I have a long skinny data matrix (size: 250,000 x 10), which I will denote X. I also have a vector p measuring the quality of my data points. My goal is to compute the following function for each row x in my data matrix X:
r(x) = min{ ||x-y|| | p[y]>p[x], y in X }
On a smaller dataset, what I would use sklearn.metrics.pairwise_distances to precompute distances, like so:
from sklearn import metrics
n = len(X);
D_full = metrics.pairwise_distances(X);
r = np.zeros((n,1));
for i in range(n):
r[i] = (D_full[i,p>p[i]]).min();
However, the above approach is memory-expensive, since I need to store D_full: a full n x n matrix. It seems like sklearn.metrics.pairwise_distances_chunked could be a good tool for this sort of problem since the distance matrix is only stored one chunk at a time. I was hoping to get some assistance in how to use it though, as I'm currently unfamiliar with generator objects. Suppose I call the following:
from sklearn import metrics
D = metrics.pairwise_distances_chunked(X);
D_chunk = next(D)
The above code yields D (a generator object) and D_chunk (a 536 x n array). Does D_chunkcorrespond to the first 536 rows of the matrix D_full from my earlier approach? If so, does next(D_chunk) correspond to the next 536 rows?
Thank you very much for your help.
This is a outline of a possible solution, but details are missing. In short, I would do the following:
Create a BallTree to query, and initialise min_quality_distance of size 250000 with say zeros.
for k=2
For each vector, find the closest k neighbour (including itself).
If vector with most distance within k found, has sufficient quality, update min_quality_distance for that point.
For remaining, repeat with k=k+1
In each iteration, we have to query less vectors. The idea is that in each iteration you nibble a few nearest neighbors with the right condition away, and it will be easier with every step. (50% easier?) I will show how to do the first iteration, and with this is should be possible to build the loop.
You can do;
import numpy as np
size = 250000
X = np.random.random( size=(size,10))
p = np.random.random( size=size)
And create a BallTree with
from sklearn.neighbors import BallTree
tree = BallTree(X, leaf_size=10, metric='minkowski')
and query it for first iteration with (this will take about 5 minutes.)
k_nearest = 2
distances, indici = tree.query(X, k=k_nearest, return_distance=True, dualtree=False, sort_results=True)
The indici of most far away point within the nearest k is
most_far_away_indici = indici[:,-1:]
And its quality
p[most_far_away_indici]
So we can
quality_closeby = p[most_far_away_indici]
And check if it is sifficient with
indici_sufficient_quality = quality_closeby > np.expand_dims(p, axis=1)
And we have
found_closeby = np.all( indici_sufficient_quality, axis=1 )
Which is True is we have found a sufficient quality nearby.
We can update the vector with
distances_nearby = distances[:,-1:]
rx = np.zeros(size)
rx[found_closeby] = distances_nearby[found_closeby][:,0]
And we now need to take care for the remaining where we were unlucky, these are
~found_closeby
so
indici_not_found = ~found_closeby
and
distances, indici = tree.query(X[indici_not_found], k=3, return_distance=True, dualtree=False, sort_results=True)
etc..
I am sure the first few loops will take minutes, but after a few iterations the speeds will quickly go to seconds.
It is a little exercise with np.argwhere() etc to make sure the right indicis get updates.
It might not be the fastest, but it is a workable approach.
Since one cannot know the dimensions of some chunk, I suggest using np.ones_like instead of np.zeros.
We have boring CSV with 10000 rows of ages (float), titles (enum/int), scores (float), ....
We have N columns each with int/float values in a table.
You can imagine this as points in ND space
We want to pick K points that would have maximised distance between each other.
So if we have 100 points in a tightly packed cluster and one point in the distance we would get something like this for three points:
or this
For 4 points it will become more interesting and pick some point in the middle.
So how to select K most distant rows (points) from N (with any complexity)? It looks like an ND point cloud "triangulation" with a given resolution yet not for 3d points.
I search for a reasonably fast approach (approximate - no precise solution needed) for K=200 and N=100000 and ND=6 (probably multigrid or ANN on KDTree based, SOM or triangulation based..).. Does anyone know one?
