상품에 대한 고객들의 구매네트워크는 고객 간의 커뮤니케이션 네트워크를 기반으로 하여 존재한다. 본 연구에서는 커뮤니케이션 네트워크와 구매네트 간의 상호의존성과 관련하여, 커뮤니케이션 네트워크 상 고객의 내향중앙성과 외향중앙성이 다른 고객으로 하여금 특정 상품을 선택, 구매함에 있어서 얼마나 영향을 주는지 그리고 어떠한 형태로 영향을 주는지를 비모수적인 방법인 다변량 적응회귀스플라인모형(MARS)을 사용하여 실증적으로 검토하였다. MARS모형의 추정결과, 내향중앙성과 외향중앙성 모두 상품의 선택확률에 유의적인 영향을 비선형적으로 준다는 것, 선택확률에 대한 내향중앙성과 외향중앙성의 효과가 위로 볼록하다는 것, 그리고 선택확률을 최대화시키는 내향중앙성과 외향중앙성의 최적 조합이 존재한다는 것을 발견하였다.
One of the most important research questions in social network analysis is to capture, or predict, differences in tie strengths across node pairs. In a majority of marketing-based networks, multiple relations between nodes are not independent from one another. For example, aside from their core communication services, mobile phone service providers provide their customers with various extra products or services, such as mobile phone games, MP3 songs, and e-books. Since these extra products or services may proliferate via communication network among mobile phone users, it is critical to understand interwoven structures between communication and purchase networks. A growing body of studies in marketing has applied sociogram-based concepts and methods to investigate various issues of interrelationships among customers. None of these extant studies, has, thus far, examined interdependencies between communication relations and purchase relations. This paper aims to empirically investigate possible interdependencies between communication network and purchase networks. To this end, we measured indegree and outdegree centralities for each node in communication network and then rigorously investigated possible effects of the two centrality indices in communication network on their relations in purchase network on the basis of multivariate adaptive regression spline (MARS). In particular, degree centralities were measured for each pair of nodes in the communication network, as follows: where denotes connected (1) or disconnected (0) directed ties starting node and ending at node . Let denotes a binary observation for whether node made purchases (1) for a given product or service. Then, MARS model was fitted to data across all the possible pairs of nodes, as follows: where denotes probability of node making purchases for the product or service, is an unknown target function to be estimated, represents candidate explanatory variables for the directed tie , is the total number of hyperplanes, denotes the -th base spline function, and is the spline coefficient associated with . MARS model was fitted to real mobile phone usage data containing information about two particular types of relations among nodes: communications (i.e. phone call records) and purchases for a non-free extra service. Major findings attained from this particular application can be summarized as follows: First, both indegree and outdegree centralities were found to increase choice probabilities for the non-free extra service. Mobile phone users who might have relatively high scores for indegree and outdegree centralities would, therefore, exhibit relatively large influence on purchase decisions of those who were communicationally connected to them. Second, choice probability responses functions of indegree and outdegree centralities were not linear but nonlinear. These nonlinearities have been found by none of all the extant studies in the marketing literature, primarily because all of them have imposed linearity presumptions a priori on the effects of centralities on adoptions of information and/or products. Third, response surface of choice probabilities were not monotonically increasing but concave. Choice probabilities for the non-free extra service were, therefore, maximed at particular combinations of indegree and outdegree centrallities, rather than at maximum values of each of them. This last finding has not been reported by any of the extant studies in the marketing literature, mainly because they have failed in accounting for decisive possibilities that indegree and outdegree centralities might affect choice probabilities nonlinearly even with higher-order interaction effects.