Maximum Mean Discrepancy (MMD)

MMD operates by embedding probability distributions into a reproducing kernel Hilbert space (RKHS). Each distribution is represented by its mean embedding—a single point in the feature space. The MMD statistic is simply the distance between these two mean embeddings.

• Avoids explicit density estimation, which is intractable in high dimensions
• The kernel choice determines which distributional features are compared
• Characteristic kernels (e.g., Gaussian RBF) ensure the embedding is injective: MMD = 0 if and only if the distributions are identical

The MMD can be expressed as the difference in expectations of a witness function evaluated on samples from each distribution. This function reveals where the distributions differ most.

• The witness function is the function in the RKHS unit ball that maximizes the mean discrepancy
• Peaks indicate regions where one distribution has higher density than the other
• In fraud detection, this pinpoints specific transaction features driving distributional shift, enabling targeted investigation

MMD provides a nonparametric statistical test for the null hypothesis that two samples are drawn from the same distribution. The test statistic converges to zero under the null and to the population MMD under the alternative.

• Unbiased estimator: Computed as a U-statistic over pairs of observations
• Quadratic-time complexity O(n²) in the number of samples, with linear-time approximations available
• Permutation-based or asymptotic null distributions yield p-values for drift detection thresholds

The choice of kernel function and its bandwidth parameter critically determines which scales of distributional difference MMD detects. A Gaussian kernel with bandwidth σ compares distributions smoothed at scale σ.

• Small bandwidths: Sensitive to fine-grained, local differences; higher variance
• Large bandwidths: Capture global distributional structure; may miss localized drift
• Median heuristic: Setting σ to the median pairwise distance in the pooled sample provides a robust default
• Multiple kernel MMD: Combining kernels at different scales yields a test sensitive to discrepancies across all scales simultaneously

In production fraud systems, MMD is applied to compare the distribution of model inputs or latent representations between a reference window and the current production window.

• Feature-level drift: Compute MMD on individual input features to isolate which variables have shifted
• Representation-level drift: Apply MMD to the embeddings from a model's penultimate layer to detect semantic drift invisible at the input level
• Prediction distribution drift: Compare output score distributions to detect concept drift without waiting for ground truth labels
• Triggers automated retraining pipelines when the MMD statistic exceeds a calibrated threshold

MMD occupies a distinct position among distributional distance measures, offering advantages over alternatives in specific contexts.

• vs. KL Divergence: KL requires density estimates and is asymmetric; MMD is symmetric and density-free
• vs. Wasserstein Distance: Both are metric distances, but MMD has a closed-form estimator and is computationally simpler for two-sample testing
• vs. Kolmogorov-Smirnov: KS operates on univariate cumulative distributions; MMD naturally handles multivariate data
• Integral Probability Metric: MMD belongs to the IPM family, defined as the supremum over a function class, providing a unified theoretical framework

An asymmetric measure of how one probability distribution diverges from a reference distribution. Unlike MMD, KL divergence requires explicit density estimation and is not defined when distributions have non-overlapping supports. In fraud detection, it is often used to compare feature distributions between training and production windows, but its sensitivity to binning choices makes it less robust than kernel-based methods for high-dimensional transaction data.

Also known as the Earth Mover's Distance, this metric computes the minimum cost to transform one distribution into another. It provides a geometrically meaningful distance even when distributions have disjoint supports—a critical advantage over KL divergence. In financial anomaly detection, Wasserstein distance is particularly effective for monitoring continuous transaction features like dollar amounts, where subtle shifts in spending patterns must be detected before they degrade model recall.

A nonparametric statistical test that measures the maximum absolute difference between two empirical cumulative distribution functions. The KS test is widely used in model monitoring dashboards for its interpretability and computational simplicity. However, it is most sensitive to shifts near the median and loses power in the tails—precisely where fraud scores concentrate. MMD offers superior tail sensitivity through its kernel embedding, making it a stronger choice for extreme-value monitoring.

A symmetric metric that bins a continuous variable and sums the logarithmic differences between expected and actual bin proportions. PSI is an industry-standard in credit risk and fraud model governance due to its regulatory acceptance. Key limitations include sensitivity to binning strategy and inability to capture multivariate distributional shifts. MMD complements PSI by operating directly on raw feature vectors without discretization, detecting complex joint-distribution drift that bin-based methods miss.

A practical technique where a binary classifier is trained to distinguish between training and production samples. If the classifier achieves high accuracy, significant distributional shift exists. This method is model-agnostic and automatically identifies the most drifted features through feature importance analysis. When combined with MMD as a formal statistical test, adversarial validation provides both a quantitative drift magnitude and actionable diagnostic insights for triggered retraining decisions.

The systematic task of identifying inference-time samples that fall outside the support of the training distribution. OOD detection is critical in fraud systems where novel attack patterns appear that the model has never encountered. MMD serves as a powerful OOD scoring function by measuring the distance between incoming transaction embeddings and the training manifold, enabling the system to abstain from prediction and escalate to human investigators rather than producing unreliable risk scores.