Personalized services for individual customers are now popular in e-commerce sites. A recommender system is a typical software solution used in e-commerce for personalized services. Traditionally, recommender systems are based on the binary purchase data. However, conventional systems usually do not work well with binary data which are typical of e-commerce data. To improve the performance of the recommender system, this thesis proposes three types of recommender systems based on navigational and behavioral patterns of customers. The proposed recommender systems include: (i) Collaborative filtering (CF)-based system; (ii) Dimensionality Reduction (DB)-based system; and (iii) Association Rule Mining (ARM)-based system.
The CF-based system consists of the following four phases. First, the data related to a customer’s purchase, navigational and behavioral patterns are collected. Second, the customers’ preference for a certain product is numerically determined. The preference level is determined by estimating the probability of reaching the point of purchase using the data gathered from the first phase. This process is carried out using the decision tree, logistic regression or artificial neural network. Third, CF is performed using the preference levels calculated in the second phase as the input values, and the preference levels of a customer for the products not clicked are predicted. Finally, a Top-N list of products is generated as a recommendation to the customer.
In the DR-based system, Kim and Yum (2005)’s approach is applied to the preference data calculated in the above CF based system. In the existing studies, the DR method has not been applied to implicit rating data (e.g., customer’s behavior data) although it has been applied to the case of explicit rating data (e.g., rating movies). This study is conducted to verify the effectiveness of the DR method for implicit rating data.
Finally, ARM-based system consists of: (i) gathering of all data related to the customers’ purchase, navigational and behavioral patterns; (ii) conversion of a numeric variable to a categorical variable in order to apply the ARM; (iii) performing ARM on the converted data; (iv) calculation of the confidence levels between the clicked products, between the products placed in the basket, and between the purchased products; (v) determination of the preference level for each pair of two products through a linear combination of the above three confidence levels; and (vi) generation of a Top-N list.
The effectiveness of the three recommender systems is assessed using an experimental e-commerce site. Computational results indicate that all three systems outperform the conventional system which utilized only the binary purchase data. Further, the ARM-based system is superior or equal to CF- and DR-based system. However, the performance difference between ARM- and DR-based systems is slight, and therefore, the DR-based system can be substituted for the ARM-based system since the former requires less memory space and computational effort. The recommender systems developed in this thesis is versatile and can be applied to a variety of e-commerce sites as long as the navigational and behavioral patterns of customers can be captured.Personalized services for individual customers are now popular in e-commerce sites. A recommender system is a typical software solution used in e-commerce for personalized services. Traditionally, recommender systems are based on the binary purchase data. However, conventional systems usually do not work well with binary data which are typical of e-commerce data. To improve the performance of the recommender system, this thesis proposes three types of recommender systems based on navigational and behavioral patterns of customers. The proposed recommender systems include: (i) Collaborative filtering (CF)-based system; (ii) Dimensionality Reduction (DB)-based system; and (iii) Association Rule Mining (ARM)-based system.
The CF-based system consists of the following four phases. First, the data related to a customer’s purchase, navigational and behavioral patterns are collected. Second, the customers’ preference for a certain product is numerically determined. The preference level is determined by estimating the probability of reaching the point of purchase using the data gathered from the first phase. This process is carried out using the decision tree, logistic regression or artificial neural network. Third, CF is performed using the preference levels calculated in the second phase as the input values, and the preference levels of a customer for the products not clicked are predicted. Finally, a Top-N list of products is generated as a recommendation to the customer.
In the DR-based system, Kim and Yum (2005)’s approach is applied to the preference data calculated in the above CF based system. In the existing studies, the DR method has not been applied to implicit rating data (e.g., customer’s behavior data) although it has been applied to the case of explicit rating data (e.g., rating movies). This study is conducted to verify the effectiveness of the DR method for implicit rating data.
Finally, ARM-based system consists of: (i) gathering of all data related to the customers’ purchase, navigational and behavioral patterns; (ii) conversion of a numeric variable to a categorical variable in order to apply the ARM; (iii) performing ARM on the converted data; (iv) calculation of the confidence levels between the clicked products, between the products placed in the basket, and between the purchased products; (v) determination of the preference level for each pair of two products through a linear combination of the above three confidence levels; and (vi) generation of a Top-N list.
