Personalized Recommendation Algorithm Based On Multi-Objective Optimization

With the rapid development of the Internet and information technology,the number of network products and Internet users is also growing rapidly.This has brought people from the previous “information lack”period into the era of “information overload”and “information explosion”.At the same time,various social media have been widely used in people's daily life and produced produce large amounts of data.Faced with these vast amounts of data and information,an urgent problem is how users can find useful information or objects quickly.Therefore,the recommender system has become an important technology in Internet services such as e-commerce,retail enterprises and Internet sales to help user reduce the burden of filtering information.In general,recommender systems model the users' interests by collecting and analyzing the historical behavior of users,thereby predicting the users' preference for other items,and then making personalized recommendations.However,there are still some bottlenecks in the field of recommendation that need to be resolved.Based on the idea of multi-objective evolutionary optimization,the thesis explores and studies the impact of various factors in the process of recommendation and the trade-off between different evaluations of the recommended results.The main research and innovation are as follows:(1)In the traditional recommendation algorithm,more consideration is given to improving the accuracy of the recommendation results,nonetheless merely considering accuracy may not be sufficient to recommend satisfactory items to the user.Other evaluation indicators of the recommender system,such as diversity,should also be taken into account when generating recommendation results.Therefore,from the perspective of user experience,a good recommendation system should take into account both accuracy and diversity.In this case,an easy to implement recommendation framework based on multi-objective immune optimization algorithm is proposed.It can generate a recommendation list with good diversity and novelty for the user without reducing the accuracy.The main idea of the proposed algorithm is to combine collaborative filtering with a multi-objective evolutionary algorithm to build a hybrid framework based on a cascade-parallel structure.Different from the existing recommendation algorithms,this new framework regards the task of the recommendation as a multi-objective optimization problem.The purpose is to solve the problem of the tradeoff between accuracy and diversity in recommendation lists.Firstly,a collaborative filtering algorithm is used to generate candidate solutions for subsequent steps.Then non-dominated neighbor immune algorithm is used to maximize a matching function and a diversity function and perform global search in the candidate data set.The algorithm generates a set of Pareto solutions for the active user that represent a set of recommendation lists.Experimental results demonstrate that the proposed algorithm is effective in solving the dilemma of tradeoff between accuracy and diversity in recommendation lists,and it can recommend more diverse and novel items to users than the traditional methods.(2)Referring to the optimization idea of the previous work,a recommendation based on parallel hybrid framework is proposed for personalized movie recommendation.In this model,the movie recommendation problem is constructed as a multi-objective optimization problem to solve the problem of conflicting accuracy and novelty in movie recommendations.In view of the fact that the actual accuracy of recommendation results cannot be known during off-line training,we adopt bipartite network probability propagation algorithm to calculate the prediction ratings of unseen movies,and construct the first objective function as an assessment of accuracy.The second objective function is the novelty of the recommendation list.The corresponding coding method is designed to encode the recommendation results of multiple users in one individual.Multiple recommendations can be provided simultaneously for all users in one run.The crossover and mutation operator that meet the requirements are designed to make evolutionary processes effectively.In order to reduce the computational complexity,a clustering technique based on similarity is firstly used to divide users into several clusters,so that users with similar ratings for the movie are in the same cluster and then run the algorithm in each cluster.The purpose of this algorithm is to recommend to the user a novel movie that he has not known before,without deviating from the user's previous viewing preferences completely.Finally,the experimental results indicate that the proposed algorithm can effectively balance the accuracy and novelty of recommendation list on the Movie Lens data sets.(3)Aiming at the shortcoming of slow convergence speed of Pareto front and a large number of iterations of algorithm in the previous chapter,a probabilistic crossover genetic operator which is more suitable for solving the multi-objective personalized movie recommendation in the real coding environment is designed.Through the multi-parent probabilistic inheritance,multiple recommended movies are passed on to offspring with greater probability.The new crossover operator is more consistent with the user selection characteristics when using the recommender system.Experimental results show that this algorithm can not only provide satisfactory recommendation lists to users,but also reduce the number of iterations and speed up the convergence compared with the algorithm in the previous chapter.(4)Social media can integrate the online data and people's off-line activities,making the Internet more realistic and thus widely used.Location recommendation in location-based social networks should not only consider the user's online behavior(such as check-ins,click,comments,etc.)and friends influence in the social networks,but also user behavior is influenced by geographic location factors off-line.In this case,this proposed algorithm puts forward a new strategy for the personalized location recommendation system to explore the potential points of interests for users in location-based social networks to overcome the drawbacks of traditional parallel weighted strategy which requires a large number of experiments to determine the weighting coefficients.In this new framework,the different factors in the personalized location recommendation are constructed into the different objective functions.A matching function of users' online check-ins and a density function which reflect geographical factors for a user to visit a new location off-line are optimized at the same time by using multi-objective evolutionary method.In this way,the process of adjusting the weight coefficient is avoided.In order to reduce the search space of optimization process and improve the timeliness of the algorithm,we also use the social information of users online.We do experiments with some real data sets which are collected from Gowalla and Brightkite to prove the effectiveness and efficiency of the proposed algorithm.Experimental results show that the proposed algorithm can provide personalized location recommendation for each user and has a good application prospect.