Doctor recommendation method based on inquiry service quality and fairness and medium

Abstract

The invention relates to the technical field of internet medical treatment, and provides a doctor recommendation method based on inquiry service quality and fairness and a medium. The method provided by the invention mainly comprises the following steps: S1, according to a disease problem uploaded by each patient, personal basic attribute information data of each doctor, and historical answer and comment data of each doctor, obtaining attribute features, interaction features and comment features, and further obtaining a doctor inquiry service quality score by using a factorization machine; s2, constructing a fair value function, obtaining a system exposure level and a target exposure level of a doctor according to the position model, and further obtaining a fairness constraint of a doctor end; and S3, constructing a doctor recommendation model, obtaining a final doctor recommendation set by using a multi-arm machine method based on improved Thompson sampling, and recommending the doctor recommendation set to the patient. According to the invention, by considering the inquiry service quality and the fairness constraint, the doctor recommendation accuracy is ensured, and the fairness of the recommendation system is effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of Internet medical technology, and in particular to a doctor recommendation method and medium based on consultation service quality and fairness. Background Art

[0002] Recommending a suitable doctor to a patient not only focuses on whether the doctor has the ability to answer the patient's questions, but also on the quality of the doctor's consultation service. However, it is impossible to directly infer the quality level of the doctor's consultation service. It is necessary to extract quality features more comprehensively and accurately from multiple dimensions and construct a dynamic evaluation method. The decision-making goal of the model is to recommend high-quality doctors to patients more accurately and reliably. Only focusing on the accuracy of the recommendation will ignore the discrimination and unfairness in the decision-making process. The existence of unfairness has brought new challenges to the design of platform recommendation algorithms, that is, not only to meet the accuracy of the recommendation, but also to ensure the fairness of the recommendation. However, the existing medical service recommendations rarely take into account the impact of the "Matthew effect" on some online doctors. The doctor recommendation mechanism is optimized by establishing fairness constraints.

[0003] Existing research on doctor recommendation is mainly conducted from the perspectives of patient preferences, doctor professional skills, doctor-patient relationship, medical data sparsity and patient privacy protection. It is implemented using collaborative filtering, hybrid recommendation, matrix decomposition recommendation, multi-attribute criterion analysis, discrete choice model, deep learning model and other technologies. Existing recommendation technology has not fully considered the dynamic characteristics of consultation service quality, and has not paid enough attention to the fairness optimization method of doctor ranking. The cross sorting proposed by the previous invention still has a certain degree of contingency in improving the fairness of recommendation.

[0004] Chinese invention patent: Publication number is "CN116705253A", and the name is "Online consultation service doctor recommendation method oriented to the principle of fairness", which discloses an online consultation service doctor recommendation method oriented to the principle of fairness, including the following steps: S1: Generate doctor feature data based on the doctor's consultation history data; S2: Extract and analyze the patient's description to generate the patient's preference feature data and disease feature data; S3: Match the patient's disease feature with the doctor's feature to obtain a professional matching candidate doctor set; S4: Match the patient's preference feature data with the doctor feature data in the candidate doctor set, and divide the candidate doctor set into a list of doctors that meet all preferences and a list of alternative doctors that meet some preferences; S5: Calculate the weight value to obtain the final weight value of each doctor; S6: Cross-sort the doctors and send them to the patient's terminal device. This technical solution can meet the patient's interests and preference requirements when recommending doctors, and can improve the accuracy and fairness of the doctors recommended to the patient. However, this technical solution only puts the doctors who can fully meet all the patient's preferences, that is, doctors with a large number of consultations and long experience, into one list; and puts the doctors who partially meet the patient's preferences, that is, new doctors or doctors with a small number of consultations, into one list, and cross-sorts the two lists and recommends them to the patients. Such a technical solution can only achieve limited fairness and will result in a mismatch between the doctors the patients need.

