Doctor searching and recommending method based on multi-behavior time sequence modeling

Abstract

The invention relates to a doctor searching and recommending method based on multi-behavior time sequence modeling, and the method comprises the steps: firstly generating an optimized query vector through a GAN and a doctor-patient relation knowledge graph, and determining the intention of a patient; secondly, mapping the high-dimensional optimized query vector to a low-dimensional space by adopting hyperplane projection index and partial homomorphic encryption, searching in an encrypted doctor database by using the projected optimized query vector, and screening out a candidate recommended doctor set matched with the intention and preference of the patient; after candidate doctors are screened, the system ranks a candidate recommended doctor set by using memory enhanced ranking in combination with a cross-patient recommendation strategy. The method aims to improve the search experience of the patient while improving the accuracy and safety of doctor retrieval, and provides an efficient and intelligent doctor recommendation scheme for an intelligent hospital guide platform.

Description

Technical Field

[0001] This application relates to the technical field of doctor recommendation, and particularly to a doctor search and recommendation method based on multi-behavior time series modeling. Background Art

[0002] When dealing with multi-behavior data, existing doctor recommendation methods often have problems such as insufficient modeling of behavior data, inability to balance query optimization and data security, personalized ranking and data sparsity, lack of recommendation interpretability, and lack of dynamic optimization and feedback mechanisms. Traditional recommendation systems usually regard patients' behavior data as independent features, ignoring the temporal dependence relationships between behaviors; the sensitivity of medical data requires that the recommendation system must ensure data security during the query process, especially the privacy protection of core medical data. However, traditional encryption methods often lead to a decrease in query efficiency. In the process of doctor recommendation, how to dynamically adjust the doctor ranking according to patients' historical behavior data, especially in the case of data sparsity, remains a difficult problem. Users (patients) of medical recommendation systems usually have relatively high requirements for the interpretability of recommendation results. Traditional recommendation systems often only provide a simple list of doctors and lack detailed explanations for the recommendation results. In addition, patients' preferences and behavior patterns may change over time. Traditional recommendation systems often lack an effective feedback mechanism and cannot dynamically adjust the recommendation strategy according to patients' real-time behaviors. Summary of the Invention

[0003] Based on this, it is necessary to provide a doctor search and recommendation method based on multi-behavior time series modeling, including: S1: Obtain patients' historical behavior data, real-time behavior data, and doctor data, respectively extract features from the real-time behavior data and its timestamp to obtain a behavior feature vector and a time feature vector; perform weighted combination on the behavior feature vector and the time feature vector to obtain a behavior embedding representation; extract features from the doctor data to obtain a doctor feature vector; S2: Use a Transformer-TCN hybrid architecture to model the behavior embedding representation to obtain a weighted behavior representation; pass the constructed doctor-patient relationship knowledge graph through a graph neural network to obtain a path embedding; input the weighted behavior representation and the path embedding into a generative adversarial network to generate an optimized query vector; S3: Encrypt the optimized query vector using a partially homomorphic encryption method, and use a hyperplane projection index to map the doctor feature vector and the encrypted optimized query vector to a low-dimensional space; calculate the matching value between the mapped doctor feature vector and the optimized query vector, and sort the doctors according to the matching value to obtain a candidate recommended doctor set; S4: Analyze the patient's preferences from the historical behavior data, model the immediate needs based on the optimized query vector, and adjust the rankings of the doctors in the candidate recommended doctor set based on the preferences and immediate needs to obtain a doctor recommendation list.

[0004] Beneficial effects: This method aims to improve the accuracy and security of doctor retrieval while enhancing the patient's search experience, providing an efficient and intelligent doctor recommendation method for the intelligent diagnosis guidance platform. Description of the Drawings

[0005] To more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0006] Figure 1 It is a flowchart of the doctor search and recommendation method based on multi-behavior time series modeling in the embodiments of the present application. Detailed Embodiments

[0007] To make the above objects, features, and advantages of the present application more obvious and understandable, the following will give a detailed description of the specific embodiments of the present application with reference to the drawings. Many specific details are set forth in the following description to fully understand the present application. However, the present application can be implemented in many other ways different from those described herein. Those skilled in the art can make similar improvements without departing from the connotation of the present application. Therefore, the present application is not limited by the specific embodiments disclosed below.

