Cross Country Model Training (Normalized)

Cross-country model for predicting wealth index

In this notebook, we normalized each model feature by country to correct country-specific feature quirks, as well as the wealth index.

This method has been done in a similar study using DHS as training data and was also documented here.

Models recover how relative values of input data (relative to other locations in that country) correlate with relative values of wealth (also relative to other locations in that country).

The current best model setup involves (1) normalizing all features per country using MinMaxScaler and (2) scaling the wealth index per country using StandardScaler

import sys
sys.path.append("../../")

import os

import uuid
import numpy as np
import geopandas as gpd
import pandas as pd
from shapely import wkt

from geowrangler import dhs
from povertymapping import settings, osm, ookla, nightlights
from povertymapping.dhs import generate_dhs_cluster_level_data
from povertymapping.osm import OsmDataManager
from povertymapping.ookla import OoklaDataManager
import getpass
import pickle

from sklearn.model_selection import train_test_split, KFold, RepeatedKFold
from sklearn.model_selection import GroupKFold, cross_val_predict, cross_val_score, LeaveOneGroupOut
from sklearn.preprocessing import MinMaxScaler, RobustScaler, StandardScaler
from sklearn.metrics import r2_score
import matplotlib.pyplot as plt
import seaborn as sns

import shap

/home/jace/workspace/unicef-ai4d-poverty-mapping/env/lib/python3.9/site-packages/geopandas/_compat.py:111: UserWarning: The Shapely GEOS version (3.11.1-CAPI-1.17.1) is incompatible with the GEOS version PyGEOS was compiled with (3.10.1-CAPI-1.16.0). Conversions between both will be slow.
  warnings.warn(
/home/jace/workspace/unicef-ai4d-poverty-mapping/env/lib/python3.9/site-packages/tqdm/auto.py:22: TqdmWarning: IProgress not found. Please update jupyter and ipywidgets. See https://ipywidgets.readthedocs.io/en/stable/user_install.html
  from .autonotebook import tqdm as notebook_tqdm

%reload_ext autoreload
%autoreload 2

Set up Data Access

# # Instantiate data managers for Ookla and OSM
# # This auto-caches requested data in RAM, so next fetches of the data are faster.
osm_data_manager = OsmDataManager(cache_dir=settings.ROOT_DIR/"data/data_cache")
ookla_data_manager = OoklaDataManager(cache_dir=settings.ROOT_DIR/"data/data_cache")

# Log-in using EOG credentials
username = os.environ.get('EOG_USER',None)
username = username if username is not None else input('Username?')
password = os.environ.get('EOG_PASSWORD',None)
password = password if password is not None else getpass.getpass('Password?') 

# set save_token to True so that access token gets stored in ~/.eog_creds/eog_access_token
access_token = nightlights.get_eog_access_token(username,password, save_token=True)

2023-02-13 16:07:15.849 | INFO     | povertymapping.nightlights:get_eog_access_token:48 - Saving access_token to ~/.eog_creds/eog_access_token
2023-02-13 16:07:15.851 | INFO     | povertymapping.nightlights:get_eog_access_token:56 - Adding access token to environmentt var EOG_ACCESS_TOKEN

Load Countries From DHS data

# Set country-specific variables
country_config = {
    'Philippines': {
        'country_osm':'philippines',
        'ookla_year': 2019,
        'nightlights_year' : 2017,
        'country_code': 'ph',
        'dhs_household_dta_path' : settings.DATA_DIR/"dhs/ph/PHHR71DT/PHHR71FL.DTA",
        'dhs_geographic_shp_path' : settings.DATA_DIR/"dhs/ph/PHGE71FL/PHGE71FL.shp"
    },
    'Timor Leste': {
        'country_osm':'east-timor',
        'ookla_year': 2019,
        'nightlights_year' : 2016,
        'country_code': 'tl',
        'dhs_household_dta_path' : settings.DATA_DIR/"dhs/tl/TLHR71DT/TLHR71FL.DTA",
        'dhs_geographic_shp_path' : settings.DATA_DIR/"dhs/tl/TLGE71FL/TLGE71FL.shp"
    },
    'Cambodia': {
        'country_osm':'cambodia',
        'ookla_year': 2019,
        'nightlights_year' : 2014,
        'country_code': 'kh',
        'dhs_household_dta_path' : settings.DATA_DIR/"dhs/kh/KHHR73DT/KHHR73FL.DTA",
        'dhs_geographic_shp_path' : settings.DATA_DIR/"dhs/kh/KHGE71FL/KHGE71FL.shp"
    },
    'Myanmar': {
        'country_osm':'myanmar',
        'ookla_year': 2019,
        'nightlights_year' : 2015,
        'country_code': 'mm',
        'dhs_household_dta_path' : settings.DATA_DIR/"dhs/mm/MMHR71DT/MMHR71FL.DTA",
        'dhs_geographic_shp_path' : settings.DATA_DIR/"dhs/mm/MMGE71FL/MMGE71FL.shp"
    }
}

Combine base features for all four countries

This section creates countries_data (pd.DataFrame) which combines DHS/Ookla/OSM/VIIRS data for all specified countries.

