arxiv: 2605.05591 · v1 · submitted 2026-05-07 · 📊 stat.ML · cs.LG· stat.CO

Recognition: unknown

In-Context Positive-Unlabeled Learning

Siyan Liu , Yi Chang , Manli Cheng , Qinglong Tian , Pengfei Li

Authors on Pith no claims yet

Pith reviewed 2026-05-08 05:51 UTC · model grok-4.3

classification 📊 stat.ML cs.LGstat.CO

keywords positive-unlabeled learningin-context learningtransformer modelsemi-supervised classificationtabular datasynthetic pretrainingstructural causal models

0 comments

The pith

A pretrained transformer performs positive-unlabeled classification in one forward pass by receiving labeled positives and unlabeled samples together as input.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents PUICL, a transformer model pretrained on synthetic positive-unlabeled datasets created from randomly generated structural causal models. At inference, the model takes the small set of known positives plus the pool of unlabeled points as a single sequence and directly outputs class probabilities for the unlabeled points. This replaces the dataset-specific training loops or iterative label estimation used by conventional PU methods. The approach matters because many practical classification problems supply only confirmed positives and a mixed unlabeled pool, and rapid deployment across many such problems requires avoiding per-task optimization.

Core claim

PUICL is a pretrained transformer that solves PU classification entirely through in-context learning. It is pretrained on synthetic PU datasets generated from randomly instantiated structural causal models, exposing it to a wide range of feature-label relationships and class-prior configurations. At inference time, PUICL receives the labeled positives and the unlabeled samples as a single input and returns class probabilities for the unlabeled rows in one forward pass, with no gradient updates or per-task fitting. On 20 semi-synthetic PU benchmarks derived from the UCI Machine Learning Repository, OpenML, and scikit-learn, PUICL outperforms four standard PU learning baselines in average AUC

What carries the argument

The PUICL transformer that ingests a concatenated sequence of labeled positives and unlabeled rows and directly emits label probabilities for the unlabeled rows via in-context learning.

If this is right

PU classification tasks can be solved without gradient updates or iterative optimization on each new dataset.
A single pretrained model can handle many different class-prior and feature-label configurations after exposure only to synthetic causal data.
The in-context learning paradigm extends from fully supervised tabular prediction to the positive-unlabeled semi-supervised setting.
Deployment across numerous PU problems becomes feasible with only one inference pass per task.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Domains such as medical screening or fraud detection, where confirmed negatives are scarce, could apply the same model to new data streams without retraining.
The method implies that explicit modeling of label noise or class priors can be replaced by implicit learning from synthetic causal examples presented in context.
Further scaling the diversity of the synthetic pretraining distribution could increase robustness when real data deviates from the causal-model assumptions.

Load-bearing premise

Pretraining on synthetic PU datasets generated from randomly instantiated structural causal models exposes the model to a sufficiently wide range of feature-label relationships and class-prior configurations that it generalizes to real-world PU problems without any dataset-specific adaptation.

What would settle it

On a held-out real-world PU dataset drawn from a domain absent from the pretraining distribution, if PUICL achieves lower average AUC than the best baseline method that performs dataset-specific fitting, the claim that one forward pass suffices without adaptation would be refuted.

Figures

Figures reproduced from arXiv: 2605.05591 by Manli Cheng, Pengfei Li, Qinglong Tian, Siyan Liu, Yi Chang.

