Skip to content

Make variable encodings and imputation regressor inputs invariant (v1.2) - #3

Merged
fabiwoller merged 1 commit into
mainfrom
invariant-variable-encodings
Nov 9, 2026
Merged

fabiwoller merged 1 commit into
mainfrom
invariant-variable-encodings

Conversation

@fabiwoller

Copy link
Copy Markdown
Contributor

Value-node inputs normalize(e_j + e_v) leave the raw variable encodings e_j non-identifiable (categorical shifts/rescalings, shared numeric shift, global scale). get_embeddings() now returns per-variable means of the normalized value-node inputs and centred, scale-normalized bin encodings; raw=True keeps the old behaviour. Adds get_value_node_embeddings, get_numeric_concept_embeddings and get_variable_centroid_norms.

The numeric regression head now uses the normalized sample input and the numeric concept encoding by default (regressor_inputs="invariant"); regressor_inputs="raw" reproduces POME <= 1.1. Tests cover invariance to the encoding symmetries.

Value-node inputs normalize(e_j + e_v) leave the raw variable encodings
e_j non-identifiable (categorical shifts/rescalings, shared numeric shift,
global scale). get_embeddings() now returns per-variable means of the
normalized value-node inputs and centred, scale-normalized bin encodings;
raw=True keeps the old behaviour. Adds get_value_node_embeddings,
get_numeric_concept_embeddings and get_variable_centroid_norms.

The numeric regression head now uses the normalized sample input and the
numeric concept encoding by default (regressor_inputs="invariant");
regressor_inputs="raw" reproduces POME <= 1.1. Tests cover invariance to
the encoding symmetries.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@fabiwoller
fabiwoller merged commit 6b375ba into main Nov 9, 2026
1 check passed
@fabiwoller
fabiwoller deleted the invariant-variable-encodings branch November 18, 2026 09:11
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant