Topic Modeling Using Graphlab



王成军

wangchengjun@nju.edu.cn

计算传播网 http://computational-communication.com

In [2]:
import graphlab
graphlab.canvas.set_target("ipynb")
In [3]:
sf = graphlab.SFrame.read_csv("/Users/datalab/bigdata/w15", 
                              header=False)
This non-commercial license of GraphLab Create for academic use is assigned to wangchengjun@nju.edu.cn and will expire on March 14, 2019.
[INFO] graphlab.cython.cy_server: GraphLab Create v2.1 started. Logging: /tmp/graphlab_server_1525423344.log
Finished parsing file /Users/datalab/bigdata/w15
Parsing completed. Parsed 100 lines in 0.693838 secs.
------------------------------------------------------
Inferred types from first 100 line(s) of file as 
column_type_hints=[str]
If parsing fails due to incorrect types, you can correct
the inferred type list above and pass it to read_csv in
the column_type_hints argument
------------------------------------------------------
Read 12278 lines. Lines per second: 11539.9
Finished parsing file /Users/datalab/bigdata/w15
Parsing completed. Parsed 72269 lines in 2.39912 secs.
In [4]:
sf
Out[4]:
X1
aynrand born and educated
in russia rand migrated ...
asphalt in american
english asphalt or ...
actinopterygii the
actinopterygii consti ...
altaiclanguages these
language families share ...
argon the name argon is
derived from the greek ...
augustderleth a 1938
guggenheim fellow der ...
amateur amateurism can be
seen in both a negative ...
assemblyline an assembly
line is a manufacturing ...
astronomicalunit an
astronomical unit ...
abbess an abbess latin
abbatissa feminine form ...
[72269 rows x 1 columns]
Note: Only the head of the SFrame is printed.
You can use print_rows(num_rows=m, num_columns=n) to print more rows and columns.

