A Kernel Method for the Two-Sample Problem

Abstract

We propose a framework for analyzing and comparing distributions, allowing us to design statistical tests to determine if two samples are drawn from different distributions. Our test statistic is the largest difference in expectations over functions in the unit ball of a reproducing kernel Hilbert space (RKHS). We present two tests based on large deviation bounds for the test statistic, while a third is based on the asymptotic distribution of this statistic. The test statistic can be computed in quadratic time, although efficient linear time approximations are available. Several classical metrics on distributions are recovered when the function space used to compute the difference in expectations is allowed to be more general (eg. a Banach space). We apply our two-sample tests to a variety of problems, including attribute matching for databases using the Hungarian marriage method, where they perform strongly. Excellent performance is also obtained when comparing distributions over graphs, for which these are the first such tests.

Cited by in corpus (77)

• Equivalence of distance-based and RKHS-based statistics in hypothesis testing
• Texture Networks: Feed-forward Synthesis of Textures and Stylized Images
• InfoVAE: Information Maximizing Variational Autoencoders
• Real-valued (Medical) Time Series Generation with Recurrent Conditional GANs
• Hilbert space embeddings and metrics on probability measures
• Learning Adversarially Fair and Transferable Representations
• How to GAN LHC Events
• The Born Supremacy: Quantum Advantage and Training of an Ising Born Machine
• Getting High: High Fidelity Simulation of High Granularity Calorimeters with High Speed
• Do We Really Need to Access the Source Data? Source Hypothesis Transfer for Unsupervised Domain Adaptation
• Universality, Characteristic Kernels and RKHS Embedding of Measures
• Simulation of electron-proton scattering events by a Feature-Augmented and Transformed Generative Adversarial Network (FAT-GAN)
• Decoding Photons: Physics in the Latent Space of a BIB-AE Generative Network
• HHHFL: Hierarchical Heterogeneous Horizontal Federated Learning for Electroencephalography
• Kernel methods on spike train space for neuroscience: a tutorial
• Implicit Maximum Likelihood Estimation
• Deep Cocktail Network: Multi-source Unsupervised Domain Adaptation with Category Shift
• On the optimal estimation of probability measures in weak and strong topologies
• Learning to Adapt Invariance in Memory for Person Re-identification
• Kernel Mean Shrinkage Estimators
• Learning Attributes Equals Multi-Source Domain Generalization
• The Information Autoencoding Family: A Lagrangian Perspective on Latent Variable Generative Models
• Measurement of the B-band Galaxy Luminosity Function with Approximate Bayesian Computation
• Kernel Distribution Embeddings: Universal Kernels, Characteristic Kernels and Kernel Metrics on Distributions
• Multi-task Mid-level Feature Alignment Network for Unsupervised Cross-Dataset Person Re-Identification
• Semi-Supervised Domain Adaptation with Non-Parametric Copulas
• Hilbert Space Embeddings of POMDPs
• DeepMatch: Balancing Deep Covariate Representations for Causal Inference Using Adversarial Training
• 3D Shape Induction from 2D Views of Multiple Objects
• Multivariate tests of association based on univariate tests
• Robust Data Geometric Structure Aligned Close yet Discriminative Domain Adaptation
• Adversarial training approach for local data debiasing
• On the Optimality of Kernel-Embedding Based Goodness-of-Fit Tests
• Domain Alignment with Triplets
• Exploiting Coarse-to-Fine Task Transfer for Aspect-level Sentiment Classification
• Label-Free Model Evaluation with Semi-Structured Dataset Representations
• Sketching for Large-Scale Learning of Mixture Models
• Domain Adaptation with Auxiliary Target Domain-Oriented Classifier
• Recent Advances of Image Steganography with Generative Adversarial Networks
• Toward a better trade-off between performance and fairness with kernel-based distribution matching
• The Three Hundred Project: Mapping The Matter Distribution in Galaxy Clusters Via Deep Learning from Multiview Simulated Observations
• Characteristic Kernels and Infinitely Divisible Distributions
• Quantile-Quantile Embedding for Distribution Transformation and Manifold Embedding with Ability to Choose the Embedding Distribution
• Minimax Estimation of Kernel Mean Embeddings
• Wasserstein Measure Coresets
• Distance Measure Machines
• Improved Texture Networks: Maximizing Quality and Diversity in Feed-forward Stylization and Texture Synthesis
• Heterogeneous Domain Adaptation via Soft Transfer Network
• Mixability of Integral Losses: a Key to Efficient Online Aggregation of Functional and Probabilistic Forecasts
• Kernel Mean Estimation and Stein's Effect
• Domain Conditioned Adaptation Network
• TLR: Transfer Latent Representation for Unsupervised Domain Adaptation
• Algebraic Geometric Comparison of Probability Distributions
• Attention-Set based Metric Learning for Video Face Recognition
• The Perturbed Variation
• DISCO Nets: DISsimilarity COefficient Networks
• Towards Robust Relational Causal Discovery
• A New Measure of Conditional Dependence
• Exemplar-based synthesis of geology using kernel discrepancies and generative neural networks
• A Geometric Algorithm for Scalable Multiple Kernel Learning
• Metric Learning across Heterogeneous Domains by Respectively Aligning Both Priors and Posteriors
• Towards Robust Pattern Recognition: A Review
• Inferring 3D Shapes from Image Collections using Adversarial Networks
• Kernel Two-Sample Hypothesis Testing Using Kernel Set Classification
• A Mean-Field Theory for Kernel Alignment with Random Features in Generative and Discriminative Models
• Post Selection Inference with Incomplete Maximum Mean Discrepancy Estimator
• Easy High-Dimensional Likelihood-Free Inference
• Strictly Proper Kernel Scoring Rules and Divergences with an Application to Kernel Two-Sample Hypothesis Testing
• Deep Learning generated observations of galaxy clusters from dark-matter-only simulations
• Learning Where to Attend Like a Human Driver
• Minimax Nonparametric Two-sample Test under Smoothing
• Level Sets Based Distances for Probability Measures and Ensembles with Applications
• A Parzen-based distance between probability measures as an alternative of summary statistics in Approximate Bayesian Computation
• Learning Shared Dynamics with Meta-World Models
• Asymptotics and practical aspects of testing normality with kernel methods
• Aggregated kernel based tests for signal detection in a regression model
• Domain Adaptation for Real-World Single View 3D Reconstruction