PaperlayerView exact revision on arXiv

Source-linked AI summary

A review of domain adaptation without target labels

Wouter M. Kouw, Marco Loog

arXiv:1901.05335v2cs.LGstat.ML

TL;DR

Domain adaptation asks how classifiers trained on a source domain can generalize to an unlabeled target domain despite distributional differences. This review categorizes methods by whether they operate on samples, feature representations, or parameter estimation, and synthesizes assumptions that support cross-domain performance bounds. It concludes that assumptions strongly influence when methods succeed or fail and that hypothesis tests, interpretability, and causal structure merit further study.

  • Problem

    Classifiers trained on collected source data may face biased samples and must generalize to a target domain where labels are difficult to obtain.

  • Method

    The review organizes domain adaptation into sample-based, feature-based, and inference-based methods and examines their theoretical assumptions and interpretability.

  • Results

    The categorization shows that assumptions strongly influence when domain-adaptation methods succeed or fail.

  • Takeaways & Limitations

    Hypothesis tests for assumption validity, interpretable transfer procedures, and causal-structure discovery are identified as important research directions.

  • Takeaways & Limitations

    The review is limited to single-source, single-target adaptation without labeled target data and is not exhaustive.

Abstract

from arXiv · show

Domain adaptation has become a prominent problem setting in machine learning and related fields. This review asks the question: how can a classifier learn from a source domain and generalize to a target domain? We present a categorization of approaches, divided into, what we refer to as, sample-based, feature-based and inference-based methods. Sample-based methods focus on weighting individual observations during training based on their importance to the target domain. Feature-based methods revolve around on mapping, projecting and representing features such that a source classifier performs well on the target domain and inference-based methods incorporate adaptation into the parameter estimation procedure, for instance through constraints on the optimization procedure. Additionally, we review a number of conditions that allow for formulating bounds on the cross-domain generalization error. Our categorization highlights recurring ideas and raises questions important to further research.

1 INTRODUCTION

Domain adaptation addresses how classifiers trained on biased source data can generalize to an unlabeled target domain. The review organizes methods into sample-based, feature-based, and inference-based approaches, while identifying assumptions relevant to performance guarantees.

  • Motivation: Collected data often differs systematically from the distribution in which a classifier must operate, limiting generalization to new samples.The review frames this mismatch as a practical consequence of biased sampling and differing populations or acquisition settings.
  • Motivation: Unlabeled target data indicates how source and target domains differ and can guide changes to classifier decisions.Target labels are harder to obtain, but unlabeled observations provide information for adaptation.
  • Method categorization: Sample-based methods correct sampling bias through data or class importance-weighting, whereas feature-based methods reshape representations for cross-domain prediction.Feature-based approaches include subspace mappings, optimal transport, domain-invariant representations, and corresponding features.
  • Method categorization: Inference-based methods incorporate adaptation into parameter estimation through approaches including robustness, minimax estimation, self-learning, empirical Bayes, and PAC-Bayes.The review treats inference-based adaptation as a distinct category from operations on samples or feature representations.
  • Theory and research directions: The review identifies conditions that permit performance guarantees but emphasizes that every adaptive classifier has settings in which it fails.It also highlights hypothesis testing and causal information as relevant to selecting domain-adaptive classifiers.
  • Scope: The survey focuses on single-source, single-target adaptation without target labels and is not exhaustive in its coverage of prior work.Multi-source adaptation, target labels, and topics covered in other reviews remain outside its scope.

2 DOMAIN ADAPTATION

Domain adaptation uses labeled source samples and unlabeled target samples drawn from different distributions, with domain dissimilarity helping characterize cross-domain generalization. Generalization bounds depend on classifier complexity, sample size, ideal joint performance, and assumptions about the domains.

  • Definitions: A domain is a probability distribution over a shared feature-label space; the source has labeled data, while the target supplies unlabeled samples for prediction.The source and target distributions are denoted pS(x, y) and pT(x, y), respectively.
  • Example setting: The running heart-disease example compares age and cholesterol measurements from hospitals in Budapest and Long Beach.Budapest is the source domain and Long Beach is the target domain.
  • Example setting: A source-trained linear classifier may perform poorly on shifted target data, and increasing domain differences can further reduce performance.The example illustrates a decision boundary trained on source samples and applied to target samples with an age shift.
  • Domain dissimilarity metrics: Domain dissimilarity can be measured using distributional metrics such as KL divergence, total variation, Wasserstein distance, or the Kolmogorov-Smirnov statistic.The choice of metric can affect the behavior of a domain-adaptive classifier.
  • Domain dissimilarity metrics: The symmetric difference hypothesis divergence measures the largest difference in classifier disagreement between source and target domains, increasing as the domains become more dissimilar.The R'enyi divergence is another example and equals KL divergence when its order α is 1.
  • Generalization error: A cross-domain error bound depends on ideal joint hypothesis error, domain dissimilarity, classifier complexity, and sample size.The stated bound applies with probability 1 − δ and uses the VC-dimension of the hypothesis space.
  • Generalization error: Tighter bounds for adaptive classifiers require assumptions such as equal source and target posterior distributions, and their implications depend on the form of domain dissimilarity.The review presents these assumptions as providing theoretical understanding of when adaptation can generalize.

