A doctor recommendation algorithm based on large-scale language representation to alleviate data sparsity

Abstract

The application discloses a doctor recommendation algorithm for relieving data sparsity based on large-scale language representation, and recommends matched doctors for patients based on the interactive behaviors of patients in an online medical platform, a large-scale language model and the introduction of diseases that doctors are good at. The forward propagation part mainly consists of five parts. The first part is to obtain the vector representation of the disease by using the large-scale language model, and initialize the doctor vector based on the dimension-reduced disease vector; the second part is to divide the patient behavior data into a conversation list, and calculate the conversation vector according to the doctor vector in the conversation; the third part is to take the historical conversation list of the current patient as input and adopt the noise reduction attention mechanism method to supplement the information of the current conversation; the fourth part is to mine similar patients of the current patient, adopt the noise reduction attention mechanism method to obtain the similar patient vector, and supplement the information of the current patient by using the similar patient vector; and the fifth part is to recommend doctors for the patient according to the patient vector.

Description

Technical Field

[0001] This 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 Technology

[0002] The application of smart healthcare services on existing online medical platforms is not yet widespread. One of the most important functions of smart healthcare services is intelligent triage. Intelligent triage uses data provided by the platform's information management system to recommend suitable doctors to users. The data in the management system mainly includes information about doctors on the platform, patient information, and interaction information between patients and doctors: namely, patient browsing, favorites, and rating data of doctors. In e-commerce platform recommendations, recommendation algorithms typically use user interaction data such as purchases, clicks, and favorites; in short video recommendations, recommendation algorithms typically use comment data as positive feedback data. Existing doctor recommendation algorithms in the medical field typically use patient rating data for recommendations, but rating data is extremely sparse. To alleviate the data sparsity problem, this method uses all patient and doctor interaction data, including browsing, favorites, and rating data, as input for recommendations. For example, if a user browses a doctor's homepage information, it indicates that the doctor is skilled in treating the patient's potential illness, meaning that the doctor's information matches the patient's needs. Since the recommendation system's recommendation is based on the matching degree between doctor information and patient needs, using patient browsing data as input is reasonable.

[0003] Existing doctor recommendation systems typically recommend matching doctors to patients based on their current conversational behavior. However, patient interactions are extremely sparse, making it difficult for recommendation algorithms to learn high-quality doctor and patient vector representations from sparse collaborative information. Large-scale language models, through massive training, have learned a wealth of open-world domain knowledge, and recommendations based on large-scale language model representations have been shown to improve the performance of traditional ID-based recommendation algorithms. However, large-scale language models have high algorithmic complexity. The challenge for recommendation algorithms is how to leverage the semantic knowledge learned from large-scale language models without significantly increasing model complexity.

[0004] To alleviate data sparsity, this method further incorporates the patient's historical interactions and those of other patients to supplement the patient's current session information. Doctor recommendation scenarios differ from other recommendation scenarios in that users have limited historical preferences, resulting in relatively singular preferences within a single session and repetitive interactions. Furthermore, many patients share similar or even identical symptoms, which forms the basis of the Mijian Medical Platform in China. Mijian's mission is to build a patient communication platform, helping patients find fellow patients, support each other, and share information. Therefore, utilizing the patient's past interactions and information from other potentially similar patients to expand the patient's current session information is reasonable. However, introducing additional information sources introduces significant noise; therefore, while mitigating data sparsity, this noise needs to be processed. Summary of the Invention

[0005] The technical challenge addressed by this method is mitigating data sparsity in doctor recommendation scenarios. The approach involves using all patient-doctor interaction data from the platform as input to the recommendation algorithm. Additionally, a large-scale language model is used to obtain doctor vectors. Then, the current patient session information is augmented based on the doctor's semantic vectors, with the augmented information sourced from the user's historical sessions and interaction data from other similar patients. However, augmenting the current patient session information with additional information inevitably introduces significant noise. Therefore, mitigating the impact of noise while addressing data sparsity is also a problem this method needs to solve.

