Discriminative learning of Bayesian networks for data mining applications

贝叶斯网络在数据挖掘中得到了广泛的应用，从数据中学习贝叶斯网络至关重要。大多数BN学习算法在所有变量上学习一个联合分布，以最好地拟合训练数据，即生成学习。判别学习直接针对特定任务模型，而不是生成模型，并且表现出优越的性能。判别学习网络包括两个步骤：结构学习和参数学习。结构学习确定BN的结构，参数学习确定每个变量的局部分布。本研究的目的是研究神经网络在各种数据挖掘任务中的判别学习：分类、排序和概率估计。在本研究中，我们将探讨以下问题。(1)有效判别BN学习：我们将研究能够有效发现底层结构的判别结构学习和能够有效计算最优参数的判别参数学习。(2)可扩展的判别BN学习：本研究的目标之一是开发可扩展到大规模数据挖掘的高效判别BN学习算法。我们将研究结构学习和参数学习的各种放大方法。(3)判别式BN学习用于排序和概率估计：在一些数据挖掘应用中，通常需要准确的排序和概率估计。在产生准确的排名和概率方面，判别学习是否更优越？如何针对这一目的进行针对性培训？应该使用什么样的评分函数和搜索策略？我们的研究将回答这些问题。由于判别学习的固有特征，我们期望本研究能够产生更有效和高效的数据挖掘算法。

Bayesian networks (BNs) have been widely used in data mining applications, in which learning a BN from data is crucial. Most BN learning algorithms learn a joint distribution over all the variables to best fit the training data, i.e. generative learning. Discriminative learning directly targets a task-specific model, instead of a generative model, and demonstrates superior performance. Discriminatively learning a BN includes two steps: structure learning and parameter learning. Structure learning determines the structure of a BN, and parameter learning sets the local distribution for each variable.The objective of this research is to investigate the discriminative learning of BNs for various data mining tasks: classification, ranking, and probability estimation. In this research, we will investigate the following issues. (1) Effective discriminative BN learning: We are going to study discriminative structure learning that can effectively discover the underlying structure, and discriminative parameter learning that can compute the optimal parameters efficiently. (2) Scalable discriminative BN learning: One objective of this research is to develop efficient discriminative BN learning algorithms scalable to large-scale data mining. We are going to study various scaling up methods for  both structure learning and parameter learning. (3) Discriminative BN learning for ranking and probability estimation: In some data mining applications, accurate rankings and probability estimates are often desired. Is discriminative learning superior in terms of yielding accurate rankings and probabilities? How can BNs be discriminatively trained for this purpose? What kind of scoring functions and search strategies should be used? Our research will answer these questions.Due to the inherent features of discriminative learning, we would expect that this research could result in more effective and efficient data mining algorithms.