Research On Item-based Hybrid Recommendation Methods

Recommender systems aim to solve the information overload problem by automatically providing users with items that they might like.Collaborative filtering uses a user's historical interaction data to predict the user's preference and have been widely used in recommender systems.Item-based collaborative filtering assumes that a user would like to choose an item similar to what she has interacted before.Therefore,how to make full use of the item correlations has become one of the key technologies.Under this background,the main work of this paper includes:(1)A user-item interaction graph provides correlation information among items to enrich user representations in item-based collaborative filtering.However,such information can not be directly used to serve a new user because the new user's interaction information is not available in the existing user-item interaction graph.This paper proposes an asymmetrical variational auto-encoder,namely AsyECF,to fill the gap.In AsyECF,we only use the graph neural network in the training phase to mine the item correlation information on the user-item interaction graph,which strengthens the representation learning.The learned item representations are used to represent the user's preference in recommendation.Therefore,in the recommendation phase,no matter whether users are present in the training data,we can be free from the graph and only use the the learned basic items embeddings to predict their intents.The results of experiments on three real-world datasets show that the recommendation performance of AsyECF is superior to baselines for users who are not participating in training.(2)The existing session-based recommendation methods with graph neural networks mine the item correlations a session to generate recommendations.They usually formulate the recommendation problem as the classification problem,and then use a specific uniform loss to learn session graph representations.Such supervised learning methods only consider the classification loss,which is insufficient to capture the item features from graph structured data.Therefore,this paper proposes the HybridGNN-SR model to combine the unsupervised and supervised graph learning to represent the item transition pattern in a session from the view of graph.Experimental results on three real-world datasets demonstrate that the combination of unsupervised and supervised graph learning methods can improve recommendation performance.(3)Most of the existing methods always embed a user or an item as a point in a vector space,and then model the user's recent behaviors as a sequence with a strict order to generate recommendations.However,both the point representation and strict order rule limit the capacity of sequential recommendation models as the diversity and uncertainty of a user's interests.This paper proposes the Box4Rec to address this issue.Box4Rec embeds a user and the user's historical items as boxes to model the user's general preference and short-term preference,and then integrates the conjunction and disjunction operations on items to generate flexible recommendation strategies.Experiments on five datasets with different scales show the recommendation performance of Box4Rec.