This repository contains Python3 code for the mean shift and subspace constrained mean shift (SCMS) algorithms in any Euclidean and/or directional (Cartesian) product space.
- Paper Reference: Mode and Ridge Estimation in Euclidean and Directional Product Spaces: A Mean Shift Approach (2021)
- Python >= 3.8 (earlier version might be applicable).
- NumPy, Matplotlib (especially the Basemap toolkit), pandas, SciPy (The speical function scipy.special.iv computes the modified Bessel function of the first kind of real order; scipy.linalg.block_diag creates a block diagonal matrix from provided arrays), pickle, datetime and time libraries.
- astropy ("Astropy" is a Python package for analyzing data among the astronomical community.
- Ray ("Ray" is a fast and simple distributed computing API for Python and Java. We use "ray[default]==1.4.0" because the lastest versions (>=1.6.0) cannot be run on our Ubuntu 16.04 server.)
- We provide an guideline of installing the Basemap toolkit on Ubuntu.
Some high-level descriptions of our Python scripts are as follows:
- Cosmic Filament Detection with Directional-Linear SCMS Algorithm (An Example).ipynb: This Jupyter Notebook contains detailed code and descriptions about how we process the SDSS-IV galaxy data (Ahumada et al., 2020) and detect cosmic filaments on them with our proposed SCMS algorithm in the directional-linear [(RA,DEC)*Redshift] product space.
- Curves_Sphere_Torus.py: This script simulates a circular-circular dataset and plot its points on a unit sphere and torus, respectively. (Figure 1 in the arxiv version of the paper).
- DirLinProdSCMS_Ray.py: This script implements the functions of KDE, component-wise/simultaneous mean shift, and subspace constrained mean shift (SCMS) algorithms with the Gaussian/von Mises product kernels in a directional/linear (mixture) product space using the parallel programming under the "Ray" environment.
- DirLinProdSCMS_fun.py: This script implements the functions of KDE, component-wise/simultaneous mean shift, and subspace constrained mean shift (SCMS) algorithms with the Gaussian/von Mises product kernels in a directional/linear (mixture) product space.
- Earthquake_Modes.py: This script contains code for applying our proposed mean shift algorithm to an Earthquake dataset (directional-linear data) (Figure 5 in the arxiv version of the paper). This script take more than 35 minutes to run on my laptop with 8 CPU cores.
- MS_SCMS_Ray.py: This script contains code for the parallel implementations of regular Euclidean/directional mean shift and SCMS algorithms.
- Mode_Seeking_Examples.py: This script contains code for mode-seeking simulation studies with our proposed mean shift algorithm (Figure 3 in the arxiv version of the paper).
- Spherical_Cone.py: This script contains code for comparing the results of the regular SCMS and our proposed SCMS algorithms on the simulated spherical cone data (Figure 4 in the arxiv version of the paper).
- Spiral_Curve.py: This script contains code for comparing the results of the regular SCMS and our proposed SCMS algorithms on the simulated spiral curve data. (Figure 2 in the arxiv version of the paper).
- Utility_fun.py: This script contains all the utility functions for our experiments.
- Varying_Stepsize.py: This script contains code for investigating the effects of varying the stepsize parameter in our proposed SCMS algorithm in Euclidean/directional product spaces. (Figures 9 and 10 in the arxiv version of the paper). The script takes more than 1.5 hours to execute due to the slow convergence of the proposed SCMS algorithm with step size "eta=1". The SCMS algorithm with our suggested choice of the step size parameter, however, does converge very fast.