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

[0007] This paper utilizes a large-scale language model to obtain vector representations of diseases, and then employs a vector dimensionality reduction algorithm to reduce the dimensionality of these disease vectors. Based on the dimensionality-reduced disease vectors, the paper initializes the vectors of doctors on the platform. Each doctor's homepage on the online platform contains information about their areas of expertise. This method statistically analyzes all disease sets on the platform and inputs the disease text into the large-scale language model to obtain high-dimensional disease vectors. This process is represented mathematically as follows: c i This involves disease text, and LLM (Large-Scale Language Model). After obtaining the high-dimensional disease vector, principal component analysis (PCA) is used to reduce its dimensionality. The dimensionality-reduced disease vector is then used... To indicate, that is PCA stands for Principal Component Analysis. In online healthcare platforms, all disease vectors are stored offline. In the recommendation algorithm, doctor vectors are initialized with the average pooling result of the disease vectors that the doctor is proficient in, mathematically expressed as: dis v It is the collection of diseases that Doctor V specializes in.

[0008] Patient behavior data is segmented into a list of sessions, and a session vector is calculated based on the doctor's vector within each session. Patient behavior data includes the patient's browsing, favorites, and rating activities on the platform. This method segments the patient behavior sequence into several sessions based on the time interval between actions. The session list for patient u is... t is the length of the session list S(u). The doctors included are indicated as n is the conversation The length of the vector is given by e. v It is initialized to emb v Vector e v Updated during model training. A sequence encoder (Encoder) is used for any session. Encode the session to obtain a vector, represented as...

[0009] The current patient's historical conversation list is used as input, and a noise reduction and attention mechanism is employed to supplement information in the current conversation. The conversation sequence of patient u is... in This is the current session. This method involves denoising the patient's historical conversations. A noise reduction attention mechanism is designed to calculate the influence vector of the patient's historical conversations on the current conversation. Gumbel softmax is then used to denoise the historical conversation weights calculated by the attention mechanism. The process is represented as follows: The noise reduction attention mechanism Attn_denois consists of three steps. The first step is to calculate any historical session of the current patient. and the current session Relevance The score function is implemented using an attention mechanism. The specific formula is:

[0010]

[0011] in, It is a conversation The vectors are denoted by φ(·), and the function φ(·) represents the inner product of the two vectors. The second step is to sample using Gumbelsoftmax to obtain the historical sessions. New weight π " , π " =Gumbel_sofotmax(p i The specific formula is:

[0012]

[0013] Among them, g iIt is Gumbel noise, and its sample value is g. i =-log(-log(U) i )),U i It is a random variable sampled from a uniform distribution U(0,1). τ>0 is a temperature coefficient that controls the smoothness of the output distribution. The third step is to obtain the influence vector of the patient's historical sessions on the current session based on the historical session vector and its weights. By fusing the influence vector and the current session vector, a new current patient vector is obtained.

[0014] This method identifies similar patients to the current patient, uses a noise reduction and attention mechanism to obtain similar patient vector representations, and then uses these similar patient vector representations to supplement information about the current patient. This approach integrates the current patient's conversation... Other patients who generate ratings and save entries for the TCM doctor are considered similar to the current patient. Patient u's current session. The doctors included are It is a conversation The length of the ... The process of supplementing the current patient's information using similar patient data involves two steps. The first step uses the current patient as the query and employs a noise reduction and attention mechanism to obtain vector representations of similar patients. The second step uses the current patient as the query and employs an attention mechanism to determine 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 formula for calculating its vector is: S his (u j ) is a similar patient u j The historical conversation list. The second step uses attention mechanisms to group similar patients. By aggregating the information, we can obtain the influence vector h of similar patients on the current patient. social :

[0015]

[0016] Where, β j Are similar patients u j The weights are then determined. Finally, the influence vectors of similar patients and the current patient vector are fused to obtain the final current patient vector.

