearn/metrics/pairwise.py in pairwise_distances(X, Y, metric, n_jobs, **kwds)
1238 func = partial(distance.cdist, metric=metric, **kwds)
1239
-> 1240 return _parallel_pairwise(X, Y, func, n_jobs, **kwds)
1241
1242
/usr/local/lib/python3.4/site-packages/sklearn/metrics/pairwise.py in _parallel_pairwise(X, Y, func, n_jobs, **kwds)
1081 if n_jobs == 1:
1082 # Special case to avoid picklability checks in delayed
-> 1083 return func(X, Y, **kwds)
1084
1085 # TODO: in some cases, backend='threading' may be appropriate
/usr/local/lib/python3.4/site-packages/sklearn/metrics/pairwise.py in euclidean_distances(X, Y, Y_norm_squared, squared, X_norm_squared)
243 YY = row_norms(Y, squared=True)[np.newaxis, :]
244
--> 245 distances = safe_sparse_dot(X, Y.T, dense_output=True)
246 distances *= -2
247 distances += XX
/usr/local/lib/python3.4/site-packages/sklearn/utils/extmath.py in safe_sparse_dot(a, b, dense_output)
184 ret = a * b
185 if dense_output and hasattr(ret, "toarray"):
--> 186 ret = ret.toarray()
187 return ret
188 else:
/usr/local/lib/python3.4/site-packages/scipy/sparse/compressed.py in toarray(self, order, out)
918 def toarray(self, order=None, out=None):
919 """See the docstring for `spmatrix.toarray`."""
--> 920 return self.tocoo(copy=False).toarray(order=order, out=out)
921
922 ##############################################################
/usr/local/lib/python3.4/site-packages/scipy/sparse/coo.py in toarray(self, order, out)
256 M,N = self.shape
257 coo_todense(M, N, self.nnz, self.row, self.col, self.data,
--> 258 B.ravel('A'), fortran)
259 return B
260
ValueError: could not convert integer scalar
I'm using Python 3.4.3 (on Red Hat), scipy 0.18.1, and scikit-learn 0.18.1.
I tried the monkey patch suggested here but that didn't work.
Googling around I found a bugfix that apparently solved the same problem for other types of sparse matrices (like csr), but not for coo.
I've tried feeding DBSCAN a sparse radius neighborhood graph (instead of a feature matrix), as suggested here, but the same error happens.
I've tried HDBSCAN, but the same error happens.
How can I fix this or bypass this?
Even if the implementation would allow it, DBSCAN would probably yield bad results on such very high dimensional data (from a statistical point of view, because of the curse of dimensionality).
Instead I would advise you to use the TruncatedSVD class to reduce the dimensionality of your TF-IDF feature vectors down to 50 or 100 components and then to apply DBSCAN on the results.