3 SAMPLE-BASED APPROACHES

Sample-based approaches adapt a source classifier by reweighting source observations to better represent the target domain, often under assumptions about domain shift. The review covers weight estimation methods, illustrative effects, and theoretical and computational limitations.

  • Data importance-weighting: Without target labels, estimating joint-density ratios is unavailable directly, so sample-based methods rely on simplifying assumptions such as covariate or prior shift.Concept shift is outside the review’s scope because it requires labeled data in both domains.
  • Data importance-weighting: Sample-based methods minimize target risk using source data by assigning observations weights related to their importance under the target distribution.Under covariate shift, the relevant weight is w(x) = pT(x)/pS(x), so target-relevant samples receive greater influence during training.
  • Data importance-weighting: Importance-weighted training can change the decision boundary and predictions toward the target domain by increasing loss for more target-relevant source samples.The example shows weighted source samples producing an adapted classifier with a decision boundary different from the original source classifier.
  • Data importance-weighting: Importance weighting requires finite expected squared weights for convergence, and larger domain divergence demands more source samples to maintain the same error difference.The review also notes that weighting is asymptotically effective for mis-specified models, whereas correctly specified models can converge optimally without weighting.
  • Data importance-weighting: Kernel Mean Matching estimates weights by minimizing a kernel-based discrepancy between weighted source and target distributions, subject to nonnegativity and near-unit-average constraints.MMD compares distributions through differences in expectations under functions, while KMM formulates the empirical optimization as a quadratic program.
  • Data importance-weighting: Weight estimation has practical limitations: some procedures require re-estimation when source samples are added, and KMM’s quadratic program is poorly suited to large datasets.Additional constraints can target effects such as low weight variance, but high variance can cause training to focus on only a few samples.
  • Data importance-weighting: Least-Squares Importance Fitting can use d variables instead of KMM’s n sample-specific variables when d < n, making its quadratic program computationally cheaper.KMM in the described form does not scale well because its quadratic program has one variable per source sample.

4 FEATURE-BASED APPROACHES

Feature-based approaches reshape, align, or represent source and target features so that classifiers trained with source labels can generalize to the target domain. The review covers deterministic and stochastic mappings, subspace and manifold methods, deep representations, and domain-invariant spaces, while noting that matching marginal data distributions need not match class-conditional distributions.

  • 4.1 Subspace mappings: Feature-based methods reshape feature space so classifiers trained on transformed source data can be applied to target samples.Mappings may translate source samples, align subspaces, model transformation paths, or learn domain-invariant representations.
  • 4.3 Domain-invariant spaces: Matching source and target marginal distributions alone does not imply matching their conditional label distributions, limiting the justification for distribution-alignment methods.The review also notes that the bound involving conditional invariant components becomes looser as θ approaches π and cannot use Jci as an upper bound at θ = π.
  • 4.1 Subspace mappings: Subspace Alignment projects each domain onto its first d principal components, aligns source and target components with a transformation matrix, and trains on mapped source data.Extensions use landmarks, distributions, or joint classifier-and-subspace learning.
  • 4.1 Subspace mappings: Manifold-based methods model intermediate domains between source and target, with Geodesic Flow Kernel integrating similarities across intermediate subspaces.The manifold assumption represents transformations as paths through a parameterized space of possible domains.
  • 4.1 Subspace mappings: Stochastic mappings adapt to feature uncertainty by training under a transfer model, reducing reliance on source features likely to be absent in the target domain.Feature-level domain adaptation estimates feature-dropout probabilities from source and target data.
  • 4.3 Domain-invariant spaces: Domain-invariant approaches seek representations whose source and target distributions are similar, using methods such as MMD-based projections, learned kernels, and conditional invariant components.Conditional invariant components require transformed class-conditional distributions to match; under the stated causal conditions, perfect transformed-distribution matching yields zero difference between transformed-source and target error.

5 INFERENCE-BASED APPROACHES

Inference-based methods adapt classifiers within the inference or parameter-estimation procedure, using optimization reformulations, target-based constraints, uncertainty, or Bayesian priors. The reviewed approaches include robust and minimax estimators, empirical Bayes, and PAC-Bayesian methods.