From past experience with a pretty similar problem, a simple solution of computing the mean Euclidean distance of all pairs within each group of K points and then taking the largest mean, works very well. As someone noted above, it's probably hard to avoid a loop on all combinations (not on all pairs). So a possible implementation of all this can be as follows:
import itertools
import numpy as np
from scipy.spatial.distance import pdist
Npoints = 3 # or 4 or 5...
# making up some data:
data = np.matrix([[3,2,4,3,4],[23,25,30,21,27],[6,7,8,7,9],[5,5,6,6,7],[0,1,2,0,2],[3,9,1,6,5],[0,0,12,2,7]])
# finding row indices of all combinations:
c = [list(x) for x in itertools.combinations(range(len(data)), Npoints )]
distances = []
for i in c:
distances.append(np.mean(pdist(data[i,:]))) # pdist: a method of computing all pairwise Euclidean distances in a condensed way.
ind = distances.index(max(distances)) # finding the index of the max mean distance
rows = c[ind] # these are the points in question
I propose an approximate solution. The idea is to start from a set of K points chosen in a way I'll explain below, and repeatedly loop through these points replacing the current one with the point, among the N-K+1 points not belonging to the set but including the current one, that maximizes the sum of the distances from the points of the set. This procedure leads to a set of K points where the replacement of any single point would cause the sum of the distances among the points of the set to decrease.
To start the process we take the K points that are closest to the mean of all points. This way we have good chances that on the first loop the set of K points will be spread out close to its optimum. Subsequent iterations will make adjustments to the set of K points towards a maximum of the sum of distances, which for the current values of N, K and ND appears to be reachable in just a few seconds. In order to prevent excessive looping in edge cases, we limit the number of loops nonetheless.
We stop iterating when an iteration does not improve the total distance among the K points. Of course, this is a local maximum. Other local maxima will be reached for different initial conditions, or by allowing more than one replacement at a time, but I don't think it would be worthwhile.
The data must be adjusted in order for unit displacements in each dimension to have the same significance, i.e., in order for Euclidean distances to be meaningful. E.g., if your dimensions are salary and number of children, unadjusted, the algorithm will probably yield results concentrated in the extreme salary regions, ignoring that person with 10 kids. To get a more realistic output you could divide salary and number of children by their standard deviation, or by some other estimate that makes differences in salary comparable to differences in number of children.
To be able to plot the output for a random Gaussian distribution, I have set ND = 2 in the code, but setting ND = 6, as per your request, is no problem (except you cannot plot it).
import matplotlib.pyplot as plt
import numpy as np
import scipy.spatial as spatial
N, K, ND = 100000, 200, 2
MAX_LOOPS = 20
SIGMA, SEED = 40, 1234
rng = np.random.default_rng(seed=SEED)
means, variances = [0] * ND, [SIGMA**2] * ND
data = rng.multivariate_normal(means, np.diag(variances), N)
def distances(ndarray_0, ndarray_1):
if (ndarray_0.ndim, ndarray_1.ndim) not in ((1, 2), (2, 1)):
raise ValueError("bad ndarray dimensions combination")
return np.linalg.norm(ndarray_0 - ndarray_1, axis=1)
# start with the K points closest to the mean
# (the copy() is only to avoid a view into an otherwise unused array)
indices = np.argsort(distances(data, data.mean(0)))[:K].copy()
# distsums is, for all N points, the sum of the distances from the K points
distsums = spatial.distance.cdist(data, data[indices]).sum(1)
# but the K points themselves should not be considered
# (the trick is that -np.inf ± a finite quantity always yields -np.inf)
distsums[indices] = -np.inf
prev_sum = 0.0
for loop in range(MAX_LOOPS):
for i in range(K):
# remove this point from the K points
old_index = indices[i]
# calculate its sum of distances from the K points
distsums[old_index] = distances(data[indices], data[old_index]).sum()
# update the sums of distances of all points from the K-1 points
distsums -= distances(data, data[old_index])
# choose the point with the greatest sum of distances from the K-1 points
new_index = np.argmax(distsums)
# add it to the K points replacing the old_index
indices[i] = new_index
# don't consider it any more in distsums
distsums[new_index] = -np.inf
# update the sums of distances of all points from the K points
distsums += distances(data, data[new_index])
# sum all mutual distances of the K points
curr_sum = spatial.distance.pdist(data[indices]).sum()
# break if the sum hasn't changed
if curr_sum == prev_sum:
break
prev_sum = curr_sum
if ND == 2:
X, Y = data.T
marker_size = 4
plt.scatter(X, Y, s=marker_size)
plt.scatter(X[indices], Y[indices], s=marker_size)
plt.grid(True)
plt.gca().set_aspect('equal', adjustable='box')
plt.show()
Output:
Splitting the data into 3 equidistant Gaussian distributions the output is this:
Assuming that if you read your csv file with N (10000) rows and D dimension (or features) into a N*D martix X. You can calculate the distance between each point and store it in a distance matrix as follows:
import numpy as np
X = np.asarray(X) ### convert to numpy array
distance_matrix = np.zeros((X.shape[0],X.shape[0]))
for i in range(X.shape[0]):
for j in range(i+1,X.shape[0]):
## We compute triangle matrix and copy the rest. Distance from point A to point B and distance from point B to point A are the same.
