Using Text Data with EvalML

In this demo, we will show you how to use EvalML to build models which use text data.

import evalml
from evalml import AutoMLSearch

Using `tqdm.autonotebook.tqdm` in notebook mode. Use `tqdm.tqdm` instead to force console mode (e.g. in jupyter console)

Dataset

We will be utilizing a dataset of SMS text messages, some of which are categorized as spam, and others which are not (“ham”). This dataset is originally from Kaggle, but modified to produce a slightly more even distribution of spam to ham.

from urllib.request import urlopen
import pandas as pd

input_data = urlopen(
    "https://featurelabs-static.s3.amazonaws.com/spam_text_messages_modified.csv"
)
data = pd.read_csv(input_data)[:750]

X = data.drop(["Category"], axis=1)
y = data["Category"]

display(X.head())

	Message
0	Free entry in 2 a wkly comp to win FA Cup fina...
1	FreeMsg Hey there darling it's been 3 week's n...
2	WINNER!! As a valued network customer you have...
3	Had your mobile 11 months or more? U R entitle...
4	SIX chances to win CASH! From 100 to 20,000 po...

The ham vs spam distribution of the data is 3:1, so any machine learning model must get above 75% accuracy in order to perform better than a trivial baseline model which simply classifies everything as ham.

y.value_counts(normalize=True)

spam    0.593333
ham     0.406667
Name: Category, dtype: float64

In order to properly utilize Woodwork’s ‘Natural Language’ typing, we need to pass this argument in during initialization. Otherwise, this will be treated as an ‘Unknown’ type and dropped in the search.

X.ww.init(logical_types={"Message": "NaturalLanguage"})

Search for best pipeline

In order to validate the results of the pipeline creation and optimization process, we will save some of our data as a holdout set.

X_train, X_holdout, y_train, y_holdout = evalml.preprocessing.split_data(
    X, y, problem_type="binary", test_size=0.2, random_seed=0
)

EvalML uses Woodwork to automatically detect which columns are text columns, so you can run search normally, as you would if there was no text data. We can print out the logical type of the Message column and assert that it is indeed inferred as a natural language column.

X_train.ww

	Physical Type	Logical Type	Semantic Tag(s)
Column
Message	string	NaturalLanguage	[]

Because the spam/ham labels are binary, we will use AutoMLSearch(X_train=X_train, y_train=y_train, problem_type='binary'). When we call .search(), the search for the best pipeline will begin.

automl = AutoMLSearch(
    X_train=X_train,
    y_train=y_train,
    problem_type="binary",
    max_batches=1,
    optimize_thresholds=True,
    verbose=True,
)

automl.search(interactive_plot=False)

AutoMLSearch will use mean CV score to rank pipelines.

*****************************
* Beginning pipeline search *
*****************************

Optimizing for Log Loss Binary.
Lower score is better.

Using SequentialEngine to train and score pipelines.
Searching up to 1 batches for a total of None pipelines.
Allowed model families:

Evaluating Baseline Pipeline: Mode Baseline Binary Classification Pipeline
Mode Baseline Binary Classification Pipeline:
        Starting cross validation
        Finished cross validation - mean Log Loss Binary: 14.658

*****************************
* Evaluating Batch Number 1 *
*****************************

Random Forest Classifier w/ Label Encoder + Natural Language Featurizer + Imputer:
        Starting cross validation
        Finished cross validation - mean Log Loss Binary: 0.212

Search finished after 6.88 seconds
Best pipeline: Random Forest Classifier w/ Label Encoder + Natural Language Featurizer + Imputer
Best pipeline Log Loss Binary: 0.212470

{1: {'Random Forest Classifier w/ Label Encoder + Natural Language Featurizer + Imputer': 5.980375528335571,
  'Total time of batch': 6.104418754577637}}

View rankings and select pipeline

Once the fitting process is done, we can see all of the pipelines that were searched.

automl.rankings

	id	pipeline_name	search_order	ranking_score	mean_cv_score	standard_deviation_cv_score	percent_better_than_baseline	high_variance_cv	parameters
0	1	Random Forest Classifier w/ Label Encoder + Na...	1	0.212470	0.212470	0.043953	98.550459	False	{'Label Encoder': {'positive_label': None}, 'I...
1	0	Mode Baseline Binary Classification Pipeline	0	14.657752	14.657752	0.104049	0.000000	False	{'Label Encoder': {'positive_label': None}, 'B...