It is intuitive that the (Cartesian) product of two Euclidean spaces is again an Euclidean space whose dimension is the sum of the dimensions of two factor (Euclidean) spaces. As the topology of such product space does not mathematically change, the regular kernel density estimator (KDE) as well as mean shift and SCMS algorithms are applicable in the Euclidean-Euclidean product space (Cheng, 1995; Comaniciu and Meer, 2002; Ozertem and Erdogmus, 2011). The Euclidean-directional and directional-directional product spaces, however, are not topologically equivalent to any of its factor spaces under any dimension. Consider, for example, a dataset with
and
being periodic. Under some renormalizations, every such circular-circular (or periodic-periodic) observation can be viewed as a point on the sphere
, where
represents the longitude and latitude, or a point on the torus
. Here,
,
where is the usual Euclidean norm in
. The supports
and
of the same dataset are topological different; see Figure 1 below. Therefore, it is worthwhile to reconsider (subspace constrained) mean shift algorithms as well as the related mode and ridge estimation problems in Euclidean/directional product spaces.
Fig 1. Simulated dataset on
and
. Each observation
is sampled uniformly from
for some integers
.
Besides the aforementioned circular-circular data, there are many real-world datasets whose observations lie on a Euclidean/directional product space. For instance, in astronomical survey data, each object has its right ascension (RA) and declination (DEC) on a celestial sphere, while its redshift measures its distance to the Earth. The collection of (RA,DEC,Redshift) tuples thus forms a directional-linear dataset.
2. Mode and Ridge Estimation on Euclidean/directional product spaces with (Subspace Constrained) Mean Shift Algorithms
Our interested data consist of independent and identically distributed (i.i.d.) observations sampled from a distribution on
, where
for
. While we only present the formulations of our proposed algorithms and related theory on (Cartesian) product spaces with two factors, our implementations (i.e., associated functions in DirLinProdSCMS_fun.py and DirLinProdSCMS_Ray.py) are adaptive to any product space with arbitrarily finte number of Euclidean/directional factor spaces.
It is natural to leverage a product kernel to construct a kernel density estimator (KDE) on as:
,
where each element of is a bandwidth parameter and the kernel functions
for
take the form as:
with and
being the profiles of linear and directional kernels, respectively. Under the Gaussian and/or von Mises kernels, i.e.,
, the KDE reduces to the following concise form as:
,
where ,
is a (block) diagonal bandwidth matrix,
is the identity matrix in
, and
is the normalizing constant.
By taking the total gradient of KDE and equating each of its components to 0, we derive two different versions of the mean shift algorithm on ; see more details in our paper.
- Version A (Simultaneous Mean Shift). This version updates all the components
simultaneously as:
for , where we require extra standardizations
and/or
if
and/or
.
- Version B (Componentwise Mean Shift). This version updates the sequence
in a two-step manner as:
with an additional standardization
if
and
with an additional standardization
if
for . The formula updates the two components
and
alternatively by first holding
, updating
, and then switching their roles. Such updating procedures borrows the spirit of the well-known coordinate ascent/descent algorithm (Wright, 2015).
Naively, one may adopt the standard SCMS iterative formula in Ozertem and Erdogmus (2011) and update the SCMS sequence as:
,
where has its columns as orthonormal eigenvectors of the (estimated) Riemannian Hessian
associated with the smallest
eigenvalues with the tangent space of
at
. This naive SCMS procedure, however, does not converge to our interested ridges of KDE
. What's worse, the incorrect ridges estimated by this naive SCMS procedure is also asymptotically invalid in estimating the ridges of the data-generating distribution.
Under the Gaussian and/or von Mises kernels, we formulate a valid SCMS iterative formula by rescaling each component of the mean shift vector with the bandwidth matrix as:
,
where is the step size parameter managing the learning rate and convergence performance of our proposed SCMS algorithm. As a guideline, we suggest taking the step size to be adaptive to bandwidth parameters as:
so that when are small,
mimics the asymptotic rate
of adaptive step sizes in Euclidean/directional (subspace constrained) mean shift algorithms (Cheng, 1995; Arias-Castro et al., 2016; Zhang and Chen, 2021). The upper bound 1 is introduced to prevent
from being too large.
The implementation of KDE in any Euclidean/directional product space is through the Python function called DirLinProdKDE in the script DirLinProdSCMS_fun.py.