%%time

# Create list of dataframes per country, and list of all columns appearing in DHS columns
country_data_list = []
dhs_columns = ['DHSCLUST']

for country, config in country_config.items():
    print(f'Loading data for {country}')

    # Load the DHS cluster data
    dhs_household_dta_path = config['dhs_household_dta_path']
    dhs_geographic_shp_path = config['dhs_geographic_shp_path']
    country_code = config['country_code']
    dhs_gdf = generate_dhs_cluster_level_data(
        dhs_household_dta_path, 
        dhs_geographic_shp_path, 
        col_rename_config=country_code,
        convert_geoms_to_bbox=True,
        bbox_size_km = 2.4
    ).reset_index(drop=True)

    dhs_columns = list(set(dhs_columns + list(dhs_gdf.columns)))
    

    # Generate base features for the dhs dataframe
    country_data = dhs_gdf.copy()
    country_osm = config['country_osm']
    ookla_year = config['ookla_year']
    nightlights_year = config['nightlights_year']
    
    # Add in OSM features
    country_data = osm.add_osm_poi_features(country_data, country_osm, osm_data_manager)
    country_data = osm.add_osm_road_features(country_data, country_osm, osm_data_manager)

    # Add in Ookla features
    country_data = ookla.add_ookla_features(country_data, 'fixed', ookla_year, ookla_data_manager)
    country_data = ookla.add_ookla_features(country_data, 'mobile', ookla_year, ookla_data_manager)

    # Add in the nighttime lights features
    country_data = nightlights.generate_nightlights_feature(country_data, nightlights_year) 

    country_data_list.append(country_data)

# Combine all country data into a single dataframe
countries_data = gpd.GeoDataFrame(pd.concat(country_data_list,ignore_index=True), crs=country_data_list[0].crs)

Loading data for Philippines

2023-02-13 16:08:25.879 | INFO     | povertymapping.osm:download_osm_country_data:187 - OSM Data: Cached data available for philippines at /home/jace/workspace/unicef-ai4d-poverty-mapping/notebooks/2023-01-24-dhs-cross-country-experiments/../../data/data_cache/osm/philippines? True
2023-02-13 16:08:25.881 | DEBUG    | povertymapping.osm:load_pois:149 - OSM POIs for philippines being loaded from /home/jace/workspace/unicef-ai4d-poverty-mapping/notebooks/2023-01-24-dhs-cross-country-experiments/../../data/data_cache/osm/philippines/gis_osm_pois_free_1.shp
2023-02-13 16:08:35.008 | INFO     | povertymapping.osm:download_osm_country_data:187 - OSM Data: Cached data available for philippines at /home/jace/workspace/unicef-ai4d-poverty-mapping/notebooks/2023-01-24-dhs-cross-country-experiments/../../data/data_cache/osm/philippines? True
2023-02-13 16:08:35.009 | DEBUG    | povertymapping.osm:load_roads:168 - OSM Roads for philippines being loaded from /home/jace/workspace/unicef-ai4d-poverty-mapping/notebooks/2023-01-24-dhs-cross-country-experiments/../../data/data_cache/osm/philippines/gis_osm_roads_free_1.shp
2023-02-13 16:09:59.097 | DEBUG    | povertymapping.ookla:load_type_year_data:68 - Contents of data cache: []
2023-02-13 16:09:59.098 | INFO     | povertymapping.ookla:load_type_year_data:83 - Cached data available at /home/jace/workspace/unicef-ai4d-poverty-mapping/notebooks/2023-01-24-dhs-cross-country-experiments/../../data/data_cache/ookla/processed/2f858b388182d50703550c8ef9d321df.csv? True
2023-02-13 16:09:59.099 | DEBUG    | povertymapping.ookla:load_type_year_data:88 - Processed Ookla data for aoi, fixed 2019 (key: 2f858b388182d50703550c8ef9d321df) found in filesystem. Loading in cache.
2023-02-13 16:09:59.538 | DEBUG    | povertymapping.ookla:load_type_year_data:68 - Contents of data cache: ['2f858b388182d50703550c8ef9d321df']
2023-02-13 16:09:59.539 | INFO     | povertymapping.ookla:load_type_year_data:83 - Cached data available at /home/jace/workspace/unicef-ai4d-poverty-mapping/notebooks/2023-01-24-dhs-cross-country-experiments/../../data/data_cache/ookla/processed/5a45dc45080a935951e6c2b6c0052b13.csv? True
2023-02-13 16:09:59.540 | DEBUG    | povertymapping.ookla:load_type_year_data:88 - Processed Ookla data for aoi, mobile 2019 (key: 5a45dc45080a935951e6c2b6c0052b13) found in filesystem. Loading in cache.
2023-02-13 16:09:59.921 | INFO     | povertymapping.nightlights:get_clipped_raster:414 - Retrieving clipped raster file /home/jace/.geowrangler/nightlights/clip/295bf47ce6753c7f06ab79012b769f2a.tif