[0008] In addition, the terms "first" and "second" are only used for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, the features defined with "first" and "second" may explicitly or implicitly include at least one of such features. In the description of the present application, "a plurality" means at least two, such as two, three, etc., unless otherwise specifically defined.

[0009] As Figure 1 shown, this embodiment provides a doctor search and recommendation method based on multi-behavior time series modeling, including: S1: Obtain the patient's historical behavior data, real-time behavior data, and doctor data, respectively extract features from the real-time behavior data and its time stamp to obtain a behavior feature vector and a time feature vector; perform weighted combination on the behavior feature vector and the time feature vector to obtain a behavior embedding representation; extract features from the doctor data to obtain a doctor feature vector.

[0010] Obtain behavioral data and doctor data from the hospital information system and the online medical platform. The behavioral data is shown in Table 1 below; Table 1 Example of behavioral data, ; The behavioral types include but are not limited to: search, consultation, appointment, evaluation; The text content includes but is not limited to: search keywords, consultation content; The numerical values include but are not limited to: appointment duration, number of consultations.

[0011] Specifically, feature extraction of real-time behavioral data includes: Convert the behavioral types in the real-time behavioral data into one-hot encoding to obtain type feature representations; Input the text content in the real-time behavioral data into a pre-trained BERT model to obtain text feature representations; Perform standardization processing on the numerical values in the real-time behavioral data to obtain numerical feature representations; Concatenate the type feature representations, text feature representations, and numerical feature representations at the same moment to obtain the patient's behavioral feature vector.

[0012] Furthermore, feature extraction of the timestamp includes: Parse the timestamp to parse out the year, month, day, and hour; Represent the year, month, and day numerically; Use sine-cosine encoding to perform periodic encoding on the hour. The calculation formula is: , ; Among them, represents the sine encoding of the hour; represents the cosine encoding of the hour; represents the sine function; represents the cosine function; hour represents the hour; Use numerical values 0 - 6 to represent the week; Adopt binary values to represent the holiday encoding, where 0 is a non-holiday and 1 is a holiday; Represent the time interval between the current behavioral data and the previous behavioral data numerically; Concatenate the numerically represented year, month, day, periodically encoded hour, numerically represented week, holiday encoding, and numerically represented time interval in sequence to obtain the time feature vector.

[0013] Even further, weighted combination of the behavioral feature vector and the time feature vector to obtain the behavioral embedding representation includes: Multiply the time feature vector by the time weight matrix, and add the resulting product to the first bias vector to obtain the time position encoding; Multiply the behavior feature vector by the behavior weight matrix, and add the resulting product to the time position encoding to obtain the behavior embedding representation.

[0014] In this embodiment, the doctor data includes but is not limited to the professional field, title, years of practice experience, academic background, and score. The professional field is represented by an embedding; the title is encoded using labels, where 0-3 are used to represent resident doctor, attending doctor, deputy chief physician, and chief physician respectively; the years of practice experience are represented by numerical values; the academic background is represented by an embedding; and the score is represented by the average score value of the corresponding doctor. Concatenate the content after the above feature extraction to obtain the doctor feature vector.

[0015] S2: Use the Transformer-TCN hybrid architecture to model the behavior embedding representation to obtain the weighted behavior representation; pass the constructed doctor-patient relationship knowledge graph through the graph neural network to obtain the path embedding; input the weighted behavior representation and the path embedding into the generative adversarial network to generate the optimized query vector.