[0005] Chinese invention patent: Publication number is "CN115050457A", and the name is "Online consultation service quality evaluation method, device, equipment, medium and product", which discloses an online consultation service quality evaluation method, device, equipment, medium and product, and the method includes: obtaining consultation information flow data, wherein the consultation information flow data includes: question information flow data corresponding to the target object and reply information flow data corresponding to the service object; based on the question time in the question information flow data and the first reply time in the reply information flow data, determine the first evaluation parameter of the service object; based on the question content in the question information flow data and the reply content in the reply information flow data, determine the second evaluation parameter and the third evaluation parameter of the service object; based on the first evaluation parameter, the second evaluation parameter and the third evaluation parameter, determine the service quality score of the service object. This technical solution solves the defect of inaccurate evaluation results of service quality caused by unilateral evaluation of housekeepers and doctors. However, the evaluation parameters of this technical solution only consider the time difference between doctors and patients asking questions and providing responses, as well as the matching degree of the response content. It does not consider the doctor’s exposure and the matching degree between the doctor’s expertise and the patient’s consultation questions, and does not solve the problem of lack of fairness in recommending doctors.

[0006] Chinese invention patent: Publication number is "CN114037321B", and the name is "Method and device for recommending crowdsourcing testers for fairness". It discloses a method and device for recommending crowdsourcing testers for fairness, including representing the demand description of a crowdsourcing task as a descriptive term vector; predicting the defect detection probability of each recommended person in the crowdsourcing task based on the descriptive term vector and the attribute information of the recommended person; setting several fairness goals including maximizing the probability of defect discovery and minimizing the difference in recommendation frequency, and performing multi-objective optimization to obtain the crowdsourcing tester recommendation results. This technical solution can alleviate the problem of unfairness and produce fairer recommendations. Moreover, on the basis of producing fairer recommendations, this technical solution method does not affect the defect detection efficiency, but also improves the defect detection efficiency. However, this technical solution does not take into account the matching degree between the doctor's expertise and the patient's consulting questions, and cannot be applied to the field of doctor recommendation in Internet medical care. Summary of the invention

[0007] In order to solve the problem that it is difficult to strike a balance between fairness and accuracy when recommending doctors in the above-mentioned prior art, the present invention proposes a doctor recommendation method and medium based on consultation service quality and fairness, aiming to effectively improve the fairness of the recommendation system while ensuring the accuracy of the recommendation by balancing consultation service quality and fairness constraints.

[0008] The present invention is achieved through the following technical solution: comprising the following steps:

[0009] S1. Obtain attribute feature F based on the disease problem uploaded by each patient, each doctor’s personal basic attribute information data, and each doctor’s historical answers and comments data A , interaction feature F I And the comment feature F R , and then use the factorization machine to obtain the doctor consultation service quality score Sco(d);

[0010] S2. Construct a fair value function, obtain the doctor's system exposure level E and target exposure level E* according to the location model, and then obtain the fairness constraint Fairness on the doctor side;

[0011] S3. Build a doctor recommendation model and use the multi-armed bandit method based on improved Thompson sampling to obtain the final doctor recommendation set and recommend it to patients.

[0012] As a further preference, the step S1 includes the following specific steps:

[0013] S1-1. Based on each doctor’s personal basic attribute information data, use unique hot encoding to extract attribute features F A And vectorized representation;

[0014] S1-2. Based on the disease questions uploaded by each patient and the historical answers of each doctor, text mining and sentiment analysis methods are used to construct interactive features of online doctor-patient behavior from information support, emotional support and response speed based on communication adaptation theory, signal theory and social presence theory. I ;

[0015] S1-3. Based on the patient comment data, the graph convolutional neural network is used to obtain the comment feature F R ;

[0016] S1-4. According to the attribute feature F A , interaction feature F I And the comment feature F R , introduce the time decay mechanism and time window mechanism, use the factor decomposition machine to get the doctor consultation service quality score Sco(d), the attribute feature F A , interaction feature F I And the comment feature F R For quality characteristics.

[0017] As a further preferred embodiment, the doctor consultation service quality score Sco(d) in step S1-4 is calculated as follows:

[0018]

[0019] Where d is the dth doctor, F a , F b ∈{F A , F I , F R}, ω0 is a model parameter, indicating the global bias, ω a represents the weight of the ath quality feature, n1 is the number of consultation service quality features, F a and F b are the values ​​of a and b quality characteristics respectively; the matrix and Represent the latent vectors of the a-th quality feature and the b-th quality feature respectively, and use k factors describing the features to describe the a-th dimension and the b-th dimension features, where the length k1 of the latent vector is a hyperparameter, and It represents the inner product of two vectors of length k1.

