Research On Metric Learning Recommendation Model Combining With Tags

Recommendation system actively searches for useful information for users in the data ocean of Internet.The core of recommendation system is recommendation model.It is of great engineering significance to research recommendation model.Matrix factorization recommendation model has been widely used because of its high recommendation precision and easy implementation.However,the point product is used to predict user preference in Matrix factorization recommendation model.The product is only a simple linear product and does not meet the trigonometric inequality,which limits the effect of Matrix factorization recommendation model.Some researchers propose a recommendation model based on metric learning,and use distance measure which satisfies trigonometric inequality instead of point product measure in Matrix factorization,then train the measurement distance between users and projects in low-dimensional space.Distance measurement satisfying trigonometric inequality can fully represent the nonlinear interaction between users and projects,construct the similarity transmission chain between users and projects,and realize similarity transfer.The model of metric learning recommendation has the advantages of strong interpretability,easy to understand and low training cost.The existing recommendation model based on metric learning has the following shortcomings: 1.It cannot fully offset the adverse effects of user rating preferences on the recommendation results.User rating preference refers to the user rating habits.Extreme rating preferences will greatly reduce the accuracy of the recommendation model.Existing metric learning recommendation models usually use bias items to express user rating preferences without considering item differences.2.The similarity between users without co-rated data is misjudged.Co-rated data refers to the rating data of two users on the same item.For users without co-rated data,the metric learning recommendation model measures the similarity of users through similarity transmission chain,but when the chain is too long,the distance between similar users will be far,which affects the metric learning recommendation model to judge the similarity of users.Based on the above analysis,this article conducts research from the following two aspects:(1)Existing related recommendation models cannot fully offset the adverse effects of user rating preference on the recommendation results.For this,a Metric Learning Recommendation Model(MLUP)based on user preference is proposed.we distinguish item differences with item tag data,design a method to calculate the user's preference for items(hereinafter referred to as preference).The higher the preference,the higher the user's preference for the item.we use user preference to train the metric space instead of rating data,which alleviates the adverse impact of user rating preferences;By calculating the distance between the user and their non-interactive items in the metric space,convert the distance into a rating,and complete the rating prediction recommendation tasks.Comparative experiments and parameter analysis experiments are conducted on three real data sets.The experimental results show that the MLUP can effectively reduce the adverse effects of user rating preferences and provide more accurate rating prediction results.(2)When there is no co-rated data among users,the existing related recommendation models are prone to misjudge the similarity between users.In this regard,a Metric Learning Recommendation Model Combining User Tag And Rating Distribution(MLTD)is proposed.By designing a user similarity calculation method that does not rely on co-rated data,we describe the user similarity relationship from two different aspects,using all the historical rating data of users,and combining these two similarities to calculate the similarity between users without co-rated data;according to the similarity of users,we determine the distance relationship between users in the measurement space.The user relationship network is constructed in the metric space to avoid the user similarity transmission chain from being too long and reduce the misjudgment rate of user similarity.Comparative experiments and parameter analysis experiments are conducted on three real data sets.The experimental results show that the MLTD can effectively improve the accuracy of Top N recommendation and rating prediction tasks.