Various string encoders: a sentiment analysis example

In this example, we explore the performance of string and categorical encoders available in skrub.

The Toxicity dataset

We focus on the toxicity dataset, a corpus of 1,000 tweets, evenly balanced between the binary labels “Toxic” and “Not Toxic”. Our goal is to classify each entry between these two labels, using only the text of the tweets as features.

from skrub.datasets import fetch_toxicity

dataset = fetch_toxicity()
X, y = dataset.X, dataset.y
X["is_toxic"] = y

Downloading 'toxicity' from https://github.com/skrub-data/skrub-data-files/raw/refs/heads/main/toxicity.zip (attempt 1/3)

When it comes to displaying large chunks of text, the TableReport is especially useful! Click on any cell below to expand and read the tweet in full.

from skrub import TableReport

TableReport(X)

Please enable javascript

The skrub table reports need javascript to display correctly. If you are displaying a report in a Jupyter notebook and you see this message, you may need to re-execute the cell or to trust the notebook (button on the top right or "File > Trust notebook").

GapEncoder

First, let’s vectorize our text column using the GapEncoder, one of the high cardinality categorical encoders provided by skrub. As introduced in the previous example, the GapEncoder performs matrix factorization for topic modeling. It builds latent topics by capturing combinations of substrings that frequently co-occur, and encoded vectors correspond to topic activations.

To interpret these latent topics, we select for each of them a few labels from the input data with the highest activations. In the example below we select 3 labels to summarize each topic.

from skrub import GapEncoder

gap = GapEncoder(n_components=30)
X_trans = gap.fit_transform(X["text"])
# Add the original text as a first column
X_trans.insert(0, "text", X["text"])
TableReport(X_trans)

Please enable javascript

The skrub table reports need javascript to display correctly. If you are displaying a report in a Jupyter notebook and you see this message, you may need to re-execute the cell or to trust the notebook (button on the top right or "File > Trust notebook").

We can use a heatmap to highlight the highest activations, making them more visible for comparison against the original text and vectors above.

import numpy as np
from matplotlib import pyplot as plt


def plot_gap_feature_importance(X_trans):
    x_samples = X_trans.pop("text")

    # We slightly format the topics and labels for them to fit on the plot.
    topic_labels = [x.replace("text: ", "") for x in X_trans.columns]
    labels = x_samples.str[:50].values + "..."

    # We clip large outliers to makes activations more visible.
    X_trans = np.clip(X_trans, a_min=None, a_max=200)

    plt.figure(figsize=(10, 10), dpi=200)
    plt.imshow(X_trans.T)

    plt.yticks(
        range(len(topic_labels)),
        labels=topic_labels,
        ha="right",
        size=12,
    )
    plt.xticks(range(len(labels)), labels=labels, size=12, rotation=50, ha="right")

    plt.colorbar().set_label(label="Topic activations", size=13)
    plt.ylabel("Latent topics", size=14)
    plt.xlabel("Data entries", size=14)
    plt.tight_layout()
    plt.show()


plot_gap_feature_importance(X_trans.head())

/home/circleci/project/examples/02_text_with_string_encoders.py:116: UserWarning: Glyph 4108 (\N{MYANMAR LETTER TTHA}) missing from font(s) DejaVu Sans.
  plt.tight_layout()

Now that we have an understanding of the vectors produced by the GapEncoder, let’s evaluate its performance in toxicity classification. The GapEncoder excels at handling categorical columns with high cardinality, but here the column consists of free-form text. Sentences are generally longer, with more unique ngrams than high cardinality categories.

To benchmark the performance of the GapEncoder against the toxicity dataset, we integrate it into a TableVectorizer, as introduced in the previous example, and create a Pipeline by appending a HistGradientBoostingClassifier, which consumes the vectors produced by the GapEncoder.

We set n_components to 30; however, to achieve the best performance, we would need to find the optimal value for this hyperparameter using either GridSearchCV or RandomizedSearchCV. We skip this part to keep the computation time for this small example.

Recall that the ROC AUC is a metric that quantifies the ranking power of estimators, where a random estimator scores 0.5, and an oracle —providing perfect predictions— scores 1.

from sklearn.ensemble import HistGradientBoostingClassifier
from sklearn.model_selection import cross_validate
from sklearn.pipeline import make_pipeline

from skrub import TableVectorizer


def plot_box_results(named_results):
    fig, ax = plt.subplots()
    names, scores = zip(
        *[(name, result["test_score"]) for name, result in named_results]
    )
    ax.boxplot(scores)
    ax.set_xticks(range(1, len(names) + 1), labels=list(names), size=12)
    ax.set_ylabel("ROC AUC", size=14)
    plt.title(
        "AUC distribution across folds (higher is better)",
        size=14,
    )
    plt.show()


results = []

y = X.pop("is_toxic").map({"Toxic": 1, "Not Toxic": 0})

gap_pipe = make_pipeline(
    TableVectorizer(high_cardinality=GapEncoder(n_components=30)),
    HistGradientBoostingClassifier(),
)
gap_results = cross_validate(gap_pipe, X, y, scoring="roc_auc")
results.append(("GapEncoder", gap_results))

plot_box_results(results)

