Skip to main content

Ensemble topic modelling with pLSA

Project description

EnsTop provides an ensemble based approach to topic modelling using pLSA. It makes use of a high performance numba based pLSA implementation to run multiple bootstrapped topic models in parallel, and then clusters the resulting outputs to determine a set of stable topics. It can then refit the document vectors against these topics embed documents into the stable topic space.

Why use EnsTop?

There are a number of advantages to using an ensemble approach to topic modelling. The most obvious is that it produces better more stable topics. A close second, however, is that, by making use of HDBSCAN for clustering topics, it can learn a “natural” number of topics. That is, while the user needs to specify an estimated number of topics, the actual number of topics produced will be determined by how many stable topics are produced over many bootstrapped runs. In practice this can either be more, or less, than the estimated number of topics.

Despite all of these extra features the ensemble topic approach is still very efficient, especially in multi-core environments (due the the embarrassingly parallel nature of the ensemble). A run with a reasonable size ensemble can be completed in around the same time it might take to fit an LDA model, and usually produces superior quality results.

In addition to this EnsTop comes with a pLSA implementation that can be used standalone (and not as part of an ensemble). So if all you are loosing for is a good fast pLSA implementation (that can run considerably faster than many LDA implementations) then EnsTop is the library for you.

How to use EnsTop

EnsTop follows the sklearn API (and inherits from sklearn base classes), so if you use sklearn for LDA or NMF then you already know how to use Enstop. General usage is very straightforward. The following example uses EnsTop to model topics from the classic 20-Newsgroups dataset, using sklearn’s CountVectorizer to generate the required count matrix.

from sklearn.datasets import fetch_20newsgroups
from sklearn.feature_extraction.text import CountVectorizer
from enstop import EnsembleTopics

news = fetch_20newsgroups(subset='all')
data = CountVectorizer().fit_transform(news.data)

model = EnsembleTopics(n_components=20).fit(data)
topics = model.components_
doc_vectors = model.embedding_

How to use pLSA

EnsTop also provides a simple to use but fast and effective pLSA implementation out of the box. As with the ensemble topic modeller it follows the sklearn API, and usage is very similar.

from sklearn.datasets import fetch_20newsgroups
from sklearn.feature_extraction.text import CountVectorizer
from enstop import PLSA

news = fetch_20newsgroups(subset='all')
data = CountVectorizer().fit_transform(news.data)

model = PLSA(n_components=20).fit(data)
topics = model.components_
doc_vectors = model.embedding_

Installation

The easiest way to install EnsTop is via pip

pip install enstop

To manually install this package:

wget https://github.com/lmcinnes/enstop/archive/master.zip
unzip master.zip
rm master.zip
cd enstop-master
python setup.py install

Help and Support

Some basic example notebooks are available here.

Documentation is coming. This project is still very young. If you need help, or have problems please open an issue and I will try to provide any help and guidance that I can. Please also check the docstrings on the code, which provide some descriptions of the parameters.

License

The EnsTop package is 2-clause BSD licensed.

Contributing

Contributions are more than welcome! There are lots of opportunities for potential projects, so please get in touch if you would like to help out. Everything from code to notebooks to examples and documentation are all equally valuable so please don’t feel you can’t contribute. To contribute please fork the project make your changes and submit a pull request. We will do our best to work through any issues with you and get your code merged into the main branch.

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

enstop-0.1.5.tar.gz (19.6 kB view details)

Uploaded Source

File details

Details for the file enstop-0.1.5.tar.gz.

File metadata

  • Download URL: enstop-0.1.5.tar.gz
  • Upload date:
  • Size: 19.6 kB
  • Tags: Source
  • Uploaded using Trusted Publishing? No
  • Uploaded via: Python-urllib/3.8

File hashes

Hashes for enstop-0.1.5.tar.gz
Algorithm Hash digest
SHA256 4965a72dd536079fc7bc87ed4012d5cf5895959e414812ccd984bfa9fae96421
MD5 8a10573d16c75db74c5554486aaebad6
BLAKE2b-256 a4c1a6230d0b0e7ad4cebc3d3902f4bd82afd6985613349a95f305f934b37e9a

See more details on using hashes here.

Supported by

AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Pingdom Monitoring Sentry Error logging StatusPage Status page