Combating Bias in AI-Driven Drug Discovery: A Novel Approach to Remove Irrelevant Attributes

In this paper, the authors propose a new method called InfoCORE that helps remove biases from machine learning models by maximizing fairness in representation. They use a technique called conditional mutual information to establish a lower bound for the training objective, which is similar to the InfoNCE loss but with unequally weighted negative samples. The authors show that this approach can be used to counteract biases from irrelevant attributes and improve the fairness of machine learning models in various applications, including image classification, natural language processing, and recommender systems.
To understand how InfoCORE works, let’s consider an analogy. Imagine you have a big box of chocolates with different flavors, and you want to ensure that each piece of chocolate is fair and equal in size. You could do this by weighing each piece individually, but this would be time-consuming and difficult. Instead, you can use a tool called a "fairness meter" that measures the weight of each piece relative to the others. The fairness meter tells you how much each piece needs to be adjusted to ensure fairness.
Similarly, in machine learning, we have a large dataset with many features (or attributes) that can affect the outcome of the model. Some of these features may be irrelevant or biased, which can lead to unfair outcomes. InfoCORE helps identify and remove these biases by using a fairness meter that measures the relevance of each feature. The algorithm adjusts the weights of each feature based on their relevance, ensuring that the model treats all individuals equally and avoids biases.
The authors demonstrate the effectiveness of InfoCORE in various applications, including image classification, natural language processing, and recommender systems. They show that by using InfoCORE, they can improve the fairness of these models without sacrificing their accuracy.
In summary, InfoCORE is a powerful tool for removing biases from machine learning models and improving their fairness. By using a fairness meter to measure the relevance of each feature, it ensures that the model treats all individuals equally and avoids biases. This can have significant implications in various applications, including image classification, natural language processing, and recommender systems.