**Figure 1.** Figure 1: Architecture of the in-context PU learning model. The input PU dataset view at source ↗

read the original abstract

Positive-unlabeled (PU) learning addresses binary classification when only a set of labeled positives is available alongside a pool of unlabeled samples drawn from a mixture of positives and negatives. Existing PU methods typically require dataset-specific training or iterative optimization, which limits their applicability when many tasks must be solved quickly or with little tuning. We introduce PUICL, a pretrained transformer that solves PU classification entirely through in-context learning. PUICL is pretrained on synthetic PU datasets generated from randomly instantiated structural causal models, exposing it to a wide range of feature-label relationships and class-prior configurations. At inference time, PUICL receives the labeled positives and the unlabeled samples as a single input and returns class probabilities for the unlabeled rows in one forward pass, with no gradient updates or per-task fitting. On 20 semi-synthetic PU benchmarks derived from the UCI Machine Learning Repository, OpenML, and scikit-learn, PUICL outperforms four standard PU learning baselines in average AUC and accuracy, and is competitive on F1-score. These results show that the in-context learning paradigm extends naturally beyond fully supervised tabular prediction to the semi-supervised PU setting.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

PUICL pretrains a transformer on synthetic SCM data for zero-shot in-context PU classification and claims gains on 20 benchmarks, but the transfer assumption and missing experimental details are the main things to check.

read the letter

The paper's main move is to pretrain a transformer on positive-unlabeled datasets drawn from randomly generated structural causal models, then feed labeled positives plus unlabeled samples into the model at inference time to get class probabilities in a single forward pass. No per-task training or optimization is needed after pretraining. This is a concrete attempt to bring the in-context learning pattern into the PU setting, where most existing methods still require dataset-specific fitting or iterative steps. The claim is that this approach beats four standard baselines on average AUC and accuracy across 20 semi-synthetic benchmarks built from UCI, OpenML, and scikit-learn data, while staying competitive on F1. If the numbers hold, the practical upside is clear: you could apply PU classification to new tables without retraining or tuning each time. The pretraining strategy itself is straightforward and avoids circularity since the synthetic data is generated independently of the evaluation sets. The soft spot is the generalization step. The model sees only data from random SCMs during pretraining, yet is tested on semi-synthetic versions of real tabular datasets. Those two distributions can differ in variable types, noise patterns, and feature correlations, so it is not obvious that the in-context predictions reflect a learned general solution to PU problems rather than incidental matches. The abstract also gives no error bars, significance tests, or implementation details on how the baselines were run, which leaves the reported outperformance hard to evaluate. A reader working on in-context methods or on practical semi-supervised tools would find this worth reading to see whether the synthetic pretraining actually transfers. The work is coherent on its own terms and engages the literature, so it deserves a serious referee to examine the experimental setup and the transfer claim directly.

Referee Report

3 major / 2 minor

Summary. The paper introduces PUICL, a transformer pretrained on synthetic positive-unlabeled (PU) datasets generated from randomly instantiated structural causal models. At inference, it performs PU classification entirely via in-context learning by taking labeled positives and unlabeled samples as input and outputting class probabilities for the unlabeled rows in a single forward pass, without gradient updates or per-task fitting. The central empirical claim is that on 20 semi-synthetic PU benchmarks derived from UCI, OpenML, and scikit-learn datasets, PUICL outperforms four standard PU baselines in average AUC and accuracy while remaining competitive on F1-score.

Significance. If the results hold under rigorous validation, the work would demonstrate that in-context learning can be successfully extended to the PU setting, providing a training-free method for solving multiple PU tasks rapidly. This could be impactful for tabular semi-supervised problems where dataset-specific optimization is costly. The pretraining strategy on diverse synthetic SCMs is a notable strength for exposing the model to varied class priors and feature-label relationships.

major comments (3)

[Experimental Results] Experimental Results section: The abstract and results claim outperformance on 20 benchmarks in AUC and accuracy, but provide no details on benchmark construction (e.g., how positives are selected from the original datasets, how class priors are set in the semi-synthetic versions, or the exact mixture ratios for unlabeled data). This information is load-bearing for assessing whether the gains reflect genuine in-context PU inference rather than artifacts of benchmark design.
[Methods and Pretraining] Methods and Pretraining section: No analysis or ablation is presented on the distribution shift between the pretraining distribution (randomly instantiated SCMs, typically yielding continuous variables with known causal structure) and the evaluation distribution (semi-synthetic versions of real tabular UCI/OpenML data, which may contain discrete features, domain artifacts, or different dependence structures). This is central to the claim that the model generalizes without adaptation.
[Results] Results section, performance tables: The reported average AUC/accuracy gains lack accompanying statistical significance tests, standard errors, or per-benchmark breakdowns with error bars. Without these, it is impossible to determine whether the outperformance is robust or driven by a few datasets, undermining the cross-benchmark superiority claim.

