Topic Modeling on PyCaret

I remember a brief conversation with my boss’ boss a while back. He said that he wouldn’t be impressed if somebody in the company built a face recognition tool from scratch because, and I quote, “Guess what? There’s an API for that.” He then goes on about the futility of doing something that’s already been done instead of just using it.

This gave me an insight into how an executive thinks. Not that they don’t care about the coolness factor of a project, but at the end of that day, they’re most concerned about how a project will add value to the business and even more importantly, how quickly it can be done.

In the real world, the time it takes to build prototype matters. And the quicker we get from data to insights, the better off we will be. These help us stay agile.

And this brings me to PyCaret.


PyCaret is an open source, low-code machine learning library in Python that allows you to go from preparing your data to deploying your model within seconds in your choice of notebook environment.[1]

Pycaret is basically a wrapper for some of the most popular machine learning libraries and frameworks scikit-learn and spaCy. Here are the things that PyCaret can do:

  • Classification
  • Regression
  • Clustering
  • Anomaly Detection
  • Natural Language Processing
  • Associate Rule Mining

If you’re interested in reading about the difference between traditional NLP approach vs. PyCaret’s NLP module, check out Prateek Baghel’s article.

Natural Language Processing

In just a few lines of code, PyCaret makes natural language processing so easy that it’s almost criminal. Like most of its other modules, PyCaret’s NLP module streamlined pipeline cuts the time from data to insights in more than half the time.

For example, with only one line, it performs text processing automatically, with the ability to customize stop words. Add another line or two, and you got yourself a language model. With yet another line, it gives you a properly formatted plotly graph. And finally, adding another line gives you the option to evaluate the model. You can even tune the model with, guess what, one line of code!

Instead of just telling you all about the wonderful features of PyCaret, maybe it’s be better if we do a little show and tell instead.


The Pipeline

For this post, we’ll create an NLP pipeline that involves the following 6 glorious steps:

  1. Getting the Data
  2. Setting up the Environment
  3. Creating the Model
  4. Assigning the Model
  5. Plotting the Model
  6. Evaluating the Model

We will be going through an end-to-end demonstration of this pipeline with a brief explanation of the functions involved and their parameters.

Let’s get started.


Housekeeping

Let us begin by installing PyCaret. If this is your first time installing it, just type the following into your terminal:

pip install pycaret

However, if you have a previously installed version of PyCaret, you can upgrade using the following command:

pip install —-upgrade pycaret

Beware: PyCaret is a big library so it’s going to take a few minutes to download and install.

We’ll also need to download the English language model because it is not included in the PyCaret installation.

python -m spacy download en_core_web_sm
python -m textblob.download_corpora

Next, let’s fire up a Jupyter notebook and import PyCaret’s NLP module:

#import nlp module
from pycaret.nlp import *

Importing the pycaret.nlp automatically sets up your environment to perform NLP tasks only.

Getting the Data

Before setup, we need to decide first how we’re going to ingest data. There are two methods of getting the data into the pipeline. One is by using a Pandas dataframe and another is by using a simple list of textual data.

Passing a DataFrame

#import pandas if we're gonna use a dataframe
import pandas as pd

# load the data into a dataframe
df = pd.read_csv('hilaryclinton.csv')

Above, we’re simply loading the data into a dataframe.

Passing a List

# read a file containing a list of text data and assign it to 'lines'
with open('list.txt') as f:
    lines = f.read().splitlines()

Above, we’re opening the file 'list.txt' and reading it. We assign the resulting list into the lines.

Sampling

From the rest of this experiment, we’ll just use a dataframe to pass textual data to thesetup() function of the NLP module. And for the sake of expediency, we’ll sample the dataframe to only select a thousand tweets.

# sampling the data to select only 1000 tweets
df = df.sample(1000, random_state=493).reset_index(drop=True)

Let’s take a quick look at our dataframe with df.head() and df.shape.

Setting Up the Environment

In the line below, we’ll initialize the setup by calling the setup() function and assign it to nlp.

# initialize the setup
nlp = setup(data = df, target = 'text', session_id = 493, custom_stopwords = [ 'rt', 'https', 'http', 'co', 'amp'])

With data and target, we’re telling PyCaret that we’d like to use the values in the 'text' column of df. Also, we’re setting the session_id to an arbitrary number of 493 so that we can reproduce the experiment over and over again and get the same result. Finally, we added custom_stopwords so that PyCaret will exclude the specified list of words in the analysis.

Note that if we want to use a list instead, we could replace df with lines and get rid of target = ‘text’ because a list has no columns for the PyCaret to target!

Here’s the output of nlp:

The output table above confirms our session id, number of documents (rows or records), and vocabulary size. It also shows up if we used custom stopwords or not.

Creating the Model

Below, we’ll create the model by calling the create_model() function and assign it to lda. The function already knows to use the dataset that we specified during setup(). In our case, the PyCaret knows we want to create a model based on the 'text' in df.

# create the model
lda = create_model('lda', num_topics = 6, multi_core = True)

In the line above, notice that w param used 'lda' as the parameter. LDA stands for Latent Dirichlet Allocation. We could’ve just as easily opted for other types of models.

Here’s the list of models that PyCaret currently supports:

  • ‘lda’: Latent Dirichlet Allocation
  • ‘lsi’: Latent Semantic Indexing
  • ‘hdp’: Hierarchical Dirichlet Process
  • ‘rp’: Random Projections
  • ‘nmf’: Non-Negative Matrix Factorization

I encourage you to research the difference between the models above, To start, check out Lettier’s awesome guide on LDA.