MinHashEncoder

We now compare these results with the MinHashEncoder, which is faster and produces vectors better suited for tree-based estimators like HistGradientBoostingClassifier. To do this, we can simply replace the GapEncoder with the MinHashEncoder in the previous pipeline using set_params().

from skrub import MinHashEncoder

minhash_pipe = make_pipeline(
    TableVectorizer(high_cardinality=MinHashEncoder(n_components=30)),
    HistGradientBoostingClassifier(),
)
minhash_results = cross_validate(minhash_pipe, X, y, scoring="roc_auc")
results.append(("MinHashEncoder", minhash_results))

plot_box_results(results)

Remarkably, the vectors produced by the MinHashEncoder offer less predictive power than those from the GapEncoder on this dataset.

TextEncoder

Let’s now shift our focus to pre-trained deep learning encoders. Our previous encoders are syntactic models that we trained directly on the toxicity dataset. To generate more powerful vector representations for free-form text and diverse entries, we can instead use semantic models, such as BERT, which have been trained on very large datasets.

TextEncoder enables you to integrate any Sentence Transformer model from the Hugging Face Hub (or from your local disk) into your Pipeline to transform a text column in a dataframe. By default, TextEncoder uses the e5-small-v2 model.

from skrub import TextEncoder

text_encoder = TextEncoder(
    "sentence-transformers/paraphrase-albert-small-v2",
    device="cpu",
)

text_encoder_pipe = make_pipeline(
    TableVectorizer(high_cardinality=text_encoder),
    HistGradientBoostingClassifier(),
)
text_encoder_results = cross_validate(text_encoder_pipe, X, y, scoring="roc_auc")
results.append(("TextEncoder", text_encoder_results))

plot_box_results(results)

SringEncoder

TextEncoder embeddings are very strong, but they are also quite expensive to use. A simpler, faster alternative for encoding strings is the StringEncoder, which works by first performing a tf-idf (computing vectors of rescaled word counts of the text wiki), and then following it with TruncatedSVD to reduce the number of dimensions to, in this case, 30.

from skrub import StringEncoder

string_encoder = StringEncoder(ngram_range=(3, 4), analyzer="char_wb")

string_encoder_pipe = make_pipeline(
    TableVectorizer(high_cardinality=string_encoder),
    HistGradientBoostingClassifier(),
)

string_encoder_results = cross_validate(string_encoder_pipe, X, y, scoring="roc_auc")
results.append(("StringEncoder", string_encoder_results))

plot_box_results(results)

The performance of the TextEncoder is significantly stronger than that of the syntactic encoders, which is expected. But how long does it take to load and vectorize text on a CPU using a Sentence Transformer model? Below, we display the tradeoff between predictive accuracy and training time. Note that since we are not training the Sentence Transformer model, the “fitting time” refers to the time taken for vectorization.

def plot_performance_tradeoff(results):
    fig, ax = plt.subplots(figsize=(5, 4), dpi=200)
    markers = ["s", "o", "^", "x"]
    for idx, (name, result) in enumerate(results):
        ax.scatter(
            result["fit_time"],
            result["test_score"],
            label=name,
            marker=markers[idx],
        )
        mean_fit_time = np.mean(result["fit_time"])
        mean_score = np.mean(result["test_score"])
        ax.scatter(
            mean_fit_time,
            mean_score,
            color="k",
            marker=markers[idx],
        )
        std_fit_time = np.std(result["fit_time"])
        std_score = np.std(result["test_score"])
        ax.errorbar(
            x=mean_fit_time,
            y=mean_score,
            yerr=std_score,
            fmt="none",
            c="k",
            capsize=2,
        )
        ax.errorbar(
            x=mean_fit_time,
            y=mean_score,
            xerr=std_fit_time,
            fmt="none",
            c="k",
            capsize=2,
        )
        ax.set_xscale("log")

        ax.set_xlabel("Time to fit (seconds)")
        ax.set_ylabel("ROC AUC")
        ax.set_title("Prediction performance / training time trade-off")

    ax.annotate(
        "",
        xy=(1.5, 0.98),
        xytext=(8.5, 0.90),
        arrowprops=dict(arrowstyle="->", mutation_scale=15),
    )
    ax.text(5.8, 0.86, "Best time / \nperformance trade-off")
    ax.legend(bbox_to_anchor=(1.02, 0.3))
    plt.show()


plot_performance_tradeoff(results)

The black points represent the average time to fit and AUC for each vectorizer, and the width of the bars represents one standard deviation

The green outlier dot on the right side of the plot corresponds to the first time the Sentence Transformers model was downloaded and loaded into memory. During the subsequent cross-validation iterations, the model is simply copied, which reduces computation time for the remaining folds.

Interestingly, StringEncoder has a performance remarkably similar to that of GapEncoder, while being significantly faster.

Conclusion

In conclusion, TextEncoder provides powerful vectorization for text, but at the cost of longer computation times and the need for additional dependencies, such as torch. StringEncoder represents a simpler alternative that can provide good performance at a fraction of the cost of more complex methods.