[0020] As a further preference, the step S2 comprises the following specific steps:

[0021] S2-1. Obtain the doctor-patient matching score Sim(i, d) according to the following formula:

[0022]

[0023] Where que represents the patient's question, ans represents the historical answer of each doctor, i represents the i-th patient, and d represents the d-th doctor;

[0024] S2-2. Arrange the doctor-patient matching scores Sim(i, d) in descending order to obtain the doctor ranking matrix R of all patients i. Use the location-based model to obtain the doctor's system exposure level E= γ R , where γ is the patience factor of the patient's browsing length in the given recommended doctor list. Assuming that the number of patients is I, the doctor recommendation set e given by the platform is obtained as follows:

[0025]

[0026] S2-3. Assuming that the correlation between the recommended doctor and the patient is a binary label, the target exposure level E* of the doctor is obtained as follows:

[0027]

[0028] In the formula, p i is the number of doctors related to the patient, D is the total number of doctors to be recommended, and γ is the patience factor of the patient's browsing length in the given recommended doctor list;

[0029] We further obtain the ideal recommendation set e* that is proportional to the actual relevance of the doctor, as follows:

[0030]

[0031] S2-4. Based on the doctor recommendation set e given by the platform and the ideal recommendation set e* proportional to the actual relevance of the doctor, the fairness constraint is obtained as

[0032] As a further preference, step S3 includes the following specific steps:

[0033] S3-1. According to the doctor consultation service quality score Sco(d) and the doctor-patient matching score Sim(i, d), the comprehensive service quality function Quality = Sco(d) + Sim(i, d) is obtained, and then combined with the fairness constraint, the model expected reward r(i, d) is obtained;

[0034] S3-2. Use the multi-armed bandit method based on improved Thompson sampling to decide whether to recommend a doctor to a patient according to the following formula:

[0035]

[0036] st∑qd≥1 (2)

[0037] ∑qd ≤D,q d =0or1 (3)

[0038] In formulas (1), (2) and (3), q d is a decision variable, if q d =1, then recommend doctor d to patient i, otherwise do not recommend;

[0039] S3-3, several q in step S3-2 d = 1 are merged into one set to obtain the final recommendation set, which is recommended to the patient.

[0040] As a further preferred embodiment, the formula of the model expected reward r(i, d) in step S3-1 is as follows:

[0041]

[0042] Where φ is the adjustment factor.

[0043] As a further preference, the value range of φ is (0, 1).

[0044] As a further preferred embodiment, the improved Thompson sampling strategy is obtained by performing a multi-objective optimization solution on the original Thompson sampling strategy with regard to consultation service quality and fairness constraints.

[0045] As a further preference, the present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium is used to store a computer program, wherein the computer program executes a doctor recommendation method based on consultation service quality and fairness as described in the present invention.

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] 1. The present invention adopts attribute features, consultation features and comment features when calculating the quality of consultation services, fully considering the professional skills and growth of doctors and the preferences of patients in choosing doctors. It is more accurate and personalized than the recommendations based on single, static consultation features in the prior art, and improves the accuracy of doctor recommendations.

[0048] 2. The present invention takes into account both service quality and fairness when making doctor recommendations, and uses the difference between system exposure and target exposure to seek fairness and rationality in doctor rankings, thereby significantly improving the fairness of the recommendation system and enhancing the fairness of doctor recommendations. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 The figure is an overall flow chart of the doctor recommendation method in the present invention. DETAILED DESCRIPTION

[0050] Advantages and features of the present invention will be illustrated and explained by the following non-limiting description of preferred embodiments thereof, which are given by way of example only with reference to the accompanying drawings.

[0051] like Figure 1 As shown, the present invention provides a doctor recommendation method based on consultation service quality and fairness, comprising the following steps:

[0052] Step 1: Obtain attribute feature F based on the disease problem uploaded by each patient, the personal basic attribute information data of each doctor, and the historical answers and comments of each doctor A , interaction feature F I And the comment feature F R , and then use the factor decomposition machine to get the doctor consultation service quality score Sco(d).

