A Doctor Recommendation Method Considering Patients' Browsing Sequences in the Field of 5G

Abstract

The present invention discloses a doctor recommendation method considering the patient's browsing sequence in the 5G field, which predicts the next doctor that the patient is interested in based on the browsing and rating data of the patient in the intelligent medical guidance platform. In order to overcome the sparsity of the patient's rating data, the present invention takes the patient's browsing data into account in the recommendation method and decouples the positive and negative emotions contained in the browsing data using the rating data. The present invention is divided into four parts: the first part is to remove the positive emotion information in the patient's browsing data based on the patient's positive rating data to obtain the negative emotion information in the patient's browsing data; the second part is to remove the negative emotion information in the patient's browsing data based on the patient's negative rating data to obtain the positive emotion information in the patient's browsing data; the third part is to combine the two emotion information of the patient's two feedback data to obtain the vector representation of the patient; the fourth part is to recommend doctors to the patient according to the patient vector representation.

Description

Technical Field

[0001] The present invention belongs to the technical field of Internet services, and in particular, relates to a doctor recommendation method considering the browsing sequence of patients in the 5G field. Background Art

[0002] The 5G network refers to the fifth-generation mobile communication network. Compared with the 4G network, the 5G network has characteristics such as ultra-high data transmission rate and low latency. The rise of the 5G network has made the types and functions of applications on mobile phones become more and more abundant. Now patients can easily find high-quality doctors in the online medical guidance software on the mobile side. However, the number of doctors on the platform is too large, and it is very difficult for patients to find a suitable doctor. For example, there are already 610,000 doctors on HaoDaiFu and 500,000 doctors on ChunYu Doctor. Therefore, a doctor recommendation system is very necessary. The recommendation system can help patients find suitable doctors from a large amount of doctor data.

[0003] The existing doctor recommendation system recommends doctors to patients based on the rating data of patients on doctors in the platform. Although the rating data is explicit feedback data and reflects the real satisfaction of patients, the rating data of patients is very sparse. A large amount of browsing data of patients on doctors in the platform is ignored. The browsing data belongs to implicit feedback data. The recommendation system in the e-commerce platform usually makes recommendations based on the browsing data of users on goods. To a certain extent, the browsing data of users in the e-commerce platform represents the interests of users, that is, the browsing data belongs to the positive feedback data of users. However, the browsing data of patients in the intelligent medical guidance platform is different from the browsing data in the e-commerce platform. The patient browses the doctor's homepage but does not choose this doctor, which to a certain extent indicates that the patient is not satisfied with this doctor. However, the decision-making of patients is complex. Since the patient chooses to browse the doctor's homepage and understand the doctor's information, it shows that the patient is interested in this doctor. The reason why the patient does not choose this doctor may be that the consultation fee of this doctor is too high or the distance is too far. In short, in the intelligent medical guidance platform, the browsing data of patients reflects both the positive emotions of patients and the negative emotions of patients. Therefore, it is necessary to decouple the information in the user browsing data.

[0004] Therefore, this paper proposes a doctor recommendation method in the 5G field that considers the patient's browsing sequence, while considering the patient's rating data and browsing data. The patient's rating data contains the patient's positive and negative emotional information. The patient's rating data in the platform ranges from 1 to 5, 1 represents very dissatisfied, and 5 represents very satisfied. This method considers that scores greater than or equal to 4 belong to positive display feedback, otherwise they are negative display feedback. This method uses the patient's positive display feedback data to remove positive emotional data from the browsing data, and retains the negative emotional information in the browsing data. At the same time, the patient's negative display feedback data is used to remove negative emotional data from the browsing data, and retain the positive emotional information in the browsing data. Finally, the two types of emotional information in the two feedback data are combined to obtain a comprehensive vector representation of the patient, and a suitable doctor is recommended to the patient. Summary of the invention