minor comments (2)

[Abstract] Abstract: The four standard PU learning baselines are not named; listing them (e.g., nnPU, PUbN, etc.) would improve clarity for readers unfamiliar with the PU literature.
[Methods] Notation: The description of the input format for in-context learning (how positives and unlabeled samples are concatenated or prompted) could be made more precise with an explicit example or pseudocode.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive and detailed feedback. We address each major comment below and indicate the revisions we will make to the manuscript.

read point-by-point responses

Referee: Experimental Results section: The abstract and results claim outperformance on 20 benchmarks in AUC and accuracy, but provide no details on benchmark construction (e.g., how positives are selected from the original datasets, how class priors are set in the semi-synthetic versions, or the exact mixture ratios for unlabeled data). This information is load-bearing for assessing whether the gains reflect genuine in-context PU inference rather than artifacts of benchmark design.

Authors: We agree that the current manuscript provides insufficient detail on benchmark construction. In the revised version we will add a dedicated subsection under Experimental Results that specifies: (i) positives are obtained by randomly subsampling a fraction of the original positive instances according to the source dataset labels; (ii) class priors are set either to the empirical positive rate of the source data or to controlled values (0.2–0.5) for semi-synthetic variants; and (iii) unlabeled sets are formed by mixing the remaining positives and negatives at explicit ratios (e.g., 30 % positive). These additions will make the evaluation protocol fully reproducible and allow readers to judge whether the reported gains arise from genuine in-context PU inference. revision: yes
Referee: Methods and Pretraining section: No analysis or ablation is presented on the distribution shift between the pretraining distribution (randomly instantiated SCMs, typically yielding continuous variables with known causal structure) and the evaluation distribution (semi-synthetic versions of real tabular UCI/OpenML data, which may contain discrete features, domain artifacts, or different dependence structures). This is central to the claim that the model generalizes without adaptation.

Authors: The referee rightly notes the absence of explicit distribution-shift analysis. While our pretraining deliberately samples SCMs with varied causal structures and includes both continuous and discretized variables, we did not quantify the shift to the UCI/OpenML evaluation sets. In the revision we will insert a discussion paragraph in the Methods section explaining how in-context learning enables adaptation to new distributions at inference time, and we will add a limited ablation on a subset of benchmarks comparing performance before and after feature discretization. A full-scale ablation across all 20 benchmarks is computationally intensive and will be noted as future work; the current revision therefore provides a targeted rather than exhaustive treatment. revision: partial
Referee: Results section, performance tables: The reported average AUC/accuracy gains lack accompanying statistical significance tests, standard errors, or per-benchmark breakdowns with error bars. Without these, it is impossible to determine whether the outperformance is robust or driven by a few datasets, undermining the cross-benchmark superiority claim.

Authors: We accept this criticism. The revised manuscript will expand the Results section to include: paired Wilcoxon signed-rank tests comparing PUICL against each baseline across the 20 benchmarks for AUC, accuracy, and F1; standard errors computed over multiple random seeds where applicable; and an appendix table with per-benchmark scores together with error bars. These additions will allow readers to assess both average performance and variability, thereby strengthening the cross-benchmark superiority claim. revision: yes

Circularity Check

0 steps flagged

No circularity: pretraining distribution independent of evaluation benchmarks; results not reduced to fitted inputs by construction.