The next parameter we used is num_topics = 6. This tells PyCaret to use six topics in the results ranging from 0 to 5. If num_topic is not set, the default number is 4. Lastly, we set multi_core to tell PyCaret to use all available CPUs for parallel processing. This saves a lot of computational time.

Assigning the Model

By calling assign_model(), we’re going to label our data so that we’ll get a dataframe (based on our original dataframe: df) with additional columns that include the following information:

  • Topic percent value for each topic
  • The dominant topic
  • The percent value of the dominant topic
# label the data using trained model
df_lda = assign_model(lda)

Let’s take a look at df_lda.

Plotting the Model

Calling the plot_model() function will give us some visualization about frequency, distribution, polarity, et cetera. The plot_model() function takes three parameters: model, plot, and topic_num. The model instructs PyCaret what model to use and must be preceded by a create_model() function. topic_num designates which topic number (from 0 to 5) will the visualization be based on.

plot_model(lda, plot='topic_distribution')
plot_model(lda, plot='topic_model')
plot_model(lda, plot='wordcloud', topic_num = 'Topic 5')
plot_model(lda, plot='frequency', topic_num = 'Topic 5')
plot_model(lda, plot='bigram', topic_num = 'Topic 5')
plot_model(lda, plot='trigram', topic_num = 'Topic 5')
plot_model(lda, plot='distribution', topic_num = 'Topic 5')
plot_model(lda, plot='sentiment', topic_num = 'Topic 5')
plot_model(lda, plot='tsne')

PyCarets offers a variety of plots. The type of graph generated will depend on the plot parameter. Here is the list of currently available visualizations:

  • ‘frequency’: Word Token Frequency (default)
  • ‘distribution’: Word Distribution Plot
  • ‘bigram’: Bigram Frequency Plot
  • ‘trigram’: Trigram Frequency Plot
  • ‘sentiment’: Sentiment Polarity Plot
  • ‘pos’: Part of Speech Frequency
  • ‘tsne’: t-SNE (3d) Dimension Plot
  • ‘topic_model’ : Topic Model (pyLDAvis)
  • ‘topic_distribution’ : Topic Infer Distribution
  • ‘wordcloud’: Word cloud
  • ‘umap’: UMAP Dimensionality Plot

Evaluating the Model

Evaluating the models involves calling the evaluate_model() function. It takes only one parameter: the model to be used. In our case, the model is stored is lda that was created with the create_model() function in an earlier step.

The function returns a visual user interface for plotting.

And voilà, we’re done!

Conclusion

Using PyCaret’s NLP module, we were able to quickly from getting the data to evaluating the model in just a few lines of code. We covered the functions involved in each step and examined the parameters of those functions.


Thank you for reading! PyCaret’s NLP module has a lot more features and I encourage you to read their documentation to further familiarize yourself with the module and maybe even the whole library!

In the next post, I’ll continue to explore PyCaret’s functionalities.

If you want to learn more about my journey from slacker to data scientist, check out the article here.

Stay tuned!

You can reach me on Twitter or LinkedIn.


[1] PyCaret. (June 4, 2020). Why PyCaret. https://pycaret.org/Towards Data

Drop It Like It’s Hot

I have a recurring dream where my instructor from a coding boot camp would constantly beat my head with a ruler telling me to read a package or library’s documentation. Hence, as a past time, I would find myself digging into Python or Panda’s documentation.

Today, I found myself wandering into pandas’ .drop() function. So, in this post, I shall attempt to make sense of panda’s documentation for the ever famous .drop().


Housekeeping

Let’s import pandas and create a sample dataframe.

import pandas as pd

data = {'fname': ['Priyanka', 'Jane', 'Sarah', 'Jake', 'Tatum', 'Shubham', 'Antonio'],
        'color': ['Red', 'Orange', 'Yellow', 'Green', 'Blue', 'Indigo', 'Violet'],
        'value': [0, 1, 2, 3, 5, 8, 13],
        'score': [345, 778, 124, 554, 864, 908, 456]
       }

df = pd.DataFrame(data)

If we type df into a cell in Jupyter notebook, this will give us the whole dataframe:

One-level DataFrame Operations

Now let’s get rid of some columns.

df.drop(['color', 'score'], axis=1)

The code above simply tells Python to get rid of the 'color' and 'score' in axis=1 which means look in the columns. Alternatively, we could’ve just as easily not used the named parameter axis because it’s confusing. So, let’s try that now:

df.drop(columns=['color', 'score'])

Both of the methods above will result in the following:

Next, we’ll get rid of some rows (or records).

df.drop([1, 2, 4, 6])

Above, we’re simply telling Python to get rid of the rows with the index of 1, 2, 4, and 6. Note that the indices are passed as a list [1, 2, 4, 6]. This will result in the following:

MultiIndex DataFrame Operations

In this next round, we’re going to work with a multi-index dataframe. Let’s set it up:

data = pd.MultiIndex(levels=[['slim jim', 'avocado', 'banana', 'pork rinds'],
                             ['carbs', 'fat', 'protein']],
                     codes=[[0, 0, 0, 1, 1, 1, 2, 2, 2, 3, 3, 3],
                            [0, 1, 2, 0, 1, 2, 0, 1, 2, 0, 1, 2]])

df = pd.DataFrame(index=data, columns=['thinghy', 'majig'],
                  data=[[45, 30], [200, 100], [1.5, 1], [30, 20],
                        [250, 150], [1.5, 0.8], [320, 250],
                        [1, 0.8], [0.3, 0.2], [34.2, 56], [33, 45.1], [67.3, 98]])

This is how the multi-index dataframe looks like:

Now, let’s get rid of the 'thinghy' column with:

df.drop(columns='thinghy')

And this is what we get:

Next, let’s get rid of 'pork rinds' because I don’t like them:

df.drop(index='pork rinds', level=0)

And this is what we get:

And finally, let’s cut the fat:

df.drop(index='fat', level=1)

Above, level=1 simply means the second level (since the first level starts with 0). In this case, it’s the carbs, fat, and protein levels. By specifying index='fat', we’re telling Python to get rid of the fat in level=1.

