Medical case knowledge recommendation method and system considering implicit feedback and human-computer interaction
By acquiring and processing historical behavior data from explicit and implicit feedback, and combining human-computer interaction and encryption technologies, the problems of protecting doctor user privacy and personalized recommendations in existing technologies have been solved, enabling accurate and flexible personalized case recommendations.
Patent Information
- Application Number
- CN202110765452.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Priority Date
- 2021-06-25
- Filing Date
- 2021-07-06
- Publication Date
- 2024-11-05
- Estimated Expiration
- 2041-07-06
AI Technical Summary
Existing medical case knowledge recommendation technologies cannot accurately recommend personalized cases to doctors while protecting their personal privacy, and lack effective human-computer interaction and personalized services.
By acquiring historical behavioral data of users browsing cases, including explicit and implicit feedback data, the data is processed using a hybrid recommendation model and encryption technology. Personalized recommendation lists are obtained by combining human-computer interaction behavior, and a differential privacy protection model is used to ensure privacy and security.
It enables doctors to receive accurate and personalized case recommendations while protecting the privacy of both doctors and patients, thus improving the accuracy and flexibility of the recommendations.
Smart Images

Figure CN113643817B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of medical data processing, and in particular to a medical case knowledge recommendation method and system considering implicit feedback and human-computer interaction. Background Art
[0002] The rapid development of the Internet has generated a large amount of data and information, and people are increasingly relying on this data and information. Most network applications recommend relevant information to people based on various recommendation systems and algorithms to meet the needs of different users. For example, in the field of medical health, based on the physiological data and disease characteristics of the human body, the case knowledge recommendation system can be used to match appropriate treatment cases from the case knowledge base and recommend them to the individual, and then treat the individual with reference to the case.
[0003] At present, in the field of medical health, the research and application of recommendation algorithms are mainly aimed at patients, and have a weak relationship with medical staff. On the one hand, when recommending cases to medical staff, due to the uneven levels and abilities of medical staff, different preferences, and the lack of necessary human-computer interaction, the existing technology cannot provide personalized case knowledge recommendation services for medical staff, and there are problems such as low accuracy and low user satisfaction; on the other hand, in order to ensure the accuracy of its recommendation results, case recommendation technology will collect various data information from doctor users and patients. If this data information cannot be properly kept, there will be a risk of privacy leakage.
[0004] This shows that the existing medical case knowledge recommendation technology cannot accurately recommend personalized cases to doctor users while protecting personal privacy. Summary of the invention
[0005] 1. Technical issues to be resolved
[0006] In view of the shortcomings of the existing technology, the present invention provides a medical case knowledge recommendation method and system that takes into account implicit feedback and human-computer interaction, which solves the problem that the existing medical case knowledge recommendation technology cannot protect personal privacy while accurately recommending personalized cases to doctor users.
[0007] (II) Technical solution
[0008] To achieve the above objectives, the present invention is implemented through the following technical solutions:
[0009] In a first aspect, the present invention first proposes a medical case knowledge recommendation method considering implicit feedback and human-computer interaction, the method comprising:
[0010] Obtaining historical behavior data of user browsing cases and performing preprocessing and encryption processing; the historical behavior data includes explicit feedback data and implicit feedback data;
[0011] Based on the explicit feedback data, a case rating matrix of the user and similarities between different cases are obtained; based on the implicit feedback data, a hybrid recommendation model of GMF and MLP fused with an attention mechanism is used to obtain a user-case interaction matrix; and the user-case interaction matrix and case data are encrypted;
[0012] Obtaining a preliminary recommendation list based on the similarity, the user's case scoring matrix, and the user-case interaction matrix;
[0013] Based on the preliminary recommendation list and in combination with user interaction behavior, a final recommendation list is obtained.
[0014] Preferably, the method further comprises: identifying and authenticating the user identity and authority when obtaining the case recommendation list.
[0015] Preferably, encrypting the user-case interaction matrix and case data comprises:
[0016] A differential privacy protection model based on random response is adopted to protect the user-case interaction matrix.
[0017] Preferably, the obtaining of the similarity between different cases includes: using modified cosine similarity and Pearson similarity to respectively calculate the similarity between different cases, and then fitting the two similarity calculation results according to a certain weight to obtain the final similarity, wherein:
[0018] The calculation formula of Pearson similarity is as follows:
[0019]
[0020] Among them, U represents the set of users who have rated both case i and case j;
[0021] R u,i represents the rating of case i by user U;
[0022] R u,j represents the rating of user U on case j;
[0023] represents the average score of case i;
[0024] represents the average score of case j;
[0025] The calculation formula of modified cosine similarity is as follows:
[0026]
[0027] Among them, U represents the set of users who have rated both case i and case j;
[0028] R u,i represents the rating of case i by user U;
[0029] R u,j represents the rating of user U on case j;
[0030] represents the average rating of user U;
[0031] Fitting the Pearson similarity and modified cosine similarity calculation results can calculate the final similarity Sim between different cases k (i,j), expressed as follows:
[0032] Sim k (i,j)=f(Sim r (i,j),Sim c (i,j))
[0033] Among them, Sim r (i,j) is the Pearson similarity between case i and case j; Sim c (i, j) is the modified cosine similarity between case i and case j.
[0034] Preferably, the method further comprises: before fitting the calculation results of Pearson similarity and modified cosine similarity, using SPSS to perform interaction analysis to obtain an interaction coefficient, and then determining a fitting method according to the value of the interaction coefficient,
[0035] If the interaction coefficient is greater than 0.05, the general linear regression equation is used to fit the Pearson similarity and modified cosine similarity calculation results:
[0036] Sim k (i,j)=a0+a1Sim r (i,j)+a2Sim c (i,j)
[0037] Among them, a0, a1, a2 are constants;
[0038] If the interaction coefficient is less than 0.05, add the interaction coefficient to fit the calculation results of Pearson similarity and modified cosine similarity:
[0039] Sim k (i,j)=a0+a1Sim r (i,j)+a2Sim c (i,j)+a3(Sim r (ij)Sim c (i,j))
[0040] Among them, a0, a1, a2 are constants, and a3 is the interaction coefficient.
[0041] In a second aspect, the present invention further proposes a medical case knowledge recommendation system considering implicit feedback and human-computer interaction, the system comprising:
[0042] A data acquisition and preprocessing module is used to acquire historical behavior data of user browsing cases and perform preprocessing and encryption processing; the historical behavior data includes explicit feedback data and implicit feedback data;
[0043] A data feature extraction and encryption module, used to obtain a user's case rating matrix and similarities between different cases based on the explicit feedback data, obtain a user-case interaction matrix based on the implicit feedback data using a hybrid recommendation model of GMF and MLP that incorporates an attention mechanism, and encrypt the user-case interaction matrix and case data;
[0044] A preliminary recommendation list acquisition module, used to acquire a preliminary recommendation list based on the similarity, the user's case scoring matrix, and the user-case interaction matrix;
[0045] The final recommendation list acquisition module is used to acquire the final recommendation list based on the preliminary recommendation list and in combination with the user interaction behavior.