Transformations

In [5]:
dir(sf['X1'])
Out[5]:
['_SArray__check_min_observations',
 '_SArray__construct_ctr',
 '__abs__',
 '__add__',
 '__and__',
 '__bool__',
 '__class__',
 '__contains__',
 '__copy__',
 '__deepcopy__',
 '__delattr__',
 '__div__',
 '__doc__',
 '__eq__',
 '__floordiv__',
 '__format__',
 '__ge__',
 '__get_content_identifier__',
 '__getattribute__',
 '__getitem__',
 '__gt__',
 '__has_size__',
 '__hash__',
 '__init__',
 '__is_materialized__',
 '__iter__',
 '__le__',
 '__len__',
 '__lt__',
 '__materialize__',
 '__mod__',
 '__module__',
 '__mul__',
 '__ne__',
 '__neg__',
 '__new__',
 '__nonzero__',
 '__or__',
 '__pos__',
 '__pow__',
 '__proxy__',
 '__radd__',
 '__rdiv__',
 '__reduce__',
 '__reduce_ex__',
 '__repr__',
 '__rfloordiv__',
 '__rmod__',
 '__rmul__',
 '__rpow__',
 '__rsub__',
 '__rtruediv__',
 '__setattr__',
 '__sizeof__',
 '__slots__',
 '__str__',
 '__sub__',
 '__subclasshook__',
 '__truediv__',
 '_count_ngrams',
 '_count_words',
 '_getitem_cache',
 '_save_as_text',
 'all',
 'any',
 'append',
 'apply',
 'argmax',
 'argmin',
 'astype',
 'clip',
 'clip_lower',
 'clip_upper',
 'contains',
 'cumulative_max',
 'cumulative_mean',
 'cumulative_min',
 'cumulative_std',
 'cumulative_sum',
 'cumulative_var',
 'date_range',
 'datetime_to_str',
 'dict_has_all_keys',
 'dict_has_any_keys',
 'dict_keys',
 'dict_trim_by_keys',
 'dict_trim_by_values',
 'dict_values',
 'dropna',
 'dtype',
 'fillna',
 'filter',
 'from_avro',
 'from_const',
 'from_sequence',
 'hash',
 'head',
 'is_in',
 'is_materialized',
 'item_length',
 'materialize',
 'max',
 'mean',
 'min',
 'nnz',
 'num_missing',
 'pixel_array_to_image',
 'random_integers',
 'random_split',
 'rolling_count',
 'rolling_max',
 'rolling_mean',
 'rolling_min',
 'rolling_stdv',
 'rolling_sum',
 'rolling_var',
 'sample',
 'save',
 'shape',
 'show',
 'size',
 'sketch_summary',
 'sort',
 'split_datetime',
 'std',
 'str_to_datetime',
 'subslice',
 'sum',
 'tail',
 'to_numpy',
 'topk_index',
 'unique',
 'unpack',
 'var',
 'vector_slice',
 'where']
In [6]:
bow = sf['X1']._count_words()
In [7]:
type(sf['X1'])
Out[7]:
graphlab.data_structures.sarray.SArray
In [8]:
type(bow)
Out[8]:
graphlab.data_structures.sarray.SArray
In [9]:
bow.dict_has_any_keys(['limited'])
Out[9]:
dtype: int
Rows: 72269
[1, 0, 0, 0, 1, 0, 0, 0, 0, 0, 1, 0, 1, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 1, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, ... ]
In [10]:
bow.dict_values()[0][:20]
Out[10]:
[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 1]
In [11]:
sf
Out[11]:
X1
aynrand born and educated
in russia rand migrated ...
asphalt in american
english asphalt or ...
actinopterygii the
actinopterygii consti ...
altaiclanguages these
language families share ...
argon the name argon is
derived from the greek ...
augustderleth a 1938
guggenheim fellow der ...
amateur amateurism can be
seen in both a negative ...
assemblyline an assembly
line is a manufacturing ...
astronomicalunit an
astronomical unit ...
abbess an abbess latin
abbatissa feminine form ...
[72269 rows x 1 columns]
Note: Only the head of the SFrame is printed.
You can use print_rows(num_rows=m, num_columns=n) to print more rows and columns.
In [12]:
sf['bow'] = bow
In [13]:
sf
Out[13]:
X1 bow
aynrand born and educated
in russia rand migrated ...
{'limited': 3,
'writings': 2, ...
asphalt in american
english asphalt or ...
{'all': 1, 'accadian': 1,
'similarity': 1, ...
actinopterygii the
actinopterygii consti ...
{'andreolepis': 1, 'all':
1, 'evolutionary': 2, ...
altaiclanguages these
language families share ...
{'sergei': 3, 'all': 6,
'todays': 1, 'chinese': ...
argon the name argon is
derived from the greek ...
{'limited': 1,
'embolism': 1, ...
augustderleth a 1938
guggenheim fellow der ...
{'evelyn': 1,
'detective': 4, ...
amateur amateurism can be
seen in both a negative ...
{'since': 1, 'subpar': 1,
'lack': 2, 'valuable' ...
assemblyline an assembly
line is a manufacturing ...
{'all': 3, 'concept': 6,
'consider': 1, 'chine ...
astronomicalunit an
astronomical unit ...
{'precise': 1, 'all': 2,
'chinese': 1, 'suns': 1, ...
abbess an abbess latin
abbatissa feminine form ...
{'kildares': 1, 'they':
4, 'founder': 1, ...
[72269 rows x 2 columns]
Note: Only the head of the SFrame is printed.
You can use print_rows(num_rows=m, num_columns=n) to print more rows and columns.
In [14]:
type(sf['bow'])
Out[14]:
graphlab.data_structures.sarray.SArray
In [15]:
len(sf['bow'])
Out[15]:
72269
In [16]:
sf['bow'][0].items()[:5]
Out[16]:
[('limited', 3),
 ('writings', 2),
 ('personally', 1),
 ('four', 1),
 ('controversial', 1)]
In [17]:
sf['tfidf'] = graphlab.text_analytics.tf_idf(sf['X1'])
In [18]:
sf
Out[18]:
X1 bow tfidf
aynrand born and educated
in russia rand migrated ...
{'limited': 3,
'writings': 2, ...
{'limited':
10.04705669672047, ...
asphalt in american
english asphalt or ...
{'all': 1, 'accadian': 1,
'similarity': 1, ...
{'all':
1.3891905239989626, ...
actinopterygii the
actinopterygii consti ...
{'andreolepis': 1, 'all':
1, 'evolutionary': 2, ...
{'andreolepis':
11.188150547181156, ...
altaiclanguages these
language families share ...
{'sergei': 3, 'all': 6,
'todays': 1, 'chinese': ...
{'sergei':
20.031873121992916, ...
argon the name argon is
derived from the greek ...
{'limited': 1,
'embolism': 1, ...
{'limited':
3.3490188989068232, ...
augustderleth a 1938
guggenheim fellow der ...
{'evelyn': 1,
'detective': 4, ...
{'evelyn':
6.7937013925087175, ...
amateur amateurism can be
seen in both a negative ...
{'since': 1, 'subpar': 1,
'lack': 2, 'valuable' ...
{'since':
1.8775124538896095, ...
assemblyline an assembly
line is a manufacturing ...
{'all': 3, 'concept': 6,
'consider': 1, 'chine ...
{'all':
4.167571571996888, ...
astronomicalunit an
astronomical unit ...
{'precise': 1, 'all': 2,
'chinese': 1, 'suns': 1, ...
{'precise':
5.491057060675752, 'a ...
abbess an abbess latin
abbatissa feminine form ...
{'kildares': 1, 'they':
4, 'founder': 1, ...
{'kildares':
11.188150547181156, ...
[72269 rows x 3 columns]
Note: Only the head of the SFrame is printed.
You can use print_rows(num_rows=m, num_columns=n) to print more rows and columns.
In [19]:
sf['tfidf'][0].items()[:5]
Out[19]:
[('limited', 10.04705669672047),
 ('writings', 9.76010421134325),
 ('personally', 5.001941923280662),
 ('four', 2.1272386886969024),
 ('controversial', 4.375805453003677)]

Text cleaning

In [20]:
docs = sf['bow'].dict_trim_by_values(2)
In [21]:
docs = docs.dict_trim_by_keys(
    graphlab.text_analytics.stopwords(),
    exclude=True)

Topic modeling

In [27]:
help(graphlab.topic_model.create)
Help on function create in module graphlab.toolkits.topic_model.topic_model:

create(dataset, num_topics=10, initial_topics=None, alpha=None, beta=0.1, num_iterations=10, num_burnin=5, associations=None, verbose=False, print_interval=10, validation_set=None, method='auto')
    Create a topic model from the given data set. A topic model assumes each
    document is a mixture of a set of topics, where for each topic some words
    are more likely than others. One statistical approach to do this is called a
    "topic model". This method learns a topic model for the given document
    collection.
    
    Parameters
    ----------
    dataset : SArray of type dict or SFrame with a single column of type dict
        A bag of words representation of a document corpus.
        Each element is a dictionary representing a single document, where
        the keys are words and the values are the number of times that word
        occurs in that document.
    
    num_topics : int, optional
        The number of topics to learn.
    
    initial_topics : SFrame, optional
        An SFrame with a column of unique words representing the vocabulary
        and a column of dense vectors representing
        probability of that word given each topic. When provided,
        these values are used to initialize the algorithm.
    
    alpha : float, optional
        Hyperparameter that controls the diversity of topics in a document.
        Smaller values encourage fewer topics per document.
        Provided value must be positive. Default value is 50/num_topics.
    
    beta : float, optional
        Hyperparameter that controls the diversity of words in a topic.
        Smaller values encourage fewer words per topic. Provided value
        must be positive.
    
    num_iterations : int, optional
        The number of iterations to perform.
    
    num_burnin : int, optional
        The number of iterations to perform when inferring the topics for
        documents at prediction time.
    
    verbose : bool, optional
        When True, print most probable words for each topic while printing
        progress.
    
    print_interval : int, optional
        The number of iterations to wait between progress reports.
    
    associations : SFrame, optional
        An SFrame with two columns named "word" and "topic" containing words
        and the topic id that the word should be associated with. These words
        are not considered during learning.
    
    validation_set : SArray of type dict or SFrame with a single column
        A bag of words representation of a document corpus, similar to the
        format required for `dataset`. This will be used to monitor model
        performance during training. Each document in the provided validation
        set is randomly split: the first portion is used estimate which topic
        each document belongs to, and the second portion is used to estimate
        the model's performance at predicting the unseen words in the test data.
    
    method : {'cgs', 'alias'}, optional
        The algorithm used for learning the model.
    