[0017] Based on the patient vector representation, recommend doctors to the patients. This involves retrieving the doctor v from the doctor set.τ The vector representation of e τ Multiply by patient vector Then use the softmax function to calculate the doctor's v. τ The score:

[0018]

[0019] in, Represents the patient vector, e τ Is a doctor v τ The vector representation of , Doctor v τ The likelihood of being viewed. For this patient sample data, the loss function is:

[0020]

[0021] Among them, y τ Representative doctor v 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 doctors 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 vector dimensionality reduction algorithm is used to reduce the dimensionality of the disease vectors and the doctor vectors are initialized according to the diseases that doctors are 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. Attached Figure Description

[0025] Figure 1 This is a flowchart illustrating a doctor recommendation algorithm based on large-scale language representation to alleviate data sparsity according to the present invention.

[0026] Figure 2 This is a schematic diagram illustrating the process of obtaining disease vectors using a large-scale language model in this invention;

[0027] Figure 3 This is a schematic diagram illustrating how the present invention divides user behavior sequences into session sequences based on the interval time of interactive behaviors;

[0028] Figure 4 This is a schematic diagram illustrating how the present invention supplements the current session with information based on the user's historical sessions using a noise reduction attention mechanism.

[0029] Figure 5 This is a schematic diagram illustrating the discovery of similar patients according to the present invention. Detailed Implementation

[0030] To further understand the present invention, the following describes in detail a doctor recommendation algorithm based on large-scale language representation to alleviate data sparsity, with reference to 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 principles of the present invention are still within the scope of protection of the present invention.

[0031] The problem defined in this method is to recommend suitable doctors to patients based on their interactive behavior on an online medical platform, a large-scale language model, and descriptions of doctors' areas of expertise. To address this, this method designs a doctor recommendation algorithm based on large-scale language representations to alleviate data sparsity. The forward propagation part of this algorithm consists of five main components. The first part uses a large-scale language model to obtain vector representations of diseases and then applies a vector dimensionality reduction algorithm to reduce the dimensionality of these disease vectors; based on the dimensionality-reduced disease vectors, the doctors' vectors in the platform are initialized. The second part segments the patient's behavioral data into a session list and calculates the session vector based on the doctor's vector in each session. The third part uses the current patient's historical session list as input and employs a denoising attention mechanism to supplement information in the current session. The fourth part identifies similar patients to the current patient, obtains similar patient vector representations using a denoising attention mechanism, and then uses these similar patient vector representations to supplement information in the current patient's session. The fifth part recommends doctors to the patient based on their vector representation.

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

[0033] S100 utilizes a large-scale language model to obtain vector representations of diseases, and employs a vector dimensionality reduction algorithm to reduce the dimensionality of these disease vectors. Based on the dimensionality-reduced disease vectors, vectors are initialized for doctors on the platform. Traditional recommendation algorithms typically use ID-based recommendations, randomly initializing doctor vectors and then training these vectors using patient collaboration information. However, patient interaction data on online medical platforms is too sparse, and training solely using collaboration data cannot yield high-quality doctor vector representations. Therefore, this method initializes doctor vectors using world knowledge learned from a large-scale language model. Doctors' homepages on the online platform include descriptions of their areas of expertise. For example, one doctor specializes in: cirrhosis, liver disease, liver fibrosis, liver cancer, ascites, chronic hepatitis B, and hepatobiliary diseases; another doctor specializes in: gastritis, Helicobacter pylori infection, gastric ulcer, hepatitis B, enteritis, hyperlipidemia, and acute enteritis. This method statistically analyzes all disease sets on the online platform and inputs the disease text into a large-scale language model to obtain high-dimensional disease vectors. The mathematical formula for this process is: The disease text c i The data is input into a large-scale language model (LLM) to obtain a high-dimensional vector of the disease. like Figure 2 As shown. Large-scale language models (LLMs) can be either open-source or closed-source commercial LLMs. For open-source LLMs, the Vicuna model can be used. By performing average pooling on the latent vectors of the last layer of the open-source LLM, the high-dimensional vector of the disease can be obtained. The Vicuna model has different parameter versions. If Vicuna-7B is used, the disease vector dimension is 4096; if Vicuna-13B is used, the disease vector dimension is 5120. If using a closed-source commercial LLM, the disease vector can be obtained directly by calling the LLM's public API interface. Closed-source commercial LLMs include text-embedding-ada-v2 and text-embedding-3-large. text-embedding-ada-v2 produces vectors with a dimension of 1536, while text-embedding-3-large outputs vectors with a dimension of 3072. This demonstrates that large-scale language models produce disease vectors with very high dimensionality. Directly using these high-dimensional vectors in the recommendation algorithm would lead to excessively high algorithm complexity. Therefore, this method uses Principal Component Analysis (PCA) to reduce the dimensionality of the disease vectors. The dimensionality-reduced disease vectors are still... To indicate, that is In this method, the high-dimensional disease vectors are 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 needing to run a large-scale language model; therefore, the large-scale language model does not increase the complexity of the recommendation algorithm. This method uses an average pooling operation to initialize the doctor vector with the doctor's specialty disease vectors. The set of specialty diseases for a given doctor v is dis. v ={Gastritis, Helicobacter pylori infection, Gastric ulcer, Hepatitis B, Enteritis, Hyperlipidemia, Acute enteritis}, use the average pooling operation to calculate the initial vector of doctor v in the disease vector of the disease set, and obtain the initial vector of doctor v. The mathematical formula is expressed as follows:

