Research On User Preference Modeling For Recommendation System

With the rapid development of the Internet in recent years,we have to receive a huge amount of information every day,and this information is full of content that has nothing to do with ourselves.For users,it is very important to obtain information that is valuable to them in the wave of data;for enterprises,obtaining information that users like from a large amount of data and recommending it to users will help enhance user stickiness.Therefore,how to use existing data to make personalized recommendations for users is a compulsory course for every company.An important process in the recommendation system is to formulate user portraits and model user preferences,that is,to use existing data to analyze user preferences.Only by accurately grasping the user's habits and preferences,can we bring a better experience to the user and thus retain the user.One of the important methods is project-based collaborative filtering.Projectbased collaborative filtering has the advantages of high recommendation accuracy and easy online recommendation,and is favored by industry recommendation systems.Item-based collaborative filtering recommends items to target users based on the similarity of the user's previous interactive items.In recent years,item-based collaborative filtering has applied advanced machine learning techniques to learn item similarity from data,and great progress has been made.Earlier methods simply treated all of the user's historical items equally,while recent methods distinguished the different importance of items to predictions.Despite the progress of the work,we believe that these project-based collaborative filtering models still ignore the different intentions of users viewing the project(for example,watching a movie because of the director,main actors,or visual effects).Therefore,they cannot estimate the product similarity at a finer-grained level,thereby predicting the user's preference for a certain product,leading to sub-optimal recommendations.In addition,because user data is limited,in a model with a large number of parameters,limited data will restrict the performance of the model,so it is necessary to rely on data from other fields for cross-domain recommendations.However,the traditional cross-domain recommendation method mainly uses the embedded mapping between the two domains,which results in the two domains not being able to better exchange data and the model may not be able to perfectly map the embedded representation of cold-start users.In order to solve the above two deficiencies,in order to better model user preferences,this article will conduct research from two directions,namely vertical modeling and horizontal modeling.Longitudinal modeling is to conduct deeper modeling on the data of interaction between users and items that already have,to fully mine user preference information;horizontal modeling is to supplement information from other fields,that is,cross-domain recommendation,which can be effective Solve problems such as data sparseness and cold start.On the issue of longitudinal modeling,we propose a general factor-level attention method based on the item-based collaborative filtering model,and combine item-level and factor-level attention to make the model more efficient;for the issue of horizontal modeling,we use The Attention Graph Neural Network performs cross-domain recommendation,so that the data between the two domains can be better integrated.We have conducted a lot of experiments on real data sets,and the results show that the performance of the factor-level attention enhancement model is always better than similar models,which proves the potential of the item-based collaborative filtering recommendation model to distinguish user intentions at the factor level;The performance of the attention-based cross-domain graph neural network model is also better than that of the traditional mapping model,and attention also greatly improves the recommendation ability of the graph neural network.