An Improved Recommendation Algorithm in Collaborative Filtering



Download 101.11 Kb.
Page1/8
Date07.04.2021
Size101.11 Kb.
  1   2   3   4   5   6   7   8




An Improved Recommendation Algorithm in Collaborative Filtering

Taek-Hun Kim, Young-Suk Ryu, Seok-In Park, and Sung-Bong Yang

Dept. of Computer Science, Yonsei University

Seoul, 120-749 Korea


{kimthun, ryu, psi93, yang}@mythos.yonsei.ac.kr




Abstract. In Electronic Commerce it is not easy for customers to find the best suitable goods as more and more information is placed on line. In order to provide information of high value a customized recommender system is required. One of the typical information retrieval techniques for recommendation systems in Electronic Commerce is collaborative filtering which is based on the ratings of other users who have similar preferences. However, collaborative filtering may not provide high quality recommendation because it does not consider customer’s preferences on the attributes of an item and the preference is calculated only between a pair of customers.

In this paper we present an improved recommendation algorithm for collaborative filtering. The algorithm uses the K-Means Clustering method to reduce the search space. It then utilizes a graph approach to the best cluster with respect to a given test user in selecting the neighbors with higher similarities as well as lower similarities. The graph approach allows us to exploit the transitivity of similarities. The algorithm also considers the attributes of each item. In the experiment the EachMovie dataset of the Digital Equipment Corporation has been used. The experimental results show that our algorithm provides better recommendation than other methods.





  1. Introduction

In Electronic Commerce it is not easy for customers to find the best suitable goods as more and more information is placed on line. In order to provide information of high quality a customized recommender system is required. A customized recommender system can help customers with fast searches for the best suitable goods by analyzing the customer’s preferences. A recommender system utilizes in general an information filtering technique called collaborative filtering(CF), even though there are several other techniques such as data mining and pattern recognition. CF is based on the ratings of other users who have similar preferences and is widely used for such recommender systems as Amazon.com and CDNow.com[1][2][3][4][5][14].

CF uses the Pearson correlation coefficient for evaluating the similarity between a pair of customer’s preferences, but it assumes that there must exist some items which have already been evaluated by both customers. On top of that, the Pearson correlation coefficient is calculated only between a pair of customers, so the prediction accuracy of CF may be degraded[1][2][5]. Another weak point of CF is that it never considers customer’s preferences on the attributes of an item.

There have been many researches to overcome these weak points of CF such as the K-Nearest Neighbor method and clustering. They are quite popular techniques for selecting neighbors who have similar preferences for certain items[6][11][12]. These techniques then predict customer’s preferences about the items through the results of evaluation on the same items by the neighbors[2][13]. The K-Means Clustering method is one of the clustering techniques and performs well on cluster numeric data sets in general[10][16]. These two techniques do improve prediction quality, but they did not overcome the fact that the Pearson correlation coefficient is calculated only between a pair of customers.

However, we assert that in order to predict customer’s preference more accurately, we should consider both high and low similarities of each customer, since a customer who has contradicting similarities may give valuable information in prediction. In this paper we present an improved recommendation algorithm for collaborative filtering. The algorithm uses the K-Means Clustering method to reduce the search space. It then utilizes a graph approach to the best cluster with respect to a given test user in selecting the neighbors with higher similarities as well as lower similarities. The graph approach allows us to exploit the transitivity of similarities. The algorithm also considers the attributes of each item.

To compare the performance of various recommendation methods including ours, the EachMovie dataset of the Digital Equipment Corporation has been used[8]. The EachMovie dataset consists of 2,811,983 preferences for 1,628 movies rated by 72,916 customers explicitly. The experimental results show that the proposed method provides better prediction accuracy than other methods.

The rest of this paper is organized as follows. Section 2 describes CF and some neighbor selection methods. Section 3 describes our graph mechanism to select neighbors. In section 4, our experimental results are given.





  1. Share with your friends:
  1   2   3   4   5   6   7   8




The database is protected by copyright ©essaydocs.org 2020
send message

    Main page