个人总结及展望.ppt

我做的工作就是简单的分类，利用已经有的样例训练出一个模型，然后根据模型来分类未知的数据。这个工作中主要的创新点就是使用集成分类器去预测，可以大大提高预测的准确性，而且使用RNAfold特征提取方法，大大减少了时间耗费。 * 下面介绍miRFam使用n-gram的特征提取方法 由于miRNA只有四个碱基，miRFam使用4个唯一的unigram，42个唯一的bigram，43个唯一的trigram，一共有4+42+43 =84个特征( A,C,G,U,AA,AC,AG,AU,CA,CC,CG,CU……) miRFam使用集中系数的概念来把这些不同的特征整合成一个特征向量，首先，定义类型i的唯一n-gram的数量用Ni表示，类型i的集中系数就是： Ci =Ni／∑3j=1Nj , i=1,2,3 当然我么可以得到： C1=4/4+16+64=0.048 C2=16/4+16+64=0.190 C3=64/4+16+64=0.762 然后特征向量可以通过下面的公式计算： fj = tj/Ti*Ci, 1=j=84 其中，tj是某一种类型i的唯一的n-gram的出现频率，Ti是类型i的所有的唯一n-gram的出现频率。特征向量包含84维，每一维对应于某一种类型i (i =1,2,3,4)的一种n-gram。 * The three layers prediction method was mentioned. The 1st layer: 19 families with the largest members are selected and each of them respectively is viewed as a?class, a total of?19 classes,?remaining?as?the last class,?the?prediction change into?a new model that the?number of the target?class has only?20?classes (the dataset is denoted as and the random forest model is noted as RF1). For the 2nd layer, we?select?the?top?99?families, respectively.?Each of them is seen as a class,?and the residual?ones are taken as?a?class. So there are total of?100?classes (the dataset is denoted as , and the random forest model is noted as RF2). For the 3rd layer, we classify the miRNA families as same as miRBase (the dataset is denoted as the random forest model is noted as RF3). Our method can identify families hierarchically and judge whether a novel miRNA belongs to the popular families. The complete forecasting process is shown in Figure 2. The predicted sequence starts from the 1st layer, and if it is predicted as one of the top 19 families, the output is considered as the last result and the process will finish. Otherwise the process will automatically continue to the 2nd layer prediction. The sequence will be predicted whether it belongs to the top 99 families. If it doesnt belong to yet, miRClassify will go into the 3rd layer and the 3rd layer predictor will give the last result. Hierarchical predictors are used in