waynnbz / patient-clinic-proximity-finder

Find the nearest travel distance clinic for each patient. Mainly used Google Geocode and Distance Matrix APIs

Final design

• Let's do two approaches

  1. Pure Google
     • Using Google Geocoding API to obtain lat/lng
     • Using geopy.distance.geodesic to filter the clinic into N nearest ones in terms of geodesic distance (N=5/6 suggested after explore the clinic distribution on map
     • Finally using Google distance matrix API to get the shortest travel distance for the N clinics and output the nearest one
  2. Possibly free:
     • Geocode.ca throttled port for lat/lng
       • (tested to be more accurate than Google but throttle control applies in free port)
     • Same N pool processs using geodesic distance
     • Use OSMnx to gain a node/edge a graph centered at Patient node with a radius by considering N pool dist range
     • Finally using Networkx to compute the shortest path locally from the OSMnx graph

• Geocode detection failed case

  • Shouldn't break the code and still should be able to recommend clinics
    • 'Postal Code' < 'FSA' < 'City' < 'Province'

• Retry on failure:

  # To retry because intermittent failures sometimes occurs
  except (GeocoderQueryError) as error:
      if retry_counter < RETRY_COUNTER_CONST:
          return geocode_address(geo_locator, line_address, component_restrictions, retry_counter + 1)
      else:
          location_result = {"Lat": 0, "Long": 0, "Error": error.message, "formatted_address": "",
                             "location_type": ""}

• During geocoding process, the will run through the fallback schema list recursively if best address failed, and in worst case scenario fall back to default values set in config file

• I've switched to the Google geocoder as Nominatim is pretty sensitive to missing information, Google hammers through it no problem

• 'Postal Code' fail rate: 3/100 for C_df

  • however, all 3 failed cases are later correctly geocoded after adding a component filter country:CA
    • One building offset, comparing the result from component added failed case to its physical address
  • TYPO, there's typo case (eg. V3Z 6S7 which mapped to V3S 6S7)
    • By adding physical 'address' GoogleV3 still able to figure out the correct address
    • Strangely enough, V3Z 6S7 is able to find a locale from OSM Geocoder.CA (and all other failed GoogleV3 case), same locale as from Google but labeled as input V3Z 6S7. Maybe the typo is from Google???
    • I guess this proves that Geocoder.CA is more accurate than GoogleV3 (w/ country component set to Canada)

• 'Address' alone fail rate 11/100 (expected)

  • However, with country component filtering, the failed cases only reduced 1, yields 10/100

• Cost and rate limit:

  • for 0-100,000: $0.005 per each
  • 50 request per second(QPS)