[0034] S200 segments patient behavior data into a list of sessions and calculates session vectors based on the doctor's vector within each session. Patient behavior data includes browsing, favorites, and rating behaviors on the platform. Patient behavior on online healthcare platforms is highly conversational; patient behavior sequences can be segmented into several sessions based on the time intervals between actions. Within a single user session on an online platform, the user's needs are typically relatively simple and clear. For example... Figure 3 As shown, this method divides the patient behavior sequence into several sessions according to a 12-hour interval. For patient u, its session sequence is represented as follows: t is the length of the session list S(u). Where the sessions... The doctors included are indicated as n is the conversation The length of the vector is given by e. v It is initialized to emb v Vector e v The model is updated during training to learn collaborative information. A sequence encoder (ENcoder) is used for any session. Encode the session to obtain a vector, represented as... The sequence encoder ENcoder can be any commonly used sequence encoder, such as a variant of the recurrent neural network RNN, the Long Short-Term Memory (LSTM) network, or a simple average pooling operation.

[0035] S300 takes the current patient's historical conversation list as input and supplements the current conversation with information using a noise reduction and attention mechanism. The previous step segmented the patient's behavioral sequence into several conversation lists and obtained a vector representation of any conversation. The conversation sequence of patient u is... in This is the current session. This involves historical sessions. A patient's range of illnesses is limited, and multiple sessions often involve consultations for similar illnesses. Therefore, this method supplements the current session with information from the patient's historical sessions. However, different sessions for the same user may arise from different and unrelated illnesses, such as a user's historical consultations for gastritis, skin allergies, and spinal diseases. Furthermore, technical analysis revealed that multiple patients share a single account, such as children consulting online for their parents. Therefore, the patient's historical session sequence contains sessions completely unrelated to the current session. Existing methods typically use an attention mechanism, treating the current session as the query and the historical session sequence as the key and value, assigning higher weights to historical sessions similar to the current session and lower weights to unrelated ones. This method, by assigning lower weights to unrelated historical sessions, introduces significant noise into the patient vector. An intuitive solution is to assign a weight of 1 to the top k most relevant historical sessions and a weight of 0 to the remaining unrelated ones. The mathematical expression is:

[0036]

[0037] The score function calculates any historical session of the patient. and the current session The relevance function and score function are implemented using an attention mechanism. However, this method is non-differentiable, making model training impossible. Therefore, this paper proposes a denoising attention mechanism to obtain the influence vector h of the patient's historical sessions on the current session. his The process is represented as The process is as follows Figure 4 As shown, thick lines represent high weights for historical sessions, while thin lines represent low weights. The noise reduction attention mechanism specifically uses Gumbel softmax to denoise the historical session weights calculated by the attention mechanism. The input to Gumbel softmax is a probability distribution, and the output is a probability distribution approximating a one-hot vector. The noise reduction attention mechanism consists of three steps: the first step calculates any historical session of the current patient. and the current session Relevance The score function is implemented using an attention mechanism. The specific formula is:

[0038]