[0053] Step 1-1: Based on the basic attribute information data of each doctor, use unique hot encoding to extract attribute features F A And vectorized representation;

[0054] Step 1-2: Based on the disease questions uploaded by each patient and the historical answers of each doctor, text mining and sentiment analysis methods are used to construct interactive features F for online doctor-patient behavior based on information support, emotional support and response speed based on communication adaptation theory, signal theory and social presence theory. I ;

[0055] Step 1-3: Based on the comment data, a graph convolutional neural network is used to obtain the comment feature F R ;

[0056] In order to capture the description of the same aspect word in multiple reviews, a graph convolutional neural network is used to generate review features F R , define the collection of comment data as Where n = 1, 2, ..., N, N represents the total number of comments, Represents the nth comment in the comment data set. Construct the keyword set X d , if the distance between two words in the comment does not exceed the set threshold, then the two words are connected with an undirected edge in the aspect word graph. The aspect word graph of doctor d is represented as G d ={X d , E d}, E d It means (x h , x n ) word pair. x h The aspect word graph G representing doctor d d A word node in x n is the neighbor node of the word node in the aspect word graph, xn The set of h ={x n |(x h , x n )∈E d or(x n , x h )EE d}.by Represents the implicit vector representation of the word node at the kth layer. Further, by performing a series of convolutional pooling operations on the k+1 layer, the neighbor node representation can be obtained as It can be expressed as:

[0057]

[0058] Among them, AVG k Indicates that the average pooling method is used from the kth layer x h Neighbor nodes of a node Aggregate to get x h This part is spliced and x h The implicit vector in Then the concatenated vector is passed through the fully connected layer to complete the convolution operation to obtain the word node x h The representation at the kth layer On this basis, the entire aspect word graph representation of doctor d is obtained, which can be expressed as:

[0059] r d =∑U h / |X d |,x h ∈X d (2)

[0060] Where X d It means that after filtering the high-frequency words in the comments, the remaining key words are a set of word nodes, then the comment feature F R The word graph representation of the entire aspect of doctor d d The value in .

[0061] Step 1-4: According to the attribute feature F A , interaction feature F I And the comment feature F R , introduce the time decay mechanism and time window mechanism, and get the doctor consultation service quality score Sco(d). The attribute feature F A , interaction feature F I And the comment feature F R For quality characteristics.

[0062] Considering the attribute feature FA , interaction feature F I And the comment feature F R The cross-relationship between them is constructed by introducing a time decay mechanism to build a dynamic service quality evaluation model based on a factor decomposition machine. When calculating feature interactions, a time window mechanism is introduced to dynamically adjust the weights. Different weight coefficients are assigned to features at positions within the corresponding time window. The closer the service behavior is to the current time, the higher its weight is, thereby enhancing the importance of recent behavior feature vectors. The specific expression is as follows:

[0063]

[0064] Where d is the dth doctor, F a , F b ∈{F A , F I , F R}, ω0 is a model parameter, indicating the global bias, ω a represents the weight of the ath quality feature, n1 is the number of consultation service quality features, F a and F b are the values ​​of a and b quality characteristics respectively; the matrix and Represent the latent vectors of the a-th quality feature and the b-th quality feature respectively, and use k factors describing the features to describe the a-th dimension and the b-th dimension features, where the length k1 of the latent vector is a hyperparameter, and It represents the inner product of two vectors of length k1.

[0065] Step 2: Construct a fair value function and obtain the doctor's system exposure level E and target exposure level E according to the location model. * , thus obtaining the fairness constraint Fairness on the doctor side.

[0066] Using the difference method to define the fair value function of doctor recommendation, the fairness constraint on the doctor side can be defined as the difference between the system and target exposure vectors, which can be expressed as:

[0067]

[0068] In the formula, vector e represents the set of doctor recommendations given by the platform, and vector e* represents the ideal recommendation set that is proportional to the actual relevance of the doctor.

[0069] Step 2-1: Obtain the doctor-patient matching score Sim(i, d) according to the following formula:

[0070]

[0071] Where que represents the patient's question, ans represents the historical answer of each doctor, i represents the i-th patient, and d represents the d-th doctor.

[0072] Doctor-patient matching describes the professional ability of doctors in solving patients' problems, indicating the correlation between doctors' professional skills and patients' diseases. It is expressed by calculating the semantic similarity between patients' questions and all doctors' historical answers. Here, the MPNet model is used to obtain the vector representation of the doctor-patient question and answer text. After vectorizing the doctor's historical answers and patients' questions respectively, the cosine similarity formula is used to calculate the similarity between doctors' expertise and patients' questions.

