Calculate Distances Between One Point in Matrix From All Other Points

Question

I am new to Python and I need to implement a clustering algorithm. For that, I will need to calculate distances between the given input data.

Consider the following input data -

    [[1,2,8],
     [7,4,2],
     [9,1,7],
     [0,1,5],
     [6,4,3]]

What I am looking to achieve here is, I want to calculate distance of [1,2,8] from ALL other points, and find a point where the distance is minimum.

And I have to repeat this for ALL other points.

I am trying to implement this with a FOR loop, but I am sure that SciPy/ NumPy must be having a function which can help me achieve this result efficiently.

I looked online, but the 'pdist' command could not get my work done.

Can someone guide me?

TIA

Answer 1

Use np.linalg.norm combined with broadcasting (numpy outer subtraction), you can do:

np.linalg.norm(a - a[:,None], axis=-1)

a[:,None] insert a new axis into a, a - a[:,None] will then do a row by row subtraction due to broadcasting. np.linalg.norm calculates the np.sqrt(np.sum(np.square(...))) over the last axis:

---

a = np.array([[1,2,8],
     [7,4,2],
     [9,1,7],
     [0,1,5],
     [6,4,3]])

np.linalg.norm(a - a[:,None], axis=-1)
#array([[ 0.        ,  8.71779789,  8.1240384 ,  3.31662479,  7.34846923],
#       [ 8.71779789,  0.        ,  6.164414  ,  8.18535277,  1.41421356],
#       [ 8.1240384 ,  6.164414  ,  0.        ,  9.21954446,  5.83095189],
#       [ 3.31662479,  8.18535277,  9.21954446,  0.        ,  7.        ],
#       [ 7.34846923,  1.41421356,  5.83095189,  7.        ,  0.        ]])

The elements [0,1], [0,2] for instance correspond to:

np.sqrt(np.sum((a[0] - a[1]) ** 2))
# 8.717797887081348

np.sqrt(np.sum((a[0] - a[2]) ** 2))
# 8.1240384046359608

respectively.

Answer 2

Here's one approach using SciPy's cdist -

from scipy.spatial.distance import cdist
def closest_rows(a):
    # Get euclidean distances as 2D array
    dists = cdist(a, a, 'sqeuclidean')

    # Fill diagonals with something greater than all elements as we intend
    # to get argmin indices later on and then index into input array with those
    # indices to get the closest rows
    dists.ravel()[::dists.shape[1]+1] = dists.max()+1
    return a[dists.argmin(1)]

Sample run -

In [72]: a
Out[72]: 
array([[1, 2, 8],
       [7, 4, 2],
       [9, 1, 7],
       [0, 1, 5],
       [6, 4, 3]])

In [73]: closest_rows(a)
Out[73]: 
array([[0, 1, 5],
       [6, 4, 3],
       [6, 4, 3],
       [1, 2, 8],
       [7, 4, 2]])

Runtime test

Other working approach(es) -

def norm_app(a): # @Psidom's soln
    dist = np.linalg.norm(a - a[:,None], axis=-1); 
    dist[np.arange(dist.shape[0]), np.arange(dist.shape[0])] = np.nan
    return a[np.nanargmin(dist, axis=0)]

Timings with 10,000 points -

In [79]: a = np.random.randint(0,9,(10000,3))

In [80]: %timeit norm_app(a) # @Psidom's soln
1 loop, best of 3: 3.83 s per loop

In [81]: %timeit closest_rows(a)
1 loop, best of 3: 392 ms per loop

---

Further performance boost

There's eucl_dist package (disclaimer: I am its author) that contains various methods to compute euclidean distances that are much more efficient than SciPy's cdist, especially for large arrays.

Thus, making use of it, we would have a more performant one, like so -

from eucl_dist.cpu_dist import dist
def closest_rows_v2(a):
    dists = dist(a,a, matmul="gemm", method="ext") 
    dists.ravel()[::dists.shape[1]+1] = dists.max()+1
    return a[dists.argmin(1)]

Timings -

In [162]: a = np.random.randint(0,9,(10000,3))

In [163]: %timeit closest_rows(a)
1 loop, best of 3: 394 ms per loop

In [164]: %timeit closest_rows_v2(a)
1 loop, best of 3: 229 ms per loop