Weighting parameters

Pass pre-computed or context-aware weights directly to test_shift or test_adverse_shift to correct for known covariate shift between your source and target groups.

Choosing an approach

Scenario	How to proceed
No weighting (default)	Omit weights
You have explicit per-sample weights	Wrap in ContextualWeights(source=..., target=...), then pass weights=
You have domain probabilities from a domain classifier	Build weights with contextual_weights(...), then pass weights=

import numpy as np
import samesame
from samesame.weights import contextual_weights

# No weighting (default)
result = samesame.test_shift(source=source_scores, target=target_scores)

# Explicit per-sample weights you computed yourself
result = samesame.test_shift(
    source=source_scores,
    target=target_scores,
    weights=samesame.ContextualWeights(source=source_weights, target=target_weights),
)

# Context-aware weights derived from domain probabilities
weights = contextual_weights(
    source_prob=source_domain_probs,  # domain probabilities for source samples
    target_prob=target_domain_probs,  # domain probabilities for target samples
    mode="source",
)
result = samesame.test_adverse_shift(
    source=source_scores,
    target=target_scores,
    direction="higher-is-worse",
    weights=weights,
)

See Sample weights for the full contextual_weights reference and guidance on choosing mode and lambda_.

For a step-by-step worked example, see the tutorial Adjust for covariate shift with importance weights. For the conceptual background on why density ratios need stabilisation and when to choose each mode, see Why importance weights stabilise shift detection.