Counterfactual Logit Pairing (CLP) is a technique within the TensorFlow Model Remediation Library that seeks to ensure that a model’s prediction doesn’t change when a sensitive attribute referenced in an example is either removed or replaced. For example, in a toxicity classifier, examples such as "I am a man" and "I am a lesbian" should not have a different prediction of toxicity.
When should you use Counterfactual Logit Pairing?
CLP addresses the scenario where a change in a sensitive attribute referenced in a feature changes the prediction (when the prediction should not have changed). In doing so, it attempts to answer the question: Is this model susceptible to changing its prediction based solely on the presence of an identity attribute? See the research paper for details on counterfactual fairness.
This issue was seen in the Perspective API, an ML tool used by developers and publishers to analyze the content of comments for potentially offensive or toxic text. The Perspective API takes comment text as input and returns a score from 0 to 1 as an indication of the probability that the comment is toxic. For example, a comment like “You are an idiot” may receive a probability score of 0.8 for toxicity, indicating how likely it is that a reader would perceive that comment as toxic.
After the initial launch of the Perspective API, external users discovered a positive correlation between identity terms containing information on race or sexual orientation and the predicted toxicity score. For example, the phrase "I am a lesbian" received a score of 0.51, while “I am a man” received a lower score of 0.2. In this case, the identity terms were not being used pejoratively, so there should not be such a significant difference in the score. For more information on the Perspective API, see the blog post on unintended bias and identity terms.
How can I measure the effect of Counterfactual Logit Pairing?
If you have assessed your machine learning model and determined that changes in predictions due to changes in specific sensitive attributes would be harmful, then you should measure the prevalence of this issue. In the case of a binary or multi-class classifier, a flip is defined as a classifier giving a different decision (such as changing a prediction from toxic to not toxic) when the sensitive attribute referenced in the example changes. When assessing the prevalence of flips, you can look at flip count and flip rate. By taking into account the potential user harm caused by a flip and the frequency that flips occur, you can determine if this is an issue that should be addressed by applying CLP. For more information about these metrics, refer to the Fairness Indicators guide.
On what model types can I apply Counterfactual Logit Pairing?
This technique can be used with binary and multi-class classifiers of different types of data such as text, images, and videos.
When is Counterfactual Logit Pairing not right for me?
CLP is not the right method for all situations. For example, it is not relevant if the presence or absence of an identity term legitimately changes the classifier prediction. This may be the case if the classifier aims to determine whether the feature is referencing a particular identity group. This method is also less impactful if the unintended correlation between classifier result and identity group has no negative repercussions on the user.
CLP is useful for testing whether a language model or toxicity classifier is changing its output in an unfair way (for example classifying a piece of text as toxic) simply because terms like “Black”, “gay”, “Muslim” are present in the text. CLP is not intended for making predictions about individuals, for example by manipulating the identity of an individual. See this paper for a more detailed discussion.
It is important to keep in mind that CLP is one technique in the Responsible AI Toolkit that is specifically designed to address the situation where sensitive attributes referenced in features changes the prediction. Depending on your model and use case, it may also be important to consider whether there are performance gaps for historically marginalized groups, particularly as CLP may affect group performance. This can be assessed with Fairness Indicators and addressed by MinDiff that is also in the TensorFlow Model Remediation Library.
You should also consider whether your product is an appropriate use for machine learning at all. If it is, your machine learning workflow should be designed to known recommended practices such as having a well defined model task and clear product needs.
How does Counterfactual Logit Pairing work?
CLP adds a loss to the original model that is provided by logit pairing an original and counterfactual example from a dataset. By calculating the difference between the two values, you penalize the differences of the sensitive terms that are causing your classifier prediction to change. This work was based on research on adversarial logit pairing and counterfactual logit pairing.