This paper proposes a method for jointly training an active feature acquisition policy network and classifier to perform cost-effective instance-based active feature acquisition for diagnosis. The method sequentially asks questions or examines the patient based on the current state of available information, acquiring only necessary features. It was tested on the physioNet 2012 medical dataset and aims to model how doctors make diagnoses by acquiring features as needed rather than statically.
1 of 2
Download to read offline
More Related Content
H. Shim, NeurIPS 2018, MLILAB, KAIST AI
1. Joint Active Feature Acquisition and
Classification with Variable-Size Set
Encoding
H. Shim, S. Hwang and E. Yang
NeurIPS 2018
Machine Learning & Intelligence Laboratory
2. Active Feature Acquisition & Classification
In some cases, we should consider cost and effect of features
e.g. Medical case
How doctors make a diagnosis?
Sequentially ask a questions or examine
Based on current information (state)
By necessity (鏝 static)
Construct Cost-effective, Instance-based active feature acquisition model
Jointly train the feature acquisition policy network & the classifier
Results on physioNet 2012 dataset
2Shim et al. NeurIPS 2018