Loading data for Timor Leste

2023-02-13 16:10:10.957 | INFO     | povertymapping.osm:download_osm_country_data:187 - OSM Data: Cached data available for east-timor at /home/jace/workspace/unicef-ai4d-poverty-mapping/notebooks/2023-01-24-dhs-cross-country-experiments/../../data/data_cache/osm/east-timor? True
2023-02-13 16:10:10.958 | DEBUG    | povertymapping.osm:load_pois:149 - OSM POIs for east-timor being loaded from /home/jace/workspace/unicef-ai4d-poverty-mapping/notebooks/2023-01-24-dhs-cross-country-experiments/../../data/data_cache/osm/east-timor/gis_osm_pois_free_1.shp
2023-02-13 16:10:12.045 | INFO     | povertymapping.osm:download_osm_country_data:187 - OSM Data: Cached data available for east-timor at /home/jace/workspace/unicef-ai4d-poverty-mapping/notebooks/2023-01-24-dhs-cross-country-experiments/../../data/data_cache/osm/east-timor? True
2023-02-13 16:10:12.046 | DEBUG    | povertymapping.osm:load_roads:168 - OSM Roads for east-timor being loaded from /home/jace/workspace/unicef-ai4d-poverty-mapping/notebooks/2023-01-24-dhs-cross-country-experiments/../../data/data_cache/osm/east-timor/gis_osm_roads_free_1.shp
2023-02-13 16:10:12.646 | DEBUG    | povertymapping.ookla:load_type_year_data:68 - Contents of data cache: ['2f858b388182d50703550c8ef9d321df', '5a45dc45080a935951e6c2b6c0052b13']
2023-02-13 16:10:12.647 | INFO     | povertymapping.ookla:load_type_year_data:83 - Cached data available at /home/jace/workspace/unicef-ai4d-poverty-mapping/notebooks/2023-01-24-dhs-cross-country-experiments/../../data/data_cache/ookla/processed/206a0323fa0e80f82339b66d0c859b4a.csv? True
2023-02-13 16:10:12.648 | DEBUG    | povertymapping.ookla:load_type_year_data:88 - Processed Ookla data for aoi, fixed 2019 (key: 206a0323fa0e80f82339b66d0c859b4a) found in filesystem. Loading in cache.
2023-02-13 16:10:12.736 | DEBUG    | povertymapping.ookla:load_type_year_data:68 - Contents of data cache: ['2f858b388182d50703550c8ef9d321df', '5a45dc45080a935951e6c2b6c0052b13', '206a0323fa0e80f82339b66d0c859b4a']
2023-02-13 16:10:12.737 | INFO     | povertymapping.ookla:load_type_year_data:83 - Cached data available at /home/jace/workspace/unicef-ai4d-poverty-mapping/notebooks/2023-01-24-dhs-cross-country-experiments/../../data/data_cache/ookla/processed/209c2544788b8e2bdf4db4685c50e26d.csv? True
2023-02-13 16:10:12.738 | DEBUG    | povertymapping.ookla:load_type_year_data:88 - Processed Ookla data for aoi, mobile 2019 (key: 209c2544788b8e2bdf4db4685c50e26d) found in filesystem. Loading in cache.
2023-02-13 16:10:12.847 | INFO     | povertymapping.nightlights:get_clipped_raster:414 - Retrieving clipped raster file /home/jace/.geowrangler/nightlights/clip/b0d0551dd5a67c8eada595334f2655ed.tif

Loading data for Cambodia

2023-02-13 16:10:17.428 | INFO     | povertymapping.osm:download_osm_country_data:187 - OSM Data: Cached data available for cambodia at /home/jace/workspace/unicef-ai4d-poverty-mapping/notebooks/2023-01-24-dhs-cross-country-experiments/../../data/data_cache/osm/cambodia? True
2023-02-13 16:10:17.429 | DEBUG    | povertymapping.osm:load_pois:149 - OSM POIs for cambodia being loaded from /home/jace/workspace/unicef-ai4d-poverty-mapping/notebooks/2023-01-24-dhs-cross-country-experiments/../../data/data_cache/osm/cambodia/gis_osm_pois_free_1.shp
2023-02-13 16:10:19.195 | INFO     | povertymapping.osm:download_osm_country_data:187 - OSM Data: Cached data available for cambodia at /home/jace/workspace/unicef-ai4d-poverty-mapping/notebooks/2023-01-24-dhs-cross-country-experiments/../../data/data_cache/osm/cambodia? True
2023-02-13 16:10:19.195 | DEBUG    | povertymapping.osm:load_roads:168 - OSM Roads for cambodia being loaded from /home/jace/workspace/unicef-ai4d-poverty-mapping/notebooks/2023-01-24-dhs-cross-country-experiments/../../data/data_cache/osm/cambodia/gis_osm_roads_free_1.shp
2023-02-13 16:10:25.192 | DEBUG    | povertymapping.ookla:load_type_year_data:68 - Contents of data cache: ['2f858b388182d50703550c8ef9d321df', '5a45dc45080a935951e6c2b6c0052b13', '206a0323fa0e80f82339b66d0c859b4a', '209c2544788b8e2bdf4db4685c50e26d']
2023-02-13 16:10:25.193 | INFO     | povertymapping.ookla:load_type_year_data:83 - Cached data available at /home/jace/workspace/unicef-ai4d-poverty-mapping/notebooks/2023-01-24-dhs-cross-country-experiments/../../data/data_cache/ookla/processed/37f570ebc130cb44f9dba877fbda74e2.csv? True
2023-02-13 16:10:25.194 | DEBUG    | povertymapping.ookla:load_type_year_data:88 - Processed Ookla data for aoi, fixed 2019 (key: 37f570ebc130cb44f9dba877fbda74e2) found in filesystem. Loading in cache.
2023-02-13 16:10:25.393 | DEBUG    | povertymapping.ookla:load_type_year_data:68 - Contents of data cache: ['2f858b388182d50703550c8ef9d321df', '5a45dc45080a935951e6c2b6c0052b13', '206a0323fa0e80f82339b66d0c859b4a', '209c2544788b8e2bdf4db4685c50e26d', '37f570ebc130cb44f9dba877fbda74e2']
2023-02-13 16:10:25.394 | INFO     | povertymapping.ookla:load_type_year_data:83 - Cached data available at /home/jace/workspace/unicef-ai4d-poverty-mapping/notebooks/2023-01-24-dhs-cross-country-experiments/../../data/data_cache/ookla/processed/1128a917060f7bb88c0a6260ed457091.csv? True
2023-02-13 16:10:25.395 | DEBUG    | povertymapping.ookla:load_type_year_data:88 - Processed Ookla data for aoi, mobile 2019 (key: 1128a917060f7bb88c0a6260ed457091) found in filesystem. Loading in cache.
2023-02-13 16:10:25.597 | INFO     | povertymapping.nightlights:get_clipped_raster:414 - Retrieving clipped raster file /home/jace/.geowrangler/nightlights/clip/4791e78094ba7e323fd5814b3f094a84.tif