full rationale

The paper pretrains a transformer on synthetic PU datasets generated from randomly instantiated structural causal models, then applies it via in-context learning (one forward pass, no per-task updates) to separate semi-synthetic benchmarks derived from UCI, OpenML, and scikit-learn repositories. Performance metrics (AUC, accuracy, F1) are reported on these held-out benchmarks without any equation or procedure that fits parameters to the evaluation data and then renames the fit as a prediction. No self-citations, uniqueness theorems, or ansatzes are invoked to justify the core method; the derivation chain from pretraining to inference remains self-contained against external data sources.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The central claim rests on one key domain assumption about the coverage of synthetic data and introduces the PUICL model itself as the primary new entity; no explicit free parameters are named in the abstract.

axioms (1)

domain assumption Synthetic PU datasets generated from randomly instantiated structural causal models cover a sufficiently broad range of feature-label relationships and class priors to enable generalization to real PU problems.
Invoked to justify why pretraining on synthetics substitutes for per-task training on real data.

invented entities (1)

PUICL pretrained transformer no independent evidence
purpose: To perform in-context positive-unlabeled classification after exposure to synthetic data.
The model is the core technical contribution; no independent evidence outside the reported experiments is provided.

pith-pipeline@v0.9.0 · 5503 in / 1457 out tokens · 63625 ms · 2026-05-08T05:51:42.307627+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

279 extracted references · 103 canonical work pages · 1 internal anchor

[1]

and Niu, Gang and Sugiyama, Masashi , journal=

du Plessis, Marthinus C. and Niu, Gang and Sugiyama, Masashi , journal=. Analysis of Learning from Positive and Unlabeled Data , volume =
[2]

and Carroll, Raymond J

Duan, Rui and Ning, Yang and Wang, Shuang and Lindsay, Bruce G. and Carroll, Raymond J. and Chen, Yong , title =. Biometrics , volume =. 2019 , month =. doi:10.1111/biom.13204 , eprint =

work page doi:10.1111/biom.13204 2019
[3]

and Escobar, Luis A

Meeker, William Q. and Escobar, Luis A. , year =. Statistical methods for reliability data , isbn =
[4]

Double/debiased machine learning for treatment and structural parameters.The Econometrics Journal, 21(1):C1–C68, 2018

Double/debiased machine learning for treatment and structural parameters , volume =. The Econometrics Journal , author =. 2018 , pages =. doi:10.1111/ectj.12097 , language =

work page doi:10.1111/ectj.12097 2018
[5]

Mathematics for

Deisenroth, Marc Peter and Faisal, A Aldo and Ong, Cheng Soon , pages =. Mathematics for
[6]

, year =

Bishop, Christopher M. , year =. Pattern recognition and machine learning , isbn =
[7]

All of nonparametric statistics , isbn =

Wasserman, Larry , year =. All of nonparametric statistics , isbn =
[8]

Understanding

Shalev-Shwartz, Shai and Ben-David, Shai , pages =. Understanding
[9]

Predicting good probabilities with supervised learning , isbn =

Niculescu-Mizil, Alexandru and Caruana, Rich , month = aug, year =. Predicting good probabilities with supervised learning , isbn =. Proceedings of the 22nd international conference on. doi:10.1145/1102351.1102430 , abstract =

work page doi:10.1145/1102351.1102430
[10]

Distribution-free

Qiu, Hongxiang and Dobriban, Edgar and Tchetgen, Eric Tchetgen , month = may, year =. Distribution-free. doi:10.48550/arXiv.2203.06126 , abstract =

work page doi:10.48550/arxiv.2203.06126
[11]

arXiv preprint arXiv:2106.09848 , year=

Park, Sangdon and Dobriban, Edgar and Lee, Insup and Bastani, Osbert , month = mar, year =. doi:10.48550/arXiv.2106.09848 , abstract =

work page doi:10.48550/arxiv.2106.09848
[12]

