System and method for demand adaptation based on individual profile

Abstract

The application discloses a demand adaptation system and method based on individual portraits, and belongs to the technical field of intelligent matching, and comprises the following steps: collecting original multi-modal asynchronous data of demand parties, inputting the processed data into a cross-modal feature fusion network to extract features, and generating a dynamic demand portrait; decomposing and incrementally learning historical service records of service parties to construct a dynamic capability portrait; then, a multi-dimensional demand-capability matching model is used to calculate a matching vector, and a comprehensive adaptation degree score, a matching decision traceability graph and a service party priority ranking list are output; finally, service feedback data is collected, and the matching decision traceability graph and the like are combined as samples to input an intelligent agent, and the matching function and the weight generation strategy in the iteration optimization framework are optimized; dynamic and accurate matching of demand and service capability is realized, and the matching mechanism can be continuously optimized through feedback, so that the effectiveness and intelligent level of adaptation are improved.

Description

Technical Field

[0001] This invention belongs to the field of intelligent matching technology, specifically a demand adaptation system and method based on personalized profiles. Background Technology

[0002] In the field of intelligent service supply and demand matching, traditional demand adaptation methods are mostly based on static labels or single-dimensional data to complete the matching. It is difficult to capture the real-time physiological and cognitive states and dynamic interaction preferences of the demand side, and it is also unable to accurately depict the multi-dimensional characteristics and temporal changes of the service provider's capabilities, resulting in insufficient accuracy and adaptability of supply and demand matching. At the same time, traditional methods lack the ability to finely decompose the historical service records of service providers, making it difficult to extract core capability units and build dynamically updated capability profiles, resulting in poor completeness and timeliness of capability representation. Existing matching frameworks have a single target optimization dimension and mostly use fixed weights to calculate the matching degree. They cannot dynamically adjust the weight coefficients of each matching dimension according to the actual situation, resulting in insufficient flexibility in adapting to the scenario. Moreover, the matching decision-making process lacks a visual traceability mechanism, resulting in poor interpretability of the matching results and difficulty in tracing the contribution of each dimension to the matching results. In addition, traditional methods have not established an effective closed-loop optimization mechanism, and cannot effectively feed the service effect feedback data back to the matching model. The model is difficult to iteratively optimize according to the actual service situation, and the long-term adaptation efficiency and effect are difficult to continuously improve, which can no longer meet the needs of intelligent and dynamic supply and demand adaptation services. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention proposes a demand adaptation system and method based on personalized profiles. It collects native multimodal asynchronous data from demanders, processes it, and then inputs it into a cross-modal feature fusion network to extract features, generating a dynamic demand profile. It then constructs a dynamic capability profile by decomposing and incrementally learning the historical service records of service providers. Subsequently, it calculates matching vectors through a multi-dimensional demand-capability matching model, outputting a comprehensive adaptation score, a matching decision tracing graph, and a service provider priority ranking list. Finally, it collects service feedback data and uses it, combined with the matching decision tracing graph, as samples to input the intelligent agent, iteratively optimizing the matching function and weight generation strategy in the framework. This achieves dynamic and accurate matching of demands and service capabilities, and continuously optimizes the matching mechanism through feedback, improving the effectiveness and intelligence of adaptation.

[0004] To achieve the above objectives, the present invention provides the following technical solution:

[0005] Personalized profile-based demand adaptation methods include:

[0006] By using a multi-source heterogeneous sensor array and interactive probes, native multimodal asynchronous data from the demand side is collected in real time to generate an enhanced synchronous sensing time sequence.

[0007] Feature extraction and modal interaction analysis are performed on enhanced synchronous sensing time-series sequences to generate structured dynamic demand profiles;

[0008] Retrieve the service provider's historical service records, break them down into the smallest capability units, and perform incremental learning to build and update the service provider's dynamic capability profile.

[0009] Input the dynamic demand profile and dynamic capability profile into the pre-built multi-dimensional demand-capability matching model, calculate the native matching vector of the preset dimensions, and output the comprehensive suitability score, the matching decision traceability diagram and the service provider priority ranking list.

[0010] After completing the service according to the service provider priority ranking list, feedback data on the service effect is collected. The feedback data, the matching decision traceability graph, and the service provider priority ranking list are used as training samples to input the agent, and the matching function and weight generation strategy in the multi-dimensional demand-capability matching model are dynamically adjusted.

[0011] Specifically, the native multimodal asynchronous data includes physiological sensing time-series streams, interactive behavior event streams, and environmental physical parameter streams;

[0012] The generation process of the enhanced synchronous sensing time sequence includes: establishing a unified time series benchmark, performing timestamp alignment and interpolation processing on the physiological sensing time series stream, interactive behavior event stream, and environmental physical parameter stream to generate time-aligned data frames; assigning a unified context identifier code to each time-aligned data frame through a preset context semantic rule base, and generating an enhanced synchronous sensing time sequence after arranging them in time order.

[0013] Specifically, the feature extraction and modal interaction analysis of the enhanced synchronous sensing time series is performed using a cross-modal feature fusion network, which is a three-layer encoder-decoder structure, specifically:

[0014] The enhanced synchronous sensing temporal sequence is input into the visual encoder, temporal encoder, and text encoder respectively for feature extraction, generating visual feature vectors, temporal feature vectors, and text feature vectors. The interaction weights between the visual feature vectors, temporal feature vectors, and text feature vectors are calculated through a cross-attention mechanism, and the visual feature vectors, temporal feature vectors, and text feature vectors are weighted and fused according to the interaction weights to generate a fused high-dimensional dynamic feature vector, which is then input into a sparse attention feature decoupling unit. The sparse attention feature decoupling unit uses a multi-head attention mechanism to calculate the sparse weights between each feature dimension within the fused high-dimensional dynamic feature vector. Based on the sparse weights, the fused high-dimensional dynamic feature vector is parsed into a state quantization vector cluster and an interaction preference probability transition matrix, which together constitute a structured dynamic demand profile.

[0015] The state quantization vector cluster is a set of multiple numerical vectors representing real-time physiological and cognitive states; the interaction preference probability transition matrix is ​​a two-dimensional probability matrix used to describe the dynamic transition probability between different interaction methods; the two-dimensional probability matrix is ​​obtained by the decoder performing a normalized exponential function operation on the fused high-dimensional dynamic feature vector.

[0016] Specifically, retrieving the service provider's historical service records and decomposing them into the smallest capability units includes: extracting service action, operation object, and result evaluation triples by performing semantic and dependency parsing analysis on the historical service records; and mapping the triples to the smallest capability units with unified semantic identifiers based on a preset capability atomization classification dictionary.

[0017] Specifically, the incremental learning process, which constructs and updates the dynamic capability profile of the service provider, includes:

[0018] The minimum capability unit is used as a node of a graph, and the temporal association and semantic co-occurrence relationship between the minimum capability units are used as edges of the graph to construct an initial capability graph. A neural network with temporal dependency graph node feature incremental update is used to aggregate and update the node features of the initial capability graph.

[0019] When a new historical service record is input, the minimum capability unit, the edge weights between nodes, and the node feature vectors in the neural network with time-dependent graph node features are updated through an incremental learning algorithm to form an updated capability graph. The updated capability graph serves as a dynamic capability profile of the service provider.

[0020] Specifically, the dynamic capability profile is a domain-pattern dual embedding network, which includes a domain subgraph and a pattern subgraph;

[0021] The domain subgraph contains multiple first nodes, each first node representing a specialty domain. The first nodes are connected by edges, and the edges between the first nodes represent the semantic association strength between the specialty domains.

[0022] The pattern subgraph contains multiple second nodes, each of which represents an operational habit pattern. The second nodes are connected by edges, and the edges between the second nodes represent the state transition probabilities between operational habit patterns.

[0023] The first node of the domain subgraph and the second node of the pattern subgraph are associated and bridged through the smallest capability unit obtained by decomposing from the historical service records of the same service provider.

[0024] Specifically, the multi-dimensional demand-capability matching model includes a matching degree calculation layer, a dynamic weight fusion layer, and a decision backtracking layer;

[0025] The matching degree calculation layer receives the dynamic demand profile and the dynamic capability profile, and calculates the cosine similarity value on multiple preset dimensions to obtain the original matching vector; each component of the original matching vector corresponds to a cosine similarity value of a preset dimension.

[0026] The dynamic weight fusion layer is implemented based on a long short-term memory network. The long short-term memory network takes the context identifier code corresponding to the enhanced synchronous perception time sequence and the historical weight sequence as input, outputs the adaptive weight coefficients of each preset dimension, and performs a weighted summation of the original matching vector and the adaptive weight coefficients to generate a comprehensive fit score.

[0027] The decision backtracking layer records the cosine similarity value of each preset dimension in the matching degree calculation layer, as well as the corresponding adaptive weight coefficient in the dynamic weight fusion layer. Based on the recorded cosine similarity value and adaptive weight coefficient, a tree-structured matching decision backtracking graph is generated to visualize the contribution path of each dimension to the comprehensive fit score.

[0028] Specifically, the process of generating the matching decision tracing graph includes:

[0029] The overall fit score is used as the root node;

[0030] Each preset dimension is connected to the root node as a child node;

[0031] Configure two leaf nodes for each child node. The first leaf node stores the cosine similarity value corresponding to the preset dimension, and the second leaf node stores the adaptive weight coefficient corresponding to the preset dimension.

[0032] Based on the magnitude of the adaptive weight coefficients, all child nodes connected to the root node are sorted and visualized to generate a tree-structured matching decision tracing graph.

[0033] Specifically, the agent is implemented using a proximal policy optimization algorithm, including an actor network and a critic network;

[0034] Adjusting the matching function in the multi-dimensional demand-ability matching model means dynamically modifying the feature weights of the dynamic demand profile and dynamic ability profile used in calculating cosine similarity in the matching degree calculation layer through the parameters output by the actor network.

[0035] Adjusting the weight generation strategy in the multi-dimensional demand-ability matching model means updating the network connection weights of the long short-term memory network in the dynamic weight fusion layer using the parameters output by the actor network.

[0036] The demand adaptation system based on personalized profiles includes: a multi-source data acquisition module, a demand profile generation module, a capability profile construction module, a multi-objective matching optimization module, and an iterative optimization module.

[0037] The multi-source data acquisition module is used to collect physiological sensor time-series streams, interactive behavior event streams, and environmental physical parameter streams from the demand side in real time, forming native multimodal asynchronous data. After timestamp alignment and context identification, it generates an enhanced synchronous sensing time-series sequence.

[0038] The demand profile generation module inputs the enhanced synchronous perception time sequence into a pre-trained cross-modal feature fusion network to generate a structured dynamic demand profile.

[0039] The capability profile building module retrieves the service provider's historical service records, decomposes them into the smallest capability units using a semantic analysis-based service capability decomposition method, builds and updates the service provider's dynamic capability profile through incremental learning, and updates it in real time based on new historical service records.

[0040] The multi-objective matching optimization module inputs dynamic demand profiles and dynamic capability profiles into a multi-dimensional demand-capability matching model, generates a comprehensive fit score and a matching decision tracing diagram, and outputs a service provider priority ranking list.

[0041] The iterative optimization module is used to collect feedback data on service effectiveness. It uses the feedback data, the matching decision traceability graph, and the service provider priority ranking list as training samples to input the agent, and dynamically adjusts the matching function and weight generation strategy of the multi-dimensional demand-capability matching model.

