Research On Drug Recommendation Algorithm

In the problem of drug recommendation,there are a large number of users and drugs,and a user usually interacts with a few drugs because of disease need.So that the problem of data sparseness is serious.Related studies have found that incorporating auxiliary information into the basic model can alleviate this problem.Therefore,this thesis starts research from the perspective of user and drug,overcomes the problem of inaccurate recommendation results when using a sparse scoring matrix for recommendation,and proposes two improved hybrid models.The main work of this thesis is as follows:(1)Propose a hybrid recommendation algorithm based on drug semantics and user behavior(DSUB-MF).First,use the drug classification information to construct a drug-category matrix,and calculate the correlation between drugs;Then,the convolutional neural network is used to learn the text features of the medicine’s main functions,so as to obtain the complete semantic knowledge representation of medicine.Secondly,by analyzing the user’s drug purchase behavior sequence,combined with the classification vector of the drug,the user’s category preference vector value for the drug is obtained,so as to calculate the similarity of the user’s behavior preference.Finally,drug semantics and user behavior characteristics are added to the probability matrix decomposition algorithm to ensure the correctness of the user and drug hidden factor characteristics learned by the model.(2)Propose a hybrid recommendation algorithm based on user profile and neural network(HRS-UD).Aiming at the problem of lack of features related to the user’s drug preference,build a portrait for the user.Use the deep neural network to extract the original features of the user and the label features of the user portrait respectively to obtain the user’s deep-level feature vector;Use deep neural network and convolutional neural network to learn the semantic information of medicines,and obtain the semantic feature representation of medicines.Finally,the user and drug characteristics are input to the neural collaborative filtering model to fit the non-linear interaction relationship between the user and the drug,and perform score prediction.