[0073] Step 2-2: Arrange the doctor-patient matching scores Sim(i, d) in descending order to obtain the doctor ranking matrix R of all patients i, and use the location-based model to obtain the doctor's system exposure level E = γ R , where γ is the patience factor of the patient's browsing length in the given recommended doctor list. Assuming that the number of patients is I, the doctor recommendation set e given by the platform is obtained as follows:

[0074]

[0075] Step 2-3: Assuming that the correlation between the recommended doctor and the patient is a binary label, the target exposure level E* of the doctor is obtained as follows:

[0076]

[0077] In the formula, p i is the number of doctors related to the patient, D is the total number of doctors to be recommended, and γ is the patience factor of the patient's browsing length in the given list of recommended doctors.

[0078] We further obtain the ideal recommendation set e* that is proportional to the actual relevance of the doctor, as follows:

[0079]

[0080] Where I is the number of patients.

[0081] Step 2-4: Based on the doctor recommendation set e given by the platform and the ideal recommendation set e* proportional to the actual relevance of the doctor, the fairness constraint is obtained as

[0082] Step 3: Build a doctor recommendation model and use the multi-armed bandit method based on improved Thompson sampling to obtain the final doctor recommendation set and recommend it to patients.

[0083] Step 3-1: According to the doctor consultation service quality score Sco(d) and the doctor-patient matching score Sim(i, d), the comprehensive service quality function Quality = Sco(d) + Sim(i, d) is obtained. Combined with the fairness constraint, the model expected reward r(i, d) is obtained as follows:

[0084]

[0085] Where φ is the adjustment factor, ranging from (0, 1).

[0086] Step 3-2: Use the multi-armed bandit method based on improved Thompson sampling to decide whether to recommend a doctor to the patient according to the following formula:

[0087]

[0088] st∑qd≥1(11)

[0089] ∑q d ≤D,q d =0or1 (12)

[0090] In formulas (10), (11) and (12), q d is a decision variable, if q d =1, doctor d is recommended to patient i, otherwise not recommended.

[0091] The recommendation problem considered in the present invention is to adaptively provide the best set of doctors for each patient, with the goal of maximizing the doctor-patient matching and fairness during the entire recommendation period. Since the Thompson sampling strategy can better balance the EvE (Exploitation versus Exploration) dilemma, the multi-armed bandit algorithm based on improved Thompson sampling is used in the present invention to implement doctor recommendation. The multi-armed bandit algorithm gradually learns the reward distribution of each choice by balancing exploration and utilization, and maximizes the cumulative reward within a limited time. Different strategies (such as Epsilon-Greedy, UCB, Thompson sampling) have different exploration and utilization balance modes and are suitable for different application scenarios. The present invention improves Thompson sampling by jointly optimizing the consultation service quality and fairness, and implements doctor recommendation in combination with the multi-armed bandit algorithm. Thompson sampling usually has better empirical performance than Epsilon-Greedy and UCB. The core idea of ​​improved Thompson sampling is to select actions based on the posterior distribution of each action (i.e., recommended options) instead of relying on point estimates. In order to improve the accuracy and fairness of the recommendation results, the present invention takes into account both the consultation service quality and fairness constraints.

[0092] Step 3-3: Replace several q d= 1 are merged into one set to obtain the final recommendation set, which is recommended to the patient.

[0093] A person of ordinary skill in the art can understand that all or part of the steps of implementing the above method embodiment can be completed by program instructions and related hardware. The aforementioned program instructions can be stored in a computer-readable storage medium. When the program instructions are executed, the steps of the above method embodiment are executed; and the aforementioned storage medium includes: a mobile storage device, a read-only memory (ROM), a magnetic disk or an optical disk, and other media that can store program codes.

[0094] In addition to the above embodiments, the present invention may also have other implementation modes. Any technical solutions formed by equivalent replacement or equivalent transformation shall fall within the protection scope required by the present invention.

Claims

1. A doctor recommendation method based on consultation service quality and fairness, characterized in that: The following steps are included: S1. Obtain attribute feature F based on the disease problem uploaded by each patient, each doctor’s personal basic attribute information data, and each doctor’s historical answers and comments data A , interaction feature F I And the comment feature F R , and then use the factorization machine to obtain the doctor consultation service quality score Sco(d); S2. Construct a fair value function and obtain the doctor’s system exposure level E and target exposure level E according to the location model. * , and then get the fairness constraint Fairness on the doctor side; S3. Build a doctor recommendation model and use the multi-armed bandit method based on improved Thompson sampling to obtain the final doctor recommendation set and recommend it to patients.

