Research On Multi-objective Evolutionary Optimization Recommendation Algorithm Based On User Stratification

The classic recommendation system focuses on the accuracy of recommendation.With the increase of users' diversified needs,the diversity of recommendation results has attracted more and more attention.The accuracy and diversity of the recommendation always conflict with each other,and the traditional recommendation algorithms often ignore the difference of users' activity in the system.Therefore,starting from the characteristics of data distribution,this thesis proposes a hierarchical recommendation model for different users.The model adopts Fast non-dominated Sorting Genetic Algorithm as the carrier of multi-objective optimization algorithm.The average prediction score of the item measures the recommendation accuracy and the coverage rate measures the recommendation diversity,which are two optimization goals.From the initialization method of the algorithm,to the crossover operation and the mutation rate,the existing multi-objective recommendation algorithm are improved.The specific contents are as follows:(1)Improvement of the probability based multi-objective evolutionary algorithmTraditional recommendation algorithms perform better in recommendation accuracy,but currently it is difficult to meet user needs in terms of the diversity of recommendation results.In order to make the recommendation results more in line with the personalized recommendation,through reading the relevant literature of multi-objective recommendation algorithm in recent years,this thesis will improve the initialization and parameters of the existing probability based multi-objective evolutionary algorithm to obtain a better way of algorithm crossover and mutation.The experimental results verify that the enhanced the probability based multi-objective evolutionary algorithm can find better results in terms of recommendation accuracy and diversity.(2)Multi-objective recommendation algorithm based on user stratificationBased on the full understanding of the traditional recommendation algorithm and the experimental test and exploration of the probability based multi-objective evolutionary algorithm,it is found that in the complete structure of multi-objective recommendation algorithm,adding three different initialization schemes proposed in this thesis in the initialization stage can obtain the recommendation results focusing on different optimization objectives.Among them,the IT100 initialization scheme pays more attention to the recommendation accuracy of the recommendation results,the IT50-IC50 initialization scheme enables the multi-objective recommendation algorithm to find the recommendation results that take into account the compromise between the recommendation accuracy and the diversity,and the IC100 initialization scheme pays more attention to the diversity of the recommendation results.Finally,through three real data sets from different dimensions and different scenarios,and based on three traditional recommendation algorithms,the pareto frontier map obtained by the multi-objective recommendation algorithm has obvious distribution characteristics,which can confirm the effectiveness of the three initial schemes proposed in this thesis.(3)Recommendation scheme based on user stratificationAccording to the characteristics of the pareto frontier graph distribution of the three initialization schemes proposed in this thesis,as well as the distribution characteristics of user evaluation data,this thesis divides users into high,medium and low user sets according to the number of user evaluations on items.Then experiments are carried out on the real data of three different dimensions and different scenarios.The results of the two optimization objectives and two evaluation indicators corresponding to the three user sets are sorted out and the recommended scheme based on user stratification is summarized.The IT100 initialization scheme is more suitable for the user set with high evaluation times,the IT50-IC50 initialization scheme is more suitable for the user set with moderate evaluation times,and the IC100 initialization scheme is more suitable for the user set with low evaluation times.