[0016] Specifically, using the Transformer-TCN hybrid architecture to model the behavior embedding representation includes: The TCN branch extracts the local temporal features of the behavior embedding representation through dilated causal convolution, and the calculation formula is: ; where, represents the local temporal feature at the t th time step; m represents the convolutional window size; represents the weight of the i th convolutional kernel; represents the bias of the i th convolutional kernel; d represents the dilation factor; represents the time steps of the behavior embedding representation between time steps t.

[0017] The Transformer branch captures the global dependence features of the behavior embedding representation through the multi-head self-attention mechanism; among them, the query vector, key vector, and value vector are all linearly transformed through the behavior embedding representation.

[0018] Pass the local temporal features and the global dependence features through the gating mechanism to output the gated features, and the calculation formula is: ; where, represents the tThe gated features at a time step; Represents the sigmoid activation function; Represents the gated weight matrix; Represents the t Global dependence features at the time step.

[0019] Element-wise multiply the gated features with the local temporal features to obtain the first product; Multiply the balance number of the gated features with respect to 1 by the global dependence features to obtain the second product; Add the first product and the second product to obtain the temporal feature vector; Multiply the temporal feature vector with the attention parameter matrix to obtain the third product; Multiply the type feature representation by the balance number of the attention parameter matrix with respect to 1 to obtain the fourth product; Add the third product and the fourth product, and pass the resulting result through the LeakyReLU activation function to obtain the attention exponent; Perform exponentiation with the natural base and the attention exponent as the exponent to obtain the attention component at any time step; Divide the attention component at any time step by the sum of the attention components at all time steps to obtain the attention weight at the corresponding time step; Sum the products of the attention weights and the temporal feature vectors at all time steps to obtain the weighted behavior representation.

[0020] Furthermore, passing the constructed doctor-patient relationship knowledge graph through a graph neural network to obtain path embeddings includes: Construct a doctor-patient relationship knowledge graph containing an entity set, a relationship set, and a triple set. The entity set includes objects in the medical field, the relationship set includes the associations between entities, and the triple set includes multiple triples, each triple representing the relationship between a head entity and a tail entity; Pass the entities and relationships in the doctor-patient relationship knowledge graph through a graph neural network to obtain path embeddings.

[0021] Input the weighted behavior representation and the path embeddings into a generative adversarial network to generate an optimized query vector. The discriminator is used to determine whether the generated query vector q is similar to the real query vector q real The discriminator loss function is expressed as: ; where represents the determination probability of the discriminator for the generated query vector, E represents the expectation, It is a real query vector, which extracts semantic information from the patient's historical records (such as search keywords, consultation content), doctors' recommendation records or paths in the knowledge graph, and uses the text embedding model BERT or the graph neural network GNN to convert it into a vector representation. Finally, the aggregated generated vector is used to represent the patient's real query intention. By adversarial training, the generator and discriminator are optimized to make the generated query vector as close as possible to the real query vector, laying a foundation for subsequent efficient retrieval.

[0022] S3: Use the partial homomorphic encryption method to encrypt the optimized query vector, and use the hyperplane projection index to map the doctor feature vector and the encrypted optimized query vector to a low-dimensional space; and calculate the matching value between the mapped doctor feature vector and the optimized query vector, and sort the doctors according to the matching value to obtain a set of candidate recommended doctors.

[0023] Specifically, this step includes: To balance data security and retrieval speed, the present invention divides the data into two levels: core medical data and non-sensitive data. For core medical data (such as doctor qualifications, patient historical consultation records, etc.), partial homomorphic encryption PHE is used to ensure that the data remains encrypted during the calculation process to prevent information leakage; for non-sensitive data (such as doctors' office hours, hospital addresses, etc.), lightweight encryption (such as hash index or fast decryption method based on symmetric encryption) is used to reduce the calculation overhead and improve the query speed.

[0024] Assume that the doctor database D consists of N doctors, and the i doctor feature vector of the th doctor is expressed as: , i represents the n th element in the doctor feature vector of the th doctor; encrypt the highly sensitive information in it with PHE: ; represents PHE encryption;

[0025] Similarly, use the partial homomorphic encryption method to encrypt the optimized query vector to obtain the encrypted optimized query vector. O ( N · M ) is reduced to O ( N / log N ), where N is the number of doctors in the doctor database and M is the feature dimension of each doctor.

