Research On Improved Neural Collaborative Filtering

With the ever-growing volume of online information,recommender systems have been an effective strategy to overcome such information overload.Recommendation system allows users to discover new items that match their tastes and enables the system to target items to the right users.In the real world,implicit feedback has widespread availability than explicit feedback,some researchers have endeavored to design recommendation systems based on implicit feedback.Traditional collaborative filtering methods,such as matrix factorization models,which simply model user-item interactions as a linear combination of their latent factor vectors,have greatly limit the ability of the model's representation.In recent years,some researchers have designed neural network-based collaborative filtering to explore the application of deep neural networks in recommendation systems,Neural Collaborative Filtering is one of the most representative methods.In this paper,we deeply researched and analyzed Neural Collaborative Filtering,and proposed the following improvement schemes for its shortcomings;(1)Neural Collaborative Filtering models user-item interactions into a complex nonlinear model to improve recommendation performance.However,Neural Collaborative Filtering uses a large number of model parameters,which results in low efficiency of model training.In this paper,we found that the embedding layer in the Neural Collaborative Filtering framework is a fully connected layer,in which the latent factor matrices of users and items are trained as part of the overall model parameters.In order to train the embedding layer separately in advance,reduce the parameters of the late model training,and improve the training efficiency.In this paper,we proposed the IFE-NCF framework by combining the Implicit Feedback Embedding model,and applied this method to the two implementations of the NCF framework — GMF and MLP,then obtained IFE-GMF and IFE-MLP,respectively.(2)However,Neural Collaborative Filtering only relies on the interactions between the users and the items,it performs poorly for “cold-users”(users with little history of such interactions)and at capturing the relationships between closely related items.In particular,Neural Collaborative Filtering does not take into account certain connections between items when modeling,and may limit the recommended effects of the model.In addition,only the interaction between the user and the item is used,which is not enough to make effective recommendations for users with less historical interaction behavior.To address these problems,we introduce context items under the framework of IFE-NCF,and proposed a Context Items-based IFE-NCF framework,which models the interaction between target users and test item in two terms: the personal preference of the user for the item,and the relationships between this item and other items clicked by the user.In this paper,the Context Items-based IFE-GMF and the Context Items-based IFE-MLP are implemented under the Context Items-based IFE-NCF framework.Finally,in this paper,we have done extensive experiments on the two implementations of the IFE-NCF framework and the two implementations of the Context Items-based IFE-NCF framework,and also performed a qualitative analysis that shows IFE-GMF and IFE-MLP can significantly improve the training efficiency without loss of prediction accuracy;Compared with GMF and MLP,Context Items-based IFE-GMF and Context Items-based IFE-MLP both have significant improvement in recommended performance and effectively alleviates the “cold-users” problem.The experimental results also show that in the training process of Context Items-based IFE-GMF and Context Items-based IFE-MLP,with the number of iterations increasing,their recommended performance is continuously improved,but when the number of iterations reaches a certain value,the recommended performance of both models tend toward stability.Moreover,the experimental results also show that increasing the number of hidden layers of the neural network can significantly improve the recommended performance of the Context Items-based IFE-MLP.