Linking losses for density ratio and class-probability estimation

Authors: Aditya Menon, Cheng Soon Ong

ICML 2016 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 8. Experimental results We present experiments2 evincing three aspects of our analysis: first, that a loss weight function w DR(ρ) dictates the range of density ratio values it focusses on; second, that existing proper losses are viable for DRE in the context of covariate shift adaptation; third; that the new application of the LSIF loss to ranking the best problems holds promise.
Researcher Affiliation Collaboration Aditya Krishna Menon ADITYA.MENON@DATA61.CSIRO.AU Cheng Soon Ong CHENGSOON.ONG@ANU.EDU.AU Data61 and the Australian National University, Canberra, ACT, Australia
Pseudocode No The paper contains mathematical derivations and equations but does not include any explicit pseudocode blocks or algorithm listings.
Open Source Code Yes 2Scripts available at first author s webpage.
Open Datasets Yes The first dataset (poly) follows the example from Shimodaira (2000); Huang et al. (2007). ... The second dataset (amazon) is the real-world Amazon review data from Blitzer et al. (2007); we used the processed data as provided by Chen et al. (2012). ... We compare these losses on several standard benchmark datasets with binary labels.
Dataset Splits Yes We set n Src = 200, n Tar = 200, and n Eval = 2000. ... We train on n Src = 3000 samples from the book domain, and test on n Tar = 3000, n Eval = 2000 samples from the electronics domain. ... Each dataset was split in the ratio 2:1, with all instances normalised to lie in the ℓ2 ball. ... for each split, we performed 5-fold cross-validation to tune the strength of regularisation...
Hardware Specification No The paper does not provide any specific hardware details such as GPU/CPU models, memory, or specific computing environments used for running the experiments.
Software Dependencies No The paper describes various statistical methods and models but does not list any specific software dependencies or libraries with their version numbers (e.g., 'PyTorch 1.9' or 'scikit-learn 0.24').
Experiment Setup Yes We set n Src = 200, n Tar = 200, and n Eval = 2000. ... For each loss, we find min θ Θ 1 n Src x S ℓ1( θ, Φ(x) ) + 1 n Tar x S ℓ1( θ, Φ(x ) ) + λDR 2 ||θ||2 2, where Θ is unconstrained for the standard CPE losses, and Θ = {θ | ( z S S ) θ, Φ(z) 0} otherwise. ... for λWLS = 10 6 and λDR = 10 4. ... Each dataset was split in the ratio 2:1, with all instances normalised to lie in the ℓ2 ball. A regularised linear model trained to score instances, where for each split, we performed 5-fold cross-validation to tune the strength of regularisation from λ {2 20, 2 19, . . . , 215}.