[0046] Preferably, the system further comprises: an identity and authority identification and authentication module, which is used to identify and authenticate the user's identity and authority when obtaining the case recommendation list.
[0047] Preferably, the data feature extraction and encryption module encrypts the user-case interaction matrix and case data including:
[0048] A differential privacy protection model based on random response is adopted to protect the user-case interaction matrix.
[0049] Preferably, the data feature extraction and encryption module obtains the similarity between different cases, including: using modified cosine similarity and Pearson similarity to calculate the similarity between different cases respectively, and then fitting the two similarity calculation results according to a certain weight to obtain the final similarity, wherein
[0050] The calculation formula of Pearson similarity is as follows:
[0051]
[0052] Among them, U represents the set of users who have rated both case i and case j;
[0053] R u,i represents the rating of case i by user U;
[0054] R u,jrepresents the rating of user U on case j;
[0055] represents the average score of case i;
[0056] represents the average score of case j;
[0057] The calculation formula of modified cosine similarity is as follows:
[0058]
[0059] Among them, U represents the set of users who have rated both case i and case j;
[0060] R u,i represents the rating of case i by user U;
[0061] R u,j represents the rating of user U on case j;
[0062] represents the average rating of user U;
[0063] Fitting the Pearson similarity and modified cosine similarity calculation results can calculate the final similarity Sim between different cases k (i,j), expressed as follows:
[0064] Sim k (i,j)=f(Sim r (i,j),Sim c (i,j))
[0065] Among them, Sim r (i,j) is the Pearson similarity between case i and case j; Sim c (i, j) is the modified cosine similarity between case i and case j.
[0066] Preferably, the data feature extraction and encryption module further comprises: before fitting the calculation results of Pearson similarity and modified cosine similarity, using SPSS to perform interactive analysis to obtain an interactive coefficient, and then determining a fitting method according to the value of the interactive coefficient,
[0067] If the interaction coefficient is greater than 0.05, the general linear regression equation is used to fit the Pearson similarity and modified cosine similarity calculation results:
[0068] Sim k (i,j)=a0+a1Sim r (i,j)+a2Sim c (i,j)
[0069] Among them, a0, a1, a2 are constants;
[0070] If the interaction coefficient is less than 0.05, add the interaction coefficient to fit the calculation results of Pearson similarity and modified cosine similarity:
[0071] Sim k (i,j)=a0+a1Sim r (i,j)+a2Sim c (i,j)+a3(Sim r (ij)Sim c (i,j))
[0072] Among them, a0, a1, a2 are constants, and a3 is the interaction coefficient.
[0073] (III) Beneficial effects
[0074] The present invention provides a medical case knowledge recommendation method and system that considers implicit feedback and human-computer interaction. Compared with the prior art, it has the following beneficial effects:
[0075] 1. This technical solution obtains the historical behavior data of users browsing cases, including explicit feedback data and implicit feedback data, and pre-processes and encrypts these data, then extracts the features of explicit and implicit feedback data to obtain the user's case rating matrix and the similarity between different cases, as well as the user-case interaction matrix, and encrypts the extraction results of implicit data features again, and then obtains a preliminary recommendation list based on the similarity, the user's case rating matrix, and the user-case interaction matrix, and then obtains the final recommendation list in combination with the user's interactive behavior. When recommending cases to doctor users, this technical solution considers the impact of explicit and implicit feedback data on the accuracy of recommendation results, and uses a similarity calculation algorithm and a collaborative filtering algorithm fused with a multi-layer neural network to process explicit and implicit feedback data respectively. At the same time, in the later generation process of the case recommendation list, it strengthens the human-computer interaction operation, taking into account the feelings and decisions of doctor users when reading cases, so that the recommendation results are more accurate and flexible, meeting the personalized needs of doctor users;
[0076] 2. In this technical solution, the hybrid recommendation model of GMF and MLP that integrates the attention mechanism is used to process the acquired implicit feedback data to obtain the user-case interaction matrix. On the one hand, the implicit behavior of doctor users paying attention to cases is subdivided into 5 types according to the actual situation. On the other hand, the attention layer is integrated into the GMF and MLP hybrid recommendation model and the corresponding loss function is improved. In this way, the real interaction behavior between users and cases generated by the real hospital case recommendation process can be considered, and weights can be automatically assigned to each feature according to the importance of the feature, so that the final case recommendation result is more accurate;
[0077] 3. This technical solution identifies and authenticates the identities of doctor users who pre-acquire the recommendation list by setting up a user identity authentication program, encrypts the data containing patient privacy in the case using the encryption algorithm of the DES algorithm and the RSA algorithm, and encrypts the feature extraction results of implicit feedback data using a differential privacy protection model based on random response. This not only protects the personal privacy of doctor users, but also ensures the personalization and specificity of the case recommendation model, while also protecting the personal privacy of patients. BRIEF DESCRIPTION OF THE DRAWINGS
[0078] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0079] Figure 1 It is an overall flow chart of a medical case knowledge recommendation method considering implicit feedback and human-computer interaction in an embodiment of the present invention;
[0080] Figure 2 is a flow chart of processing explicit feedback data in an embodiment of the present invention;
[0081] Figure 3 is a process diagram of processing implicit feedback data in an embodiment of the present invention;
[0082] Figure 4 This is a flowchart of online recommendation cases for doctor users in an embodiment of the present invention. DETAILED DESCRIPTION
[0083] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention are clearly and completely described. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0084] The embodiments of the present application provide a medical case knowledge recommendation method and system that take implicit feedback and human-computer interaction into consideration, thereby solving the problem that existing medical case knowledge recommendation technologies are unable to protect personal privacy while accurately recommending personalized cases to doctor users, thereby achieving the purpose of protecting the privacy of doctor users and patients when recommending cases to doctors.
[0085] The technical solution in the embodiment of the present application is to solve the above technical problems, and the overall idea is as follows:
[0086] This technical solution collects, pre-processes and encrypts the historical behavior data of users browsing cases (including explicit feedback data and implicit feedback data), and then uses a similarity calculation algorithm and a collaborative filtering algorithm that integrates a multi-layer neural network with an attention mechanism to process the explicit and implicit feedback data respectively, and uses a differential privacy protection model to encrypt the data processing results. Then, based on the processing results of the explicit and implicit feedback data, combined with human-computer interaction behavior, a personalized and accurate case recommendation list suitable for the user is determined. In addition, in the process of case recommendation, user identity authentication is set to ensure that the user's personalized case recommendation model is not arbitrarily changed by other users, which once again strengthens the privacy protection of doctor users. This technical solution protects the personal privacy of users and patients while accurately recommending personalized cases to users, and has excellent overall performance.