The effectiveness of the three recommender systems is assessed using an experimental e-commerce site. Computational results indicate that all three systems outperform the conventional system which utilized only the binary purchase data. Further, the ARM-based system is superior or equal to CF- and DR-based system. However, the performance difference between ARM- and DR-based systems is slight, and therefore, the DR-based system can be substituted for the ARM-based system since the former requires less memory space and computational effort. The recommender systems developed in this thesis is versatile and can be applied to a variety of e-commerce sites as long as the navigational and behavioral patterns of customers can be captured.
최근에 전자상거래 사이트에서는 각 고객에게 개별화된 서비스를 제공하고 있다. 그러한 개별화된 서비스 중 하나가 바로 추천 시스템이다. 기존의 전자 상거래에서 사용된 추천 시스템의 경우 고객의 구매 여부만을 다룬 binary 데이터를 사용하였으나, 본 연구에서는 고객의 행동 및 탐색 패턴 데이터까지 이용함으로써 기존의 추천 시스템의 성능을 향상시켰다. 본 논문에서는 다음의 세 종류의 추천 시스템을 개발하였다. 첫째, 협업적 필터링을 이용한 추천 시스템 (CF-based system), 둘째, 차원 감소 기법을 이용한 추천 시스템 (DR-based system), 셋째, 연관성 규칙을 이용한 추천 시스템 (ARM-based system) 이다.
CF-based system은 다음의 네 단계로 구성된다. 첫째, 고객의 구매, 탐색 및 행동 패턴에 대한 모든 데이터를 수집한다. 둘째, 특정 상품에 대한 고객의 선호도를 추정한다. 만일 특정 상품이 구매되었다면, 이때의 선호도는 1로 한다. 반면에 구매가 이루어지지 않은 상품들에 대해서는 첫번째 단계에서 수집된 데이터를 이용하여 구매가 이루어질 확률을 추정하여 이를 선호도로 결정한다. 이 과정에서는 의사결정나무, 로지스틱 회귀분석, 인공신경망 등을 이용한다. 셋째, 두번째 단계에서 구한 선호도를 입력값으로 하여 협업적 필터링을 수행한다. 넷째, 상품을 높은 예측값을 갖는 순서로 나열한 후, Top_N 목록을 만들어 각 고객에게 해당되는 상품을 추천한다. DR-based system에서는 CF-based system에서 구한 고객 선호도 데이터에 대해 Kim and Yum (2005)의 방법을 적용하였다. 기존 연구에서는, 차원 감소 기법이 영화 선호도 데이터와 같은 explicit rating 데이터에 대해서만 적용하였을 뿐, 고객의 행동 데이터인 implicit rating 데이터에 대해 적용하지는 않았다. 본 연구는 차원 감소 기법이 implicit rating 데이터에 효과적으로 적용가능한지 검증하기 위해 수행하였다. ARM-based system은 다음의 여섯 단계로 구성되어 있다. 첫째, 고객의 구매, 탐색 및 행동 패턴에 대한 모든 데이터를 수집한다. 둘째, 연관성 규칙을 적용하기 위해 수치형 변수를 범주형 변수로 변환한다. 셋째, 변환된 데이터에 대해 연관성 분석을 수행한다. 넷째, 클릭된 상품, 장바구니에 담겨진 상품, 구매된 상품간의 신뢰도를 계산한다. 다섯째, 앞서 구한 세 종류의 신뢰도의 선형조합을 통해 두 상품간의 선호도를 결정한다. 여섯째, 상품을 높은 예측값을 갖는 순서로 나열한 후, Top_N 목록을 만들어 각 고객에게 해당되는 상품을 추천한다. 세 가지 추천 시스템의 성능은 실험용 전자상거래 사이트를 이용하여 평가하였다. 본 연구에서 개발한 세가지 추천 시스템의 성능이 구매 데이터만을 이용한 기존의 추천 시스템보다 우수하다는 것을 실험적으로 검증하였다. 또한, ARM-based system이 CF-based system과 DR-based system보다 우수한 성능을 나타냈으나, DR-based system의 성능과 큰 차이를 나타내지는 않았다. 따라서, DR-based system이 메모리 공간을 적게 차지하고 계산 시간도 적게 걸리므로, ARM-based system을 DR-based system으로 대체하는 것이 효과적이다.
