Personalized Recommendation Based On User Interest And Time Information

The development of information technology and the internet has led people into the era of artificial intelligence.In this era,the Internet has recorded vast amounts of information.Every day,people actively or passively receive a large amount of information.This is the so-called "information overload." In order to solve this problem,the researchers proposed many solutions.For example,classified catalogs,searching engines,and recommender systems.All of them are tools that help people quickly find useful information.The recommender system differs from the previous two tools in that it can help users find information of interest without the user's definite demand.Based on the empirical data of the recommender system,we analyze the user behavior pattern,which is used to improve the recommendation algorithm.In addition,the influence of data size and data time on the accuracy of the recommendation algorithm is studied.The main research work is as follows:(1)We study the personalized recommendation system based on user interest model.By studying the correlation between the user activity and object popularity of the four empirical recommendation system datasets(Netflix,SMovieLens,LMovieLens and RYM),the users with high activeness are found to have a stronger preference to unpopular objects.Users with low activeness show a wider interest.We introduce the user interest pattern into the personalized recommendation algorithm,and propose a general function form to improve seven recommendation algorithms based on different similarity functions,and finally get four new recommendation algorithms.We test the four new recommendation algorithms on the above four empirical datasets,and find that due to the high diversity and low accuracy of the Heat Conduction(HC),the diversity of the algorithm is reduced for the new algorithm,but its accuracy is greatly improved.The accuracy and diversity of the rest of the proposed algorithms are better than those before the improvement.In addition,one of the new algorithms(P-CN)is compared with two other excellent algorithms,namely,a hybrid algorithm of heat conduction and mass diffusion(HHM)and a biased heat conduction(BHC),and find that the P-CN has better diversity and higher accuracy.The recommendation effect of four new algorithms on objects with different popularity is tested on four empirical recommendation system data sets.The new algorithms are found to effectively improve the recommendation accuracy of objects with low popularity.(2)We study a personalized recommendation system based on time effect.Based on two empirical recommendation system datasets,MovieLens and Netflix,the effect of data volume and data time on the recommendation effect is studied for three recommendation algorithms.It is found that increasing the amount of data used for personalized recommendation does not always improve the accuracy of recommendation,and the data with time close to the current recommendation data is very important for improving the accuracy of recommendation.The further study of Mass Diffusion(MD)and HHM algorithms shows that,when using a dataset with time close to the current recommendation data,the function form of the HHM algorithm is the same as the MD algorithm when it obtains the optimal recommendation result.The MD algorithm is compared with the above-mentioned three algorithms(CN,AA and SOR),and the MD algorithm is found to perform better than the CN,AA and SOR algorithms.It suggests that when using a dataset that is close to the current recommendation data,the HHM algorithm does not need to adjust the optimization parameter to get a good recommendation.This can significantly reduce the recommendation time.In summary,we analyze the user interest pattern and its effect on personalized recommendation,and investigate the time and data volume influence on the recommendation performance,which can provide an evidence for the improvement of recommendation algorithm.