Property Label Correlation#

This notebook provides an overview for using and understanding the “Property Label Correlation” check.


What Is The Purpose of the Check?#

The check estimates for every text property (such as text length, language etc.) its ability to predict the label by itself.

This check can help find a potential bias in the dataset - the labels being strongly correlated with simple text properties such as percentage of special characters, sentiment, toxicity and more.

This is a critical problem that can result in a phenomenon called “shortcut learning”, where the model is likely to learn this property instead of the actual textual characteristics of each class, as it’s easier to do so. In this case, the model will show high performance on text collected under similar conditions (e.g. same source), but will fail to generalize on other data (for example, when production receives new data from another source). This kind of correlation will likely stay hidden without this check until tested on the actual problem data.

For example, in a classification dataset of true and false statements, if only true facts are written in detail, and false facts are written in a short and vague manner, the model might learn to predict the label by the length of the statement, and not by the actual content. In this case, the model will perform well on the training data, and may even perform well on the test data, but will fail to generalize to new data.

The check is based on calculating the predictive power score (PPS) of each text property. In simple terms, the PPS is a metric that measures how well can one feature predict another (in our case, how well can one property predict the label). For further information about PPS you can visit the ppscore github or the following blog post: RIP correlation. Introducing the Predictive Power Score

Run the Check#

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

# 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.

# Load Data:
dataset = tweet_emotion.load_data(as_train_test=False)

Let’s see how our data looks like:

text label user_age gender days_on_platform user_region
0 “Worry is a down payment on a problem you may ... optimism 30.73 Male 5614 Americas
1 My roommate: it's okay that we can't spell bec... anger 42.29 Female 4308 Europe
2 No but that's so cute. Atsu was probably shy a... happiness 24.97 Male 2729 Middle East/Africa
3 Rooneys fucking untouchable isn't he? Been fuc... anger 21.66 Male 1376 Asia Pacific
4 it's pretty depressing when u hit pan on ur fa... sadness 35.07 Female 4631 Europe

Now lets run the check:

result = PropertyLabelCorrelation().run(dataset)
Property-Label Correlation

We can see that in our example of tweet emotion dataset, the label is correlated with the “sentiment” property, which makes sense, as the label is the emotion of the tweet, and the sentiment expresses whether the tweet is positive or negative. Also, there’s some correlation with the “toxciity” property, which is a measure of how toxic the tweet is. This is also reasonable, as some emotions are more likely to be expressed in a toxic way. However, these correlation may indicate that a model may learn to predict the label by curse words, for instance, instead of the actual content of the tweet, which could lead it to fail on new tweets that don’t contain curse words.

Total running time of the script: (0 minutes 0.390 seconds)

Gallery generated by Sphinx-Gallery