| With the development of the Internet,it provides convenience for users to obtain diversified information.However,the problem of information overload also comes along,and users need to quickly find the content that meets their needs in the massive amount of information.To solve this problem,recommendation system comes into being.With the increasing emphasis of the business community on user experience,in the field of commodities with sparse individual data,the recommendation system has poor recommendation performance at the early stage of its launch,which is called the difficult cold start problem,and it is an emerging and concerned problem in the field of recommendation system research.To solve this problem,existing studies have the following problems: 1)Recommendations are made with the help of data in this field,and the comments of users in different fields are not fully utilized;2)More attention is paid to extracting information about comments from the model level,while the effectiveness of extracting implicit features of comments in the pre-training stage is ignored;3)Insufficient attention is paid to the data relationship between ratings and reviews in cross-domain recommendations to improve the quality of data characteristics;4)Emotional feature extraction and fusion are not performed at the level of emphasis between cross-domain reviews.To solve the above problems,this thesis proposes two recommendation algorithms based on cross-domain information extraction and fusion,and on this basis,adopts various methods of enterprise application development technology to design and implement a high availability movie recommendation system.Major contributions include:1.A cross-domain recommendation algorithm of deep emotion perception integrating domain information is proposed.Firstly,the fusion method of domain information and user information is designed to extract the common features of the comments of the same user in multiple fields,fill the target field,and alleviate the problem of data sparsity.Secondly,BERT pre-training method was introduced to extract the emotional features of exotic comments in the pre-training stage,so as to fully explore the cross-domain emotional features.Compared with the baseline model,a comparative experiment was conducted on five public data sets.With MSE as the evaluation index,compared with the optimal results of the baseline model,the results increased by 12.1%,7.6%,13.4%,10.4% and 18.2%,respectively.2.A rating and comment fusion enhancement cross-domain emotion recommendation algorithm is proposed.First of all,the correlation between user comments and ratings is modeled,which not only captures the potential correlation between comments and ratings more fully,but also further fills in the feature domain with sparse data effectively.Secondly,a critical comment filter is introduced to screen important comment features to avoid overfitting problems.Finally,gating mechanism is introduced to extract emotion features at aspect level to further enrich cross-domain recommendation features.Compared with the baseline model,a comparative experiment was conducted on five public data sets.With MSE as the evaluation index,compared with the optimal results of the baseline model,the results increased by 6.8%,4.2%,5.4%,6.8% and 7.2%,respectively.3.Design and implement a available online movie recommendation system,which is based on the recommendation algorithm in the thesis,uses a variety of enterpriselevel application development technologies,combines the movie recommendation application scenarios,and introduces the distributed development idea. |
