Fusing Explicit And Implicit Feedback For Deep Collaborative Filtering Recommendation

The emergence of Internet technology makes all kinds of information service websites flourish,bringing the explosive growth of Internet information.While enjoying the convenience of various information services,people have to deal with the problem of "information overload".The recommendation system is an important information filtering technology,playing an increasingly important role in dealing with information overload,and is widely used in a variety of online service systems,including e-commerce,online news,and social networking sites.Collaborative Filtering has been widely used in the recommendation system due to its outstanding ability to make use of user-item interaction information.Although the collaborative filtering algorithm has achieved great success,it still faces problems such as data sparsity and cold start.In recent years,due to its powerful representational ability,deep learning has been widely used in-depth recommendation methods based on collaborative filtering,which greatly alleviates the problem of data sparsity.However,most of the existing in-depth recommendation models focus on the design of neural network models,and little attention is paid to the research of input data construction and interaction function design.Aiming at the above problems,this thesis studies the deep collaborative filtering algorithm as follows:(1)Fusing explicit and implicit feedback.Most of the existing deep recommendation models only consider one of the two feedbacks,which may impair the performance of recommendation models.To address the issue,we proposed a novel deep recommendation model that captures user-user and user-item interaction features by combining both explicit and implicit feedback.Explicit feedback is considered as a high-dimensional vector to help us learn about local information in relation to user-item interaction;implicit feedback is implemented to learn about independent user-item latent vectors.We then merged such two models to maximize the advantages.(2)United cosine similarity interaction and inner product interaction.In view of the fact that most of the existing researches only uses one interaction function to learn the similarity between users and items,we propose a general framework to combine multiple interaction functions.We design an algorithm example of joint cosine similarity interaction and inner product interaction and use the advantages of the two interaction modes to improve the performance of the model.According to the characteristic that the user's interest changes with time,a loss function of time series attenuation is designed according to the interest drift theory to assign different weights to the user's history records at different times.