Further, the implementations of simultaneous and componentwise mean shift algorithms are encapsulated into two Python functions called DirLinProdMS and DirLinProdMSCompAsc in the script DirLinProdSCMS_fun.py, respectively. The input arguments of DirLinProdMS and DirLinProdMSCompAsc are the same, and we notice that their outputs are identical, though the simultaneous version seems to be faster in the convergence speed. Finally, we implement our proposed SCMS algorithm in any Euclidean/directional product space on the Python functions DirLinProdSCMS and DirLinProdSCMSLog under log-density in the same script DirLinProdSCMS_fun.py. As the input arguments of DirLinProdSCMSLog subsume the ones of DirLinProdKDE and DirLinProdMS/DirLinProdMSCompAsc, we combine the descriptions of their arguments as follows:
def DirLinProdKDE(x, data, h=[None,None], com_type=['Dir', 'Lin'], dim=[2,1]):
def DirLinProdMS(mesh_0, data, h=[None,None], com_type=['Dir','Lin'], dim=[2,1], eps=1e-7, max_iter=1000):
def DirLinProdMSCompAsc(mesh_0, data, h=[None,None], com_type=['Dir','Lin'], dim=[2,1], eps=1e-7, max_iter=1000):
def DirLinProdSCMSLog(mesh_0, data, d=1, h=[None,None], com_type=['Dir','Lin'], dim=[2,1], eps=1e-7, max_iter=1000, eta=None):
-
Parameters:
-
mesh_0: (m, sum(dim)+sum(com_type=='Dir'))-array ---- Eulidean coordinates of m query points in the product space, where (dim[0]+1) / dim[0] is the Euclidean dimension of a directional/linear component (first (dim[0]+1) columns), and so on.
-
data: (n, sum(dim)+sum(com_type=='Dir'))-array ---- Euclidean coordinates of n random sample points in the product space, where (dim[0]+1) / dim[0] is the Euclidean dimension of a directional/linear component (first (dim[0]+1) columns), and so on.
-
d: int ---- The order of the density ridge. (Default: d=1.)
-
h: list of floats ---- Bandwidth parameters for all the components. (Default: h=[None]*K, where K is the number of components in the product space. Whenever h[k]=None for some k=1,...,K, then a rule of thumb for directional KDE with the von Mises kernel in Garcia-Portugues (2013) is applied to that directional component or the Silverman's rule of thumb is applied to that linear component; see Chen et al.(2016) for details.)
-
com_type: list of strings ---- Indicators of the data type for all the components. If com_type[k]='Dir', then the corresponding component is directional. If com_type[k]='Lin', then the corresponding component is linear.
-
dim: list of ints ---- Intrinsic data dimensions of all the directional/linear components.
-
eps: float ---- The precision parameter. (Default: eps=1e-7.)
-
max_iter: int ---- The maximum number of iterations for the SCMS algorithm on each initial point. (Default: max_iter=1000.)
-
eta: float ---- The step size parameter for the SCMS algorithm. (Default: eta=None, then eta=np.min([np.min(h) * np.max(h), 1]).)
-
-
Return:
- SCMS_path: (m, sum(dim)+sum(com_type=='Dir'), T)-array ---- The entire iterative SCMS sequence for each initial point.
We also provide the corresponding implementations of the above functions under the Ray parallel programming environment as DirLinProdKDE_Fast, DirLinProdMS_Fast, DirLinProdMSCompAsc_Fast, and DirLinProdSCMSLog_Fast in the script DirLinProdSCMS_Ray.py.