To select the best pipeline we can call automl.best_pipeline.

best_pipeline = automl.best_pipeline

Describe pipeline

You can get more details about any pipeline, including how it performed on other objective functions.

automl.describe_pipeline(automl.rankings.iloc[0]["id"])

*************************************************************************************
* Random Forest Classifier w/ Label Encoder + Natural Language Featurizer + Imputer *
*************************************************************************************

Problem Type: binary
Model Family: Random Forest

Pipeline Steps
==============
1. Label Encoder
         * positive_label : None
2. Natural Language Featurizer
3. Imputer
         * categorical_impute_strategy : most_frequent
         * numeric_impute_strategy : mean
         * boolean_impute_strategy : most_frequent
         * categorical_fill_value : None
         * numeric_fill_value : None
         * boolean_fill_value : None
4. Random Forest Classifier
         * n_estimators : 100
         * max_depth : 6
         * n_jobs : -1

Training
========
Training for binary problems.
Total training time (including CV): 6.0 seconds

Cross Validation
----------------
             Log Loss Binary  MCC Binary  Gini   AUC  Precision    F1  Balanced Accuracy Binary  Accuracy Binary # Training # Validation
0                      0.213       0.876 0.948 0.974      0.916 0.927                     0.940            0.940        400          200
1                      0.168       0.865 0.976 0.988      0.947 0.917                     0.928            0.935        400          200
2                      0.256       0.782 0.920 0.960      0.896 0.868                     0.887            0.895        400          200
mean                   0.212       0.841 0.948 0.974      0.920 0.904                     0.918            0.923          -            -
std                    0.044       0.051 0.028 0.014      0.026 0.032                     0.028            0.025          -            -
coef of var            0.207       0.061 0.030 0.014      0.028 0.035                     0.030            0.027          -            -

best_pipeline.graph()

Notice above that there is a Natural Language Featurizer as the first step in the pipeline. AutoMLSearch uses the woodwork accessor to recognize that 'Message' is a text column, and converts this text into numerical values that can be handled by the estimator.

Evaluate on holdout

Now, we can score the pipeline on the holdout data using the ranking objectives for binary classification problems.

scores = best_pipeline.score(
    X_holdout, y_holdout, objectives=evalml.objectives.get_ranking_objectives("binary")
)
print(f'Accuracy Binary: {scores["Accuracy Binary"]}')

Accuracy Binary: 0.9333333333333333

As you can see, this model performs relatively well on this dataset, even on unseen data.

What does the Natural Language Featurizer do?

Machine learning models cannot handle non-numeric data. Any text must be broken down into numeric features that provide useful information about that text. The Natural Natural Language Featurizer first normalizes your text by removing any punctuation and other non-alphanumeric characters and converting any capital letters to lowercase. From there, it passes the text into featuretools’ nlp_primitives dfs search, resulting in several informative features that replace the original column in your dataset: Diversity Score, Mean Characters per Word, Polarity Score, LSA (Latent Semantic Analysis), Number of Characters, and Number of Words.

Diversity Score is the ratio of unique words to total words.

Mean Characters per Word is the average number of letters in each word.

Polarity Score is a prediction of how “polarized” the text is, on a scale from -1 (extremely negative) to 1 (extremely positive).

Latent Semantic Analysis is an abstract representation of how important each word is with respect to the entire text, reduced down into two values per text. While the other text features are each a single column, this feature adds two columns to your data, LSA(column_name)[0] and LSA(column_name)[1].

