Dirichlet Multinomial Mixture Model做短文本聚类（包括代码）

#原文地址
http://blog.csdn.net/qy20115549/article/details/79429127
#论文来源
Yin J, Wang J. A dirichlet multinomial mixture model-based approach for short text clustering[C]//Proceedings of the 20th ACM SIGKDD international conference on Knowledge discovery and data mining. ACM, 2014: 233-242.

#论文理解及公式推理
<center>
![这里写图片描述](http://img.blog.csdn.net/20180303112906858?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvcXkyMDExNTU0OQ==/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast)
![这里写图片描述](http://img.blog.csdn.net/20180303112917136?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvcXkyMDExNTU0OQ==/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast)
![这里写图片描述](http://img.blog.csdn.net/20180303112925508?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvcXkyMDExNTU0OQ==/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast)
![这里写图片描述](http://img.blog.csdn.net/20180303112954863?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvcXkyMDExNTU0OQ==/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast)
![这里写图片描述](http://img.blog.csdn.net/20180303113003768?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvcXkyMDExNTU0OQ==/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast)
</center>

#核心源码
`//模型初始化
	public void intialize(DocumentSet documentSet)
	{
		
		D = documentSet.D;	//获取文档总数目
		z = new int[D]; //文档对应的主题数目
		for(int d = 0; d < D; d++){
			//获取每一篇文档的内容
			Document document = documentSet.documents.get(d);
			//针对每一篇文档随机初始化一个主题
			int cluster = (int) (K * Math.random());
			z[d] = cluster;
			//每个主题对应的文档数目统计
			m_z[cluster]++;
			//对文档的每个单词进行循环
			for(int w = 0; w < document.wordNum; w++){
				//获取文档每个单词的编号
				int wordNo = document.wordIdArray[w];
				//获取文档每个单词出现的数目
				int wordFre = document.wordFreArray[w];
				//统计每个主题下，每个单词出现的数目
				n_zv[cluster][wordNo] += wordFre;
				//统计每个主题下所有单词的数目
				n_z[cluster] += wordFre; 
			}
		}
	}
	//gibbs采样
	public void gibbsSampling(DocumentSet documentSet)
	{
		for(int i = 0; i < iterNum; i++){
			//每篇文档循环
			for(int d = 0; d < D; d++){
				Document document = documentSet.documents.get(d);
				//获取文档对应的主题
				int cluster = z[d];
				//移除该文档，该主题对应的文档数目减去1
				m_z[cluster]--;
				for(int w = 0; w < document.wordNum; w++){
					int wordNo = document.wordIdArray[w];
					int wordFre = document.wordFreArray[w];
					//该主题对应的文档中的单词的数目，减少文档该单词出现的数目
					n_zv[cluster][wordNo] -= wordFre;
					//该主题对应的单词总数减去了该文档对应单词的总数
					n_z[cluster] -= wordFre;
				}
				//抽取该文档所属的新主题
				cluster = sampleCluster(d, document);
				//分配该文档对应的新主题后，重新统计相关词频
				z[d] = cluster;
				m_z[cluster]++;
				for(int w = 0; w < document.wordNum; w++){
					int wordNo = document.wordIdArray[w];
					int wordFre = document.wordFreArray[w];
					n_zv[cluster][wordNo] += wordFre; 
					n_z[cluster] += wordFre; 
				}
			}
		}
	}

private int sampleCluster(int d, Document document)
	{ 
		double[] prob = new double[K];
		//统计是哪个主题
		int[] overflowCount = new int[K];
		//对所有主题进行循环，计算该文档属于每个主题的概率
		for(int k = 0; k < K; k++){
			//依照计算公式计算文档d属于每个单词的概率
			prob[k] = (m_z[k] + alpha) / (D - 1 + alpha0);
			double valueOfRule2 = 1.0;
			int i = 0;
			for(int w=0; w < document.wordNum; w++){
				//获取该文档中某一单词的编号及出现的频率
				int wordNo = document.wordIdArray[w];
				int wordFre = document.wordFreArray[w];
				//文档的每个单词进行计算，这里有防止连乘积概率过小的判断及处理
				for(int j = 0; j < wordFre; j++){
					if(valueOfRule2 < smallDouble){
						overflowCount[k]--;
						valueOfRule2 *= largeDouble;
					}
					
					valueOfRule2 *= (n_zv[k][wordNo] + beta + j) 
							 / (n_z[k] + beta0 + i);
					i++;
				}
			}
			prob[k] *= valueOfRule2;			
		}
		//重新计算概率
		reComputeProbs(prob, overflowCount, K);
		//使用轮盘赌，分配新的主题
		for(int k = 1; k < K; k++){
			prob[k] += prob[k - 1];
		}
		double thred = Math.random() * prob[K - 1];
		int kChoosed;
		for(kChoosed = 0; kChoosed < K; kChoosed++){
			if(thred < prob[kChoosed]){
				break;
			}
		}
		
		return kChoosed;
	}
	//重新计算概率值
	private void reComputeProbs(double[] prob, int[] overflowCount, int K)
	{
		int max = Integer.MIN_VALUE;
		for(int k = 0; k < K; k++){
			if(overflowCount[k] > max && prob[k] > 0){
				max = overflowCount[k];
				System.out.println("max:"+max);
			}
		}
		//概率值统一扩展
		for(int k = 0; k < K; k++){			
			if(prob[k] > 0){
				prob[k] = prob[k] * Math.pow(largeDouble, overflowCount[k] - max);
			}
		}		
	}
`

Dirichlet Multinomial Mixture Model做短文本聚类（包括代码）

相关博客

最热博客