Example code:
import numpy as np
import scipy.special as sp
import matplotlib.pyplot as plt
from DirLinProdSCMS_fun import DirLinProdKDE, DirLinProdMS, DirLinProdSCMSLog
from Utility_fun import vMF_Gauss_mix
## Simulation 1: Mode-seeking on a directional-linear space $\Omega_1 \times \mathbb{R}$
np.random.seed(123) ## Set an arbitrary seed for reproducibility
prob1 = [2/5, 1/5, 2/5] ## Mixture probabilities
mu_N1 = np.array([[0], [1], [2]]) ## Means of the Gaussian component
cov1 = np.array([1/4, 1, 1]).reshape(1,1,3) ## Variances of the Gaussian components
mu_vMF1 = np.array([[1, 0], [0, 1], [-1, 0]]) ## Means of the vMF components
kappa1 = [3, 10, 3] ## Concentration parameters of the vMF components
# Sample 1000 points from the vMF-Gaussian mixture model
vMF_Gau_data = vMF_Gauss_mix(1000, q=1, D=1, mu_vMF=mu_vMF1, kappa=kappa1,
mu_N=mu_N1, cov=cov1, prob=prob1)
# Convert the vMF components of the simulated data to their angular coordinates
Angs = np.arctan2(vMF_Gau_data[:,1], vMF_Gau_data[:,0])
vMF_Gau_Ang = np.concatenate([Angs.reshape(-1,1), vMF_Gau_data[:,2].reshape(-1,1)], axis=1)
# Bandwidth selection
data = vMF_Gau_data
n = vMF_Gau_data.shape[0]
q = 1
D = 1
data_Dir = data[:,:(q+1)]
data_Lin = data[:,(q+1):(q+1+D)]
## Rule-of-thumb bandwidth selector for the directional component
R_bar = np.sqrt(sum(np.mean(data_Dir, axis=0) ** 2))
kap_hat = R_bar * (q + 1 - R_bar ** 2) / (1 - R_bar ** 2)
h = ((4 * np.sqrt(np.pi) * sp.iv((q-1) / 2 , kap_hat)**2) / \
(n * kap_hat ** ((q+1) / 2) * (2 * q * sp.iv((q+1)/2, 2*kap_hat) + \
(q+2) * kap_hat * sp.iv((q+3)/2, 2*kap_hat)))) ** (1/(q + 4))
bw_Dir = h
print("The current bandwidth for directional component is " + str(h) + ".\n")
## Normal reference rule of bandwidth selector for the linear component
b = (4/(D+2))**(1/(D+4))*(n**(-1/(D+4)))*np.mean(np.std(data_Lin, axis=0))
bw_Lin = b
print("The current bandwidth for linear component is "+ str(b) + ".\n")
# Set up a set of mesh points and estimate the density values on it
nrows, ncols = (100, 100)
ang_qry = np.linspace(-np.pi-0.1, np.pi+0.1, nrows)
lin_qry = np.linspace(-2, 5.5, ncols)
ang_m1, lin_m1 = np.meshgrid(ang_qry, lin_qry)
X = np.cos(ang_m1.reshape(-1,1))
Y = np.sin(ang_m1.reshape(-1,1))
mesh1 = np.concatenate([X.reshape(-1,1), Y.reshape(-1,1),
lin_m1.reshape(-1,1)], axis=1)
d_DirLin = DirLinProdKDE(mesh1, data=vMF_Gau_data, h=[bw_Dir, bw_Lin],
com_type=['Dir','Lin'], dim=[1,1]).reshape(nrows, ncols)
# below 5% density quantile
d_DirLin_dat = DirLinProdKDE(vMF_Gau_data, vMF_Gau_data, h=[bw_Dir, bw_Lin],
com_type=['Dir','Lin'], dim=[1,1])
vMF_Gau_data_thres = vMF_Gau_data[d_DirLin_dat > np.quantile(d_DirLin_dat, 0.05)]
# Mode-seeking on the denoised data with our proposed mean shift algorithm
DLMS_path = DirLinProdMS(vMF_Gau_data, vMF_Gau_data_thres, h=[bw_Dir, bw_Lin], com_type=['Dir','Lin'],
dim=[1,1], eps=1e-7, max_iter=3000)