[0087] In order to better understand the above technical solution, the above technical solution will be described in detail below in conjunction with the accompanying drawings and specific implementation methods.
[0088] Embodiment 1:
[0089] First, see Figure 1 The present invention first proposes a medical case knowledge recommendation method considering implicit feedback and human-computer interaction, the method comprising:
[0090] S1. Obtain historical behavior data of user browsing cases and perform preprocessing and encryption processing; the historical behavior data includes explicit feedback data and implicit feedback data;
[0091] S2. Based on the explicit feedback data, a case rating matrix of the user and similarities between different cases are obtained; based on the implicit feedback data, a hybrid recommendation model of GMF and MLP combined with an attention mechanism is used to obtain a user-case interaction matrix; and the user-case interaction matrix and case data are encrypted;
[0092] S3, obtaining a preliminary recommendation list based on the similarity, the user's case scoring matrix, and the user-case interaction matrix;
[0093] S4. Based on the preliminary recommendation list and in combination with the user interaction behavior, a final recommendation list is obtained.
[0094] It can be seen that the technical solution obtains the historical behavior data of users browsing cases, including explicit feedback data and implicit feedback data, and preprocesses and encrypts these data, and then extracts the features of the explicit and implicit feedback data to obtain the user's case rating matrix and the similarity between different cases, as well as the user-case interaction matrix, and encrypts the extraction results of the implicit data features again, and then obtains a preliminary recommendation list based on the similarity, the user's case rating matrix and the user-case interaction matrix, and then obtains the final recommendation list in combination with the user's interaction behavior. When recommending cases for doctor users, this technical solution considers the impact of explicit and implicit feedback data on the accuracy of recommendation results, and uses a similarity calculation algorithm and a collaborative filtering algorithm fused with a multi-layer neural network to process the explicit and implicit feedback data respectively. At the same time, in the later generation of the case recommendation list, the human-computer interaction operation is strengthened, and the feelings and decisions of doctor users when reading cases are taken into account, so that the recommendation results are more accurate and flexible, meeting the personalized needs of doctor users; in this technical solution, the hybrid recommendation model of GMF and MLP fused with the attention mechanism is used to process the acquired implicit feedback data to obtain the user-case interaction matrix. On the one hand, the implicit behavior of doctor users paying attention to cases is subdivided into 5 types according to the actual situation; on the other hand, in GMF and MLP, the implicit behavior of doctor users paying attention to cases is subdivided into 5 types according to the actual situation. The LP hybrid recommendation model integrates the attention layer and improves the corresponding loss function, which can take into account the real interaction behavior between users and cases generated by the real hospital case recommendation process, and can automatically assign weights to each feature according to the importance of the feature, so that the final case recommendation result is more accurate; this technical solution identifies and authenticates the identity of the doctor user who pre-acquires the recommendation list by setting a user identity authentication program, encrypts the data containing patient privacy in the case using encryption algorithms such as the DES algorithm and the RSA algorithm, and encrypts the feature extraction results of the implicit feedback data using a differential privacy protection model based on random response, which not only protects the personal privacy of the doctor user, but also ensures the personalization and specificity of the case recommendation model, while also protecting the personal privacy of the patient.
[0095] In an embodiment of the present invention, in order to protect the personalized case recommendation model of doctor users, a better processing method is that when implementing the medical case knowledge recommendation method that considers implicit feedback and human-computer interaction, the method also includes: identifying and authenticating the user identity and authority when obtaining the case recommendation list.
[0096] Secondly, in order to further enhance the protection of the personalized case recommendation model for doctor users, a better processing method is that the encryption of the user-case interaction matrix and case data includes:
[0097] A differential privacy protection model based on random response is adopted to protect the user-case interaction matrix.
[0098] In practice, in order to more accurately obtain the similarity between different cases browsed by the user, a better processing method is that when obtaining the similarity between different cases, it includes: using modified cosine similarity and Pearson similarity to respectively calculate the similarity between different cases, and then fitting the two similarity calculation results according to a certain weight to obtain the final similarity, wherein,
[0099] The calculation formula of Pearson similarity is as follows:
[0100]
[0101] Among them, U represents the set of users who have rated both case i and case j;
[0102] R u,i represents the rating of case i by user U;
[0103] R u,j represents the rating of user U on case j;
[0104] represents the average score of case i;
[0105] represents the average score of case j;
[0106] The calculation formula of modified cosine similarity is as follows:
[0107]
[0108] Among them, U represents the set of users who have rated both case i and case j;
[0109] R u,i represents the rating of case i by user U;
[0110] R u,j represents the rating of user U on case j;
[0111] represents the average rating of user U;
[0112] Fitting the Pearson similarity and modified cosine similarity calculation results can calculate the final similarity Sim between different cases k (i,j), expressed as follows:
[0113] Sim k (i,j)=f(Sim r (i,j),Sim c (i,j))
[0114] Among them, Sim r(i,j) is the Pearson similarity between case i and case j; Sim c (i, j) is the modified cosine similarity between case i and case j.
[0115] Furthermore, in order to obtain the final similarity between different cases based on Pearson similarity and modified cosine similarity, a better processing method is that the method further includes: before fitting the calculation results of Pearson similarity and modified cosine similarity, using SPSS to perform interaction analysis to obtain an interaction coefficient, and then determining a fitting method according to the value of the interaction coefficient,
[0116] If the interaction coefficient is greater than 0.05, the general linear regression equation is used to fit the Pearson similarity and modified cosine similarity calculation results:
[0117] Sim k (i,j)=a0+a1Sim r (i,j)+a2Sim c (i,j)
[0118] Among them, a0, a1, a2 are constants;
[0119] If the interaction coefficient is less than 0.05, add the interaction coefficient to fit the calculation results of Pearson similarity and modified cosine similarity:
[0120] Sim k (i,j)=a0+a1Sim r (i,j)+a2Sim c (i,j)+a3(Sim r (ij)Sim c (i,j))
[0121] Among them, a0, a1, a2 are constants, and a3 is the interaction coefficient.
[0122] The implementation process of an embodiment of the present invention is described in detail below in combination with the specific explanation of each step S1-S4.
[0123] S1. Obtain various historical behavior data of doctor users browsing cases and perform preprocessing and encryption processing; the historical behavior data includes explicit feedback data and implicit feedback data.
[0124] When users (mainly doctors) browse individual medical and health cases, a large amount of historical behavior data of browsing cases will be generated. These historical behavior data can be divided into explicit feedback data and implicit feedback data. Explicit feedback data refers to data that expresses clear meanings and can directly reflect users' preferences for products. For example, users' scoring and marking of cases can be used as explicit feedback data; implicit feedback data refers to users' indirect reflection of their preferences for products. Here, it refers to users' browsing behaviors of browsing cases, such as browsing and evaluating.
