Doctor recommendation algorithm for relieving data sparsity based on large-scale language representation

Abstract

The invention discloses a doctor recommendation algorithm for relieving data sparsity based on large-scale language representation, which is used for recommending matched doctors to patients on the basis of interactive behaviors of the patients in an online medical platform, a large-scale language model and introduction of diseases at which the doctors are good. The forward transmission part is mainly composed of five parts. In the first part, vector representation of a disease is obtained by using a large-scale language model, and a doctor vector is initialized based on a disease vector after dimension reduction; in the second part, the patient behavior data is divided into session lists, and session vectors are calculated according to doctor vectors in sessions; the third part is that the historical session list of the current patient is used as input, and a noise reduction attention mechanism method is adopted to carry out information supplement on the current session; fourthly, similar patients of the current patient are mined, similar patient vectors are obtained through a noise reduction attention mechanism method, and the similar patient vectors are used for supplementing information to the current patient; and the fifth part is to recommend doctors to the patients according to the patient vectors.

Description

Technical Field

[0001] The present invention belongs to the field of Internet service technology, and in particular relates to a doctor recommendation algorithm based on large-scale language representation to alleviate data sparsity. Background Art

[0002] The application of smart medical services in existing online medical platforms is not yet widespread. One of the most important functions of smart medical services is intelligent medical guidance. Intelligent medical guidance is to recommend suitable doctors to users using the data provided in the platform information management system. The data in the management system mainly includes the information of platform doctors, the information of patients, and the interactive information between patients and doctors: that is, the browsing, collection and rating data of patients on doctors. In e-commerce platform recommendations, recommendation algorithms usually use interactive data such as user purchases, clicks and collections; in short video recommendations, recommendation algorithms usually use comment data as positive feedback data. Existing doctor recommendation algorithms in the medical field usually use patients' rating data on doctors for recommendation, but the rating data is extremely sparse. In order to alleviate the problem of data sparsity, this method uses all the interactive data between patients and doctors, including browsing, collection and rating data, as input for recommendation. Among them, the user browses the doctor's homepage information, which means that the doctor is good at treating the diseases that the patient may have, that is, the doctor's information matches the patient's needs. The recommendation system recommends based on the matching degree between doctor information and patient needs, so it is reasonable to use patients' browsing data on doctors as input.

[0003] Existing doctor recommendation systems usually recommend matching doctors to patients based on their current conversational behaviors. However, patients’ interactive behaviors are extremely sparse, and recommendation algorithms cannot learn high-quality vector representations of doctors and patients through sparse collaborative information. Large-scale language models have learned a large amount of open-world domain knowledge through massive training. Recommendations based on large-scale language model representations have been proven to improve the performance of traditional ID-based recommendation algorithms. However, the algorithm complexity of large-scale language models is relatively high. How to use the semantic knowledge learned from large-scale language models without significantly increasing the model complexity is a problem that recommendation algorithms need to solve.

[0004] In order to alleviate data sparsity, this method further introduces the patient's historical interaction behavior and other patients' interaction behavior to supplement the patient's current session information. The difference between the doctor recommendation scenario and other recommendation scenarios is that the user's historical preferences are limited, the preferences in a single session are relatively single, and there are repeated interactions. In addition, there are a large number of patients whose symptoms are similar or even the same. This is also the basis of China's Mijian Medical Platform. The purpose of the Mijian Medical Platform is to build a patient communication platform to help patients find fellow patients, help each other, and share information. Therefore, it is reasonable to use the patient's past interaction information and the information of other potentially similar patients to expand the patient's current session information. However, the introduction of other sources of information will introduce more noise, and while alleviating the problem of data sparsity, the noise needs to be processed. Summary of the invention

[0005] The technical difficulty that this method aims to solve is to alleviate the data sparsity problem in the doctor recommendation scenario. The idea to alleviate this problem is to use all the interaction data between patients and doctors in the platform as the input of the recommendation algorithm, and use a large-scale language model to obtain the doctor's vector, and then expand the patient's current session information based on the doctor's semantic vector. The expanded information source is the user's historical conversation and the interaction data of other similar patients. Using other information to expand the patient's current session information will inevitably bring more noise. While alleviating data sparsity, how to reduce the impact of noise is also a problem that this method needs to solve.