[0026] The hyperplane projection index maps the doctor feature vector and the encrypted optimized query vector to a low-dimensional space through a projection matrix. The calculation formula is as follows: , ; Among them, represents the doctor feature vector of the i th doctor in the low dimension; represents the optimized query vector in the low dimension; W represents the projection matrix; T represents the transpose; represents the i th doctor's doctor feature vector; q represents the optimized query vector.

[0027] Since the behavior patterns of different patients are different, some patients are more concerned about the doctor's appointment situation, and some patients may be more concerned about the doctor's patient evaluation. To adapt to these different behavior patterns, this embodiment proposes to dynamically adjust the projection matrix W: ; ; ; Among them, represents the between-class scatter of cross-behavior features, which is used to distinguish features of different patient needs, is the mean of class k, is the overall mean, is the number of samples in class k. represents the within-class scatter of the same patient's behavior (the tightness of similar behaviors); represents the trace.

[0028] Transpose the mapped doctor feature vector and multiply it by the mapped optimized query vector to obtain a matching value; Sort the doctors corresponding to the doctor feature vectors in descending order of the matching value to obtain a candidate recommended doctor set.

[0029] To balance data security and retrieval speed, the data is divided into two levels: core medical data and non-sensitive data, and partial homomorphic encryption and lightweight encryption are used respectively. The query is mapped from high dimension to low dimension through the hyperplane projection index (HPI) to reduce the computational complexity, and the projection matrix is dynamically adjusted according to the patient behavior pattern to improve the query efficiency.

[0030] S4: Analyze the patient's preferences from the historical behavior data, model the immediate needs based on the optimized query vector, and adjust the ranking of the doctors in the candidate recommended doctor set based on the preferences and immediate needs to obtain a doctor recommendation list.

[0031] Rank personalized candidate doctors to make the ranking more in line with patients' preferences; utilize patients' historical behavior data to mine deep matching features; adopt a cross-patient recommendation strategy to improve the recommendation effect when data is sparse. The present invention proposes Memory-Augmented Ranking (MARS), which combines patients' historical behavior, dynamically adjusts doctor rankings and cross-patient recommendation strategies, and uses the behavior of similar patients to optimize rankings in the case of insufficient data.

[0032] Memory-Augmented Ranking is a ranking method that combines patients' long-term and short-term behaviors. Long-term memory refers to patients' historical behavior patterns, reflecting their overall preferences; short-term memory refers to patients' current search intentions, reflecting their immediate needs.

[0033] Specifically, this step includes: Search for the interaction records between patients and doctors from historical behavior data. The interaction records include the doctors with whom the interaction has occurred and the interaction time; Calculate the preference based on the doctors with whom the interaction has occurred and the interaction time. The calculation formula is: ; Where, represents the preference of patient p ; represents the i th doctor with whom the interaction has occurred in the set of candidate recommended doctors; represents the interaction time with the i th doctor with whom the interaction has occurred; represents the interaction record; t represents the current time; represents the importance coefficient; represents the i th doctor feature vector of the doctor with whom the interaction has occurred; thus the preference is mainly determined by the interactions in the recent period while retaining certain historical information.

[0034] Calculate the immediate need based on the optimized query vector and the most recent interaction record. The calculation formula is: ; Where, represents the immediate need of patient p ; represents the fusion ratio; represents the optimized query vector corresponding to patient p ; represents the set of doctors with whom patient p has most recently interacted; represents the j th doctor feature vector of the doctor with whom the interaction has occurred most recently; represents the jThe weight of the doctors interacted recently; usually, it can be calculated by the time decay of the most recent interaction, by controlling variables Controlling the time impact of short-term behavior: ; Among them, represents the interaction time with the doctor interacted with for the j th time; represents the interaction time with the doctor interacted with for the th time.

