(a)Content-based Filtering (b) Item-based Collaborative Filtering (c) User-based Collaborative Filtering ples of different Filtering With item-based collaborative filtering as shown in 1. Implementing collaborative and content-based Figure Ib, if many users(User 2 and User3 in this methods separately and combining their predictions, example)like Item A and Item D, we assume that Item 2. Incorporating some content-based characteristics into A is highly correlated with Item D. If Userl likes Item a collaborative approach, A, he should also like Item D given the strong link of 3. Incorporating some collaborative characteristics into a association between items a and d content-based approach, and Item based Collaborative Filtering is quite common 4. Constructing a general unifying model that for current recommendation systems, which have been ncorporates both content-based and collaborative wildly used by Movie Lens and Yahoo. According to characteristics papers [3][4], item-based first analyze One thing in common across the above four fusion user-item matrix to identify relationships between approaches is that a large amount of input data is different items, and then use these relationships to required. Combining several filtering approache directly compute recommendations for users implies more than one aspect of input data might need With user-based collaborative filtering as shown in to be incorporated into the recommender system, such Figure Ic, the taste of User3 is very similar to that of s the users'information, item contents and users Userl because they have preferred Item A and Itemb ratings. A table summarizing what input data is in common. They are deemed to be like-minded users. below. d for the filtering approa Hence user 3 likes item d so will Userl Furthermore, it was argued in [5] that in real life the way in which two people are said to be similar is no Table 1: Input data required by different CF based solely on whether they have complimentary t data pinions on a specific subject, e. g, movie ratings, bu also on other factors, such as their background and User ratings(by a single user) Item-based CF lifestyles. Therefore, when doing the profile matching, User ratings(by multi issues such as age, gender and preferences of movie User-based CF genres must also be taken into account. User ratings(by multiple users)+ a novel framework for user-based collaborative Hybrid CF emographic inforn filtering is proposed in [6] that enables (e.g. GA-based CF) en contents recommendation by groups of closely related User ratings(by multiple users) individuals. By using the rating information from a dividual user in a group can be predicted s of the group of closely related users, unrated iter During filtering, the input data are processed as features by some heuristic algorithms. In some hybrid recommendation systems, all features are included. But 1. 1. Hybrid Model not all features contribute significantly to the quality of recommendation output; some are noises that bring Several recommendation systems(e.g. [7J[8) use a down the prediction accuracy. Moreover,every user hybrid approach by combining collaborative and places a different importance priority on each feature content-based methods, which helps to avoid certain These priorities are enumerated as referred to feature Different ways to combine collaborative and content based methods into a hybrid recommender system can mature users, then his feature weight for age would be be classified as follows. higher than other features(a) Content-based Filtering (b) Item-based Collaborative Filtering (c) User-based Collaborative Filtering Figure 1. Examples of different Filtering With item-based collaborative filtering as shown in Figure 1b, if many users (User2 and User3 in this example) like Item A and Item D, we assume that Item A is highly correlated with Item D. If User1 likes Item A, he should also like Item D given the strong link of association between Items A and D. Item based Collaborative Filtering is quite common for current recommendation systems, which have been wildly used by Movie Lens and Yahoo. According to papers [3][4], item-based techniques first analyze a user-item matrix to identify relationships between different items, and then use these relationships to indirectly compute recommendations for users. With user-based collaborative filtering as shown in Figure 1c, the taste of User3 is very similar to that of User1 because they have preferred Item A and Item B in common. They are deemed to be like-minded users. Hence User 3 likes Item D so will User1. Furthermore, it was argued in [5] that in real life the way in which two people are said to be similar is not based solely on whether they have complimentary opinions on a specific subject, e.g., movie ratings, but also on other factors, such as their background and lifestyles. Therefore, when doing the profile matching, issues such as age, gender and preferences of movie genres must also be taken into account. A novel framework for user-based collaborative filtering is proposed in [6] that enables recommendation by groups of closely related individuals. By using the rating information from a group of closely related users, unrated items of the individual user in a group can be predicted. 1.1. Hybrid Model Several recommendation systems (e.g. [7][8]) use a hybrid approach by combining collaborative and content-based methods, which helps to avoid certain limitations of content-based and collaborative systems. Different ways to combine collaborative and content￾based methods into a hybrid recommender system can be classified as follows: 1. Implementing collaborative and content-based methods separately and combining their predictions, 2. Incorporating some content-based characteristics into a collaborative approach, 3. Incorporating some collaborative characteristics into a content-based approach, and 4. Constructing a general unifying model that incorporates both content-based and collaborative characteristics. One thing in common across the above four fusion approaches is that a large amount of input data is required. Combining several filtering approaches implies more than one aspect of input data might need to be incorporated into the recommender system, such as the users’ information, item contents and users’ ratings. A table summarizing what input data is required for the filtering approach to work on is shown below. Table 1: Input data required by different CF Filtering approach Input data Content-based analysis Item contents + User ratings (by a single user) Item-based CF Item contents + User ratings (by multiple users) User-based CF User demographic information + User ratings (by multiple users) + Hybrid CF (e.g. GA-based CF) User demographic information + Item contents + User ratings (by multiple users) During filtering, the input data are processed as ‘features’ by some heuristic algorithms. In some hybrid recommendation systems, all features are included. But not all features contribute significantly to the quality of recommendation output; some are noises that bring down the prediction accuracy. Moreover, every user places a different importance priority on each feature. These priorities are enumerated as referred to feature weights. For example, if a mature user prefers to be given recommendations based on the opinions of other mature users, then his feature weight for age would be higher than other features