Loading data for Myanmar

2023-02-13 16:10:31.819 | INFO     | povertymapping.osm:download_osm_country_data:187 - OSM Data: Cached data available for myanmar at /home/jace/workspace/unicef-ai4d-poverty-mapping/notebooks/2023-01-24-dhs-cross-country-experiments/../../data/data_cache/osm/myanmar? True
2023-02-13 16:10:31.820 | DEBUG    | povertymapping.osm:load_pois:149 - OSM POIs for myanmar being loaded from /home/jace/workspace/unicef-ai4d-poverty-mapping/notebooks/2023-01-24-dhs-cross-country-experiments/../../data/data_cache/osm/myanmar/gis_osm_pois_free_1.shp
2023-02-13 16:10:34.204 | INFO     | povertymapping.osm:download_osm_country_data:187 - OSM Data: Cached data available for myanmar at /home/jace/workspace/unicef-ai4d-poverty-mapping/notebooks/2023-01-24-dhs-cross-country-experiments/../../data/data_cache/osm/myanmar? True
2023-02-13 16:10:34.205 | DEBUG    | povertymapping.osm:load_roads:168 - OSM Roads for myanmar being loaded from /home/jace/workspace/unicef-ai4d-poverty-mapping/notebooks/2023-01-24-dhs-cross-country-experiments/../../data/data_cache/osm/myanmar/gis_osm_roads_free_1.shp
2023-02-13 16:10:51.602 | DEBUG    | povertymapping.ookla:load_type_year_data:68 - Contents of data cache: ['2f858b388182d50703550c8ef9d321df', '5a45dc45080a935951e6c2b6c0052b13', '206a0323fa0e80f82339b66d0c859b4a', '209c2544788b8e2bdf4db4685c50e26d', '37f570ebc130cb44f9dba877fbda74e2', '1128a917060f7bb88c0a6260ed457091']
2023-02-13 16:10:51.603 | INFO     | povertymapping.ookla:load_type_year_data:83 - Cached data available at /home/jace/workspace/unicef-ai4d-poverty-mapping/notebooks/2023-01-24-dhs-cross-country-experiments/../../data/data_cache/ookla/processed/d72ec7e4d144b750e1c0950ecad081e0.csv? True
2023-02-13 16:10:51.604 | DEBUG    | povertymapping.ookla:load_type_year_data:88 - Processed Ookla data for aoi, fixed 2019 (key: d72ec7e4d144b750e1c0950ecad081e0) found in filesystem. Loading in cache.
2023-02-13 16:10:51.740 | DEBUG    | povertymapping.ookla:load_type_year_data:68 - Contents of data cache: ['2f858b388182d50703550c8ef9d321df', '5a45dc45080a935951e6c2b6c0052b13', '206a0323fa0e80f82339b66d0c859b4a', '209c2544788b8e2bdf4db4685c50e26d', '37f570ebc130cb44f9dba877fbda74e2', '1128a917060f7bb88c0a6260ed457091', 'd72ec7e4d144b750e1c0950ecad081e0']
2023-02-13 16:10:51.741 | INFO     | povertymapping.ookla:load_type_year_data:83 - Cached data available at /home/jace/workspace/unicef-ai4d-poverty-mapping/notebooks/2023-01-24-dhs-cross-country-experiments/../../data/data_cache/ookla/processed/2aff65fdf8072457cba0d42873b7a9c2.csv? True
2023-02-13 16:10:51.742 | DEBUG    | povertymapping.ookla:load_type_year_data:88 - Processed Ookla data for aoi, mobile 2019 (key: 2aff65fdf8072457cba0d42873b7a9c2) found in filesystem. Loading in cache.
2023-02-13 16:10:51.911 | INFO     | povertymapping.nightlights:get_clipped_raster:414 - Retrieving clipped raster file /home/jace/.geowrangler/nightlights/clip/7a58f067614b6685cd9bb62d4d15a249.tif

CPU times: user 2min 21s, sys: 1min 9s, total: 3min 30s
Wall time: 3min 37s

Inspect the combined target country data

countries_data.info()

