← Back to the benchmark

NeoLDM · 03 — cortex embeddings

cortex's per-transaction hidden state, fed to the same XGBoost, on the full test:

Embeddings: luizcoroo/cortex-ibm-tabformer-embeddings (finetune split).

import json
m=json.load(open('results/fulltest_finetune.json')); t=m['test']
print(f"full test ({t['rows']:,} rows, {t['fraud']:,} fraud):")
labels={'raw_13d':'raw','finetune_hidden':'cortex','raw+finetune_hidden':'cortex + raw'}
for k,v in m['arms'].items():
    print(f"  {labels.get(k,k):22s} AUPRC {v['auprc_mean']:.3f} ± {v['auprc_std']:.3f}  ({v['auprc_mean']/t['rate']:.0f}x random)   F1 {v['f1_mean']:.3f}")
raw=m['arms']['raw_13d']['auprc_mean']; comb=m['arms']['raw+finetune_hidden']['auprc_mean']
print(f"\n  raw {raw:.3f} -> cortex + raw {comb:.3f}  ({comb/raw:.1f}x)")

full test (2,412,326 rows, 2,698 fraud):
  raw                    AUPRC 0.176 ± 0.034  (157x random)   F1 0.301
  cortex                 AUPRC 0.896 ± 0.034  (801x random)   F1 0.870
  cortex + raw           AUPRC 0.955 ± 0.007  (854x random)   F1 0.906

  raw 0.176 -> cortex + raw 0.955  (5.4x)

Note

• AUPRC 0.955 — 854× random, 5.4× the raw baseline.
• cortex alone reaches 0.896 (801× random); adding the raw features on top lifts it to 0.955.