Handling sequential pattern decay: Developing a two-stage collaborative recommender system

Abstract

This study proposes a sequential pattern based collaborative recommender system that predicts the customer’s time-variant purchase behavior in an e-commerce environment where the customer’s purchase patterns may change gradually. A new two-stage recommendation process is developed to predict customer purchase behavior for the product categories, as well as for product items. The time window weight is introduced to produce sequential patterns closer to the current time period that possess a larger impact on the prediction than patterns relatively far from the current time period. This study is the first to propose time-decaying sequential patterns within a collaborative recommender system. The experimental results show that the proposed system outperforms the traditional collaborative system using a public food mart dataset and a synthetic dataset.

Introduction

Recommender systems are developed to deal with information overload and provide personalized recommendations, content and services to users (Adomavicius and Tuzhilin, 2005, Rashid et al., 2002). These software systems have been applied in many areas including e-commerce, news, advertisement, document management and e-learning. In the e-commerce application, recommender systems can potentially turn browsers into buyers by providing personalized shopping information that interests the customer, thus improving cross-sales and attaining customer loyalty (Schafer et al., 1999, Schafer et al., 2001).

Recommender systems make recommendations using three basic steps: acquiring preferences from the customers’ input data, computing recommendations using proper techniques and presenting the recommendations to customers (Wei et al., 2007). Customer preferences can be captured using four types of customer data: demographic data, rating data, browsing pattern data and transaction data (Wei et al., 2007). Transaction data provides sequences of purchased items and can be used to predict future customer purchasing preferences. In the rapid changing business environment, product life cycle has become shorter with customer buying behaviors changing more frequently. Therefore, some researchers (Cho et al., 2005, Min and Han, 2005) applied the customer’s time-variant buying sequence as a dynamic customer profile to improve recommender system performance. The periodic transaction data for a customer can be called the “dynamic customer profile”, which was originated by Cho et al. (2005). Traditional recommender systems incorporate customer transaction data from only a single temporal period, known as the static customer profile. These systems do not make use of customer purchase sequences (Cho et al., 2005). The dynamic nature of a customer’s purchase sequences can improve the recommendation quality of recommender systems.

Although the dynamic customer profile provides a basis for generating recommendations, in an environment in which the customer gradually changes purchase patterns (interests), the transaction data close to the current temporal period are usually more important than that temporally far from the current period. This study introduces time window weights to provide higher importance on the patterns generated in more recent time periods to improve the recommendation quality based on the dynamic customer profile.

In addition to the time-variant dynamic customer profile, predicting the customer’s purchasing using various techniques plays an important role in designing a recommendation system. Many different approaches categorized as heuristic-based or model-based (Adomavicius and Tuzhilin, 2005) have been applied to the problem of making accurate recommendations. These techniques include nearest neighbor algorithms, Bayesian analysis, neural networks and association rules. Association rules have been used for many years in e-commerce merchandising, both to analyze preference patterns across products and recommend products to consumers based on other products that they have selected (Schafer et al., 2001). The association rule expresses the relationship that one product is often purchased along with other products.

Recently, temporal data mining – sequential data mining – was introduced by Agrawal and Srikant (1995). A sequence database consists of a series of sequences. Each sequence is composed of several transactions (also called events) sorted in time ascending order. Sequential pattern mining is the mining of frequently occurring ordered events or sequences (Han and Kamber, 2006). Unlike association rules, sequential patterns may suggest that a consumer who buys a new product in the current time period is likely to buy another product in the next time period. Sequential pattern mining has been applied to sales promotions, targeted marketing, production processes, web access pattern analysis, network intrusion detection and DNA sequence analysis. In e-commerce, sequential patterns are useful for personalizing product recommendations and product related advertisements to improve customer satisfaction. There has not been any research on a collaborative system that uses sequential purchase patterns based on a multi-period time-variant customer profile. Researchers have addressed time decay in the importance of previous purchases (Cho et al., 2005, Min and Han, 2005), but until now, no system has yet been developed that applies time decay to sequential patterns.

Collaborative recommender systems recommend items that are similar to items already purchased by people in the same preference group. The collaborative recommender systems have data sparsity limitation problems. That is, the quantity of products that customers may have purchased is relatively few, compared with large product sets (e.g. Amazon.com). In this situation, the recommender systems may be unable to make any product recommendations for a particular customer, as these customers may have purchased too few common products, thus leading to poor recommendations. This is known as the data sparsity problem (Adomavicius and Tuzhilin, 2005, Sarwar et al., 2000). To reduce this limitation, this study uses a two-stage recommendation: predicting the top-M product categories according to sequential patterns and deriving the top-N product items among the predicted product categories. In the first stage, the sequential purchasing patterns for the product categories are established in advance. The target customers’ possible favorite product categories can then be predicted based on the sequential purchase patterns. Focusing on product category prediction has the merits of reducing the data dimension, increasing the scalability and reducing the data sparsity limitation. In the second stage, the product items that the target customers are likely to purchase are generated using the top-N approach based on the categories predicted in the first stage. The proposed two-stage recommender system based on the customers’ time-variant customer profile can improve the recommendation quality.

The rest of this paper is organized as follows. Section 2 introduces related works. Section 3 describes the framework of the proposed sequential pattern based collaborative recommender system. Section 4 examines the performance of the proposed CF system using a public food mart dataset and a simulated dataset. Section 5 provides conclusions.