Research On Deep Learning Recommendation Methods Combining Context Information

In recent years,with the advent of the era of big data,the continuous progress of the mobile Internet and the widespread popularity of intelligent mobile devices have brought great convenience to people.People can access the network anytime and anywhere,while obtaining a rich service experience,it also generates a huge amount of information data.The rapid development of the mobile Internet allows users to interact with the outside world through various forms of multimedia information.The diverse and complex contextual information directly reflects the user's emotional characteristics of items,and even indirectly implies the user's potential interest preferences.The recommender system is proactive to meet users' needs for various mobile Internet application services,help users make decisions,and save the energy and time required to search for candidate services.In the current environment of the mobile big data era,the final direction in the recommender system is to provide users with multi-level resource recommendations.Among them,information such as product ratings,location,time,review text,social networks,and emotional representation are the most common and the most easily contextual data that can be mined,analyzed,and utilized.However,existing context-aware recommender systems are often accompanied by problems such as incomplete context information mining,inadequate utilization of item-side information,too sparse user-item interaction matrix,and poor model interpretation capabilities.Therefore,research on context-aware recommendation has important theory and reality significance.The main work of this paper includes:(1)Fully studied the current research situation of context-aware recommendation,analyzed a series of problems contained in the current context recommendation,and analyzed on this basis to study effective solutions;(2)Propose a hybrid collaborative filtering recommendation method HDMF(Hybrid Deep Matrix Factorization),which uses deep learning structure for text semantic perception-based recommendation.By using the deep matrix factorization model to reduce the dimension of the high-dimensional sparse user-item rating matrix,combined with the recurrent neural network to extract the semantic feature representation of the text,and merged into the structure of the deep neural network to optimize the representation vector of the item.A large number of comparison experiments and additional hyperparameter analysis experiments were performed on the Cite ULike data set to verify the effectiveness of HDMF;(3)A collaborative filtering method ICDMF(Item Classification based Deep Matrix Factorization)is proposed,which is combined with item classification information from the perspective of the item side.The item classification not only reflects the function and semantic information of the item,but also has a dimension much smaller than that of the item sparse vector in the user-item rating matrix.It also improves the explainability of the algorithm itself.A large number of comparison experiments and additional hyperparameter analysis experiments were performed on the Movie Lens-1M and JD-Smartphone data sets to verify the effectiveness of ICDMF;(4)Based on the above research,the two algorithms are fused and implemented,and a recommender prototype system Rec Movie based on context awareness is implemented.