Long tail problem and implicit user feedback in e-commerce marketplace recommender systems Bartlomiej Twardowski Warsaw University of Technology, Institute of Computer Science, Nowowiejska 15/19, 00-665 Warsaw, Poland, e-mail: B.Twardowski@ii.pw.edu.pl Abstract. This article describes how to cope with long-tail problem and lack of explicit user feedback in online e-commerce marketplaces. Presented approach describes flexible way of classification new items to meta-items. Then, using meta-item well- know, state of the art recommendation algorithms can be used. User preferences are implicitly calculated based on user behavior on marketplace web-site or application. User event model with aging is proposed. Keywords: recommender system, clustering, big data, collaborative filtering, implicit user feedback Introduction Online marketplaces are one of the fastest growing e-commerce business branches. Big companies like eBay, Amazon and Naspers providing online platforms where buyers and sellers exchange goods, like in the old fash- ion marketplaces. However, in the online web trading systems resources needed to successfully sell good are far less lower than in standard merchandising. What leads to exponential new items-information growth and long tail phenomenon[2]. In such environments, searching in vast amount of goods is hard. Still, non-popular things in the long tail brings a lot of income. Thus, recommender systems are the core services of online e-commerce platforms. The are many recommendation algorithms based on neighborhood, collaborative filtering and content. But in online market- places the main problems with applying existing recom- mender systems are: high data sparsity, different data na- tures between product categories, description of items is in a free form text with pictures, users does note rate items, vast of items are in long tail - unpopular and unique. In this paper an approach to cope with this selected marketplaces problem is presented. Related Work Recommender systems is a rapidly developing field of science. There is a lot of academic and industrial work on recommender systems. From simple collaborative filtering base with item-to-item approach in Amazon[8] many new, well-performing algorithms emerged[3] which are used in most e-commerce application with success[5]. One of the most important researches in area of rec- ommender systems was connected to the Netflix Prize com- petition. Competition has fueled the development of new recommendation techniques like matrix factorization col- laborative filtering[11, 7]. Long tail problem in recommender systems was re- Figure 1: Main stages in recommendation with items classification and implicit feedback processing. search in one of the biggest e-commerce company - eBay[10]. Large scale and items temporariness was also ad- dressed in the eBay researchers work[6, 4]. Lack of explicit user feedback and modeling user pref- erences with aging model was also subject of researchers in- terest [1]. Proposed Approach In this work available, well-performing in e-commerce recommendation algorithms will be used(collaborative filtering[8], ALS[11]). Research emphasis is place on the data preparation stages: items classification and ratings, pro- cessing user events. In order to recommend new items for the user, rec- ommender system should know its preferences. What user likes or dislikes. That is the basic of every recommender. In the marketplace platform, there is no explicit user feedback like rating. The main entities are: Item - marketplace offer in a form of auction or retail, User - which can be buyer and seller on marketplace platform, User Event - user actions made on marketplace e.g. clickstream from online website or interactions in mobile application. On the figure 1 main stages in recommendation pro- cess are presented. First stage is data collection of main enti- ties: items, users and user events. Other supported data can be collected too, e.g. product description for specific item The challenges of contemporary science. Theory and applications, ISBN 978-83-935118-1-5 33