[0006] Specifically, the present invention proposes a doctor recommendation algorithm based on large-scale language representation to alleviate data sparsity, and the process is as follows:

[0007] The large-scale language model is used to obtain the vector representation of the disease, and the vector dimension reduction algorithm is used to reduce the dimension of the disease vector; the vector of the doctors in the platform is initialized based on the reduced dimension disease vector. The homepage of the doctor in the online platform has an introduction to the diseases he is good at. This method counts all the disease sets in the platform and inputs the disease text into the large-scale language model to obtain the high-dimensional vector of the disease. The process is expressed in mathematical symbols as c i is the disease text, and LLM is a large-scale language model. After obtaining the high-dimensional disease vector, the principal component analysis method is used to reduce the dimension of the disease vector. The disease vector after dimension reduction is still used To express, that is PCA stands for principal component analysis. All disease vectors in the online medical platform are stored offline in the system. In the recommendation algorithm, the doctor vector is initialized as the average pool result of the disease vectors that the doctor is good at. The mathematical formula is expressed as dis v It is a collection of diseases that doctor V is good at.

[0008] The patient behavior data is divided into a conversation list, and the conversation vector is calculated based on the vector of the doctor in the conversation. The patient behavior data includes the patient's browsing, collection and rating behaviors in the platform. This method divides the patient behavior sequence into several conversations based on the interval time of the behavior. The conversation list of patient u is t is the length of the conversation list S(u). The included doctors are represented by n is the session The length of any doctor v is represented by e v , which is initialized to emb v , vector e v Updated during model training. Use sequence encoder Encoder for any session Encode and get the vector of the session, expressed as

[0009] The historical conversation list of the current patient is taken as input and the noise reduction attention mechanism is used to supplement the current conversation information. The conversation sequence of patient u is in is the current session, This method designs a denoising attention mechanism to calculate the influence vector of the patient's historical conversation on the current conversation, and uses Gumbel softmax to denoise the historical conversation weights calculated by the attention mechanism. The process is expressed as The specific process of the denoising attention mechanism Attn_denois is divided into three steps. The first step is to calculate any historical session of the current patient. and the current session The correlation The score function is implemented by the attention mechanism. The specific formula is:

[0010]

[0011] in, It is a conversation The second step is to use Gumbelsoftmax to sample and get the historical session The new weight π " , π " =Gumbel_sofotmax(p i ) is as follows:

[0012]

[0013] Among them, g iis Gumbel noise, whose sampling value is g i =-log(-log(U i )),U i is a random variable sampled from the uniform distribution U(0,1). τ>0 is the temperature coefficient, which controls the smoothness of the output distribution. The third step is to obtain the influence vector of the patient's historical session on the current session based on the historical session vector and its weight. Fusion of the influence vector and the current session vector gives the new current patient vector:

[0014] Mining similar patients of the current patient, using the denoising attention mechanism method to obtain similar patient vector representations, and then using similar patient vector representations to supplement the current patient’s information. Other patients who have generated ratings and favorites in the current session are considered similar to the current patient. The doctors included are It is a conversation The length of . The number of patients who have rated and collected any doctor v is denoted as N(v), so the current similar patients of patient u are denoted as The process of using similar patients' information to supplement the current patient's information consists of two steps. The first step is to use the current patient as a query and use the denoising attention mechanism to obtain the vector representation of similar patients. The second step is to use the current patient as a query and use the attention mechanism to obtain the influence of similar patients on the current patient. The output of the first step is the vector representation of similar patients. Any similar patient u j , the calculation formula of its vector is S his (u j ) is a similar patient u j The second step is to use the attention mechanism to sort similar patients The information in the data is aggregated to obtain the influence vector h of similar patients on the current patient. social :

[0015]

[0016] Among them, β j Similar patients j Finally, the influence vector of similar patients and the current patient vector are fused to obtain the final current patient vector

[0017] Recommend doctors to patients based on their vector representation.τ The vector representation of e τ Multiply by the patient vector Then use the softmax function to calculate the doctor v τ Score:

[0018]

[0019] in, represents the patient vector, e τ Doctor τ The vector representation of Indicates doctor v τ The possibility of being browsed. For this patient sample data, the loss function is:

[0020]

[0021] Among them, y τ Representative Doctor τ One-hot encoding. The function is optimized using the Adam optimizer.

