Hey everyone,

I am working on a real-world data quality problem and would appreciate feedback on my modeling approach.

Context:

I have a dataset of meters and their associated transformers (utility infrastructure). Some of these associations are incorrect, and the goal is to both detect and correct them.

Training data:

I’m using ~20,000 manually reviewed meter–transformer associations:

- Correct association → label = 1

- Incorrect association → label = 0

For incorrect cases, I also augment the data with the correct transformer, e.g.:

Meter1 | Trans1 | 0 (incorrect)

Meter1 | Trans2 | 1 (corrected)

Meter2 | Trans3 | 1 (correct)

Current baseline:

I started with a logistic regression model (class_weight="balanced" due to ~37% incorrect vs 63% correct).

Using a 0.20 threshold gives strong true negative performance (~98%), but only moderate recall.

Candidate generation:

For inference, I generate candidate transformers within a 550 ft radius for each meter (including the currently assigned…