A Study on New Product Recommendation Using Multi-Label CVAE for Fresh Flowers
Aya Kitasato, Gendo Kumoi, Masayuki Goto · 2021
In recent years, it has become very popular to use purchase history data on e-commerce sites for marketing measures to increase sales. Under such a situation, this paper considers measures using the purchase history data of a company providing delivering services of fresh flower products through an e-commerce site. This site deals mainly with fresh flowers, and the majority of items are purchased for gifts. The demands of flower gifts are usually strongly related with certain events, such as birthday, Mother’s day, opening celebration, etc. Since each customer often makes purchase only at certain event when purchasing a flower gift, and it is important to encourage them to make purchases at other events from marketing viewpoint. In addition, the appearance of fresh flowers is important, so product recommendation with product images is necessary. It is relatively easy to develop floral gifts because they consist of certain patterns such as types of fresh flowers and shapes such as bouquets. However, there is no development of product which quantitatively uses purchase history information, The purpose of this research is, therefore, to generate product images that are preferred by customers in another event, considering the characteristics of product images purchased in individual event, where it is also possible to create new product images that are not contained in existing items. The proposed model is based on Conditional Variational Auto Encoader (CVAE) and can generate image outputs by inputting product images as multi-labels of events and attributes such as age and gender of customers that greatly affect product selection. Then, after learning a generator model, we consider to analyze what kinds of new products a customer with certain attributes who purchased at certain event would newly prefer at other events by changing the labels. Furthermore, in this study, we demonstrate the validity of the model by analyzing an actual data set.