[0005] The problem of this method is defined as predicting the next doctor that the patient is interested in based on the patient's browsing and rating data. The patient's browsing data belongs to the patient's implicit feedback data, and the patient's rating data belongs to the explicit feedback data. The mathematical symbols involved are: the patient set is U, and the doctor set is V. Patient u i The browsing behavior sequence is The subscript l represents the patient u i Browsing behavior sequence The length of the sequence is , the superscript b indicates the browse data, and the jth doctor in the browsing behavior sequence is represented by Patient i The score data of is divided into two sequences: positive score sequence and negative score sequence. When the patient score is greater than or equal to 4, the score is positive, otherwise it is negative. The positive score sequence is expressed as The subscript m1 represents the patient u i Positive Rating Sequence The length of the sequence is , the superscript p represents the positive score data, and the jth doctor in the positive score sequence is represented by The negative rating sequence is represented as The subscript m2 represents the patient u i Negative Rating Sequence The length of , the superscript n represents negative score data, and the jth doctor in the negative score sequence is represented by Existing recommendation methods recommend doctors to patients based on the rating data of doctors by patients on the platform. Although the rating data is display feedback data reflecting the true satisfaction of patients, the rating data of patients is very sparse. These methods ignore the browsing data of patients on doctors accumulated on the platform, and the browsing data reflects both the positive and negative emotions of patients. The patient browses the homepage of the doctor but does not choose the doctor, which to a certain extent indicates that the patient is not satisfied with the doctor. However, the decision-making of patients is complex. Since the patient chooses to browse the homepage of the doctor and understand the information of the doctor, it shows that the patient is interested in the doctor. The reason for not choosing the doctor may be that the consultation fee of the doctor is too high or the distance is too far. Therefore, it is necessary to decouple the information in the user browsing data. For this purpose, the present invention adopts the following technical solutions:

[0006] A doctor recommendation method considering the patient browsing sequence in the 5G field, comprising the following steps:

[0007] Based on the positive rating data of patients, the positive emotion information in the patient browsing data is removed to obtain the negative emotion information in the patient browsing data. This method is to solve the sparsity of the patient rating data by introducing the patient browsing data into the recommendation system. The patient browsing data reflects both the positive and negative emotions of patients. In this method, the patient set is U and the doctor set is V. The browsing behavior sequence of patient u i is where the subscript l represents the length of the browsing behavior sequence of patient u i browsing behavior sequence and the superscript b represents the browsing data. The j-th doctor in the browsing behavior sequence is denoted as The rating data of patient u i is divided into two sequences: a positive rating sequence and a negative rating sequence. When the patient rating is greater than or equal to 4, the rating is a positive rating, otherwise it is a negative rating. The positive rating sequence is denoted as where the subscript m1 represents the length of the positive rating sequence of patient u i positive rating sequence and the superscript p represents the positive rating data. The j-th doctor in the positive rating sequence is denoted as The negative rating sequence is denoted as where the subscript m2 represents the length of the negative rating sequence of patient u i negative rating sequence and the superscript n represents the negative rating data. The j-th doctor in the negative rating sequence is denoted as The scoring sequence and browsing sequence of the patient both contain the patient's positive and negative sentiment information. To distinguish, the sentiment information extracted from the patient's scoring sequence is called strong sentiment information in this method, and the sentiment information extracted from the patient's browsing sequence is called weak sentiment information. According to the patient's positive scoring data, this method eliminates the positive sentiment information in the patient's browsing data to obtain the negative sentiment information in the patient's browsing data, and this process is called the anti-attention mechanism. The specific steps include two steps: The first step is to obtain the patient's strong positive sentiment information according to the patient's positive scoring sequence The formula is as follows: The formula is as follows:

[0008]

[0009] Wherein, is the vector representation of the j-th doctor in the positive scoring sequence The second step is to extract the weak negative sentiment information from the patient's browsing behavior sequence based on the patient's strong positive sentiment information The formula is as follows:

[0010]

[0011]

[0012] Wherein, is the vector representation of the j-th doctor in the browsing behavior sequence α j is the weight of the j-th doctor The function φ(·) represents the inner product of two vectors. It can be seen from the calculation formula of α j that the more similar the browsing behavior sequence i of patient u is to the strong positive sentiment information , the lower the weight. Therefore, the extracted represents the weak negative sentiment information of the patient.

[0013] Based on the patient's negative scoring data, the negative sentiment information in the patient's browsing data is eliminated to obtain the positive sentiment information in the patient's browsing data. The specific steps include two steps: The first step is to obtain the patient's strong negative sentiment information according to the patient's negative scoring sequence The formula is as follows: The formula is as follows:

[0014]

[0015] Wherein, is the negative scoring sequence the j-th doctor in vector representation. In the second step, based on the strong negative emotional information of the patient extract weak positive emotional information from the patient's browsing behavior sequence using inverse attention, as follows: The formula is as follows:

[0016]

[0017]

[0018] where is the vector representation of the j-th doctor in the browsing behavior sequence . β j is the weight of the j-th doctor , and the φ(·) function represents the inner product of two vectors. It can be seen from the calculation formula of β j that the more similar the browsing behavior sequence i of patient u is to the strong negative emotional information , the lower the weight, so the extracted represents the weak positive emotional information of the patient. Combining the two emotional information from the two feedback data of the patient to obtain the vector representation of the patient. This method assigns weights to the two emotional information extracted from the two feedback data and then sums them to obtain the comprehensive vector representation p of the patient, specifically:

[0019]

[0020]

[0021]

[0021] where and represent the weights of strong positive emotional information and strong negative emotional information; and represent the weights of weak positive emotional information and weak negative emotional information. and are hyperparameters. p is the vector representation of the patient.

