research: Add unseen-entity generalization experiment - #5
Merged
Conversation
Network features (out-degree, counterparty HHI, lifetime counts) drive most of the ablation lift (PR-AUC 0.09 -> 0.94), which raises an obvious question: does that hold for accounts the model never saw during training, or is it propped up by account history that only exists for already-observed entities? scripts/unseen_entity_evaluation.py reuses the existing artifact (no retraining) and partitions the held-out test set by whether both sender and receiver appeared in the training window: - Standard out-of-time (full test set): PR-AUC 0.9859, lift 198.6x - Both parties seen during training: PR-AUC 0.9918, lift 198.6x - At least one party unseen: PR-AUC 0.7912, lift 198.4x PR-AUC and precision drop sharply for unseen entities, but ROC-AUC (0.9998), recall @ 0.5% (~99.2%), and lift over base rate (~198x) are essentially unchanged across all three partitions. The PR-AUC drop tracks an 8x lower positive prevalence in the unseen-entity subset, not a loss of ranking ability -- PR-AUC's baseline is the prevalence itself, so a rarer-positive subset produces a lower score even under identical discrimination. Full writeup: docs/research/unseen_entity_generalization.md Also notes what this does NOT test: a connected-component/network holdout (colluding clusters never seen together) is a stricter and still-open question, left as further work. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Network features (out-degree, counterparty HHI, lifetime counts) drive most of the ablation lift (PR-AUC 0.09 -> 0.94), which raises an obvious question: does that hold for accounts the model never saw during training, or is it propped up by account history that only exists for already-observed entities?
scripts/unseen_entity_evaluation.py reuses the existing artifact (no retraining) and partitions the held-out test set by whether both sender and receiver appeared in the training window:
PR-AUC and precision drop sharply for unseen entities, but ROC-AUC (0.9998), recall @ 0.5% (~99.2%), and lift over base rate (~198x) are essentially unchanged across all three partitions. The PR-AUC drop tracks an 8x lower positive prevalence in the unseen-entity subset, not a loss of ranking ability -- PR-AUC's baseline is the prevalence itself, so a rarer-positive subset produces a lower score even under identical discrimination. Full writeup: docs/research/unseen_entity_generalization.md
Also notes what this does NOT test: a connected-component/network holdout (colluding clusters never seen together) is a stricter and still-open question, left as further work.