[0125] A large amount of historical behavior data of user browsing can be obtained through web crawling and other forms, and then these historical behavior data are preprocessed. The preprocessing process includes:
[0126] Read the user's historical behavior data and case scoring results.
[0127] Process missing values, outliers, and duplicate records in the data.
[0128] Continuous variables were normalized. No less than 30 pieces of data viewed by each doctor were retained. Each piece of data recorded should include the doctor ID, case release time, case score, browsing time, browsing content, number of views, and the time interval between previous and subsequent views of the same piece of information.
[0129] Vectorize the discrete data. Use One-Hot encoding to represent the doctor browsing data sequence, that is, only activate the click browsing information, and then record the corresponding reading time of each user. If the user browsing time exceeds 20 seconds, it is marked as 1, and if it is less than 20 seconds, it is marked as 0. See Table 1 and Table 2:
[0130] Table 1 Statistics of doctors browsing information
[0131] Case 1 Case 2 Case 3 Doctor A 1min 30s 15s Doctor B 45s Not reviewed 70s
[0132] According to the one-hot encoding, the value of each case is calculated to obtain the vector of the doctor. See Table 2, which shows the word vectors of Doctor A and Doctor B.
[0133] Table 2 Doctor's vector
[0134] Doctor A 1 1 0 Doctor B 1 0 1
[0135] In order to protect the privacy of patients, the DES algorithm and RSA algorithm are used to encrypt the case data containing patient secrets and privacy. Specifically:
[0136] The case data contains patient secrets and privacy of personal information (for example, the patient's name) data deformation processing, while other information in the case is not encrypted, which not only protects the patient's personal privacy, but also does not affect the recommendation value of the case. The DES algorithm is used to encrypt the data set, and the public key used by the DES algorithm is encrypted by the RSA algorithm. The packaged ciphertext and the encrypted public key are transmitted to the user. After the user obtains the data packet, the RSA decryption operation is performed on the key ciphertext to obtain the private key that can decrypt the DES algorithm encryption. Among them, the public key is responsible for encrypting and protecting the patient's personal medical records after diagnosis and treatment. The private key is responsible for decryption and restoring data information. The personal private key can only be obtained through the user's fingerprint or identity login. If other users want to access, they need to obtain the doctor's temporary key. At the same time, the doctor can set the validity period and access rights of the temporary key.
[0137] S2. Based on the explicit feedback data, the user's case rating matrix and the similarities between different cases are obtained; based on the implicit feedback data, a hybrid recommendation model of GMF and MLP is used to obtain a user-case interaction matrix; and the user-case interaction matrix and case data are encrypted.
[0138] 1) See Figure 2 , which is a flowchart of processing explicit feedback data in an embodiment of the present invention, processing the explicit feedback data to obtain the user's case rating matrix and the similarities between different cases, specifically includes:
[0139] 1.1) The server tracks the visit records of different doctors and obtains the doctors' interest information based on the doctors' scores for different cases. It also sets scoring principles and uses the obtained interest information to establish a case scoring matrix for each doctor's user.
[0140] The average score of each doctor, the number of users who scored each case, and the average score of each case are calculated based on the user's case scoring matrix. The average score of each doctor refers to the average of the scores given by a doctor to several cases; the number of users who scored each case refers to the number of all doctors who scored the same case; and the average score of each case refers to the average of the scores given by all doctors to the same case.
[0141] A common evaluation user number matrix is determined based on the rating information and calculation results to record the number of common users who have ratings for any two cases.
[0142] 1.2) Calculate the similarity between any two cases. Specifically, when calculating the similarity, cosine similarity and Pearson similarity are used to calculate respectively, and then the two similarity calculation results are matched according to a certain weight to obtain the final similarity.
[0143] The calculation formula of Pearson similarity is as follows:
[0144]
[0145] Among them, U represents the set of users who have rated both case i and case j;
[0146] R u,i represents the rating of case i by user U;
[0147] R u,j represents the rating of user U on case j;
[0148] represents the average score of case i;
[0149] represents the average score of case j.
[0150] The calculation formula of modified cosine similarity is as follows:
[0151]
[0152] Among them, U represents the set of users who have rated both case i and case j;
[0153] R u,i represents the rating of case i by user U;
[0154] R u,j represents the rating of user U on case j;
[0155] Represents the average rating of user U.
[0156] Using the Pearson similarity and modified cosine similarity formula, we can calculate the similarity Sim based on explicit feedback. k (i,j), the calculation formula is as follows:
[0157] Sim k (i,j)=f(Sim r (i,j),Sim c (i,j))
[0158] Among them, Sim r (i,j) is the Pearson similarity between case i and case j; Sim c (i, j) is the modified cosine similarity between case i and case j.
[0159] In order to determine whether there is a moderating effect between two similarities, and to flexibly select different regression equations for similarity fitting according to whether there is a moderating effect, so that the final similarity fitting result is more accurate, SPSS is used for interaction analysis before using the f function to fit the similarity, considering the interaction of the two similarities, and calculating the interaction coefficient value. If the interaction coefficient is greater than 0.05, it means that the significant level of interaction between the two similarities is not obvious, and the general linear regression equation is used for fitting:
[0160] Sim k (i,j)=a0+a1Sim r (i,j)+a2Sim c (i,j)
[0161] Among them, a0, a1, a2 are constants.
[0162] If the interaction coefficient is less than 0.05, it means that there is a regulatory effect between the two similarities. The interaction coefficient is added for fitting:
[0163] Sim k (i,j)=a0+a1Sim r (i,j)+a2Sim c (i,j)+a3(Sim r (ij)Sim c (i,j))
[0164] Among them, a0, a1, a2 are constants, and a3 is the interaction coefficient.
[0165] By calculating the cosine similarity and Pearson similarity between any two cases respectively and assigning weights to the two similarities according to the actual situation, the similarity between any two cases is finally obtained and stored.
[0166] 2) See Figure 3 When the implicit feedback data is processed using a hybrid recommendation model that integrates GMF and MLP with an attention mechanism to obtain the user-case interaction matrix, a personalized recommendation algorithm model is established to process the implicit feedback data. NCF is a general framework for collaborative filtering based on neural networks proposed on the basis of implicit feedback data. It learns interaction functions from data to simulate the potential characteristics of users and projects.