[0022] Recommend doctors to the patient according to the patient vector representation. Multiply the vector representation x τ of doctor v in the doctor set by the patient vector representation p, and then use the softmax function to calculate the score of doctor v τ : τ The formula is:

[0023]

[0024] where p represents the patient vector representation, p TThe superscript T represents the transpose operation of a vector, and x τ is the vector representation of doctor v τ . represents the possibility that doctor v τ is viewed. For this patient sample data, the loss function is:

[0025]

[0026] where y τ represents the one-hot encoding of doctor v τ . The function is optimized using the gradient descent method.

[0027] The beneficial technical effects of the present invention are as follows:

[0028] (1) This method innovatively proposes to simultaneously consider the patient's rating data and browsing data in the doctor recommendation scenario, and uses the patient's browsing data to make up for the sparsity of the rating data.

[0029] (2) Considering the particularity of the patient's browsing behavior in the doctor recommendation scenario, this method proposes that in the doctor recommendation scenario, the patient's browsing behavior contains both the user's positive emotional information and negative emotional information. In the e-commerce scenario, the recommendation system usually believes that the browsing data only contains the user's positive emotional information.

[0030] (3) This method proposes to decouple the information in the user's browsing behavior data, and according to the positive and negative rating data, adopt an anti-attention mechanism to simultaneously extract the negative and positive emotional information from the browsing data. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 is a schematic flow chart of a doctor recommendation method considering the patient's browsing sequence in the 5G field of the present invention;

[0032] Figure 2 is a schematic model diagram of a doctor recommendation method considering the patient's browsing sequence in the 5G field of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0033] To further understand the present invention, the following specifically describes a doctor recommendation method considering the patient's browsing sequence in the 5G field provided by the present invention in combination with specific embodiments. However, the present invention is not limited thereto. Non-essential improvements and adjustments made by those skilled in the art under the core guiding ideology of the present invention still fall within the protection scope of the present invention.

[0034] The problem of this method is defined as predicting the next doctor that a patient is interested in based on the patient's browsing and rating data. Among them, the patient's browsing data belongs to the patient's implicit feedback data, and the patient's rating data belongs to the explicit feedback data. The mathematical symbols involved are: the set of patients is U, and the set of doctors is V. Patient u i 's browsing behavior sequence is where the subscript l represents the length of the browsing behavior sequence of patient u i The superscript b represents browse data, and the j-th doctor in the browsing behavior sequence is denoted as The rating data of patient u is divided into two sequences: a positive rating sequence and a negative rating sequence. When the patient's rating is greater than or equal to 4, the rating is a positive rating; otherwise, it is a negative rating. The positive rating sequence is denoted as i where the subscript m1 represents the length of the positive rating sequence of patient u i The superscript p represents positive rating data, and the j-th doctor in the positive rating sequence is denoted as The negative rating sequence is denoted as where the subscript m2 represents the length of the negative rating sequence of patient u i The superscript n represents negative rating data, and the j-th doctor in the negative rating sequence is denoted as

[0035] Figure 1 A doctor recommendation method considering the patient's browsing sequence in the 5G field mainly consists of four parts. The first part is to remove the positive emotional information from the patient's browsing data based on the patient's positive rating data to obtain the negative emotional information in the patient's browsing data; the second part is to remove the negative emotional information from the patient's browsing data based on the patient's negative rating data to obtain the positive emotional information in the patient's browsing data; the third part is to combine the two emotional information of the patient's two feedback data to obtain the vector representation of the patient; the fourth part is to recommend doctors to the patient according to the patient's vector representation.

