Sentence classification

In this tutorial we will try to predict whether an SMS is a spam or not. To train our model, we will use the SMSSpam dataset. This dataset is unbalanced, there is only 13.4% spam. Let's look at the data:

from river import datasets

datasets.SMSSpam()

SMS Spam Collection dataset.

The data contains 5,574 items and 1 feature (i.e. SMS body). Spam messages represent
13.4% of the dataset. The goal is to predict whether an SMS is a spam or not.

      Name  SMSSpam                                                                              
      Task  Binary classification                                                                
   Samples  5,574                                                                                
  Features  1                                                                                    
    Sparse  False                                                                                
      Path  /home/runner/river_data/SMSSpam/SMSSpamCollection                                    
       URL  https://archive.ics.uci.edu/ml/machine-learning-databases/00228/smsspamcollection.zip
      Size  466.71 KiB                                                                           
Downloaded  True

from pprint import pprint

X_y = datasets.SMSSpam()

for x, y in X_y:
    pprint(x)
    print(f'Spam: {y}')
    break

{'body': 'Go until jurong point, crazy.. Available only in bugis n great world '
         'la e buffet... Cine there got amore wat...\n'}
Spam: False

Let's start by building a simple model like a Naive Bayes classifier. We will first preprocess the sentences with a TF-IDF transform that our model can consume. Then, we will measure the accuracy of our model with the AUC metric. This is the right metric to use when the classes are not balanced. In addition, the Naive Bayes models can perform very well on unbalanced datasets and can be used for both binary and multi-class classification problems.

from river import feature_extraction
from river import naive_bayes
from river import metrics

X_y = datasets.SMSSpam()

model = (
    feature_extraction.TFIDF(on='body') | 
    naive_bayes.BernoulliNB(alpha=0)
)

metric = metrics.ROCAUC()
cm = metrics.ConfusionMatrix()

for x, y in X_y:

    y_pred = model.predict_one(x)

    if y_pred is not None:
        metric.update(y_pred=y_pred, y_true=y)
        cm.update(y_pred=y_pred, y_true=y)

    model.learn_one(x, y)

metric

ROCAUC: 93.00%

The confusion matrix:

cm

        False   True  
False   4,809     17  
 True     102    645

The results are quite good with this first model.

Since we are working with an imbalanced dataset, we can use the imblearn module to rebalance the classes of our dataset. For more information about the imblearn module, you can find a dedicated tutorial here.

from river import imblearn

X_y = datasets.SMSSpam()

model = (
    feature_extraction.TFIDF(on='body') | 
    imblearn.RandomUnderSampler(
        classifier=naive_bayes.BernoulliNB(alpha=0),
        desired_dist={0: .5, 1: .5},
        seed=42
    )
)

metric = metrics.ROCAUC()
cm = metrics.ConfusionMatrix()

for x, y in X_y:

    y_pred = model.predict_one(x)

    if y_pred is not None:
        metric.update(y_pred=y_pred, y_true=y)
        cm.update(y_pred=y_pred, y_true=y)

    model.learn_one(x, y)

metric

ROCAUC: 94.61%

The imblearn module improved our results. Not bad! We can visualize the pipeline to understand how the data is processed.

The confusion matrix:

cm

        False   True  
False   4,570    255  
 True      41    706

model

TFIDF

TFIDF ( normalize=True on="body" strip_accents=True lowercase=True preprocessor=None stop_words=None tokenizer_pattern="(?u)\b\w[\w\-]+\b" tokenizer=None ngram_range=(1, 1) )

RandomUnderSampler

RandomUnderSampler ( classifier=BernoulliNB ( alpha=0 true_threshold=0. ) desired_dist={0: 0.5, 1: 0.5} seed=42 )

BernoulliNB

BernoulliNB ( alpha=0 true_threshold=0. )

Now let's try to use logistic regression to classify messages. We will use different tips to make my model perform better. As in the previous example, we rebalance the classes of our dataset. The logistics regression will be fed from a TF-IDF.

from river import linear_model
from river import optim
from river import preprocessing

