ArwindhC N

Match each of 1.7M businesses to every record of the same business among 10M records from two other sources. Names arrive with typos, abbreviations, reordered addresses and Indian scripts, and one test country (France) never appears in training. The metric is macro F0.5, so a false match costs twice as much as a missed one.

What moved the score

• Rarity-based blocking over a packed index. Keys come from each record's rarest words. A 40-bit key hash and a 24-bit record ID share one sorted uint64 array per country, so every lookup is a binary search.
• Bidirectional blocking with a hub cap. Searching backwards from every S2/S3 record lifted candidate recall from 0.953 to 0.971. The cap stopped one business from collecting 142,645 candidates through the word "com".
• A fine-tuned transformer as one feature. A 22.7M-parameter MiniLM cross-encoder, fine-tuned in folds split by entity so it never scores entities it trained on. Alone it scored 0.961; as a LightGBM feature it added +0.005.
• Learned transliteration. A 1,305-word Indian-script to Latin dictionary learned only from training pairs. Median name similarity on true pairs went from 0.10 to 1.00.
• An expected-F0.5 decision layer. Each record goes to at most one business, then each business keeps the candidate subset that maximises expected F0.5, including predicting nothing.

I logged what failed too. A three-model LightGBM ensemble scored 0.9660 against 0.9661 for one model. Half precision on the Apple GPU was slower (435 vs 548 pairs/s). Error analysis on v7 showed 41.5% of remaining errors were true matches that never reached the candidate set, which is why later versions went after recall in blocking.

Five anti-obesity drugs (semaglutide, liraglutide, tirzepatide, orlistat, phentermine) do not suit every patient equally. Type 2 diabetes, cardiovascular disease or chronic kidney disease changes which one is safest and most effective. The model has to rank them for each patient and agree with what large clinical trials established.

1. Heterogeneous graphpatient, gene, protein, drug, disease nodes
2. Pathway edgesgenes linked by Reactome co-membership
3. Type projectionsper-type linear layers to one hidden space
4. 2 × HANConvmulti-head attention over meta-paths
5. Drug–disease scoressample-weighted BCE loss

Auditing my own pipeline

A perfect score is a reason to look harder. Before trusting it, I audited the labels and features for ways the model could be learning my assumptions instead of the evidence.

Ablation studies on within-group personalization then set the boundaries of what graph-based decision support can claim here. Research prototype on synthetic patients, not for clinical use.

Given a patient's medications, conditions and genetic variants, flag the risks a prescriber should see, and show the graph path that explains each flag. Six public biomedical datasets are normalised and loaded into one Neo4j graph, which a FastAPI service queries for a React dashboard.

What the engine flags

• Drug–drug interactionsHarmful co-prescribed combinations
• ContraindicationsDrugs ruled out by an existing condition
• PolypharmacyRisk from many concurrent drugs
• Duplicate therapyTwo drugs on the same receptor
• PharmacogenomicsVariants that change drug response
• Side-effect riskSevere effects matched to the patient