2.2. GA Recommender System Architecture Each User record We constructed a GA-based recommendation System as a prototype for conducting experiments. The Movie database experimental protocol is capable of gathering, disseminating, and using ratings from some users to Accumulate the no of times predict other users'interest in movies. Figure 3 shows appear for these genres the architecture of our system that can be deployed as online hybrid CF application implemented in GA gorithm. The process is divided into three phases Phase 1: Collect User Information For a new visitor, the system requests him to register an account as to collect his personal particulars as well Find maximum no. of genres as the ratings of a set of movies. Each time the user has watched or purchased a movie, he would be asked to Update "Prefer Film Type ate how good it is. Phase ll: Create User Profile features Figure 4. Flow chart of creating a feature For those movies the user rated we search for the corresponding genres and characters from the movie Phase III: GA Recommender database. In addition to using users' ratings, we try to This phase uses Ga functions to search for the e apture what types of items are most preferred by the appropriate recommendation by selecting and weighing ser. In the case of movie recommendation. the the features. Figure 5 shows the workflow of the preferred types of items are the movie genres such as recommender that the best subset can be gained from the following, Preferred Film Type, Preferred Director, the full data set by processing Feature Selection Preferred Actress, Preferred Actor, Preferred Producer, Feature Weighting and Recommendation Generation Preferred Writer, Preferred Editor and Preferred Language They form the additional 8 features to be stored in the FullDataSet ser profile(c f. Figure 2), which will be coded into the GA chromosome Accuracy abstract of what the user pree tings alone, an explicit feature weight that helps in the recommendation search process. This is particularly effective when the ratin records were in scarce during the early stage This is how it works on extracting the 8 features ith respect to each user, we sum up the total number (frequency) appears on each genre, the one with Algorithm maximum number is assumed to be most favoured by the user and marked as Prefer_ Film Type. If the sam igure 5. Workflow of GA Recommender lumber of genres exists, we would get the latest one as we suppose that the last record reflects the latest user Feature selection is the process that chooses an preference. A flowchart that shows how optimal subset of features according to a certain Prefer_ FiIm Type is computed is shown in Figure 4. The criterion. as there are thousands of features in the same logic applies to the other features can reduce the dimensionality. For example, in a users record, the most frequent eliminate noise and save search time ype is Musical; then we assume that the preferred film The first step in the GA Recommender phase is to ype of this user is Musical. In another example, if most prepare the features that are needed. For hybrid GA- of the movies that the user watched are found to be based CF. all the 37 features should be included directed by"Steven Spielberg,, then we assume the However, from our evaluation experiments, we found preferred director of the user is'Steven Spielberg that some of the user profile attributes are more This concept of extracting user preferences from the effective than the others. Good quality of majority of items that he frequented, should be generic recommendation can be produced by using a minimum enough to apply ther recommend 12 features out of ning, Ag han movie advisor Prefer FilmType, Prefer Director, Prefer Actress, Prefer2.2. GA Recommender System Architecture We constructed a GA-based Recommendation System as a prototype for conducting experiments. The experimental protocol is capable of gathering, disseminating, and using ratings from some users to predict other users' interest in movies. Figure 3 shows the architecture of our system that can be deployed as an online hybrid CF application implemented in GA algorithm. The process is divided into three phases: • Phase I: Collect User Information For a new visitor, the system requests him to register an account as to collect his personal particulars as well as the ratings of a set of movies. Each time the user has watched or purchased a movie, he would be asked to rate how good it is. • Phase II: Create User Profile features For those movies the user rated, we search for the corresponding genres and characters from the movie database. In addition to using users’ ratings, we try to capture what types of items are most preferred by the user. In the case of movie recommendation, the preferred types of items are the movie genres such as the following, Preferred Film Type, Preferred Director, Preferred Actress, Preferred Actor, Preferred Producer, Preferred Writer, Preferred Editor and Preferred Language. They form the additional 8 features to be stored in the user profile (c.f. Figure 2), which will be coded into the GA chromosome. Instead of relying on users’ ratings alone, an explicit abstract of what the user prefers often offers a good feature weight that helps in the recommendation search process. This is particularly effective when the rating records were in scarce during the early stage. This is how it works on extracting the 8 features: with respect to each user, we sum up the total number (frequency) appears on each genre, the one with maximum number is assumed to be most favoured by the user and marked as Prefer_FilmType. If the same number of genres exists, we would get the latest one as we suppose that the last record reflects the latest user preference. A flowchart that shows how Prefer_FilmType is computed is shown in Figure 4. The same logic applies to the other features. For example, in a user’s record, the most frequent type is Musical; then we assume that the preferred film type of this user is Musical. In another example, if most of the movies that the user watched are found to be directed by ‘Steven Spielberg’, then we assume the preferred director of the user is ‘Steven Spielberg’. This concept of extracting user preferences from the majority of items that he frequented, should be generic enough to apply on other recommendation domains than movie advisor. Figure 4. Flow chart of creating a feature • Phase III: GA Recommender This phase uses GA functions to search for the appropriate recommendation by selecting and weighing the features. Figure 5 shows the workflow of the recommender that the best subset can be gained from the full data set by processing Feature Selection, Feature Weighting and Recommendation Generation. Figure 5. Workflow of GA Recommender Feature selection is the process that chooses an optimal subset of features according to a certain criterion. As there are thousands of features in the database, selection can reduce the dimensionality, eliminate noise and save search time. The first step in the GA Recommender phase is to prepare the features that are needed. For hybrid GA￾based CF, all the 37 features should be included. However, from our evaluation experiments, we found that some of the user profile attributes are more effective than the others. Good quality of recommendation can be produced by using a minimum 12 features out of 37: Rating, Age, Gender, Occupation, Prefer FilmType, Prefer Director, Prefer Actress, Prefer