Here’s what we get:

Staying Put

So far, with all the playing that we did, somehow, if we type df into a cell, the output that we’re going to get is the original dataframe without modifications. this is because all the changes that we’ve been making take effect only on the display.

But what if we want to make the changes permanent? Enter: inplace.

df.drop(index='fat', level=1, inplace=True)

Above, we added inplace=True in the parameter. This signals Python that we want the changes to be taken in place so that when we output df, this is what we’ll get:

We had permanently cut the fat off. LOL!


Thank you for reading! That’s it for today.

Stay tuned!

You can reach me on Twitter or LinkedIn.

Selecting Rows with .loc

As data scientists, we spent most of our time wrangling knee-deep in manipulating data using Pandas. In this post, we’ll be looking at the .loc property of Pandas to select rows based on some predefined conditions.

Let’s open up a Jupyter notebook, and let’s get wrangling!


The Data

We will be using the 311 Service Calls dataset¹ from the City of San Antonio Open Data website to illustrate how the different .loc techniques work.

Housekeeping

Before we get started, let’s do a little housekeeping first.

import pandas as pd

# to print out all the outputs
from IPython.core.interactiveshell import InteractiveShell
InteractiveShell.ast_node_interactivity = "all"

# set display options
pd.set_option('display.max_columns', None)
pd.set_option('display.max_rows', None)
pd.set_option('display.max_colwidth', -1)

Nothing fancy going on here. We’re just importing the mandatory Pandas library and setting the display options so that when we inspect our dataframe, the columns and rows won’t be truncated by Jupyter. We’re setting it up so that every output within a single cell is displayed and not just the last one.

def show_missing(df):
    """
    Return the total missing values and the percentage of
    missing values by column.
    """
    null_count = df.isnull().sum()
    null_percentage = (null_count / df.shape[0]) * 100
    empty_count = pd.Series(((df == ' ') | (df == '')).sum())
    empty_percentage = (empty_count / df.shape[0]) * 100
    nan_count = pd.Series(((df == 'nan') | (df == 'NaN')).sum())
    nan_percentage = (nan_count / df.shape[0]) * 100
    return pd.DataFrame({'num_missing': null_count, 'missing_percentage': null_percentage,
                         'num_empty': empty_count, 'empty_percentage': empty_percentage,
                         'nan_count': nan_count, 'nan_percentage': nan_percentage})

In the code above, we’re defining a function that will show us the number of missing or null values and their percentage.

Getting the Data

Let’s load the data into a dataframe.

Doing a quick df.head() we’ll see the first five rows of the data:

And df.info() will let us see the dtypes of the columns.

Then, show_missing(df) shows us if there are any missing values in the data.

Selecting rows where the column is null or not.

Let’s select rows where the 'Dept' column has null values and also filtering a dataframe where null values are excluded.

df['Dept'].value_counts(dropna=False)

df_null = df.loc[df['Dept'].isnull()]
df_null.head()
df_null.shape

df_notnull = df.loc[df['Dept'].notnull()]
df_notnull.head()
df_notnull.shape

First, we did a value count of the column ‘Dept’ column. The method .value_counts() returns a panda series listing all the values of the designated column and their frequency. By default, the method ignores NaN values and will not list it. However, if you include the parameter dropna=False it will include any NaN values in the result.

Next, the line df_null = df.loc[df['Dept'].isnull()] tells the computer to select rows in df where the column 'Dept' is null. The resulting dataframe is assigned to df_null , and all its rows will NaN as values in the ‘Dept’ column.

Similarly, the line df_notnull = df.loc[df['Dept'].notnull()] tells the computer to select rows in df where the column 'Dept' is not null. The resulting dataframe is assigned to df_notnull , and all its rows will not have any NaN as values in the ‘Dept’ column.

The general syntax for these two techniques are:

df_new = df_old.loc[df_old['Column Name'].isnull()]
df_new = df_old.loc[df_old['Column Name'].notnull()]

Selecting rows where the column is a specific value.

The 'Late (Yes/No)' column looks interesting. Let’s take a look at it!

df['Late (Yes/No)'].value_counts(dropna=False)

df_late = df.loc[df['Late (Yes/No)'] == 'YES']
df_late.head()
df_late.shape

df_notlate = df.loc[df['Late (Yes/No)'] == 'NO']
df_notlate.head()
df_notlate.shape

Again, we did a quick value count on the 'Late (Yes/No)' column. Then, we filtered for the cases that were late with df_late = df.loc[df['Late (Yes/No)'] == 'YES']. Similarly, we did the opposite by changing 'YES' to 'NO' and assign it to a different dataframe df_notlate.

The syntax is not much different from the previous example except the addition of == sign between the column and the value we want to compare. It basically asks, for every row, if the value on a particular column (left side) matches the value that we specified (right-side). If the match is True, it includes that row in the result. If the match is False, it ignores it.

Here’s the resulting dataframe for df_late:

And here’s the one for df_notlate:

The general syntax for this technique is:

df_new = df_old.loc[df_old['Column Name'] == 'some_value' ]

Selecting rows where the column is not a specific value.

