Understanding and Scaling-Up Machine Learning Algorithms

Dietterich 一些机器学习算法已经变得非常流行和广泛使用-特别是决策树的自顶向下分区算法和前馈神经网络的反向传播算法。 本研究旨在加深我们对这些算法的理解，解释为什么各种提升和投票技术可以提高其准确性，并确定将其应用于实践的最佳程序。 研究的第二个目标是找到扩展这些算法（及其相关的提升技术）的方法，以处理数十亿个训练示例和数千个输出类的问题。 根据我们在投票算法方面的经验，我们将开发一些方法来降低投票成本，扩大输出类的数量，处理连续输出值，以及处理大量的训练示例。 现有的学习算法是为数据非常昂贵而计算机时间相对便宜的条件而设计的。 数据挖掘中的新兴应用表现出相反的条件：数据量大，用户需要交互速度。 这项研究将导致新的算法，是必不可少的，以支持新兴的应用程序在大规模数据挖掘。

Dietterich Understanding and Scaling-Up Machine Learning Algorithms Several machine learning algorithms have become very popular and widely used -- particularly the top-down partitioning algorithms for decision trees and the backpropagation algorithm for feed-forward neural networks. This research seeks to deepen our understanding of these algorithms, to explain why various boosting and voting techniques improve their accuracy, and to determine the best procedures for applying them in practice. A second goal of the research is to find ways of scaling up these algorithms (and their associated boosting techniques) to handle problems with billions of training examples and thousands of output classes. Based on our experience with voting algorithms, we will develop methods for reducing the cost of voting, scaling up the number of output classes, handling continuous output values, and handling large numbers of training examples. Existing learning algorithms have been designed for conditions where data is very expensive and computer time is relatively cheap. Emerging applications in data mining exhibit the opposite conditions: data is voluminous and users want interactive speeds. This research will result in new algorithms that are essential to supporting emerging applications in large-scale data mining.