Classifying and Extracting Information from Promotions for Demand Forecasting Using Topic Modelling with BERTopic.

抄録

In the e-commerce industry, sales promotions significantly influence demand. Extracting essential information from promotions, such as promotion type, duration, discount rate, target customers, and product categories, is a crucial factor of feature engineering for demand forecasting. However, promotional information is usually stored in text format, making it challenging to extract essential information for generating features. In this paper, we leverage the topic model BERTopic, which is robust in context analysis, to appropriately classify each promotion and extract necessary information for promotion feature generation based on the classification's topic. We conducted experiments on past data of a major Japanese e-commerce company. The result shows this method can achieve better performance compared to existing topic modelling baselines like LDA and NMF, and it was confirmed that relevant information for feature generation could be extracted based on the topics corresponding to each classification.