4.1 Stop Problems

Stop problems are detected after the matching pipeline completes and identify data quality issues regarding individual stops.

Detection uses a predicate pipeline mirroring the matching pipeline architecture. During database import, ProblemContext.build() precomputes shared indexes (KDTrees, UIC counts, duplicate maps, and route evidence lookups) from the MatchingOutput. Each stop is then evaluated by the five problem predicates.

Code: matching_and_import_db/problem_detection/

How Detection Runs

1. Context construction — ProblemContext.build(MatchingOutput)

All three lists from the matching output are consumed once to precompute shared indexes (KDTrees, UIC counts, duplicate maps). The resulting ProblemContext is then passed as the ctx argument to every predicate call.

2. Invocation — which records are evaluated by which predicates

Each record is passed individually as the stop_dict argument. Predicates use isinstance checks to decide whether they apply and return [] when the record type is irrelevant.

Invocation Paths

Matched records use MatchRecord.evaluate_matched_problems(), which calls each predicate with the MatchRecord itself. The predicates access record.atlas_node and record.osm_node directly:

current_match.evaluate_matched_problems(problem_ctx, STOP_PROBLEM_PIPELINE)

Unmatched records are passed as bare AtlasNode or OsmNode entities to evaluate_unmatched_problems():

problems = evaluate_unmatched_problems(STOP_PROBLEM_PIPELINE, problem_ctx, atlas_node)

Each predicate uses isinstance checks to decide whether it applies — distance_problem returns [] for bare nodes, unmatched_problem returns [] for MatchRecord, etc.

ProblemContext

Built once from PipelineResult via ProblemContext.build(), providing precomputed indexes:

Stop Problem Types

4.1.1. Distance Problems

Flag matched pairs where physical distance exceeds tolerance. This typically indicates either a matching error or significant coordinate discrepancy between datasets.

The predicate reads record.distance_m and record.atlas_node.business_org_abbr directly from the MatchRecord.

Thresholds

DISTANCE_THRESHOLD_P1 = 80   # meters
DISTANCE_THRESHOLD_P2 = 25   # meters
DISTANCE_THRESHOLD_P3 = 15   # meters

Priority Logic

Example: A bus stop matched with 85m distance would be flagged as P1 (critical), while a train platform with the same distance would be P3 (minor).

4.1.2. Unmatched Problems

Identify stops that failed to match. The predicate receives bare AtlasNode or OsmNode entities and uses ProblemContext spatial indexes to compute isolation.

ATLAS Unmatched Priority

Uses ctx.nearest_osm_distance() (KDTree query) and ctx.osm_count_by_uic:

OSM Unmatched Priority

Uses ctx.nearest_atlas_distance() (KDTree query) and ctx.atlas_count_by_uic:

Isolation Detection

Isolation is computed using ProblemContext.nearest_osm_distance() / nearest_atlas_distance(), which query the precomputed KDTrees. Separately from problem detection, the importer marks unmatched ATLAS entries with no OSM node within 50m as match_type='no_nearby_counterpart' in stops_matched.

4.1.3. Attribute Problems

Flag inconsistencies between matched pairs. The predicate reads fields directly from record.atlas_node and record.osm_node on the MatchRecord.

Priority Logic

Note: Name and local_ref comparisons are case-insensitive. UIC comparisons are exact. Each check can be individually toggled via ENABLE_*_CHECK constants in context.py.

4.1.4. Contradicts Route Matching Problems

Flag matched pairs whose own route evidence contradicts the chosen ATLAS↔OSM stop pairing. This predicate reads route evidence from ProblemContext.atlas_routes_by_sloid, ProblemContext.osm_node_routes, and ProblemContext.osm_name_dirs, then reuses the same route-alignment classifier as the matching pipeline.

Priority Logic

4.1.5. Duplicate Problems

Identify redundant entries in either dataset. This predicate is polymorphic — it handles MatchRecord, AtlasNode, and OsmNode.

OSM duplicates are only flagged for nodes with OsmNode.public_transport equal to platform or stop_position. When both ATLAS and OSM duplicates exist, only the OSM duplicate is flagged (OSM-side takes precedence). ATLAS duplicates that already produced a duplicate_propagation match are deliberately suppressed to avoid double-reporting the same grouping behavior.