[0022] The beneficial technical effects of the present invention are as follows:

[0023] (1) This method is based on a large-scale language model to obtain the initial vector of the doctor in the doctor recommendation algorithm. In order not to increase the complexity of the recommendation algorithm, this method obtains the vectors of all diseases in the system through a large-scale language model and stores them offline in the system. Then, the disease vector is reduced in dimension using a vector dimensionality reduction algorithm and the doctor vector is initialized according to the disease that the doctor is good at.

[0024] (2) This method proposes a noise reduction attention mechanism, which uses the patient's historical interaction behavior and the interaction data of other similar patients to expand the patient's current conversation information, and reduces noise while expanding the information. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 A flowchart of a doctor recommendation algorithm based on large-scale language representation to alleviate data sparsity in the present invention;

[0026] Figure 2 A schematic diagram of the process of obtaining a disease vector using a large-scale language model in the present invention;

[0027] Figure 3 A schematic diagram of dividing a user behavior sequence into session sequences based on the interval time of the interactive behavior according to the present invention;

[0028] Figure 4 This is a schematic diagram of the present invention using the user's historical conversation to supplement the current conversation information based on the noise reduction attention mechanism;

[0029] Figure 5 Schematic diagram of similar patient mining of the present invention. DETAILED DESCRIPTION

[0030] In order to further understand the present invention, a doctor recommendation algorithm based on large-scale language representation to alleviate data sparsity provided by the present invention is described in detail below in combination with a specific implementation method, but the present invention is not limited to this. Non-essential improvements and adjustments made by technical personnel in this field under the core guiding idea of ​​the present invention still fall within the scope of protection of the present invention.

[0031] The problem definition of this method is to recommend matching doctors to patients based on the interactive behavior of patients in the online medical platform, large-scale language models, and the introduction of the diseases that doctors are good at. To this end, this method designs a doctor recommendation algorithm based on large-scale language representation to alleviate data sparsity. The forward propagation part of the algorithm mainly consists of five parts. The first part is to use the large-scale language model to obtain the vector representation of the disease, and use the vector dimension reduction algorithm to reduce the dimension of the disease vector; based on the disease vector after dimension reduction, the vector of the doctor in the platform is initialized; the second part is to divide the patient behavior data into a conversation list, and calculate the conversation vector according to the vector of the doctor in the conversation; the third part is to take the historical conversation list of the current patient as input and use the denoising attention mechanism method to supplement the information of the current conversation; the fourth part is to mine similar patients of the current patient, use the denoising attention mechanism method to obtain the vector representation of similar patients, and then use the vector representation of similar patients to supplement the information of the current patient; the fifth part is to recommend doctors to patients based on the patient vector representation.

[0032] like Figure 1 As shown, according to one embodiment of the present invention, the method includes the following steps:

[0033] S100, using a large-scale language model to obtain a vector representation of the disease, and using a vector dimensionality reduction algorithm to reduce the dimension of the disease vector; based on the disease vector after dimensionality reduction, the vectors of doctors in the platform are initialized. Traditional recommendation algorithms usually use ID-based recommendations, randomly initialize the doctor vector, and then use the patient's collaborative information to train the vector. However, the patient interaction data in the online medical platform is too sparse, and it is impossible to obtain a high-quality doctor vector representation using only collaborative data training. Therefore, this method uses the world knowledge learned from the large-scale language model to initialize the doctor vector. The doctor's homepage in the online platform has an introduction to the diseases he is good at. For example, the diseases that one doctor is good at are: cirrhosis, liver disease, liver fibrosis, liver cancer, ascites, chronic hepatitis B and hepatobiliary diseases; the diseases that another doctor is good at are: gastritis, Helicobacter pylori infection, gastric ulcer, hepatitis B, enteritis, hyperlipidemia and acute enteritis. This method counts all the disease sets in the online platform, and inputs the disease text into the large-scale language model to obtain a high-dimensional vector of the disease. The mathematical formula of this process is Disease text c i Input into the large-scale language model LLM to obtain a high-dimensional vector of the disease like Figure 2 As shown. The large-scale language model LLM can use an open source language model or a closed source commercial large-scale language model. For the open source large-scale language model, the Vicuna model can be used. The latent vector of the last layer of the open source LLM is averaged and pooled to obtain the high-dimensional vector of the disease. The Vicuna model has different parameter versions. If Vicuna-7B is used, the disease vector dimension is 4096, and if Vicuna-13B is used, the disease vector dimension is 5120. If a closed-source commercial LLM is used, directly call the public API interface of LLM to obtain the disease vector The closed-source commercial LLMs include text-embedding-ada-v2 and text-embedding-3-large. The vector dimension obtained by text-embedding-ada-v is 1536, and the vector dimension output by text-embedding-3-large is 3072. It can be seen that the dimension of the disease vector obtained by the large-scale language model is very high. Directly using this high-dimensional vector to participate in the model recommendation will lead to excessive complexity of the recommendation algorithm. Therefore, this method uses principal component analysis (PCA) to reduce the dimension of the disease vector, and the disease vector after dimension reduction is still used To express, that is In this method, the high-dimensional disease vector is reduced to 64 dimensions. All disease vectors in the online medical platform are stored offline in the system. The recommendation algorithm directly reads the offline stored disease vectors without running a large-scale language model. Therefore, the large-scale language model does not increase the complexity of the recommendation algorithm. This method uses the average pooling operation to initialize the doctor vector with the disease vector that the doctor is good at. The disease set that a certain doctor v is good at is dis v = {gastritis, Helicobacter pylori infection, gastric ulcer, hepatitis B, enteritis, hyperlipidemia, acute enteritis}, use the average pooling operation to calculate the disease vectors in the disease set, and get the initialization vector of the doctor v. The mathematical formula is expressed as

