Research On Personalized Recommendation Algorithm Based On Multi-source Data

With the rapid development of Internet technology,the amount of data information is increasing day by day,and people are becoming more and more powerless in the face of huge and complicated data.So in order to solve the problem of information overload,a recommendation system was proposed.Nowadays,recommendation systems are widely used in e-commerce,music,video,advertising and other fields,and they play an irreplaceable role.In the recommendation system,two types of user historical feedback data are used for recommendation,one is explicit feedback data,and the other is implicit feedback data.In the explicit feedback recommendation,the collaborative filtering algorithm acts as one of the most prestigious algorithms,but there are always problems of data sparsity and cold start.In order to improve and deal with these problems,in recent years,more and more collaborative filtering algorithms have been integrated and recommended with other algorithms,and these algorithms with their respective advantages and disadvantages and complementary can be effectively integrated,and their respective advantages can be fully utilized to achieve better results.Good recommendation effect.In implicit feedback recommendation,one of the most well-known effective algorithms is Bayesian Personalized Ranking(BPR).In order to better predict the user's preference ranking of product items,many researchers have further explored to optimize and improve their algorithms.It is good to construct the user's preference behavior and enhance the feature extraction ability.Therefore,these two parts are one of the important research directions of the current recommendation system.At present,this article uses the user's explicit feedback or implicit feedback data to find the content that users are really interested in.Based on the collected data,the user interest model is constructed to provide users with accurate and personalized recommendations.On the one hand,when the explicit feedback data is easy to obtain,this article improves the traditional collaborative filtering algorithm to alleviate the cold start problem and improve the recommendation accuracy;on the other hand,when the explicit feedback data is not easy to obtain,but it is implicit When the formula data is rich,this paper proposes a new MF ranking recommendation algorithm by improving and optimizing the BPR recommendation algorithm to improve the recommendation accuracy of the product item ranking.Therefore,the explicit feedback data and the implicit feedback data are studied separately.The main research contents are as follows:1.Aiming at explicit feedback data,a fusion recommendation algorithm based on user-based collaborative filtering and popularity and word-of-mouth rankings is proposed.According to the group effect of people,they will often pay attention to recent TV series,movies,books or hot events;see a lot of people watching together,I want to go up to the fun,check it out;when shopping online,they will also tend to buy good reviews.More products.Therefore,when people do not have a strong purpose to choose a product item,they will consider highly popular products or products with good reputation to a large extent.This can solve the cold start problem to a certain extent.When we don't know what the user needs at the beginning,we can recommend products that are popular or have a good reputation.When users have historical behavior records,they can recommend the products they need through the user's historical behavior records and the psychology of group effects.2.In view of the implicit feedback data,a matrix factorization recommendation for optimizing the ranking in implicit feedback is proposed.It is found that in BPR,the score of products that are both positive and negative feedback will become higher than that of products with only positive feedback,which leads to inaccurate rankings of products recommended to each user and degradation of recommendation performance.Therefore,a new optimized ranking MF method is proposed to effectively learn from implicit feedback.This study will use the TOP1 ranking loss function to train the MF model for the first time.And for the problem that the ranking ratio of products with only positive feedback is both positive feedback and negative feedback products,a new optimization criterion is proposed.This article compares with a number of representative recommendation algorithms and conducts multiple comparative experiments,and all have achieved better experimental results,which can verify the effectiveness of the improved method proposed in this article.