Python code to filter closest distance pairs

Question

This is my code. Please note that this is just a toy dataset, my real set contains about a 1000 entries in each table.

import pandas as pd
import numpy as np
import sklearn.neighbors

locations_stores = pd.DataFrame({
    'city_A' :     ['City1', 'City2', 'City3', 'City4', ],
    'latitude_A':  [ 56.361176, 56.34061, 56.374749, 56.356624],
    'longitude_A': [ 4.899779, 4.871195, 4.893847, 4.912281]
})
locations_neigh = pd.DataFrame({
    'neigh_B':      ['Neigh1', 'Neigh2', 'Neigh3', 'Neigh4','Neigh5'],
    'latitude_B' : [ 53.314, 53.318, 53.381, 53.338,53.7364],
    'longitude_B': [ 4.955,4.975,4.855,4.873,4.425]
})

/some calc code here/

##df_dist_long.loc[df_dist_long.sort_values('Dist(km)').groupby('neigh_B')['city_A'].min()]##


df_dist_long.to_csv('dist.csv',float_format='%.2f')

When i add df_dist_long.loc[df_dist_long.sort_values('Dist(km)').groupby('neigh_B')['city_A'].min()] . I get this error

 File "C:\Python\Python38\lib\site-packages\pandas\core\groupby\groupby.py", line 656, in wrapper                                                    
    raise ValueError                                                                                                                                  
ValueError    

                                                                        
                                                           

Without it, the output is like so...

    city_A  neigh_B Dist(km)
0   City1   Neigh1  6.45
1   City2   Neigh1  6.42
2   City3   Neigh1  7.93
3   City4   Neigh1  5.56
4   City1   Neigh2  8.25
5   City2   Neigh2  6.67
6   City3   Neigh2  8.55
7   City4   Neigh2  8.92
8   City1   Neigh3  7.01   ..... and so on

What I want is another table that filters the city closest to the Neighbour. So as an example, for 'Neigh1', City4 is the closest(least in distance). So I want the table as below

city_A  neigh_B Dist(km)
0   City4   Neigh1  5.56
1   City3   Neigh2  4.32
2   City1   Neigh3  7.93
3   City2   Neigh4  3.21
4   City4   Neigh5  4.56
5   City5   Neigh6  6.67
6   City3   Neigh7  6.16
 ..... and so on

Doesn't matter if the city name gets repeated, I just want the closest pair saved to another csv. How can this be implemented, experts, please help!!

Answer 1

You don't want to calculate the full distance matrix if you just want the closest city for each neighbourhood.

Here is a working code example, though I get different output than yours. Maybe a lat/long mistake.

I used your data

import pandas as pd
import numpy as np
import sklearn.neighbors

locations_stores = pd.DataFrame({
    'city_A' :     ['City1', 'City2', 'City3', 'City4', ],
    'latitude_A':  [ 56.361176, 56.34061, 56.374749, 56.356624],
    'longitude_A': [ 4.899779, 4.871195, 4.893847, 4.912281]
})
locations_neigh = pd.DataFrame({
    'neigh_B':      ['Neigh1', 'Neigh2', 'Neigh3', 'Neigh4','Neigh5'],
    'latitude_B' : [ 53.314, 53.318, 53.381, 53.338,53.7364],
    'longitude_B': [ 4.955,4.975,4.855,4.873,4.425]
})

Created a BallTree we can querie

from sklearn.neighbors import BallTree
import numpy as np

stores_gps = locations_stores[['latitude_A', 'longitude_A']].values
neigh_gps = locations_neigh[['latitude_B', 'longitude_B']].values

tree = BallTree(stores_gps, leaf_size=15, metric='haversine')

And for each Neigh we want to closest ( k=1 ) City/Store:

distance, index = tree.query(neigh_gps, k=1)
 
earth_radius = 6371

distance_in_km = distance * earth_radius

We can create a DataFrame of the result with

pd.DataFrame({
    'Neighborhood' : locations_neigh.neigh_B,
    'Closest_city' : locations_stores.city_A[ np.array(index)[:,0] ].values,
    'Distance_to_city' : distance_in_km[:,0]
})

This gave me

  Neighborhood Closest_city  Distance_to_city
0       Neigh1        City2      19112.334106
1       Neigh2        City2      19014.154744
2       Neigh3        City2      18851.168702
3       Neigh4        City2      19129.555188
4       Neigh5        City4      15498.181486

Since our output is different, there is some mistake to correct. Maybe swapped lat/long, I am just guessing here. But this is the approach you want, especially for the amounts of your data.

---

Edit: For the Full matrix, use

from sklearn.neighbors import DistanceMetric

dist = DistanceMetric.get_metric('haversine')

earth_radius = 6371

haversine_distances = dist.pairwise(np.radians(stores_gps), np.radians(neigh_gps) )
haversine_distances *= earth_radius

This will give the full matrix, but be aware, for largers numbers it will take long, and expect hit memory limitation.

You could use numpy's np.argmin(haversine_distances, axis=1) to get similar results from the BallTree. It will result in the index of the closest in distance, which can be used just like in the BallTree example.