On 7 February 2017 at 15:19, Kanchana Ranasinghe ***@***.***> wrote: Can I try working on this? I am new though. — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#8306 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAEz64YNY5YHNCJWjtB8I88z6xFHqX1nks5rZ_C_gaJpZM4L5Arq> .

On 7 Feb 2017 7:19 pm, "Kanchana Ranasinghe" ***@***.***> wrote: But scipy.spatial.distance.is_valid_dm() verifies this as an acceptable distance matrix. I even tried plotting the corresponding points and got: [0.1008,0.2193],[1.6069,0.0202],[-0.5538,-0.1653],[0. 2431,-0.1523],[-1.3969,0.0781] However, distance matrices (at least for Euclidean) aren't allowed to have zero values off the diagonal, are they? — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#8306 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAEz62G_wPmeIK_XsR1cGAdPC3c_RgABks5raCj9gaJpZM4L5Arq> .

On 8 Feb 2017 12:10 am, "Alexey Grachev" ***@***.***> wrote: Let I be more concrete. Problem is specific to sparse matrix. Error happens if and only if the first line in matrix is empty. Example sparse matrix synthetically created from dense input for easy reproduction. All zero-values there are treated as non-existent elements. Incorrect behavior is in this part of code https://github.com/scikit-learn/scikit-learn/blob/14031f6/sklearn/cluster/ dbscan_.py#L122 if metric == 'precomputed' and sparse.issparse(X): neighborhoods = np.empty(X.shape[0], dtype=object) X.sum_duplicates() # XXX: modifies X's internals in-place X_mask = X.data <= eps masked_indices = astype(X.indices, np.intp, copy=False)[X_mask] masked_indptr = np.cumsum(X_mask)[X.indptr[1:] - 1] # <<== UNVALIDATED SHIFT-DECREMENT # insert the diagonal: a point is its own neighbor, but 0 distance # means absence from sparse matrix data masked_indices = np.insert(masked_indices, masked_indptr, np.arange(X.shape[0])) masked_indptr = masked_indptr[:-1] + np.arange(1, X.shape[0]) # split into rows neighborhoods[:] = np.split(masked_indices, masked_indptr) In sparce matrix row indexer first element is always zero. If first row is empty then second element in row indexer is also zero. There is shift-decrement operation in code. After this operation second element becomes first with value equal -1 and thus points to the tail of column index. This treats first row as neighbor of all other rows. In my project I reimplemented your dbscan() function with this code if metric == 'precomputed' and sparse.issparse(X): X.sum_duplicates() # XXX: modifies X's internals in-place X.data = (X.data <= eps).astype(np.intp) X.eliminate_zeros() # Convert to linked list for better performance and to use built-in .rows object X_lil = X.tolil() # insert the diagonal: a point is its own neighbor. X_lil.setdiag(1) # split into rows neighborhoods[:] = X_lil.rows I doubt it's the best option, but it works correctly and don't look like some David Blane street magic. — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#8306 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAEz6zgL1S5LMsmg3H4zLrXXSZnzAU0sks5raG1RgaJpZM4L5Arq> .