Neural Collaborative Filtering Framework Based On Knowledge Map Representation Learning Method

The purpose of the recommendation system is to generate a personalized ranking list of items that may be of interest to end users.With the unprecedented success of deep learning in the field of computer vision and speech recognition,how to reasonably introduce deep learning into the recommendation system has also aroused the thinking of researchers.Knowledge map,as a new research hotspot,contains abundant new auxiliary information of entity semantic association.The researchers found that when the knowledge map is introduced into the recommendation system,it can reduce the data sparsity and cold start problem,and it is a good assistant for neural network in the recommendation system.In the traditional recommendation system,because it relies on the matrix decomposition and collaborative filtering algorithm for recommendation,there will inevitably be problems of cold start and data sparsity.The problem of data sparsity often refers to the large number of users and projects in platforms such as large-scale e-commerce,but in the user project matrix obtained,the average number of users interacting with the project is small,which will cause the user project matrix to be sparse.The cold start problem refers to how to make personalized recommendation for new users without a large number of user data.The sparsity of data will eventually lead to the inability to capture the relationship between different users and different projects,thus reducing the accuracy of the recommendation system.As an implicit expression,implicit feedback can get users' preferences in many ways,rather than limited to the display of expression preferences,so as to enrich the user item matrix and alleviate the problem of data sparsity.Neural network can analyze the relationship between things from a higher dimension,and improve the data sparsity.In the final analysis,the problem of cold start is that the information dimension of data is not enough.The knowledge map contains the fact relationship of a thing in the real world,which is equivalent to providing additional information dimension for the data needed to be trained in the model,so as to solve the cold start problem to a certain extent.This paper proposes a collaborative filtering recommendation ranking framework kncr based on implicit feedback data and knowledge map representation learning and neural network.First,the framework introduces deep learning neural network based on traditional matrix decomposition algorithm,and is verified on MLP and CNN respectively.Secondly,theframework uses knowledge map to represent the learning method,which enriches the information dimension of the existing data set by extracting the existing semantic data from the public knowledge base.Therefore,we propose the low-dimensional dense vector represented by knowledge map as the neural collaborative filtering model embedded in the middle layer.Thirdly,in order to be more close to the practical significance of recommendation system,we study a learning method of pairwise sorting,which optimizes the loss function and improves the ability of model learning.We have carried out a number of comparative experiments on open datasets,and the results show that kncr effectively improves the accuracy of the recommendation system.