[0042] Compared with the prior art, the beneficial effects of the present invention are:

[0043] 1. This invention proposes a demand adaptation system based on personalized profiles, and optimizes and improves its architecture, operation steps and processes. The system has the advantages of simple process, low investment and operation costs and low production and working costs.

[0044] 2. This invention proposes a demand matching method based on personalized profiles. Through multi-source heterogeneous data collection and professional algorithm processing, it achieves the accurate construction of dynamic demand profiles for demanders and dynamic capability profiles for service providers. Relying on technologies such as deep cross-modal comparative learning and time-dependent graph node feature incremental update neural networks, the profiles can reflect the physiological cognitive state and interaction preferences of demanders in real time, as well as core information such as the expertise and operating modes of service providers. At the same time, the multi-dimensional demand-capability matching model, combined with cosine similarity calculation and adaptive weight fusion, can accurately calculate the supply-demand fit from multiple dimensions, outputting a visualized matching decision traceability graph and a service provider priority ranking list, realizing dynamic and accurate matching between supply and demand, greatly improving the scientific nature and visualization of matching, and making matching decisions traceable and explainable.

[0045] 3. This invention proposes a demand matching method based on personalized profiles, constructing a closed-loop matching optimization mechanism. Service effect feedback data and matching-related data are used as training samples. Through continuous iteration and optimization of the matching function and weight generation strategy by an intelligent agent, the multi-dimensional demand-capability matching model can dynamically adjust according to actual service effects, continuously improving matching accuracy. Simultaneously, the dynamic capability profile of the service provider supports incremental learning and can be continuously updated with historical service records, allowing the matching system to adapt to the dynamic changes of both supply and demand sides. It possesses strong self-learning and self-optimization capabilities, and in the long run, can continuously improve the efficiency and effectiveness of demand matching, promoting the intelligent and dynamic development of matching services. Attached Figure Description

[0046] Figure 1 This is a schematic diagram of the demand adaptation method based on personalized profiles according to the present invention;

[0047] Figure 2 This is a flowchart illustrating the principle of the demand adaptation method based on personalized profiles in this invention.

[0048] Figure 3 This is a diagram of the system architecture for adapting to the needs of personalized profiles in this invention. Detailed Implementation

[0049] Example 1:

[0050] Please see Figure 1 and Figure 2 The present invention provides an embodiment of a demand adaptation method based on a personalized profile, the method comprising S1 to S5, including the following steps:

[0051] S1: Through a multi-source heterogeneous sensor array and interactive probes, the native multimodal asynchronous data of the demand side is collected in real time to generate an enhanced synchronous sensing time sequence;

[0052] Furthermore, the multi-source heterogeneous sensor array is deployed in the user's interactive terminal and surrounding environment, including physiological sensing devices and environmental perception sensors, which respectively collect physiological sensing time-series streams and environmental physical parameter streams. Interactive probes collect interactive behavior event streams, which together constitute native multimodal asynchronous data. The interactive probes are embedded in the terminal's operation interface and voice interaction module to capture the user's interactive behaviors such as clicks, swipes, and voice commands in real time, and mark the behavior type and trigger timestamp for each behavior. The environmental perception sensors collect physical parameters such as temperature, humidity, light intensity, and environmental noise to generate environmental data frames.

[0053] Furthermore, this embodiment takes patient demand adaptation in online medical consultation as a typical application scenario. When a patient expresses their consultation needs on the intelligent consultation terminal, a multi-source heterogeneous sensor array synchronously collects data: physiological sensing devices collect the patient's heart rate and blood pressure data at a frequency of 1 second / time through the terminal's heart rate sensor and blood pressure detection module, forming a physiological sensing time-series stream; environmental perception sensors collect the light intensity of 350 lx, temperature of 26℃, and noise level of 40 dB in the consultation environment, generating an environmental physical parameter stream at a frequency of 5 seconds / time, where ℃ represents Celsius temperature, lx represents lux, used to characterize the visible light luminous flux received per unit area, reflecting the brightness of the irradiated surface, dB represents decibel; the interactive probe captures the patient's operational behavior on the terminal, such as clicking on the "stomach pain" symptom category, entering the text "abdominal pain lasting 3 days", issuing the voice command "want to find a gastroenterologist", etc., and marks each behavior with information such as [behavior type: click, timestamp: 1698865200], [behavior type: voice, timestamp: 1698865215], forming an interactive behavior event stream.

[0054] Furthermore, a multimodal data temporal alignment and context labeling algorithm based on a unified time reference is used to process the physiological sensing time series, interactive behavior event series, and environmental physical parameter series. First, a unified time reference with a 1-second unit is established. Then, timestamp alignment and linear interpolation are performed on the physiological sensing time series, environmental physical parameter series, and interactive behavior event series. For example, 5-second interval data of the environmental physical parameter series are interpolated to generate 1-second interval environmental data, and discrete timestamp data of the interactive behavior event series are completed to generate continuous time frames. Subsequently, a pre-defined contextual semantic rule base is used to... Time-aligned data frames are assigned a unified context identifier code. In this embodiment, the patient's heart rate during the consultation is 85 beats / min, blood pressure is 120 / 80 mmHg, the operation behavior is symptom selection, and the environment is a quiet indoor environment. Matching the context of "medical consultation - symptom description - normal physiological state" in the context semantic rule base, the context identifier code M001 is assigned. After all data frames with context identifier codes are arranged in chronological order, an enhanced synchronous perception time sequence is generated. This enhanced synchronous perception time sequence contains synchronous data in five dimensions: timestamp, physiological data, interaction behavior, environmental parameters, and context identifier code.

[0055] Furthermore, the pre-built contextual semantic rule base is a set of pre-constructed contextual judgment rules for service scenarios. These rules are developed by domain experts in conjunction with business scenarios. Each rule includes five items: physiological state range, interaction behavior type, environmental parameter range, contextual semantic description, and unique contextual identifier code. The rule base adopts a condition combination matching mode. When a data frame simultaneously meets all the conditions of any rule in terms of physiological range, behavior type, and environmental range, it matches the corresponding context and assigns a contextual identifier code.

[0056] S2: Perform feature extraction and modal interaction analysis on the enhanced synchronous sensing time series to generate a structured dynamic demand profile;

[0057] Furthermore, in this embodiment, the enhanced synchronous perception time-series sequence with M001 context identifier code is input into the cross-modal feature fusion network. First, the visual encoder extracts the visual features of the patient's interface click trajectory and symptom classification selection on the terminal, generating a 128-dimensional visual feature vector. Then, the time-series encoder extracts features from the time-series data of heart rate and blood pressure to capture the changing patterns of physiological state, generating a 128-dimensional time-series feature vector. Finally, the text encoder performs semantic encoding on text information such as "stomach pain", "abdominal pain lasting for 3 days", and "want to see a gastroenterologist", generating a 256-dimensional text feature vector.

[0058] Furthermore, the network calculates the interaction weights between the three feature vectors through a cross-attention mechanism. In this embodiment, the interaction weight between the text feature vector and the visual feature vector is 0.45, the interaction weight between the text feature vector and the temporal feature vector is 0.3, and the interaction weight between the visual feature vector and the temporal feature vector is 0.25. Based on these weights, the three feature vectors are weighted and fused to generate a 512-dimensional fused high-dimensional dynamic feature vector. This feature vector is input into a sparse attention feature decoupler. The decoupler uses a multi-head attention mechanism to calculate the sparse weights of each dimension within the feature vector, resolving the state quantization vector cluster and the interaction preference probability transition matrix. The state quantization vector cluster consists of four 12-dimensional vectors. The set of 8-dimensional numerical vectors represents the patient's physiological state, symptom representation state, cognitive need state, and interaction state, respectively. The interaction preference probability transition matrix is ​​a 10×10 two-dimensional probability matrix that describes the transition probability of different interaction methods for the patient. For example, the transition probability from "text input" to "voice interaction" is 0.1, and the transition probability from "symptom click" to "doctor department selection" is 0.92. This matrix is ​​obtained by the decoder by performing a softmax normalization exponential function operation on the fused high-dimensional dynamic feature vectors. The state quantization vector cluster and the interaction preference probability transition matrix together constitute the patient's structured dynamic need profile, accurately depicting the patient's consultation needs, physiological state, and interaction preferences.

[0059] S3: Retrieve the service provider's historical service records, break them down into the smallest capability units, perform incremental learning, and build and update the service provider's dynamic capability profile.

[0060] Furthermore, in this embodiment, the service provider is a gastroenterologist on an online medical platform. The system retrieves the physician's historical service records, including structured and unstructured data such as consultation cases, treatment plans, patient evaluations, service duration, and interaction methods. Based on a service capability decomposition method using semantic analysis, the system performs semantic and dependency parsing analysis on the historical service records, extracting service actions, operation objects, and result evaluation triples. Then, based on a capability atomization classification dictionary in the medical field, the triples are mapped to the smallest capability unit with a unified semantic identifier.

[0061] Furthermore, taking the historical service records of gastroenterologist A within the platform as an example, the engine extracts multiple triples, such as [Service Action: Diagnosis, Target: Patients with chronic gastritis, Result Evaluation: Diagnostic Rate 98%], [Service Action: Medication Guidance, Target: Helicobacter pylori Infected Individuals, Result Evaluation: Satisfaction 95%], [Service Action: Voice Consultation, Target: Elderly Patients, Result Evaluation: Interaction Fluency 90%], [Service Action: Text Answer, Target: Patients with Abdominal Pain, Result Evaluation: Timely Response Rate 99%]. Based on the medical capability atomized classification dictionary, these triples are mapped to the smallest capability units with unified semantic identifiers, such as D001 (Diagnosis of Gastroenterological Diseases - Chronic Gastritis), M002 (Medication Guidance - Helicobacter pylori), I003 (Interaction Method - Voice Consultation), I004 (Interaction Method - Text Answer), etc., with each smallest capability unit accompanied by an effect evaluation value.

[0062] Furthermore, an incremental learning approach using a graph node feature incremental update neural network with temporal dependence is employed to construct and update the dynamic capability profile of physician A. First, the aforementioned minimum capability units are used as nodes in a graph. Temporal associations between units, such as guidance on diagnosis followed by medication, and semantic co-occurrence relationships, such as abdominal pain diagnosis often being accompanied by textual explanations, are used as edges in the graph. The weights of the edges are determined by the association frequency and effectiveness evaluation, thus constructing an initial capability graph. The initial capability graph is then aggregated and updated using the graph node feature incremental update neural network with temporal dependence, strengthening the feature values ​​of nodes with high effectiveness and high relevance. If a new service record is input, such as physician A adding a service record for "abdominal pain emergency diagnosis," the minimum capability unit D005 (gastroenterology disease diagnosis - abdominal pain emergency) is added using the incremental learning algorithm. The edge weights of related nodes and the node feature vectors of the graph node feature incremental update neural network with temporal dependence are then updated, forming the updated capability graph. The updated capability graph is a domain-pattern dual embedding network, comprising a domain subgraph and a pattern subgraph. The first node of the domain subgraph represents Physician A's expertise domain, such as "diagnosis of chronic gastritis," "treatment of Helicobacter pylori," and "diagnosis of abdominal pain emergency." The edges between the first nodes represent the semantic association strength of the expertise domain; for example, the association strength between "diagnosis of abdominal pain emergency" and "diagnosis of chronic gastritis" is 0.75. The second node of the pattern subgraph represents Physician A's operational habit pattern, such as "voice consultation," "text answer," and "quick diagnosis." The edges between the second nodes represent the state transition probability of the operational habit; for example, the transition probability from "text answer" to "quick diagnosis" is 0.88. The domain subgraph and the pattern subgraph are bridged through minimum capability units; for example, "diagnosis of abdominal pain emergency" and "text answer" are associated through units D005 and I004, with a bridging weight of 0.95. This dual embedding network is Physician A's dynamic capability profile.

