毕业论文 外文翻译 求好人帮忙（2）

毕业论文 外文翻译 求好人帮忙（2）
Recent work in the active learning field suggests an optimal way to identify which instances to label next by minimizing expected classification error [29]. This approach is also known as empirical risk minimization (ERM). The approach,while having very good performance, is also computationally expensive as one need to consider, for each instance, what the classification error would be if one was to know the correct label for that instance. In other words, look at the expected classification error as the instance takes on each of the possible class-labels and weight it by the (predicted) likelihood that the instance takes on that class label. This means that one needs to induce a new classifier for each unlabeled instance in order to estimate its risk.
One recently advocated method for semi-supervised learning on relational data uses Gaussian fields and harmonic functions [37]. This particular method, and graph-based methods like it (e.g., [5, 18, 37, 34]), assume that data is presented in the form of a partially labeled graph, where all the data is interconnected. This setting is also known as within-network learning in statistical relational learning
[21], although the graph-based methods consider only the univariate case (i.e., the graph consists only of one type of node and edge and the only variable is the class label, whose value is known for a subset of the nodes in the graph）.
The Gaussian fields and harmonic functions approach has been used in conjunction with active learning with some success [38]. To our knowledge, this is the only published work where these two learning methodologies have been
combined. Various active learning strategies were tested in this work, where they found that using ERM did indeed perform the best among the various strategies. While an efficient way of computing the risk was proposed (to minimize the computational cost), one still needs to consider all the unlabeled points, which is an expensive task. One key observation made in the work was that ERM often picked
nodes that were centers of “groups” in the graphs.

最近的工作在主动学习领域提出了最佳方式辨识哪种情况下,通过最小化的标签将分类错误(29岁).这种方法也被称为经验风险最小化范围之内.该方法都有很好的效果,也是求取作为一个需要考虑的,因为每个实例的分类错误,如果一个是知道正确的标签,这种情况下.换句话说,看看预期的分类误差为实例呈现每种可能class-labels和重量,它由(预测)可能性实例呈现那个类标签.这意味着一个人需要引起一种新的分类器为每个unlabeled实例以估计其风险.
倡导的一种新方法在半监督学习的关系数据使用高斯领域和谐波功能[37].这个特殊方法,基于图论方法如(如[5,18岁,37岁,34]),假设数据的形式给出了部分标记,所有的数据图表所示,相互联系的.这个设置也被称为within-network学习在统计关联的学习
(21),虽然基于图论方法只考虑的案例.数理统计与管理.1999,18(例如,图形都是由一种节点和边缘和唯一的变量在班上标签,其价值是知道中节点的子集下图).
高斯领域和谐波功能方案已经结合使用,与主动学习都能成功[38].根据我们所知,这是唯一的已经发表的作品,在这两种学习方法
结合.各种活性的学习策略进行了测试,在这部作品中,他们在那儿发现使用嗯确实执行最好的在各种策略.当一个有效的方式提出了风险计算的计算成本降到最低的(),你还需要考虑到所有的unlabeled积分,这是一个昂贵的任务.一个关键的观察结果在工作中是风险管理经常被选中
节点中心的“组”在图.