        - *cgs:* Collapsed Gibbs sampling
        - *alias:* AliasLDA method.
    
    Returns
    -------
    out : TopicModel
        A fitted topic model. This can be used with
        :py:func:`~TopicModel.get_topics()` and
        :py:func:`~TopicModel.predict()`. While fitting is in progress, several
        metrics are shown, including:
    
        +------------------+---------------------------------------------------+
        |      Field       | Description                                       |
        +==================+===================================================+
        | Elapsed Time     | The number of elapsed seconds.                    |
        +------------------+---------------------------------------------------+
        | Tokens/second    | The number of unique words processed per second   |
        +------------------+---------------------------------------------------+
        | Est. Perplexity  | An estimate of the model's ability to model the   |
        |                  | training data. See the documentation on evaluate. |
        +------------------+---------------------------------------------------+
    
    See Also
    --------
    TopicModel, TopicModel.get_topics, TopicModel.predict,
    graphlab.SArray.dict_trim_by_keys, TopicModel.evaluate
    
    References
    ----------
    - `Wikipedia - Latent Dirichlet allocation
      <http://en.wikipedia.org/wiki/Latent_Dirichlet_allocation>`_
    
    - Alias method: Li, A. et al. (2014) `Reducing the Sampling Complexity of
      Topic Models. <http://www.sravi.org/pubs/fastlda-kdd2014.pdf>`_.
      KDD 2014.
    
    Examples
    --------
    The following example includes an SArray of documents, where
    each element represents a document in "bag of words" representation
    -- a dictionary with word keys and whose values are the number of times
    that word occurred in the document:
    
    >>> docs = graphlab.SArray('https://static.turi.com/datasets/nytimes')
    
    Once in this form, it is straightforward to learn a topic model.
    
    >>> m = graphlab.topic_model.create(docs)
    
    It is also easy to create a new topic model from an old one  -- whether
    it was created using GraphLab Create or another package.
    
    >>> m2 = graphlab.topic_model.create(docs, initial_topics=m['topics'])
    
    To manually fix several words to always be assigned to a topic, use
    the `associations` argument. The following will ensure that topic 0
    has the most probability for each of the provided words:
    
    >>> from graphlab import SFrame
    >>> associations = SFrame({'word':['hurricane', 'wind', 'storm'],
                               'topic': [0, 0, 0]})
    >>> m = graphlab.topic_model.create(docs,
                                        associations=associations)
    
    More advanced usage allows you  to control aspects of the model and the
    learning method.
    
    >>> import graphlab as gl
    >>> m = gl.topic_model.create(docs,
                                  num_topics=20,       # number of topics
                                  num_iterations=10,   # algorithm parameters
                                  alpha=.01, beta=.1)  # hyperparameters
    
    To evaluate the model's ability to generalize, we can create a train/test
    split where a portion of the words in each document are held out from
    training.
    
    >>> train, test = gl.text_analytics.random_split(.8)
    >>> m = gl.topic_model.create(train)
    >>> results = m.evaluate(test)
    >>> print results['perplexity']

In [29]:
help(graphlab.text_analytics.random_split)
Help on function random_split in module graphlab.toolkits.text_analytics._util:

random_split(dataset, prob=0.5)
    Utility for performing a random split for text data that is already in
    bag-of-words format. For each (word, count) pair in a particular element,
    the counts are uniformly partitioned in either a training set or a test
    set.
    
    Parameters
    ----------
    dataset : SArray of type dict, SFrame with columns of type dict
        A data set in bag-of-words format.
    
    prob : float, optional
        Probability for sampling a word to be placed in the test set.
    
    Returns
    -------
    train, test : SArray
        Two data sets in bag-of-words format, where the combined counts are
        equal to the counts in the original data set.
    
    Examples
    --------
    >>> docs = graphlab.SArray([{'are':5, 'you':3, 'not': 1, 'entertained':10}])
    >>> train, test = graphlab.text_analytics.random_split(docs)
    >>> print train
    [{'not': 1.0, 'you': 3.0, 'are': 3.0, 'entertained': 7.0}]
    >>> print test
    [{'are': 2.0, 'entertained': 3.0}]

In [31]:
train, test = graphlab.text_analytics.random_split(docs, .8)
In [33]:
m = graphlab.topic_model.create(train, 
                                num_topics=100,       # number of topics
                                num_iterations=100,   # algorithm parameters
                                alpha=None, beta=.1)  # hyperparameters
Learning a topic model
       Number of documents     72269
           Vocabulary size    107782
   Running collapsed Gibbs sampling
+-----------+---------------+----------------+-----------------+
| Iteration | Elapsed Time  | Tokens/Second  | Est. Perplexity |
+-----------+---------------+----------------+-----------------+
| 10        | 2.03s         | 5.015e+06      | 0               |
| 20        | 3.86s         | 5.37519e+06    | 0               |
| 30        | 5.64s         | 5.12154e+06    | 0               |
| 40        | 7.49s         | 5.30373e+06    | 0               |
| 50        | 9.29s         | 5.00983e+06    | 0               |
| 60        | 11.22s        | 5.00837e+06    | 0               |
| 70        | 13.12s        | 4.94135e+06    | 0               |
| 80        | 15.02s        | 4.60452e+06    | 0               |
| 90        | 16.96s        | 4.7366e+06     | 0               |
| 100       | 18.93s        | 4.84249e+06    | 0               |
+-----------+---------------+----------------+-----------------+
In [34]:
results = m.evaluate(test)
print results['perplexity']
4527.13385741
In [23]:
m
Out[23]:
Class                          : TopicModel