[0063] S4: Input the dynamic demand profile and dynamic capability profile into the pre-built multi-dimensional demand-capability matching model, calculate the native matching vector of the preset dimensions, and output the comprehensive suitability score, matching decision traceability diagram and service provider priority ranking list.

[0064] Furthermore, the method for generating the service provider priority ranking list is as follows: sort all service providers' comprehensive suitability scores in descending order, and adjust the constraints based on the real-time load status of the service providers to generate the final service provider priority ranking list.

[0065] Furthermore, the multi-dimensional demand-capability matching model includes a matching degree calculation layer, a dynamic weight fusion layer, and a decision backtracking layer. For medical consultation scenarios, it presets four core matching dimensions: expertise matching, treatment effect matching, interaction method matching, and service efficiency matching. The framework outputs accurate matching results through multi-dimensional calculation and fusion.

[0066] Furthermore, in this embodiment, the dynamic needs profile of the patient and the dynamic ability profiles of Physician A and the remaining three gastroenterologists in the platform are input into the framework: the matching degree calculation layer calculates the cosine similarity values ​​of four preset dimensions to obtain the original matching vector. The patient's core need is "abdominal pain diagnosis, preference for text interaction". In Physician A's dynamic ability profile, "abdominal pain emergency diagnosis" is a specialty and the text answering effect is good. Therefore, Physician A's original matching vector is [Specialty domain matching: 0.92, Treatment effect matching: 0.89, Interaction method matching: 0.95, Service efficiency matching: 0.90]; the original matching vectors of the other three physicians are [0.75, 0.82, 0.60, 0.85], [0.80, 0.86, 0.70, 0.88], and [0.85, 0.90, 0.80, 0.82], respectively.

[0067] Furthermore, the dynamic weight fusion layer is implemented based on a Long Short-Term Memory (LSTM) network. It takes the context identifier M001 corresponding to the enhanced synchronous perception time sequence and the historical weight sequence as input, and combines the contextual features of the medical consultation scenario to output the adaptive weight coefficients for each dimension. In this embodiment, because the patient clearly describes their symptoms and prefers a fixed interaction method, the weight coefficient for interaction method matching is increased, ultimately outputting [Specialty domain matching: 0.3, Treatment effect matching: 0.25, Interaction method matching: 0.35, Service efficiency matching: 0.1]. The original matching vector and the adaptive weight coefficients are weighted and summed to calculate the comprehensive fit score. The comprehensive fit score for Physician A is: 0.92×0.3+0.89×0.25+0.95×0.35+0.90×0.1=0.923, and the scores for the other three physicians are 0.743, 0.791, and 0.836, respectively.

[0068] Furthermore, the decision backtracking layer records the cosine similarity value and adaptive weight coefficient of each dimension, generating a tree-structured matching decision backtracking graph: the root node is the comprehensive fit score of 0.923; the four preset matching dimensions are connected to the root node as child nodes; each child node is configured with two leaf nodes, storing the cosine similarity value and adaptive weight coefficient of that dimension respectively, such as the leaf nodes of the "interaction method matching" child node being [0.95, 0.35]; according to the size of the adaptive weight coefficient, the four child nodes are sorted, where interaction method matching > expertise domain matching > treatment effect matching > service efficiency matching, and then visualized. The larger the adaptive weight coefficient of the child node, the larger the node size displayed in the matching decision backtracking graph, intuitively showing the contribution path of each dimension to the comprehensive fit score. Finally, the framework sorts the four gastroenterologists from highest to lowest overall suitability score and outputs a priority list of service providers: First, Physician A, 0.923; Second, Physician B, 0.836; Third, Physician C, 0.791; Fourth, Physician D, 0.743.

[0069] S5: After completing the service according to the service provider priority ranking list, collect feedback data on the service effect, and use the feedback data, the matching decision traceability graph, and the service provider priority ranking list as training samples to input into the agent, dynamically adjusting the matching function and weight generation strategy in the multi-dimensional demand-capability matching model.

[0070] Furthermore, the feedback data on the service effect is a multi-dimensional quantitative vector, which includes at least the task completion score, the demander satisfaction score, and the service efficiency index.

[0071] Furthermore, the intelligent agent includes an actor network and a critic network, and adaptively adjusts the matching parameters using service feedback as a reward signal. Specifically, the critic network receives the training samples, which at least include service performance feedback data, a matching decision source graph, and a service provider priority ranking list. The critic network evaluates the long-term value function of the current matching strategy based on the training samples. The actor network receives the long-term value function and the matching decision source graph, calculates the strategy gradient, and updates the network parameters of the actor network. Using the updated network parameters of the actor network, the matching function and weight generation strategy in the multi-dimensional demand-ability matching model based on cosine similarity and dynamic weights are dynamically adjusted.

[0072] Furthermore, in this embodiment, the patient selects Physician A based on the service provider priority ranking list to complete the online consultation service. After the service is completed, the platform collects service effect feedback data, including the patient's rating of diagnostic accuracy (9.8 points), interaction fluency (9.9 points), and service satisfaction (9.8 points). At the same time, it records objective data such as service duration (8 minutes) and diagnostic results consistent with those of offline hospitals. This feedback data, Physician A's matching decision traceability graph, and service provider priority ranking list are integrated into training samples and input into the intelligent agent.

[0073] Furthermore, the agent's critic network evaluates the matching effect of the training sample and calculates the reward value of the current matching strategy. In this embodiment, the reward value is +0.95 because the service effect is excellent. The actor network outputs adjustment parameters based on the reward value. By using service feedback as a reward signal, the matching parameters are adaptively adjusted, and the matching function and weight generation strategy in the multi-dimensional demand-ability matching model based on cosine similarity and dynamic weights are dynamically adjusted. When adjusting the matching function, the feature weights in the matching degree calculation layer when calculating cosine similarity are corrected through the parameters output by the actor network, strengthening the weight ratio of features such as "abdominal pain diagnosis" and "text interaction". When adjusting the weight generation strategy, the network connection weights of the LSTM network in the dynamic weight fusion layer are updated through the parameters output by the actor network, so that the adaptive weight coefficients of the interaction method matching dimension are further optimized under the same context identifier code. If a training sample with poor service effect occurs, such as when the patient feedback does not match the doctor's interaction method, the agent will output a negative reward value, and the actor network will adjust the matching function and weight generation strategy in reverse to achieve dynamic iterative optimization of the framework parameters and improve the accuracy of demand matching.

[0074] Furthermore, the method also includes a cold start processing mechanism. When the dynamic demand profile or the dynamic capability profile data is insufficient, a pre-filling algorithm based on collaborative filtering is used to initialize the filling using the average features of the group profiles. The pre-filling algorithm based on collaborative filtering is a user-based collaborative filtering algorithm.

[0075] Furthermore, the method also includes privacy protection processing, which adds noise to the native multimodal asynchronous data and sensitive information in historical service records using differential privacy technology before transmission and computation.

[0076] Example 2:

[0077] In this embodiment, a cross-modal feature fusion network is used for feature extraction and modal interaction analysis of enhanced synchronous sensing time-series sequences. The cross-modal feature fusion network is a three-layer encoder-decoder structure, specifically as follows:

[0078] S2.1: The enhanced synchronous sensing temporal sequence is input into the visual encoder, temporal encoder and text encoder respectively for feature extraction, generating visual feature vector, temporal feature vector and text feature vector;

[0079] Furthermore, the visual encoder is used to extract visual features from the enhanced synchronous perception temporal sequence. This encoder is a convolutional neural network structure, consisting of four convolutional layers, four pooling layers, and two fully connected layers. The network input is the format-normalized visual feature data from the enhanced synchronous perception temporal sequence, with the input data size uniformly set to 224×224 pixels. All convolutional layers of the network use two-dimensional convolution operations. The first convolutional layer is configured with 64 convolutional kernels of size 3×3, with a stride of 1 and uniform padding. After the convolution operation, a ReLU activation function is applied for non-linear transformation, followed by the first pooling layer. The pooling layer uses max pooling with a kernel size of 2×2 and a stride of 2. No padding operation; the second convolutional layer has 128 3×3 convolutional kernels with a stride of 1 and the same padding method. After being transformed by the ReLU activation function, it is connected to the second pooling layer. The max pooling kernel of the pooling layer is 2×2 with a stride of 2 and no padding. The third convolutional layer has 256 3×3 convolutional kernels with a stride of 1 and the same padding method. After being transformed by the ReLU activation function, it is connected to the third pooling layer. The pooling layer parameters are consistent with the first two pooling layers. The fourth convolutional layer has 512 3×3 convolutional kernels with a stride of 1 and the same padding method. After being transformed by the ReLU activation function, it is connected to the fourth pooling layer. The pooling layer parameters are still 2×2 max pooling kernel, stride of 2, and no padding. After the above convolution and pooling operations, the visual feature data is transformed into a multi-dimensional feature matrix, which is then sequentially connected to two fully connected layers for feature dimensionality reduction and integration. The number of neurons in the first fully connected layer is set to 2048. After processing by the ReLU activation function, it is connected to the second fully connected layer. The number of neurons in the second fully connected layer is set to 512, and no activation function is applied. After the operation of the second fully connected layer, a 512-dimensional visual feature vector is finally generated. The ReLU activation function is existing technology in this field and is not an inventive solution of this application, so it will not be described in detail here.

[0080] Furthermore, the temporal encoder is used to extract temporal features from the enhanced synchronous sensing temporal sequence. The encoder is a long short-term memory network structure, which consists of two unidirectional long short-term memory layers and one fully connected layer. The input of the network is the normalized temporal feature data in the enhanced synchronous sensing temporal sequence. The input time step is set to 60, and the input feature dimension is set to 64. That is, the network receives temporal data for 60 time steps each time, and each time step contains 64-dimensional feature information. The first long short-term memory (LSM) layer has 256 hidden layer neurons. The forget gate, input gate, and output gate all use the Sigmoid activation function, while the candidate cell states use the Tanh activation function. A dropout layer with a dropout rate of 0.2 is included within the layer to prevent overfitting. The output of the first LSM layer is a 256-dimensional hidden layer state sequence. This hidden layer state sequence serves as the input to the second LSM layer, which has 128 hidden layer neurons. The activation function configuration is consistent with the first LSM layer, and a dropout layer with a dropout rate of 0.2 is also included within the layer. The second LSM layer only extracts the hidden layer state from the last time step as its output, resulting in a 128-dimensional output. The output of the second long short-term memory layer is connected to a fully connected layer for feature integration. The number of neurons in this fully connected layer is set to 512, and no activation function is connected. After the operation of the fully connected layer, a 512-dimensional temporal feature vector is finally generated, realizing the extraction and vectorization representation of the temporal pattern of the temporal features. The Sigmoid activation function and the Tanh activation function are existing technologies in this field and are not the inventive solutions of this application, and will not be described in detail here.