## Simulation 2: Ridge-finding on a directional-linear space $\Omega_1 \times \mathbb{R}$
N = 1000
sigma = 0.3
np.random.seed(123) ## Set an arbitrary seed for reproducibility
# Simulated a curve with additive Gaussian noises on a cylinder (directional-linear case)
t = np.random.rand(N)*2*np.pi - np.pi
t_p = t + np.random.randn(1000) * sigma
X_p = np.cos(t_p)
Y_p = np.sin(t_p)
Z_p = t/2 + np.random.randn(1000) * sigma
cur_dat = np.concatenate([X_p.reshape(-1,1), Y_p.reshape(-1,1),
Z_p.reshape(-1,1)], axis=1)
# Use the default bandwidths
bw_Dir = None
bw_Lin = None
# Create a set of mesh points and estimate the density value on it
nrows, ncols = (100, 100)
ang_qry = np.linspace(-np.pi, np.pi, nrows)
lin_qry = np.linspace(-2.5, 2.5, ncols)
ang_m2, lin_m2 = np.meshgrid(ang_qry, lin_qry)
X = np.cos(ang_m2.reshape(-1,1))
Y = np.sin(ang_m2.reshape(-1,1))
qry_pts = np.concatenate((X.reshape(-1,1),
Y.reshape(-1,1),
lin_m2.reshape(-1,1)), axis=1)
d_DirLinProd = DirLinProdKDE(qry_pts, cur_dat, h=[bw_Dir, bw_Lin],
com_type=['Dir','Lin'], dim=[1,1]).reshape(ncols, nrows)
# Proposed SCMS algorithm with our rule-of-thumb step size eta=h1*h2
ProdSCMS_DL_p, lab_DL_p = DirLinProdSCMSLog(cur_dat, cur_dat, d=1, h=[bw_Dir,bw_Lin],
com_type=['Dir','Lin'], dim=[1,1],
eps=1e-7, max_iter=5000, eta=None)
## Plotting the results
fig = plt.figure(figsize=(16,10))
# Create a cylinder for the directional-linear space
theta = np.linspace(-np.pi, np.pi, 100)
z = np.linspace(-2, 5, 100)
th_m, Zc = np.meshgrid(theta, z)
Xc = np.cos(th_m)
Yc = np.sin(th_m)
# Plot the simulated data points and local modes on the cylinder
step = DLMS_path.shape[2] - 1
Modes_angs = np.arctan2(DLMS_path[:,1,step], DLMS_path[:,0,step])
ax = fig.add_subplot(221, projection='3d')
ax.view_init(30, 60)
ax.plot_surface(Xc, Yc, Zc, alpha=0.2, color='grey')
ax.scatter(vMF_Gau_data[:,0], vMF_Gau_data[:,1], vMF_Gau_data[:,2],
alpha=0.2, color='deepskyblue')
ax.scatter(DLMS_path[:,0,step], DLMS_path[:,1,step], DLMS_path[:,2,step],
color='red', s=40)
ax.axis('off')
plt.title('Simulated vMF-Gaussian mixture data and local modes \n estimated '\
'by our mean shift algorithm on a cylinder')
# Plot the local modes on the contour plot of the estimated density
step = DLMS_path.shape[2] - 1
Modes_angs = np.arctan2(DLMS_path[:,1,step], DLMS_path[:,0,step])
plt.subplot(222)
plt.scatter(Angs, vMF_Gau_data[:,2], alpha=1)
plt.contourf(ang_m1, lin_m1, d_DirLin, 10, cmap='OrRd', alpha=0.7)
plt.colorbar()
plt.scatter(Modes_angs, DLMS_path[:,2,step], color='red', s=40)
plt.title('Estimated local modes on the contour plot of KDE')
# Plot the simulated data and estimated ridge on a cylinder
step_DL_p = ProdSCMS_DL_p.shape[2] - 1
ax = fig.add_subplot(223, projection='3d')
ax.view_init(30, 10)
## Mesh points on the cylinder
theta = np.linspace(-np.pi, np.pi, 100)
z = np.linspace(-2, 2, 100)
th_m, Zc = np.meshgrid(theta, z)
Xc = np.cos(th_m)
Yc = np.sin(th_m)
