Accountability infrastructure for NYC housing

Anatomy of an Over-Merge — 235 River Avenue

Why seven owners collapse into one WoW portfolio · a single shared Lakewood office
Data question Explain why Who Owns What groups 27 buildings into a single owner, when they belong to seven different people — and how Watchline separates them.

Seven Owners, One Portfolio — How a Shared Office Fakes Common Ownership

Who Owns What fuses 27 buildings into one portfolio because their owners all file from a single back-office suite — 235 River Ave #220, Lakewood NJ. But those owners are seven different people. Watchline refuses the merge: it treats a high-degree shared office as non-probative and lands on 7 owners. This is the mirror image of the Escobar case — there a typo split one owner apart; here a shared address fuses many together.

◆ The over-merge, at a glance

27
Buildings WoW groups as one owner
1 → 7
WoW's one portfolio → Watchline's owners
45
Distinct owners filing from the one suite
51
Cross-owner address edges Watchline declines
7
Owners Watchline resolves
~6
Real people (the 7th is an unmerged typo)
Who Owns What — one portfolio
27 buildings
1 owner · 10 name slots · fused on the shared office
→ignore the aggregator office
Watchline — seven distinct owners
Joe Nebenzahl
7 BLDG
Abraham Miller
7 BLDG
Jeanette Padilla
4 BLDG
Nathan Obstfeld
4 BLDG
Vanessa Soria
3 BLDG
Nathan Obtfeld*
1 BLDG
Benyomin Fisch
1 BLDG
OwnerOwnerGroupBuildingsMerged name variants
Joe NebenzahlOG-558207JOE + JOSEPH NEBENZAHL
Abraham MillerOG-2337ABRAHAM + ABARAHAM MILLER
Jeanette PadillaOG-529164PADILLA / PADILLa
Nathan ObstfeldOG-850304
Vanessa SoriaOG-1102003VANESSA + VANEESA SORIA
Nathan Obtfeld*singleton1* OBTFELD typo, didn't merge with Obstfeld
Benyomin Fischsingleton1

I. The tie: one shared office

Every one of the 27 buildings files its head officer from the same back-office suite:

235 RIVER AVENUE #220, Lakewood NJ 08701
45 distinct owners · 33 registrations · 31 buildings file from here

That's not a home address — it's a shared registration office (a Lakewood back-office used by dozens of unrelated landlords), with a full degree of 45 distinct owners across the city. It's the housing-data equivalent of a mailing-list address: presence there says nothing about who owns what. Yet a shared string is exactly what registration-contact graphs link on.

II. How Who Owns What over-merges

WoW builds a node per (owner name, business address) and draws edges two ways: same name, or same business address. Here the names are all different — and typo-split into 10 name slots — so the name edges don't span them. But the address is identical for all 27. That single shared-office edge is enough: connected components sweeps the whole set into one portfolio.

Same office ≠ same owner WoW isn't claiming these seven people are one landlord — its graph simply can't distinguish "same office" from "same owner." A single aggregator address does all the merging; the differing names are ignored by the address edge.

III. How Watchline corrects it

The correction is subtler than "delete the shared address." The office does wire these owners together in the graph — there are 55 CONNECTED_BY_ADDRESS edges among them, 51 of which cross owner boundaries. Watchline's owner-identity layer (OwnerGroup) simply declines every one of the 51:

A high-degree office is non-probative The resolver treats an address shared by dozens of landlords as an operational nexus, not evidence of common ownership — so it refuses to merge on it. Result: 7 distinct owners, not one. And it still merges the within-person typos (JOE/JOSEPH, ABRAHAM/ABARAHAM, VANESSA/VANEESA) — the exact opposite discipline from WoW, which merges across the office but leaves each person's typos as separate slots.
Mind the layer The KG holds both answers. The Portfolio / APPARENT_CONTROL layer is built by connected-components over those 55 address edges, so it still over-merges (it reports one "Abraham Miller" controlling all 27) — mirroring WoW. Only OwnerGroup makes the correction. Read owner identity from the owner-group layer, never from the raw address edges.