[0081] Furthermore, the text encoder is used to extract text-like features from the enhanced synchronous sensing temporal sequence. The encoder is a Transformer-based encoding structure, which consists of 6 Transformer encoding blocks and one fully connected layer. The input of the network is the text feature data from the enhanced synchronous sensing temporal sequence after word segmentation and vectorization. The length of the input sequence is set to 128, and the dimension of the input word vector is set to 128. That is, the network receives a text sequence of length 128 each time, and each word corresponds to a 128-dimensional word vector. Each Transformer encoding block contains two core modules: a multi-head self-attention mechanism and a feedforward neural network. The multi-head self-attention mechanism has 8 heads, meaning it splits the 128-dimensional word vector into eight 16-dimensional sub-vectors for self-attention calculation. The scaling factor for self-attention calculation is set to 4, and the attention weights are calculated using dot product attention. A masking operation is used to block information about future positions. The output of the multi-head self-attention mechanism is processed by residual connections and layer normalization before being input into the feedforward neural network. The feedforward neural network contains two linear transformation layers. The first linear transformation increases the feature dimension from 128 to 512 and applies the ReLU activation function. The second linear transformation restores the feature dimension from 512 to 128 without applying the activation function. The output of the feedforward neural network is also processed by residual connections and layer normalization before being used as the output of the Transformer encoding block. Six Transformer coding blocks are concatenated sequentially, with the output of the previous Transformer coding block serving as the input to the next. After feature extraction by the six Transformer coding blocks, the text feature data is transformed into a 128×128 feature matrix. This matrix undergoes global average pooling to integrate the sequence-dimensional features, generating a 128-dimensional feature vector. This vector is then fed into a fully connected layer with 512 neurons and no activation function. After processing by the fully connected layer, a 512-dimensional text feature vector is finally generated, achieving semantic information extraction and vectorized representation of text features.

[0082] Furthermore, during the training phase of the visual encoder, temporal encoder, and text encoder, all three encoders employ batch gradient descent for parameter optimization. The batch size is uniformly set to 32, and the initial learning rate is set to 0.001. A learning rate decay strategy is adopted, with the learning rate decreasing to 0.9 every 10 training epochs. The total number of training epochs is set to 100 epochs, and the mean squared error loss function is used for all three. The ratio of the training set to the validation set is set to 8:2. When the loss value on the validation set no longer decreases for 10 consecutive epochs, the encoder training is stopped, and the optimal parameters of each encoder are saved. In the actual feature extraction process, all three encoders load the optimal parameters after training and perform forward propagation operations on the corresponding feature data in the enhanced synchronous perception temporal sequence, generating fixed-dimensional visual feature vectors, temporal feature vectors, and text feature vectors, respectively. The dimensions of the three feature vectors are consistent. The batch gradient descent method is existing technology in this field and is not an inventive solution of this application, so it will not be described in detail here.

[0083] S2.2: Calculate the interaction weights between the visual feature vector, temporal feature vector, and text feature vector through a cross-attention mechanism, and perform weighted fusion of the visual feature vector, temporal feature vector, and text feature vector based on the interaction weights to generate a fused high-dimensional dynamic feature vector.

[0084] Furthermore, the specific steps of S2.2 include:

[0085] (1) Perform layer normalization on the visual feature vector, temporal feature vector and text feature vector respectively. Set the mean of the normalization operation to zero and the variance to 1. This operation eliminates the differences in numerical distribution of the three types of feature vectors and ensures the numerical stability of cross attention calculation.

[0086] (2) After layer normalization, the visual feature vector is used as the query feature, and the temporal feature vector and text feature vector are concatenated as the key feature and value feature. At the same time, the temporal feature vector is used as the query feature, and the visual feature vector and text feature vector are concatenated as the key feature and value feature. Then, the text feature vector is used as the query feature, and the visual feature vector and temporal feature vector are concatenated as the key feature and value feature, thus completing the combination construction of three sets of query features, key features and value features.

[0087] (3) Construct a cross-attention calculation module. The cross-attention calculation module is a multi-head cross-attention structure. It includes four steps: attention head configuration, attention weight calculation, attention weight normalization, and attention feature aggregation. The number of attention heads in the cross-attention calculation module is set to 8. The 512-dimensional query features, key features, and value features are divided into 8 independent sub-feature groups according to each 64-dimensional segment. Each attention head processes a set of sub-features to realize multi-dimensional parallel attention calculation. In the operation of a single attention head, a learnable query transformation matrix is ​​configured for the query sub-feature group with a dimension of 64 rows and 640 columns. A learnable key transformation matrix is ​​configured for the key sub-feature group with a dimension of 640 rows and 64 columns. A learnable value transformation matrix is ​​configured for the value sub-feature group with a dimension of 640 rows and 64 columns. The query, key, and value sub-feature groups are linearly transformed by the three transformation matrices to obtain dimension-matched query transformation features, key transformation features, and value transformation features.

[0088] (4) After completing the linear transformation of the features, calculate the attention weight of a single attention head. Perform matrix multiplication on the transpose of the query transformation feature and the key transformation feature to obtain the initial attention score matrix. Scale the initial attention score matrix with the scaling factor set to the square root of 64, i.e., 8. This operation avoids the problem of excessively large attention score values ​​due to high feature dimensions. After scaling, perform normalization exponential function operation on the initial attention score matrix to map all values ​​in the matrix to the interval between zero and one, and the sum of all values ​​is 1, thus obtaining the normalized attention weight matrix of a single attention head. Subsequently, the normalized attention weight matrix and the value transformation feature are multiplied together to obtain the attention output feature of a single attention head. After the eight attention heads complete parallel processing, eight 64-dimensional attention output features are obtained. These eight attention output features are concatenated in dimensional order to obtain a 512-dimensional multi-head cross-attention fusion feature. A learnable output transformation matrix is ​​then configured for this fusion feature, with the matrix dimension set to 512 rows and 512 columns. The concatenated features are linearly transformed using the output transformation matrix to obtain an attention feature matrix of pairwise interactions between the three types of feature vectors. This matrix is ​​a 512-row, 512-column square matrix, and each value in the matrix represents the interaction strength between the corresponding feature dimensions.

[0089] (5) Solving and normalizing the interaction weights based on the attention feature matrix. First, the attention feature matrix is ​​subjected to global average pooling by row, and the 512-row, 512-column matrix is ​​transformed into a 512-dimensional one-dimensional vector. This vector is the initial interaction weight vector of the three types of feature vectors. The normalization exponential function operation is performed on the initial interaction weight vector again to obtain the global interaction weights of the visual feature vector, temporal feature vector and text feature vector respectively. The sum of the three weights is 1, which respectively represents the importance of the three types of feature vectors in modal interaction. Thus, the calculation of the interaction weights of the three types of feature vectors by the cross attention mechanism is completed.

[0090] (6) Then, based on the interaction weights obtained from the solution, the visual feature vector, temporal feature vector and text feature vector are weighted and fused. Each value of the visual feature vector is multiplied with the corresponding visual interaction weight to obtain the weighted visual feature vector. Using the same operation method, the temporal feature vector and text feature vector are multiplied with their respective corresponding interaction weights element by element to obtain the weighted temporal feature vector and the weighted text feature vector. The three types of weighted feature vectors are still 512-dimensional.

[0091] (7) The three weighted feature vectors are added element by element to obtain a 512-dimensional weighted fusion basic feature vector. To further improve the representation ability of the features, two fully connected layers are built to perform high-dimensional feature mapping on the weighted fusion basic feature vector. The number of neurons in the first fully connected layer is set to 1024, and the activation function is a linear rectified function. After the operation of the first fully connected layer, a 1024-dimensional feature vector is obtained. The number of neurons in the second fully connected layer is set to 2048, and no activation function is set. After the linear transformation of the second fully connected layer, a 2048-dimensional fusion feature vector is obtained. This vector is the fusion high-dimensional dynamic feature vector.

[0092] Furthermore, the training process of the cross-attention module needs to be synchronized with the overall training of the cross-modal feature fusion network. The batch size during training is set to 32, the initial learning rate is set to 0.0001, and a cosine annealing learning rate decay strategy is adopted. The learning rate decay period is set to 20 training epochs, and the minimum learning rate is set to 0.000001. The cross-modal contrastive loss function is adopted, treating different modal features of the same sample as positive sample pairs and the same modal features of different samples as negative sample pairs. The module's parameters are optimized by maximizing the similarity of positive sample pairs and minimizing the similarity of negative sample pairs. The ratio of the training set to the validation set is set to 8:2, and the total number of training epochs is set to 100. A gradient clipping strategy is adopted during training, and the maximum norm of the gradient is set to 5 to prevent gradient explosion. When the contrastive loss value on the validation set no longer decreases for 10 consecutive epochs, training is stopped and the optimal parameters of the cross-attention module are saved. In the actual feature fusion process, these optimal parameters are loaded to complete all operations of interaction weight calculation and feature fusion.

[0093] S2.3: The fused high-dimensional dynamic feature vector is input into a sparse attention feature decoupling unit. The sparse attention feature decoupling unit utilizes a multi-head attention mechanism to calculate the sparse weights between the various feature dimensions within the fused high-dimensional dynamic feature vector. Based on these sparse weights, the fused high-dimensional dynamic feature vector is parsed into a state quantization vector cluster and an interaction preference probability transition matrix. The state quantization vector cluster is a set of multiple numerical vectors representing real-time physiological and cognitive states. The interaction preference probability transition matrix is ​​a two-dimensional probability matrix used to describe the dynamic transition probabilities between different interaction methods. The two-dimensional probability matrix is ​​obtained by performing a normalized exponential function operation on the fused high-dimensional dynamic feature vector using a decoder.

[0094] Furthermore, the specific steps in S2.3 include:

[0095] (1) Perform input preprocessing operation on the fused high-dimensional dynamic feature vector. The fused high-dimensional dynamic feature vector is a continuous numerical vector with a fixed dimension of 2048. First, perform layer normalization on the vector. The mean of the normalization operation is set to zero and the variance is set to one to eliminate the difference in numerical distribution of different dimensions within the feature vector. After completing the layer normalization, perform linear projection on the vector. Configure a learnable linear projection matrix with a dimension of 2048×2048. Perform linear transformation on the high-dimensional dynamic feature vector through this matrix to obtain a projection feature vector with consistent dimensions. This projection feature vector serves as the formal input data of the sparse attention feature decoupling device.

[0096] (2) The core structure of the sparse attention feature decoupler is constructed. The sparse attention feature decoupler is based on the multi-head attention mechanism and includes four core links: attention head configuration, sparse attention mask construction, sparse weight calculation, and sparse weight normalization. The number of attention heads in the decoupler is set to 16. The 2048-dimensional projected feature vector is divided into 16 independent sub-feature groups according to each 128-dimensional segment. Each attention head processes a set of sub-features to realize parallel sparse attention calculation of multiple dimensions. Each attention head is configured with independent learnable parameters to ensure the independence of sparse weight calculation of each dimension segment. In the operation of a single attention head, learnable query transformation matrix, key transformation matrix, and value transformation matrix are configured for the 128-dimensional sub-feature groups respectively. The dimensions of the three matrices are all set to 128×128. The query, key, and value sub-features are linearly transformed by the three transformation matrices respectively to obtain the dimension-matched query transformation features, key transformation features, and value transformation features.

