EDA Findings Summary

1. Dataset Scale

Training data:

A naive comparison of S1 against S2+S3 would require approximately 22.8 trillion pairs. Therefore, candidate generation / blocking is essential.


2. Matching Structure

The problem is not one-to-one.

An S1 entity can have:

Therefore, the final system must support one-to-many matching.


3. Missing Values

S1:

Among true S2/S3 matches:

Conclusion:


4. Name Noise

Manual inspection of true matches showed:

Examples:

Therefore, exact string equality is insufficient.


5. Multilingual / Cross-Script Names

Non-ASCII names:

Among true matches:

For true matches with non-ASCII matched names:

Conclusion:

Cross-script matching is a significant part of the problem. Plain token similarity can completely fail for some genuine matches.

We have not yet concluded that multilingual embeddings are necessary.


6. Name Similarity

Using Unicode-preserving normalization and token Jaccard:

True-match distribution:

Name Jaccard Share
0–0.10 14.53%
0.10–0.25 1.61%
0.25–0.50 19.74%
0.50–0.75 32.35%
0.75–0.90 4.01%
0.90–0.99 ~0%
0.99–1.00 27.73%

Only ~21.85% of true matches have exactly equal names after the initial ASCII-oriented normalization.

Conclusion:

Name similarity is useful, but exact name equality is too restrictive.


7. Address Similarity

For true matches, token Jaccard on normalized addresses:

Conclusion:

Address similarity is useful, but addresses contain substantial noise and missingness.


Blocking Experiments

8. Exact Normalized-Name Blocking

We generated S1-S2 candidates using exact normalized business name.

Results:

Conclusion:

Exact-name blocking reduces the search space dramatically, but captures only about 1 in 5 true S1-S2 matches.

It is useful as one blocker, but cannot be the only blocking strategy.


9. What Exact-Name Blocking Misses

Among true S1-S2 pairs:

For the missed true matches:

Distribution of missed matches:

Name Jaccard Share
0–0.10 20.81%
0.10–0.25 1.72%
0.25–0.50 23.70%
0.50–0.75 40.83%
0.75–0.90 5.32%
0.90–1.00 7.63%

About 53.8% of true matches missed by exact-name blocking still have name Jaccard ≥ 0.5.

Conclusion:

Approximate name retrieval is strongly justified as a second blocking experiment.

However, about 20.8% of missed matches have name Jaccard < 0.1, so name-based blocking alone will still miss a substantial group.


10. Exact Normalized-Address Blocking

We generated S1-S2 candidates using exact normalized address.

Results:

Conclusion:

Exact-address blocking is highly precise but has low recall.

It produces far fewer candidates than exact-name blocking and a much higher proportion of true matches, but it misses many genuine matches because addresses are noisy.


11. What Exact-Address Blocking Misses

Among true S1-S2 pairs missed by exact-address blocking:

Distribution:

Address Jaccard Share
0–0.10 5.17%
0.10–0.25 2.14%
0.25–0.50 22.92%
0.50–0.75 40.61%
0.75–0.90 16.41%
0.90–1.00 12.75%

About 69.8% of true matches missed by exact-address blocking still have address Jaccard ≥ 0.5.

Conclusion:

Approximate address retrieval is also strongly justified.


12. Current Blocking Picture

Blocker Candidates Precision Recall
Exact normalized name 10.35M 7.66% 21.44%
Exact normalized address 564K 81.68% 12.48%

The two blockers have different strengths:

No single blocking rule appears sufficient.

The likely direction is to combine multiple candidate-generation strategies:

  1. Exact normalized name
  2. Approximate / character-level name retrieval
  3. Exact normalized address
  4. Approximate / token-level address retrieval
  5. Union the candidate sets
  6. Pass the resulting candidates to a matching model

The approximate approaches are experiments to test, not yet confirmed solutions.


13. Current Status

Completed

Next

  1. Test approximate name candidate generation.
  2. Test approximate address candidate generation.
  3. Measure candidate recall, candidate count, and runtime.
  4. Combine complementary blockers.
  5. Establish a fixed S1-level train/validation split before serious model comparison.
  6. Build hard-negative candidate datasets.
  7. Develop the matching model and optimize macro F0.5.