[0034] S200, divide the patient behavior data into a conversation list, and calculate the conversation vector based on the doctor's vector in the conversation. The patient behavior data includes the patient's browsing, collection and rating behaviors in the platform. The patient behavior in the online medical platform is highly conversational, and the patient behavior sequence can be divided into several sessions according to the interval time of the patient's behavior. In a user conversation in the online platform, the user's needs are usually relatively simple and clear. Figure 3 As shown in Figure 2, this method divides the patient behavior sequence into several sessions at intervals of 12 hours. For patient u, the session sequence is represented as t is the length of the session list S(u). The included doctors are represented by n is the session The length of any doctor v is represented by e v , which is initialized to emb v , vector e v Updated during model training to learn collaborative information. Encode and get the vector of the session, expressed as The sequence encoder ENcoder can be any commonly used sequence encoder, and can use a variant of the recurrent neural network RNN, the long short-term memory network LSTM, or a simple average pooling operation.

[0035] S300, taking the historical conversation list of the current patient as input and using the noise reduction attention mechanism method to supplement the information of the current conversation. The previous step divides the patient's behavior sequence into several conversation lists and obtains the vector representation of any conversation. The conversation sequence of patient u is in is the current session, is a historical session. A patient has a limited number of diseases, and there are usually cases where similar diseases are consulted between multiple sessions. Therefore, this method uses the patient's historical session information to supplement the current session. However, there are also different sessions of the same user that are generated due to different and unrelated diseases, such as a user's historical consultation diseases include gastritis, skin allergies, and spondylosis. And according to technical analysis, there is a phenomenon in the system where multiple patients share one account, such as children consulting their parents online. Therefore, there are sessions in the patient's historical session sequence that are completely unrelated to the current session. Existing methods usually use an attention mechanism, taking the current session as a query and the historical session sequence as a key and value, giving a larger weight to historical sessions similar to the current session, and a smaller weight to unrelated historical sessions. This method gives a smaller weight to historical sessions that are unrelated to the current session, which introduces more noise to the patient vector. The intuitive solution is to give a weight of 1 to the top k most relevant historical sessions and a weight of 0 to other unrelated historical sessions. The mathematical expression is:

[0036]

[0037] The score function is used to calculate the patient's historical session and the current session The score function is implemented by the attention mechanism. However, this method is not differentiable and the model cannot be trained. Therefore, this method proposes a denoising attention mechanism method to obtain the influence vector h of the patient's historical session on the current session his , the process is expressed as The process is as Figure 4 As shown in the figure, the thick line indicates that the weight of the historical session is high, and the thin line indicates that the weight of the historical session is low. The denoising attention mechanism specifically uses Gumbel softmax to denoise the weight of the historical session calculated by the attention mechanism. The input of Gumbel softmax is a probability distribution, and the output is a probability distribution that approximates a one-hot vector. The specific process of the denoising attention mechanism is divided into three steps. The first step is to calculate any historical session of the current patient. and the current session The correlation The score function is implemented by the attention mechanism. The specific formula is:

[0038]