We’ve learned how to select rows based on ‘yes’ and ‘no.’ But what if the values are not binary? For example, let’s look at the ‘Category’ column:

One hundred ninety-two thousand one hundred ninety-seven rows or records do not have a category assigned, but instead of NaN, empty, or null value, we get 'No Category' as the category itself. What if we want to filter these out? Enter: the != operator.

df.Category.value_counts(dropna=False)

df_categorized = df.loc[df['Category'] != 'No Category']
df_categorized.head()
df_categorized.shape

df_categorized.Category.value_counts(dropna=False)

As usual, we did customary value counts on the 'Category' column to see what we’re working with. Then, we created the df_categorized dataframe to include any records in the the df dataframe that don’t have 'No Category' as their value in the 'Category' column.

Here’s the result of doing a value count on the 'Category' column of the df_categorized dataframe:

As the screenshot above shows, the value counts retained everything but the ‘No Category.’

The general syntax for this technique is:

df_new = df_old.loc[df_old['Column Name'] != 'some_value' ]

Select rows based on multiple conditions.

Let’s consider the following columns, 'Late (Yes/No)' and 'CaseStatus':

What if we wanted to know which open cases right now are already passed their SLA (service level agreement)? We would need to use multiple conditions to filter the cases or rows in a new dataframe. Enter the & operator.

df_late_open = df.loc[(df['Late (Yes/No)'] == 'YES') & (df['CaseStatus'] == 'Open')]

df_late_open.head()
df_late_open.shape

The syntax is similar to the previous ones except for the introduction of the & operator in between parenthesis. In the line df_late_open = df.loc[(df[‘Late (Yes/No)’] == ‘YES’) & (df[‘CaseStatus’] == ‘Open’)], there are two conditions:

  1. (df[‘Late (Yes/No)’] == ‘YES’)
  2. (df[‘CaseStatus’] == ‘Open’)

We want both of these to be true to match a row, so we included the operator & in between them. In plain speak, the & bitwise operator simply means AND. Other bitwise operators include pipe| sign for OR and the tilde ~ for NOT. I encourage you to experiment using these bitwise operators to get a good feel of what all they can do. Just remember to enclose each condition between parenthesis so that you don’t confuse Python.

The general syntax for this technique is:

df_new = df_old.loc[(df_old['Column Name 1'] == 'some_value_1') & (df['Column Name 2'] == 'some_value_2')]

Select rows having a column value that belongs in some list of values.

Let’s look at the value count for the 'Council District' column:

What if we wanted to focus on districts #2, #3, #4, and #5 because they’re in south San Antonio, and they’re known for getting poor service from the city? (I’m so totally making this up by the way!) In this case, we could use the .isin() method like so:

df['Council District'].value_counts(dropna=False)

df_south = df.loc[df['Council District'].isin([2,3,4,5])]
df_south.head()
df_south.shape

df_south['Council District'].value_counts()

Remember to pass your choices inside the .isin() method as a list like ['choice1', 'choice2', 'choice3'] because otherwise, it will cause an error. For integers like in our example, it is not necessary to include quotation marks because quotation marks are for string values only.

Here’s the result of our new dataframe df_south:

The general syntax for this technique is:

df_new = df_old.loc[df_old[Column Name'].isin(['choice1', 'choice2', 'choice3'])]

Conclusion

And that’s it! In this post, we loaded the 311 service calls data into a dataframe and created subsets of data using the .loc method.


Thanks for reading! I hope you enjoyed today’s post. Data wrangling, at least for me, is a fun exercise because this is the phase where I first get to know the data and it gives me a chance to hone my problem-solving skills when faced with really messy data. Happy wrangling folks!

Stay tuned!

You can reach me on Twitter or LinkedIn.

[1] City of San Antonio Open Data. (May 31, 2020). 311 Service Calls. https://data.sanantonio.gov/dataset/service-calls

Into the Heart of Darkness - Pt. 2

Exploring the Trump Twitter Archive with spaCy.


In a previous post, we set out to explore the dataset provided by the Trump Twitter Archive. My initial goal was to do something fun by using a very interesting dataset. However, it didn’t quite turn out that way.

On this post, we’ll continue our journey but this time we’ll be using spaCy.


For this project, we’ll be using pandas for data manipulation, spaCy for natural language processing, and joblib to speed things up.

Let’s get started by firing up a Jupyter notebook!

Housekeeping

Let’s import pandas and also set the display options so Jupyter won’t truncate our columns and rows. Let’s also set a random seed for reproducibility.

# for manipulating data
import pandas as pd
# setting the random seed for reproducibility
import random
random.seed(493)
# to print out all the outputs
from IPython.core.interactiveshell import InteractiveShell
InteractiveShell.ast_node_interactivity = "all"
# set display options
pd.set_option('display.max_columns', None)
pd.set_option('display.max_rows', None)
pd.set_option('display.max_colwidth', -1)

Getting the Data

Let’s read the data into a dataframe. If you want to follow along, you can download the cleaned dataset here along with the file for stop words¹. This dataset contains Trump’s tweets from the moment he took office on January 20, 2017 to May 30, 2020.

df = pd.read_csv('trump_20200530_clean.csv', parse_dates=True, index_col='datetime')

Let’s take a quick look at the data.

df.head()
df.info()

Using spaCy

Now let’s import spaCy and begin natural language processing.