Test-time

Ma, Wenao and Chen, Cheng and Zheng, Shuang and Qin, Jing and Zhang, Huimao and Dou, Qi , month = jul, year =. Test-time. doi:10.48550/arXiv.2207.00769 , abstract =

work page doi:10.48550/arxiv.2207.00769
[13]

Adaptive

Gibbs, Isaac and Candes, Emmanuel , year =. Adaptive. Advances in
[14]

Yang, Yachong and Kuchibhotla, Arun Kumar and Tchetgen, Eric Tchetgen , month = may, year =. Doubly. doi:10.48550/arXiv.2203.01761 , abstract =

work page doi:10.48550/arxiv.2203.01761
[15]

Proceedings of the

He, Yue and Wang, Zimu and Cui, Peng and Zou, Hao and Zhang, Yafeng and Cui, Qiang and Jiang, Yong , month = apr, year =. Proceedings of the. doi:10.1145/3485447.3511969 , abstract =

work page doi:10.1145/3485447.3511969
[16]

Learning from a

Sahoo, Roshni and Lei, Lihua and Wager, Stefan , month = jan, year =. Learning from a. doi:10.48550/arXiv.2209.01754 , abstract =

work page doi:10.48550/arxiv.2209.01754
[17]

40 puzzles and problems in probability and mathematical statistics , isbn =

Schwarz, Wolfgang , year =. 40 puzzles and problems in probability and mathematical statistics , isbn =
[18]

The Econometrics Journal , author =

The practice of non-parametric estimation by solving inverse problems: the example of transformation models , volume =. The Econometrics Journal , author =. 2010 , note =

2010
[19]

Yu, Xiyu and Liu, Tongliang and Gong, Mingming and Zhang, Kun and Batmanghelich, Kayhan and Tao, Dacheng , month = nov, year =. Label-. Proceedings of the 37th
[20]

A Gentle Introduction to Conformal Prediction and Distribution-Free Uncertainty Quantification

Angelopoulos, Anastasios N. and Bates, Stephen , month = dec, year =. A. doi:10.48550/arXiv.2107.07511 , abstract =

work page internal anchor Pith review doi:10.48550/arxiv.2107.07511
[21]

and Ravikumar, Pradeep , month = feb, year =

Aragam, Bryon and Dan, Chen and Xing, Eric P. and Ravikumar, Pradeep , month = feb, year =. Identifiability of
[22]

Journal of Machine Learning Research , volume=

Semi-supervised novelty detection , author=. Journal of Machine Learning Research , volume=
[23]

Detecting and

Lipton, Zachary and Wang, Yu-Xiang and Smola, Alexander , month = jul, year =. Detecting and. Proceedings of the 35th
[24]

doi:10.1198/jcgs.2009.07175 , file =

An. doi:10.1198/jcgs.2009.07175 , file =

work page doi:10.1198/jcgs.2009.07175 2009
[25]

Neural Computation , author =

Practical identifiability of finite mixtures of multivariate bernoulli distributions , volume =. Neural Computation , author =. 2000 , pmid =. doi:10.1162/089976600300015925 , abstract =

work page doi:10.1162/089976600300015925 2000
[26]

, month = oct, year =

Zhu, Zheqi and Fan, Pingyi and Peng, Chenghui and Letaief, Khaled B. , month = oct, year =
[27]

Communication-

Zecchin, Matteo and Kountouris, Marios and Gesbert, David , month = oct, year =. Communication-. doi:10.48550/arXiv.2205.15614 , abstract =

work page doi:10.48550/arxiv.2205.15614
[28]

The Annals of Statistics , author =

Identifiability of. The Annals of Statistics , author =. 2009 , pages =

2009
[29]

Yu, Shuang and Li, Chunping , editor =. Machine. 2007 , keywords =. doi:10.1007/978-3-540-73499-4_43 , abstract =

work page doi:10.1007/978-3-540-73499-4_43 2007
[30]

and Tao, Dacheng , month = mar, year =

He, Fengxiang and Liu, Tongliang and Webb, Geoffrey I. and Tao, Dacheng , month = mar, year =. Instance-
[31]

