Mining changes in customer buying behavior for collaborative recommendations

Abstract

The preferences of customers change over time. However, existing collaborative filtering (CF) systems are static, since they only incorporate information regarding whether a customer buys a product during a certain period and do not make use of the purchase sequences of customers. Therefore, the quality of the recommendations of the typical CF could be improved through the use of information on such sequences.

In this study, we propose a new methodology for enhancing the quality of CF recommendation that uses customer purchase sequences. The proposed methodology is applied to a large department store in Korea and compared to existing CF techniques. Various experiments using real-world data demonstrate that the proposed methodology provides higher quality recommendations than do typical CF techniques, with better performance, especially with regard to heavy users.

Introduction

Recommender systems have been a recent focus of researchers and practitioners. Many companies hope that the use of recommender systems may be a means of surviving in a competitive environment. Recommender systems are particularly suited to retail business, as compared to other types of business, since retail markets are distinguished by several characteristics, such as repeated buying over a particular time horizon, large numbers of customers, and a wealth of information detailing past customer purchases (Schmittlein & Peterson, 1994).

In general, retail companies operate purchase databases in a longitudinal way, such that all transactions are stored in chronological order. A record-of-transaction database typically contains the transaction date for and the products bought in the course of, a given transaction. Usually, each record also contains a customer ID, particularly when the purchase was made using a credit card or a frequent-buyer card. Therefore, the purchasing sequence of a customer in the database who has made repeat purchases can easily be determined. This purchase sequence provides a description of the changes in a customer's preferences over time. However, in our domain of knowledge, there has been little study of the question of whether recommendations based on purchase sequences may be more accurate than existing recommender system predictions, based on non-sequential patterns. In this study, for the purpose of enhancing the quality of recommendations, we propose a new methodology that considers the way in which a customer's purchase sequence evolves over time.

To date, a variety of recommender systems (Balabanović and Shoham, 1997, Basu et al., 1998, Hill et al., 1995, Lawrence et al., 2001, Resnick et al., 1994, Sarwar et al., 2001, Shardanand and Maes, 1995) has been developed. Collaborative filtering (CF) has thus far been the most successful recommendation technique and has been used in a number of different applications, such as in the recommendation of web pages, movies, articles and products (Hill et al., 1995, Resnick et al., 1994, Shardanand and Maes, 1995). Collaborative filtering works by recommending products to a target customer through a process of identifying people who share similar preferences for products and looking for those products that target customers are most likely to purchase.

The recommendation processes of typical collaborative filtering in retail business consist of the following three steps (Sarwar et al., 2000, Sarwar et al., 2001).

(1) Customer profile construction

The purchase transaction records of a customer for a certain period are used to build a customer profile describing his or her likes and dislikes. The system represents the customer profile, A, such that a_ij is one if a customer i has purchased a product j, and zero, otherwise.

(2) Neighborhood formation

This is the most important part of the CF-based recommender systems. The system finds a set of customers, known as neighbors, who, in the past, have exhibited similar behavior (i.e. bought a similar set of products), through calculating the correlations among customers for the customer profile. A set of K customers is usually found (a neighborhood of size K), which is formed according to the degree of similarity between each of the neighbors and a target customer.

(3) Recommendation generation

Once a neighborhood is formed for a target customer, the system generates a set of the top N products that the target customer is most likely to purchase, by searching for products that the neighbors have purchased and that the target customer has not yet purchased.

As mentioned above, typical collaborative filtering is static, since it only makes use of information relating to whether the customer bought a product during a certain period, and does not use information on the purchase sequences of customers in the determination of the neighbors of a target customer. However, customers in retail business are not static, and their buying behavior changes over time. Thus, the quality of the recommendation of the typical CF could be further improved through the use of the available information on the purchase sequences of customers.

To illustrate the importance of this potential improvement in accuracy, let us consider the following example. Table 1 presents typical transaction records for a retail company and the customer profile is provided in Table 2. This example determines products that target customer ID011 is likely to buy, using transaction records for consumers CID001 through CID011.

Assume that the typical CF algorithm is used for solving the problem, that the neighborhood size (K) is three, and that the number of products recommended (N) is two. The typical collaborative filtering algorithm considers the correlation of preferences between the target customer and the other customers. All of the four customers, CID011, CID001, CID002 and CID003, commonly bought ‘Perfumes’ and ‘Skincare Products.’ The similarities between CID011 and the other three customers are equivalent; that is, the Pearson correlation coefficient is 0.67. Therefore, a recommender system based on the collaborative filtering algorithm will determine that CID001, CID002 and CID003 are the nearest neighbors and have the same preferences as the target customer. However, it is quite difficult to select two products that should be recommended to CID011, because CID001, CID002 and CID003 each purchased different additional products: ‘Dresses,’ ‘Shoes,’ and ‘Knits,’, respectively. In this case, two different products to be recommended to CID011 would have to be selected randomly. Accordingly, the recommendations would not necessarily be very appropriate for the preferences of the target customer.

Table 3 provides a customer profile rearranged with regard to each customer's transaction time. From the table, it can be seen that the purchase sequence of the target customer CID011 occurred in the order of Perfumes, followed by Skincare Products. Similarly, the purchase sequences of CID001, CID002 and CID003 were Perfumes→Skincares→Dresses, Perfumes→Shoes→Skincares, and Skincares→Perfumes→Knits, respectively. Assume now that the neighbors of the target customer are determined based on the purchase sequence of each customer. Customers who have purchase sequences similar to that of CID011 include CID 001 and CID002. The purchase sequences of CID011 and CID001 are exactly same, as both bought the same products during the same month; therefore, the nearest neighbor of the target customer CID011 is CID001. The next nearest neighbor is CID002. Therefore, the products recommendable to CID011 are ‘Dresses,’ and ‘Shoes.’ As mentioned above, when the past purchase sequences of each customer are available, this knowledge can be used to enhance the quality of the recommendations made.

However, if a customer profile rearranged by time is applied directly, this application results in a significant problem, namely, the sparsity problem. It is well known that a sparse data set, having few nonzero entries, decreases recommendation accuracy (Mobasher, Cooly, & Srivastava, 2000). The sparsity level, defined as 1- (nonzero entries/total entries), of Table 2 is 60% (=1−(22/55)), while that of Table 3 is 87% (=1−(22/165)). In general, the sparsity level of a typical data set in the field of recommendation is over 95%. Rearrangement of an input matrix by time results in an increased time dimension, as compared to that of a typical customer profile, and thereby makes the input data set sparser. Therefore, a new solution to the sparsity problem must be found.

In our research, we employ a clustering technique that groups the transactions of customers into homogeneous subgroups. The SOM (Self-Organizing Map) technique, which has been applied frequently of late, is used for clustering (Kohonen, 1990). With the aid of the SOM, all the transactions of customers may be allocated to a certain cluster and a cluster number imposed. The change in the cluster number resulting from each transaction determines a customer purchase sequence.

By observing changes in the cluster number of each customer over time, a buying sequence can be built for each customer. These buying sequences are potentially capable of predicting the future purchases of a target customer. However, since not all buying sequences have a statistical validity sufficient to guarantee the generalization of the prediction, the association rule mining technique may be used to extract the sequential patterns from the buying sequences (Agrawal, Imielinski, & Swami, 1993).