Research On Recommendation Technology Based On Multi-source Data Fusion

In the era of big data,inundated by the overwhelmingly massive information,people can hardly distinguish valuable information effectively for further decision-making or choices,which has become an inevitable issue in the Internet era,known as information overload.With the development of recommendation technology,a feasible solution to the information overload is available now,which has been extensively applied in news information,e-commerce or education.To provide a more personalized recommendation system for each user with more targeted recommendations,a more fine-grained modeling on user interests and preferences should be firstly performed because in-depth understanding of the user is the foundation for offering specific and value recommendations.However,for most of the existing algorithms,the abovementioned modeling is realized merely by means of user-item historical behavior interaction information,without fully utilizing other multi-source data.The recommendation effect is undoubtedly hard to meet user requirements and also affects user experience.Therefore,the study on how to more effectively integrate the multi-source data related to the recommendation task,and perform modeling of user interest and preferences accurately will be of great research significance for providing better recommendations for users.Through summary and analysis of the previous recommendation algorithms,it is found that better user interest modeling and more effective feature combinations will be more conductive to the improvement of the recommendation system performance.Thus,the study is conducted from two basic perspectives.To obtain a more effective feature combination,a recommendation algorithm based on deep feature overlapping is designed.A more fine-grained feature expression is achieved by introducing the concept of feature domain,thereby realizing the deep overlapping of multi-domain features and solving parameter problem arising from the previous step through the method of sharing parameters.Further,the model can learn the complex non-linear relationship between the user and the item by means of deep learning,thereby enhancing the expression capability of the model and improving the generalization ability of the model to obtain better recommendation results.For accurate modeling on user interest preferences through effective utilization of multi-source auxiliary data,a basic model structure capable of effectively integrating multi-source data is designed with model scalability and the ability to effectively implement multi-source data fusion.Based on the fusion framework proposed in this paper,we put forward a self-attention-based representation model recommendation method Firstly,through self-attention mechanism,it is found that model modelling of multiple features of the same item varies in terms of the importance for different users.Next,in terms of the transfer of user interest preferences,such sequential relationship is dealt with by using a two-way gated recurrent unit network for modeling of user interest transfer.From above two basic perspectives,deep feature combination overlapping is realized and more effective feature combinations are discovered.A more fine-grained modeling on user interest preferences is realized.Thus,the designed algorithm is verified to be effective through experiments.