Schema
------
Vocabulary Size                : 171005

Settings
--------
Number of Topics               : 10
alpha                          : 5.0
beta                           : 0.1
Iterations                     : 10
Training time                  : 3.6698
Verbose                        : False

Accessible fields             : 
m['topics']                   : An SFrame containing the topics.
m['vocabulary']               : An SArray containing the words in the vocabulary.
Useful methods                : 
m.get_topics()                : Get the most probable words per topic.
m.predict(new_docs)           : Make predictions for new documents.
In [37]:
m.get_topics()
Out[37]:
topic word score
0 time 0.00422209137382
0 2001 0.00201402456976
0 worked 0.00181329122394
0 gonzales 0.00174638010866
0 stunt 0.00167946899339
1 god 0.0134496294644
1 century 0.0111281287498
1 temple 0.00998245307251
1 people 0.00690721835975
1 great 0.00648512732074
[500 rows x 3 columns]
Note: Only the head of the SFrame is printed.
You can use print_rows(num_rows=m, num_columns=n) to print more rows and columns.
In [39]:
help(m.get_topics)
Help on method get_topics in module graphlab.toolkits.topic_model.topic_model:

get_topics(self, topic_ids=None, num_words=5, cdf_cutoff=1.0, output_type='topic_probabilities') method of graphlab.toolkits.topic_model.topic_model.TopicModel instance
    Get the words associated with a given topic. The score column is the
    probability of choosing that word given that you have chosen a
    particular topic.
    
    Parameters
    ----------
    topic_ids : list of int, optional
        The topics to retrieve words. Topic ids are zero-based.
        Throws an error if greater than or equal to m['num_topics'], or
        if the requested topic name is not present.
    
    num_words : int, optional
        The number of words to show.
    
    cdf_cutoff : float, optional
        Allows one to only show the most probable words whose cumulative
        probability is below this cutoff. For example if there exist
        three words where
    
        .. math::
           p(word_1 | topic_k) = .1
    
           p(word_2 | topic_k) = .2
    
           p(word_3 | topic_k) = .05
    
        then setting :math:`cdf_{cutoff}=.3` would return only
        :math:`word_1` and :math:`word_2` since
        :math:`p(word_1 | topic_k) + p(word_2 | topic_k) <= cdf_{cutoff}`
    
    output_type : {'topic_probabilities' | 'topic_words'}, optional
        Determine the type of desired output. See below.
    
    Returns
    -------
    out : SFrame
        If output_type is 'topic_probabilities', then the returned value is
        an SFrame with a column of words ranked by a column of scores for
        each topic. Otherwise, the returned value is a SArray where
        each element is a list of the most probable words for each topic.
    
    Examples
    --------
    Get the highest ranked words for all topics.
    
    >>> docs = graphlab.SArray('https://static.turi.com/datasets/nips-text')
    >>> m = graphlab.topic_model.create(docs,
                                        num_iterations=50)
    >>> m.get_topics()
    +-------+----------+-----------------+
    | topic |   word   |      score      |
    +-------+----------+-----------------+
    |   0   |   cell   |  0.028974400831 |
    |   0   |  input   | 0.0259470208503 |
    |   0   |  image   | 0.0215721599763 |
    |   0   |  visual  | 0.0173635081992 |
    |   0   |  object  | 0.0172447874156 |
    |   1   | function | 0.0482834508265 |
    |   1   |  input   | 0.0456270024091 |
    |   1   |  point   | 0.0302662839454 |
    |   1   |  result  | 0.0239474934631 |
    |   1   | problem  | 0.0231750116011 |
    |  ...  |   ...    |       ...       |
    +-------+----------+-----------------+
    
    Get the highest ranked words for topics 0 and 1 and show 15 words per
    topic.
    
    >>> m.get_topics([0, 1], num_words=15)
    +-------+----------+------------------+
    | topic |   word   |      score       |
    +-------+----------+------------------+
    |   0   |   cell   |  0.028974400831  |
    |   0   |  input   | 0.0259470208503  |
    |   0   |  image   | 0.0215721599763  |
    |   0   |  visual  | 0.0173635081992  |
    |   0   |  object  | 0.0172447874156  |
    |   0   | response | 0.0139740298286  |
    |   0   |  layer   | 0.0122585145062  |
    |   0   | features | 0.0115343177265  |
    |   0   | feature  | 0.0103530459301  |
    |   0   | spatial  | 0.00823387994361 |
    |  ...  |   ...    |       ...        |
    +-------+----------+------------------+
    
    If one wants to instead just get the top words per topic, one may
    change the format of the output as follows.
    