[0097] (3) After completing the linear transformation of the sub-features, the sparse attention mask is constructed and the sparse weights of individual attention heads are calculated. First, the key transformation feature is subjected to global average pooling to convert the 128-dimensional vector into a scalar value. The scalar value is compared with a preset sparse threshold, which is set to 0.05. If the scalar value is less than the sparse threshold, the attention weight at the corresponding position is reset to zero. If the scalar value is greater than or equal to the sparse threshold, the attention weight at the corresponding position is retained. In this way, a 128×128 sparse attention mask matrix is ​​constructed to mask the weak correlation between feature dimensions and strengthen the attention weight ratio of strong correlation. Then, the transpose of the query transformation feature and the key transformation feature are multiplied by matrix to obtain the initial sparse attention score matrix. The initial sparse attention score matrix is ​​scaled, and the scaling factor is set to 0.05. To avoid the problem of excessively large attention scores due to high feature dimensions, the scaled initial sparse attention score matrix is ​​multiplied element-wise with the sparse attention mask matrix to obtain the masked sparse attention score matrix. The masked sparse attention score matrix is ​​then subjected to a normalized exponential function operation to map all values ​​in the matrix to the interval between zero and one, with the sum of the values ​​in each row being 1, to obtain the normalized sparse weight matrix of a single attention head. Each value in this matrix represents the sparse correlation strength between the corresponding feature dimensions.

[0098] (4) After the 16 attention heads complete the parallel operation, 16 normalized sparse weight matrices of 128 rows and 128 columns are obtained. The 16 sparse weight matrices are spliced ​​in dimensional order to obtain a 2048×2048 global sparse weight matrix, which is the global sparse weight between each feature dimension inside the fused high-dimensional dynamic feature vector. The global sparse weight matrix is ​​subjected to global max pooling in the row dimension to transform the 2048×2048 matrix into a 2048-dimensional one-dimensional sparse weight vector. Each value in the one-dimensional sparse weight vector corresponds to the global sparse weight value of a feature dimension in the fused high-dimensional dynamic feature vector. Then, the one-dimensional sparse weight vector is normalized to map the values ​​to the interval between zero and one, thus completing the final calculation of the sparse weight.

[0099] (5) Based on the global sparse weights, the fused high-dimensional dynamic feature vector is analyzed. First, the state quantization vector cluster is extracted. The 2048-dimensional sparse weight vector is numerically sorted, and the top 1024 feature dimensions with the highest weight values ​​are selected according to the order of weight values ​​from high to low. These high-weight feature dimensions are extracted from the fused high-dimensional dynamic feature vector to form a 1024-dimensional high-weight feature vector. The feature dimensions of the high-weight feature vector are divided according to the feature attributes representing physiological and cognitive states. It is divided into four 256-dimensional sub-feature vectors. Each sub-feature vector is a numerical vector representing any real-time state of the demander. The four sub-feature vectors together form a state quantification vector cluster. The first 256-dimensional sub-feature vector represents the real-time physiological basic state of the demander, the second 256-dimensional sub-feature vector represents the real-time physiological stress state of the demander, the third 256-dimensional sub-feature vector represents the real-time cognitive demand state of the demander, and the fourth 256-dimensional sub-feature vector represents the real-time cognitive interaction state of the demander. This achieves a quantitative representation of the real-time physiological and cognitive state of the demander.

[0100] (6) Generate the interaction preference probability transition matrix. The generation of the interaction preference probability transition matrix is ​​achieved by the dedicated decoder configured by the decoupler. The decoder is a three-layer fully connected network structure. First, it extracts the remaining 1024-dimensional low-weight feature dimension after the above screening from the fused high-dimensional dynamic feature vector to form a 1024-dimensional low-weight feature vector. This low-weight feature vector is used as the input data of the decoder to characterize the correlation characteristics between different interaction methods of the demand side. The number of neurons in the first fully connected layer of the decoder is set to 512. The activation function is a linear rectified function. The 1024-dimensional low-weight feature vector is subjected to feature dimensionality reduction and nonlinear transformation to obtain a 512-dimensional feature vector. The number of neurons in the second fully connected layer is set to 256. The activation function is also a linear rectified function. The feature vector is subjected to dimensionality reduction and nonlinear transformation to obtain a 256-dimensional feature vector. The number of neurons in the third fully connected layer is set to 625. No activation function is set. The 256-dimensional feature vector is subjected to linear transformation to obtain a 625-dimensional feature vector. This dimension corresponds to a 25×25 two-dimensional matrix dimension, which meets the dimension requirements of the interaction preference probability transition matrix.

[0101] (7) The 625-dimensional feature vector output by the third layer of the decoder is reshaped into a 25×25 two-dimensional original matrix, which is the original matrix of the interaction preference probability transfer matrix. Each value in the original matrix represents the initial transfer probability between two different interaction methods of the demand side. The normalization exponential function operation is performed on the two-dimensional original matrix, and the values ​​in each row of the matrix are normalized respectively. All values ​​in each row are mapped to the interval between zero and one, and the sum of the values ​​in each row is 1. The normalization calibration of the values ​​is completed. The 25×25 two-dimensional matrix obtained after this operation is the final interaction preference probability transfer matrix. The rows and columns of the matrix correspond to different interaction methods of the demand side. Each value in the matrix represents the probability value of the demand side transferring from one interaction method to another, thereby realizing the quantitative representation of the dynamic transfer probability between different interaction methods of the demand side.

[0102] Furthermore, the sparse attention feature decoupler and the cross-modal feature fusion network are jointly trained. The batch size during training is set to 32, the initial learning rate is set to 0.0001, and a step-wise learning rate decay strategy is adopted, with the learning rate decaying to 0.85 after every 15 training epochs, for a total of 100 training epochs. The loss function is a joint loss function, obtained by weighted summation of state quantization loss and matrix generation loss. The state quantization loss uses the mean squared error loss function to constrain the numerical representation accuracy of the state quantization vector cluster, and the matrix generation loss uses the cross-entropy loss function to constrain the probability distribution of the interaction preference probability transition matrix. The weighting coefficients of both losses are set to 0.5. The ratio of training set to validation set is set to 8:2. A gradient pruning strategy is adopted during training, with the maximum norm of the gradient set set to 5 to prevent gradient explosion. When the joint loss value on the validation set no longer decreases for 10 consecutive epochs, training is stopped and the optimal parameters of the sparse attention feature decoupler, including all transformation matrix parameters of the multi-head attention mechanism and all fully connected layer parameters of the decoder, are saved.

[0103] S2.4: The state quantization vector cluster and the interaction preference probability transition matrix together constitute a structured dynamic demand profile.

[0104] The incremental learning process, which constructs and updates the dynamic capability profile of the service provider, includes:

[0105] S3.1: Using the minimum capability unit as the node of the graph, and the temporal association and semantic co-occurrence relationship between the minimum capability units as the edge of the graph, an initial capability graph is constructed;

[0106] It should be noted that the smallest capability units obtained by decomposing service capabilities based on semantic analysis all have a unified semantic identifier and corresponding attribute information. The attribute information includes core content such as service action type, scope of operation object, result evaluation value, and service record timestamp.

[0107] Furthermore, the specific steps of S3.1 include:

[0108] (1) Perform semantic uniqueness verification on all minimum capability units, remove duplicate minimum capability units, and then classify and organize the remaining minimum capability units according to the service domain of the service provider to ensure that the minimum capability units in the same domain are grouped into the same set. After the classification is completed, feature vector representation is performed on each minimum capability unit. Based on the preset capability unit feature encoding rules, the discrete attributes such as semantic identifier, service action type, and operation object category of the minimum capability unit are transformed into numerical features. The continuous attributes such as result evaluation value are normalized and then integrated into the feature vector. Finally, a fixed-dimensional feature vector is generated for each minimum capability unit. This feature vector serves as the core feature of the graph node and is used to represent the capability attributes of the node. Thus, the basic construction of the initial capability graph node is completed. Each unique minimum capability unit corresponds to an independent node in the initial capability graph. The node identifier is the unified semantic identifier of the minimum capability unit, and the node feature is the fixed-dimensional feature vector generated above.

[0109] Furthermore, the capability unit feature encoding rules are pre-constructed based on the capability atomization classification dictionary, service business attributes, and evaluation index system, including discrete attribute encoding rules and continuous attribute normalization rules; among them, discrete attributes are mapped to numerical vectors using one-hot encoding, and continuous attributes are normalized to the [0,1] interval using maximum-minimum normalization. The feature dimension is fixed at 512 dimensions based on the total number of capability attributes. One-hot encoding and maximum-minimum normalization are existing technologies in this field and are not inventive solutions of this application, so they will not be elaborated here.

[0110] (2) Extract the service record timestamps corresponding to all minimum capability units. Sort the minimum capability units under the same service scenario and the same service object in chronological order according to the timestamps from earliest to latest to form a sequence of minimum capability units arranged in chronological order. For the sequence of minimum capability units, set a time-series association window. The length of the window is set according to the complexity of the service process of the service provider. Take the minimum capability units that appear consecutively as the analysis object and judge the direct time-series association between the units. If two minimum capability units appear consecutively in the time series and belong to the preceding and following links of the same service process, it is determined that there is a direct time-series association between the two. Construct a time-series association edge for the nodes corresponding to the two units. The direction of the edge is from the unit that appears first. The node pointing to the subsequent unit node reflects the temporal sequence of the service process. Then, the weight of the temporal association edge is calculated based on the frequency of consecutive occurrences of two minimum capability units in historical service records. Higher frequencies indicate stronger temporal associations. This frequency is divided by the total number of records for that type of service process by the service provider to obtain a normalized temporal association weight value. This value is assigned to the corresponding temporal association edge, completing the construction and weight assignment of a single temporal association edge. If multiple minimum capability units form a continuous temporal association chain in the time series, weighted temporal association edges are constructed sequentially for adjacent unit nodes, forming a chain-like temporal association edge structure, thus completing the mining of temporal association relationships between all minimum capability units and the construction of temporal association edges.

[0111] (3) Semantic vectorization is performed on the semantic identifiers and attribute information of all minimum capability units. Based on the pre-trained semantic coding model, the semantic content of each minimum capability unit is transformed into a high-dimensional semantic vector. The semantic coding model adopts a Transformer-based text coding structure with a total of 6 coding layers, 8 multi-head attention heads, a fixed word embedding dimension of 128, a input sequence length of 128, and a fixed output semantic vector dimension of 128. The cosine similarity between the high-dimensional semantic vectors of any two minimum capability units is calculated to determine the degree of semantic association between them. The semantic similarity threshold is set to 0.7 according to the semantic closeness requirement of the service domain. If the cosine similarity value of the semantic vectors of two units is greater than the semantic similarity threshold, it is determined that there is a semantic association between them. On this basis, the semantic association between the two minimum capability units in the service domain is statistically analyzed. The frequency of co-occurrence among all historical service records of the same entity is used to determine the co-occurrence frequency. Co-occurrence refers to two entities appearing in the same service record or in consecutive service records of the same service object. The weight of the semantic co-occurrence edge is calculated by combining the semantic association degree and the co-occurrence frequency. The semantic co-occurrence weight value is obtained by multiplying the semantic vector cosine similarity value by the normalized co-occurrence frequency. If two minimum capability units have both a semantic association and a non-zero co-occurrence frequency, an undirected semantic co-occurrence edge is constructed for their corresponding nodes, and the calculated semantic co-occurrence weight value is assigned to this edge, thus completing the construction and weight assignment of a single semantic co-occurrence edge. For minimum capability units with a semantic vector cosine similarity value lower than the semantic similarity threshold or no co-occurrence frequency, no semantic co-occurrence edge is constructed, thus completing the mining of semantic co-occurrence relationships and the construction of semantic co-occurrence edges among all minimum capability units.