X_y = datasets.SMSSpam()

model = (
    feature_extraction.TFIDF(on='body') | 
    preprocessing.Normalizer() | 
    imblearn.RandomUnderSampler(
        classifier=linear_model.LogisticRegression(
            optimizer=optim.SGD(.9), 
            loss=optim.losses.Log()
        ),
        desired_dist={0: .5, 1: .5},
        seed=42
    )
)

metric = metrics.ROCAUC()
cm = metrics.ConfusionMatrix()

for x, y in X_y:

    y_pred = model.predict_one(x)

    metric.update(y_pred=y_pred, y_true=y)
    cm.update(y_pred=y_pred, y_true=y)

    model.learn_one(x, y)

metric

ROCAUC: 93.80%

The confusion matrix:

cm

        False   True  
False   4,584    243  
 True      55    692

model

TFIDF

TFIDF ( normalize=True on="body" strip_accents=True lowercase=True preprocessor=None stop_words=None tokenizer_pattern="(?u)\b\w[\w\-]+\b" tokenizer=None ngram_range=(1, 1) )

Normalizer

Normalizer ( order=2 )

RandomUnderSampler

RandomUnderSampler ( classifier=LogisticRegression ( optimizer=SGD ( lr=Constant ( learning_rate=0.9 ) ) loss=Log ( weight_pos=1. weight_neg=1. ) l2=0. l1=0. intercept_init=0. intercept_lr=Constant ( learning_rate=0.01 ) clip_gradient=1e+12 initializer=Zeros () ) desired_dist={0: 0.5, 1: 0.5} seed=42 )

LogisticRegression

LogisticRegression ( optimizer=SGD ( lr=Constant ( learning_rate=0.9 ) ) loss=Log ( weight_pos=1. weight_neg=1. ) l2=0. l1=0. intercept_init=0. intercept_lr=Constant ( learning_rate=0.01 ) clip_gradient=1e+12 initializer=Zeros () )

The results of the logistic regression are quite good but still inferior to the naive Bayes model.

Let's try to use word embeddings to improve our logistic regression. Word embeddings allow you to represent a word as a vector. Embeddings are developed to build semantically rich vectors. For instance, the vector which represents the word python should be close to the vector which represents the word programming. We will use spaCy to convert our sentence to vectors. spaCy converts a sentence to a vector by calculating the average of the embeddings of the words in the sentence.

You can download pre-trained embeddings in many languages. We will use English pre-trained embeddings as our SMS are in English.

The command below allows you to download the pre-trained embeddings that spaCy makes available. More informations about spaCy and its installation may be found here here.

python -m spacy download en_core_web_sm

Here, we create a custom transformer to convert an input sentence to a dict of floats. We will integrate this transformer into our pipeline.

import spacy

from river.base import Transformer

class Embeddings(Transformer):
    """My custom transformer, word embedding using spaCy."""

    def __init__(self, on: str):
        self.on = on
        self.embeddings = spacy.load('en_core_web_sm')

    def transform_one(self, x, y=None):
        return {dimension: xi for dimension, xi in enumerate(self.embeddings(x[self.on]).vector)}

Let's train our logistic regression:

X_y = datasets.SMSSpam()

model = (
    Embeddings(on='body') | 
    preprocessing.Normalizer() |
    imblearn.RandomOverSampler(
        classifier=linear_model.LogisticRegression(
            optimizer=optim.SGD(.5), 
            loss=optim.losses.Log()
        ),
        desired_dist={0: .5, 1: .5},
        seed=42
    )
)

metric = metrics.ROCAUC()
cm = metrics.ConfusionMatrix()

for x, y in X_y:

    y_pred = model.predict_one(x)

    metric.update(y_pred=y_pred, y_true=y)
    cm.update(y_pred=y_pred, y_true=y)

    model.learn_one(x, y)

metric

ROCAUC: 91.86%

The confusion matrix:

cm

        False   True  
False   4,545    282  
 True      78    669

model

Embeddings

Embeddings ( on="body" )

Normalizer

Normalizer ( order=2 )

RandomOverSampler

RandomOverSampler ( classifier=LogisticRegression ( optimizer=SGD ( lr=Constant ( learning_rate=0.5 ) ) loss=Log ( weight_pos=1. weight_neg=1. ) l2=0. l1=0. intercept_init=0. intercept_lr=Constant ( learning_rate=0.01 ) clip_gradient=1e+12 initializer=Zeros () ) desired_dist={0: 0.5, 1: 0.5} seed=42 )

LogisticRegression

LogisticRegression ( optimizer=SGD ( lr=Constant ( learning_rate=0.5 ) ) loss=Log ( weight_pos=1. weight_neg=1. ) l2=0. l1=0. intercept_init=0. intercept_lr=Constant ( learning_rate=0.01 ) clip_gradient=1e+12 initializer=Zeros () )

The results of the logistic regression using spaCy embeddings are lower than those obtained with TF-IDF values. We could surely improve the results by cleaning up the text. We could also use embeddings more suited to our dataset. However, on this problem, the logistic regression is not better than the Naive Bayes model. No free lunch today.