[0039] in, It is a conversation The second step is to use Gumbelsoftmax to sample and get the historical session The new weight πi , π i =Gumbel_softmax(p i ) is as follows:

[0040]

[0041] Among them, g i is Gumbel noise, whose sampling value is g i =-log(-log(U i )),U i is a random variable sampled from the uniform distribution U(0,1). τ>0 is the temperature coefficient, which controls the smoothness of the output distribution. The temperature coefficient τ→0 will make the distribution closer to the one-hot vector, and the temperature coefficient τ→∞ will make the distribution smoother. The third step is to obtain the influence vector of the patient's historical session on the current session based on the historical session vector and its weight, that is, Fusion of the influence vector and the current session vector gives the new current patient vector: The purpose of the Merge function in this method is to merge two vectors. This method uses a simple vector addition method, that is, Technicians can use other more complex vector fusion methods, such as weighted summation or MLP layer, according to their own scenarios and data.

[0042] S400, mining similar patients of the current patient, using the noise reduction attention mechanism method to obtain similar patient vector representation, and then using the similar patient vector representation to supplement the information of the current patient. Other patients who have generated ratings and favorites in the current session are considered similar to the current patient. The doctors included are It is a conversation For any doctor v, the number of patients who have rated and collected the doctor is N(v). Note that in order to reduce the impact of noise, N(v) does not include patients who browsed the doctor's homepage. Based on the above mathematical symbols, the current similar patients of patient u are expressed as The process of mining similar patients is as follows: Figure 5 shown.

[0043] The process of using similar patients' information to supplement the current patient's information consists of two steps. The first step is to use the current patient as a query and use the denoising attention mechanism to obtain the vector representation of similar patients. The second step is to use the current patient as a query and use the attention mechanism to obtain the influence of similar patients on the current patient. The output of the first step is the vector representation of similar patients. Any similar patient u j , the calculation formula of its vector is S his (u j ) is a similar patient u j The calculation process of the denoising attention mechanism Attn_denoise has been described in the previous step, that is, first use the attention mechanism to u As a query, get S his (u j ) in the session; then Gumbel Softmax is used according to S his (u j ) to resample the session weights to obtain new weights; finally, combined with S his (u j ) to obtain the vectors of similar patients.

[0044] In the second step, the attention mechanism is used to group similar patients The information in the data is aggregated to obtain the influence vector h of similar patients on the current patient. social :

[0045]

[0046] Among them, β j Similar patients j Finally, the influence vector of similar patients and the current patient vector are fused to obtain the final current patient vector The purpose of all Merge functions in this method is to merge two vectors. This method uses a simple vector addition method, that is, Technicians can use other more complex vector fusion methods, such as weighted summation or MLP layer, according to their own scenarios and data.

[0047] S500, recommend a doctor to the patient based on the patient vector representation. τ The vector representation of e τ Multiply by the patient vector Then use the softmax function to calculate the doctor v τ Score:

[0048]

[0049] in, represents the patient vector, e τ Doctor τ The vector representation of Indicates doctor v τ The possibility of being browsed. For this patient sample data, the loss function is:

[0050]

[0051] Among them, y τ Representative Doctor τ One-hot encoding. The function is optimized using the Adam optimizer.

[0052] The above description of the embodiments is to facilitate the understanding and application of the present invention by those skilled in the art. 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 work. Therefore, the present invention is not limited to the above embodiments, and improvements and modifications made to the present invention by those skilled in the art based on the disclosure of the present invention should be within the scope of protection of the present invention.