Estimating the class prior and posterior from noisy positives and unlabeled data , url =

Jain, Shantanu and White, Martha and Radivojac, Predrag , month = jan, year =. Estimating the class prior and posterior from noisy positives and unlabeled data , url =
[32]

Alexandari, Amr and Kundaje, Anshul and Shrikumar, Avanti , month = nov, year =. Maximum. Proceedings of the 37th
[33]

Journal of the Royal Statistical Society

Prediction and outlier detection in classification problems , volume =. Journal of the Royal Statistical Society. Series B, Statistical Methodology , author =. 2022 , pmid =. doi:10.1111/rssb.12443 , abstract =

work page doi:10.1111/rssb.12443 2022
[34]

A general theory of identification , url =

Basse, Guillaume and Bojinov, Iavor , month = feb, year =. A general theory of identification , url =
[35]

Handbook of

Cure. Handbook of. 2013 , note =

2013
[36]

Annual Review of Statistics and Its Application , author =

Cure. Annual Review of Statistics and Its Application , author =. 2018 , note =. doi:10.1146/annurev-statistics-031017-100101 , abstract =

work page doi:10.1146/annurev-statistics-031017-100101 2018
[37]

Recommendations as

Schnabel, Tobias and Swaminathan, Adith and Singh, Ashudeep and Chandak, Navin and Joachims, Thorsten , pages =. Recommendations as
[38]

Proceedings of the AAAI Conference on Artificial Intelligence , author =

Modelling. Proceedings of the AAAI Conference on Artificial Intelligence , author =. 2015 , note =. doi:10.1609/aaai.v29i1.9612 , abstract =

work page doi:10.1609/aaai.v29i1.9612 2015
[39]

Neurocomputing , author =

A generalised label noise model for classification in the presence of annotation errors , volume =. Neurocomputing , author =. 2016 , keywords =. doi:10.1016/j.neucom.2015.12.106 , abstract =

work page doi:10.1016/j.neucom.2015.12.106 2016
[40]

Covariate shift adaptation on learning from positive and unlabeled data
[41]

Proceedings of the

Zhang, Xin and Fang, Minghong and Liu, Zhuqing and Yang, Haibo and Liu, Jia and Zhu, Zhengyuan , month = oct, year =. Proceedings of the. doi:10.1145/3492866.3549723 , abstract =

work page doi:10.1145/3492866.3549723
[42]

Modeling delayed feedback in display advertising , isbn =

Chapelle, Olivier , month = aug, year =. Modeling delayed feedback in display advertising , isbn =. Proceedings of the 20th. doi:10.1145/2623330.2623634 , abstract =

work page doi:10.1145/2623330.2623634
[43]

Yasui, Shota and Morishita, Gota and Fujita, Komei and Shibata, Masashi , month = apr, year =. A. Proceedings of. doi:10.1145/3366423.3380032 , abstract =

work page doi:10.1145/3366423.3380032
[44]

Garg, Saurabh and Wu, Yifan and Balakrishnan, Sivaraman and Lipton, Zachary , year =. A. Advances in
[45]

Courty, Nicolas and Flamary, Rémi and Tuia, Devis and Rakotomamonjy, Alain , month = jun, year =. Optimal
[46]

Le, Trung and Do, Dat and Nguyen, Tuan and Nguyen, Huy and Bui, Hung and Ho, Nhat and Phung, Dinh , month = mar, year =. On
[47]

Deng, Songgaojun and Wang, Shusen and Rangwala, Huzefa and Wang, Lijing and Ning, Yue , month = oct, year =. Cola-. Proceedings of the 29th. doi:10.1145/3340531.3411975 , abstract =

work page doi:10.1145/3340531.3411975
[48]

