reweight

Reweight is the act of assigning weights to observations or samples to reflect a target distribution or objective, often by multiplying by a ratio or adjusting loss contributions.

In statistics and computational methods, reweighting is central to importance sampling, where samples drawn from a proposal distribution q(x) are used to estimate expectations under a target distribution p(x). The weight for each sample is w(x) = p(x)/q(x); estimators use normalized weights to reduce variance.

In data analysis and causal inference, inverse probability weighting uses weights based on the probability of treatment assignment or censoring to correct for confounding or missing data.

In machine learning, reweighting appears as class weights in loss functions to address class imbalance or misclassification costs, and as sample weights in training that scale individual examples. In boosting algorithms, iterative reweighting increases the weights of misclassified instances to focus learning on harder cases.

In physics and chemistry, reweighting adjusts Monte Carlo samples when model parameters or experimental conditions change. Boltzmann reweighting, for example, modifies ensemble averages as a function of energy differences without generating new samples.

Limitations include high variance when the target distribution differs substantially from the sampling distribution, numerical instability with extreme weights, and the need for careful normalization or regularization.

See also: importance sampling, weighted averages, propensity score weighting, inverse probability weighting.