Similarity-aware Recommendations For Different Scenarios

With the development of Internet information technology,people are faced with the increasingly serious problem of information overload.It’s difficult for users to pick an item from a large number of candidate items.The recommendation system can discover users’ potential interest by using historical data such as user information,item information,and interactive information.The recommendation system helps users find their favorite item and provides users with personalized recommendations.First,through the study and analysis of collaborative filtering recommendation algorithm and autoencoder algorithm,we propose an algorithm called Similarity-Aware Recommendations based on Denoising Auto Encoder(SARDAE).The SARDAE algorithm integrates collaborative filtering and denoising autoencoder.Through the paired items similarity discrimination task,the model can hold the similarity information between items while reconstructing item vectors.So our model can pay more attention to the user ratings for similar items.Secondly,when constructing paired samples,we delete some popular items.The method can alleviate the Matthew effect in the recommendation system.When constructing positive and negative samples,we use the sampling method to select the similar item instead of the most similar item.It efficiently reduces the time complexity of sample construction.Through experiments,we analyze the impact of different hyper-parameters on the model.The experiment results show that the SARDAE algorithm can improve prediction accuracy.However,the SARDAE model is unable to provide personalized recommendations in the cold-start scenario,it can only use a certain fixed value as user rating.Therefore,for the cold-start scenario,this paper proposes a similarity-aware neural network model(SANN).The model uses user attribute information,item attribute information,and paired items similarity information.Through the paired items similarity discrimination task,SANN learns the scoring modes of different types of items.It improves the model’s prediction ability in the cold-start scenario.The SANN algorithm mainly uses the attribute information of users and items,and is suitable for recommendation in the cold-start scenario.The SARDAE algorithm makes full use of user-item rating information,and the prediction accuracy is higher than SANN in the normal scenario.Therefore,this paper combines the SARDAE algorithm with the SANN algorithm and proposes the SARDAE-CS algorithm.If it is a new user or new item,we will use the SANN model to predict the rating,otherwise we will use the SARDAE model.