Learning Bayesian Network Classifiers by Maximizing Conditional Likelihood

Abstract

Bayesian networks are a powerful probabilistic representation, and their use for classification has received considerable attention. However, they tend to perform poorly when learned in the standard way. This is attributable to a mismatch between the objective function used (likelihood or a function thereof) and the goal of classification (maximizing accuracy or conditional likelihood). Unfortunately, the computational cost of optimizing structure and parameters for conditional likelihood is prohibitive. In this paper we show that a simple approximation – choosing structures by maximizing conditional likelihood while setting parameters by maximum likelihood – yields good results. On a large suite of benchmark datasets, this approach produces better class probability estimates than naïve Bayes, TAN, and generatively-trained Bayesian networks.

1. Introduction

The simplicity and surprisingly high accuracy of the naïve Bayes classifier have led to its wide use, and to many attempts to extend it. In particular, naïve Bayes is a special case of a Bayesian network, and learning the structure and parameters of an unrestricted Bayesian network would appear to be a logical means of improvement. However, Friedman et al. found that naive Bayes easily outperforms such unrestricted Bayesian network classifiers on a large sample of benchmark datasets. This explanation was that the scoring functions used in standard Bayesian network learning attempt to optimize the likelihood of the entire data, rather than just the conditional likelihood of the class given the attributes. Such scoring results in suboptimal choices during the search process whenever the two functions favor differing changes to the network. The natural solution would then be to use conditional likelihood as the objective function. Unfortunately, Friedman et al. observed that, while maximum likelihood parameters can be efficiently computed in closed form, this is not true of conditional likelihood. The latter must be optimized using numerical methods, and doing so at each search step would be prohibitively expensive. Friedman et al. thus abandoned this avenue, leaving the investigation of possible heuristic alternatives to it as an important direction for future research. In this paper, we show that the simple heuristic of setting the parameters by maximum likelihood while choosing the structure by conditional likelihood is accurate and efficient.

Friedman et al. chose instead to extend naive Bayes by allowing a slightly less restricted structure (one parent per variable in addition to the class) while still optimizing likelihood. They showed that TAN, the resulting algorithm, was indeed more accurate than naive Bayes on benchmark datasets. We compare our algorithm to TAN and naive Bayes on the same datasets, and show that it outperforms both in the accuracy of class probability estimates, while outperforming naive Bayes and tying TAN in classification error.

2. Bayesian Networks

A Bayesian network encodes the joint probability distribution of a set of

2.1 Learning Bayesian Networks

Given an i.i.d. training set

When the structure of the network is known, this reduces to estimating

Since on average adding an arc never decreases likelihood on the training data, using the log likelihood as the scoring function can lead to severe overfitting. This problem can be overcome in a number of ways. The simplest one, which is often surprisingly effective, is to limit the number of parents a variable can have. Another alternative is to add a complexity penalty to the log-likelihood. For example, the MDL method minimizes Bayesian Dirichlet (BD) score:

where

2.2 Bayesian Network Classifiers

The goal of classification is to correctly predict the value of a designated discrete class variable predictors or attributes naïve Bayes classifier is a Bayesian network where the class has no parents and each attribute has the class as its sole parent. Friedman et al.'s TAN algorithm uses a variant of the Chow and Liu method to produce a network where each variable has one other parent in addition to the class. More generally, a Bayesian network learned using any of the methods described above can be used as a classifier. All of these are generative models in the sense that they are learned by maximizing the log likelihood of the entire data being generated by the model, conditional log likelihood

Notice discriminative learning, because it would focus on correctly discriminating between classes. The problem with this approach is that, unlike

3. The BNC Algorithm

We now introduce BNC, an algorithm for learning the structure of a Bayesian network classifier by maximizing conditional likelihood. BNC is similar to the hill climbing algorithm of Heckerman et al. except that it uses the conditional log likelihood of the class as the primary objective function. BNC starts from an empty network, and at each step considers adding each possible new arc (i.e., all those that do not create cycles) and deleting or reversing each current arc. BNC pre-discretizes continuous values and ignores missing values in the same way that TAN does.

We consider two versions of BNC. The first,

The second version,

The goal of