Journal of Convergence Information Technology Volume 5. Number 8 October 2010 nearest neighbors of u 4. Multi-criteria collaborative filtering The CF technique has been successfully applied to solve the recommendation problem with single riterion. When the problem is extended to be multiple conflicting criteria, new techniques are needed in order to effectively incorporate the multi-criteria rating information into the recommendation process. Particularly, in this section describe how to incorporate the multi-criteria rating information into the CF process following the aggregation function-based approach proposed in [13] Aggregation Module(AG), and the Recommendation Model(RC)as shown in Figure dule(MP), the The proposed system consists of three main modules: the Multi-Criteria Prediction Mo The overall system can be viewed as a blackbox which takes, as input, the ratings matrix, and produces, as output, a recommendation list. The ratings matrix R contains the rating set ru drg, i,,..,"y) standing for the ratings of the user u, for item iy, where rg represents the overall rating value and i represents the rating value for criterion k. The MP module uses the single-criterion CF algorithm to produce the recommendation list for each individual criterion. The AG module computes the user's preference for each criterion in terms of weight using GA algorithms. The rC module produces an overall recommendation list based on the results from MP and AG modules Multi-Criteria Agge空ator Prediction Module TOp-N Recommendation List Figure 1. System architecture 4. 1. Aggregation module In the multi-criteria rating environment, different people may place different emphases on these interrelated features. The goal of AG is to find the relationship between the overall rating and the underlying multi-criteria ratings for each user. More specifically, given the ratings data of a user, AG computes his/her preference model in terms of feature weights using GAs In this study, each chromosome x in the ga process is expressed as a set of feature weights where W, =(wI, w2, w,, w4)with 4 genes. Each gene is represented as a movie feature weight and encod with 8 bits. The Ga begins with a random population of chromosomes. For an active user l,, each chromosome is assigned alternately and tested by the fitness function. The fitness function measures the prediction accuracy of products using the weights as defined in the current chromosome (5) Fitness(x)=l-Journal of Convergence Information Technology Volume 5, Number 8, October 2010 nearest neighbors of ua . 4. Multi-criteria collaborative filtering The CF technique has been successfully applied to solve the recommendation problem with single criterion. When the problem is extended to be multiple conflicting criteria, new techniques are needed in order to effectively incorporate the multi-criteria rating information into the recommendation process. Particularly, in this section, we describe how to incorporate the multi-criteria rating information into the CF process following the aggregation function-based approach proposed in [13]. The proposed system consists of three main modules: the Multi-Criteria Prediction Module (MP), the Aggregation Module (AG), and the Recommendation Model (RC) as shown in Figure 1. The overall system can be viewed as a blackbox which takes, as input, the ratings matrix, and produces, as output, a recommendation list. The ratings matrix R contains the rating set rij = { k ijijij ,,, rrr 10 } standing for the ratings of the user ui for item ij, where 0 ij r represents the overall rating value and k ij r represents the rating value for criterion k. The MP module uses the single-criterion CF algorithm to produce the recommendation list for each individual criterion. The AG module computes the user’s preference for each criterion in terms of weight using GA algorithms. The RC module produces an overall recommendation list based on the results from MP and AG modules. Figure 1. System architecture 4. 1. Aggregation Module In the multi-criteria rating environment, different people may place different emphases on these interrelated features. The goal of AG is to find the relationship between the overall rating and the underlying multi-criteria ratings for each user. More specifically, given the ratings data of a user, AG computes his/her preference model in terms of feature weights using GAs. In this study, each chromosome x in the GA process is expressed as a set of feature weights where ),,,( 4321 = wwwwW xxxxx with 4 genes. Each gene is represented as a movie feature weight and encoded with 8 bits. The GA begins with a random population of chromosomes. For an active user ua , each chromosome is assigned alternately and tested by the fitness function. The fitness function measures the prediction accuracy of products using the weights as defined in the current chromosome. l rp itness xF l j ∑ jaja = − −= 1 00 1 ,, )( (5) 129