# for natural language processing: named entity recognition
import spacy
import en_core_web_sm

We’re only going to use spaCy’s ner functionality or named-entity recognition so we’ll disable the rest of the functionalities. This will save us a lot of loading time later.

nlp = spacy.load(‘en_core_web_sm’, disable=[‘tagger’, ‘parser’, ‘textcat’])

Now let’s load the contents stopwords file into the variable stopswords. Note that we converted the list into a set to also save some processing time later.

with open(‘twitter-stopwords — TA — Less.txt’) as f:
contents = f.read().split(‘,’)
stopwords = set(contents)

Next, we’ll import joblib and define a few functions to help with parallel processing.

from joblib import Parallel, delayed

def chunker(iterable, total_length, chunksize):
    return (iterable[pos: pos + chunksize] for pos in range(0, total_length, chunksize))

def flatten(list_of_lists):
    "Flatten a list of lists to a combined list"
    return [item for sublist in list_of_lists for item in sublist]

def process_chunk(texts):
    preproc_pipe = []
    for doc in nlp.pipe(texts, batch_size=20):
        preproc_pipe.append([(ent.text) for ent in doc.ents if ent.label_ in ['NORP', 'PERSON', 'FAC', 'ORG', 'GPE', 'LOC', 'PRODUCT', 'EVENT']])
    return preproc_pipe

def preprocess_parallel(texts, chunksize=100):
    executor = Parallel(n_jobs=7, backend='multiprocessing', prefer="processes")
    do = delayed(process_chunk)
    tasks = (do(chunk) for chunk in chunker(texts, len(df), chunksize=chunksize))
    result = executor(tasks)
    return flatten(result)

In the code above², the function preprocess_parallel executes the other function process_chunks in parallel to help with speed. The function process_chunks iterates through a series of texts — in our case, the column 'tweet' of our the df dataframe — and inspects the entity if it belongs to either NORP, PERSON, FAC, ORG, GPE, LOC, PRODUCT, or EVENT. If it is, the entity is then appended to 'preproc_pipe' and subsequently returned to its caller. Prashanth Rao has a very good article on making spaCy super fast.

Let’s call the main driver for the functions now.

df['entities'] = preprocess_parallel(df['tweet'], chunksize=1000)

Doing a quick df.head() will reveal the new column 'entities' that we added earlier to hold the entities found in the 'tweet' column.

Prettifying the Results

In the code below, we’re making a list of lists called 'entities' and then flattening it for easier processing. We’re also converting it into a set called 'entities_set'.

entities = [entity for entity in df.entities if entity != []]
entities = [item for sublist in entities for item in sublist]
entities_set = set(entities)

Next, let’s count the frequency of the entities and append it to the list of tuples entities_counts. Then let’s convert the results into a dataframe df_counts.

df_counts = pd.Series(entities).value_counts()[:20].to_frame().reset_index()
df_counts.columns=['entity', 'count']
df_counts

For this step, we’re going to reinitialize an empty list entity_counts and manually construct a list of tuples with a combined set of entities and the sum of their frequencies or count.

entity_counts = []

entity_counts.append(('Democrats', df_counts.loc[df_counts.entity.isin(['Democrats', 'Dems', 'Democrat'])]['count'].sum()))
entity_counts.append(('Americans', df_counts.loc[df_counts.entity.isin(['American', 'Americans'])]['count'].sum()))
entity_counts.append(('Congress', df_counts.loc[df_counts.entity.isin(['House', 'Senate', 'Congress'])]['count'].sum()))
entity_counts.append(('America', df_counts.loc[df_counts.entity.isin(['U.S.', 'the United States', 'America'])]['count'].sum()))
entity_counts.append(('Republicans', df_counts.loc[df_counts.entity.isin(['Republican', 'Republicans'])]['count'].sum()))

entity_counts.append(('China', 533))
entity_counts.append(('FBI', 316))
entity_counts.append(('Russia', 313))
entity_counts.append(('Fake News', 248))
entity_counts.append(('Mexico', 213))
entity_counts.append(('Obama', 176))

Let’s take a quick look before continuing.

Finally, let’s convert the list of tuples into a dataframe.

df_ner = pd.DataFrame(entity_counts, columns=["entity", "count"]).sort_values('count', ascending=False).reset_index(drop=True)

And that’s it!

We’ve successfully created a ranking of the named entities that President Trump most frequently talked about in his tweets since taking office.


Thank you for reading! Exploratory data analysis uses a lot of techniques and we’ve only explored a few on this post. I encourage you to keep practicing and employ other techniques to derive insights from data.

In the next post, we shall continue our journey into the heart of darkness and do some topic-modeling using LDA.

Stay tuned!

You can reach me on Twitter or LinkedIn.

[1] GONG Wei’s Homepage. (May 30, 2020). Stop words for tweets. https://sites.google.com/site/iamgongwei/home/sw

[2] Towards Data Science. (May 30, 2020). Turbo-charge your spaCy NLP pipeline. https://towardsdatascience.com/turbo-charge-your-spacy-nlp-pipeline-551435b664ad

Into the Heart of Darkness - Pt. 1

Exploring the Trump Twitter Archive with Python. For beginners.


In this post, we’ll explore the dataset provided by the Trump Twitter Archive. My goal was to do something fun by using a very interesting dataset. However, as it turned out, exposure to Trump’s narcissism and shenanigans were quite depressing — if not traumatic.

You’d been warned!


For this project, we’ll be using pandas and numpy for data manipulation, matplotlib for visualizations, datetime for working with timestamps, unicodedata and regex for processing strings, and finally, nltk for natural language processing.

Let’s get started by firing up a Jupyter notebook!

Environment

We’re going to import pandas and matplotlib, and also set the display options for Jupyter so that the rows and columns are not truncated.