    >>> topics = m.get_topics(output_type='topic_words')
    dtype: list
    Rows: 10
    [['cell', 'image', 'input', 'object', 'visual'],
     ['algorithm', 'data', 'learning', 'method', 'set'],
     ['function', 'input', 'point', 'problem', 'result'],
     ['model', 'output', 'pattern', 'set', 'unit'],
     ['action', 'learning', 'net', 'problem', 'system'],
     ['error', 'function', 'network', 'parameter', 'weight'],
     ['information', 'level', 'neural', 'threshold', 'weight'],
     ['control', 'field', 'model', 'network', 'neuron'],
     ['hidden', 'layer', 'system', 'training', 'vector'],
     ['component', 'distribution', 'local', 'model', 'optimal']]

In [40]:
topics = m.get_topics(num_words=10).unstack(['word','score'], \
                                new_column_name='topic_words')['topic_words'].apply(lambda x: x.keys())
for topic in topics:
    print topic
['goldstone', '1990', 'school', 'group', 'district', 'list', 'arnold', 'glass', 'hercules', 'time']
['foundation', 'full', 'bar', 'charles', 'trees', 'research', 'men', 'small', '2008', 'walk']
['water', 'plant', 'made', 'food', 'traditional', 'small', 'iron', 'popular', 'found', 'called']
['sault', 'wisconsin', 'lewis', 'percent', 'number', 'sierra', 'call', 'time', 'campbell', '2008']
['jones', 'bar', 'commission', '1966', '1971', 'arbroath', 'engineering', 'represented', 'served', 'pangasinan']
['states', 'bars', 'pcbs', 'start0', 'high', 'visitors', 'part', 'including', 'plotdata', 'sharp']
['school', 'made', 'group', 'singapore', 'malaysian', 'years', 'malaysia', '2007', 'association', 'left']
['division', 'force', 'army', 'german', 'men', 'general', 'forces', 'battle', 'military', 'war']
['miami', 'gonzales', 'due', 'height', 'greek', '2001', 'time', 'stunt', 'worked', 'ha']
['city', 'francisco', 'san', 'chicago', 'angeles', 'los', 'california', 'york', 'time', 'served']
['foundation', 'lankan', 'scales', 'sri', 'fitzgerald', 'due', 'rinpoche', 'time', 'medicine', 'lanka']
['album', 'released', 'song', 'recorded', 'tour', 'band', 'single', 'music', 'records', 'songs']
['building', 'city', 'north', 'built', 'area', 'west', 'park', 'town', 'street', 'road']
['swedish', 'danish', 'made', 'finnish', 'denmark', 'finland', 'norwegian', 'sweden', 'time', 'norway']
['lessons', 'feet', 'lee', 'bell', 'national', '2002', 'society', 'part', 'originally', 'williams']
['championship', 'champion', 'title', 'wrestling', 'won', 'team', 'world', 'ring', 'round', 'match']
['operations', 'training', 'united', 'force', 'group', 'aircraft', 'base', 'flight', 'air', 'war']
['europe', 'school', 'club', 'phillip', 'alan', 'roberts', 'phillips', 'valid', 'beth', 'springfield']
['division', 'set', 'skanderbeg', 'university', 'member', 'street', 'college', 'cloud', 'fleetwood', 'red']
['1991', 'city', 'major', 'morley', 'cj', 'village', 'pipe', 'child', 'davis', '2010']
['tara', 'cameras', 'award', 'illinois', 'lens', 'camera', 'master', '1936', 'part', 'red']
['house', 'dudley', 'award', 'mcphee', 'part', '5', 'long', 'brunswick', 'babylon', 'centauri']
['television', 'network', 'show', 'tv', 'broadcast', 'station', 'radio', 'news', 'local', 'channel']
['head', '1996', 'national', 'point', 'wilson', 'years', 'bernard', 'points', '2007', 'hat']
['school', 'made', 'brothers', 'clark', 'wright', 'club', 'rico', 'served', 'puerto', '2010']
['group', 'wallace', 'post', 'company', 'tradition', 'place', 'members', 'model', 'john', 'korean']
['naval', 'island', 'ships', 'japanese', 'coast', 'sea', 'islands', 'navy', 'fleet', 'ship']
['united', 'people', 'german', 'national', 'government', 'political', 'states', 'union', 'state', 'war']
['salford', 'head', 'orlando', 'long', 'youth', 'born', 'street', 'manchester', 'listed', 'call']
['dna', 'cells', 'proteins', 'site', 'cell', 'enzyme', 'protein', 'acid', 'called', 'structure']
['canada', 'toronto', 'ontario', 'iowa', 'university', 'canadian', 'hospital', 'health', 'dr', 'medical']
['blue', 'named', 'gmina', 'wisconsin', '1979', '1997', 'country', 'broken', 'wayne', 'martin']
['copper', 'mining', 'oil', 'gold', 'company', 'gas', 'mine', 'mines', 'coal', 'part']
['des', 'la', 'paris', 'louis', 'de', 'french', 'france', 'le', 'jean', 'du']
['hong', 'korea', 'chinese', 'modern', 'japanese', 'dynasty', 'china', 'kong', 'japan', 'emperor']
['mr', 'smith', 'smiths', 'years', 'project', 'famous', 'state', 'london', 'allen', 'plates']
['robert', 'regional', 'award', 'director', 'connecticut', 'wood', 'place', 'international', 'burns', 'university']
['brazil', 'city', 'portuguese', 'mexico', 'la', 'de', 'spanish', 'el', 'spain', 'italian']
['great', 'filipino', 'marcos', 'philippines', 'philippine', 'including', 'manila', 'college', 'post', 'university']
['simon', 'stores', 'rights', 'british', '2000', '2001', 'ghostbusters', 'closed', 'worked', 'store']
['named', 'family', '1981', 'rose', 'cat', 'born', 'grid', 'rangers', 'beaver', 'george']
['ukraine', 'ukrainian', 'moscow', 'war', 'late', 'part', 'polish', 'russian', 'austrian', 'poland']
['made', 'working', 'michigan', 'act', 'york', 'time', 'half', 'walker', 'soft', 'bond']
['town', 'khmer', '1995', 'rescue', 'unknown', 'national', 'year', 'joined', 'rouge', 'williams']
['engine', 'built', 'engines', 'models', 'car', 'power', 'high', 'production', 'design', 'model']
['pakistan', 'khan', 'district', 'preov', 'region', 'india', 'indian', 'bystrica', 'singh', 'koice']
['king', 'great', 'ancient', 'century', 'god', 'people', 'jesus', 'found', 'temple', 'called']
['cornish', 'van', 'village', 'cornwall', 'cats', 'harrison', 'part', '2008', '2009', '2010']
['200', 'university', 'miller', 'national', 'southern', 'british', 'years', 'professor', '2009', 'originally']
['irish', 'parade', 'northern', 'dublin', 'cork', 'county', '2006', 'ireland', 'senior', 'local']
['television', 'theatre', 'show', 'appeared', 'movie', 'production', 'role', 'series', 'films', 'film']
