Tailoring task arithmetic to address bias in models trained on multi-institutional datasets.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Natural Language Processing

Background:

• Multi-institutional datasets enhance machine learning model generalization but can introduce confounding by provenance bias.
• Deep learning models may learn data source indicators rather than true pathology, leading to poor generalization.
• Confounding by provenance is a significant concern in natural language processing due to pervasive linguistic indicators of data origin.

Purpose of the Study:

• To evaluate the effectiveness of task arithmetic in mitigating confounding by provenance in deep learning models.
• To propose novel model-agnostic methods for reducing provenance-related bias in natural language processing.
• To extend the task vectors approach to address bias in composite clinical datasets.

Main Methods:

• Representing trained deep network weights as task vectors for arithmetic composition.
• Developing Task Arithmetic for Provenance Effect Reduction (TAPER) and Dominance-Aligned Polarized Provenance Effect Reduction (DAPPER).
• Evaluating TAPER and DAPPER on RoBERTa and Llama-2 models across three datasets.

Main Results:

• Task vector approaches (TAPER and DAPPER) improved model robustness to confounding by provenance.
• Enhanced performance was observed at the extremes of distribution shift.
• Both RoBERTa and Llama-2 models demonstrated improved generalization with the proposed methods.

Conclusions:

• Adjusting for confounding by provenance is crucial, particularly in cases of extreme distribution shift.
• TAPER and DAPPER offer efficient strategies for mitigating bias in deep learning models for NLP.
• These methods provide a novel approach applicable to various biases in composite clinical data.