[0035] Multiply the preference by the adaptive fusion weight, multiply the immediate demand by the balance number of the adaptive fusion weight with respect to 1, add the two products to obtain the doctor preference of the patient; The adaptive fusion weight is adaptively adjusted by the number of historical behaviors of the patient: ; When the patient has more historical behaviors , then , the long-term preference dominates; when the patient has fewer historical behaviors , then , the short-term preference dominates; is the threshold of the number of behaviors, controlling the switching rate.

[0036] Transpose the doctor preference of the patient and the doctor weight matrix constructed by the weights of each doctor, multiply the transposed doctor preference and doctor weight matrix by the doctor feature vector of the doctor to obtain the fifth product; Multiply the adjustment parameter by the memory enhancement term to obtain the sixth product; the memory enhancement term is expressed as: ; Among them, represents the preference memory of the patient for the doctor ; represents the weight of the h th doctor interacted with; represents the cosine similarity between the h th doctor interacted with and the i th doctor interacted with in the candidate recommended doctor set; Add the fifth product and the sixth product to obtain the final matching score; Adjust the ranking of the doctors in the candidate recommended doctor set in descending order of the final matching score to obtain the doctor recommendation list.

[0037] Using the patient's historical behavior data, combined with long-term and short-term preferences, dynamically adjust the doctor ranking through Memory-Augmented Ranking (MARS). When the data is sparse, introduce a cross-patient recommendation strategy to use the data of similar patients to improve the ranking effect and make the recommendation results more in line with the patient's preferences.

[0038] In this embodiment, S4 further includes: when the interaction records are sparse, introduce a cross-patient recommendation strategy, calculate the similarity between the target patient and other patients, and use the doctor preferences of similar patients to update the doctor preferences of the target patient. Calculate the second final matching score based on the updated doctor preferences; adjust the ranking of the doctors in the candidate recommended doctor set from largest to smallest according to the second final matching score to obtain a doctor recommendation list. The doctor preference update formula for the target patient is: ; ; Among them, represents the doctor preference of the updated target patient p ; represents the doctor preference of the target patient p ; represents the weight, which determines the fusion ratio of the target patient and similar patients; represents the set of patients similar to the target patient p ; represents the target patient p and the similar patient the similarity between; represents the doctor preference of the similar patient ; represents the modulus; T represents the transpose, using instead of for doctor ranking to make the recommendation results more accurate.

[0039] In this embodiment, the method further includes: After the sorted doctor recommendation list, the list needs a certain recommendation explanation to allow patients to more intuitively understand the recommendation reasons and improve the trust and acceptance of the recommendation results. Generate natural language recommendation explanations based on the structured information (matching score), text description (area of expertise) of the doctor, and patient evaluations. The generation of the recommendation explanation can be regarded as a weighted fusion model: ; Among them, represents the u th patient; represents the natural language generation model; represents the patient u and the doctor the final matching score of.

[0040] The structured score includes the preference weights of each dimension of the doctor's characteristics by the patient, including the doctor's field, geographical location, doctor experience, etc. Using NLP to parse the doctor's profile, representing the text information of the doctor (including the fields of expertise, research directions, etc.), and using BERT to extract the core keywords: ; Among them, used to obtain the context embedding of the doctor's text, select the most important K keywords. representing the patient's evaluation, and using sentiment analysis to calculate the sentiment score: ; Among them, is the doctor 's z sentiment score of the nth evaluation, calculated by LSTM, ranging from [-1, 1], and Z is the total number of evaluations.

[0041] After the sorted doctor recommendation list, based on the doctor's structured information, text description and patient evaluation, generate a natural language recommendation explanation to improve the trust and acceptance of the recommendation result.