['zealand', 'cricket', 'australia', 'england', 'played', 'made', 'wales', 'australian', 'day', 'south']
['shown', 'group', 'groups', 'number', 'site', 'society', '2', 'members', 'action', 'jon']
['made', 'end', 'began', 'year', 'years', 'early', 'including', 'time', 'continued', 'called']
['named', 'america', 'michigan', 'top', 'malta', 'hayes', 'design', 'maltese', 'john', 'side']
['jones', 'school', 'carter', 'journal', 'st', 'member', 'pa', 'college', 'years', 'crawford']
['study', 'disease', 'research', 'patients', 'health', 'blood', 'treatment', 'found', 'children', 'effects']
['1984', 'dean', 'miss', 'mildred', 'location', '2005', 'claimed', 'position', 'dr', 'mode']
['october', 'march', 'september', 'june', '1', '2', '2007', '2008', '2009', '2010']
['lao', 'weld', 'torture', 'forum', 'club', 'years', 'jordan', '4', 'miranda', 'orange']
['star', 'center', 'launch', 'sun', 'space', 'mission', 'moon', 'solar', 'time', 'earth']
['king', 'england', 'family', 'sir', 'married', 'son', 'william', 'london', 'john', 'died']
['town', 'city', 'named', 'episode', '1994', 'tree', '1974', 'carroll', 'group', '2009']
['government', 'general', 'member', 'elected', 'state', 'election', 'council', 'president', 'party', 'minister']
['story', 'episode', 'released', 'series', 'character', 'player', 'game', 'version', 'characters', 'original']
['management', 'financial', 'business', 'corporation', 'company', 'million', 'companies', 'market', 'sold', 'bank']
['life', 'death', 'love', 'family', 'time', 'father', 'back', 'mother', 'tells', 'man']
['championship', 'won', 'series', 'car', 'team', 'points', 'time', 'race', 'grand', 'racing']
['major', 'dance', 'opera', 'work', 'orchestra', 'festival', 'music', 'performed', 'works', 'musical']
['city', 'famous', 'minnesota', 'met', 'minister', 'martin', 'served', 'women', 'guide', 'engineer']
['service', 'evans', 'fire', 'university', 'stamp', 'stamps', '2007', 'duck', 'served', 'postal']
['florida', 'family', 'moved', 'hurricane', 'damage', 'tropical', 'gang', 'storm', 'texas', 'york']
['case', 'united', 'court', 'legal', 'states', 'state', 'members', 'act', 'police', 'law']
['nfl', 'goal', 'season', 'football', 'bowl', 'field', 'game', 'yards', 'pass', 'quarter']
['city', 'german', 'west', 'von', 'munich', 'mark', 'berlin', 'germany', 'east', 'syracuse']
['life', 'theory', 'people', 'work', 'research', 'society', 'human', 'social', 'world', 'history']
['arts', 'art', 'house', 'museum', 'work', 'style', 'collection', 'design', 'artists', 'works']
['states', 'united', 'washington', 'virginia', 'county', 'american', 'black', 'york', 'white', 'war']
['level', 'union', 'cemetery', 'professor', 'hammond', 'kitty', 'porky', 'youth', 'university', 'stage']
['living', 'family', 'people', '18', 'age', 'median', 'years', 'income', 'average', 'population']
['league', 'played', 'cup', 'club', 'season', 'football', 'game', 'won', 'games', 'team']
['olympic', 'won', 'gold', 'events', 'held', 'championships', 'world', 'international', 'medal', 'event']
['species', 'family', 'long', 'order', 'large', 'black', 'small', 'found', 'white', 'birds']
['mountain', 'valley', 'north', 'creek', 'area', 'park', 'lake', 'water', 'river', 'south']
['israel', 'jews', 'palestinian', 'jewish', 'romanian', 'camp', 'israeli', 'years', 'palestine', '1939']
['catholic', 'christian', 'saint', 'century', 'parish', 'st', 'roman', 'council', 'church', 'bishop']
['horses', 'dates', '1994', 'family', 'horse', 'link', 'year', 'orange', 'called', 'preferences']
['school', 'students', 'university', 'research', 'high', 'program', 'college', 'year', 'schools', 'education']
['garden', 'form', '1981', 'crow', 'point', 'ii', 'middle', 'japan', 'nerman', 'cotton']
['iran', 'ali', 'muslim', 'al', 'years', 'arab', 'islamic', 'muhammad', 'muslims', 'islam']
['body', 'city', 'de', 'junior', 'beer', 'kent', 'year', 'east', 'open', 'brewery']
['life', 'work', 'books', 'story', 'writing', 'magazine', 'book', 'stories', 'published', 'wrote']
['sound', 'register', 'club', 'house', 'hotel', 'county', 'historic', 'listed', 'home', 'properties']
['service', 'route', 'lines', 'station', 'london', 'trains', 'services', 'railway', 'line', 'class']
['school', 'fair', 'rogers', 'dallas', 'work', 'andrews', 'time', 'eugene', 'sidney', 'left']
['development', 'government', 'energy', 'countries', 'system', 'trade', 'states', 'economic', 'world', 'national']
['letters', 'language', 'die', 'der', 'languages', 'written', 'names', 'words', 'english', 'word']
['information', 'code', 'users', 'technology', 'system', 'computer', 'systems', 'internet', 'data', 'software']
['case', 'function', 'set', 'form', 'light', 'energy', 'number', 'surface', '1', 'method']
['town', 'city', 'reverse', 'people', 'coins', 'list', '1933', 'green', 'date', 'coin']
In [26]:
m['vocabulary']
Out[26]:
dtype: str
Rows: 171005
['duke', 'studies', 'journal', 'chris', 'research', 'matthew', 'crisis', 'financial', 'paul', '1987', 'reagan', 'traditional', 'rightwing', 'nominee', 'libertarianism', 'cato', 'chief', 'smith', 'line', 'south', 'nick', '1999', 'documentary', 'animated', 'shows', 'references', 'commentator', 'powerful', 'ethics', 'rush', 'neil', 'lives', 'cited', 'produced', 'night', 'originality', 'interest', '2007', 'individual', 'authors', 'admirer', 'married', 'club', 'library', 'essays', 'recent', '2009', 'burns', 'inspiration', 'artist', 'women', 'early', 'barbara', 'organized', 'gave', 'referred', 'company', 'personalist', 'criticism', 'john', 'reviewers', 'language', 'understanding', 'writes', 'fewer', 'attention', 'positive', 'masterful', 'review', 'times', 'critic', 'praise', 'theory', 'randian', 'importance', 'calling', 'nonfiction', 'academics', 'kant', 'philosophers', 'italian', 'remarked', 'wife', 'house', 'subject', 'scholarly', 'edward', 'system', 'influence', 'acknowledged', '100', 'branden', 'criticized', 'sacrificing', 'exist', 'selfinterest', 'rational', 'communism', 'journals', 'copies', ... ]
In [41]:
m['topics']
Out[41]:
