* (C) Copyright 2005, Gregor Heinrich (gregor :: arbylon : net) (This file is
* part of the org.knowceans experimental software packages.)
*/
/*
* LdaGibbsSampler is free software; you can redistribute it and/or modify it
* under the terms of the GNU General Public License as published by the Free
* Software Foundation; either version 2 of the License, or (at your option) any
* later version.
*/
/*
* LdaGibbsSampler is distributed in the hope that it will be useful, but
* WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
* FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for more
* details.
*/
/*
* You should have received a copy of the GNU General Public License along with
* this program; if not, write to the Free Software Foundation, Inc., 59 Temple
* Place, Suite 330, Boston, MA 02111-1307 USA
*/
/*
* Created on Mar 6, 2005
*/
package com.xh.lda;
import java.text.DecimalFormat;
import java.text.NumberFormat;
/**
* Gibbs sampler for estimating the best assignments of topics for words and
* documents in a corpus. The algorithm is introduced in Tom Griffiths' paper
* "Gibbs sampling in the generative model of Latent Dirichlet Allocation"
* (2002).
*
* @author heinrich
*/
public class LdaGibbsSampler {
/**
* document data (term lists)
*/
int[][] documents;
/**
* vocabulary size
*/
int V;
/**
* number of topics
*/
int K;
/**
* Dirichlet parameter (document--topic associations)
*/
double alpha;
/**
* Dirichlet parameter (topic--term associations)
*/
double beta;
/**
* topic assignments for each word.
*/
int z[][];
/**
* cwt[i][j] number of instances of word i (term?) assigned to topic j.
*/
int[][] nw;
/**
* na[i][j] number of words in document i assigned to topic j.
*/
int[][] nd;
/**
* nwsum[j] total number of words assigned to topic j.
*/
int[] nwsum;
/**
* nasum[i] total number of words in document i.
*/
int[] ndsum;
/**
* cumulative statistics of theta
*/
double[][] thetasum;
/**
* cumulative statistics of phi
*/
double[][] phisum;
/**
* size of statistics
*/
int numstats;
/**
* sampling lag (?)
*/
private static int THIN_INTERVAL = 20;
/**
* burn-in period
*/
private static int BURN_IN = 100;
/**
* max iterations
*/
private static int ITERATIONS = 1000;
/**
* sample lag (if -1 only one sample taken)
*/
private static int SAMPLE_LAG;
private static int dispcol = 0;
/**
* Initialise the Gibbs sampler with data.
*
* @param V
* vocabulary size
* @param data
*/
public LdaGibbsSampler(int[][] documents, int V) {
this.documents = documents;
this.V = V;
}
/**
* Initialisation: Must start with an assignment of observations to topics ?
* Many alternatives are possible, I chose to perform random assignments
* with equal probabilities
*
* @param K
* number of topics
* @return z assignment of topics to words
*/
public void initialState(int K) {
int i;
int M = documents.length;
// initialise count variables.
nw = new int[V][K];
nd = new int[M][K];
nwsum = new int[K];
ndsum = new int[M];
// The z_i are are initialised to values in [1,K] to determine the
// initial state of the Markov chain.
z = new int[M][];
for (int m = 0; m < M; m++) {
int N = documents[m].length;
z[m] = new int[N];
for (int n = 0; n < N; n++) {
int topic = (int) (Math.random() * K);
z[m][n] = topic;
// number of instances of word i assigned to topic j
nw[documents[m][n]][topic]++;
// number of words in document i assigned to topic j.
nd[m][topic]++;
// total number of words assigned to topic j.
nwsum[topic]++;
}
// total number of words in document i
ndsum[m] = N;
}
}
/**
* Main method: Select initial state ? Repeat a large number of times: 1.
* Select an element 2. Update conditional on other elements. If
* appropriate, output summary for each run.
*
* @param K
* number of topics
* @param alpha
* symmetric prior parameter on document--topic associations
* @param beta
* symmetric prior parameter on topic--term associations
*/
public void gibbs(int K, double alpha, double beta) {
this.K = K;
this.alpha = alpha;
this.beta = beta;
// init sampler statistics
if (SAMPLE_LAG > 0) {
thetasum = new double[documents.length][K];
phisum = new double[K][V];
numstats = 0;
}
// initial state of the Markov chain:
initialState(K);
System.out.println("Sampling " + ITERATIONS
+ " iterations with burn-in of " + BURN_IN + " (B/S="
+ THIN_INTERVAL + ").");
for (int i = 0; i < ITERATIONS; i++) {
// for all z_i
for (int m = 0; m < z.length; m++) {
for (int n = 0; n < z[m].length; n++) {
// (z_i = z[m][n])
// sample from p(z_i|z_-i, w)
int topic = sampleFullConditional(m, n);
z[m][n] = topic;
}
}
if ((i < BURN_IN) && (i % THIN_INTERVAL == 0)) {
// System.out.print("B");
dispcol++;
}
// display progress
if ((i > BURN_IN) && (i % THIN_INTERVAL == 0)) {
// System.out.print("S");
dispcol++;
}
// get statistics after burn-in
if ((i > BURN_IN) && (SAMPLE_LAG > 0) && (i % SAMPLE_LAG == 0)) {
updateParams();
// System.out.print("|");
if (i % THIN_INTERVAL != 0)
dispcol++;
}
if (dispcol >= 100) {
// System.out.println();
dispcol = 0;
}
}
}
/**
* Sample a topic z_i from the full conditional distribution: p(z_i = j |
* z_-i, w) = (n_-i,j(w_i) + beta)/(n_-i,j(.) + W * beta) * (n_-i,j(d_i) +
* alpha)/(n_-i,.(d_i) + K * alpha)
*
* @param m
* document
* @param n
* word
*/
private int sampleFullConditional(int m, int n) {
// remove z_i from the count variables
int topic = z[m][n];
nw[documents[m][n]][topic]--;
nd[m][topic]--;
nwsum[topic]--;
ndsum[m]--;
// do multinomial sampling via cumulative method:
double[] p = new double[K];
for (int k = 0; k < K; k++) {
p[k] = (nw[documents[m][n]][k] + beta) / (nwsum[k] + V * beta)
* (n
评论0