[0112] (4) Construct the topology of the initial capability graph. First, create the node set and edge set of the graph based on the construction rules of the graph data structure. Include all nodes representing the smallest capability unit in the node set and assign a unique node ID to each node. The node ID corresponds one-to-one with the unified semantic identifier of the smallest capability unit. At the same time, store the feature vector and attribute information of the node in association with the node ID to ensure the integrity of the node information. Include all completed temporal association edges and semantic co-occurrence edges in the edge set and assign a unique edge ID to each edge. Record the edge type, i.e., temporal association edge or semantic co-occurrence edge, the connecting nodes of the edge, i.e., the starting node ID and the ending node ID, the weight value of the edge, and the association information corresponding to the edge, such as temporal association edge and semantic co-occurrence edge. The time window for sequential association, the cosine similarity value of semantic co-occurrence, and the co-occurrence frequency are used to ensure the integrity of edge information. Then, the topological structure of the graph is built based on the node set and the edge set. According to the connection relationship of the edges, the nodes are connected to each other through the corresponding edges. For directed temporally associated edges, the connection is strictly in the direction from the first-appearing unit node to the second-appearing unit node. For undirected semantic co-occurrence edges, a bidirectional connection is made between the two associated nodes to form the graph topological structure. The built graph topological structure is optimized as a whole. The graph is divided into subgraphs according to the service provider's expertise. The nodes and edges corresponding to the smallest capability unit in the same expertise constitute a domain subgraph, which facilitates graph analysis and feature extraction. After optimization, the initial capability graph is stored according to the storage specifications of graph data and managed using a graph database. Node information, edge information, and topology information are stored in the corresponding database tables. At the same time, an association index between nodes and edges is established to ensure that data can be quickly retrieved when performing operations such as node feature aggregation and edge weight update on the initial capability graph. This completes the entire construction process of the initial capability graph with the smallest capability unit as the node and the temporal association and semantic co-occurrence relationship between units as the edge.

[0113] S3.2: The initial capability map is aggregated and updated using a time-dependent graph node feature incremental update neural network.

[0114] It should be noted that each node in the initial capability map corresponds to a minimum capability unit, and each node has an initial feature vector, the dimension of which is uniformly set to 512.

[0115] Furthermore, the specific steps in S3.2 include:

[0116] (1) Perform layer normalization on the initial feature vectors of all nodes to ensure the numerical stability of the convolution operation; then construct the adjacency matrix of the initial capability graph. The dimension of the adjacency matrix is ​​consistent with the number of nodes in the graph. Each element in the matrix represents the connection relationship between two corresponding nodes. If there is a temporally related edge or a semantically co-occurring edge between two nodes, the element value at the corresponding position is assigned the weight value of the edge between the two nodes. If there is no connecting edge between two nodes, the element value at the corresponding position is assigned zero. At the same time, add an identity matrix to the adjacency matrix to realize the self-connection aggregation of the node's own features and avoid the node's own features being weakened during the aggregation process. Perform two matrix multiplication operations on the adjacency matrix and the inverse square root of the node degree matrix to obtain the symmetric normalized adjacency matrix, which is used to represent the weighted neighborhood relationship between nodes. The node degree matrix is ​​a diagonal matrix with the same dimension as the adjacency matrix. Each element on the diagonal of the matrix is ​​the node degree of the corresponding node. The elements at the off-diagonal positions are all zero. The node degree is equal to the sum of the weights of the edges connected to a single node.

[0117] (2) Construct a time-dependent graph node feature incremental update neural network architecture for node feature aggregation and updating. The time-dependent graph node feature incremental update neural network architecture is a multi-layer graph convolutional layer structure, which includes three graph convolutional layers and one fully connected layer. Each graph convolutional layer is configured with independent learnable convolutional kernel parameters and bias terms. The fully connected layer is used to complete the final mapping and updating of the aggregated features. The convolutional kernel dimension of the first graph convolutional layer is set to 512×512, and the bias term dimension is set to 512. The input of this layer is the normalized initial feature vector of the node and the symmetric normalized adjacency matrix. The core implementation is the aggregation of first-order neighbor node features, that is, each node only aggregates the features of directly connected neighbor nodes. The convolutional kernel dimension of the second graph convolutional layer is set to 512×256, and the bias term dimension is set to 256. The input of this layer is the output of the first graph convolutional layer. The core of the architecture is the aggregation of second-order neighborhood node features, using the features and symmetrically normalized adjacency matrix. This aggregation further aggregates the features of neighboring nodes based on the first-order neighborhood features of each node. The third graph convolutional layer has a kernel dimension of 256×128 and a bias term dimension of 128. Its input consists of the output features of the second graph convolutional layer and the symmetrically normalized adjacency matrix. This aggregation achieves the aggregation of third-order neighborhood node features, enabling multi-range correlation and fusion of node features within the graph. Furthermore, each graph convolutional layer is followed by a linear rectified activation function to achieve non-linear transformation of features, enhancing the network's feature representation capability. The fully connected layer has 512 neurons and no additional activation function; it only performs linear mapping of features, restoring the aggregated low-dimensional features to the same dimension as the initial feature vector, ensuring dimensionality consistency after node feature updates.

[0118] (3) After completing the construction of the neural network architecture for incremental update of graph node features with time-dependent graph, the first-order neighborhood node feature aggregation operation of the first graph convolutional layer is performed, including: performing matrix multiplication operation between the symmetric normalized adjacency matrix and the normalized initial feature vector of the node to obtain the first-order neighborhood feature fusion matrix of all nodes. Each row vector of this matrix represents the fusion feature of all first-order neighborhood node features of the corresponding node after weighting by edge weight. Then, the first-order neighborhood feature fusion matrix is ​​multiplied by the convolution kernel of the first graph convolutional layer to complete the linear transformation of the features. Subsequently, the bias term of this layer is added for numerical calibration to obtain the linear feature vector after first-order aggregation. The vector is input into the linear rectified activation function for nonlinear transformation to suppress invalid values ​​in the features and enhance the representation ability of the features. Finally, a 512-dimensional first-order aggregated feature vector is obtained. This vector integrates the node's own features and the weighted features of the first-order neighborhood nodes to achieve the initial aggregation of node features.

[0119] (4) Perform the second-order neighborhood node feature aggregation operation of the second-layer graph convolutional layer, including: performing matrix multiplication operation between the symmetric normalized adjacency matrix and the first-order aggregated feature vector output by the first-layer graph convolutional layer to obtain the second-order neighborhood feature fusion matrix of all nodes. Each row vector of this matrix represents the fusion feature of all second-order neighborhood node features of the corresponding node after weighting by edge weight. Perform matrix multiplication operation between the second-order neighborhood feature fusion matrix and the convolution kernel of the second-layer graph convolutional layer to complete the feature reduction and linear transformation. After adding the bias term of this layer, input the linear rectified activation function for nonlinear transformation, and finally obtain a 256-dimensional second-order aggregated feature vector. This vector further integrates the weighted features of the second-order neighborhood nodes on the basis of the first-order aggregation, realizing the deep aggregation of node features.

[0120] (5) Perform the third-order neighborhood node feature aggregation operation of the third-layer graph convolutional layer, including: performing matrix multiplication operation between the symmetric normalized adjacency matrix and the second-order aggregated feature vector output by the second-layer graph convolutional layer to obtain the third-order neighborhood feature fusion matrix of all nodes. Each row vector of this matrix represents the fusion feature of all third-order neighborhood node features of the corresponding node after weighting by edge weight. Perform matrix multiplication operation between the third-order neighborhood feature fusion matrix and the convolution kernel of the third-layer graph convolutional layer to complete the further dimensionality reduction and linear transformation of the features. After adding the bias term of this layer, input the linear rectified activation function for nonlinear transformation, and finally obtain a 128-dimensional third-order aggregated feature vector. This third-order aggregated feature vector integrates the node's own features and the weighted features of the first to third-order neighborhood nodes, fully explores the multi-level correlation between nodes in the initial capability graph, and realizes the comprehensive aggregation of node features.

[0121] (6) Input the 128-dimensional third-order aggregated feature vector output by the third graph convolutional layer into the fully connected layer. Through the linear transformation of the fully connected layer, the 128-dimensional aggregated feature is mapped back to 512-dimensional, resulting in a 512-dimensional updated feature vector. This updated feature vector is the final feature vector after the time-dependent graph node feature incremental update neural network aggregates and updates the node features of the initial capability graph. It retains the core capability features of the node itself and integrates the associated capability features of the multi-level neighborhood nodes in the initial capability graph. At the same time, it combines edge weights to achieve feature weighting with different association strengths, which more accurately represents the features and value of each minimum capability unit in the overall capability system of the service provider.

[0122] Furthermore, the time-dependent graph node feature incremental update neural network for aggregating and updating node features in the initial capability map requires independent training and parameter optimization. The training process aims to optimize the service provider's capability matching performance. A 512-dimensional capability feature label vector is pre-constructed based on the service provider's historical service records and a capability atomization classification dictionary. This vector maintains the same dimension as the initial node feature vector and is generated by domain experts based on the service provider's actual expertise level, operational proficiency, and historical service performance. This vector serves as a supervisory reference signal for the node feature aggregation results. The training batch size is set to 32, the initial learning rate is set to 0.0005, and a cosine annealing learning rate decay strategy is adopted. The learning rate decay period is set to 20 training epochs, and the minimum learning rate is set to 0.000001 to ensure the efficiency of parameter optimization during training. The loss function is the mean squared error loss function, using the mean squared error between the node-aggregated and updated feature vector and the pre-defined capability feature label vector as the optimization metric to constrain the accuracy and representational ability of node feature aggregation. The training set to validation set ratio was set to 8:2, and the total number of training epochs was set to 80. A gradient pruning strategy was employed during training, with the maximum norm of the gradient set to 5 to effectively prevent gradient explosion and ensure network training stability. When the loss value on the validation set no longer decreased for 8 consecutive epochs, network training was stopped, and the optimal parameters of each layer of the neural network, including convolutional kernel parameters, bias term parameters, and weight parameters of fully connected layers, were updated incrementally using graph node features with time-dependent dependencies.

[0123] S3.3: When a new historical service record is input, the minimum capability unit, the edge weights between nodes, and the node feature vectors in the neural network with time-dependent graph node features are updated through an incremental learning algorithm to form an updated capability graph. The updated capability graph is used as a dynamic capability profile of the service provider.

[0124] Furthermore, the specific steps of S3.3 include:

[0125] (1) Newly input historical service records first undergo text cleaning and structuring to remove invalid characters and redundant information, while retaining core information such as service actions, operation objects, result evaluations, service timestamps, and service scenarios. The processed records are then input into a service capability decomposition method based on semantic analysis. This method extracts triples of service actions, operation objects, and result evaluations through semantic analysis and dependency parsing. Subsequently, based on a preset capability atomization classification dictionary, the extracted triples are mapped to the smallest capability unit with a unified semantic identifier, and a result evaluation normalization number is assigned to this unit. The evaluation results are normalized to map the original evaluation values ​​to the range of zero to one, with a mapping coefficient of 0.01. After the mapping is completed, the newly extracted minimum capability unit is compared with the minimum capability unit in the original initial capability map to determine whether it is a newly added minimum capability unit or a new service record of the original minimum capability unit. If the semantic identifier does not appear in the original initial capability map, it is determined to be a newly added minimum capability unit. If the semantic identifier already exists, it is determined to be a new service record of the original minimum capability unit. Different update processes are performed for the two types of determination results.

