Simulation Study on E-commerce Recommender System Based on a Customer-Product Purchase-Matrix
This paper investigates the efficiencies of CF method and SVD-based recommender system for producing useful recommendations to customers when large-scale customer-product purchase data are available. Simulation experiments on synthetic transaction data sh
- PDF / 258,816 Bytes
- 10 Pages / 430 x 660 pts Page_size
- 26 Downloads / 223 Views
Abstract. This paper investigates the efficiencies of CF method and SVD-based recommender system for producing useful recommendations to customers when large-scale customer-product purchase data are available. Simulation experiments on synthetic transaction data show SVD-based recommender system yields a better performance than the CF method. Reduced product dimensionality from SVD may be more effective in generating a reliable neighborhood than CF method, and thereby it may improve the efficiency of recommendation performance. In applying SVD-based recommender system, the recommendation quality increases as the size of the neighborhood increase up to a certain point, but after that point, the improvement gains diminish. Our simulation results also show that an appropriate number of products for recommendation would be 10 in term of the error of false positives since around this point, the recall is not small, and both precision and F1 metric appear to be maximal. Even though the recommendation quality depends upon the dimension and structure of transaction data set, we consider such information may be useful in applying recommender system
1 Introduction The large E-commerce sites offer lots of products for sale. Choosing among so many options is challenging for customers. Recommender systems have emerged to resolve this problem. A recommender system receives information from customers about which products they are interested and recommends products they likely need. While a customer is at the E-commerce site, these systems find the customer’s behavior, develop a model of behavior and apply this model to recommend products to the customers. Collaborative filtering (CF) is the most successful system technology to date, and it is used in many of recommender system on the Web [15]. CF systems build a database of preferences for products by customers and then recommend products to a target customer base on the opinions of other customers. In large E-commerce sites, customer-product purchase data present that even active customers may have purchased well under 1% of the products. With association rules it is common to find rules having support and confidence higher than a user-defined * **
This paper was supported by Dong-A research fund in 2004. Corresponding author.
J.-W. Park, T.-G. Kim, and Y.-B. Kim (Eds.): AsiaSim 2007, CCIS 5, pp. 327–336, 2007. © Springer-Verlag Berlin Heidelberg 2007
328
C.-m. Kwon and S.-y. Kim
minimum. When the frequency of a customer’s purchased-product is low, even we may find a strong association rule for product recommendation, rules with very low support are often uninteresting since they do not describe sufficiently large populations. (From simulation results applying E-miner version 9.5 of SAS to synthetic product-purchase data, the minimum support is 0.4%). In applying CF, the problem space can be formulated as a matrix of users versus items, with each cell representing user’s rating on a specific item. This matrix is generally very sparse since each user rates only a small per
Data Loading...