<class 'geopandas.geodataframe.GeoDataFrame'>
RangeIndex: 2720 entries, 0 to 2719
Data columns (total 86 columns):
 #   Column                             Non-Null Count  Dtype   
---  ------                             --------------  -----   
 0   DHSCLUST                           2720 non-null   int64   
 1   Wealth Index                       2720 non-null   float64 
 2   DHSID                              2720 non-null   object  
 3   DHSCC                              2720 non-null   object  
 4   DHSYEAR                            2720 non-null   float64 
 5   CCFIPS                             2720 non-null   object  
 6   ADM1FIPS                           2720 non-null   object  
 7   ADM1FIPSNA                         2720 non-null   object  
 8   ADM1SALBNA                         2720 non-null   object  
 9   ADM1SALBCO                         2720 non-null   object  
 10  ADM1DHS                            2720 non-null   float64 
 11  ADM1NAME                           2720 non-null   object  
 12  DHSREGCO                           2720 non-null   float64 
 13  DHSREGNA                           2720 non-null   object  
 14  SOURCE                             2720 non-null   object  
 15  URBAN_RURA                         2720 non-null   object  
 16  LATNUM                             2720 non-null   float64 
 17  LONGNUM                            2720 non-null   float64 
 18  ALT_GPS                            2720 non-null   float64 
 19  ALT_DEM                            2720 non-null   float64 
 20  DATUM                              2720 non-null   object  
 21  geometry                           2720 non-null   geometry
 22  poi_count                          2720 non-null   float64 
 23  atm_count                          2720 non-null   float64 
 24  atm_nearest                        2720 non-null   float64 
 25  bank_count                         2720 non-null   float64 
 26  bank_nearest                       2720 non-null   float64 
 27  bus_station_count                  2720 non-null   float64 
 28  bus_station_nearest                2720 non-null   float64 
 29  cafe_count                         2720 non-null   float64 
 30  cafe_nearest                       2720 non-null   float64 
 31  charging_station_count             2720 non-null   float64 
 32  charging_station_nearest           2720 non-null   float64 
 33  courthouse_count                   2720 non-null   float64 
 34  courthouse_nearest                 2720 non-null   float64 
 35  dentist_count                      2720 non-null   float64 
 36  dentist_nearest                    2720 non-null   float64 
 37  fast_food_count                    2720 non-null   float64 
 38  fast_food_nearest                  2720 non-null   float64 
 39  fire_station_count                 2720 non-null   float64 
 40  fire_station_nearest               2720 non-null   float64 
 41  food_court_count                   2720 non-null   float64 
 42  food_court_nearest                 2720 non-null   float64 
 43  fuel_count                         2720 non-null   float64 
 44  fuel_nearest                       2720 non-null   float64 
 45  hospital_count                     2720 non-null   float64 
 46  hospital_nearest                   2720 non-null   float64 
 47  library_count                      2720 non-null   float64 
 48  library_nearest                    2720 non-null   float64 
 49  marketplace_count                  2720 non-null   float64 
 50  marketplace_nearest                2720 non-null   float64 
 51  pharmacy_count                     2720 non-null   float64 
 52  pharmacy_nearest                   2720 non-null   float64 
 53  police_count                       2720 non-null   float64 
 54  police_nearest                     2720 non-null   float64 
 55  post_box_count                     2720 non-null   float64 
 56  post_box_nearest                   2720 non-null   float64 
 57  post_office_count                  2720 non-null   float64 
 58  post_office_nearest                2720 non-null   float64 
 59  restaurant_count                   2720 non-null   float64 
 60  restaurant_nearest                 2720 non-null   float64 
 61  social_facility_count              2720 non-null   float64 
 62  social_facility_nearest            2720 non-null   float64 
 63  supermarket_count                  2720 non-null   float64 
 64  supermarket_nearest                2720 non-null   float64 
 65  townhall_count                     2720 non-null   float64 
 66  townhall_nearest                   2720 non-null   float64 
 67  road_count                         2720 non-null   float64 
 68  fixed_2019_mean_avg_d_kbps_mean    1443 non-null   float64 
 69  fixed_2019_mean_avg_u_kbps_mean    1443 non-null   float64 
 70  fixed_2019_mean_avg_lat_ms_mean    1443 non-null   float64 
 71  fixed_2019_mean_num_tests_mean     1443 non-null   float64 
 72  fixed_2019_mean_num_devices_mean   1443 non-null   float64 
 73  mobile_2019_mean_avg_d_kbps_mean   1707 non-null   float64 
 74  mobile_2019_mean_avg_u_kbps_mean   1707 non-null   float64 
 75  mobile_2019_mean_avg_lat_ms_mean   1707 non-null   float64 
 76  mobile_2019_mean_num_tests_mean    1707 non-null   float64 
 77  mobile_2019_mean_num_devices_mean  1707 non-null   float64 
 78  avg_rad_min                        2720 non-null   float64 
 79  avg_rad_max                        2720 non-null   float64 
 80  avg_rad_mean                       2720 non-null   float64 
 81  avg_rad_std                        2720 non-null   float64 
 82  avg_rad_median                     2720 non-null   float64 
 83  F21                                0 non-null      object  
 84  F22                                0 non-null      object  
 85  F23                                0 non-null      object  
dtypes: float64(69), geometry(1), int64(1), object(15)
memory usage: 1.8+ MB

# Check if DHS CC can be used as a country grouping column
countries_data['DHSCC'].value_counts()

PH    1213
KH     611
TL     455
MM     441
Name: DHSCC, dtype: int64

# Explore as a map with first 30 columns in tooltip
#countries_data.iloc[:, :30].explore()

Data Preparation

Split into labels and features

# Set parameters
label_col = 'Wealth Index'
group_col = 'DHSCC'
##normalize_labels = 'single-country' # False, 'cross-country' or 'single_country'
normalize_labels = 'cross-country'

drop_cols = ['DHSID', 'DHSYEAR', 'CCFIPS',
       'ADM1FIPS', 'ADM1FIPSNA', 'ADM1SALBNA', 'ADM1SALBCO', 'ADM1DHS',
       'ADM1NAME', 'DHSREGCO', 'DHSREGNA', 'SOURCE', 'LATNUM',
       'LONGNUM', 'ALT_GPS', 'ALT_DEM', 'DATUM','F21','F22','F23','URBAN_RURA']

