Enhanced Pairwise Ranking Factorization Machine By Neural Network

In the field of early recommendation systems,matrix decomposition was generally used to process the User-Item matrix.However,with the development of the Internet,the dimension of the User-Item matrix in the system becames larger and larger,the data becames more and more sparse.Matrix decomposition is difficult to efficiently process high-dimensional sparse matrices and lacks the utilization of contextual information.In 2010,a factorization machine model that can model contextual information was proposed,and this model has quickly been widely used in prediction,recommendation and other fields.Since then,with the improvement of computer computing power and the development of neural networks,the factorization machine model based on neural network enhancement can capture more interactions of high-order features,thus making the prediction results more accurate,and has become a research hotspot of current recommendation algorithms.Among them,the personalized neural network-enhanced pairwise ranking factorization machine model NPRFM,which combines implicit feedback with content information,has received extensive attention.In the field of personalized ranking with implicit feedback data,pairwise learning methods are an important technique that is widely used.The method not only learns user interests from recorded user feedback,but also learns user preferences well from potential interactions between users and items.However,the existing methods do not comprehensively consider high-level interaction features and original input features when modeling the interaction features between users and items,resulting in inaccurate recommendation results.In this paper,the NPRFM model is deeply studied,and a novel deep neural network-enhanced pairwise ranking factorization machine Deep PRFM is proposed and implemented.The main research contents are as follows:(1)In view of the fact that many current methods pay more attention to the high-level interaction of features and ignore the original user-item input feature interaction problem,which can best reflect the first-hand information of users and items,this paper fuses the high-level interaction features between users and items with the original input features,and proposed and implemented a new personalized deep neural network-based pairwise ranking factorization machine model Deep PRFM by using the pairwise learning method.(2)Aiming at the widespread cold-start problem in recommender systems,this paper proposes a pairwise learning method with multiple negative samples,which makes full use of items in the system that users have not interacted with to improve model accuracy,speed up model optimization,and alleviate the cold-start problem of system users.(3)In order to alleviate the high time complexity brought by the pairwise learning method of multiple negative samples,a sampling strategy based on typical negative samples is introduced in the pairwise learning method of multiple negative samples.The sampling strategy can dynamically select the sample with the highest gain to the model effect as a typical negative sample according to different users.In this way,a smaller proportion of negative samples can be used,and the excellent performance of the factorization machine enhanced by pairwise learning with multiple negative samples can be achieved,and the model convergence speed can be accelerated.Extensive experimental results on two real datasets show that the Deep PRFM model proposed in this paper outperforms a large number of current methods in personalized ranking,and when the multi-negative sample pairwise learning method and the typical negative sample sampling strategy are applied to other pairwise ranking factorization machines,it can also get better results than the original model.