[0042] In this embodiment, the method further includes: After the patient gets the doctor recommendation list, an interaction behavior is generated; the interaction behavior includes clicking on the doctor's details page, directly making an appointment with the doctor, browsing repeatedly but not selecting, and jumping out of the page; Based on the interaction behavior of the patient with each doctor in the doctor recommendation list, calculate the user feedback score, and the calculation formula is: ; Among them, represents the u th patient's C user feedback score for the d th doctor in the doctor recommendation list; represents clicking on the details page of doctor d ; represents directly making an appointment with doctor d ; represents browsing repeatedly but not selecting doctor d ; represents jumping out of the page of doctor ; , , , are the behaviors , , , the importance, and .

[0043] According to the patient's behavioral feedback, dynamically adjust the weights of doctor features to make future search rankings more in line with the patient's preferences. If a certain feature has a greater impact on the patient's behavior, then update the weight of the corresponding doctor based on the user feedback score. The update formula is: ; where, represents the weight of the a th doctor after update; represents the weight of the a th doctor; represents the learning rate; represents the number of patients; represents the u th patient's user feedback score for the C th doctor in the doctor recommendation list i ; represents the expected value of the user feedback score a under the condition of the doctor feature vector of the th doctor.

[0044] Update the projection matrix based on the user feedback score. The update formula is: ; where, represents the updated projection matrix; W represents the projection matrix; represents the update amplitude of the projection matrix; C represents the doctor recommendation list; represents the C th i doctor's doctor feature vector in the doctor recommendation list; T represents the transpose.

[0045] According to the patient's interaction behavior with the recommendation list, dynamically adjust the weights of doctor features and the projection matrix to make future search rankings more in line with the patient's preferences and optimize the recommendation effect.

[0046] The doctor search and recommendation method based on multi-behavior time series modeling provided in this embodiment has the following beneficial effects: By analyzing the behavior sequences of patients such as searches, consultations, and appointments, this method retrieves matching doctors in the doctor database and finally provides personalized doctor recommendation results for patients. Specifically: First, use GAN and the doctor-patient relationship knowledge graph to generate an optimized query vector to clarify the patient's intention; then, use hyperplane projection indexing (HPI) and partial homomorphic encryption (PHE) to map the high-dimensional optimized query vector to a low-dimensional space, and use the projected optimized query vector to search in the encrypted doctor database to screen out a set of candidate recommended doctors that match the patient's intention and preferences. After screening the candidate doctors, the system uses memory-augmented ranking (MARS) combined with a cross-patient recommendation strategy to rank the set of candidate recommended doctors. After the ranked recommendation result list is returned, based on the structured information, text description, and patient evaluation of the doctor, a natural language recommendation explanation is generated to improve the trust and acceptance of the recommendation result. Finally, adjust the doctor feature weights and projection matrix according to the patient behavior feedback. This method aims to improve the accuracy and security of doctor retrieval while enhancing the patient's search experience, providing an efficient and intelligent doctor recommendation solution for the intelligent medical guidance platform.

[0047] The technical features of the above-described embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope recorded in this specification.

[0048] The above-described embodiments only represent several implementation manners of the present application. The description is relatively specific and detailed, but it should not be construed as a limitation on the scope of the patent application. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several modifications and improvements can still be made, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application should be subject to the appended claims.

Claims

1. A doctor search and recommendation method based on multi-behavior time series modeling, characterized in that, Including: S1: Obtain the historical behavior data, real-time behavior data, and doctor data of the patient. Respectively extract features from the real-time behavior data and its timestamp to obtain a behavior feature vector and a time feature vector; perform weighted combination on the behavior feature vector and the time feature vector to obtain a behavior embedding representation; extract features from the doctor data to obtain a doctor feature vector; S2: Use a Transformer-TCN hybrid architecture to model the behavior embedding representation to obtain a weighted behavior representation; pass the constructed doctor-patient relationship knowledge graph through a graph neural network to obtain a path embedding; input the weighted behavior representation and the path embedding into a generative adversarial network to generate an optimized query vector; S3: Encrypt the optimized query vector using a partial homomorphic encryption method, and use a hyperplane projection index to map the doctor feature vector and the encrypted optimized query vector to a low-dimensional space; calculate the matching value between the mapped doctor feature vector and the optimized query vector, and sort the doctors according to the matching value to obtain a set of candidate recommended doctors; S4: Analyze the patient's preferences from the historical behavior data, and model the immediate needs based on the optimized query vector. Adjust the ranking of the doctors in the set of candidate recommended doctors based on the preferences and immediate needs to obtain a doctor recommendation list.