[0039] in, It is a conversation The vectors are denoted by φ(·), and the function φ(·) represents the inner product of the two vectors. The second step is to sample using Gumbelsoftmax to obtain the historical sessions. New weight πi , π i =Gumbel_softmax(p i The specific formula is:

[0040]

[0041] Among them, g i It is Gumbel noise, and its sample value is g. i =-log(-log(U) i )),U i It is a random variable sampled from a uniform distribution U(0,1). τ>0 is a temperature coefficient that controls the smoothness of the output distribution. A temperature coefficient τ→0 makes the distribution closer to a one-hot vector, while a temperature coefficient τ→∞ makes the distribution smoother. The third step is to obtain the influence vector of the patient's historical sessions on the current session based on the historical session vector and its weights. By fusing the influence vector and the current session vector, a new current patient vector is obtained. The purpose of the Merge function in this method is to merge two vectors. This method uses a simple vector addition approach, i.e. Technicians can use other more complex vector fusion methods, such as weighted summation or MLP layers, depending on their own scenarios and data.

[0042] S400 involves identifying similar patients to the current patient, employing a noise reduction and attention mechanism to obtain similar patient vector representations, and then using these similar patient vector representations to supplement information about the current patient. This method integrates the current patient's conversation... Other patients who generate ratings and save entries for the TCM doctor are considered similar to the current patient. Patient u's current session. The doctors included are It is a conversation The length of the term is given. For any doctor v, the number of patients who rate and favorite that doctor is denoted as N(v). Note that to reduce noise, N(v) does not include patients who have viewed the doctor's homepage. Based on the above mathematical notation, the current similar patients of patient u are denoted as... The process of identifying similar patients, such as Figure 5 As shown.

[0043] The process of supplementing the current patient's information using similar patient data involves two steps. The first step uses the current patient as the query and employs a noise reduction and attention mechanism to obtain vector representations of similar patients. The second step uses the current patient as the query and employs an attention mechanism to determine 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 formula for calculating its vector is: S his (u j ) is a similar patient u j The historical session list. The calculation process of the noise reduction attention mechanism Attn_denoise has been described in the previous step, that is, first use the attention mechanism to denoise h u As a query, S is obtained. his (u j The weights of the sessions are determined in S; then Gumbel Softmax is used based on S. his (u j The session weights are resampled to obtain new weights; finally, they are combined with S. his (u j The session vector representation and weights are used to obtain vectors of similar patients.

[0044] The second step uses attention mechanisms to connect similar patients. By aggregating the information, we can obtain the influence vector h of similar patients on the current patient. social :

[0045]

[0046] Where, β j Are similar patients u j The weights are then determined. Finally, the influence vectors 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 approach, i.e. Technicians can use other more complex vector fusion methods, such as weighted summation or MLP layers, depending on their own scenarios and data.

[0047] S500, based on the patient vector representation, recommends doctors to patients. It then selects doctors v from the doctor set. τ The vector representation of e τ Multiply by patient vector Then use the softmax function to calculate the doctor's v. τ The score:

[0048]

[0049] in, Represents the patient vector, e τ Is a doctor v τ The vector representation of , Doctor v τ The likelihood of being viewed. For this patient sample data, the loss function is:

[0050]