2. A doctor recommendation method based on consultation service quality and fairness according to claim 1, characterized in that: The step S1 comprises the following specific steps: S1-1. Based on each doctor’s personal basic attribute information data, use unique hot encoding to extract attribute features F A And vectorize it; S1-2. Based on the disease questions uploaded by each patient and the historical answers of each doctor, text mining and sentiment analysis methods are used to construct interactive features of online doctor-patient behavior from information support, emotional support and response speed based on communication adaptation theory, signal theory and social presence theory. I ; S1-3. Based on the patient comment data, the graph convolutional neural network is used to obtain the comment feature F R ; S1-4. According to the attribute feature F A , interaction feature F I And the comment feature F R , introduce the time decay mechanism and time window mechanism, use the factor decomposition machine to get the doctor consultation service quality score Sco(d), the attribute feature F A , interaction feature F I And the comment feature F R For quality characteristics.

3. A doctor recommendation method based on consultation service quality and fairness according to claim 2, characterized in that: The doctor consultation service quality score Sco(d) formula in step S1-4 is as follows: Where d is the dth doctor, F a , F b ∈{F A , F I , F R }, ω0 is a model parameter, indicating the global bias, ω a represents the weight of the ath quality feature, n1 is the number of consultation service quality features, F a and F b are the values ​​of a and b quality characteristics respectively; matrix and Represent the latent vectors of the a-th quality feature and the b-th quality feature respectively, and use k factors describing the features to describe the a-th dimension and the b-th dimension features, where the length k1 of the latent vector is a hyperparameter, and It represents the inner product of two vectors of length k1.

4. A doctor recommendation method based on consultation service quality and fairness according to claim 3, characterized in that: The step S2 comprises the following specific steps: S2-1. Obtain the doctor-patient matching score Sim(i, d) according to the following formula: Where que represents the patient's question, ans represents the historical answer of each doctor, i represents the i-th patient, and d represents the d-th doctor; S2-2. Arrange the doctor-patient matching scores Sim(i, d) in descending order to obtain the doctor ranking matrix R of all patients i. Use the location-based model to obtain the doctor's system exposure level E=γ R , where γ is the patience factor of the patient's browsing length in the given recommended doctor list. Assuming that the number of patients is I, the doctor recommendation set e given by the platform is obtained as follows: S2-3. Assuming that the correlation between the recommended doctor and the patient is a binary label, the target exposure level E* of the doctor is obtained as follows: In the formula, p i is the number of doctors related to the patient, D is the total number of doctors to be recommended, and γ is the patience factor of the patient's browsing length in the given recommended doctor list; Further calculations are performed to obtain the ideal recommendation set e* that is proportional to the actual relevance of the doctor, as follows: S2-4. Based on the doctor recommendation set e given by the platform and the ideal recommendation set e* proportional to the actual relevance of the doctor, the fairness constraint is obtained as 5. A doctor recommendation method based on consultation service quality and fairness according to claim 4, characterized in that: The step S3 comprises the following specific steps: S3-1. According to the doctor consultation service quality score Sco(d) and the doctor-patient matching score Sim(i, d), the comprehensive service quality function Quality = Sco(d) + Sim(i, d) is obtained, and then combined with the fairness constraint, the model expected reward r(i, d) is obtained; S3-2. Use the multi-armed bandit method based on improved Thompson sampling to decide whether to recommend a doctor to a patient according to the following formula: s.t.∑q d ≥1 (2) ∑q d ≤D,q d =0 or 1 (3) In formulas (1), (2) and (3), q d is a decision variable, if q d =1, then recommend doctor d to patient i, otherwise do not recommend; S3-3, several q in step S3-2 d = 1 are merged into one set to obtain the final recommendation set, which is recommended to the patient.

6. A doctor recommendation method based on consultation service quality and fairness according to claim 5, characterized in that: The formula of the model expected reward r(i, d) in step S3-1 is as follows: Where φ is the adjustment factor.

7. A doctor recommendation method based on consultation service quality and fairness according to claim 6, characterized in that: The value range of φ is (0, 1).

8. A doctor recommendation method based on consultation service quality and fairness according to claim 7, characterized in that: The improved Thompson sampling strategy is obtained by solving the multi-objective optimization of the original Thompson sampling strategy with regard to the quality of consultation service and fairness constraints.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium is used to store a computer program, and the computer program executes the doctor recommendation method based on consultation service quality and fairness as described in any one of claims 1-8.

Images