[0167] 2.1) Generation of training set. The training set consists of positive samples and negative samples, and each positive sample pos(y ui ) corresponds to 5 negative samples. The positive samples are all the interactive data (u,i), and 5 negative samples are generated for each positive sample (u,i), specifically:
[0168]
[0169] pos(y ui )'s oth (o=1,2,3,4,5)th negative sample
[0170] for And order
[0171] for And order
[0172] … … …
[0173] for And order
[0174] 2.2) Using the NCF general framework, the user-case interaction matrix Y obtained from implicit feedback data is defined as:
[0175] y ui =[y ui1 ,y ui2 ,y ui3 ,y ui4 ,y ui5 ] T
[0176]
[0177] Among them, y ui represents the element in the user u-case i interaction matrix Y, y uij The value is 1, indicating that user U has some interaction with case i; uij A value of 0 indicates that user U does not have any interaction with case i. The interaction between user U and case i specifically includes the user's browsing behavior on the case, commenting behavior on the case, forwarding behavior on the case, collecting behavior on the case, and disliking behavior on the case. Specifically:
[0178] (y ui1 ,y ui2 ,y ui3 ,y ui4 ,y ui5 ) = (browse, comment, forward, favorite, dislike)
[0179] 2.3) The processing of implicit feedback data includes the following steps:
[0180] 2.3.1) Input layer. Assume that the number of users is M and the number of cases is N. User (u) and case Item (i) are both one-hot encoded and converted into sparse vectors.
[0181] For example: the vector user of the i-th user i =[0,0,…,0,1,0,…,0],user i The dimension is 1*M,user i [i]=1,user i [j] = 0 (i≠j);
[0182] The vector item of the i-th case i =[0,0,…,0,1,0,…,0],item i The dimension is 1*N,item i [i]=1,item i [j]=0(i≠j).
[0183] 2.3.2) Embedding layer. i Vectors and items i The vectors are embedded into a space of smaller dimension. Specifically, let the embedding matrices of GMF and MLP be: (P G ,Q G ),(P M ,Q M ),in:
[0184]
[0185]
[0186]
[0187]
[0188] For user i With item i , the GMF embedding vector is The embedding vector of MLP is
[0189] 2.3.3)①GMF layer
[0190] (⊙ represents the inner product, )
[0191]
[0192]
[0193] ②MLP layer
[0194]
[0195] Where: relu(x) = max(0,x)
[0196] in,
[0197] 2.3.4) Attention layer. The attention mechanism is introduced into the neural network to identify the key features that contribute significantly to the prediction results and assign them higher weights. Correspondingly, unimportant features are assigned lower weights. Specifically, the attention layer is:
[0198]
[0199]
[0200] (where z2∈R 3×K , )
[0201]
[0202] (where z3∈R 3×K )
[0203] (where w∈R 5×K ,y ui ∈R 5×1 )
[0204] 2.3.5) Output layer:
[0205] (where w∈R 5×K , )
[0206]
[0207] 2.3.6) Calculation of loss function:
[0208]
[0209] 2.4) In order to protect the privacy of the personalized recommendation model for doctor users, a differential privacy protection model based on random response is used to encrypt and protect the personalized recommendation model for doctor users.
[0210] The processing results of the implicit feedback data output by the output layer of the above model (i.e., the final user-case interaction matrix, i.e., the personalized recommendation model for doctor users) are protected by a differential privacy protection model based on random response, thereby achieving the purpose of protecting the personalized recommendation model for doctor users. Add a randomization module and add differential privacy to the neural network. The specific steps are as follows:
[0211] Convert the output values into binary strings and standardize them;
[0212] Randomizing each of the m binary strings will result in a privacy loss of m*ε. In the random response, the difference between adjacent inputs is d, and the sensitivity is d. Randomizing the binary string results in a privacy loss of the input value ε.
[0213] Among them, the length of the binary string is L = (p + q + 1), p represents the number of binary digits in the integer part, q represents the number of binary digits in the decimal part, so the length of the combined binary string is L*s, and s is the number of convolutional layer outputs. The difference between two consecutive inputs is at most L*s, so the sensitivity is L*s, and the randomization probability is:
[0214]
[0215] 2.5) By splicing and combining the last hidden layer of the GMF model and the MLP model, the processing results of the two models for implicit feedback data are obtained, that is, the final user-case interaction matrix is obtained to represent the personalized matrix of user preferences and habits, and the result is stored.
[0216] S3. Obtain a preliminary recommendation list based on the similarity, the user's case scoring matrix, and the user-case interaction matrix.
[0217] See also Figure 4 , when pushing the preliminary recommendation list to the user, the following process is included:
[0218] For each case, according to the similarity obtained above, the cases are sorted from high to low, and m cases with similarity greater than the threshold are selected from the sorted sequence to form the most recent case set, so as to achieve quality control of the cases. The default value of the threshold can be set manually according to the doctor's long-term reading habits, and can be manually changed according to the actual situation. If the number of cases with a final similarity greater than the threshold exceeds m, then m cases are selected from high to low. When recommending cases to the target user, first, according to the user's case scoring matrix obtained above, the user's scored cases and unscored cases are obtained. For the scored cases, if the score is higher than the k value (the default value of the K value is set manually, and can also be changed manually according to the doctor's long-term reading habits, and can be changed manually) and belongs to the most recent case set, this case can be recommended first. Through the previous scoring records, the final score of the case is calculated, and then the similarity of different cases is combined to obtain the grade of each case, which can be subdivided into grade A, grade B, and grade C. Among them, Grade A indicates cases with high similarity and high case scores, which are worth reading for all user groups; Grade B indicates cases with high similarity but low case scores or low similarity but high case scores, which are worth reading for a specific user group; Grade C indicates cases with low similarity and low case scores, which will not be automatically recommended but users can check them by themselves.
[0219] After classifying the cases, combined with the GMF and MLP models established above, the doctor's diagnostic preferences are obtained, and the similarity calculation results and neural network deep learning results are combined to obtain a preliminary recommendation list. For example, when obtaining the preliminary recommendation list, users can first obtain similar case retrieval according to their needs, and then filter out cases that meet their personalized requirements based on the existing personalized recommendation model.
[0220] S4. Based on the preliminary recommendation list and in combination with the user interaction behavior, a final recommendation list is obtained.
[0221] See also Figure 4 , based on the preliminary recommendation list and the interaction behavior of the doctor user, the final recommendation list is obtained. Several interaction behaviors and methods of doctor users include:
[0222] 1) Doctors actively screen the preliminary recommendation list. Collect relevant information about doctors actively intervening in the preliminary recommendation list, such as removing, screening, and sorting unnecessary cases in the preliminary recommendation list, and actively replying to comments on cases in the preliminary list (actively replying to comments means that the doctor user is interested in this case, and the priority in the preliminary recommendation list will be adjusted, etc.);
[0223] 2) Identify and process multimodal information such as voice, gestures, and images during real-time consultations by doctors, and adjust the preliminary recommendation list based on these interactive information. Real-time transmission and processing of doctor users' feelings and feedback after receiving the preliminary case recommendation list service saves time and energy in processing a large amount of case information and manual operations during consultations. Voice can be the retrieval or evaluation of cases; gestures can be the doctor user's feedback on the satisfaction of the case and the capture of the preference habits during consultations; images can be case photos uploaded and shared by doctors in real time. For example, the voice is extracted by the microphone through voice recognition technology, and the evaluation of the case by the doctor user and the patient is obtained, and then the preliminary recommendation list is adjusted based on these evaluations; the millimeter wave radar technology is used to collect the user's feedback on the satisfaction of the case, so that specific instructions can be implemented to adjust the preliminary recommendation list; image recognition technology is used to identify the case photos and images uploaded by the doctor in real time, and the poor quality images are enhanced and reconstructed to improve the image quality, and then intervene in the preliminary recommendation list.
