Hybrid Recommendation Algorithm Based On User Profiles And Item Profiles

With the rapid expansion of E-commerce and short video business,the data information in the cyberspace has achieved the exponential growth and the "information overload" has become the major challenge nowadays.The recommended system greatly alleviates the difficulties.It takes the optimization of the user experience and increase of the merchant’s profits as the goals.The users’ behavior logs are analyzed to construct the user profile information,and then actively generate recommendation results.Although the recommended system has achieved the certain progress,it also faces the great challenges in the cold start,sparse co-occurrence matrix,follow-up of the users’ interest changes and other aspects.To alleviate the impact of above problems on the recommended system,this thesis divides the profile information into behavior profile and attribute profile.The self-encoder model is adopted to extract the effective information in the co-occurrence matrix,to reduce the influence of sparse co-occurrence matrix on the algorithm and generate behavior profiles at the same time.The semantic embedding model is applied to encode the demographic information,item description information and other information into the vectors,and generate the attribute profiles,mitigating the impact of cool start on the recommended algorithm.Inspired by the collaborative filtering,the UPHR algorithm and IPHR algorithm are proposed on the basis of user profiles and item profiles,respectively.Meanwhile,this thesis explores two methods integrating the UPHR algorithm and IPHR algorithm,namely the weighting method and logistic regression method,putting forward the UP-IPHR algorithm.In addition,this thesis improves the UPHR algorithm model from the behavior profiles and hobbies,and puts forward the T-UPHR algorithm,to further improve the model interpretability and refine the users’ profile information.Firstly,this algorithm introduces the attention mechanism and learning strengthening in the UPHR algorithm,to explore the important attributes affecting the users to make the choices and refine the users’ profile information.Then,this algorithm introduces the Ebbinghaus forgetting curve and time context information in the recommended model,to improve the defects that the traditional recommended model could not track the users’ interest changes,so as to improve the recommended quality of the model.To sum up,this thesis proposes four recommended models based on the profile information,namely,UPHR algorithm,IPHR algorithm,UP-IPHR algorithm and T-UPHR algorithm.Based on the experimental results of four publicly recommended data sets,it could be concluded that three kinds of the algorithm(UPHR algorithm,UP-IPHR algorithm and TUPHR algorithm)could better mitigate the impact of the cool start and sparse cooccurrence matrix on the recommended algorithm.Its performance in the accuracy and recalling rates is better than the classic recommended algorithm model and the Top-N recommended model based on the deep learning.