[0126] Furthermore, the preset capability atomization classification dictionary is a standardized semantic dictionary pre-built for the corresponding service domain. It contains four core entries: capability type, service action, operation object, and evaluation dimension. Each entry corresponds to a unique unified semantic identifier, which is used to map unstructured service record triples into standardized minimum capability units.

[0127] (2) Execute the update of the minimum capability unit, including the construction of nodes for the newly added minimum capability unit and the attribute update of the original minimum capability unit, including: For the newly added minimum capability unit, construct the corresponding graph node for it. First, assign a unique node ID and unified semantic identifier to the node. Then, based on the capability unit feature encoding rules, convert its discrete attributes such as service action type, operation object category, and service scenario label into numerical features. Integrate the continuous attributes such as the result evaluation normalized value directly into the feature vector. Finally, generate an initial feature vector of 512 dimensions. This vector serves as the initial feature of the newly added node and is associated with and stored with the node ID, semantic identifier, and attribute information. For the newly added service record that is determined to be the original minimum capability unit, do not construct a new node. Only update the attribute information of the original node. Take the weighted average of the result evaluation normalized value of the newly added service record and the result evaluation mean of the original node. Set the weighting coefficient to the ratio of the number of original records to the total number of records and the ratio of the number of newly added records to the total number of records. Update the result evaluation mean attribute of the node and supplement the newly added service timestamp and service scenario label.

[0128] (3) Incremental update of edge weights between nodes, including weight update of temporal association edges and semantic co-occurrence edges. At the same time, construct association edges between new nodes and existing nodes and assign weights. The edge weight update is based on the temporal information and semantic co-occurrence information in the new service records. First, for the temporal association edges of the existing nodes, count the frequency of consecutive new occurrences between the original minimum capability units in the new service records. Add the new frequency to the original frequency to get the total frequency. Then divide the total frequency by the total number of records of this type of service process to get the updated temporal association weight value, which replaces the original edge weight value. The weight value of the temporal association edge is kept in the range of zero to one. For the semantic co-occurrence edges of the existing nodes, first count the frequency of new co-occurrence of the original minimum capability units in the new service records. Add it to the original co-occurrence frequency to get the total co-occurrence frequency. Then recalculate the two units. The cosine similarity of semantic vectors is used to multiply the normalized total frequency of occurrence by the cosine similarity to obtain the updated semantic co-occurrence weight value, which replaces the original edge weight value. The weight value of the semantic co-occurrence edge also remains in the range of zero to one. For new nodes, the temporal association and semantic co-occurrence relationship between them and the original nodes are first explored. If the new minimum capability unit and the original minimum capability unit have a temporal order in the new service record and belong to the same service process, a temporal association edge is constructed for the two, and the initial weight value is calculated according to the frequency of addition. If the cosine similarity of the semantic vectors of the two is greater than the preset cosine similarity threshold of 0.7 and they both appear in the new service record, a semantic co-occurrence edge is constructed for the two, and the initial weight value is calculated according to the co-occurrence frequency and semantic cosine similarity. All new edges are labeled with edge type and association basis to complete the overall incremental update of edge weights.

[0129] (4) Perform incremental learning and updating of node feature vectors in the graph node feature incremental update neural network with temporal dependence. This process adopts an incremental learning algorithm based on local neighborhood aggregation, which only updates the feature vectors of newly added nodes and associated nodes, and does not update the features of unassociated nodes. The graph node feature incremental update neural network with temporal dependence still adopts the structure of three convolutional layers and one fully connected layer. The parameters of the convolutional layers and the fully connected layers use the optimal parameters after training. Local aggregation of features is only performed on newly added information. Specifically, it includes: first, performing layer normalization on the 512-dimensional initial feature vector of the newly added nodes, and then constructing the local adjacency matrix between the newly added nodes and associated nodes. The dimensions of the local adjacency matrix are... The degree is set according to the number of associated nodes, and the matrix elements are the edge weights between corresponding nodes. An identity matrix is ​​added to the local adjacency matrix to achieve self-connection, and then symmetric normalization is performed. The normalized initial feature vector and the local adjacency matrix are input into the first graph convolutional layer. This layer has a kernel dimension of 512 rows and 512 columns, and a bias term dimension of 512. After matrix multiplication and linear rectified activation function transformation, the first-order neighborhood aggregation feature vector of the newly added nodes is obtained. The first-order neighborhood aggregation feature vector and the local adjacency matrix are input into the second graph convolutional layer. This layer has a kernel dimension of 512×256, and a bias term dimension of 256. After operation and activation, the second-order neighborhood aggregation feature vector is obtained. The second-order neighborhood aggregation feature vector and the local adjacency matrix are then input into the third graph convolutional layer, which has a kernel dimension of 256×128 and a bias term dimension of 128. After operation and activation, the third-order neighborhood aggregation feature vector is obtained. Finally, the third-order neighborhood aggregation feature vector is input into the fully connected layer, which has 512 neurons. After linear transformation, a 512-dimensional updated feature vector is obtained, which serves as the final feature vector of the new node. For nodes in the original nodes that are related to the new node, the same local neighborhood aggregation method is used to integrate the features of the new node into its neighborhood aggregation process, recalculate and update its feature vector. The feature vectors of the original nodes that are not related remain unchanged. This completes the incremental update of the feature vector of the neural network node with temporal dependence graph node feature incremental update. The batch size of the entire incremental calculation process is set to 16, and the learning rate is set to 0.0001 to ensure the stability of feature update.

[0130] (5) Integrate all updated information to reconstruct the capability graph and form a dynamic capability profile of the service provider, including: First, update the node set of the capability graph, add new nodes to the node set, and update the attribute information and feature vectors of the original nodes to ensure the latest and completeness of all node information in the node set; then update the edge set of the capability graph, add new temporal association edges and semantic co-occurrence edges to the edge set, replace the updated weight values ​​of the original edges, remove invalid edges with weight values ​​of zero, and update the association information of all edges, such as frequency and semantic similarity; based on the updated node set and edge set, rebuild the topology of the graph, connect the new nodes to the original nodes according to the connection relationship and direction of the edges, keep the topological connection relationship of the original nodes unchanged, and only update the association. The local topology of the newly added nodes is adjusted. After the topology is built, the graph is divided into subgraphs. Following the domain-pattern dual embedding network structure, the updated nodes are assigned to the domain subgraph and the pattern subgraph respectively. The first node of the domain subgraph represents the specialty domain, and the weights of the edges between nodes are updated to the latest semantic association strength. The second node of the pattern subgraph represents the operation habit pattern, and the weights of the edges between nodes are updated to the latest state transition probability. At the same time, the nodes of the domain subgraph and the pattern subgraph are bridged through the minimum capability unit, and the bridging weights are updated to the latest association strength. After the subgraph is divided and bridged, the validity of the entire graph is verified. The uniqueness of the node ID and edge ID, as well as the rationality of the association relationship between the node and the edge, are checked, and invalid node and edge connections are removed.

[0131] (6) After verification, the updated capability graph is stored in accordance with the storage specifications of the graph database. The indexes of related data such as node feature vectors, edge weights, and topology are updated to ensure the efficiency of graph calling and analysis. The updated capability graph is the latest dynamic capability profile of the service provider. It retains the core features and relationships of the original capability graph and incorporates capability information from new historical service records, realizing the dynamic incremental update of the service provider's capability profile.

[0132] Furthermore, the algorithm training and parameter configuration of the entire incremental learning update process are designed around the local update requirements of the neural network for incrementally updating graph node features with time-dependent dependencies. The training batch size of the incremental learning algorithm is set to 32, the initial learning rate is set to 0.0005, and a step-wise learning rate decay strategy is adopted, with the learning rate decaying to 0.9 after every 10 training epochs. The total number of training epochs is set to 60 epochs. The loss function is the mean squared error loss function, with the mean squared error between the updated node feature vector and the actual capability feature label vector as the optimization index. The ratio of the training set to the validation set is set to 8:2. A gradient pruning strategy is adopted during training, and the maximum norm of the gradient is set to 5 to prevent gradient explosion. When the loss value on the validation set no longer decreases for 6 consecutive epochs, the training of the incremental learning algorithm is stopped, and the local update parameters of the algorithm are saved. In the actual graph update process, these parameters are loaded to complete the efficient incremental update of the minimum capability unit, edge weights, and node feature vectors, ensuring that the dynamic capability profile can keep up with the actual service capability changes of the service provider in real time.

[0133] The dynamic capability profile is specifically a domain-pattern dual embedding network, which includes a domain subgraph and a pattern subgraph.

[0134] S3.3.1: The domain subgraph contains multiple first nodes, each first node represents a specialty domain, and the first nodes are connected by edges, which represent the semantic association strength between specialty domains;

[0135] S3.3.2: The pattern subgraph contains multiple second nodes, each second node representing an operation habit pattern. The second nodes are connected by edges, and the edges between the second nodes represent the state transition probabilities between operation habit patterns.

[0136] S3.3.3: The first node of the domain subgraph and the second node of the pattern subgraph are associated and bridged through the smallest capability unit obtained by decomposing from the historical service records of the same service provider.

[0137] Furthermore, the update triggering condition of the domain-pattern dual embedding network is as follows: when the number of new minimum capability units detected exceeds a preset threshold, or when the total number of services completed by the service provider reaches a preset period, the incremental learning process of the time-dependent graph node feature incremental update neural network is automatically triggered. The preset threshold is set according to the update frequency and data accumulation stability requirements of the service domain. For medical services and education tutoring scenarios, it is set to 10 by default. When the number of newly added minimum capability units reaches 10 or more, it is determined that the data accumulation meets the update requirements. The preset period is set according to the service frequency and capability iteration rules of the service provider. For regular service scenarios, it is set to 30 by default. When the service provider has completed 30 services in total, it is determined that the capability status needs to be updated. Incremental learning can be started as long as either condition is met.

[0138] The multi-dimensional demand-capability matching model includes a matching degree calculation layer, a dynamic weight fusion layer, and a decision backtracking layer.

[0139] The matching degree calculation layer receives the dynamic demand profile and the dynamic capability profile, and calculates the cosine similarity value on multiple preset dimensions to obtain the original matching vector. Each component of the original matching vector corresponds to a cosine similarity value of a preset dimension. The cosine similarity calculation formula is the prior art in this field and is not an inventive solution of this application, so it will not be described in detail here.

[0140] The dynamic weight fusion layer is implemented based on a long short-term memory network. The long short-term memory network takes the context identifier code corresponding to the enhanced synchronous perception time sequence and the historical weight sequence as input, and outputs the adaptive weight coefficients of each preset dimension. The original matching vector and the adaptive weight coefficients are weighted and summed to generate a comprehensive fit score.

[0141] Furthermore, the native matching vector is an eight-dimensional vector composed of cosine similarity values ​​calculated by the matching degree calculation layer on eight preset matching dimensions. Each value represents the native matching degree of the corresponding dimension, and the value ranges from zero to one.

