감리지적기업의 분류적 특성에 관한 연구: 베이지안 망과 C5.0, 그리고 앙상블 방법간의 비교를 중심으로 (A Study on the Classification Properties of Firms to beSubject to Accounting Disclosure Reviews andInvestigations: Comparison of Bayesian Network, C5.0, andEnsemble Prediction Methods)

서지정보

· 발행기관 : 한국경영학회
· 수록지 정보 : 경영학연구 / 36권 / 3호 / 705 ~ 737페이지
· 저자명 : 이건창, 최관

초록

본 연구는 감사보고서 감리지적기업과 감리비지적기업을 효율적으로 구분할 수 있는 분류방법에 대한 연구이다. 본 연구에서는 선행연구에서 널리 사용되어온 분류방법인 로짓회귀분석 방법이 종속변수와 설명변수간에 획일적인 선형함수만을가정하는 데에서 나오는 문제점을 두 가지 측면에서 극복하고자 한다. 첫째는 감리지적 여부에 영향을 미치는 설명변수간에 존재하는 인과관계(causal relationship)를 도출할 필요가 있다. 이는 어떤 변수가 다른 어떤 변수와 직접 또는 간접적 인과관계를 통하여 감리지적 여부에 영향을 주는지를 의사결정자에게 알려줌으로써 보다 효과적인 감리작업을 할 수있도록 지원할 수 있다. 이를 위하여 본 연구에서는 일반 베이지안 망(GBN: General Bayesian Network)을 제안하고 GBN에서 유도되는 마코프 블랭킷(Markov Blanket)을 제시한다. 둘째는 감리지적 예측을 보다 정확하게 하기 위하여 기존에 사용되던 분류방법인 GBN과 나이브 베이지안 망(NBN: Naive Bayesian Network) 및 C5.0을 결합한 앙상블 방법을 제시한다. 1990년에서 1999년까지의 감리지적 및 감리비지적기업의 자료를 기초로 실험한 결과 본 연구에서 제안하는 두 가지 방법이 모두 통계적으로 유의한 결과를 제공한다는 것이 실증적으로 검증되었다.

영어초록

As many people is increasingly investing in those firms listed in the stock market, it becomes important than ever to review and evaluate the reliability of audit reports. To secure the reliability, the Korean Financial Supervisory Service has conducted the ADRI (accounting disclosure reviews and investigations) of the audit reports selectively. However, since there is a lack of systematic as well as highly predictive method that can classify the firms subject to ADRI precisely, the Korean Financial Supervisory Service needs more refined classification methods.
This study proposes using an ensemble method which is based on combining three predictive methods such as general Bayesian network (GBN), naive Bayesian network (NBN) and C5.0. Especially, in the process of proposing the ensemble method, we revealed that the Markov Blanket induced from GBN can show the underlying structure hidden in the data in terms of cause-effect relationships between minimal set of relevant variables explaining the target variables. Rigorous experiments with the ADRI sample data ranging from 1990 to 1999 showed that the proposed ensemble method surpasses other methods in terms of prediction accuracy, and it can be used significantly for the purpose of performing ADRI activities very systematically. In addition, we found the usefulness of the Markov Blanket obtained from GBN in explaining how a certain variable affects the target variables through the causal relationships with other related variables.
Basically, this study is based on using Bayesian network as a main vehicle of developing an ensemble method to deal with the ADRI problem. Therefore, let us explain the meaning of Bayesian Network. BN or Bayesian Network is denoted as B=<G, P>, is a directed acyclic graph (DAG) G with a set of conditional probability distributions P, that satisfies the Markov condition. Given its parents, each node is conditionally independent of the set of all of its non-descendants. Bayesian network classifiers theoretically seek an optimal error rate based on posteriori probabilities. If all child nodes were relevant to the root node (target concept) and conditionally independent of each other given the root, the naïve Bayesian network (NBN) would be the optimal classifier. However, in real-world applications its performance can be degraded since variables are often related to each other. The tree augmented naïve Bayesian network (TAN) introduced by Friedman et al. (1997) relaxes the unrealistic independence assumption of NBN by allowing arcs between the child nodes. The GBN can be viewed as a generalized form of TAN, where the nodes are allowed to form an arbitrary graph (rather than just a tree) and each child node need not be connected to the class node. Hence, the GBN is clearly more suitable to finding an underlying structure, since it can reveal true interdependencies among all variables. In this paper, we use the GBN structure as the basis from which to reveal the underlying structure of churn behavior.
Its two byproducts, the Markov blanket and evidential beliefs, are then utilized for assisting complex models.
The proposed method for solving ADRI problem is based on the selection of salient features by Markov Blanket (MB) of GBN. The Markov blanket introduced here has four major advantages over traditional feature selection methods. First, it provides an explicit criterion for selecting salient features. Second, it can identify redundant features. Third, it can identify nonlinear interdependencies among features of interest, allowing the underlying structure of a problem to be found. Fourth, it produces a minimal set of variables required for learning a target concept, ultimately reducing the costs for collecting and maintaining variables. The concept of the Markov blanket was first introduced by Pearl in 1988, but has recently received renewed attention in the areas of Bayesian learning (Cheng and Greiner 1999, Margaritis and Thrun 1999) and features selection (Koller and Sahami 1996, Tsamardinos et al. 2003). The probabilistic nature of the Markov blanket can be explained using the concept of d-separation (direction dependent separation) (Pearl, 1988), a graphicalcriterion related to the blocking of information flow among variables. In a faithful A DAG G is faithful to a joint probability distribution P over set of variables if and only if every conditional independency entailed by G is also present in P (Sprites et al. 1993). Bayesian network, d-separation captures all of the conditional independence relationships encoded in the network. If all variables in the Markov blanket of a node are instantiated, then we can say that the node is d-separated from the rest of the network.
When the Bayesian network is faithful, MB(T) probabilistically shields T from the rest of the variables. Given this property, knowledge about the variables belonging to MB(T) is enough to determine the probability distribution of T, thus rendering all other variables superfluous. In other words, MB(T) is a minimal feature subset required to predict T, which graphically corresponds to a set neighborhood of T: its parents, its children and the other parents of its children. It is straightforward that this set d-separates T from the set of all other nodes. See Neapolitan (2004) for the proof.
In conclusion, we used the Markov blanket here to show the underlying structure of ADRI problem under consideration. On the other hand, the ensemble prediction method was proposed to predict the ADRI better than the existing methods like GBN, NBN, C5.0, and even logistic regression. Rigorous experiment with real ADRI data showed that the proposed MB administered by GBN can reveal the underlying structure of the ADRI problem, and the proposed ensemble prediction method can yield more accurate ADRI prediction results.

참고자료

· 없음