Research On Context-aware Recommendation Systems Based On Attentive Interaction Network

In the era of big data,with the flourishing of mobile network,cloud computing,and the Internet of things,the amount of Internet information and the number of visitors have grown exponentially.Although rich information has brought us more choices,it also brings serious information overload problem.Faced with this data deluge,users are looking to find high-quality information that they are interested in from the vast amount of information.And information producers hope to targe the rightproducts and services to users to increase revenue and user loyalty.Recommender systems(RS)play a pivotal role in alleviating the information overload problem and help users identify relevant information on Web.By mining user's history or user's profile,RS recommends items to a user that he may be intorested in.Traditional recommendation methods only consider user-item interactions when recommending items to users and do not take into account any additional contextual information.However,these kinds of contexts can have subtle but powerful effect on users behaviors.Compared to conventional recommendation methods,context-aware recommender systems(CARS)can better model user preferences for items and improve the performance of recommendations by taking into account the effects of contextsRecent state-of-the-art CARS methods are mainly based on latent factor mod-els and represent the common effects of contexts as a tensor.To model the effects of contexts on users and items,these methods generate context-aware user/item representations from the tensor,user latent vector,and context latent vector by linear operation.The context-aware user representations and item representations can be interpreted as the change of user interests and item properties under the effects of certain contexts.The context-aware user representations and item representations are then combine to predict users' preference scores for items in certain contexts.Although the existing CARS methods have greatly improved the performance compared with the traditional RS methods,the following short-comings still exist.First,the true effects of contexts may be much more complex in real world.The existing approaches rely on linear operations may not be sufficient to capture the complex effects.Second,different contexts usually have different effects.The existing approaches are unable to distinguish the different effects of different contextsIn this thesis,we explore the use of deep learning in the context-aware recommendation task and propose a novel neural network——Attentive Interaction Network(AIN).Specifically,AIN consists of three core components,namely Interaction-Centric Module,Effect-Level Module,and User/Item-Centric Mod-ule.Interaction-Centric Module explicitly captures the respective effects of con-texts on users and items.Effect-Level Module learns the importance of each context effect acting on users,and aggregate them to get the overall effect acting on users.Similarly,we can get the overall effect acting on items.Next,we model how the two overall effects influence user interests and item attributes,and get the context-aware user representations and item representations.At last,the user and item representations under context effects are combined to predict the recommendation scoresWe conduct extensive experiments on two explicit feedback datasets and one implicit feedback dataset.The experimental results show that AIN outperforms state-of-the-art CARS methods on both rating prediction and personalized ranking tasks.In addition,we also find that AIN provides better explainable recommendations.