Context-aware Recommendation Based On Attribute Boosting And Preference Integration

In recent years,Context-Aware Recommender Systems(CARS)are very popular.Because such recommender systems can do better in implementing personalized recommendation by mining contextual information.Two core entities in recommender systems,users and items,have close relationship with contextual information.From this perspective,contextual information can be called attribute,which includes user attribute and item attribute.The technique for modeling attributes is evolving and changing,however,attribute modeling still faces these problems:(1)Existing methods are not flexible enough for handling attributes;(2)Some interaction models are too complex and not targeted;(3)Different recommendation tasks have different requirements in attribute modeling;(4)General algorithms are influenced by attributes of different domains;(5)The impact of different records on user preferences varies.To overcome these shortcomings,this paper proposes two types of strategies to optimize context-aware recommendations,which are attribute boosting and preference integration.First,attribute boosting can complete two major tasks,rating prediction and item recommendation,in recommender systems separately.Then,preference integration is able to refine the prediction results based on attribute boosting.Finally,preference integration can be improved as a generic framework for common recommendation.The contents of this paper are as follows:(1)Research on local boosting and preference integration for rating prediction tasksBased on boosting items,local boosting technology implements attribute interaction through three aspects,including user,item,and attribute type.By combining gradient descent with sampling techniques,local learning strategy is proposed to train the framework efficiently.After training,the local predicted preference consists of three aspects of preferences.Based on local boosting,preference integration method adopts gradient boosted regression trees to integrate multi-preferences for representing overall interests of users.Finally,the rating prediction value is generated by local predicted preference and user overall interest.The experimental results demonstrate that the local predicted preference is more accurate than popular context-aware methods such as factorization machine.Moreover,with the help of user overall interest,the accuracy of rating prediction has been further improved.(2)Research on global boosting and preference integration for item recommendationThe main idea of global boosting is to reduce the negative impact of specific domains on attribute modeling by attribute neighbors.First,user neighbors and item neighbors are found by similarity calculations.Then,a single-neighbor integrating strategy is proposed to make attributes domain-independent.At last,three new interaction methods are developed to form an item recommender.To improve the global item recommender,local low-rank approximation technology is utilized in preference integration.More specifically,each neighbor is given a flexible weight to express its contribution.Then,multi-neighbors are integrated and interact with each other for recommendation.Experiments prove that global boosting approach is less sensitive to specific domain.After joining global preferences integration part,the performance of global recommender has outperformed advanced item recommenders,such as pairwise interaction tensor factorization.(3)A common preference integration framework based on division learningUnlike the two types of preference integration technologies above,the common preference integration approach focuses on the division and modeling of records.The core of this framework is the division learning strategy,which carries out recommendation through three ordered steps:dividing records,mining preferences,and integrating preferences.First of all,the algorithm proposes group division tree to complete the division of original records and generate group preferences.Then,for each group,a lightweight regression model is built to capture local preferences of a particular user in that group.At last,by integrating group and local preferences,the common recommender gets the overall interests of users.Experiments show that by flexibly adjusting the model granularity,this recommender can accommodate different datasets to achieve its optimal performance.Moreover,such framework is able to finish two recommendation tasks in high accuracy.To summarize,through the proposed attribute boosting and preference integration strategies,this paper optimizes context-aware recommendation in a number of ways,to fulfill two main tasks in recommender systems.This is a new attempt to attribute modeling and preference prediction,which provides an idea to mine deeper user interests.Therefore,this topic has a large theoretical research value and positive significance of practical application.