# Split train/test data into features and labels

# For labels, we just select the target label column
labels = countries_data[[label_col]]

# For features, drop all columns from the input dhs files
# If you need the cluster data, refer to country_data / country_test
features = countries_data.drop(drop_cols, axis=1).copy()

features.shape, labels.shape

((2720, 65), (2720, 1))

Normalize labels and features

def scale_col(X,scaler):
    X_ = np.atleast_2d(X)
    return pd.DataFrame(scaler.fit_transform(X_), X.index)

feature_cols = features.columns[4:].tolist()
print(feature_cols)

['poi_count', 'atm_count', 'atm_nearest', 'bank_count', 'bank_nearest', 'bus_station_count', 'bus_station_nearest', 'cafe_count', 'cafe_nearest', 'charging_station_count', 'charging_station_nearest', 'courthouse_count', 'courthouse_nearest', 'dentist_count', 'dentist_nearest', 'fast_food_count', 'fast_food_nearest', 'fire_station_count', 'fire_station_nearest', 'food_court_count', 'food_court_nearest', 'fuel_count', 'fuel_nearest', 'hospital_count', 'hospital_nearest', 'library_count', 'library_nearest', 'marketplace_count', 'marketplace_nearest', 'pharmacy_count', 'pharmacy_nearest', 'police_count', 'police_nearest', 'post_box_count', 'post_box_nearest', 'post_office_count', 'post_office_nearest', 'restaurant_count', 'restaurant_nearest', 'social_facility_count', 'social_facility_nearest', 'supermarket_count', 'supermarket_nearest', 'townhall_count', 'townhall_nearest', 'road_count', 'fixed_2019_mean_avg_d_kbps_mean', 'fixed_2019_mean_avg_u_kbps_mean', 'fixed_2019_mean_avg_lat_ms_mean', 'fixed_2019_mean_num_tests_mean', 'fixed_2019_mean_num_devices_mean', 'mobile_2019_mean_avg_d_kbps_mean', 'mobile_2019_mean_avg_u_kbps_mean', 'mobile_2019_mean_avg_lat_ms_mean', 'mobile_2019_mean_num_tests_mean', 'mobile_2019_mean_num_devices_mean', 'avg_rad_min', 'avg_rad_max', 'avg_rad_mean', 'avg_rad_std', 'avg_rad_median']

#scale features
scaler = StandardScaler()
for col in feature_cols:
    print(col)
    features[col+'_scaled'] = features.groupby(group_col, group_keys=False)[[col]].apply(scale_col, scaler=scaler)
print(features.columns)

poi_count
atm_count
atm_nearest
bank_count
bank_nearest
bus_station_count
bus_station_nearest
cafe_count
cafe_nearest
charging_station_count
charging_station_nearest
courthouse_count
courthouse_nearest
dentist_count
dentist_nearest
fast_food_count
fast_food_nearest
fire_station_count
fire_station_nearest
food_court_count
food_court_nearest
fuel_count
fuel_nearest
hospital_count
hospital_nearest
library_count
library_nearest
marketplace_count
marketplace_nearest
pharmacy_count
pharmacy_nearest
police_count
police_nearest
post_box_count
post_box_nearest
post_office_count
post_office_nearest
restaurant_count
restaurant_nearest
social_facility_count
social_facility_nearest
supermarket_count
supermarket_nearest
townhall_count
townhall_nearest
road_count
fixed_2019_mean_avg_d_kbps_mean
fixed_2019_mean_avg_u_kbps_mean
fixed_2019_mean_avg_lat_ms_mean
fixed_2019_mean_num_tests_mean
fixed_2019_mean_num_devices_mean
mobile_2019_mean_avg_d_kbps_mean
mobile_2019_mean_avg_u_kbps_mean
mobile_2019_mean_avg_lat_ms_mean
mobile_2019_mean_num_tests_mean
mobile_2019_mean_num_devices_mean
avg_rad_min
avg_rad_max
avg_rad_mean
avg_rad_std
avg_rad_median
Index(['DHSCLUST', 'Wealth Index', 'DHSCC', 'geometry', 'poi_count',
       'atm_count', 'atm_nearest', 'bank_count', 'bank_nearest',
       'bus_station_count',
       ...
       'mobile_2019_mean_avg_d_kbps_mean_scaled',
       'mobile_2019_mean_avg_u_kbps_mean_scaled',
       'mobile_2019_mean_avg_lat_ms_mean_scaled',
       'mobile_2019_mean_num_tests_mean_scaled',
       'mobile_2019_mean_num_devices_mean_scaled', 'avg_rad_min_scaled',
       'avg_rad_max_scaled', 'avg_rad_mean_scaled', 'avg_rad_std_scaled',
       'avg_rad_median_scaled'],
      dtype='object', length=126)

# fill nas with 0
features = features.fillna(0)

# # scale labels
scaler = StandardScaler()
labels = countries_data[[group_col,label_col]].groupby(group_col, group_keys=False)[[label_col]].apply(scale_col, scaler=scaler)
plt.hist(labels)

(array([ 16.,  85., 442., 702., 533., 379., 308., 179.,  68.,   8.]),
 array([-2.65317484, -2.0616477 , -1.47012056, -0.87859342, -0.28706627,
         0.30446087,  0.89598801,  1.48751515,  2.0790423 ,  2.67056944,
         3.26209658]),
 <BarContainer object of 10 artists>)