distance_matrix[i][j]= np.linalg.norm(X[i]-X[j]) ## Here I am calculating Eucledian distance. Other distance measures can also be used.
#distance_matrix = distance_matrix + distance_matrix.T - np.diag(np.diag(distance_matrix)) ## This syntax can be used to get the lower triangle of distance matrix, which is not really required in your case.
K = 5 ## Number of points that you want to pick
indexes = np.unravel_index(np.argsort(distance_matrix.ravel())[-1*K:], distance_matrix.shape)
print(indexes)
Bottom Line Up Front: Dealing with multiple equally distant points and the Curse of Dimensionality are going to be larger problems than just finding the points. Spoiler alert: There's a surprise ending.
I think this an interesting question but I'm bewildered by some of the answers. I think this is, in part, due to the sketches provided. You've no doubt noticed the answers look similar -- 2d, with clusters -- even though you indicated a wider scope was needed. Because others will eventually see this, I'm going to step through my thinking a bit slowly so bear with me for the early part.
It makes sense to start with a simplified example to see if we can generalize a solution with data that's easy to grasp and a linear 2D model is easiest of the easy.
We don't need to calculate all the distances though. We just need the ones at the extremes. So we can then take the top and bottom few values:
right = lin_2_D.nlargest(8, ['x'])
left = lin_2_D.nsmallest(8, ['x'])
graph = sns.scatterplot(x="x", y="y", data=lin_2_D, color = 'gray', marker = '+', alpha = .4)
sns.scatterplot(x = right['x'], y = right['y'], color = 'red')
sns.scatterplot(x = left['x'], y = left['y'], color = 'green')
fig = graph.figure
fig.set_size_inches(8,3)
What we have so far: Of 100 points, we've eliminated the need to calculate the distance between 84 of them. Of what's left we can further drop this by ordering the results on one side and checking the distance against the others.
You can imagine a case where you have a couple of data points way off the trend line that could be captured by taking the greatest or least y values, and all that starts to look like Walter Tross's top diagram. Add in a couple of extra clusters and you get what looks his bottom diagram and it appears that we're sort of making the same point.
The problem with stopping here is the requirement you mentioned is that you need a solution that works for any number of dimensions.
The unfortunate part is that we run into four challenges:
Challenge 1: As you increase the dimensions you can run into a large number of cases where you have multiple solutions when seeking midpoints. So you're looking for k furthest points but have a large number of equally valid possible solutions and no way prioritizing them. Here are two super easy examples illustrate this:
A) Here we have just four points and in only two dimensions. You really can't get any easier than this, right? The distance from red to green is trivial. But try to find the next furthest point and you'll see both of the black points are equidistant from both the red and green points. Imagine you wanted the furthest six points using the first graphs, you might have 20 or more points that are all equidistant.
edit: I just noticed the red and green dots are at the edges of their circles rather than at the center, I'll update later but the point is the same.
B) This is super easy to imagine: Think of a D&D 4 sided die. Four points of data in a three-dimensional space, all equidistant so it's known as a triangle-based pyramid. If you're looking for the closest two points, which two? You have 4 choose 2 (aka, 6) combinations possible. Getting rid of valid solutions can be a bit of a problem because invariably you face questions such as "why did we get rid of these and not this one?"
Challenge 2: The Curse of Dimensionality. Nuff Said.