IV. The mirror of Escobar

The same address-edge machinery fails in opposite directions, and Watchline's precision-first resolver fixes both:

 Escobar (under-merge)Miller (over-merge)
What the record doesa typo & blanked ZIP break one owner aparta shared office fuses seven owners together
WoW's mistakemisses links → one owner as 2 portfoliosover-links → seven owners as 1 portfolio
The failing signalexact address match (too strict)shared address match (too loose on an aggregator)
Watchline's fixmerges the typo'd same owner (deed/identity)declines the aggregator-address different owners
One principle, both directions Precision-first owner identity: merge on real evidence (a shared deed, a rare-name match), and refuse to merge on weak signals (a shared back-office). It reunites the owner the record fragments and splits the owners the record only appears to connect.

V. Method & provenance

office: 235 RIVER AVE #220, Lakewood NJ 27 buildings · 1 WoW portfolio (wow #183) owner identity: OwnerGroup
  • The tie (raw HPD): all 27 buildings' head-officer contacts in hpd_contacts file from 235 River Ave #220, Lakewood NJ; that suite has a degree of 45 distinct owners.
  • WoW's over-merge: wow_portfolios places all 27 in one portfolio, spanning 10 distinct name slots — linked purely by the shared business address.
  • Watchline's split: grouping the 27 buildings' registered landlords by IN_OWNER_GROUP yields 7 owner units (5 owner groups + 2 singletons); the 51 cross-owner CONNECTED_BY_ADDRESS edges are declined.
  • Explore: the interactive map — toggle Who Owns What's one portfolio ↔ Watchline's seven owners.
Watchline reports 7 owner units, but two are singletons — one (NATHAN OBTFELD) is an unmerged typo of Nathan Obstfeld — so it's really about six people. Precision-first: it under-merges a typo rather than risk a bad merge.
WoW orig_ids renumber between dataset refreshes (this portfolio has appeared as #183 and #210); the stable anchors are the office address and the BBL set. Ownership links are algorithmic inferences from public records — leads to verify, not determinations of legal ownership.

VI. See it in the graph (Cypher)

The exact subgraph behind this analysis, ready to paste into Neo4j Browser. The shared office isn't a KG node — it's the Landlord.bizaddr property behind the CONNECTED_BY_ADDRESS edges — so the query synthesizes a virtual :Office hub (in-memory, nothing written) and wires every landlord to it with a FILES_FROM edge. It returns the 27 buildings, their 11 landlords all MEMBER_OF one over-merged Portfolio, the shared-office web of CONNECTED_BY_ADDRESS edges, and the OwnerGroup split.

// === Miller over-merge subgraph + synthesized shared-office hub — 235 River Ave #220 ===
// The office is NOT a KG node (it's the Landlord.bizaddr property behind the
// CONNECTED_BY_ADDRESS edges), so we create a VIRTUAL :Office node and wire every
// landlord to it with a virtual FILES_FROM edge (in-memory; nothing is written).
WITH ['1004360028','1004360029','1004360129','1004370027','1019880024','1020700019',
      '2032780028','2032780054','2032790006','2032830022','2032890032','2032930142',
      '2032930144','2032930146','2033010063','3003050032','3003930043','3012680001',
      '3012820053','3027570036','3027890011','3032210031','3032370001','3050880046',
      '3051890045','3052150042','4031040070'] AS bbls
MATCH (b0:Building)<-[:REGISTERED_FOR]-(l0:Landlord) WHERE b0.bbl IN bbls
WITH bbls, collect(DISTINCT l0) AS lls
CALL apoc.create.vNode(['Office'], {name:'235 River Ave #220, Lakewood NJ', owners_on_file:45}) YIELD node AS office
UNWIND lls AS l
CALL apoc.create.vRelationship(l, 'FILES_FROM', {}, office) YIELD rel AS files
OPTIONAL MATCH (l)-[rf:REGISTERED_FOR]->(b:Building)          WHERE b.bbl IN bbls
OPTIONAL MATCH (l)-[inre:IN_OWNER_GROUP]->(re:OwnerGroup)
OPTIONAL MATCH (l)-[mem:MEMBER_OF]->(p:Portfolio)
OPTIONAL MATCH (l)-[conn:CONNECTED_BY_ADDRESS|CONNECTED_BY_NAME|CONNECTED_BY_SPLINK|CONNECTED_BY_DEED]-(l2:Landlord)
  WHERE l2 IN lls
RETURN office, files, l, rf, b, inre, re, mem, p, conn, l2

Renders ~44 nodes: 27 Building · 11 Landlord · 5 OwnerGroup (+ 2 unattached singletons) · 1 Portfolio · 1 virtual :Office. Colour relationships by type to see the CONNECTED_BY_ADDRESS tangle crossing the owner-group clusters.