[0035] As Figure 1 shown, according to an embodiment of the present invention, this method includes the following steps:

[0036] S100. Based on the positive rating data of patients, the positive sentiment information in the patient browsing data is removed to obtain the negative sentiment information in the patient browsing data. The existing doctor recommendation system recommends doctors to patients based on the rating data of patients on doctors in the platform, ignoring the huge amount of patient browsing data of doctors accumulated in the intelligent diagnosis guidance platform. To solve the sparsity of patient rating data, this method introduces the implicit feedback data of patients, that is, the patient browsing data, into the recommendation system. The patient browsing data reflects both the positive sentiment and the negative sentiment of patients. In this method, the patient set is U and the doctor set is V. The browsing behavior sequence of patient u i is where the subscript l represents the length of the browsing behavior sequence of patient u i of the browsing behavior sequence and the superscript b represents the browsing data. The j-th doctor in the browsing behavior sequence is denoted as Patient u i 's rating data is divided into two sequences: the positive rating sequence and the negative rating sequence. When the patient's rating is greater than or equal to 4, the rating is a positive rating; otherwise, it is a negative rating. The positive rating sequence is denoted as where the subscript m1 represents the length of the positive rating sequence of patient u i of the positive rating sequence and the superscript p represents the positive rating data. The j-th doctor in the positive rating sequence is denoted as The negative rating sequence is denoted as where the subscript m2 represents the length of the negative rating sequence of patient u i of the negative rating sequence and the superscript n represents the negative rating data. The j-th doctor in the negative rating sequence is denoted as Both the patient's rating sequence and browsing sequence contain both the positive sentiment information and the negative sentiment information of the patient. For distinction, this method calls the sentiment information extracted from the patient rating sequence strong sentiment information, and the sentiment information extracted from the patient browsing sequence weak sentiment information. Since the patient's rating data belongs to explicit feedback data, representing the stronger sentiment of the patient; while the patient's browsing data belongs to implicit feedback data. The positive rating data of patients is too sparse. If the attention mechanism is used to extract the content related to the patient's positive rating data from the patient browsing data, very one-sided information will be obtained. Therefore, this method removes the positive sentiment information in the patient browsing data according to the patient's positive rating data to obtain the negative sentiment information in the patient browsing data. This is called the anti-attention mechanism in this paper. The specific steps include two steps: The first step is to obtain the strong positive sentiment information of the patient from the positive rating sequence of the patient The formula is as follows:

[0037]

[0038] Among them, is the positive scoring sequence of the j-th doctor in. The second step is to use the strong positive emotional information of the patient as a basis to extract weak negative emotional information from the patient's browsing behavior sequence by using inverse attention. The formula is as follows:

[0039]

[0040]

[0041] Among them, is the browsing behavior sequence of the j-th doctor in. α j is the weight of the j-th doctor in. The ≤(·) function represents the inner product of two vectors. It can be seen from the calculation formula of α j that the more similar the browsing behavior sequence i of patient u is to the strong positive emotional information , the lower the weight. Therefore, the extracted represents the weak negative emotional information of the patient.

[0042] S200. Based on the negative scoring data of the patient, the negative emotional information in the patient's browsing data is removed to obtain the positive emotional information in the patient's browsing data. The specific steps include two steps: The first step is to obtain the strong negative emotional information of the patient according to the negative scoring sequence of the patient. The formula is as follows:

[0043]

[0044] Among them, is the negative scoring sequence of the j-th doctor in. The second step is to use the strong negative emotional information of the patient as a basis to extract weak positive emotional information from the patient's browsing behavior sequence by using inverse attention.

[0045]

[0046]

[0047] Among them, is the browsing behavior sequence of the j-th doctor in the sequence. β j is the weight of the j-th doctor and the φ(·) function represents the inner product of two vectors. From the calculation formula of β j , it can be seen that for patient u i in the browsing behavior sequence , the more similar it is to strong negative sentiment information , the lower the weight. Therefore, the extracted represents the weak positive sentiment information of the patient. Steps S100 and S200 are the innovative points of this method. The schematic diagram of the model is shown in Figure 2 .

[0048] S300, combining the two sentiment information of the two feedback data of the patient to obtain the vector representation of the patient. This method assigns weights to the two sentiment information extracted from the two feedback data and then sums them to obtain the comprehensive vector representation p of the patient, specifically:

[0049]

[0050] where and represent the weights of strong positive sentiment information and strong negative sentiment information; and represent the weights of weak positive sentiment information and weak negative sentiment information. and are hyperparameters. This method sets and and p is the vector representation of the patient.

[0051] S400, recommending doctors to the patient according to the patient vector representation. Multiply the vector representation x τ of doctor v τ in the doctor set by the patient vector representation p, and then use the softmax function to calculate the score of doctor v τ :

[0052]

[0053] where p represents the patient vector representation, x τ is the vector representation of doctor v τ , represents the possibility that doctor v τ is browsed. For this patient sample data, the loss function is:

[0054]

[0055] where yτ One-hot encoding of the representative doctor v τ . The function is optimized using the gradient descent method.

