Computational Complexity Reduction for Factorization-Based Collaborative Filtering Algorithms

In order to draw users’ attention and to increase their satisfaction toward online information search results, search-engine developers and vendors try to predict user preferences based on users’ behavior. Recommendations are provided by the search engines or online vendors to the users. Recommendation systems are implemented on commercial and nonprofit Web sites to predict user preferences. For commercial Web sites, accurate predictions may result in higher selling rates. The main functions of recommendation systems include analyzing user data and extracting useful information for further predictions. Recommendation systems are designed to allow users to locate preferable items quickly and to avoid possible information overload. Recommendation systems apply data-mining techniques to determine the similarity among thousands or even millions of data. Collaborative-filtering techniques have been successful in enabling the prediction of user preferences in recommendation systems (Hill, Stead, Rosenstein, & Furnas, 1995, Shardanand & Maes, 1995). There are three major processes in recommendation systems: object data collections and representations, similarity decisions, and recommendation computations. Collaborative filtering aims at finding the relationships among new individual data and existing data in order to further determine their similarity and provide recommendations. How to define the similarity is an important issue. How similar should two objects be in order to finalize the preference prediction? Similarity decisions are concluded differently by collaborative-filtering techniques. For example, people that like and dislike movies in the same categories would be considered as the ones with similar behavior (Chee, Han, & Wang, 2001). The concept of the nearest-neighbor algorithm has been included in the implementation of recommendation systems (Resnick, Iacovou, Suchak, Bergstrom, & Riedl, 1994). The designs of pioneer recommendation systems focus on entertainment fields (Dahlen, Konstan, Herlocker, Good, Borchers, & Riedl, 1998; Resnick et al.; Shardanand & Maes; Hill et al.). The challenge of conventional collaborative-filtering algorithms is the scalability issue (Sarwar, Karypis, Konstan, & Riedl, 2000a). Conventional algorithms explore the relationships among system users in large data sets. User data are dynamic, which means the data vary within a short time period. Current users may change their behavior patterns, and new users may enter the system at any moment. Millions of user data, which are called neighbors, are to be examined in real time in order to provide recommendations (Herlocker, Konstan, Borchers, & Riedl, 1999). Searching among millions of neighbors is a time-consuming process. To solve this, item-based collaborative-filtering algorithms are proposed to enable reductions of computations because properties of items are relatively static (Sarwar, Karypis, Konstan, & Riedl, 2001). Suggest is a top-N recommendation engine implemented with item-based recommendation algorithms (Deshpande & Karypis, 2004; Karypis, 2000). Meanwhile, the amount of items is usually less than the number of users. In early 2004, Amazon Investor Relations (2004) stated that the Amazon.com apparel and accessories store provided about 150,000 items but had more than 1 million customer accounts that had ordered from this store. Amazon.com employs an item-based algorithm for collaborative-filtering-based recommendations (Linden, Smith, & York, 2003) to avoid the disadvantages of conventional collaborative-filtering algorithms.

Download Full-text