论文部分内容阅读
结构化支持向量机是机器学习中描述结构化输出问题的一种新模型,对其进行训练是一个典型的非光滑凸优化问题,最常用的训练算法是切平面法。切平面法中原问题的目标函数往往会发生振荡,因此一般需要加入线搜索环节。但是还没有针对结构化支持向量机的高效的线搜索方法。该文提出了一种优化的切平面法,通过二次插值来进行近似线搜索,并将其应用到结构化支持向量机的训练中。在多类分类上的实验表明:该算法的迭代次数接近精确线搜索,而每次迭代的计算量保持不变。在序列分类上的实验表明:该算法在训练其他复杂类型的结构化支持向量机时仍然比当前主流算法效率高很多。
Structured SVM is a new model for describing structured output problems in machine learning. Training it is a typical non-smooth convex optimization problem. The most common training algorithm is the tangent plane method. The objective function of the original problem in the tangent plane method often oscillates, so it usually needs to join the line search. However, there is no efficient line search method for structured support vector machines. In this paper, an optimized slicing method is proposed, which performs approximate line search by quadratic interpolation and applies it to the training of structured support vector machines. Experiments on various categories show that the number of iterations of this algorithm is close to the exact line search, and the computation amount of each iteration remains unchanged. Experiments on sequence classification show that the algorithm is still much more efficient than the current mainstream algorithms in training other complex types of structured support vector machines.