Wilson, Tyler and McDonald, Andrew and Galib, Asadullah Hill and Tan, Pang-Ning and Luo, Lifeng , month = aug, year =. Beyond. Proceedings of the 28th. doi:10.1145/3534678.3539464 , abstract =

work page doi:10.1145/3534678.3539464
[49]

Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences , author =

Time. Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences , author =. 2021 , note =. doi:10.1098/rsta.2020.0209 , abstract =

work page doi:10.1098/rsta.2020.0209 2021
[50]

Signal Processing , author =

A review of novelty detection , volume =. Signal Processing , author =. 2014 , pages =. doi:10.1016/j.sigpro.2013.12.026 , abstract =

work page doi:10.1016/j.sigpro.2013.12.026 2014
[51]

Garg, Saurabh and Wu, Yifan and Smola, Alexander J and Balakrishnan, Sivaraman and Lipton, Zachary , journal =. Mixture. 2021 , pages =

2021
[52]

Na, Byeonghu and Kim, Hyemi and Song, Kyungwoo and Joo, Weonyoung and Kim, Yoon-Yeong and Moon, Il-Chul , month = oct, year =. Deep. Proceedings of the 29th. doi:10.1145/3340531.3411971 , abstract =

work page doi:10.1145/3340531.3411971
[53]

IEEE Transactions on Pattern Analysis and Machine Intelligence , author =

Instance-. IEEE Transactions on Pattern Analysis and Machine Intelligence , author =. 2021 , keywords =. doi:10.1109/TPAMI.2021.3061456 , language =

work page doi:10.1109/tpami.2021.3061456 2021
[54]

Joint estimation of posterior probability and propensity score function for positive and unlabelled data , url =

Furmańczyk, Konrad and Mielniczuk, Jan and Rejchel, Wojciech and Teisseyre, Paweł , month = sep, year =. Joint estimation of posterior probability and propensity score function for positive and unlabelled data , url =
[55]

Learning from positive and unlabeled data: a survey

Learning from positive and unlabeled data: a survey , volume =. Machine Learning , author =. 2020 , keywords =. doi:10.1007/s10994-020-05877-5 , abstract =

work page doi:10.1007/s10994-020-05877-5 2020
[56]

Beyond the

Bekker, Jessa and Robberechts, Pieter and Davis, Jesse , editor =. Beyond the. 2020 , note =. doi:10.1007/978-3-030-46147-8_5 , abstract =

work page doi:10.1007/978-3-030-46147-8_5 2020
[57]

Smola, Alex and Gretton, Arthur and Song, Le and Scholkopf, Bernhard , pages =. A
[58]

Gretton, Arthur and Borgwardt, Karsten and Rasch, Malte and Schölkopf, Bernhard and Smola, Alex , year =. A. Advances in
[59]

Efficient error models for fault-tolerant architectures and the Pauli twirling approximation

Demystifying. The American Statistician , author =. 2022 , pages =. doi:10.1080/00031305.2021.2021984 , abstract =

work page Pith review doi:10.1080/00031305.2021.2021984 2022
[60]

Journal of Statistical Planning and Inference , volume = 90, number = 2, pages =

Improving predictive inference under covariate shift by weighting the log-likelihood function , volume =. Journal of Statistical Planning and Inference , author =. 2000 , keywords =. doi:10.1016/S0378-3758(00)00115-4 , abstract =

work page doi:10.1016/s0378-3758(00)00115-4 2000
[61]

, year =

Tsiatis, Anastasios A. , year =. Semiparametric theory and missing data , isbn =
[62]

and MacKay, R

Steiner, Stefan H. and MacKay, R. Jock , year =. Statistical engineering: an algorithm for reducing variation in manufacturing processes , isbn =
[63]

, year =

McLachlan, Geoffrey J. , year =. Discriminant analysis and statistical pattern recognition , isbn =
[64]

Sun, Baochen and Saenko, Kate , editor =. Deep. Computer. 2016 , keywords =. doi:10.1007/978-3-319-49409-8_35 , abstract =

work page doi:10.1007/978-3-319-49409-8_35 2016
[65]