# keep only scaled columns
features = features[[col+'_scaled' for col in feature_cols]]

Base Features List

The features can be subdivided by the source dataset

OSM

• <poi type>_count: number of points of interest (POI) of a specified type in that area
  • ex. atm_count: number of atms in cluster
  • poi_count: number of all POIs of all types in cluster
• <poi_type>_nearest: distance of nearest POI of the specified type
  • ex. atm_nearest: distance of nearest ATM from that cluster
• OSM POI types included: atm, bank, bus_stations, cafe, charging_station, courthouse, dentist (clinic), fast_food, fire_station, food_court, fuel (gas station), hospital, library, marketplace, pharmacy, police, post_box, post_office, restaurant, social_facility, supermarket, townhall, road

Ookla

The network metrics features follow the following name convention:

<type>_<year>_<yearly aggregate>_<network variable>_<cluster aggregate>

• type: kind of network connection measured
  • fixed: connection from fixed sources (landline, fiber, etc.)
  • mobile: connection from mobile devices
• year: Year of source data
• yearly aggregate: How data was aggregated into yearly data
  • Note: Ookla provides data per quarter, so a yearly mean takes the average across 4 quarters
  • For this model, we only aggregate by yearly mean
• network variable: network characteristic described
  • avg_d_kbps: average download speed in kbps
  • avg_u_kbps: average upload speed in kbps
  • avg_lat_ms: average latency in ms
  • num_devices: number of devices measured
• cluster aggregate: how the data was aggregated per cluster aggregate
  • Types: min, mean, max, median, std.
    • For this model: only mean is used
  • This is calculated using area zonal stats, which weighs the average by the intersection of the Ookla tile with the cluster geometry.

Ex. fixed_2019_mean_avg_d_kbps_median takes the cluster median of 2019 yearly average download speed.

Nightlights (VIIRS)

All nightlights features are taken as the zonal aggregate of the raster data per cluster

• ex. avg_rad_mean: cluster mean of the average radiance
• aggregations used: min, mean, max, median

features.info()

<class 'pandas.core.frame.DataFrame'>
RangeIndex: 2720 entries, 0 to 2719
Data columns (total 61 columns):
 #   Column                                    Non-Null Count  Dtype  
---  ------                                    --------------  -----  
 0   poi_count_scaled                          2720 non-null   float64
 1   atm_count_scaled                          2720 non-null   float64
 2   atm_nearest_scaled                        2720 non-null   float64
 3   bank_count_scaled                         2720 non-null   float64
 4   bank_nearest_scaled                       2720 non-null   float64
 5   bus_station_count_scaled                  2720 non-null   float64
 6   bus_station_nearest_scaled                2720 non-null   float64
 7   cafe_count_scaled                         2720 non-null   float64
 8   cafe_nearest_scaled                       2720 non-null   float64
 9   charging_station_count_scaled             2720 non-null   float64
 10  charging_station_nearest_scaled           2720 non-null   float64
 11  courthouse_count_scaled                   2720 non-null   float64
 12  courthouse_nearest_scaled                 2720 non-null   float64
 13  dentist_count_scaled                      2720 non-null   float64
 14  dentist_nearest_scaled                    2720 non-null   float64
 15  fast_food_count_scaled                    2720 non-null   float64
 16  fast_food_nearest_scaled                  2720 non-null   float64
 17  fire_station_count_scaled                 2720 non-null   float64
 18  fire_station_nearest_scaled               2720 non-null   float64
 19  food_court_count_scaled                   2720 non-null   float64
 20  food_court_nearest_scaled                 2720 non-null   float64
 21  fuel_count_scaled                         2720 non-null   float64
 22  fuel_nearest_scaled                       2720 non-null   float64
 23  hospital_count_scaled                     2720 non-null   float64
 24  hospital_nearest_scaled                   2720 non-null   float64
 25  library_count_scaled                      2720 non-null   float64
 26  library_nearest_scaled                    2720 non-null   float64
 27  marketplace_count_scaled                  2720 non-null   float64
 28  marketplace_nearest_scaled                2720 non-null   float64
 29  pharmacy_count_scaled                     2720 non-null   float64
 30  pharmacy_nearest_scaled                   2720 non-null   float64
 31  police_count_scaled                       2720 non-null   float64
 32  police_nearest_scaled                     2720 non-null   float64
 33  post_box_count_scaled                     2720 non-null   float64
 34  post_box_nearest_scaled                   2720 non-null   float64
 35  post_office_count_scaled                  2720 non-null   float64
 36  post_office_nearest_scaled                2720 non-null   float64
 37  restaurant_count_scaled                   2720 non-null   float64
 38  restaurant_nearest_scaled                 2720 non-null   float64
 39  social_facility_count_scaled              2720 non-null   float64
 40  social_facility_nearest_scaled            2720 non-null   float64
 41  supermarket_count_scaled                  2720 non-null   float64
 42  supermarket_nearest_scaled                2720 non-null   float64
 43  townhall_count_scaled                     2720 non-null   float64
 44  townhall_nearest_scaled                   2720 non-null   float64
 45  road_count_scaled                         2720 non-null   float64
 46  fixed_2019_mean_avg_d_kbps_mean_scaled    2720 non-null   float64
 47  fixed_2019_mean_avg_u_kbps_mean_scaled    2720 non-null   float64
 48  fixed_2019_mean_avg_lat_ms_mean_scaled    2720 non-null   float64
 49  fixed_2019_mean_num_tests_mean_scaled     2720 non-null   float64
 50  fixed_2019_mean_num_devices_mean_scaled   2720 non-null   float64
 51  mobile_2019_mean_avg_d_kbps_mean_scaled   2720 non-null   float64
 52  mobile_2019_mean_avg_u_kbps_mean_scaled   2720 non-null   float64
 53  mobile_2019_mean_avg_lat_ms_mean_scaled   2720 non-null   float64
 54  mobile_2019_mean_num_tests_mean_scaled    2720 non-null   float64
 55  mobile_2019_mean_num_devices_mean_scaled  2720 non-null   float64
 56  avg_rad_min_scaled                        2720 non-null   float64
 57  avg_rad_max_scaled                        2720 non-null   float64
 58  avg_rad_mean_scaled                       2720 non-null   float64
 59  avg_rad_std_scaled                        2720 non-null   float64
 60  avg_rad_median_scaled                     2720 non-null   float64
dtypes: float64(61)
memory usage: 1.3 MB

