Learning to Maximize Mutual Information for Dynamic Feature Selection

Authors: Ian Connick Covert, Wei Qiu, Mingyu Lu, Na Yoon Kim, Nathan J White, Su-In Lee

ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our experiments evaluate the proposed method on numerous datasets, and the results show that it outperforms many recent dynamic and static feature selection methods.
Researcher Affiliation Academia 1Paul G. Allen School of Computer Science & Engineering, University of Washington 2Department of Emergency Medicine, University of Washington.
Pseudocode Yes Algorithm 1: Training pseudocode
Open Source Code Yes Code for reproducing our experiments is available online: https://github.com/iancovert/ dynamic-selection.
Open Datasets Yes Next, we conducted experiments using three publicly available tabular datasets: spam classification (Dua & Graff, 2017), particle identification (Mini Boo NE) (Roe et al., 2005) and diabetes diagnosis (Miller, 1973).
Dataset Splits Yes We used the standard train-test splits, and we split the train set to obtain a validation set with the same size as the test set (10,000 examples).
Hardware Specification No The paper does not provide specific hardware details such as GPU/CPU models, processor types, or memory amounts used for running the experiments.
Software Dependencies No The paper mentions using 'Py Torch' and the 'Captum package' but does not specify their exact version numbers (e.g., PyTorch 1.9) used in the experiments.
Experiment Setup Yes The dropout probability is set to 0.3, the networks have two hidden layers of width 128, and we performed early stopping using the validation loss. For MNIST, we used fully connected architectures with two layers of width 512 and the dropout probability set to 0.3. For CIFAR-10, we used a shared Res Net backbone (He et al., 2016b) consisting of several residually connected convolutional layers.