topic_probabilities vocabulary
[6.69111152745e-06,
3.01493599291e-06, ...
affair
[6.69111152745e-06,
0.000515554054787, ...
writings
[6.69111152745e-06,
3.01493599291e-06, ...
collectivism
[6.69111152745e-06,
3.01493599291e-06, ...
rosenbaum
[6.69111152745e-06,
3.01493599291e-06, ...
reviewers
[0.000140513342076,
3.01493599291e-06, ...
rest
[6.69111152745e-06,
3.01493599291e-06, ...
years
[6.69111152745e-06,
3.01493599291e-06, ...
produced
[6.69111152745e-06,
0.000123612375709, ...
held
[6.69111152745e-06,
3.01493599291e-06, ...
including
[107782 rows x 2 columns]
Note: Only the head of the SFrame is printed.
You can use print_rows(num_rows=m, num_columns=n) to print more rows and columns.
In [46]:
def print_topics(m):
    topics = m.get_topics(num_words=5)
    topics = topics.unstack(['word','score'], new_column_name='topic_words')['topic_words']
    topics = topics.apply(lambda x: x.keys())
    for topic in topics:
        print topic
print_topics(m)
['1990', 'list', 'group', 'school', 'district']
['small', '2008', 'charles', 'bar', 'full']
['food', 'water', 'made', 'called', 'iron']
['campbell', 'lewis', '2008', 'wisconsin', 'time']
['represented', 'arbroath', 'bar', '1971', 'served']
['high', 'sharp', 'start0', 'part', 'visitors']
['school', 'group', '2007', 'association', 'singapore']
['battle', 'division', 'war', 'forces', 'army']
['gonzales', 'worked', '2001', 'time', 'stunt']
['angeles', 'city', 'california', 'san', 'los']
['lanka', 'rinpoche', 'sri', 'fitzgerald', 'time']
['album', 'band', 'song', 'music', 'released']
['town', 'city', 'street', 'road', 'area']
['swedish', 'norwegian', 'sweden', 'made', 'time']
['society', 'national', 'part', 'lee', 'bell']
['title', 'world', 'championship', 'match', 'team']
['force', 'aircraft', 'base', 'flight', 'air']
['club', 'phillip', 'school', 'beth', 'roberts']
['member', 'university', 'set', 'college', 'red']
['davis', 'city', '2010', 'morley', 'village']
['part', 'camera', 'master', 'red', 'award']
['part', 'brunswick', '5', 'dudley', 'award']
['news', 'station', 'radio', 'network', 'show']
['head', 'years', 'hat', 'national', 'wilson']
['puerto', 'wright', 'brothers', 'clark', 'rico']
['place', 'model', 'group', 'post', 'members']
['island', 'ship', 'ships', 'sea', 'islands']
['state', 'people', 'political', 'war', 'government']
['salford', 'born', 'street', 'manchester', 'orlando']
['cell', 'cells', 'dna', 'enzyme', 'protein']
['canada', 'hospital', 'health', 'medical', 'canadian']
['country', 'named', 'gmina', 'wisconsin', 'martin']
['coal', 'mining', 'oil', 'gas', 'gold']
['paris', 'la', 'de', 'french', 'france']
['japan', 'emperor', 'china', 'japanese', 'chinese']
['project', 'famous', 'state', 'smith', 'mr']
['robert', 'wood', 'place', 'university', 'award']
['spanish', 'city', 'de', 'mexico', 'la']
['university', 'including', 'philippine', 'college', 'philippines']
['simon', '2001', 'stores', 'ghostbusters', 'british']
['rose', 'named', 'george', 'family', '1981']
['russian', 'polish', 'poland', 'part', 'war']
['walker', 'bond', 'time', 'made', 'act']
['town', 'national', 'joined', 'williams', 'rouge']
['engine', 'models', 'model', 'production', 'power']
['pakistan', 'region', 'indian', 'india', 'district']
['god', 'great', 'people', 'temple', 'century']
['2008', '2009', '2010', 'cornwall', 'village']
['miller', 'national', '2009', 'university', 'years']
['county', 'irish', 'dublin', 'northern', 'ireland']
['films', 'series', 'role', 'film', 'show']
['zealand', 'wales', 'australian', 'australia', 'south']
['2', 'group', 'number', 'members', 'site']
['early', 'years', 'made', 'end', 'time']
['maltese', 'michigan', 'america', 'side', 'malta']
['jones', 'member', 'school', 'carter', 'years']
['found', 'patients', 'blood', 'study', 'disease']
['miss', 'position', 'dr', 'location', 'mode']
['2008', '2009', 'june', '2007', '2010']
['lao', 'weld', '4', 'torture', 'years']
['earth', 'time', 'star', 'mission', 'space']
['son', 'king', 'john', 'died', 'sir']
['episode', 'named', '2009', 'carroll', 'group']
['elected', 'president', 'state', 'election', 'party']
['episode', 'game', 'character', 'characters', 'series']
['market', 'company', 'million', 'business', 'bank']
['life', 'mother', 'father', 'back', 'man']
['car', 'race', 'racing', 'won', 'team']
['dance', 'festival', 'opera', 'orchestra', 'music']
['city', 'met', 'women', 'minnesota', 'engineer']
['fire', 'served', 'stamps', 'service', 'evans']
['tropical', 'texas', 'storm', 'family', 'york']
['states', 'police', 'state', 'law', 'court']
['field', 'game', 'yards', 'season', 'pass']
['german', 'city', 'berlin', 'germany', 'von']
['work', 'society', 'social', 'theory', 'history']
['design', 'arts', 'work', 'art', 'museum']
['states', 'american', 'united', 'white', 'york']
['youth', 'professor', 'kitty', 'university', 'stage']
['years', '18', 'age', 'population', 'income']
['club', 'league', 'game', 'team', 'season']
['world', 'championships', 'medal', 'won', 'gold']
['small', 'white', 'black', 'species', 'family']
['water', 'river', 'north', 'lake', 'area']
['jewish', 'israel', 'camp', 'israeli', 'jews']
['catholic', 'roman', 'st', 'bishop', 'church']
['horses', 'orange', 'horse', 'link', 'year']
['students', 'school', 'education', 'college', 'university']
['middle', 'japan', 'point', 'garden', 'cotton']
['arab', 'islamic', 'islam', 'al', 'ali']
['body', 'city', 'beer', 'brewery', 'year']
['work', 'magazine', 'book', 'books', 'published']
['register', 'home', 'hotel', 'historic', 'house']
['line', 'station', 'london', 'railway', 'service']
['school', 'left', 'time', 'fair', 'eugene']
['development', 'trade', 'economic', 'system', 'government']
['languages', 'word', 'words', 'language', 'english']
['information', 'data', 'system', 'systems', 'software']
['function', 'energy', 'set', 'number', 'form']
['date', 'city', 'coins', 'green', 'people']
In [25]:
pred = m.predict(another_data) # do not run
pred = m.predict(another_data, output_type='probabilities')