## True curve structure
t = np.linspace(-np.pi, np.pi, 200)
X_cur = np.cos(t)
Y_cur = np.sin(t)
Z_cur = t/2
ax.plot_surface(Xc, Yc, Zc, alpha=0.2)
ax.plot(X_cur, Y_cur, Z_cur, linewidth=5, color='green')
ax.scatter(ProdSCMS_DL_p[:,0,step_DL_p], ProdSCMS_DL_p[:,1,step_DL_p],
ProdSCMS_DL_p[:,2,step_DL_p], alpha=0.5, color='deepskyblue')
ax.axis('off')
plt.title('Simulated data and density ridges \n estimated '\
'by our SCMS algorithm on a cylinder')
# Plot the estimated ridge on the contour plot of estimated density
plt.subplot(224)
plt.contourf(ang_m2, lin_m2, d_DirLinProd, 10, cmap='OrRd', alpha=0.5)
plt.colorbar()
Ridges_angs_p = np.arctan2(ProdSCMS_DL_p[:,1,step_DL_p],
ProdSCMS_DL_p[:,0,step_DL_p])
plt.scatter(Ridges_angs_p, ProdSCMS_DL_p[:,2,step_DL_p], color='deepskyblue', alpha=0.6)
plt.xlabel('Directional Coordinate')
plt.ylabel('Linear Coordinate')
plt.title('Estimated density ridges on the contour plot of KDE')
fig.tight_layout()
fig.savefig('./Figures/DirLin_example.png')
Fig 2. Illustrative plots of applying our proposed mean shift and SCMS algorithms to simulated directional-linear datasets.
Other simulation studies on directional-directional data (e.g., observations on a torus) and real-world applications of our proposed (subspace constrained) mean shift algorithms can be found in our paper.
- R. Ahumada, C. A.Prieto, A. Almeida, F. Anders, S. F. Anderson, B. H. Andrews, B. Anguiano, R. Arcodia, E. Armengaud, M. Aubert, et al. The 16th data release of the sloan digital sky surveys: first release from the apogee-2 southern survey and full release of eboss spectra. The Astrophysical Journal Supplement Series, 249(1):3, 2020.
- Y. Cheng. Mean shift, mode seeking, and clustering. IEEE Transactions on Pattern Analysis and Machine Intelligence, 17(8):790–799, 1995.
- D. Comaniciu and P. Meer. Mean shift: a robust approach toward feature space analysis. IEEE Transactions on Pattern Analysis and Machine Intelligence, 24(5):603–619, 2002.
- U. Ozertem and D. Erdogmus. Locally defined principal curves and surfaces. Journal of Machine Learning Research, 12(34):1249–1286, 2011.
- S. J. Wright. Coordinate descent algorithms. Mathematical Programming, 151(1):3–34, 2015.
- E. Arias-Castro, D. Mason, and B. Pelletier. On the estimation of the gradient lines of a density and the consistency of the mean-shift algorithm. Journal of Machine Learning Research, 17(43):1–28, 2016.
- Y. Zhang and Y.-C. Chen. Linear convergence of the subspace constrained mean shift algorithm: From euclidean to directional data. arXiv preprint arXiv:2104.14977, 2021.
- E. Garcı́a-Portugués (2013). Exact risk improvement of bandwidth selectors for kernel density estimation with directional data. Electronic Journal of Statistics 7 1655–1685.
- Y.-C. Chen, C. Genovese, and L. Wasserman (2016). A comprehensive approach to mode clustering. Electronic Journal of Statistics 10(1) 210-241.