Claims

1. A doctor recommendation algorithm based on large-scale language representation to alleviate data sparsity, characterized by: A large-scale language model is used to obtain a vector representation of the disease, and a vector dimensionality reduction algorithm is used to reduce the dimension of the disease vector. The vectors of doctors in the platform are initialized based on the reduced disease vector. The homepage of the doctor in the online platform has an introduction to the diseases he is good at. This method counts all the disease sets in the platform and inputs the disease text into the large-scale language model to obtain a high-dimensional vector of the disease. The process is expressed in mathematical symbols as c " is the disease text, LLM is the large-scale language model; after obtaining the high-dimensional disease vector, the principal component analysis method is used to reduce the dimension of the disease vector, and the disease vector after dimension reduction is still used To express, PCA stands for principal component analysis; All disease vectors in the online medical platform are stored offline in the system; the doctor vector is initialized as the average pool result of the disease vectors that the doctor is good at, and the mathematical formula is expressed as dis # It is the set of diseases that doctor v is good at; The patient behavior data is divided into a conversation list, and the conversation vector is calculated based on the vector of the doctor in the conversation; the patient behavior data includes the patient's browsing, collection and rating behaviors in the platform; This method divides the patient's behavior sequence into several sessions according to the interval time of the behavior; the session list of patient u is t is the length of the session list S(u); The included doctors are represented by n is the session The length of any doctor v is represented by e # , which is initialized to emb # , vector e # Updated during model training; using a sequence encoder for any session Encode and get the vector of the session, expressed as The historical conversation list of the current patient is taken as input and the noise reduction attention mechanism method is used to supplement the information of the current conversation; the conversation sequence of patient u is in is the current session, is a historical conversation; this method designs a denoising attention mechanism method to calculate the influence vector of the patient's historical conversation on the current conversation, and uses Gumbel softmax to denoise the historical conversation weights calculated by the attention mechanism. The process is expressed as The specific process of the denoising attention mechanism Attn_denoise is divided into three steps. The first step is to calculate any historical session of the current patient. and the current session The correlation The score function is implemented by the attention mechanism; the specific formula is: in, It is a conversation The φ(·) function represents the inner product of two vectors. The second step is to use Gumbel softmax to sample and get the historical session The new weight π " , π " =Gumbel_softmax(p " ) is as follows: Among them, g " is Gumbel noise, whose sampling value is g " =-log(-log(U " )),U " is a random variable sampled from the uniform distribution U(0,1); τ>0 is the temperature coefficient, which controls the smoothness of the output distribution; the third step is to obtain the influence vector of the patient's historical session on the current session based on the historical session vector and its weight Fusion of the influence vector and the current session vector gives the new current patient vector: Mining similar patients of the current patient, using the denoising attention mechanism method to obtain similar patient vector representations, and then using similar patient vector representations to supplement the current patient's information; this method uses the current patient's conversation Other patients who have generated ratings and collection behaviors in the doctor are considered to be similar patients to the current patient; the current session of patient u The doctors included in It is a conversation Length; The patients who have rated and collected any doctor v are denoted as N(v), so the current similar patients of patient u are denoted as The process of using similar patients' information to supplement the current patient's information consists of two steps. The first step is to use the current patient as a query and use the denoising attention mechanism to obtain the vector representation of similar patients. The second step is to use the current patient as a query and use the attention mechanism to obtain the influence of similar patients on the current patient. The output of the first step is the vector representation of similar patients. For any similar patient u2, the calculation formula of its vector is S +", Ju2K is a list of historical sessions of similar patient u2; In the second step, the attention mechanism is used to group similar patients The information in the data is aggregated to obtain the influence vector h of similar patients on the current patient. ,3c"45 : Among them, β2 is the weight of the similar patient u2; finally, the influence vector of the similar patient and the current patient vector are fused to obtain the final current patient vector Recommend doctors to patients based on the patient vector representation; recommend doctors v in the doctor set - The vector representation of e - Multiply by the patient vector Then use the softmax function to calculate the doctor v - Score: in, represents the patient vector, e - Doctor - The vector representation of Indicates doctor v - The possibility of being browsed; for this patient sample data, the loss function is: Among them, y - Representative Doctor - One-hot encoding; The function is optimized using the Adam optimizer.

2. The doctor recommendation algorithm based on large-scale language representation to alleviate data sparsity according to claim 1 is characterized by: Said The medium-scale language model LLM is an open source large-scale language model. The latent vector of the last layer of the open source LLM is averaged and pooled to obtain the disease vector 3. The doctor recommendation algorithm based on large-scale language representation to alleviate data sparsity according to claim 1 is characterized by: Said The medium-scale language model LLM is a closed-source commercial LLM. The disease vector is obtained by directly calling the public API interface of the closed-source LLM.

4. The doctor recommendation algorithm based on large-scale language representation to alleviate data sparsity according to claim 1 is characterized by: The sequence encoder Encoder is a long short-term memory network LSTM.

5. The doctor recommendation algorithm based on large-scale language representation to alleviate data sparsity according to claim 1 is characterized by: The sequence encoder Encoder is an average pooling operation.

6. The doctor recommendation algorithm based on large-scale language representation to alleviate data sparsity according to claim 1 is characterized by: The Merge function is an addition operation of two vectors.

Images