Mercurial Hosting > traffic-intelligence
view python/ml.py @ 309:80cbafd69109
Added spectral clustering function
author | Mohamed Gomaa |
---|---|
date | Tue, 25 Dec 2012 02:24:21 -0500 |
parents | 8bafd054cda4 |
children | 6c068047edbf |
line wrap: on
line source
#! /usr/bin/env python '''Libraries for machine learning algorithms''' import numpy as np __metaclass__ = type class Centroid: 'Wrapper around instances to add a counter' def __init__(self, instance, nInstances = 1): self.instance = instance self.nInstances = nInstances # def similar(instance2): # return self.instance.similar(instance2) def add(self, instance2): self.instance = self.instance.multiply(self.nInstances)+instance2 self.nInstances += 1 self.instance = self.instance.multiply(1/float(self.nInstances)) def average(c): inst = self.instance.multiply(self.nInstances)+c.instance.multiply(instance.nInstances) inst.multiply(1/(self.nInstances+instance.nInstances)) return Centroid(inst, self.nInstances+instance.nInstances) def draw(self, options = ''): from matplotlib.pylab import text self.instance.draw(options) text(self.instance.position.x+1, self.instance.position.y+1, str(self.nInstances)) def clustering(data, similar, initialCentroids = []): '''k-means algorithm with similarity function Two instances should be in the same cluster if the sameCluster function returns true for two instances. It is supposed that the average centroid of a set of instances can be computed, using the function. The number of clusters will be determined accordingly data: list of instances averageCentroid: ''' from random import shuffle from copy import copy, deepcopy localdata = copy(data) # shallow copy to avoid modifying data shuffle(localdata) if initialCentroids: centroids = deepcopy(initialCentroids) else: centroids = [Centroid(localdata[0])] for instance in localdata[1:]: i = 0 while i<len(centroids) and not similar(centroids[i].instance, instance): i += 1 if i == len(centroids): centroids.append(Centroid(instance)) else: centroids[i].add(instance) return centroids def spectralClustering(similarityMatrix,k): ''' Steps of Spectral Clustering''' n= len(similarityMatrix) # create Laplacian matrix rowsum = np.sum(similarityMatrix,axis=0) D = np.diag(1 / np.sqrt(rowsum)) I = np.identity(n) L = I - np.dot(D,np.dot(similarityMatrix,D)) # compute eigenvectors of L U,sigma,V = np.linalg.svd(L) # create feature vector from k first eigenvectors # by stacking eigenvectors as columns features = np.array(V[:k]).T # k-means from scipy.cluster.vq import kmeans, whiten, vq features = whiten(features) centroids,distortion = kmeans(features,k,iter=20) # default iter = 20 code,distance = vq(features,centroids) # code starting from 0 (represent first cluster) to k-1 (last cluster) return code,sigma