Embeddings Drift

This notebooks provides an overview for using and understanding the embeddings drift check.

Structure:

• What Is Embeddings Drift?

• Loading the Data

• Run the Check

What Is Embeddings Drift?

Drift is simply a change in the distribution of data over time, and it is also one of the top reasons why machine learning model’s performance degrades over time.

In unstructured data such as text, we cannot measure the drift of the data directly, as there’s no “distribution” to measure. In order to measure the drift of the data, we can use the model’s embeddings as a proxy for the data distribution.

For more on embeddings, see the Text Embeddings Guide.

This detects embeddings drift by using a domain classifier. For more information on drift, see the Drift Guide.

How Does This Check Work?

This check detects the embeddings drift by using a domain classifier, and uses the AUC score of the classifier as the basis for the measure of drift. For efficiency, the check first reduces the dimensionality of the embeddings, and then trains the classifier on the reduced embeddings. By default, the check uses UMAP for dimensionality reduction, but you can also use PCA by setting the dimension_reduction_method parameter to pca.

The check also provides a scatter plot of the embeddings, which is a 2D representation of the embeddings space. This is achieved by further reducing the dimensionality, using UMAP.

How To Use Embeddings in Deepchecks?

See how to calculate default embeddings or setting your own embeddings in the Embeddings Guide.

from deepchecks.nlp.datasets.classification import tweet_emotion
from deepchecks.nlp.checks import TextEmbeddingsDrift

Load Data

For this example, we’ll use the tweet emotion dataset, which is a dataset of tweets labeled by one of four emotions: happiness, anger, sadness and optimism.

train_ds, test_ds = tweet_emotion.load_data()
train_embeddings, test_embeddings = tweet_emotion.load_embeddings(as_train_test=True)

# Set the embeddings in the datasets:
train_ds.set_embeddings(train_embeddings)
test_ds.set_embeddings(test_embeddings)

Let’s see how our data looks like:

train_ds.head()

	text	label	user_age	gender	days_on_platform	user_region
0	No but that's so cute. Atsu was probably shy a...	happiness	24.97	Male	2729	Middle East/Africa
1	Rooneys fucking untouchable isn't he? Been fuc...	anger	21.66	Male	1376	Asia Pacific
2	Tiller and breezy should do a collab album. Ra...	happiness	37.29	Female	3853	Americas
3	@user broadband is shocking regretting signing...	anger	15.39	Female	1831	Europe
4	@user Look at those teef! #growl	anger	54.37	Female	4619	Europe

Run Check

As there’s natural drift in this dataset, we can expect to see some drift in the data:

check = TextEmbeddingsDrift()
result = check.run(train_dataset=train_ds, test_dataset=test_ds)
result

n_jobs value -1 overridden to 1 by setting random_state. Use no seed for parallelism.
n_jobs value -1 overridden to 1 by setting random_state. Use no seed for parallelism.

Observing the results

We can see that the check found drift in the data. Moreover, we can investigate the drift by looking at the scatter plot, which is a 2D representation of the embeddings space. We can see that there are a few clusters in the graph where there are more tweets from the test dataset than the train dataset. This is a sign of drift in the data. By hovering over the points, we can see the actual tweets that are in the dataset, and see for example that there are clusters of tweets about motivational quotes, which are more common in the test dataset than the train dataset.

previous

Label Drift

next

NLP Property Drift