[0056] The above description of the embodiments is provided to enable those of ordinary skill in the art to understand and apply the present invention. It is obvious that those skilled in the art can easily make various modifications to the above embodiments and apply the general principles described herein to other embodiments without creative efforts. Therefore, the present invention is not limited to the above embodiments, and all improvements and modifications made by those skilled in the art based on the disclosure of the present invention should fall within the protection scope of the present invention.

Claims

1. A doctor recommendation method considering the patient browsing sequence in the 1.5G field, characterized in that: Based on the positive rating data of patients, the positive sentiment information in the patient browsing data is removed to obtain the negative sentiment information in the patient browsing data; this method introduces the patient browsing data into the recommendation system to solve the sparsity of patient rating data; the patient browsing data reflects both the positive sentiment and the negative sentiment of patients; in this method, the patient set is U and the doctor set is V; for patient u i 's browsing behavior sequence is where the subscript l represents the length of the browsing behavior sequence of patient u i of the browsing behavior sequence and the superscript b represents the browsing data. The j-th doctor in the browsing behavior sequence is denoted as For patient u i 's rating data is divided into two sequences: a positive rating sequence and a negative rating sequence; when the patient's rating is greater than or equal to 4, the rating is a positive rating, otherwise it is a negative rating; the positive rating sequence is denoted as where the subscript m1 represents the length of the positive rating sequence of patient u i of the positive rating sequence and the superscript p represents the positive rating data. The j-th doctor in the positive rating sequence is denoted as The negative rating sequence is denoted as where the subscript m2 represents the length of the negative rating sequence of patient u i of the negative rating sequence and the superscript n represents the negative rating data. The j-th doctor in the negative rating sequence is denoted as Both the patient's rating sequence and browsing sequence contain both the positive sentiment information and the negative sentiment information of the patient. For the sake of distinction, this method calls the sentiment information extracted from the patient rating sequence strong sentiment information and the sentiment information extracted from the patient browsing sequence weak sentiment information; this method removes the positive sentiment information in the patient browsing data based on the patient's positive rating data to obtain the negative sentiment information in the patient browsing data, and this process is called the anti-attention mechanism; the specific steps include two steps: the first step is to obtain the strong positive sentiment information of the patient according to the patient's positive rating sequence The formula is as follows: The formula is as follows: Among them, is the vector representation of the j-th doctor in the positive scoring sequence In the second step, based on the strong positive emotion information of the patient, anti-attention is used to extract weak negative emotion information from the patient's browsing behavior sequence The formula is as follows: Among them, is the vector representation of the j-th doctor in the browsing behavior sequence ; α j is the weight of the j-th doctor ; the φ(·) function represents the inner product of two vectors; from the calculation formula of α j , it can be seen that the more similar the browsing behavior sequence i of patient u is to the strong positive sentiment information , the lower the weight, so the extracted represents the weak negative sentiment information of the patient; Based on the patient's negative rating data, the negative sentiment information in the patient's browsing data is removed to obtain the positive sentiment information in the patient's browsing data. The specific steps include two steps: The first step is to obtain the patient's strong negative sentiment information according to the patient's negative rating sequence The formula is as follows: Among them, is the vector representation of the j-th doctor in the negative scoring sequence; in the second step, based on the strong negative emotional information of the patient, anti-attention is used to extract weak positive emotional information from the patient's browsing behavior sequence The formula is as follows: The formula is as follows: Among them, is the vector representation of the j-th doctor in the browsing behavior sequence ; β j is the weight of the j-th doctor ; the φ(·) function represents the inner product of two vectors; from the calculation formula of β j , it can be seen that the more similar the browsing behavior sequence i of patient u is to strong negative emotion information , the lower the weight, so the extracted represents the weak positive emotion information of the patient; Combining two kinds of emotional information of two kinds of feedback data of the patient to obtain a vector representation of the patient; in this method, weights are assigned to the two kinds of emotional information extracted from the two kinds of feedback data, and then summed to obtain a comprehensive vector representation p of the patient, specifically: Among them, and represent the weights of strong positive sentiment information and strong negative sentiment information; and represent the weights of weak positive sentiment information and weak negative sentiment information; and are hyperparameters; p is the patient vector representation; Recommend doctors for the patient according to the patient vector representation; multiply the vector representation x τ of doctor v in the doctor set τ by the patient vector representation p, and then use the softmax function to calculate the score of doctor v τ : Among them, p represents the patient vector representation, The superscript of represents the transpose operation of the vector, x τ is the vector representation of doctor v τ ; represents the possibility that doctor v τ is viewed; for this patient sample data, the loss function is: Among them, y τ represents the one-hot encoding of doctor v τ ; The function is optimized using the gradient descent method.

Images