Challenge 3 Revenge of The Curse of Dimensionality Because you're looking for the most distant points, you have to x,y,z ... n coordinates for each point or you have to impute them. Now, your data set is much larger and slower.
Challenge 4 Because you're looking for the most distant points, dimension reduction techniques such as ridge and lasso are not going to be useful.
So, what to do about this?
Nothing.
Wait. What?!?
Not truly, exactly, and literally nothing. But nothing crazy. Instead, rely on a simple heuristic that is understandable and computationally easy. Paul C. Kainen puts it well:
Intuitively, when a situation is sufficiently complex or uncertain,
only the simplest methods are valid. Surprisingly, however,
common-sense heuristics based on these robustly applicable techniques
can yield results which are almost surely optimal.
In this case, you have not the Curse of Dimensionality but rather the Blessing of Dimensionality. It's true you have a lot of points and they'll scale linearly as you seek other equidistant points (k) but the total dimensional volume of space will increase to power of the dimensions. The k number of furthest points you're is insignificant to the total number of points. Hell, even k^2 becomes insignificant as the number of dimensions increase.
Now, if you had a low dimensionality, I would go with them as a solution (except the ones that are use nested for loops ... in NumPy or Pandas).
If I was in your position, I'd be thinking how I've got code in these other answers that I could use as a basis and maybe wonder why should I should trust this other than it lays out a framework on how to think through the topic. Certainly, there should be some math and maybe somebody important saying the same thing.
Let me reference to chapter 18 of Computer Intensive Methods in Control and Signal Processing and an expanded argument by analogy with some heavy(-ish) math. You can see from the above (the graph with the colored dots at the edges) that the center is removed, particularly if you followed the idea of removing the extreme y values. It's a though you put a balloon in a box. You could do this a sphere in a cube too. Raise that into multiple dimensions and you have a hypersphere in a hypercube. You can read more about that relationship here.
Finally, let's get to a heuristic:
Select the points that have the most max or min values per dimension. When/if you run out of them pick ones that are close to those values if there isn't one at the min/max. Essentially, you're choosing the corners of a box For a 2D graph you have four points, for a 3D you have the 8 corners of the box (2^3).
More accurately this would be a 4d or 5d (depending on how you might assign the marker shape and color) projected down to 3d. But you can easily see how this data cloud gives you the full range of dimensions.
Here is a quick check on learning; for purposes of ease, ignore the color/shape aspect: It's easy to graphically intuit that you have no problem with up to k points short of deciding what might be slightly closer. And you can see how you might need to randomize your selection if you have a k < 2D. And if you added another point you can see it (k +1) would be in a centroid. So here is the check: If you had more points, where would they be? I guess I have to put this at the bottom -- limitation of markdown.
So for a 6D data cloud, the values of k less than 64 (really 65 as we'll see in just a moment) points are pretty easy. But...
If you don't have a data cloud but instead have data that has a linear relationship, you'll 2^(D-1) points. So, for that linear 2D space, you have a line, for linear 3D space, you'd have a plane. Then a rhomboid, etc. This is true even if your shape is curved. Rather than do this graph myself, I'm using the one from an excellent post on by Inversion Labs on Best-fit Surfaces for 3D Data
If the number of points, k, is less than 2^D you need a process to decide what you don't use. Linear discriminant analysis should be on your shortlist. That said, you can probably satisfice the solution by randomly picking one.
For a single additional point (k = 1 + 2^D), you're looking for one that is as close to the center of the bounding space.
When k > 2^D, the possible solutions will scale not geometrically but factorially. That may not seem intuitive so let's go back to the two circles. For 2D you have just two points that could be a candidate for being equidistant. But if that were 3D space and rotate the points about the line, any point in what is now a ring would suffice as a solution for k. For a 3D example, they would be a sphere. Hyperspheres (n-spheres) from thereon. Again, 2^D scaling.
One last thing: You should seriously look at xarray if you're not already familiar with it.
Hope all this helps and I also hope you'll read through the links. It'll be worth the time.
*It would be the same shape, centrally located, with the vertices at the 1/3 mark. So like having 27 six-sided dice shaped like a giant cube. Each vertice (or point nearest it) would fix the solution. Your original k+1 would have to be relocated too. So you would select 2 of the 8 vertices. Final question: would it be worth calculating the distances of those points against each other (remember the diagonal is slightly longer than the edge) and then comparing them to the original 2^D points? Bluntly, no. Satifice the solution.
