Boston University
Active Feature Acquisition for efficient & transparent medical diagnoses
Abstract
dc:description.abstractIn many real-world machine learning scenarios, such as clinical diagnostics, input features are acquired sequentially and selectively, due to practical constraints like cost and time. This necessitates accurate predictions from partially observed data and the strategic selection of informative features. Active Feature Acquisition (AFA), also referred to as Dynamic Feature Selection (DFS), addresses this challenge by dynamically determining which features to acquire at each step. We propose three novel methodologies to enhance the performance and robustness of Active Feature Acquisition (AFA) frameworks. First, we introduce a feature imputation strategy that supplies imputed values for unobserved features during the estimation of conditional mutual information (CMI), serving as a heuristic to support more informed acquisition decisions by approximating missing feature values. Second, we develop a non-greedy acquisition mechanism that leverages the predicted discounted cumulative CMI to forecast the value of future acquisitions, allowing for more strategic and forward-looking feature selection. Finally, we present an ensemble acquisition policy that trains multiple policies and averages their outputs, thereby improving the robustness, stability, and generalization of the feature selection process across different datasets and tasks. Empirical evaluations show that these strategies significantly enhance both acquisition efficiency and predictive accuracy.
Degree
thesis:*- Name thesis:degree_name
- Master of Science
- Level thesis:degree_level
- masters
- Discipline thesis:degree_discipline
- Computer Science
- Grantor
- Boston University
- Year dc:date.issued
- 2025
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Saichandran, Ketan Suhaas
- Advisor dc:contributor.advisor
-
- Kolachalama, Vijaya B.
- ORCID dc:identifier.orcid
Subjects
dc:subject × 8Rights
dc:rights- Statement dc:rights
-
- Attribution 4.0 International
- Licence dc:rights.uri
- Language dc:language.iso
- en_US
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/2144/53424