# for manipulating data
import pandas as pd
import numpy as np
# for visualizations
%matplotlib inline
import matplotlib.pyplot as plt
# to print out all the outputs
from IPython.core.interactiveshell import InteractiveShell
InteractiveShell.ast_node_interactivity = "all"
# set display options
pd.set_option('display.max_columns', None)
pd.set_option('display.max_rows', None)
pd.set_option('display.max_colwidth', -1)

Getting the Data

Let’s read the data into a dataframe. If you want to follow along, you can download the dataset here. This dataset contains Trump’s tweets from the moment he took office on January 20, 2017 to May 30, 2020.

df = pd.read_csv('trump_20200530.csv')

Let’s look at the first five rows and see the number of records (rows) and fields (columns).

df.head()
df.shape

Let’s do a quick renaming of the columns to make it easier for us later.

df.columns=['source', 'tweet', 'date_time', 'retweets', 'favorites', 'is_retweet', 'id']

Let’s drop the id column since it’s not really relevant right now.

df = df.drop(columns=['id'])

Let’s do a quick sanity check, this time let’s also check the dtypes of the columns.

df.head()
df.info()

Working with Timestamps

We can see from the previous screenshot that the ‘date_time’ column is a string. Let’s parse it to a timestamp.

# for working with timestamps
from datetime import datetime
from dateutil.parser import parse
dt = []
for ts in df.date_time:
dt.append(parse(ts))
dt[:5]

Let’s add a column with ‘datetime’ that contains the timestamp information.

df['datetime'] = df.apply(lambda row: parse(row.date_time), axis=1)

Let’s double-check the data range of our dataset.

df.datetime.min()
df.datetime.max()

Trimming the Data

Let’s see how many sources there are for the tweets.

df.source.value_counts()

Let’s only keep the ones that were made using the ‘Twitter for iPhone’ app.

df = df.loc[df.source == 'Twitter for iPhone']

We should drop the old ‘date_time’ column and the ‘source’ column as well.

df = df.drop(columns=['date_time', 'source'])

Separating the Retweets

Let’s see how many are retweets.

df.is_retweet.value_counts()

Let’s make another dataframe that contains only retweets and drop the ‘is_retweet’ column.

df_retweets = df.loc[df.is_retweet == True]
df_retweets = df_retweets.drop(columns=['is_retweet'])

Sanity check:

df_retweets.head()
df_retweets.shape

Back on the original dataframe, let’s remove the retweets from the dataset and drop the ‘is_retweet’ column altogether.

df = df.loc[df.is_retweet == False]
df = df.drop(columns=['is_retweet'])

Another sanity check:

df.head()
df.shape

Exploring the Data

Let’s explore both of the dataframes and answer a few questions.

What time does the President tweet the most? What time does he tweet the least?

The graph below shows that the President most frequently tweets around 12pm. He also tweets the least around 8am. Maybe he’s not a morning person?

title = 'Number of Tweets by Hour'
df.tweet.groupby(df.datetime.dt.hour).count().plot(figsize=(12,8), fontsize=14, kind='bar', rot=0, title=title)
plt.xlabel('Hour')
plt.ylabel('Number of Tweets')

What day does the President tweet the most? What day does he tweet the least?

The graph below shows that the President most frequently tweets on Wednesday. He also tweets the least on Thursday.

title = 'Number of Tweets by Day of the Week'
df.tweet.groupby(df.datetime.dt.dayofweek).count().plot(figsize=(12,8), fontsize=14, kind='bar', rot=0, title=title)
plt.xlabel('Day of the Week')
plt.ylabel('Number of Tweets')
plt.xticks(np.arange(7),['Mon', 'Tue', 'Wed', 'Thu', 'Fri', 'Sat', 'Sun'])

Isolating Twitter Handles from the Retweets

Let’s import regex so we can use it to parse the text and isolate the Twitter handles of the original tweets. In the code below, we add another column that contains the Twitter handle.

import re
pattern = re.compile('(?<=RT @).*?(?=:)')
df_retweets['original'] = [re.search(pattern, tweet).group(0) for tweet in df_retweets.tweet]

Let’s create another dataframe that will contain only the original Twitter handles and their associated number of retweets.

df_originals = df_retweets.groupby(['original']).sum().sort_values('retweets').reset_index().sort_values('retweets', ascending=False)

Let’s check the data real quick:

df_originals.head()
df_originals.shape

Let’s visualize the results real quick so we can get an idea if the data is disproportionate or not.

df_originals = df_retweets.groupby(['original']).sum().sort_values('retweets').reset_index().sort_values('retweets', ascending=False)[:10].sort_values('retweets')
df_originals.plot.barh(x='original', y='retweets', figsize=(16,10), fontsize=16)
plt.xlabel("Originating Tweet's Username")
plt.xticks([])

Which Twitter user does the President like to retweet the most?

The graph below shows that the President likes to retweet the tweets from ‘@realDonaldTrump’. Does this mean the president likes to retweet himself? You don’t say!

The interesting handle on this one is ‘@charliekirk11’. Charlie Kirk is the founder of Turning Point USA. CBS News described the organization as a far-right organization that is “shunned or at least ignored by more established conservative groups in Washington, but embraced by many Trump supporters”.¹

The Top 5 Retweets

Let’s look at the top 5 tweets that were retweeted the most by others based on the original Twitter handle.

Let’s start with the ones with ‘@realDonaldTrump’.

df_retweets.loc[df_retweets.original == 'realDonaldTrump'].sort_values('retweets', ascending=False)[:5]

And another one with ‘@charliekirk11’.

df_retweets.loc[df_retweets.original == 'charliekirk11'].sort_values('retweets', ascending=False)[:5]

Examining Retweets’ Favorites count

Let’s find out how many of the retweets are favorited by others.

df_retweets.favorites.value_counts()

Surprisingly, none of the retweets seemed to have been favorited by anybody. Weird.

