How akash-bardia/amazon-ml-challenge-2026 works
This repository implements a multi-source business entity resolution pipeline: it normalizes vendor records, generates likely cross-source pairs, scores them, applies calibrated thresholds and target exclusivity, and writes matching and candidate TSVs. A separate training workflow builds hard-negative examples, fits a matcher, and evaluates it. The graph includes the major README stages; some subsystem links are based on documented capabilities because their source was not sampled.
Main entry points: inference.py Country-wise inference · train.py Training workflow · main.py Inference entry point · normalization.py Record normalization
Components 17
Matching runtime
- Inference entry point
code/business_entity_resolution/src/main.py - Country-wise inference
code/business_entity_resolution/src/inference.py
Resolution logic
- Record normalization
code/business_entity_resolution/src/normalization.py - Candidate blocking
code/business_entity_resolution/src/blocking.py - Pair features
code/business_entity_resolution/src/features.py - Matcher model
code/business_entity_resolution/src/model.py - Threshold and dedup
code/business_entity_resolution/src/thresholding.py - Evaluation metrics
code/business_entity_resolution/src/evaluation.py
Model training
- Training workflow
train.py - TSV and truth loading
code/business_entity_resolution/src/data_loader.py
Models and results
- Trained matcher
models/final_entity_matcher.joblib - Threshold metadata
models/model_metadata.json - Match and candidate TSVs
code/business_entity_resolution/src/inference.py - Submission validator
utils/validate_submission.py - Output utilities
code/business_entity_resolution/src/output.py
Other components
- Pipeline operator
- Vendor TSV inputs
Connections 23
- Pipeline operator → Inference entry point: runs
code/business_entity_resolution/src/main.py - Vendor TSV inputs → Country-wise inference: reads records
code/business_entity_resolution/src/inference.py - Inference entry point → Country-wise inference: runs inference
code/business_entity_resolution/src/main.py - Country-wise inference → Record normalization: normalizes records
code/business_entity_resolution/src/inference.py - Country-wise inference → Candidate blocking: generates candidates
code/business_entity_resolution/src/inference.py - Country-wise inference → Pair features: extracts pair features
code/business_entity_resolution/src/inference.py - Country-wise inference → Matcher model: scores pairs
code/business_entity_resolution/src/inference.py - Country-wise inference → Threshold and dedup: filters and deduplicates
code/business_entity_resolution/src/inference.py - Country-wise inference → Trained matcher: loads model
- Country-wise inference → Threshold metadata: loads thresholds
code/business_entity_resolution/src/inference.py - Country-wise inference → Match and candidate TSVs: writes TSVs
code/business_entity_resolution/src/inference.py - Inference entry point → Submission validator: runs validation
code/business_entity_resolution/src/main.py - Record normalization → Candidate blocking: provides normalized values
code/business_entity_resolution/src/blocking.py - Record normalization → Pair features: provides normalized values
code/business_entity_resolution/src/features.py - Training workflow → Candidate blocking: mines candidates
train.py - Training workflow → Pair features: builds training features
train.py - Training workflow → Matcher model: fits model
train.py - Training workflow → Evaluation metrics: evaluates predictions
train.py - Training workflow → Threshold and dedup: optimizes thresholds
train.py - Training workflow → Trained matcher: saves model
- Training workflow → Threshold metadata: saves metadata
- Threshold and dedup → Evaluation metrics: evaluates thresholds
code/business_entity_resolution/src/thresholding.py - TSV and truth loading → Vendor TSV inputs: loads TSV data