Interpretability-driven active feature acquisition in learning systems
Keywords: Active feature acquisition, Model interpretability
TL;DR: We utilized explanation methods to identify instance-wise feature importance and proposed an alternative active feature acquisition framework
Abstract: In real-world applications like medicine, machine learning models must often work with a limited number of features due to the high cost and time required to acquire all relevant data. While several static feature selection methods exist, they are suboptimal due to their inability to adapt to varying feature importance across different instances. A more flexible approach is active feature acquisition (AFA), which dynamically selects features based on their relevance for each individual case. Here, we introduce an AFA framework that leverages SHapley Additive exPlanations (SHAP) to generate instance-specific feature importance rankings. By reframing the AFA problem as a feature prediction task, we propose a policy network based on a decision transformer architecture, trained to predict the next most informative feature based on SHAP values. This method allows us to sequentially acquire features in order of their predictive significance, resulting in more efficient feature selection and acquisition. Extensive experiments across multiple datasets show that our approach achieves superior performance compared to current state-of-the-art AFA techniques, both in terms of predictive accuracy and feature acquisition efficiency. These results demonstrate the potential of explainability-driven AFA for applications where feature acquisition cost is a critical consideration.
Primary Area: interpretability and explainable AI
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 10870
Loading