If you're interested in getting the most distant points you can take advantage of all of the methods that were developed for nearest neighbors, you just have to give a different "metric".
For example, using scikit-learn's nearest neighbors and distance metrics tools you can do something like this
import numpy as np
from sklearn.neighbors import BallTree
from sklearn.neighbors.dist_metrics import PyFuncDistance
from sklearn.datasets import make_blobs
from matplotlib import pyplot as plt
def inverted_euclidean(x1, x2):
# You can speed this up using cython like scikit-learn does or numba
dist = np.sum((x1 - x2) ** 2)
# We invert the euclidean distance and set nearby points to the biggest possible
# positive float that isn't inf
inverted_dist = np.where(dist == 0, np.nextafter(np.inf, 0), 1 / dist)
return inverted_dist
# Make up some fake data
n_samples = 100000
n_features = 200
X, _ = make_blobs(n_samples=n_samples, centers=3, n_features=n_features, random_state=0)
# We exploit the BallTree algorithm to get the most distant points
ball_tree = BallTree(X, leaf_size=50, metric=PyFuncDistance(inverted_euclidean))
# Some made up query, you can also provide a stack of points to query against
test_point = np.zeros((1, n_features))
distance, distant_points_inds = ball_tree.query(X=test_point, k=10, return_distance=True)
distant_points = X[distant_points_inds[0]]
# We can try to visualize the query results
plt.plot(X[:, 0], X[:, 1], ".b", alpha=0.1)
plt.plot(test_point[:, 0], test_point[:, 1], "*r", markersize=9)
plt.plot(distant_points[:, 0], distant_points[:, 1], "sg", markersize=5, alpha=0.8)
plt.show()
Which will plot something like:
There are many points that you can improve on:
I implemented the inverted_euclidean distance function with numpy, but you can try to do what the folks of scikit-learn do with their distance functions and implement them in cython. You could also try to jit compile them with numba.
Maybe the euclidean distance isn't the metric you would like to use to find the furthest points, so you're free to implement your own or simply roll with what scikit-learn provides.
The nice thing about using the Ball Tree algorithm (or the KdTree algorithm) is that for each queried point you have to do log(N) comparisons to find the furthest point in the training set. Building the Ball Tree itself, I think also requires log(N) comparison, so in the end if you want to find the k furthest points for every point in the ball tree training set (X), it will have almost O(D N log(N)) complexity (where D is the number of features), which will increase up to O(D N^2) with the increasing k.
I want to get the k-means cost (inertia in scikit kmeans).
Just to remind:
The cost is the sum of squared distanctes from each point to the nearest cluster.
I get a strange difference between the cost calc of scikit('inertia'),
and my own trivial way for computing the cost
Please see the following example:
p = np.random.rand(1000000,2)
from sklearn.cluster import KMeans
a = KMeans(n_clusters=3).fit(p)
print a.inertia_ , "****"
means = a.cluster_centers_
s = 0
for x in p:
best = float("inf")
for y in means:
if np.linalg.norm(x-y)**2 < best:
best = np.linalg.norm(x-y)**2
s += best
print s, "*****"
Where for my run the output is:
66178.4232156 ****
66173.7928716 *****
Where on my own dataset, the result is more significant(20% difference).
Is this a bug in scikit's implementation?
First - it does not seem to be a bug (but for sure ugly inconsistency). Why is that? You need to take a closer look into what is the code actually doing. For this general purpose it calls the cython code from _k_means.pyx
(lines 577-578)
inertia = _k_means._assign_labels_array(
X, x_squared_norms, centers, labels, distances=distances)
and what it does is essentialy exactly your code, but... using doubles in C. So maybe it is just a numerical issue? Let us test your code but now, with clear clusters structure (thus there are no points which might be assigned to many centers - depending on numerical accuracy).
import numpy as np
from sklearn.metrics import euclidean_distances
p = np.random.rand(1000000,2)
p[:p.shape[0]/2, :] += 100 #I move half of points far away
from sklearn.cluster import KMeans
a = KMeans(n_clusters=2).fit(p) #changed to two clusters
print a.inertia_ , "****"
means = a.cluster_centers_
s = 0
for x in p:
best = float("inf")
for y in means:
d = (x-y).T.dot(x-y)
if d < best:
best = d
s += best
print s, "*****"
results
166805.190832 ****
166805.190946 *****
makes sense. Thus the problem is with existance of samples "near the boundary" which might be assigned to multiple clusters depending on arithmetic accuracy. Unftunately I was not able to trace exactly where the difference comes from.
