Purchasing behavior prediction model for real stores considering product cluster and time-series patterns

Yuiko Sakuma, Masahiro Yoshida, Hiroaki Nishi

Research output: Contribution to journalArticlepeer-review

Abstract

Recently, cashless payment services have been rapidly introduced, increasing the demand for recommender systems for real stores. Recommender systems have been well-studied and successful in many Internet services. While factors such as the diversity and coverage of products should be considered, the prediction accuracy of user purchasing behaviors is the most crucial factor in a recommender system. This research proposes a purchasing behavior prediction model for real stores by using purchase history data collected from a supermarket in Saitama City, Japan, as a part of a smart city project. Time-series patterns in purchasing behavior indicate the sales characteristics of supermarkets, where daily products are predominately purchased. For example, some users repeatedly purchase specific products in a certain period or purchase different products from their previous visit. Many previous studies have only considered the compatibility between the features of users and items; however, this work also models the time-series patterns of purchasing behavior. The proposed prediction model adopts ensemble learning, in which weak learners learn the two factors (the compatibility between users and items) and time-series patterns of purchasing behavior separately. Moreover, product names in supermarkets often contain meta-information of the product alongside the item name. For instance, the product name “three tomatoes from Gunma Prefecture” includes information about the quantity and manufacturer as well as the item name “tomato”. However, products that are the same item but have different product names are likely to be purchased together in the context of purchasing activity, thus deteriorating the prediction accuracy. Therefore, a product classification method is proposed to categorize the products for each item. The experimental results show 3.2% and 10.6% improvements in the precision-recall area under the curve (PR-AUC) for the proposed method compared to models that consider only the compatibility between users and items and the time-series patterns of purchasing behavior, respectively.

Original languageEnglish
Pages (from-to)471-482
Number of pages12
JournalIEEJ Transactions on Electronics, Information and Systems
Volume141
Issue number3
DOIs
Publication statusPublished - 2021 Mar 1

Keywords

  • Purchasing behavior prediction
  • Recommender system
  • Time-series modeling

ASJC Scopus subject areas

  • Electrical and Electronic Engineering

Fingerprint Dive into the research topics of 'Purchasing behavior prediction model for real stores considering product cluster and time-series patterns'. Together they form a unique fingerprint.

Cite this