Topic modeling

.travis.yml

@@ -5,6 +5,7 @@ python:
 install:
 - pip install ".[testing]"
 - pip install ".[nlp]"
+- pip install ".[topic_modeling]"
 - python -m nltk.downloader punkt stopwords wordnet
 script: pytest
 deploy:

quantgov/estimator/candidate_sets.py

@@ -17,9 +17,22 @@
 import sklearn.multioutput
 import sklearn.pipeline
 import sklearn.feature_extraction
+from . import structures
+
+try:
+    import gensim
+except ImportError:
+    gensim = None
+try:
+    import gensim
+    import spacy
+except ImportError:
+    spacy = None
+    gensim = None
 
 import quantgov.estimator
 
+
 classification = [
     quantgov.estimator.CandidateModel(
         name="Random Forests",
@@ -69,3 +82,16 @@
         }
     ),
 ]
+
+if gensim and spacy:
+    topic_modeling = [
+        quantgov.estimator.CandidateModel(
+            name="LDA",
+            model=structures.QGLdaModel(),
+            parameters={
+                'eta': [0.1, 0.05, 0.01],
+                'passes': [1, 2, 3],
+                'num_topics': [10, 50, 100]
+            }
+        ),
+    ]

quantgov/estimator/structures.py

@@ -5,6 +5,34 @@
 """
 import collections
 import joblib as jl
+from sklearn.base import BaseEstimator, TransformerMixin
+from six import iteritems
+from decorator import decorator
+import re
+
+try:
+    from spacy.lang.en.stop_words import STOP_WORDS
+    from gensim.corpora import Dictionary
+    from gensim import sklearn_api
+    import gensim
+    spacy = True
+except ImportError:
+    spacy = None
+    gensim = None
+
+
+@decorator
+def check_spacy(func, *args, **kwargs):
+    if spacy is None:
+        raise RuntimeError('Must install spacy to use {}'.format(func))
+    return func(*args, **kwargs)
+
+
+@decorator
+def check_gensim(func, *args, **kwargs):
+    if gensim is None:
+        raise RuntimeError('Must install gensim to use {}'.format(func))
+    return func(*args, **kwargs)
 
 
 class _PersistanceMixin(object):
@@ -85,3 +113,41 @@ class CandidateModel(
             parameter values to test as values
     """
     pass
+
+
+class QGLdaModel(BaseEstimator, TransformerMixin):
+    @check_gensim
+    @check_spacy
+    def __init__(self, word_regex=r'\b[A-z]{2,}\b', stop_words=STOP_WORDS):
+        self.stop_words = stop_words
+        self.word_regex = re.compile(word_regex)
+
+    def transform(self, driver):
+        self.test_corpus = self.create_corpus(driver)
+        return self.model.transform(self.test_corpus)
+
+    def create_corpus(self, driver):
+        return [self.dictionary.doc2bow([i.group(0).lower()
+                for i in self.word_regex.finditer(doc.text)])
+                for doc in driver.stream()]
+
+    def fit(self, driver, alpha=None, eta=None, num_topics=1, passes=1):
+        self.dictionary = Dictionary([[i.group(0).lower()
+                                      for i in self.word_regex
+                                        .finditer(doc.text)]
+                                      for doc in driver.stream()])
+        stop_ids = [self.dictionary.token2id[stopword] for stopword
+                    in self.stop_words if stopword in self.dictionary.token2id]
+        once_ids = [tokenid for tokenid, docfreq in
+                    iteritems(self.dictionary.dfs) if docfreq == 1]
+        self.dictionary.filter_tokens(stop_ids + once_ids)
+        self.corpus = self.create_corpus(driver)
+        self.model = sklearn_api.ldamodel.LdaTransformer(
+            alpha=alpha,
+            eta=eta,
+            num_topics=num_topics,
+            passes=passes,
+            id2word=self.dictionary
+        )
+        self.model.fit(self.corpus)
+        return self

setup.py

@@ -65,6 +65,10 @@ def find_version(*file_paths):
         'nlp': [
             'textblob',
             'nltk',
+        ],
+        'topic_modeling': [
+            'gensim',
+            'spacy'
         ]
     },
     entry_points={

tests/test_estimators.py

@@ -0,0 +1,22 @@
+# import pytest
+import subprocess
+import quantgov.estimator
+import quantgov
+
+from pathlib import Path
+
+PSEUDO_CORPUS_PATH = Path(__file__).resolve().parent.joinpath('pseudo_corpus')
+driver = quantgov.load_driver(PSEUDO_CORPUS_PATH)
+
+
+def test_topic_model():
+    sample = quantgov.estimator.structures.QGLdaModel()
+    sample.fit(driver, num_topics=2)
+    sample.transform(driver)
+
+
+def check_output(cmd):
+    return (
+        subprocess.check_output(cmd, universal_newlines=True)
+        .replace('\n\n', '\n')
+    )