Recommendation system development for fashion retail e-commerce

Highlights

• Develop a real-world recommender for a large Korean fashion company.

• Extend item-based CF by reflecting fashion industry-specific characteristics.

• Show time and purchase intention are important in recommending fashion items.

Abstract

This study presents a real-world collaborative filtering recommendation system implemented in a large Korean fashion company that sells fashion products through both online and offline shopping malls. The company’s recommendation environment displays the following unique characteristics: First, the company’s online and offline stores sell the same products. Second, fashion products are usually seasonal, so customers’ general preference changes according to the time of year. Last, customers usually purchase items to replace previously preferred items or purchase items to complement those already bought. We propose a new system called K-RecSys, which extends the typical item-based collaborative filtering algorithm by reflecting the above domain characteristics. K-RecSys combines online product click data and offline product sale data weighted to reflect the online and offline preferences of customers. It also adopts a preference decay function to reflect changes in preferences over time, and finally recommends substitute and complementary products using product category information. We conducted an A/B test in the actual operating environment to compare K-RecSys with the existing collaborative filtering system implemented with only online data. Our experimental results show that the proposed system is superior in terms of product clicks and sales in the online shopping mall and its substitute recommendations are adopted more frequently than complementary recommendations.

Introduction

Recently, e-commerce has become an important channel for many retail businesses. The eMarketer (2017), an e-business marketing company, estimates worldwide retail e-commerce sales will increase from $2.29 trillion in 2017 to $4.48 trillion by the end of 2021. In spite of its success, e-commerce has a significant market limitation. While there is staff available to assist customers in offline stores, there is no staff to help buyers in online stores. In order to overcome this limitation, online stores provide various features, for instance, a “search and directory” to assist customers. Although these services can enhance purchase experience, online customers can only take advantage of them if they use them.

Recommendation systems are an innovative solution that overcomes the limitations of e-commerce services. Recommendation systems use customer behavior and information, and product information to identify customer preferences, and proactively suggest products that they are likely to buy. Many studies have been conducted to develop such recommendation systems and many practical systems have been successfully implemented in various businesses (Choi et al., 2012, Koren, 2009a, Linden et al., 2003, Wei et al., 2016).

Recommendation systems used in e-commerce have been developed to reflect unique domain characteristics (Portugal et al., 2015, Schafer et al., 1999, Sivapalan et al., 2014, Wang and Zhang, 2013, Zhao et al., 2015). This study aims to develop a recommendation system for a company – referred to as Company K in the following discussion – that sells fashion products through an online shopping mall as well as through offline shopping outlets. Company K has an average of 5 million members and sells around 40,000 products per year in the online shopping mall. There are around 1.5 million clicks and around 10,000 transactions per month. Company K also operates about 1300 offline stores in Korea and sells around 20,000 products per year.

Company K possesses the following unique operation environment:

• When purchasing fashion products, customers buy to replace or supplement their previous purchases, or preferred products.

• Demand for fashion products generally decreases over time due to seasonal changes. In general, people buy fashion products appropriate to the current season. Such a pattern in purchases is found frequently in fashion items, while purchases of other products like books and music do not display any significant relationship with the changing of the seasons.

• Fashion products can be sold in both online and offline stores. Most previous studies focus mainly on online stores. However, online and offline stores usually sell the same fashion products. Usually, customers first decide on potential purchases at online stores and then purchase them online.

Company K has already recognized that recommendation systems are a key success factor for its business and has used a system that employs a conventional item-based collaborative filtering for its online shopping mall. However, Company K wants a new recommendation system to be developed to reflect the fashion industry-specific characteristics discussed above. In the development of this new recommendation system, we address the following recommendation requirements.

First, the recommendation system should reflect the decline in preference for fashion products over time. Previous temporal recommendation studies assume that the intensity of preference decreases as time passes (Campos et al., 2013, Ding and Li, 2005, Ding et al., 2006, Hong et al., 2012, Koren, 2010, Larrain et al., 2015, Lathia et al., 2010, Xu et al., 2016). That is, these studies assume that recent preference-indicating behaviors, such as clicks or purchases for the same product, reflect stronger preferences than older ones. However, this study focuses on the decline in preference for fashion products which occurs over time following their release.

Second, the recommendation system needs to combine both offline customer preference data and online customer preference data. Purchases in offline shopping malls reflect the preferences of offline customers. Therefore, combining offline purchase data with that of online customers can improve the performance of online recommendation. Some researchers emphasize that online and offline information can help predict customer preferences, but have not applied this finding to the development of recommendation systems. (Cheema and Papatla, 2010, Dzyabura et al., 2016). Only a few studies consider the problem of integrating online and offline shopping mall data into recommendation systems (Adomavicius and Tuzhilin, 2001, Cantador et al., 2015, Kim et al., 2016, Nilashi et al., 2014a). The most important reason seems to be that it is difficult to find a domain where it is important to integrate both online and offline preference data. However, fashion products are sold through both online and offline stores and therefore preference data can be collected from these two stores. This study therefore aims to propose a way to combine online and offline preferences for recommending fashion products.

Third, customer purchase intent should be reflected by the recommendation system. When buying a product, the customer chooses an item that can be used with, or an item that replaces something that he or she had previously preferred. In this paper, the former is called a complementary product, and the latter is called a substitute product. This study proposes a method that recommends complementary products and substitute products separately using product category information.