idgf

pipeline/transform/transform_poi.py

@@ -6,6 +6,10 @@ import polars as pl
 from pipeline.utils.england_geometry import in_england_mask
 
 DROP_CATEGORIES = {
+    # GEOLYTIX Grocery Retail Points is the authoritative supermarket source
+    # (transform_grocery_retail_points), so drop OSM supermarkets to avoid
+    # double-counting each store as both a GEOLYTIX brand and an OSM "Supermarket".
+    "shop/supermarket",
     # Street furniture & infrastructure
     "amenity/advice",
     "amenity/atm",
@@ -364,14 +368,6 @@ _CATEGORIES: list[tuple[str, str, str, list[str]]] = [
             "leisure/yes",
         ],
     ),
-    (
-        "Groceries",
-        "Supermarket",
-        "🛒",
-        [
-            "shop/supermarket",
-        ],
-    ),
     (
         "Groceries",
         "Convenience Store",
@@ -1534,6 +1530,14 @@ def transform(
         pl.col("category").replace_strict(emoji_mapping).alias("emoji"),
     )
 
+    # A single OSM object can carry several tag keys that map to the same
+    # friendly category (e.g. amenity/pharmacy + shop/chemist -> "Pharmacy"),
+    # which pois.py emits as multiple raw rows sharing one id. Collapse those
+    # duplicates so they don't inflate downstream proximity counts; rows sharing
+    # an id with DIFFERENT categories are preserved. Other sources are
+    # pre-deduplicated.
+    lf = lf.unique(subset=["id", "category"], keep="first", maintain_order=True)
+
     naptan_df = pl.scan_parquet(naptan_path).collect()
     mask = in_england_mask(
         boundary_path,