topsbm / topsbm

Scikit-learn compatible Topic Modelling with Hierarchical Statistical Block Models (Gerlach, Peixoto and Altmann, 2018)

topsbm - A scikit-learn extension for Topic Models based on Stochastic Block Models

Martin Gerlach, Tiago P. Peixoto, and Eduardo G. Altmann, “A network approach to topic models,” Science Advances (2018)

Software ported to Scikit-learn format by the Sydney Informatics Hub at the University of Sydney.

topsbm's Issues

Consider a mode with non-hierarchical inference

Quoting Eduardo

Control over the hierarchy:

The non-hierarchical model we considered in our paper is obtained
calling gt.minimize_blockmodel_dl instead of
gt.minimize_nested_blockmodel_dl.

Instead of a set of solutions in different levels of the hierarchy, we
only obtain one solution. We would have to adapt the 'get_groups'
function in the following way:
- now, the solution of the hierarchical model (state) is projected onto
a given level in the hierachy (state_l)
- if we use the non-hierarchical model, we would skip the projection-part
We could also set the upper and lower limit for the total number of groups.
The advantage of this would be to have more control over the number of
groups and to not have to decide about which level in the hierarchy.
However, the main advantage of the hierarchical model is that it
provides a better prior increasing the resolution of small clusters
(keyword: 'resolution limit' in community detection).

Rename this repository to TopSBM

Starting with `fit_transform`

I think you should initially work on defining fit_transform. I would give it the following docstring:

def fit_transform(self, X, y=None):
    """Fit the hSBM topic model

    Constructs a graph representation of X, infers clustering, and reports
    the cluster probability for each sample in X.

    Parameters
    ----------
    X : ndarray or sparse matrix of shape (n_samples, n_features)
        Word frequencies for each document, represented as  non-negative
        integers.
    y : ignored

    Returns
    -------
    Xt : ndarray of shape (n_samples, n_components)
    """

This would do basically everything for the simple case of just getting a topic decomposition of some X. It would call private methods as appropriate to modularise the work. The main idea would be to encapsulate the logic for transformation as a function of numeric array X and any hyper-parameters provided in __init__.

An initial test would be:

import pytest
from sklearn.datasets import make_multilabel_classification
from hSBM import hSBMTransformer

@pytest.mark.parametrize('sparse', [False, True])
@pytest.mark.parametrize('n_components', [5, 10])
def test_basic_fit_transform(sparse, n_components):
    X, y = make_multilabel_classification(random_state=0, n_features=150)
    est = hSBMTransformer(n_components=n_components)
    Xt = est.fit_transform(X)
    assert Xt.shape == (X.shape[0], n_components)

def test_continuous_unacceptable():
    # hSBMTransformer should refuse to transform anything but integer data
    est = hSBMTransformer()
    with pytest.raises(ValueError):
        est.fit_transform(np.linspace(0, 1, 1000).reshape(20, 5))

Once this is done, we can further work out what we need to do to make the estimator more useful.

Fix Licence

Licence currently has incorrect copyright attribution, and incorrect licence. Licence must be GPL v3. Each file with substantive work should have a copyright notice plus:

    This file is part of TopSBM.

    TopSBM is free software: you can redistribute it and/or modify
    it under the terms of the GNU General Public License as published by
    the Free Software Foundation, either version 3 of the License, or
    (at your option) any later version.

    TopSBM is distributed in the hope that it will be useful,
    but WITHOUT ANY WARRANTY; without even the implied warranty of
    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
    GNU General Public License for more details.

    You should have received a copy of the GNU General Public License
    along with TopSBM.  If not, see <https://www.gnu.org/licenses/>.

Eventually want to put docs on readthedocs or similar

But only after we've got the name right

Rename transformer and the module

hSBMTransformer is:

• unconventional in its use of initial lowercase. Need uppercased camel case for Python classes
• this isn't just hSBM. hSBM is an algorithm over graphs.

Candidate names:

• HSBMTopicModel
• GerlachAltmannTransformer
• HSBMDecomposition
• GraphTopicModel
• BipartiteSBMTopicModel
• BipartiteSBMDecomposition

We need to ask the client about naming.

Document the structure and semantics of `groups_` better

Expose b_{min,max} to the user as {min,max}_groups

Does np.dot(p_tw_w, docs) = p_tw_d?

If so, it is trivial to implement transform.

transform to different levels of hierarchy

There are different groups (i.e. sets of topics) at different levels of the hierarchy. All of these should be available as transform targets.

By default we might transform redundantly to all levels of the hierarchy, or to the finest, and provide an option to change this (with a memoised model?).

Make sure random state can be set

We need a parameter random_state that controls randomization with a numpy RandomState. This requires investigating how graph_tool performs random number generation or allows this to be controlled.

See if there are any simple ways to speed up processing

It is very slow.

Add/refine tests

Add n_jobs parameter that temporarily uses openmp_set_num_threads

perhaps expand groups_ to multiple attributes of consistent type/semantics

Expose the output of inference in a more accessible way

Add n_init param to control multiple initialisations

See https://github.com/martingerlach/hSBM_Topicmodel/blob/2ba7658ac513b10d9bc2b1c2a63dd6512449e7ca/sbmtm.py#L145-L153

Topics of a single document.

Can I also run topSBM on a single document?

I am not interested in how topics are distributed over different documents. I just simply want to extract the main topics per document.

I know that the pd.DataFrame(model.groups_[1]['p_tw_d'], columns=titles) gives a matrix with how much each topic correspond per document, but this topic is based on words from all the documents....

Any suggestions?

Make sure docs compile and show example appropriately

Eventually we should have them continuously deployed, with either Circle CI or readthedocs.

Recompile examples automatically and check they are consistent with expectations

This would be a valuable end-to-end test beyond the unit test suite.

Unfortunately, compiling the example on readthedocs results in a timeout.

Add flake8 to CI and make sure code conforms to PEP8

Add an example to the TopSBM docstring

Store the number of topics in n_components_ after fitting

Use nbsphinx instead of sphinx-gallery to precompute results

Running examples kills readthedocs as it runs out of memory. We'll need to precompute the example results :(

rename variables l and L to level and n_levels

Deploy to pypi eventually

(Need to get the name right (#3) first)

Add a conda environment.yml

avoid indvidual configuration

Evaluation metrics?

Is there any way I could measure how well topsbm is doing quantitatively on my data except for just eyeballing the topics returned?

Rewrite the example ipynb to use the new interface

Add badges to readme

conda-forge now has graph-tool

The conda installation can be simplified now that conda-forge has a graph-tool build on it: https://anaconda.org/conda-forge/graph-tool. This may better support cross-platform installs (still no windows), but should be tested.

Add OSX instances to travis

Deploy to conda

(Need to get the name right (#3) first)

Change docs CSS to improve display of parameter/attrib types

Sphinx/docutils produce spans with "delimiter" and "classifier" classes in dt nodes in HTML. These are not styled in the rtd default stylesheet.

Consider a more scalable inference algorithm

There are problems with both memory and time consumption of the implementation. We could consider implementing the variational inference described at https://arxiv.org/pdf/1711.05150.pdf. I've previously found an R implementation of this, but have now lost it.

Test the contents of groups_