Research On Deep Interest Network Recommendation Algorithm For Networked Collaborative Manufacturing Cloud Servic

As one of the algorithms closest to people’s life,recommendation algorithm is widely used in various fields such as e-commerce,videos,musics,news blogs and job recruitment,etc.Its research is of great significance.The development of Internet technology has led to the spread of networked collaborative manufacturing mode and increasingly refined division of labor,especially in the field of aviation equipment manufacturing,the networked collaborative manufacturing cloud service platform can greatly promote the refinement of aviation equipment manufacturing process and industrial prosperity.In order to better serve customers,the cloud platform transforms from selling products to selling services.Cloud services,especially aviation equipment manufacturing cloud services,have complex and diverse characteristics,which makes it difficult for users to obtain the cloud services they are interested in.In order to solve this problem,this paper investigates the cloud service recommendation algorithm.However,cloud service recommendation is more different from other recommendation scenarios: cloud services often have unspecific and missing information descriptions.In addition,the process of cloud service recommendation also has problems such as slow recommendation response,more neighboring candidate services lost in the recall layer,and lack of diversity in recommendation results.To address the above problems,the thesis first improves the recall layer local sensitive hash algorithm bucket boundary and user cold start problems,adds a random bias term with normal distribution to the local sensitive hash algorithm,extracts information such as image text and items and applies it to the recall algorithm and multimodal recommendation,and finally uses integrated learning and distillation to improve the diversity of recommendation results.The main work of this thesis is as follows:(1)To address the problem that the neighborhood points are easily lost due to the solidification of the bucket boundary of the locally sensitive hash function or the use of uniformly distributed random bias terms,we propose to improve the recall accuracy of the locally sensitive hash algorithm by adding the random bias terms with normal distribution.In addition,to address the user cold start problem that occurs in the recall of locally sensitive hash algorithm,we propose to extract multimodal information using algorithms such as item recognition and image text recognition,and input to BERTKMeans algorithm for encoding and clustering in bucket recall,which solves the problem that new users who have not rated items cannot make recommendations.(2)To address the problems of unspecific and missing cloud service information descriptions,a method of extracting picture items,text,and color features for multimodal recommendations is proposed to improve the accuracy of fine-ranking layer recommendations.In addition,we add multi-head mechanism and use distributed computing framework spark respectively to strengthen the model fitting expression ability and improve the computational efficiency.(3)To address the problem that the recommendation results of cloud services are too single and boring,we propose to integrate multiple frontier deep recommendation models using integration learning,initialize and regularize the parameters of the integration layer,and design a more appropriate activation function for activation integration of multiple sub-models.Finally,the integrated large model is compressed by knowledge distillation,which improves the model inference speed and recommendation result diversity.