[0224] Combined with the interactive information collected above, the preliminary recommendation list can be updated for a second time to obtain a personalized final recommendation list with doctor preferences and behavioral habits.
[0225] In addition, since each doctor user has his or her own personal habits and preferences, the case recommendation list obtained according to his or her personal habits and preferences is the most suitable for the user. Therefore, in order to ensure the exclusivity and privacy of the case recommendation list of the doctor user, the medical case knowledge recommendation method considering implicit feedback and human-computer interaction in this embodiment also includes:
[0226] S5. Identify and authenticate the user's identity and authority when obtaining the case recommendation list.
[0227] A user identity and authority identification step is set. When a doctor user consults again, the doctor user can only obtain the case recommendation list after passing the identity authentication. This ensures that the case recommendation model of each doctor user will not be tampered with due to the random data input of other doctor users, and the case recommendation list will not be viewed by others unexpected to the doctor user. This strengthens the protection of the specificity, accuracy and privacy of the doctor user's personalized case recommendation model and personalized case recommendation list. Specifically, the identity and authority identification step includes but is not limited to password login, fingerprint recognition, facial recognition, and iris recognition. For example, when fingerprint recognition is used to authenticate user identity authority, the doctor user enters his or her fingerprint to obtain a recommendation list that meets his or her personalized needs.
[0228] Embodiment 2:
[0229] In a second aspect, the present invention further proposes a medical case knowledge recommendation system considering implicit feedback and human-computer interaction, the system comprising:
[0230] A data acquisition and preprocessing module is used to acquire historical behavior data of user browsing cases and perform preprocessing and encryption processing; the historical behavior data includes explicit feedback data and implicit feedback data;
[0231] A data feature extraction and encryption module, used to obtain a user's case rating matrix and similarities between different cases based on the explicit feedback data, obtain a user-case interaction matrix based on the implicit feedback data using a hybrid recommendation model of GMF and MLP that incorporates an attention mechanism, and encrypt the user-case interaction matrix and case data;
[0232] A preliminary recommendation list acquisition module, used to acquire a preliminary recommendation list based on the similarity, the user's case scoring matrix, and the user-case interaction matrix;
[0233] The final recommendation list acquisition module is used to acquire the final recommendation list based on the preliminary recommendation list and in combination with the user interaction behavior.
[0234] Furthermore, the system also includes: an identity and authority identification and authentication module, which is used to identify and authenticate the user's identity and authority when obtaining the case recommendation list.
[0235] Furthermore, the data feature extraction and encryption module encrypts the user-case interaction matrix and case data, including:
[0236] A differential privacy protection model based on random response is adopted to protect the user-case interaction matrix.
[0237] Furthermore, the data feature extraction and encryption module obtains the similarity between different cases, including: using modified cosine similarity and Pearson similarity to calculate the similarity between different cases respectively, and then fitting the two similarity calculation results according to a certain weight to obtain the final similarity, wherein
[0238] The calculation formula of Pearson similarity is as follows:
[0239]
[0240] Among them, U represents the set of users who have rated both case i and case j;
[0241] R u,i represents the rating of case i by user U;
[0242] R u,j represents the rating of user U on case j;
[0243] represents the average score of case i;
[0244] represents the average score of case j;
[0245] The calculation formula of modified cosine similarity is as follows:
[0246]
[0247] Among them, U represents the set of users who have rated both case i and case j;
[0248] R u,i represents the rating of case i by user U;
[0249] R u,j represents the rating of user U on case j;
[0250] represents the average rating of user U;
[0251] Fitting the Pearson similarity and modified cosine similarity calculation results can calculate the final similarity Sim between different cases k (i,j), expressed as follows:
[0252] Sim k (i,j)=f(Sim r (i,j),Sim c (i,j))
[0253] Among them, Sim r (i,j) is the Pearson similarity between case i and case j; Sim c (i, j) is the modified cosine similarity between case i and case j.
[0254] Furthermore, the data feature extraction and encryption module further includes: before fitting the calculation results of Pearson similarity and modified cosine similarity, using SPSS to perform interaction analysis to obtain an interaction coefficient, and then determining a fitting method according to the value of the interaction coefficient,
[0255] If the interaction coefficient is greater than 0.05, the general linear regression equation is used to fit the Pearson similarity and modified cosine similarity calculation results:
[0256] Sim k (i,j)=a0+a1Sim r (i,j)+a2Sim c (i,j)
[0257] Among them, a0, a1, a2 are constants;
[0258] If the interaction coefficient is less than 0.05, add the interaction coefficient to fit the calculation results of Pearson similarity and modified cosine similarity:
[0259] Sim k (i,j)=a0+a1Sim r (i,j)+a2Sim c (i,j)+a3(Sim r (ij)Sim c (i,j))
[0260] Among them, a0, a1, a2 are constants, and a3 is the interaction coefficient.
[0261] It is understandable that the medical case knowledge recommendation system considering implicit feedback and human-computer interaction provided by the embodiment of the present invention corresponds to the medical case knowledge recommendation method considering implicit feedback and human-computer interaction. The explanation, examples, beneficial effects, etc. of the relevant contents can refer to the corresponding contents in the medical case knowledge recommendation method considering implicit feedback and human-computer interaction, and will not be repeated here.
[0262] In summary, compared with the prior art, the present invention has the following beneficial effects:
[0263] 1. This technical solution obtains the historical behavior data of users browsing cases, including explicit feedback data and implicit feedback data, and pre-processes and encrypts these data, then extracts the features of explicit and implicit feedback data to obtain the user's case rating matrix and the similarity between different cases, as well as the user-case interaction matrix, and encrypts the extraction results of implicit data features again, and then obtains a preliminary recommendation list based on the similarity, the user's case rating matrix, and the user-case interaction matrix, and then obtains the final recommendation list in combination with the user's interactive behavior. When recommending cases to doctor users, this technical solution considers the impact of explicit and implicit feedback data on the accuracy of recommendation results, and uses a similarity calculation algorithm and a collaborative filtering algorithm fused with a multi-layer neural network to process explicit and implicit feedback data respectively. At the same time, in the later generation process of the case recommendation list, it strengthens the human-computer interaction operation, taking into account the feelings and decisions of doctor users when reading cases, so that the recommendation results are more accurate and flexible, meeting the personalized needs of doctor users;
[0264] 2. In this technical solution, the hybrid recommendation model of GMF and MLP that integrates the attention mechanism is used to process the acquired implicit feedback data to obtain the user-case interaction matrix. On the one hand, the implicit behavior of doctor users paying attention to cases is subdivided into 5 types according to the actual situation. On the other hand, the attention layer is integrated into the GMF and MLP hybrid recommendation model and the corresponding loss function is improved. In this way, the real interaction behavior between users and cases generated by the real hospital case recommendation process can be considered, and weights can be automatically assigned to each feature according to the importance of the feature, so that the final case recommendation result is more accurate;
[0265] 3. This technical solution identifies and authenticates the identities of doctor users who pre-acquire the recommendation list by setting up a user identity authentication program, encrypts the data containing patient privacy in the case using the encryption algorithm of the DES algorithm and the RSA algorithm, and encrypts the feature extraction results of implicit feedback data using a differential privacy protection model based on random response. This not only protects the personal privacy of doctor users, but also ensures the personalization and specificity of the case recommendation model, while also protecting the personal privacy of patients.