Number of Characters is the number of characters in the text.

Number of Words is the number of words in the text.

Let’s see what this looks like with our spam/ham example.

best_pipeline.input_feature_names

{'Label Encoder': ['Message'],
 'Natural Language Featurizer': ['Message'],
 'Imputer': ['DIVERSITY_SCORE(Message)',
  'MEAN_CHARACTERS_PER_WORD(Message)',
  'NUM_CHARACTERS(Message)',
  'NUM_WORDS(Message)',
  'POLARITY_SCORE(Message)',
  'LSA(Message)[0]',
  'LSA(Message)[1]'],
 'Random Forest Classifier': ['DIVERSITY_SCORE(Message)',
  'MEAN_CHARACTERS_PER_WORD(Message)',
  'NUM_CHARACTERS(Message)',
  'NUM_WORDS(Message)',
  'POLARITY_SCORE(Message)',
  'LSA(Message)[0]',
  'LSA(Message)[1]']}

Here, the Natural Language Featurizer takes in a single “Message” column, but then the next component in the pipeline, the Imputer, receives five columns of input. These five columns are the result of featurizing the text-type “Message” column. Most importantly, these featurized columns are what ends up passed in to the estimator.

If the dataset had any non-text columns, those would be left alone by this process. If the dataset had more than one text column, each would be broken into these five feature columns independently.

The features, more directly

Rather than just checking the new column names, let’s examine the output of this component directly. We can see this by running the component on its own.

natural_language_featurizer = evalml.pipelines.components.NaturalLanguageFeaturizer()
X_featurized = natural_language_featurizer.fit_transform(X_train)

Now we can compare the input data to the output from the Natural Language Featurizer:

X_train.head()

	Message
296	Sunshine Hols. To claim ur med holiday send a ...
652	Yup ü not comin :-(
526	Hello hun how ru? Its here by the way. Im good...
571	I tagged MY friends that you seemed to count a...
472	What happened to our yo date?

X_featurized.head()

	DIVERSITY_SCORE(Message)	MEAN_CHARACTERS_PER_WORD(Message)	NUM_CHARACTERS(Message)	NUM_WORDS(Message)	POLARITY_SCORE(Message)	LSA(Message)[0]	LSA(Message)[1]
296	1.0	4.344828	154.0	29.0	0.003	0.150556	-0.072443
652	1.0	3.000000	16.0	4.0	0.000	0.017340	-0.005411
526	1.0	3.363636	143.0	33.0	0.162	0.169954	0.022670
571	0.8	4.083333	60.0	12.0	0.681	0.144713	0.036799
472	1.0	3.833333	28.0	6.0	0.000	0.109373	-0.042754

These numeric values now represent important information about the original text that the estimator at the end of the pipeline can successfully use to make predictions.

Why encode text this way?

To demonstrate the importance of text-specific modeling, let’s train a model with the same dataset, without letting AutoMLSearch detect the text column. We can change this by explicitly setting the data type of the 'Message' column in Woodwork to Categorical using the utility method infer_feature_types.

from evalml.utils import infer_feature_types

X = infer_feature_types(X, {"Message": "Categorical"})
X_train, X_holdout, y_train, y_holdout = evalml.preprocessing.split_data(
    X, y, problem_type="binary", test_size=0.2, random_seed=0
)

automl_no_text = AutoMLSearch(
    X_train=X_train,
    y_train=y_train,
    problem_type="binary",
    max_batches=1,
    optimize_thresholds=True,
    verbose=True,
)

automl_no_text.search(interactive_plot=False)

AutoMLSearch will use mean CV score to rank pipelines.

*****************************
* Beginning pipeline search *
*****************************

Optimizing for Log Loss Binary.
Lower score is better.

