AN ADAPTIVE RECOMMENDATION SYSTEM 175 the confidence values. These confidence values are corrected by the gA-based learning mechanisms using users' future navigation behaviors. Our experimental results indicate a significant increase in the accuracy of recommendation results due to the integration of the proposed learning mechanism. The remainder of this paper is organized as follows. Section 2 summarizes the related work. In Section 3, we provide an overview on the functionality of Yoda In Section 4, we discuss the detailed design of Yoda. Section 5 depicts the learning mechanism of Yoda. Section 6 describes the results of our evaluations as well as the details of the system implementation and our benchmarking method Section 7 concludes the paper 2. Related work Recommendation systems are designed either based on content-based filtering or collab- orative filtering. Both types of systems have inherent strengths and weaknesses, where content-based approaches directly exploit the product information, and the collaboration filtering approaches utilize specific user rating information Content-based filtering approaches are derived from the concepts introduced by the In- formation Retrieval (IR)community. Content-based filtering systems are usually criticized or two weaknesses 1. Content limitation: IR methods can only be applied to a few kinds of content, such as text and image, and the extracted features can only capture certain aspects of the content Over-specialization: Content-based recommendation system provides recommendations merely based on user profiles. Therefore, users have no chance of exploring new items at are not similar to those items included in their profiles The collaborative filtering(CF) approach remedies for these two problems. Typically, CF-based recommendation systems do not use the actual content of the items for recommen- dation. Moreover, since other user profiles are also considered, user can explore new items The nearest-neighbor algorithm is the earliest CF-based technique used in recommendation systems [20, 28]. With this algorithm, the similarity between users is evaluated based on their ratings of products, and the recommendation is generated considering the items visited by nearest neighbors of the user. In its original form, the nearest-neighbor algorithm uses a two-dimensional user-item matrix to represent the user profiles. This original form of CF-based recommendation systems suffers from three problems 1. Scalability: The time complexity of executing the nearest-neighbor algorithm grows linearly with the number of items and the number of users. Thus, the recommendation system cannot support large-scale applications such as Amazon. com, which provides more than 18 million unique items for over 20 million users 2. Sparsity: Due to large number of items and user reluctance to rate the items, usually the profile matrix is sparse. Therefore, the system cannot provide recommendations for some users, and the generated recommendations are not accurate.AN ADAPTIVE RECOMMENDATION SYSTEM 175 the confidence values. These confidence values are corrected by the GA-based learning mechanisms using users’ future navigation behaviors. Our experimental results indicate a significant increase in the accuracy of recommendation results due to the integration of the proposed learning mechanism. The remainder of this paper is organized as follows. Section 2 summarizes the related work. In Section 3, we provide an overview on the functionality of Yoda. In Section 4, we discuss the detailed design of Yoda . Section 5 depicts the learning mechanism of Yoda. Section 6 describes the results of our evaluations as well as the details of the system implementation and our benchmarking method. Section 7 concludes the paper. 2. Related work Recommendation systems are designed either based on content-based filtering or collab￾orative filtering. Both types of systems have inherent strengths and weaknesses, where content-based approaches directly exploit the product information, and the collaboration filtering approaches utilize specific user rating information. Content-based filtering approaches are derived from the concepts introduced by the In￾formation Retrieval (IR) community. Content-based filtering systems are usually criticized for two weaknesses: 1. Content limitation: IR methods can only be applied to a few kinds of content, such as text and image, and the extracted features can only capture certain aspects of the content. 2. Over-specialization: Content-based recommendation system provides recommendations merely based on user profiles. Therefore, users have no chance of exploring new items that are not similar to those items included in their profiles. The collaborative filtering (CF) approach remedies for these two problems. Typically, CF-based recommendation systems do not use the actual content of the items for recommen￾dation. Moreover, since other user profiles are also considered, user can explore new items. The nearest-neighbor algorithm is the earliest CF-based technique used in recommendation systems [20, 28]. With this algorithm, the similarity between users is evaluated based on their ratings of products, and the recommendation is generated considering the items visited by nearest neighbors of the user. In its original form, the nearest-neighbor algorithm uses a two-dimensional user-item matrix to represent the user profiles. This original form of CF-based recommendation systems suffers from three problems: 1. Scalability: The time complexity of executing the nearest-neighbor algorithm grows linearly with the number of items and the number of users. Thus, the recommendation system cannot support large-scale applications such as Amazon.comTM, which provides more than 18 million unique items for over 20 million users. 2. Sparsity: Due to large number of items and user reluctance to rate the items, usually the profile matrix is sparse. Therefore, the system cannot provide recommendations for some users, and the generated recommendations are not accurate