마할라노비스 거리를 이용한 자료융합전략의 성과측정 (Performance Evaluation of Data FusionStrategy using Mahalanobis Distance)

서지정보

· 발행기관 : 한국경영학회
· 수록지 정보 : 경영학연구 / 34권 / 6호 / 1853 ~ 1867페이지
· 저자명 : 김성호, 조성빈

초록

본 연구는 자료융합에서 효과적인 데이터베이스의 축소와 설계에 기여할 수 있는 여러 가지 기증자선택전략을 탐구하고그 성과를 비교하고자 한다. 공통속성변수를 미리 정하여 놓은 계획된 자료융합을 설정하였으며, 자료융합의 영역에 최초로 마할라노비스 거리의 개념을 도입하고 응답자간의 비유사성을 측정하였다. 본 연구에서는 다섯 가지의 기증자선택전략이 설계되었다: 상관계수를 적용하는 전략; 유클리디안 거리를 적용하는 전략; 마할라노비스 거리를 적용하는 전략; 대응일치분석을 거쳐 유클리디안 거리를 적용하는 전략; 대응일치분석을 거쳐 마할라노비스 거리를 적용하는 전략. 몬테카를로 시뮬레이션을 실시한 결과, 대응일치분석을 거쳐 마할라노비스 거리를 적용하는 전략의 성과가 가장 우수하게 도출되었으며, 다음으로는 대응일치분석을 거쳐 유클리디안 거리를 적용하는 전략으로 나타났다. 본 연구의 실험설계와 결과에서 발견한 점으로는 첫째, 대응일치분석은 공통속성변수의 차원을 축소하고 보다 변별력을 향상시키는데 기여하였으며, 둘째, 마할라노비스 거리의 도입은 공통속성변수 간에 실재하는 통계적 종속성을 모델에 반영함으로써 예측의 성과를 높이는데 유효한 역할을 한 것으로 판단된다.

영어초록

For successful implementation of customer relationship management, corporations today do their best to understand their customer's needs. Conjoint analysis has been used to analyze consumer behaviors toward goods and services. However, it is also notorious for its exponentially increasing number of hypothetical products to evaluate. The most economical and fastest way of gathering information about customers is through questionnaire.
Especially today a common form of survey is moving to Internet survey in which respondents can easily get distracted during answering and thereby, survey results might lose sincerity.
This study insists that one way of gathering sincere answers from customers is to give them a smaller amount of questions so that they can answer quickly, maintaining constant attention.
Data fusion plays a key role in merging more than two databases and creating an integrated one. Few studies have investigated the intentional data missing where common attribute variables must be determined before collecting data. This study examines various donor location strategies in the area of intentional, preplanned data fusion. We newly introduce the concept of Mahalanobis distance in measuring dissimilarity between respondents in data fusion. In particular, the experiments are accomplished using the following five strategies: Strategy 1 - locating donors by correlation coefficient; Strategy 2 - by Euclidean distance; Strategy 3 - by Mahalanobis distance; Strategy 4 - by Euclidean distance after employing correspondence analysis; and Strategy 5 - by Mahalanobis distance after employing correspondence analysis. By increasing the level of missing, we evaluate the performance of the above five strategies.
A part-worth data composed of 12 attributes and 35 attribute levels is used for the experiment. The sample is divided into two groups. The common attribute variables are selected by the size of variance of attributes. The missing variables are randomly selected among the non-common attributes and its all attribute levels are deleted. In the analysis, this type of systematic missing is designed to simulate preplanned data missing. Missing values in one group are substituted from the other group and vice versa.
In the experimental design, the concept of ideal point is introduced. The ideal point means that the maximum attribute level represents the corresponding attribute. Here, continuous variables are converted into categorical variables. To measure the distance from these variables, correspondence analysis is performed in which the coordinates of (p - 1) dimensions are computed where p is the number of attributes. The purpose of ideal point is to decrease the number of dimensions since 19 common attribute levels exist. After applying the correspondence analysis, Euclidean distance and Mahalanobis distance are measured in the last two strategies, compared to purely measuring both distances in Strategies 2 and 3.
A Monte Carlo simulation is conducted 20 times for the five strategies and three levels of missing. The results show that donor location strategy and the level of missing are both statistically significant. By comparing the means of experiment factors, Strategy 5 outperforms other strategies. Next accurate strategy is Strategy 4, and then Strategy 1, Strategy 2, Strategy 3. In the light of the above fact, correspondence analysis seems to play an essential role in decreasing the number of common variables while enhancing explanatory power since Strategies 4 and 5 are better than others.
The purpose of introducing Mahalanobis distance is to reflect the statistical dependence structure of common variables because it has not been considered in the existing Euclidean distance-based data fusion techniques. In the analysis, Mahalanobis distance is effective only when a moderate number of common variables are included in the model. This study explores the possibility of applying Mahalanobis distance for measuring dissimilarity of customers and jointly using correspondence analysis in the area of data fusion. Future extension of this study might apply the proposed donor location strategies into various types of data sets and try to reach a comprehensive conclusion.

참고자료

· 없음