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

[0052] The above description of the embodiments is provided to enable those skilled in the art to understand and apply the present invention. It will be apparent to those skilled in the art that various modifications can be made to the above embodiments, and the general principles described herein can be applied to other embodiments without inventive effort. Therefore, the present invention is not limited to the above embodiments, and any improvements and modifications made to the present invention by those skilled in the art based on the disclosure thereof 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 in that: This paper utilizes a large-scale language model to obtain vector representations of diseases, and then employs a vector dimensionality reduction algorithm to reduce the dimensionality of these disease vectors. Based on the dimensionality-reduced disease vectors, vector initialization is performed on doctors' data within the platform. Since doctors' homepages on the online platform contain descriptions of their areas of expertise, this method statistically analyzes all disease sets within the platform and inputs the disease text into the large-scale language model to obtain high-dimensional disease vectors. This process is represented mathematically as follows: , It is a disease text. It is a large-scale language model; after obtaining the high-dimensional disease vectors, principal component analysis is used to reduce the dimensionality of the disease vectors. The dimensionality-reduced disease vectors are still used... To indicate, , Principal component analysis; In the online medical platform, all disease vectors are stored offline in the system; doctor vectors are initialized with the average pooling result of the disease vectors that the doctor is proficient in, expressed by the mathematical formula as follows: , He is a doctor The collection of diseases that are its strengths; Patient behavior data is segmented into a list of sessions, and a session vector is calculated based on the doctor's vector in the session; patient behavior data includes patients' browsing, favorites, and rating behaviors on the platform; This method segments the patient's behavioral sequence into several sessions based on the interval between behaviors; patient The conversation list is , It is a list of conversations. Length of the session; The doctors included are indicated as , It is a conversation Length; any doctor The vector representation is Its initialization is ,vector Updated during model training; using a sequence encoder. For any session Encode the session to obtain a vector, represented as... ; The current patient's historical conversation list is used as input, and a noise reduction and attention mechanism is employed to supplement information in the current conversation; patient The session sequence is ,in This is the current session. This method involves designing a noise reduction attention mechanism to calculate the influence vector of the patient's historical conversations on the current conversation. Gumbel softmax is then used to denoise the historical conversation weights calculated by the attention mechanism. The process is represented as follows: Noise reduction and attention mechanism The specific process consists of three steps. The first step is to calculate any historical session of the current patient. and the current session Relevance , The function is implemented using an attention mechanism; the specific formula is: in, It is a conversation The vector, The function represents the inner product of two vectors; the second step is to sample using Gumbel softmax to obtain the historical session. New weights , The specific formula is: in, It is Gumbel noise, and its sampled value is , From uniform distribution Random variables sampled from the middle; The first step is to use a temperature coefficient to control the smoothness of the output distribution; the third step is to obtain the influence vector of the patient's historical sessions on the current session based on the historical session vector and its weights. By fusing the influence vector and the current session vector, a new current patient vector is obtained. ; This method identifies similar patients to the current patient, uses a noise reduction and attention mechanism to obtain similar patient vector representations, and then uses these similar patient vector representations to supplement information about the current patient. This approach integrates the current patient's conversation... Other patients who generate ratings and save entries for TCM doctors consider themselves similar to the current patient; Current session The doctors included are , It is a conversation Length; for any doctor Patients who rated and saved the data were considered as follows: Therefore, the patient Current similar patients are represented as The process of supplementing the current patient's information using similar patient information involves two steps. The first step is to use the current patient as the query and employ a noise reduction and attention mechanism to obtain vector representations of similar patients. The second step is to use the current patient as the query and employ an attention mechanism to determine 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 The formula for calculating its vector is: , Similar patients The second step is to use an attention mechanism to group similar patients through a historical conversation list. By aggregating the information, we can obtain the influence vector of similar patients on the current patient. : in, Similar patients The weights are then determined; finally, the influence vectors of similar patients and the current patient vector are fused to obtain the final current patient vector. ; Based on the patient vector representation, recommend doctors to patients; and select doctors from the doctor set. vector representation Multiply by patient vector Then use The function calculates the doctor The score: in, Represents the patient vector. He is a doctor The vector representation of , Indicates doctor The probability of being viewed; for this patient sample data, the loss function is: in, 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, characterized in that: The Medium and large-scale language models It is an open-source, large-scale language model, which will open source The latent vectors in the last layer are subjected to average pooling to obtain the disease vector. .

3. The doctor recommendation algorithm based on large-scale language representation to alleviate data sparsity according to claim 1, characterized in that: The Medium and large-scale language models It is a closed-source business By directly calling the public API interface of the closed-source LLM, disease vectors can be obtained. .

4. The doctor recommendation algorithm based on large-scale language representation to alleviate data sparsity according to claim 1, characterized in that: The sequence encoder It is a Long Short-Term Memory (LSTM) network.

5. The doctor recommendation algorithm based on large-scale language representation to alleviate data sparsity according to claim 1, characterized in that: The sequence encoder It is an average pool operation.

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

Images