Double-Weighting for Covariate Shift Adaptation

Authors: José I. Segovia-Martín, Santiago Mazuelas, Anqi Liu

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

Reproducibility Variable Result LLM Response
Research Type Experimental This section shows experimental results for the proposed approach in comparison with existing methods on synthetic and real datasets.
Researcher Affiliation Academia 1Basque Center for Applied Mathematics (BCAM), Bilbao, Spain 2IKERBASQUE-Basque Foundation for Science 3CS department, Whiting School of Engineering, Johns Hopkins University, Baltimore, Maryland, USA.
Pseudocode Yes Algorithm 1 The proposed algorithm: DW-GCS Input: Training samples (x1, y1), (x2, y2), . . . , (xn, yn) Testing instances xn+1, xn+2, . . . , xn+t, D Output: Weights ˆβ and ˆα Classifier parameters µ , Minimax risk R(U) 1: ˆβ, ˆα solution of (25) 2: τ 1 n Pn i=1 ˆβ(i)Φ(xi, yi) 3: λ solution of (31) 4: µ solution of (30) using (12) for 0-1-loss, and (13) for log-loss 5: R(U) τ Tµ + 1 t Pt i=1 ϕℓ(µ , xn+i, ˆα(i)) + λT|µ |
Open Source Code Yes The source code for the methods presented is publicly available in the library MRCpy (Bondugula et al., 2023) and the experimental setup in https://github.com/MachineLearningBCAM/MRCs-for-Covariate-Shift-Adaptation.
Open Datasets Yes For the experiments in Section 6, we have considered four binary classification datasets, available in the UCI repository (Dua & Graff, 2017)... In addition, we use the dataset News20groups that is intrinsically affected by covariate shift (Zhang et al., 2013).
Dataset Splits No The paper mentions "training and testing samples" and "100 random partitions" but does not explicitly specify a distinct validation set split (e.g., percentages, counts, or k-fold cross-validation setup) for model training or hyperparameter tuning in a reproducible manner. It states "standard cross-validation is not valid under covariate shift" for its approach.
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 No The paper mentions the library "MRCpy" but does not provide specific version numbers for any software dependencies, including MRCpy itself or other programming languages/libraries used.
Experiment Setup Yes For the results obtained using the flattening method in (Shimodaira, 2000) and the Ru LSIF method in (Yamada et al., 2011) we considered the hyperparameter γ = 0.5, which is the default value used in those papers. The table also shows the parameter σ used in the computation of the kernel matrix K for the Ru LSIF, KMM and DW-KMM methods, which is determined using the common heuristic based on nearest neighbors with K = 50, as is done in (Wen et al., 2014). and Specifically, we select the value of D to achieve the lowest minimax risk over a certain range D ≥ 1. and The second hyperparameter λ is determined solving min p,λ 1Tλ s.t. P y∈Y p(y|xn+i)Φα(xn+i, y) τ + λ and P y∈Y p(y|xn+i) = 1/t for i = 1, . . . , t (31)