딥러닝 기법에 기반한 인터넷 상점 클릭스트림 데이터를 통한 구매 예측

Abstract

최근 빅 데이터가 이슈가 되면서 데이터들을 어떻게 분석할 것인지가 이슈가 되고 있다. 이 중 딥러닝(Deep Learning)이라 불리는 심층신경망 기법들이 이미지 인식이나 음성 인식에서 뛰어난 성능을 보여 주목을 받고 있는데, 이는 기존 기법들이 구현하기 어려웠던 사람 얼굴과 같은 고차원적인 특징을 추출하는 것에서 강점을 가지고 있기 때문이다. 다만 지금까지 딥러닝에 관한 연구들은 대부분 컴퓨터 분야에서 딥러닝의 성능을 높이거나 더 빠르게 훈련할 수 있는지에 맞춰져 있었다. 본 연구는 이러한 딥러닝의 강점을 이미지나 음성 데이터 외에도 비즈니스 데이터를 분석하는데 활용한다면 좋은 성과를 보일 수 있을 것이라 생각하였다. 이를 확인해 보기 위하여 한 인터넷 서점의 클릭스트림 데이터를 사용해 고객의 구매를 예측하는 모형들을 만들고 그 성능을 비교해 보기로 하였다. 실험을 위해 딥러닝 기법 중 DBM을 기반으로 한 구매 예측 모형들과, 성능 비교를 위한 나이브 베이지안(Naive Bayesian) 분류법을 기반으로 한 구매 예측 모형들을 작성하였다. 두 종류의 모형들의 예측 성능을 비교할 때에는 정밀도와 재현율을 동시에 고려할 수 있는 F1-score를 사용하였다. 실험 결과, 딥러닝을 기반으로 한 예측 모형이 나이브 베이지안 분류법을 기반으로 한 모형에 비해 더 좋은 예측 성능을 보여준 것을 확인하였다. 이를 통해 비즈니스 데이터를 분석할 때에 딥러닝 기법을 활용한다면 더 좋은 결과를 보일 수 있을 것이란 가능성을 볼 수 있었다. 그리고 위 실험에서 가장 좋은 성능을 보인 딥러닝 예측 모형을 하나 선택하여, 해당 모형이 학습한 구매자들의 페이지 이동 경로 특징을 시각화하여 분석하였다. 그 결과 구매자들은 검색 페이지와 홍보 페이지, 검색 페이지와 장바구니 페이지 사이를 왕복한다는 특징과, 계정 페이지에서 장바구니 페이지로 이동한다는 특징을 가지고 있다는 것을 확인할 수 있었다.|Recently, as big data becomes a buzz word, how to analyse enterprise data is a big issue in many companies. Among big data learning techniques, Deep Neural Network, also called Deep Learning has been received a spotlight. The reason is its better image and voice recognition capabilities. For example, it can extract high dimensional features from human faces, which was not an easy task for existing image recognition methodologies. Until now, studies on deep learning has been mainly focused on improving learning process speed and training in the field of computer engineering. This study is a trial to broaden the existing concerns over deep learning methods from voice or image recognition to business data analysis. To proceed this study, click stream data of an online bookstore has been collected to build up models to predict whether a customer purchase or not. Two methodologies are applied for the purpose. One was based on DBM, one of deep learning methodology, and the other one was Naive Bayesian classification method. This is to compare the prediction capabilities of the two methods, deep learning method and Naive Bayesian method. F1-score, which is enable to consider both of precision and recall values, has been applied to compare prediction capabilities of the models from two methods. The test results showed that the models based on deep learning method perform better than models based on Naive Bayesian method. This means deep learning techniques can be applied to business data analysis to provide better performance in practice. Following the results, we tried to visualize purchaser’s page movement route on the online bookstore learned by the best model based on deep learning method. The results of the visualization suggested that there are purchasers’ patterns to move between search page and promotion page, and, search page and shopping cart page. In addition, they also showed relatively weak tendency to move from account page to shopping cart page.; Recently, as big data becomes a buzz word, how to analyse enterprise data is a big issue in many companies. Among big data learning techniques, Deep Neural Network, also called Deep Learning has been received a spotlight. The reason is its better image and voice recognition capabilities. For example, it can extract high dimensional features from human faces, which was not an easy task for existing image recognition methodologies. Until now, studies on deep learning has been mainly focused on improving learning process speed and training in the field of computer engineering. This study is a trial to broaden the existing concerns over deep learning methods from voice or image recognition to business data analysis. To proceed this study, click stream data of an online bookstore has been collected to build up models to predict whether a customer purchase or not. Two methodologies are applied for the purpose. One was based on DBM, one of deep learning methodology, and the other one was Naive Bayesian classification method. This is to compare the prediction capabilities of the two methods, deep learning method and Naive Bayesian method. F1-score, which is enable to consider both of precision and recall values, has been applied to compare prediction capabilities of the models from two methods. The test results showed that the models based on deep learning method perform better than models based on Naive Bayesian method. This means deep learning techniques can be applied to business data analysis to provide better performance in practice. Following the results, we tried to visualize purchaser’s page movement route on the online bookstore learned by the best model based on deep learning method. The results of the visualization suggested that there are purchasers’ patterns to move between search page and promotion page, and, search page and shopping cart page. In addition, they also showed relatively weak tendency to move from account page to shopping cart page.