Collaborative and

Zhang, Weichen and Ouyang, Wanli and Li, Wen and Xu, Dong , month = jun, year =. Collaborative and. 2018. doi:10.1109/CVPR.2018.00400 , abstract =

work page doi:10.1109/cvpr.2018.00400 2018
[66]

Machine Learning , author =

Optimal transport for conditional domain matching and label shift , volume =. Machine Learning , author =. 2022 , keywords =. doi:10.1007/s10994-021-06088-2 , abstract =

work page doi:10.1007/s10994-021-06088-2 2022
[67]

Discriminative

Bickel, Steffen and De, Cs Uni-Potsdam and De, Cs Uni-Potsdam and De, Cs Uni-Potsdam , keywords =. Discriminative
[68]

Zhang, Kun and Scholkopf, Bernhard and Muandet, Krikamol and Wang, Zhikun , keywords =. Domain
[69]

, month = jul, year =

Garg, Saurabh and Balakrishnan, Sivaraman and Lipton, Zachary C. , month = jul, year =. Domain. doi:10.48550/arXiv.2207.13048 , abstract =

work page doi:10.48550/arxiv.2207.13048
[70]

Flexible

Wang, Xuezhi and Schneider, Jeff , year =. Flexible. Advances in
[71]

IEEE Transactions on Pattern Analysis and Machine Intelligence , author =

Importance. IEEE Transactions on Pattern Analysis and Machine Intelligence , author =. 2021 , note =. doi:10.1109/TPAMI.2021.3086060 , abstract =

work page doi:10.1109/tpami.2021.3086060 2021
[72]

Minimax optimal approaches to the label shift problem , url =

Maity, Subha and Sun, Yuekai and Banerjee, Moulinath , month = apr, year =. Minimax optimal approaches to the label shift problem , url =. doi:10.48550/arXiv.2003.10443 , abstract =

work page doi:10.48550/arxiv.2003.10443 2003
[73]

Moreno-Torres and Troy Raeder and Roc

A unifying view on dataset shift in classification , volume =. Pattern Recognition , author =. 2012 , keywords =. doi:10.1016/j.patcog.2011.06.019 , abstract =

work page doi:10.1016/j.patcog.2011.06.019 2012
[74]

Covariate

Gretton, Arthur and Smola, Alex and Huang, Jiayuan and Schmittfull, Marcel and Borgwardt, Karsten and Schölkopf, Bernhard , editor =. Covariate. Dataset. 2008 , doi =

2008
[75]

Bridging

Zhang, Yuchen and Liu, Tianle and Long, Mingsheng and Jordan, Michael , month = may, year =. Bridging. Proceedings of the 36th
[76]

Covariate

Sugiyama, Masashi and Krauledat, Matthias , keywords =. Covariate
[77]

IEEE Transactions on Pattern Analysis and Machine Intelligence , author =

A review of domain adaptation without target labels , volume =. IEEE Transactions on Pattern Analysis and Machine Intelligence , author =. 2021 , note =. doi:10.1109/TPAMI.2019.2945942 , abstract =

work page doi:10.1109/tpami.2019.2945942 2021
[78]

Rabanser, Stephan and Günnemann, Stephan and Lipton, Zachary , year =. Failing. Advances in
[79]

Advances in

Sohn, Kihyuk and Berthelot, David and Carlini, Nicholas and Zhang, Zizhao and Zhang, Han and Raffel, Colin A and Cubuk, Ekin Dogus and Kurakin, Alexey and Li, Chun-Liang , year =. Advances in
[80]

Domain Adaptation in Computer Vision Applications , author =

Domain-. Domain Adaptation in Computer Vision Applications , author =. 2017 , pages =. doi:10.1007/978-3-319-58347-1_10 , abstract =

work page doi:10.1007/978-3-319-58347-1_10 2017

Showing first 80 references.