[0266] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.
[0267] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A medical case knowledge recommendation method considering implicit feedback and human-computer interaction, characterized in that: The method comprises: Obtaining historical behavior data of user browsing cases and performing preprocessing and encryption processing; the historical behavior data includes explicit feedback data and implicit feedback data; Based on the explicit feedback data, a case rating matrix of the user and similarities between different cases are obtained; based on the implicit feedback data, a hybrid recommendation model of GMF and MLP fused with an attention mechanism is used to obtain a user-case interaction matrix; and the user-case interaction matrix and case data are encrypted; Obtaining a preliminary recommendation list based on the similarity, the user's case scoring matrix, and the user-case interaction matrix; Based on the preliminary recommendation list, a final recommendation list is obtained in combination with user interaction behavior; Wherein, obtaining a user-case interaction matrix based on the implicit feedback data by using a hybrid recommendation model of GMF and MLP integrated with an attention mechanism, and encrypting the user-case interaction matrix and case data, including: 2.1) Generate a training set, which consists of positive samples and negative samples, and each positive sample pos(y ui ) corresponds to 5 negative samples, and the positive samples are all the interactive data (u,i). For each positive sample (u,i), 5 negative samples are generated. Specifically: pos(y ui ) is the oth negative sample Where o = 1, 2, 3, 4, 5 for And order for And order LLL for And order 2.2) Using the NCF general framework, the user-case interaction matrix Y obtained from implicit feedback data is defined as: and ui =[and ui1 ,and ui2 ,and ui3 ,and ui4 ,and ui5 ] T Among them, y ui represents the element in the user u-case i interaction matrix Y, y uij The value is 1, indicating that user U has some interaction with case i; uij A value of 0 indicates that there is no interaction between user U and case i. The interaction between user U and case i includes browsing, commenting, forwarding, collecting, and disliking of the case. Specifically: (y ui1 ,y ui2 ,y ui3 ,y ui4 ,y ui5 ) = (browse, comment, forward, favorite, dislike) 2.3) The processing of implicit feedback data by the hybrid recommendation model of GMF and MLP includes the following process: 2.3.1) The input layer converts both the user User (u) and the case Item (i) into sparse vectors through one-hot encoding; The vector user of the i-th user i =[0,0,…,0,1,0,…,0],user i The dimension is 1*M,user i [i]=1,user i [j] = 0, i ≠ j; The vector item of the i-th case i =[0,0,…,0,1,0,…,0],item i The dimension is 1*N,item i [i]=1,item i [j] = 0, i≠j; M is the number of users, N is the number of cases; 2.3.2) Embedding layer puts user i Vectors and items i The vectors are embedded into a space of smaller dimension. Specifically, let the embedding matrices of GMF and MLP be: (P G ,Q G ),(P M ,Q M ),in: For user i With item i , the GMF embedding vector is The embedding vector of MLP is 2.3.3)①GMF layer Where e represents the inner product, ②MLP layer in, Where: relu(x) = max(0,x) in, 2.3.4) Attention layer introduces the attention mechanism into the neural network to identify key features that contribute significantly to the prediction results. Specifically, the attention layer is: Where z2∈R 3×K , Among them, z3∈R 3×K Where w∈R 5×K ,y ui ∈R 5×1 2.3.5) Output layer: in, 2.3.6) Calculation of loss function: 2.4) The doctor-user personalized recommendation model is encrypted and protected using a differential privacy protection model based on random responses, including: The processing results of the implicit feedback data output by the output layer are protected by a differential privacy protection model based on random response, so as to achieve the purpose of protecting the personalized recommendation model of doctor users. A randomization module is added to add differential privacy to the neural network. The specific steps are as follows: Convert the output values into binary strings and standardize them; Randomize each of the m binary strings, resulting in a privacy loss of m*ε. In a random response, the difference between adjacent inputs is d, and the sensitivity is d. Randomize the binary string, and the privacy loss is the input value ε. Among them, the length of the binary string L = (p + q + 1), p represents the number of binary digits of the integer part, q represents the number of binary digits of the decimal part, so the length of the combined binary string is L*s, s is the number of convolutional layer outputs; the difference between two consecutive inputs is at most L*s, so the sensitivity is L*s, and the randomization probability is:
2. The method according to claim 1, characterized in that The method also includes: identifying and authenticating the user identity and authority when obtaining the case recommendation list.
3. The method according to claim 1, characterized in that The obtaining of the similarity between different cases includes: using modified cosine similarity and Pearson similarity to respectively calculate the similarity between different cases, and then fitting the two similarity calculation results according to a certain weight to obtain the final similarity, wherein: The calculation formula of Pearson similarity is as follows: Among them, U represents the set of users who have rated both case i and case j; R u,i represents the rating of case i by user U; R u,j represents the rating of user U on case j; represents the average score of case i; represents the average score of case j; The calculation formula of modified cosine similarity is as follows: Among them, U represents the set of users who have rated both case i and case j; R u,i represents the rating of case i by user U; R u,j represents the rating of user U on case j; represents the average rating of user U; Fitting the Pearson similarity and modified cosine similarity calculation results can calculate the final similarity Sim between different cases k (i,j), expressed as follows: Sim k (i,j)=f(Sim r (i,j),Sim c (i,j)) Among them, Sim r (i,j) is the Pearson similarity between case i and case j; Sim c (i, j) is the modified cosine similarity between case i and case j.
4. The method according to claim 3, characterized in that The method further comprises: before fitting the calculation results of the Pearson similarity and the modified cosine similarity, using SPSS to perform interaction analysis to obtain an interaction coefficient, and then determining a fitting method according to the value of the interaction coefficient, If the interaction coefficient is greater than 0.05, the general linear regression equation is used to fit the Pearson similarity and modified cosine similarity calculation results: Sim k (i,j)=a0+a1Sim r (i,j)+a2Sim c (i,j) Among them, a0, a1, a2 are constants; If the interaction coefficient is less than 0.05, add the interaction coefficient to fit the calculation results of Pearson similarity and modified cosine similarity: Sim k (i,j)=a0+a1Sim r (i,j)+a2Sim c (i,j)+a3(Sim r (i.j)Sim c (i,j)) Among them, a0, a1, a2 are constants, and a3 is the interaction coefficient.
