Research On Recommendation Algorithm Based On Factorization Machine

People are living in an era of big data.In the face of massive information resources,how to quickly and accurately match information has become particularly important,and recommendation system plays an important role in realizing the balance of benefits between information producers and consumers.The user behavior data studied by recommender system can be divided into explicit feedback and implicit feedback.Implicit feedback data is the mainstream training data of recommender algorithms at present,and matrix factorization algorithm is still one of the most widely used techniques in recommender systems.This thesis improves the dot product defect of the matrix factorization method based on implicit feedback and the imbalance of the dataset.At the same time,it combines neural network technology to enhance the learning ability of the model.The main work and contributions of this thesis are as follows:(1)A recommendation algorithm for metric factorization based on full weighting is proposed.The dot product of matrix factorization does not satisfy the triangle inequality,which restricts the improvement of its recommendation effect.This thesis draws on the distance factor of metric learning,and converts the determinant of the correlation between users and items from the dot product size of matrix factorization to the distance of metric factorization.At the same time,in most of the current public datasets,the number of positive examples is much smaller than the number of negative examples.This unbalanced composition will affect the accuracy of the recommendation model.Inspired by the positive semi-definite matrix of Mahalanobis distance popular in the field of metric learning,this thesis proposes a full weighted matrix based on the interactive information between users and items.(2)This thesis proposes a fully-weighted metric factorization recommendation algorithm based on neural network.Factorization technology has derived many classic and effective recommendation algorithms,but it is still a traditional linear model in essence,and its ability to learn data and fit interactive functions is limited.In order to better learn and mine the potential nonlinear relationship of training data,this thesis introduces the neural network architecture to expand the model,which greatly enhances the expressive ability of the model.Finally,this thesis organically combines the above improved technologies to propose the All-Weighted Neural Metric Factorization(AWNMF)algorithm,and apply it to personalized item ranking tasks.This thesis has done a lot of simulation experiments on three common real datasets.The results show that the AWNMF algorithm model proposed in this thesis is superior to the competitive baseline model in mainstream ranking evaluation indicators.