  • Inference-based methods: Inference-based methods incorporate adaptation into inference through objective reformulations, target-domain constraints, or Bayesian uncertainty.This category is more diverse than sample- and feature-based approaches because it includes many forms of adaptation.
  • Robust methods: λ-shift models bounded changes between source and target posterior probabilities, ranging from identical posteriors at λ = 0 to arbitrary differences at λ = 1.The condition is defined region by region and class by class.
  • Robust methods: λ-SVMA constrains source-sample weights within each feature-space region using the empirical target probability, preventing concentration where target samples are scarce.Its pessimistic variant sets λ = 1, imposing no posterior-difference restriction.
  • Minimax estimators: Minimax estimators optimize classifier risk against adversarial uncertainty, producing more conservative adaptation under maximal uncertainty.The Robust Bias-Aware classifier constrains adversarial posteriors to match source feature-statistic moments.
  • Minimax estimators: The Target Contrastive Robust risk estimator compares target risk with the source classifier and avoids adaptation when no guaranteed improvement exists, thereby avoiding negative transfer.For discriminant analysis, its empirical target risk is strictly smaller than the source classifier’s for the given target samples, but it can perform poorly when the source classifier is a bad target choice.
  • Empirical Bayes and PAC-Bayes: Empirical Bayes uses source data to construct an informative prior over target-model parameters, while PAC-Bayesian bounds can account for target regions where the source is uninformative.An informative prior can be harmful when source and target domains differ substantially.

6 DISCUSSION

The discussion emphasizes that domain adaptation depends on unverifiable domain assumptions, motivating hypothesis tests, causal information, and interpretable methods. It also identifies practical failure modes and scope limits, including support mismatch, high-dimensional weighting, and non-exhaustive coverage.

  • Assumptions, tests, and no-free-lunch: Without target labels, assumptions such as covariate shift or prior shift cannot be verified, so performance on a given dataset cannot be predicted reliably.The review states that every adaptive classifier has settings in which learning fails dramatically.
  • Assumptions, tests, and no-free-lunch: Hypothesis tests and causal information are discussed as tools for selecting domain-adaptation classifiers and assessing whether assumptions are plausible.The review presents such tests as important for practical solutions because invalid assumptions create a risk of negative transfer.
  • Insights and interpretability: Interpretability helps inspect why domain-adaptation methods are likely to succeed or fail, generating insights that deepen understanding.The review links explicit transfer descriptions to comparisons of typical success and failure cases.
  • Shrinking the search space: Source-informed parameter restrictions can interfere with target learning when parameters useful to the target are ignored.This limitation applies to the interpretation of adaptation as shrinking the parameter search space.
  • General limitations: Importance-weighting can fail in high dimensions, while many methods require high-probability target support to be contained within source support.When target mass lies outside source support, pathological solutions can arise.
  • Scope: The survey is not exhaustive because it prioritizes papers presenting original ideas and may miss work using inconsistent terminology.The authors nevertheless describe it as a thorough summary.

7 CONCLUSION

The review organizes domain adaptation methods into sample-based, feature-based, and inference-based categories, and concludes that assumptions strongly shape when methods succeed or fail. It also identifies hypothesis testing, interpretability, and causal information as important directions for selecting and developing adaptive classifiers.

  • 7 CONCLUSION: Domain adaptation methods operate on individual observations, representations of observation sets, or the parameter estimator.These are the review’s three coarse categories: sample-based, feature-based, and inference-based methods.
  • 7 CONCLUSION: Sample-based methods divide into data importance-weighting under covariate shift and class importance-weighting under prior shift.The review also notes that diverse weight estimators address different problem settings and data types.
  • 7 CONCLUSION: Feature-based methods include subspace mappings, optimal transport, domain-invariant spaces, deep domain adaptation, and correspondence learning.These approaches reshape or transform source and target representations or distributions to support adaptation.
  • 7 CONCLUSION: Assumptions are necessary in domain adaptation without target labels and strongly influence when a method succeeds or fails.The review therefore emphasizes developing hypothesis tests for assumption validity.
  • 7 CONCLUSION: Future work should study interpretable transfer procedures and causal structure to improve adaptive classifier selection.The review connects explicit comparisons of success and failure cases with possible insights into transfer, while identifying causal discovery as a way to resolve ambiguities.

http://arxiv.org/ps/1901.05335v2

The passage indicates that the figure “biophoto_wouter.jpg” is available in JPG format.

  • The figure “biophoto_wouter.jpg” is available in JPG format.
Loading 1901.05335v2…