소규모 전자상거래를 위한 추천 시스템의 시간 차이에 따른 추천 효과 측정에 관한 연구

Due to the development of ICT and expansion of the internet, users face information overload problem. To mitigate the issue of information overload problem, studies proposed the recommender systems and continued to improve the performance of the said model, however, which have the following problems. (1) Most of the studies are conducted using large-scale datasets. Studies verify performances using large-scale datasets such as MovieLens, Pinterest, etc. Such experiments may not perform well in cases that do not use large datasets such as small e-commerce. (2) It is difficult to measure a duration effect of the recommender systems because using only fixed datasets for measuring performance. Most of the studies measure the performance by using Cross Validation (CV) or by randomly dividing the datasets. However, these approaches make it difficult to measure the duration effect of performance between time when user recommended items and time user purchased items. In this study, the performance of various recommender systems for small e-commerce compared to the same performance of the recommender systems proposed in previous studies. In addition, to effectively measure the duration effect, we design a multi-period datasets constructed by extending collection period of the test datasets for one week. As a result, we show that the difference between the results of using recommender systems in the previous studies and the results of the systems implemented in small e-commerce exists. Moreover, we also manifest that the duration effect can exist in previous recommender systems proposed in other studies. Ultimately, we conclude the paper with ideas for novel applications in small e-commerce.