e-GAI: e-value-based Generalized $α$-Investing for Online False Discovery Rate Control
Authors: Yifan Zhang, Zijian Wei, Haojie Ren, Changliang Zou
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Both simulated and real data experiments demonstrate the advantages of both e-LORD and e-SAFFRON in FDR control and power. In this section, we evaluate the performance of our online testing framework on both synthetic and real data. We compare e-LORD, e-SAFFRON, p L-RAI, and p S-RAI with e-LOND, LORD++, SAFFRON, and Sup LORD in terms of FDR and power. |
| Researcher Affiliation | Academia | 1School of Mathematical Sciences, Shanghai Jiao Tong University, Shanghai, China 2School of Statistics and Data Sciences, LPMC, KLMDASR and LEBPS, Nankai University, Tianjin, China. Correspondence to: Haojie Ren <EMAIL>. |
| Pseudocode | Yes | Algorithm 1 e-LORD 1: Input: target FDR level α, initial allocation coefficient ω1 (0, 1), parameters φ and ψ (0, 1), sequence of e-values e1, e2, . . .. 2: Calculate α1 = αω1 and decide δ1 = 1 n e1 1 α1 3: Update R1 = δ1 and ω2 by (9); 4: for t = 2, 3, . . . do 5: Update testing level αt by (8); 6: Make decision δt = 1 n et 1 αt 7: Update Rt = Rt 1 + δt and ωt+1 by (9); 8: end for 9: Output: decision set {δ1, δ2, . . .}. Algorithm 2 e-SAFFRON 1: Input: target FDR level α, initial allocation coefficient ω1 (0, 1), parameters λ, φ and ψ (0, 1), sequence of e-values e1, e2, . . .. 2: Calculate α1 = α(1 λ)ω1 and decide δ1 = 1 n e1 1 α1 3: Update R1 = δ1 and ω2 by (9); 4: for t = 2, 3, . . . do 5: Update testing level αt by (10); 6: Make decision δt = 1 n et 1 αt 7: Update Rt = Rt 1 + δt and ωt+1 by (9); 8: end for 9: Output: decision set {δ1, δ2, . . .}. |
| Open Source Code | Yes | The code for all numerical experiments in this paper is available at https://github.com/zijianwei01/e-GAI. |
| Open Datasets | Yes | We analyze the NYC taxi dataset from the Numenta Anomaly Benchmark (NAB) repository (Lavin & Ahmad, 2015). |
| Dataset Splits | Yes | The first 2000 time points are taken as the initial sequence for model calibration. The calibration is implemented by using the first 1/3 observations, assuming the related period to be free of bubbles. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., GPU/CPU models, memory amounts, or detailed computer specifications) used for running its experiments. |
| Software Dependencies | Yes | We use default parameters from the R package online FDR (Robertson et al., 2022) for other benchmarks. R package 2.5.1. |
| Experiment Setup | Yes | We take ω1 = 0.005, φ = ψ = 0.5 in e-LORD and p L-RAI, and additionally λ = 0.1 in e-SAFFRON and p S-RAI, while we use default parameters from the R package online FDR (Robertson et al., 2022) for other benchmarks. The target FDR level is set as α = 0.05. |