Model Training

# Set parameters
groupkfold_col = 'DHSCC'
# cv_num_splits = 5
# cv_num_repeats = 5
train_test_seed = 42
test_size = 0.2

Create train/test cross-validation indices

# train_features, test_features, train_labels, test_labels = train_test_split(
#     features, labels, test_size=test_size, random_state=train_test_seed
# )

# Cross validation
print(f"Performing GroupKFold CV with groups based on DHSCC...")
groups = countries_data[groupkfold_col].values
cv = GroupKFold(n_splits = len(set(groups)))

print(cv.split(features, groups=groups))
print(f'Number of splits based on DHSCC unique values: {cv.get_n_splits()}')

Performing GroupKFold CV with groups based on DHSCC...
<generator object _BaseKFold.split at 0x7f6d50d022e0>
Number of splits based on DHSCC unique values: 4

Instantiate model

For now, we will train a simple random forest model

from sklearn.ensemble import RandomForestRegressor
model = RandomForestRegressor(n_estimators=100, random_state=train_test_seed, verbose=0)
model

RandomForestRegressor(random_state=42)

In a Jupyter environment, please rerun this cell to show the HTML representation or trust the notebook.
On GitHub, the HTML representation is unable to render, please try loading this page with nbviewer.org.

Evaluate model training using cross-validation

We evalute the model’s generalizability when training over different train/test splits

Ideally for R^2

• We want a high mean: This means that we achieve a high model performance over the different train/test splits
• We want a low standard deviation (std): This means that the model performance is stable over multiple training repetitions

split_r2_list = []

for i, (train, test) in enumerate(cv.split(features, labels,groups=groups)):
    print(f'Split# {i+1}')

    # Print info about current split
    train_labels = set(groups[train])
    test_labels = set(groups[test])
    print(f'Train countries (num samples): {train_labels} ({len(train)})')
    print(f'Test countries (num samples): {test_labels} ({len(test)})')

    # Split data into train/test
    X_train = features.values[train]
    X_test = features.values[test]
    y_train = labels.values.ravel()[train]
    y_test = labels.values.ravel()[test]

    # print('Input shapes (X_train, X_test, y_train, y_test): ',np.shape(X_train), np.shape(X_test), np.shape(y_train), np.shape(y_test))

    # Train model and get r2
    model.fit(X_train, y_train)
    split_r2 = r2_score(y_test, model.predict(X_test))
    split_r2_list.append(split_r2)
    print(f'Split r^2: {split_r2} \n')

    # Plot histogram of labels
    fig, ax = plt.subplots(2,1, sharex=True, figsize=(5,4))
    ax[0].hist(y_train)
    ax[0].set_title(f'train countries: {train_labels}')
    ax[1].hist(y_test)
    ax[1].set_title(f'test countries: {test_labels}')
    ax[1].set_xlabel('Wealth Index')
    fig.suptitle(f'Split# {i+1} Labels (model $r^2 = {round(split_r2,2)}$)')
    plt.tight_layout()
    plt.show()

    # Plot scatter plot
    fig, ax = plt.subplots(1,1, figsize=(5,4))
    ax.scatter(y_train, model.predict(X_train), label="train")
    ax.scatter(y_test, model.predict(X_test), label="test")
    ax.plot([0,1],[0,1], transform=ax.transAxes, linestyle='dashed', color='k')
    ax.set_xlabel("True Value")
    ax.set_ylabel("Predicted Value")
    ax.legend()    
    fig.suptitle(f'Split#{i+1} (test $r^2 = {round(split_r2,2)}$)')
    plt.tight_layout()
    plt.show()

split_r2_mean = round(np.array(split_r2_list).mean(), 4)
split_r2_std = round(np.array(split_r2_list).std(), 4)
print(f'Mean split r^2  (std): {split_r2_mean} ({split_r2_std})')

Split# 1
Train countries (num samples): {'KH', 'MM', 'TL'} (1507)
Test countries (num samples): {'PH'} (1213)
Split r^2: 0.48054801637734346 

Split# 2
Train countries (num samples): {'MM', 'TL', 'PH'} (2109)
Test countries (num samples): {'KH'} (611)
Split r^2: 0.5834437437650173 

Split# 3
Train countries (num samples): {'KH', 'MM', 'PH'} (2265)
Test countries (num samples): {'TL'} (455)
Split r^2: 0.5350537480791098 

Split# 4
Train countries (num samples): {'KH', 'TL', 'PH'} (2279)
Test countries (num samples): {'MM'} (441)
Split r^2: 0.5055486937348812 

Mean split r^2  (std): 0.5261 (0.0383)

# explainer = shap.TreeExplainer(model)
# shap_values = explainer.shap_values(features)

# shap.summary_plot(shap_values, features, feature_names=features.columns, plot_type="bar")