Initializing from other models

In [30]:
m2 = graphlab.topic_model.create(docs,
                                 num_topics=10,
                                 initial_topics=m['topics'])
Initializing from provided topics and vocabulary.
Learning a topic model
       Number of documents     72269
           Vocabulary size    171005
   Running collapsed Gibbs sampling
+-----------+---------------+----------------+-----------------+
| Iteration | Elapsed Time  | Tokens/Second  | Est. Perplexity |
+-----------+---------------+----------------+-----------------+
| 10        | 2.41s         | 8.85404e+06    | 0               |
+-----------+---------------+----------------+-----------------+

Seeding the model with prior knowledge

In [31]:
associations = graphlab.SFrame()
associations['word'] = ['recognition']
associations['topic'] = [0]
In [32]:
m2 = graphlab.topic_model.create(docs,
                                 num_topics=20,
                                 num_iterations=50,
                                 associations=associations, 
                                 verbose=False)
Learning a topic model
       Number of documents     72269
           Vocabulary size    171005
   Running collapsed Gibbs sampling
+-----------+---------------+----------------+-----------------+
| Iteration | Elapsed Time  | Tokens/Second  | Est. Perplexity |
+-----------+---------------+----------------+-----------------+
| 10        | 2.75s         | 8.25271e+06    | 0               |
| 20        | 5.15s         | 8.35785e+06    | 0               |
| 30        | 7.50s         | 7.61449e+06    | 0               |
| 40        | 9.90s         | 7.93222e+06    | 0               |
| 50        | 12.24s        | 8.52982e+06    | 0               |
+-----------+---------------+----------------+-----------------+
In [33]:
m2.get_topics(num_words=10)
Out[33]:
topic word score
0 court 0.0177612734598
0 king 0.0127195614005
0 police 0.0107441537516
0 war 0.00842083848682
0 people 0.00717072976569
0 city 0.00638056670611
0 emperor 0.00618007757159
0 order 0.00594125963194
0 made 0.00587934386981
0 french 0.0055638683199
[200 rows x 3 columns]
Note: Only the head of the SFrame is printed.
You can use print_rows(num_rows=m, num_columns=n) to print more rows and columns.
In [34]:
print_topics(m2)
['king', 'court', 'war', 'police', 'people']
['information', 'aircraft', 'network', 'service', 'system']
['states', 'region', 'united', 'state', 'government']
['la', 'de', 'india', 'indian', 'france']
['military', 'army', 'force', 'war', 'air']
['model', 'set', 'number', 'power', 'system']
['large', 'found', 'small', 'species', 'family']
['club', 'league', 'football', 'year', 'season']
['party', 'company', 'election', 'council', 'served']
['town', 'age', 'years', 'school', 'population']
['work', 'band', 'book', 'art', 'published']
['students', 'university', 'national', 'college', 'state']
['death', 'father', 'time', 'family', 'son']
['season', 'team', 'final', 'won', 'played']
['city', 'line', 'river', 'road', 'area']
['roman', 'church', 'people', 'language', 'century']
['years', 'john', 'york', 'race', 'time']
['album', 'released', 'music', 'film', 'song']
['world', 'game', 'championship', 'games', 'team']
['engine', 'production', 'company', 'made', 'design']