Property Label Correlation#

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

Structure:

What is the purpose of the check?#

The check estimates for every image property (such as brightness, contrast 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 image properties such as color, brightness, aspect ratio and more.

This is a critical problem, sometimes referred to as shortcut learning, where the model is likely to learn this property instead of the actual visual characteristics of each class, as it’s easier to do so. In this case, the model will show high performance on images taken in similar conditions, but will fail in the wild, where the simple properties don’t hold true. This kind of correlation will likely stay hidden without this check until tested in the wild.

A famous example is the case of wolves vs. dogs classification, where a model needs to classify whether an image contains a wolf or a dog, and can learn to do it by the background instead of the actual animal - in the dataset most of the wolves were photographed in the snow and therefore had a white background while all the dogs were photographed in the grass and therefore had a green background.

The check is based on calculating the predictive power score (PPS) of each image property. For more details you can read here how the PPS is calculated.

What is a problematic result?#

Image properties with a high predictive score can indicate that there is a bias in the dataset, as a single property can be used to predict the label successfully (e.g. using simple classic ML algorithms).

This means that a deep learning algorithm may accidentally learn these properties instead of more accurate complex abstractions. For example, in the dataset of wolves and dogs photographs, the brightness of the image may be used to predict the label “wolf” easily.

How do we calculate the predictive power for different vision tasks?#

  • For classification tasks, this check uses PPS to predict the class by image properties.

  • For object detection tasks, this check uses PPS to predict the class of each bounding box, by the image properties of that specific bounding box. This means that for each image, this check crops all the sub-images defined by bounding boxes, and uses them as inputs as though they were regular classification dataset images.

How is the Predictive Power Score (PPS) calculated?#

The properties’ predictive score results in a numeric score between 0 (feature has no predictive power) and 1 (feature can fully predict the label alone).

The process of calculating the PPS is the following:

The properties’ predictive score results in a numeric score between 0 (feature has no predictive power) and 1 (feature can fully predict the label alone).

The process of calculating the PPS is the following:

  1. Extract from the data only the label and the feature being tested

  2. Drop samples with missing values

  3. Keep 5000 (this is configurable parameter) samples from the data

  4. Preprocess categorical columns. For the label using sklearn.LabelEncoder and for the feature using sklearn.OneHotEncoder

  5. Partition the data with 4-fold cross-validation

  6. Train decision tree

  7. Compare the trained model’s performance with naive model’s performance as follows:

Regression: The naive model always predicts the median of the label column, the metric being used is MAE and the PPS calculation is: 1 - \frac{\text{MAE model}}{\text{MAE naive}}

Classification: The naive model always predicts the most common class of the label column, The metric being used is F1 and the PPS calculation is: \frac{\text{F1 model} - \text{F1 naive}}{1 - \text{F1 naive}}

Note

All the PPS parameters can be changed by passing to the check the parameter ppscore_params

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#

In this example we will run the check on the dataset of wolves vs. dogs. For example purposes we picked 10 images of dogs and 10 images of wolves out of the full dataset. The original data was downloaded from https://www.kaggle.com/datasets/harishvutukuri/dogs-vs-wolves, which is licensed under DbCL v1.0.

from deepchecks.vision.checks import PropertyLabelCorrelation
from deepchecks.vision.simple_classification_data import classification_dataset_from_directory
import albumentations as A
import urllib.request
import zipfile

url = 'https://figshare.com/ndownloader/files/36671001'
urllib.request.urlretrieve(url, 'wolves_vs_dogs_mini.zip')

with zipfile.ZipFile('wolves_vs_dogs_mini.zip', 'r') as zip_ref:
    zip_ref.extractall('.')

dataset = classification_dataset_from_directory(
    'wolves_vs_dogs_mini', object_type='VisionData', transforms=A.Resize(64, 64))

check_result = PropertyLabelCorrelation().run(dataset)
check_result.show()
Validating Input:
|     | 0/1 [Time: 00:00]
Validating Input:
|#####| 1/1 [Time: 00:00]

Ingesting Batches:
|     | 0/1 [Time: 00:00]

Ingesting Batches:
|#####| 1/1 [Time: 00:00]


Computing Check:
|     | 0/1 [Time: 00:00]


Computing Check:
|#####| 1/1 [Time: 00:00]


Computing Check:
|#####| 1/1 [Time: 00:00]
Property Label Correlation


We can see that both the “Brightness” property and the “Mean Green Relative Intensity” property have a significant ability to predict the label.

This is as expected - pictures of wolves have higher brightness because they appear with a white background, while dogs appear with a green background, making “Green-ness” a strong predictor for an image containing a dog. Using this check we can be made aware of these artifacts, and can solve them (for example by collecting images with different backgrounds) before training any kind of model.

Define a condition#

We can define a condition to verify that the results are less than a certain threshold.

check_result = PropertyLabelCorrelation().add_condition_property_pps_less_than(0.5).run(dataset)
check_result.show(show_additional_outputs=False)
Validating Input:
|     | 0/1 [Time: 00:00]
Validating Input:
|#####| 1/1 [Time: 00:00]

Ingesting Batches:
|     | 0/1 [Time: 00:00]

Ingesting Batches:
|#####| 1/1 [Time: 00:00]


Computing Check:
|     | 0/1 [Time: 00:00]


Computing Check:
|#####| 1/1 [Time: 00:00]


Computing Check:
|#####| 1/1 [Time: 00:00]
Property Label Correlation


We can now see that the condition failed because the results here are above the threshold.

Total running time of the script: ( 0 minutes 1.611 seconds)

Gallery generated by Sphinx-Gallery