We should drop it.

Counting N-Grams

To do some n-gram ranking, we need to import unicodedata and nltk. We also need to specify additional stopwords that we may need to exclude from our analysis.

# for cleaning and natural language processing
import unicodedata
import nltk
# add appropriate words that will be ignored in the analysis
ADDITIONAL_STOPWORDS = ['rt']

Here are a few of my favorite functions for natural language processing:

def clean(text):
  """
  A simple function to clean up the data. All the words that
  are not designated as a stop word is then lemmatized after
  encoding and basic regex parsing are performed.
  """
  wnl = nltk.stem.WordNetLemmatizer()
  stopwords = nltk.corpus.stopwords.words('english') + ADDITIONAL_STOPWORDS
  text = (unicodedata.normalize('NFKD', text)
    .encode('ascii', 'ignore')
    .decode('utf-8', 'ignore')
    .lower())
  words = re.sub(r'[^\w\s]', '', text).split()
  return [wnl.lemmatize(word) for word in words if word not in stopwords]

def get_words(df, column):
    """
    Takes in a dataframe and columns and returns a list of
    words from the values in the specified column.
    """
    return clean(''.join(str(df[column].tolist())))

def get_bigrams(df, column):
    """
    Takes in a list of words and returns a series of
    bigrams with value counts.
    """
    return (pd.Series(nltk.ngrams(get_words(df, column), 2)).value_counts())[:10]

def get_trigrams(df, column):
    """
    Takes in a list of words and returns a series of
    trigrams with value counts.
    """
    return (pd.Series(nltk.ngrams(get_words(df, column), 3)).value_counts())[:10]

def viz_bigrams(df ,column):
    get_bigrams(df, column).sort_values().plot.barh(color='blue', width=.9, figsize=(12, 8))

    plt.title('20 Most Frequently Occuring Bigrams')
    plt.ylabel('Bigram')
    plt.xlabel('# Occurances')

def viz_trigrams(df, column):
    get_trigrams(df, column).sort_values().plot.barh(color='blue', width=.9, figsize=(12, 8))

    plt.title('20 Most Frequently Occuring Trigrams')
    plt.ylabel('Trigram')
    plt.xlabel('# Occurances')
 

Let’s look at the top 10 bigrams in the df dataframe using the ‘tweet’ column.

get_bigrams(df, 'tweet')

And now, for the top 10 trigrams:

Let’s use the viz_bigrams() function and visualize the bigrams.

viz_bigrams(df, ‘tweet’)

Similarly, let’s use the viz_trigrams() function and visualize the trigrams.

viz_trigrams(df, 'tweet')

And there we have it!

From the moment that Trump took office, we can confidently say that the “fake news media” has been on top of the president’s mind.

Conclusion

Using basic Python and the nltk library, we’ve explored the dataset from the Trump Twitter Archive and did some n-gram ranking out of it.


Thank you for reading! Exploratory data analysis uses a lot of techniques and we’ve only explored a few on this post. I encourage you to keep practicing and employ other techniques to derive insights from data.

In the next post, we shall continue our journey into the heart of darkness and use spaCy to extract named-entities from the same dataset.

Stay tuned!

You can reach me on Twitter or LinkedIn.

[1] CBS News. “Trump speaks to conservative group Turning Point USA”. www.cbsnews.com. Archived from the original on July 31, 2019. Retrieved August 5, 2019.

Populating a Network Graph with Named-Entities

An early attempt of using networkx to visualize the results of natural language processing.


I do a lot of natural language processing and usually, the results are pretty boring to the eye. When I learned about network graphs, it got me thinking, why not use keywords as nodes and connect them together to create a network graph?

Yupp, why not!

In this post, we’ll do exactly that. We’re going to extract named-entities from news articles about coronavirus and then use their relationships to connect them together in a network graph.


A Brief Introduction

Network graphs are a cool visual that contains nodes (vertices) and edges (lines). It’s often used in social network analysis and network analysis but data scientists also use it for natural language processing.

Photo by Anders Sandberg on Flicker

Natural Language Processing or NLP is a branch of artificial intelligence that deals with programming computers to process and analyze large volumes of text and derive meaning out of them.¹ In other words, it’s all about teaching computers how to understand human language… like a boss!

Photo by brewbooks on Flickr

Enough introduction, let’s get to coding!


To get started, let’s make sure to take care of all dependencies. Open up a terminal and execute the following commands:

pip install -U spacy
python -m spacy download en
pip install networkx
pip install fuzzywuzzy

This will install spaCy and download the trained model for English. The third command installs networkx. This should work for most systems. If it doesn’t work for you, check out the documentation for spaCy and networkx. Also, we’re using fuzzywuzzy for some text preprocessing.

With that out of the way, let’s fire up a Jupyter notebook and get started!


Imports

Run the following code block into a cell to get all the necessary imports into our Python environment.

import pandas as pd
import numpy as np
import pickle
from operator import itemgetter
from fuzzywuzzy import process, fuzz# for natural language processing
import spacy
import en_core_web_sm# for visualizations
%matplotlib inline
from matplotlib.pyplot import figureimport networkx as nx

Getting the Data

If you want to follow along, you can download the sample dataset here. The file was created using newspaper to import news articles from the npr.org. If you’re feeling adventurous, use the code snippet below to build your own dataset.

import requests
import json
import time
import newspaper
import pickle

npr = newspaper.build('https://www.npr.org/sections/coronavirus-live-updates')

corpus = []
count = 0
for article in npr.articles:
    time.sleep(1)
    article.download()
    article.parse()
    text = article.text
    corpus.append(text)
    if count % 10 == 0 and count != 0:
        print('Obtained {} articles'.format(count))
    count += 1

corpus300 = corpus[:300]

with open("npr_coronavirus.txt", "wb") as fp:   # Pickling
    pickle.dump(corpus300, fp)