2. The doctor search and recommendation method based on multi-behavior time series modeling according to claim 1, wherein Feature extraction of the real-time behavior data includes: Convert the behavior type in the real-time behavior data into a one-hot encoding to obtain a type feature representation; Input the text content in the real-time behavior data into a pre-trained BERT model to obtain a text feature representation; Perform normalization processing on the numerical values in the real-time behavior data to obtain a numerical feature representation; Concatenate the type feature representation, text feature representation, and numerical feature representation at the same moment to obtain the patient's behavior feature vector.

3. The doctor search and recommendation method based on multi-behavior time series modeling according to claim 1, characterized in that Feature extraction of the timestamp includes: Parse the timestamp to parse out the year, month, day, and hour; Represent the year, month, and day numerically; Use sine-cosine encoding to perform periodic encoding on the hour; Use numerical values 0-6 to represent the week; Use binary values to represent the holiday encoding, 0 for non-holiday, 1 for holiday; Represent the time interval between the current behavior data and the previous behavior data numerically; Concatenate the numerically represented year, month, day, periodically encoded hour, numerically represented week, holiday encoding, and numerically represented time interval in sequence to obtain a time feature vector.

4. The doctor search and recommendation method based on multi-behavior time series modeling according to claim 1, characterized in that Performing weighted combination on the behavior feature vector and the time feature vector to obtain a behavior embedding representation includes: Multiply the time feature vector by a time weight matrix, and add the obtained product to the first bias vector to obtain a time position encoding; Multiply the behavior feature vector by a behavior weight matrix, and add the obtained product to the time position encoding to obtain a behavior embedding representation.

5. The doctor search and recommendation method based on multi-behavior time series modeling according to claim 2, wherein Using a Transformer-TCN hybrid architecture to model the behavior embedding representation includes: The TCN branch extracts local temporal features of the behavior embedding representation through dilated causal convolution; The Transformer branch captures global dependency features of the behavior embedding representation through a multi-head self-attention mechanism; Pass the local temporal features and the global dependency features through a gating mechanism to output gating features; Element-wise multiply the gated feature and the local temporal feature to obtain the first product; Multiply the balance number of the gated feature with respect to 1 and the global dependence feature to obtain the second product; Add the first product and the second product to obtain the temporal feature vector; Multiply the temporal feature vector with the attention parameter matrix to obtain the third product; Multiply the type feature representation and the balance number of the attention parameter matrix with respect to 1 to obtain the fourth product; Add the third product and the fourth product, and pass the obtained result through the LeakyReLU activation function to obtain the attention exponent; Perform exponentiation with the natural number as the base and the attention exponent as the exponent to obtain the attention component at any time step; Divide the attention component at any time step by the sum of the attention components at all time steps to obtain the attention weight at the corresponding time step; Sum the products of the attention weights and the temporal feature vectors at all time steps to obtain the weighted behavior representation.

6. The doctor search and recommendation method based on multi-behavior time series modeling according to claim 1, characterized in that The constructed doctor-patient relationship knowledge graph passes through a graph neural network to obtain path embeddings, including: Construct a doctor-patient relationship knowledge graph including an entity set, a relationship set, and a triple set. The entity set includes objects in the medical field, the relationship set includes associations between entities, and the triple set includes multiple triples, where each triple represents the relationship between a head entity and a tail entity; Pass the entities and relationships in the doctor-patient relationship knowledge graph through a graph neural network to obtain path embeddings.