The funny thing is that there is actually an inconsistency coming from the fact, that inertia_ field is filled with Cython code, and .score calls NumPy one. Thus if you call
print -a.score(p)
you will get exactly your inertia.
I have experimental data of the form (X,Y) and a theoretical model of the form (x(t;*params),y(t;*params)) where t is a physical (but unobservable) variable, and *params are the parameters that I want to determine. t is a continuous variable, and there is a 1:1 relationship between x and t and between y and t in the model.
In a perfect world, I would know the value of T (the real-world value of the parameter) and would be able to do an extremely basic least-squares fit to find the values of *params. (Note that I am not trying to "connect" the values of x and y in my plot, like in 31243002 or 31464345.) I cannot guarantee that in my real data, the latent value T is monotonic, as my data is collected across multiple cycles.
I'm not very experienced doing curve fitting manually, and have to use extremely crude methods without easy access to a basic scipy function. My basic approach involves:
Choose some value of *params and apply it to the model
Take an array of t values and put it into the model to create an array of model(*params) = (x(*params),y(*params))
Interpolate X (the data values) into model to get Y_predicted
Run a least-squares (or other) comparison between Y and Y_predicted
Do it again for a new set of *params
Eventually, choose the best values for *params
There are several obvious problems with this approach.
1) I'm not experienced enough with coding to develop a very good "do it again" other than "try everything in the solution space," of maybe "try everything in a coarse grid" and then "try everything again in a slightly finer grid in the hotspots of the coarse grid." I tried doing MCMC methods, but I never found any optimum values, largely because of problem 2
2) Steps 2-4 are super inefficient in their own right.
I've tried something like (resembling pseudo-code; the actual functions are made up). There are many minor quibbles that could be made about using broadcasting on A,B, but those are less significant than the problem of needing to interpolate for every single step.
People I know have recommended using some sort of Expectation Maximization algorithm, but I don't know enough about that to code one up from scratch. I'm really hoping there's some awesome scipy (or otherwise open-source) algorithm I haven't been able to find that covers my whole problem, but at this point I am not hopeful.
import numpy as np
import scipy as sci
from scipy import interpolate
X_data
Y_data
def x(t,A,B):
return A**t + B**t
def y(t,A,B):
return A*t + B
def interp(A,B):
ts = np.arange(-10,10,0.1)
xs = x(ts,A,B)
ys = y(ts,A,B)
f = interpolate.interp1d(xs,ys)
return f
N = 101
lsqs = np.recarray((N**2),dtype=float)
count = 0
for i in range(0,N):
A = 0.1*i #checks A between 0 and 10
for j in range(0,N):
B = 10 + 0.1*j #checks B between 10 and 20
f = interp(A,B)
y_fit = f(X_data)
squares = np.sum((y_fit - Y_data)**2)
lsqs[count] = (A,b,squares) #puts the values in place for comparison later
count += 1 #allows us to move to the next cell
i = np.argmin(lsqs[:,2])
A_optimal = lsqs[i][0]
B_optimal = lsqs[i][1]
If I understand the question correctly, the params are constants which are the same in every sample, but t varies from sample to sample. So, for example, maybe you have a whole bunch of points which you believe have been sampled from a circle
x = a+r cos(t)
y = b+r sin(t)
at different values of t.
In this case, what I would do is eliminate the variable t to get a relation between x and y -- in this case, (x-a)^2+(y-b)^2 = r^2. If your data fit the model perfectly, you would have (x-a)^2+(y-b)^2 = r^2 at each of your data points. With some error, you could still find (a,b,r) to minimize
sum_i ((x_i-a)^2 + (y_i-b)^2 - r^2)^2.
Mathematica's Eliminate command can automate the procedure of eliminating t in some cases.
PS You might do better at stats.stackexchange, math.stackexchange or mathoverflow.net . I know the last one has a scary reputation, but we don't bite, really!