Using SequentialEngine to train and score pipelines.
Searching up to 1 batches for a total of None pipelines.
Allowed model families:

Evaluating Baseline Pipeline: Mode Baseline Binary Classification Pipeline
Mode Baseline Binary Classification Pipeline:
        Starting cross validation
        Finished cross validation - mean Log Loss Binary: 14.658

*****************************
* Evaluating Batch Number 1 *
*****************************

Random Forest Classifier w/ Label Encoder + Natural Language Featurizer + Imputer:
        Starting cross validation
        Finished cross validation - mean Log Loss Binary: 0.212

Search finished after 4.87 seconds
Best pipeline: Random Forest Classifier w/ Label Encoder + Natural Language Featurizer + Imputer
Best pipeline Log Loss Binary: 0.212470

{1: {'Random Forest Classifier w/ Label Encoder + Natural Language Featurizer + Imputer': 4.276244878768921,
  'Total time of batch': 4.401290655136108}}

Like before, we can look at the rankings and pick the best pipeline.

automl_no_text.rankings

	id	pipeline_name	search_order	ranking_score	mean_cv_score	standard_deviation_cv_score	percent_better_than_baseline	high_variance_cv	parameters
0	1	Random Forest Classifier w/ Label Encoder + Na...	1	0.212470	0.212470	0.043953	98.550459	False	{'Label Encoder': {'positive_label': None}, 'I...
1	0	Mode Baseline Binary Classification Pipeline	0	14.657752	14.657752	0.104049	0.000000	False	{'Label Encoder': {'positive_label': None}, 'B...

best_pipeline_no_text = automl_no_text.best_pipeline

Here, changing the data type of the text column removed the Natural Language Featurizer from the pipeline.

best_pipeline_no_text.graph()

automl_no_text.describe_pipeline(automl_no_text.rankings.iloc[0]["id"])

*************************************************************************************
* Random Forest Classifier w/ Label Encoder + Natural Language Featurizer + Imputer *
*************************************************************************************

Problem Type: binary
Model Family: Random Forest

Pipeline Steps
==============
1. Label Encoder
         * positive_label : None
2. Natural Language Featurizer
3. Imputer
         * categorical_impute_strategy : most_frequent
         * numeric_impute_strategy : mean
         * boolean_impute_strategy : most_frequent
         * categorical_fill_value : None
         * numeric_fill_value : None
         * boolean_fill_value : None
4. Random Forest Classifier
         * n_estimators : 100
         * max_depth : 6
         * n_jobs : -1

Training
========
Training for binary problems.
Total training time (including CV): 4.3 seconds

Cross Validation
----------------
             Log Loss Binary  MCC Binary  Gini   AUC  Precision    F1  Balanced Accuracy Binary  Accuracy Binary # Training # Validation
0                      0.213       0.876 0.948 0.974      0.916 0.927                     0.940            0.940        400          200
1                      0.168       0.865 0.976 0.988      0.947 0.917                     0.928            0.935        400          200
2                      0.256       0.782 0.920 0.960      0.896 0.868                     0.887            0.895        400          200
mean                   0.212       0.841 0.948 0.974      0.920 0.904                     0.918            0.923          -            -
std                    0.044       0.051 0.028 0.014      0.026 0.032                     0.028            0.025          -            -
coef of var            0.207       0.061 0.030 0.014      0.028 0.035                     0.030            0.027          -            -

# get standard performance metrics on holdout data
scores = best_pipeline_no_text.score(
    X_holdout, y_holdout, objectives=evalml.objectives.get_ranking_objectives("binary")
)
print(f'Accuracy Binary: {scores["Accuracy Binary"]}')

Accuracy Binary: 0.9333333333333333

Without the Natural Language Featurizer, the 'Message' column was treated as a categorical column, and therefore the conversion of this text to numerical features happened in the One Hot Encoder. The best pipeline encoded the top 10 most frequent “categories” of these texts, meaning 10 text messages were one-hot encoded and all the others were dropped. Clearly, this removed almost all of the information from the dataset, as we can see the best_pipeline_no_text performs very similarly to randomly guessing “ham” in every case.