# with open("npr_coronavirus.txt", "rb") as fp:   # Unpickling
#     corpus = pickle.load(fp)

Let’s get our data.

with open('npr_coronavirus.txt', 'rb') as fp:   # Unpickling
corpus = pickle.load(fp)

Extract Entities

Next, we’ll start by loading spaCy’s English model:

nlp = en_core_web_sm.load()

Then, we’ll extract the entities:

entities = []for article in corpus[:50]:
tokens = nlp(''.join(article))
gpe_list = []
for ent in tokens.ents:
if ent.label_ == 'GPE':
gpe_list.append(ent.text)
entities.append(gpe_list)

In the above code block, we created an empty list called entities to store a list of lists that contains the extracted entities from each of the articles. In the for-loop, we looped through the first 50 articles of the corpus. For each iteration, we converted each articles into tokens (words) and then we looped through all those words to get the entities that are labeled as GPE for countries, states, and cities. We used ent.text to extract the actual entity and appended them one by one to entities.

Here’s the result:

Note that North Carolina has several variations of its name and some have “the” prefixed in their names. Let’s get rid of them.

articles = []for entity_list in entities:
cleaned_entity_list = []
for entity in entity_list:
cleaned_entity_list.append(entity.lstrip('the ').replace("'s", "").replace("’s",""))
articles.append(cleaned_entity_list)

In the code block above, we’re simply traversing the list of lists articles and cleaning the entities one by one. With each iteration, we’re stripping the prefix “the” and getting rid of 's.

Optional: FuzzyWuzzy

Looking at the entities, I’ve noticed that there are also variations in the “United States” is represented. There exists “United States of America” while some are just “United States”. We can trim these down into a more standard naming convention.

FuzzyWuzzy can help with this.

Described by pypi.org as “string matching like a boss,” FiuzzyWuzzy uses Levenshtein distance to calculate the similarities between words.¹ For a really good tutorial on how to use FuzzyWuzzy, check out Thanh Huynh’s article.FuzzyWuzzy: Find Similar Strings within one column in PythonToken Sort Ratio vs. Token Set Ratiotowardsdatascience.com

Here’s the optional code for using FuzzyWuzzy:

choices = set([item for sublist in articles for item in sublist])

cleaned_articles = []
for article in articles:
    article_entities = []
    for entity in set(article):
        article_entities.append(process.extractOne(entity, choices)[0])
    cleaned_articles.append(article_entities)

For the final step before creating the network graph, let’s get rid of the empty lists within our list of list that were generated by articles who didn’t have any GPE entity types.

articles = [article for article in articles if article != []]

Create the Network Graph

For the next step, we’ll create the world into which the graph will exist.

G = nx.Graph()

Then, we’ll manually add the nodes with G.add_nodes_from().

for entities in articles:
G.add_nodes_from(entities)

Let’s see what the graph looks like with:

figure(figsize=(10, 8))
nx.draw(G, node_size=15)

Next, let’s add the edges that will connect the nodes.

for entities in articles:
if len(entities) > 1:
for i in range(len(entities)-1):
G.add_edges_from([(str(entities[i]),str(entities[i+1]))])

For each iteration of the code above, we used a conditional that will only entertain a list of entities that has two or more entities. Then, we manually connect each of the entities with G.add_edges_from().

Let’s see what the graph looks like now:

figure(figsize=(10, 8))
nx.draw(G, node_size=10)

This graph reminds me of spiders! LOL.

To organize it a bit, I decided to use the shell version of the network graph:

figure(figsize=(10, 8))
nx.draw_shell(G, node_size=15)

We can tell that some nodes are heavier on connections than others. To see which nodes have the most connections, let’s use G.degree().

G.degree()

This gives the following degree view:

Let’s find out which node or entity has the most number of connections.

max(dict(G.degree()).items(), key = lambda x : x[1])

To find out which other nodes have the most number of connections, let’s check the top 5:

degree_dict = dict(G.degree(G.nodes()))
nx.set_node_attributes(G, degree_dict, 'degree')sorted_degree = sorted(degree_dict.items(), key=itemgetter(1), reverse=True)

Above, sorted_degrees is a list that contains all the nodes and their degree values. We only wanted the top 5 like so:

print("Top 5 nodes by degree:")
for d in sorted_degree[:5]:
print(d)

Bonus Round: Gephi

Gephi is an open-source and free desktop application that lets us visualize, explore, and analyze all kinds of graphs and networks.²

Let’s export our graph data into a file so we can import it into Gephi.

nx.write_gexf(G, "npr_coronavirus_GPE_50.gexf")

Cool beans!

Next Steps

This time, we only processed 50 articles from npr.org. What would happen if we processed all 300 articles from our dataset? What will we see if we change the entity type from GPE to PERSON? How else can we use network graphs to visualize natural language processing results?

There’s always more to do. The possibilities are endless!


I hope you enjoyed today’s post. The code is not perfect and we have a long way to go towards realizing insights from the data. I encourage you to dive deeper and learn more about spaCynetworkxfuzzywuzzy, and even Gephi.

Stay tuned!

You can reach me on Twitter or LinkedIn.

[1]: Wikipedia. (May 25, 2020). Natural language processing https://en.wikipedia.org/wiki/Natural_language_processing

[2]: Gephi. (May 25, 2020). The Open Graph Viz Platform https://gephi.org/

This article was first published in Towards Data Science‘ Medium publication.