The scenario describes a classic machine learning problem where the model shows high performance on the majority class ("legitimate claim") but fails to correctly identify the minority class ("fraudulent claim"). The key piece of evidence is the SageMaker Clarify pretraining bias analysis, which revealed a "significant skew." This skew directly indicates a disproportionate representation of classes in the dataset, a condition known as Class Imbalance (CI). Models trained on such data are inherently biased toward the majority class to minimize overall error, leading to poor predictive power for the underrepresented minority class.

A. Insufficient text extraction would likely cause poor performance across all classes by creating incomplete feature sets, not a specific bias against the minority class.

B. Overfitting results in poor generalization on all new, unseen data. The model's good performance on the majority class contradicts a general overfitting problem.

C. A high learning rate typically leads to unstable training and poor convergence across all classes, rather than a systematic failure on only the minority class.

1. Amazon SageMaker Developer Guide: Under the section "Detect pre-training bias in datasets," the guide defines the Class Imbalance (CI) metric. It states, "A value of 1 indicates that the dataset is balanced... Values greater than 1 indicate that the dataset is imbalanced and that there is a bias towards the majority class." This directly links the "significant skew" found by Clarify to Class Imbalance.

Source: Amazon SageMaker Developer Guide, Section: "Pre-training bias metrics", Subsection: "Class Imbalance (CI)".

2. AWS Machine Learning Blog: In the article "Handling imbalanced data with Amazon SageMaker," it is explained that "In imbalanced classification problems, the model tends to be biased towards the majority class, causing poor classification of the minority class." This perfectly matches the behavior described in the question's scenario.

Source: AWS Machine Learning Blog, "Handling imbalanced data with Amazon SageMaker," Nov 12, 2021.

3. He, H., & Garcia, E. A. (2009). Learning from Imbalanced Data. This foundational academic paper explains that with imbalanced datasets, "the learning algorithm is not motivated to resolve the classification problem for the minority class examples since they contribute very little to the overall accuracy."

Source: IEEE Transactions on Knowledge and Data Engineering, vol. 21, no. 9, pp. 1263-1284, Sept. 2009, Section I (Introduction). DOI: https://doi.org/10.1109/TKDE.2008.239