Binary Losses for Density Ratio Estimation

Authors: Werner Zellinger

ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our novel loss functions outperform related approaches for resolving parameter choice issues of 11 deep domain adaptation algorithms in average performance across 484 real-world tasks including sensor signals, texts, and images.
Researcher Affiliation Academia Werner Zellinger LIT AI Lab, Institute for Machine Learning Johannes Kepler University Linz, Austria EMAIL
Pseudocode Yes Algorithm 1: Density ratio estimation by binary classification Setup : Loss function ℓ: { 1, 1} R R, invertible probability link function Ψ 1 : R [0, 1] and function class F {f : R R}. Input : Observations (xi)n i=1 P and (x i,1)m i=1 Q. Output: Estimator bβ : X R of Radon-Nikod ym derivative β = d P d Q. Step 1 : Compute binary classifier bf := arg min f F (x,y) (xi,1)n i=1 (x i, 1)m i=1 ℓ(y, f(x)) (2) and estimate by Ψ 1 bf(x) the probability2ρ(y = 1|x) that x is drawn from P. Step 2 : Construct bβ using Bayes theorem bβ(x) := Ψ 1( b f(x)) 1 Ψ 1( b f(x)) ρ(y=1|x) ρ(y= 1|x) = β(x).
Open Source Code Yes Our framework extends the Pytorch framework5 of Gruber et al. (2024), by our exponential weight method in Eq. 12. The Pytorch framework of Gruber et al. (2024) itself extends the one6of Dinu et al. (2023) which relies on the Ada Time benchmark suite7 of Ragab et al. (2023). In all our remaining parameters, we exactly follow the statistical setup and data splits of Dinu et al. (2023) together with the cross-validation based parameter choice for density ratio of Gruber et al. (2024), see Section C for details. 5https://github.com/lugruber/dre_iter_reg
Open Datasets Yes The first dataset is the Amazon Reviews dataset from Blitzer et al. (2006) which consists of text reviews from four domains: books (B), DVDs (D), electronics (E), and kitchen appliances (K). The second dataset is the Heterogeneity Human Activity Recognition dataset (HHAR) from Stisen et al. (2015) which investigates sensor-, deviceand workload-specific heterogeneities from 36 smartphones and smartwatches, consisting of 13 different device models from four manufacturers. Our third dataset is the Mini Domain Net dataset from Zellinger et al. (2021) which is a reduced version of the large scale Domain Net-2019 from Peng et al. (2019) consisting of six different image domains (Quickdraw, Real, Clipart, Sketch, Infograph, and Painting).
Dataset Splits Yes For training and selecting the density ratio estimation methods within this pipeline we follow Gruber et al. (2024) perform an additional train/val split of 80/20 on the datasets that are used for training the domain adaption methods.
Hardware Specification No The paper mentions training "9174 neural networks" and using "deep domain adaptation algorithms" but does not specify any particular hardware like GPU or CPU models, memory, or cloud computing instances used for these experiments.
Software Dependencies No The paper mentions using a 'Pytorch framework' and the 'BFGS algorithm', but does not provide specific version numbers for Pytorch or any other software libraries or dependencies used in the experiments.
Experiment Setup Yes The regularization parameter λ is selected from {10, 10^-1, 10^-3} and the number of iterations of the BFGS algorithm is fixed with 100. We follow Kanamori et al. (2012) in using the Gaussian kernel with kernel width set according to the median heuristic for all compared density ratio estimation methods.