Research On Multi-objective Commercial Recommendation Systems Based On Evolutionary Algorithm

Nowadays,personalized commercial recommendation systems driven by big data have been widely applied in large Internet companies such as Amazon and e Bay.Their main purpose is to maximize profits by using the recommendation system,which is closely related to the accuracy of recommendation,the profit value of recommended items and the sales of long tail items.However,both the traditional recommendation algorithm and the existing multi-objective recommendation algorithm only consider the user's experience,but do not consider the profit on which the company is most concerned.In addition,the existing multi-objective recommendation algorithm based on evolutionary algorithm suffers from hypervolume fluctuation and too high runtime overhead.In view of the above problems,the main research and innovation work of this thesis is as follows:(1)A multi-objective business recommendation model integrating expected total profit and novelty is constructed.In this model,the expected total profit function comprehensively considers the accuracy of the recommendation results and the profit value of the recommended items.Increasing the expected total profit function value can improve the true profit of the recommendation list.The novelty function is defined as the proportion of new items and long-tail items in the recommended list.Increasing the novelty function value can improve the visibility of long-tail items and alleviate the cold start problem.(2)A novel hybrid probabilistic multi-objective evolutionary algorithm(HP-MOEA)is proposed to optimize the two conflicting objectives in(1).By integrating the classic NSGA-II and SMS-EMOA frameworks,HP-MOEA makes complementary advantages and enables the algorithm to increase the hypervolume rapidly and stably during the whole iteration process.In addition,this thesis proposes a new probability crossover operator,which makes the excellent parent individual have greater genetic probability.Hence,the possibility of inheriting the excellent genes is increased,and it is more likely for the population to produce excellent offspring.HP-MOEA algorithm is compared with MOEA-EPG,PMOEA,MOEA-PROBS,CF and MF in three benchmark data sets.The experimental results show that HP-MOEA outperforms other algorithms.(3)A multi-objective recommendation model based on parallel evolutionary algorithm is constructed for the first time,and a novel parallel multi-objective evolutionary algorithm based on collaboration(CC-MOEA)is proposed.CC-MOEA makes up for the defects of the traditional master-slave model and island model through a kind of novel island model,which reduces the time overhead of the algorithm significantly and ensures the hypervolume not lower than that of the serial algorithm.The experimental results show that CC-MOEA is superior to other algorithms in hypervolume,and its time cost is much lower than that of the serial evolutionary algorithm.