5. A medical case knowledge recommendation system considering implicit feedback and human-computer interaction, characterized in that: The system comprises: A data acquisition and preprocessing module is used to acquire historical behavior data of user browsing cases and perform preprocessing and encryption processing; the historical behavior data includes explicit feedback data and implicit feedback data; A data feature extraction and encryption module, used to obtain a user's case rating matrix and similarities between different cases based on the explicit feedback data, obtain a user-case interaction matrix based on the implicit feedback data using a hybrid recommendation model of GMF and MLP that incorporates an attention mechanism, and encrypt the user-case interaction matrix and case data; A preliminary recommendation list acquisition module, used to acquire a preliminary recommendation list based on the similarity, the user's case scoring matrix, and the user-case interaction matrix; A final recommendation list acquisition module, used to acquire a final recommendation list based on the preliminary recommendation list and in combination with user interaction behavior; Wherein, obtaining a user-case interaction matrix based on the implicit feedback data by using a hybrid recommendation model of GMF and MLP integrated with an attention mechanism, and encrypting the user-case interaction matrix and case data, including: 2.1) Generate a training set, which consists of positive samples and negative samples, and each positive sample pos(y ui ) corresponds to 5 negative samples, and the positive samples are all the interactive data (u,i). For each positive sample (u,i), 5 negative samples are generated. Specifically: pos(y ui ) is the oth negative sample Where o = 1, 2, 3, 4, 5 for And order for And order LLL for And order 2.2) Using the NCF general framework, the user-case interaction matrix Y obtained from implicit feedback data is defined as: and ui =[and ui1 ,and ui2 ,and ui3 ,and ui4 ,and ui5 ] T Among them, y ui represents the element in the user u-case i interaction matrix Y, y uij The value is 1, indicating that user U has some interaction with case i; uij A value of 0 indicates that there is no interaction between user U and case i. The interaction between user U and case i includes browsing, commenting, forwarding, collecting, and disliking of the case. Specifically: (y ui1 ,y ui2 ,y ui3 ,y ui4 ,y ui5 ) = (browse, comment, forward, favorite, dislike) 2.3) The processing of implicit feedback data by the hybrid recommendation model of GMF and MLP includes the following process: 2.3.1) The input layer converts both the user User (u) and the case Item (i) into sparse vectors through one-hot encoding; The vector user of the i-th user i =[0,0,…,0,1,0,…,0],user i The dimension is 1*M,user i [i]=1,user i [j] = 0, i ≠ j; The vector item of the i-th case i =[0,0,…,0,1,0,…,0],item i The dimension is 1*N,item i [i]=1,item i [j] = 0, i≠j; M is the number of users, N is the number of cases; 2.3.2) Embedding layer puts user i Vectors and items i The vectors are embedded into a space of smaller dimension. Specifically, let the embedding matrices of GMF and MLP be: (P G ,Q G ),(P M ,Q M ),in: For user i With item i , the GMF embedding vector is The embedding vector of MLP is 2.3.3)①GMF layer Where e represents the inner product, ②MLP layer in, Where: relu(x) = max(0,x) in, 2.3.4) Attention layer introduces the attention mechanism into the neural network to identify key features that contribute significantly to the prediction results. Specifically, the attention layer is: Where z2∈R 3×K , Among them, z3∈R 3×K Where w∈R 5×K ,y ui ∈R 5×1 2.3.5) Output layer: in, 2.3.6) Calculation of loss function: 2.4) The doctor-user personalized recommendation model is encrypted and protected using a differential privacy protection model based on random responses, including: The processing results of the implicit feedback data output by the output layer are protected by a differential privacy protection model based on random response, so as to achieve the purpose of protecting the personalized recommendation model of doctor users. A randomization module is added to add differential privacy to the neural network. The specific steps are as follows: Convert the output values into binary strings and standardize them; Randomize each of the m binary strings, resulting in a privacy loss of m*ε. In a random response, the difference between adjacent inputs is d, and the sensitivity is d. Randomize the binary string, and the privacy loss is the input value ε. Among them, the length of the binary string L = (p + q + 1), p represents the number of binary digits of the integer part, q represents the number of binary digits of the decimal part, so the length of the combined binary string is L*s, s is the number of convolutional layer outputs; the difference between two consecutive inputs is at most L*s, so the sensitivity is L*s, and the randomization probability is:
6. The system according to claim 5, characterized in that The system also includes: an identity and authority identification and authentication module, which is used to identify and authenticate the user's identity and authority when obtaining the case recommendation list.
7. The system according to claim 5, characterized in that The data feature extraction and encryption module obtains the similarity between different cases, including: using modified cosine similarity and Pearson similarity to calculate the similarity between different cases respectively, and then fitting the two similarity calculation results according to a certain weight to obtain the final similarity, wherein The calculation formula of Pearson similarity is as follows: Among them, U represents the set of users who have rated both case i and case j; R u,i represents the rating of case i by user U; R u,j represents the rating of user U on case j; represents the average score of case i; represents the average score of case j; The calculation formula of modified cosine similarity is as follows: Among them, U represents the set of users who have rated both case i and case j; R u,i represents the rating of case i by user U; R u,j represents the rating of user U on case j; represents the average rating of user U; Fitting the Pearson similarity and modified cosine similarity calculation results can calculate the final similarity Sim between different cases k (i,j), expressed as follows: Sim k (i,j)=f(Sim r (i,j),Sim c (i,j)) Among them, Sim r (i,j) is the Pearson similarity between case i and case j; Sim c (i, j) is the modified cosine similarity between case i and case j.
8. The system according to claim 7, characterized in that The data feature extraction and encryption module also includes: before fitting the calculation results of Pearson similarity and modified cosine similarity, using SPSS to perform interactive analysis to obtain an interactive coefficient, and then determining a fitting method according to the value of the interactive coefficient, If the interaction coefficient is greater than 0.05, the general linear regression equation is used to fit the Pearson similarity and modified cosine similarity calculation results: Sim k (i,j)=a0+a1Sim r (i,j)+a2Sim c (i,j) Among them, a0, a1, a2 are constants; If the interaction coefficient is less than 0.05, add the interaction coefficient to fit the calculation results of Pearson similarity and modified cosine similarity: Sim k (i,j)=a0+a1Sim r (i,j)+a2Sim c (i,j)+a3(Sim r (i.j)Sim c (i,j)) Among them, a0, a1, a2 are constants, and a3 is the interaction coefficient.
