Online Adaptive Anomaly Thresholding with Confidence Sequences

Authors: Sophia Huiwen Sun, Abishek Sankararaman, Balakrishnan Murali Narayanaswamy

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We complement our theoretical results with empirical evidence that our method outperforms commonly used baselines across synthetic and real world datasets.
Researcher Affiliation Collaboration 1University of California, San Diego. Work done when interning with Amazon Web Services. 2Amazon Web Services, Santa Clara CA.
Pseudocode Yes Algorithm 1 Decision making with confidence set; Algorithm 3 Thresholding without offline data; Algorithm 5 General thresholding algorithm
Open Source Code No The code to generate the synthetic dataset as well as implementations of our algorithms will be open sourced.
Open Datasets Yes We tested our algorithms on the MNIST dataset for one-class anomaly detection to demonstrate their real-world efficacy.
Dataset Splits No The paper mentions 'first 1500 as holdout' for Forest Cover and 'first 300 as holdout' for Mammography, but does not provide general or comprehensive train/validation/test dataset splits with specific percentages or counts for all datasets used, nor for the main MNIST experiments or real-world datasets.
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, memory, or cloud instance types) used for running its experiments.
Software Dependencies No The paper does not specify any software dependencies with version numbers needed for replication.
Experiment Setup Yes The metrics are abstention percentages (Abs. %) and mistake counts (FP + FN), averaged over 1000 streams each with 2000 samples drawn from Normal distributions with random parameters, with p = 1 10 2 and α = 10 3.