7. The doctor search and recommendation method based on multi-behavior time series modeling according to claim 1, wherein S3 includes: Use the partial homomorphic encryption method to encrypt the optimized query vector to obtain the encrypted optimized query vector; The hyperplane projection index maps the doctor feature vector and the encrypted optimized query vector to a low-dimensional space through a projection matrix; Transpose the mapped doctor feature vector and multiply it with the mapped optimized query vector to obtain a matching value; Sort the doctors corresponding to the doctor feature vectors in descending order of the matching value to obtain a candidate recommended doctor set.

8. The doctor search and recommendation method based on multi-behavior time series modeling according to claim 7, wherein S4 includes: Search for the interaction records between patients and doctors from the historical behavior data. The interaction records include the doctors with whom the interaction occurred and the interaction time; Calculate the preference based on the doctors with whom the interaction occurred and the interaction time. The calculation formula is: ； Among them, represents the preference of the patient p ; represents the i th doctor interacted with in the set of candidate recommended doctors; represents the interaction time with the i th doctor interacted with; represents the interaction record; t represents the current time; represents the importance coefficient; represents the doctor feature vector of the i th doctor interacted with; Calculate the immediate need based on the optimized query vector and the most recent interaction record. The calculation formula is: ； Among them, represents the immediate needs of the patient p ; represents the fusion ratio; represents the optimized query vector corresponding to the patient p ; represents the set of doctors with whom the patient p has interacted most recently; represents the doctor feature vector of the j th doctor with whom the patient has interacted most recently; represents the weight of the j th doctor with whom the patient has interacted most recently; Multiply the preference by the adaptive fusion weight, multiply the immediate need by the balance number of the adaptive fusion weight with respect to 1, and add the two products to obtain the patient's doctor preference; Transpose the patient's doctor preference and the doctor weight matrix constructed by the weights of each doctor, and multiply the transposed doctor preference and doctor weight matrix with the doctor's doctor feature vector to obtain the fifth product; Multiply the adjustment parameter and the memory enhancement term. The memory enhancement term is expressed as: ； Among them, represents the patient's preference memory for the doctor; represents the weight of the nth h interacted doctor; represents the h cosine similarity between the i nth interacted doctor and the nth interacted doctor in the candidate recommended doctor set; Add the fifth product and the sixth product to obtain the final matching score; Adjust the ranking of the doctors in the candidate recommended doctor set in descending order of the final matching score to obtain a doctor recommendation list.

9. The doctor search and recommendation method based on multi-behavior time series modeling according to claim 8, wherein S4 also includes: when the interaction records are sparse, introducing a cross-patient recommendation strategy, calculating the similarity between the target patient and other patients, and using the doctor preferences of similar patients to update the doctor preferences of the target patient, calculating a second final matching score based on the updated doctor preferences; adjusting the ranking of doctors in the candidate recommended doctor set in descending order of the second final matching score to obtain a doctor recommendation list.

10. The doctor search and recommendation method based on multi-behavior time series modeling according to claim 8, characterized in that, It also includes: After the patient obtains the doctor recommendation list, an interaction behavior is generated; The interaction behavior includes clicking on the doctor details page, directly making an appointment with the doctor, repeatedly browsing but not selecting, and leaving the page; Based on the interaction behavior of the patient with each doctor in the doctor recommendation list, a user feedback score is calculated; Updating the weight of the corresponding doctor based on the user feedback score, and the update formula is: ； Among them, represents the weight of the a -th doctor after update; represents the weight of the a -th doctor; represents the learning rate; represents the number of patients; represents the u -th patient's user feedback score for the C -th doctor in the doctor recommendation list i ; represents the expected value of the user feedback score a -th doctor's doctor feature vector is given, the user feedback score ; Updating the projection matrix based on the user feedback score, and the update formula is: ； Among them, represents the updated projection matrix; W represents the projection matrix; represents the update amplitude of the projection matrix; C represents the doctor recommendation list; represents the doctor recommendation list C in the i doctor feature vector of the T represents the transpose.

Images