Include "Administrative Area attribute" in Address Variants for ABP

[dodgydave67]

Another option to dramatically increase match weight would be to create variants with the "Administrative Area attribute", which can include the county, local area etc.

For example, below. If the ABP variant included Essex, it would increase the match weight for already pretty close addresses, despite the main address not really changing. Getting more already "good addresses" to have increased the match_weight would mean more addresses pass the threshold set, reducing the unnecessary ones potentially needed for manual review. Some messy address also contain multiple areas, like "Bristol, Gloucestershire".

In addition, when looking at the cleaned messy address, it's clear counties - like Essex, Bristol, Bath, Somerset - are not removed prior to matching. I do see that there's the trailing token logic in ukam, but I will be benchmarking by comparing including "Administrative Area" (by changing lpi.py) and without.

unique_id	matched_uprn	supplied_address	supplied_postcode	matched_address	matched_postcode	match_reason	match_weight	distinguishability
82	100091448086	3 Cripsey Avenue, Ongar, Essex, CM5 0AT	CM5 0AT	3 CRIPSEY AVENUE ONGAR	CM5 0AT	splink: probabilistic match	32.18106171899265	21.39156316236841

[dodgydave67]

I see the peeled stage in ukam, and potentially to reduce false positives, only values that exist in a set administrative area list (can probably be sourced from the ABP data), can be peeled. This might be more compute efficient than adding more "Administrative Area" variants.

[dodgydave67]

I made an administrative area patch, and I did a comparison on around 7000 addresses and did see 367 more exact full matches.

I can send you the comparison of the before and after.

It's potentially not worth the extra compute, unless you want more exact matches, and a slight increase in average match weights