Research On Two-step Prediction-based Movie Recommendation System

In today's information age,recommendation system plays an important role in all kinds of industries.To achieve accurate marketing is the ultimate goal of recommendation.According to the historical activity record of the user,as much as possible,the relevant behavior patterns of the user and some characteristics of the commodity are excavated,the corresponding user portrait is constructed,the interest of the user is accurately located,and the goods most likely to be interested are matched for the user according to the interest of the user,so as to recommend the most suitable commodity for the most suitable person.The research of the system is also very important in the film industry.The research of the film recommendation system has a certain theoretical meaning and some application value.In theory,the angle of the improved algorithm proposed and the means of feature mining can provide some reference for other research;from the practical application value,it can help improve the film recommendation system and realize more efficient marketing.According to the research of film recommendation system,the existing problems are as follows:(1)In view of the film review data,there is a lack of comparison and discussion on the applicability of different methods,and there is a certain possibility of misuse;(2)The mining of features is not deep enough,and more feature mining is not done by combining text data and browsing data at the same time;(3)The traditional film recommendation only uses a score prediction,with the disadvantage that it includes a scoring prediction with a low viewing probability.In view of the existing problems,the core work of this paper is to compare the applicability of different methods according to the text data.At the same time,combined with the graph embedding algorithm to deeply mine the features,a two-step prediction film recommendation system is proposed.First of all,this paper uses Latent Dirichlet Allocation topic model(assuming that each article is composed of a series of topics,the distribution of words corresponding to each topic is different,and generates each document through Gibbs sampling)and the topic word embedding algorithm(word2vec algorithm combined with clustering algorithm)to mine the topic.Because the Latent Dirichlet Allocation(LDA)topic model realizes the mining of the topic through the co-occurrence relationship between the words,the algorithm has some disadvantages in the short text data.However,the topic embedding algorithm doesn't ask for high requirements for text data and does not need to aggregate short text data in advance.Through the comparison of the two algorithms,the following conclusions are drawn:(1)The topic embedding algorithm is more detailed than the LDA topic model at the level of topic mining;(2)The topic embedding algorithm can automatically extract the corresponding stop words,which can provide a certain reference for the construction of professional stop words corpus.(3)The topic feature extracted by topic embedding algorithm has better effect on the improvement of recommendation system.Then,the feature is further expanded by graph embedding algorithm(using user and movie as nodes to build a network,and the graph embedding vector of nodes is obtained by random walk),and it is found that the feature shows a specific pattern on similar users or similar movies,and the interactive information of different users' browsing records is more considered,and the feature performs best in the first step prediction task in the recommendation framework of the subsequent two-step prediction.Finally,in view of the shortcomings of the traditional recommendation system,a two-step recommendation framework is proposed,that is,the first step is the probability prediction of the user clicking on the film,the second step is the prediction of the user's score of the film.The final score is obtained by the predicted probability value multiplied by the predicted score.According to the order of the final score,we can get the corresponding recommendation list.By comparing the two-step recommendation framework with the traditional one-step recommendation framework,it is found that the two-step recommendation framework has achieved better performance,and the rationality of the two-step recommendation framework proposed in this paper is also explained.In this paper,the deep mining of text data is realized,and the performance and applicability of different methods are discussed.Combined with the graph embedding algorithm,the deep mining of features can provide some reference for other research,such as the mining of user comment data by e commerce platform in other scenarios.Moreover,the two-step recommendation learning framework proposed in this paper performs better than the traditional one-step learning framework,and significantly improves the recommendation effect.