[0142] Furthermore, the dynamic weight fusion layer, based on a long short-term memory network, implements adaptive weight coefficient output and completes the comprehensive fit score calculation. This requires first standardizing and preprocessing the input context identifier code and historical weight sequence, then building a long short-term memory network architecture to meet the weight generation requirements. Leveraging the network's temporal feature learning ability, it mines the correlation between context and weights, outputting the adaptive weight coefficients for each preset matching dimension. Finally, it performs a weighted summation operation between these coefficients and the original matching vector to generate the comprehensive fit score. The specific process includes:

[0143] (1) Perform standardized encoding preprocessing on the input data. The preprocessed input data includes the context identifier code and historical weight sequence corresponding to the enhanced synchronous sensing time sequence. First, perform one-hot encoding on the context identifier code. The context identifier code is a preset discrete semantic identifier. Its encoding dimension is consistent with the total number of preset context identifier codes in the system. In the one-hot vector generated after encoding, only the position corresponding to the context identifier code has a value of one, and the values ​​of the other positions are all zero. After encoding, a fixed-dimensional context encoding vector is obtained. The dimension of this vector is set to 128 dimensions to realize the numerical representation of discrete context identifiers. Then, preprocess the historical weight sequence. The historical weight sequence is time-series data composed of adaptive weight coefficients of each preset matching dimension output from multiple consecutive time steps in the preceding sequence. The number of preset matching dimensions is set to 8, and the time step length of the historical weight sequence is fixed at 60. That is, the sequence contains 8-dimensional adaptive weight coefficient data of 60 time steps. First, all weight coefficients in the historical weight sequence are normalized to eliminate the distribution difference of weight values ​​in different time steps. Then, the normalized time-series data is dimensionally reshaped to form a 60×8 historical weight time-series matrix. After completion, the matrix is ​​flattened into a 480-dimensional one-dimensional historical weight vector to achieve dimensional regularization of the time-series weight data.

[0144] (2) After completing the preprocessing of the context encoding vector and the historical weight vector, the 128-dimensional context encoding vector and the 480-dimensional historical weight vector are concatenated to obtain a 608-dimensional fusion input vector, which serves as the formal input data of the long short-term memory network. At the same time, the fusion input vector is subjected to layer normalization to further ensure the stability of the numerical distribution.

[0145] (3) Construct a long short-term memory network architecture for outputting adaptive weight coefficients. The long short-term memory network architecture consists of two long short-term memory layers and two fully connected layers connected in series. Each long short-term memory layer is configured with independent learnable parameters, including the weight matrix and bias terms for the input gate, forget gate, output gate, and cell state. The number of hidden layer neurons in the first long short-term memory layer is set to 512. The input gate, forget gate, and output gate all use the Sigmoid activation function, and the candidate values ​​for the cell state use the Tanh activation function. A dropout layer with a dropout rate of 0.2 is set within the layer to prevent overfitting during network training. The output of this layer... The input is a 608-dimensional fused input vector after normalization, and the output is a 512-dimensional hidden layer state sequence, realizing the initial extraction of temporal correlation features of the input data. The number of hidden layer neurons in the second long short-term memory layer is set to 256, and the activation function and Dropout layer configuration are consistent with the first layer, with the dropout rate also set to 0.2. The input of this layer is the 512-dimensional hidden layer state sequence output by the first layer, and only the hidden layer state of the last time step is extracted as the output to obtain a 256-dimensional core temporal feature vector. This vector integrates the temporal correlation features of contextual features and historical weights to achieve in-depth mining of core correlation patterns.

[0146] (4) Two fully connected layers are connected in series after two long short-term memory layers. The number of neurons in the first fully connected layer is set to 128. The activation function is a linear rectified function. The 256-dimensional core temporal feature vector is reduced in dimension and nonlinearly transformed to obtain a 128-dimensional feature vector, which further enhances the representation ability of the features. The number of neurons in the second fully connected layer is consistent with the number of preset matching dimensions, which is set to 8. No additional activation function is configured. Only linear feature mapping is performed to map the 128-dimensional feature vector to an 8-dimensional adaptive weight coefficient vector. This adaptive weight coefficient vector is the initial weight coefficient of each preset matching dimension.

[0147] (5) Perform element-wise multiplication of the cosine similarity value of each dimension in the 8-dimensional original matching vector with the adaptive weight coefficient of the corresponding dimension in the adaptive weight coefficient vector to obtain the weighted matching value of the 8 dimensions. Then sum the weighted matching values ​​of all dimensions. The summation result is the comprehensive fit score. The value range of the comprehensive fit score is from zero to one. The higher the value, the higher the matching degree between the dynamic demand profile and the dynamic capability profile.

[0148] Furthermore, this Long Short-Term Memory (LSTM) network requires independent training and parameter optimization. The training process aims to improve the prediction accuracy of the weight coefficients. The batch size is set to 32, the initial learning rate is set to 0.0001, and a cosine annealing learning rate decay strategy is employed. The decay period is set to 20 training epochs, and the minimum learning rate is set to 0.000001 to ensure the efficiency and stability of parameter optimization. The mean squared error loss function is used, with a training set to validation set ratio of 8:2. The total number of training epochs is 100, and a gradient pruning strategy is used during training. The maximum norm of the gradient is set to 5 to effectively prevent gradient explosion. When the loss value on the validation set no longer decreases for 10 consecutive epochs, network training is stopped, and the optimal parameters of each layer of the LSM network are saved, including the gated weight matrix of the LSM layer, the bias term, and the weight parameters of the fully connected layer. During the actual weight coefficient generation process, the network loads these optimal parameters to complete the forward propagation operation and outputs accurate adaptive weight coefficients.

[0149] The decision backtracking layer records the cosine similarity value of each preset dimension in the matching degree calculation layer, as well as the corresponding adaptive weight coefficient in the dynamic weight fusion layer; a tree-structured matching decision backtracking graph is generated based on the recorded cosine similarity value and adaptive weight coefficient to visualize the contribution path of each dimension to the comprehensive fit score.

[0150] The process of generating the matching decision tracing graph includes:

[0151] S4.1: The comprehensive fit score is used as the root node;

[0152] S4.2: Connect each preset matching dimension as a child node to the root node;

[0153] S4.3: Configure two leaf nodes for each child node. The first leaf node stores the cosine similarity value corresponding to the preset matching dimension, and the second leaf node stores the adaptive weight coefficient corresponding to the preset matching dimension.

[0154] S4.4: Based on the magnitude of the adaptive weight coefficients, sort and visualize all child nodes connected to the root node to generate a tree-structured matching decision tracing graph.

[0155] The agent is implemented using a proximal policy optimization algorithm, including an actor network and a critic network;

[0156] Adjusting the matching function in the multi-dimensional demand-ability matching model based on cosine similarity and dynamic weights means dynamically modifying the feature weights of the dynamic demand profile and dynamic ability profile used in the matching degree calculation layer when calculating cosine similarity through the parameters output by the actor network.

[0157] Furthermore, the actor network adopts a three-layer fully connected network structure to output the correction parameters of the feature weights in the matching degree calculation layer. The input of the actor network is the feature vector after fusing the service effect feedback data, the matching decision traceability graph, and the service provider priority ranking list. The output is a correction parameter matrix with the same dimension as the feature weight matrix. The parameter configuration of each layer of the network is designed around the accuracy requirements of feature weight correction. The input feature vector of the actor network is set to 1024 dimensions. The number of neurons in the first fully connected layer is set to 512, and the activation function is a linear rectified function, which completes the first dimensionality reduction and nonlinear transformation of the input features, resulting in a 512-dimensional feature vector. The number of neurons in the second fully connected layer is set to 256, and the activation function is also a linear rectified function, which completes the second dimensionality reduction and nonlinear transformation, resulting in a 256-dimensional feature vector. The number of neurons in the third fully connected layer is consistent with the total number of elements in the feature weight matrix, which is set to 2048. No activation function is configured, and only linear feature mapping is performed, outputting a 2048-dimensional one-dimensional vector. This vector dimension is reshaped into a 64×32 correction parameter matrix. This matrix is ​​the feature weight correction parameter output by the actor network, which perfectly matches the dimension of the initial feature weight matrix in the matching degree calculation layer, realizing the accurate correction of the weight of each feature dimension.

[0158] Furthermore, the actor network is trained using the proximal policy optimization algorithm. The batch size is set to 32, the initial learning rate is set to 0.0003, and a step-wise learning rate decay strategy is adopted, with the learning rate decreasing to 0.9 every 15 training epochs, for a total of 80 epochs. The loss function is the advantage function of the proximal policy optimization algorithm, and a gradient pruning strategy is used during training, with the maximum norm of the gradient set to 5. When the advantage function loss value on the validation set no longer decreases for 10 consecutive epochs, training is stopped, and the optimal parameters of each layer of the actor network, including the weight matrix and bias terms of the fully connected layers, are saved. The proximal policy optimization algorithm is prior art in this field and is not an inventive solution of this application; therefore, it will not be described in detail here.

[0159] Furthermore, the actor network output parameters are fused with the initial feature weight matrix to obtain the corrected feature weight matrix. This process uses a linear fusion method, performing element-wise operations on the corrected parameter matrix and the initial feature weight matrix to ensure the smoothness and rationality of the weight correction. First, the 64×32 corrected parameter matrix output by the actor network is normalized using a max-min normalization method, mapping all values ​​in the matrix to the range of -0.2 to 0.2. This range represents the correction range of the feature weights, avoiding abrupt changes in the cosine similarity calculation results due to excessively large correction amplitudes, and ensuring the stability of the matching function. The normalized corrected parameter matrix is ​​then added element-wise with the initialized and normalized initial feature weight matrix to obtain the preliminary corrected feature weight matrix. Next, a normalized exponential function operation is performed on this preliminary corrected feature weight matrix, remapping all values ​​in the matrix to the range of zero to one, ensuring that the sum of all element values ​​in the matrix is ​​1, resulting in the final corrected feature weight matrix. This matrix is ​​the latest feature weight matrix used for cosine similarity calculation in the matching degree calculation layer, realizing dynamic correction of the feature weights.

[0160] Furthermore, the modified feature weight matrix is ​​applied to the cosine similarity calculation of the matching degree layer. In the specific cosine similarity calculation process, 64×32 feature vectors of the dynamic demand profile and dynamic capability profile are first extracted under each preset matching dimension. The modified 64×32 feature weight matrix is ​​then multiplied element-wise with the feature vectors of the demand side and the service side, respectively, to obtain the weighted feature vectors of both sides. The weighted feature vectors of the demand side and the service side are then standardized to eliminate differences in numerical magnitude. Finally, according to the cosine similarity calculation formula, the cosine value of the angle between the two weighted standardized feature vectors is calculated. This value is the cosine of the angle between the two weighted standardized feature vectors. The native matching similarity value under the pre-defined matching dimension should be preset. The similarity values ​​of all preset dimensions constitute the native matching vector. In the entire cosine similarity calculation process, the corrected feature weight matrix participates in the calculation as the core parameter. The correction parameter output by the actor network directly adjusts the operation logic of the matching function by changing the feature weight matrix. When the service effect feedback data shows that the matching of any feature dimension has a greater impact on the service effect, the actor network will output the corresponding correction parameter to increase the weight value of that feature dimension, so that the matching function focuses more on the matching calculation of that dimension. Conversely, it will decrease the weight value of the corresponding feature dimension, so as to realize the dynamic iterative optimization of the matching function based on service feedback.

[0161] Furthermore, during the dynamic adjustment of the matching function, a feature weight update cycle is set. Every thirty sets of service performance feedback data, matching decision tracing graphs, and service provider priority ranking list training samples are collected, triggering a parameter output and feature weight correction operation for the actor network. This ensures the matching function can adapt to changes in service scenarios and client matching needs in a timely manner. Simultaneously, the validity of each corrected feature weight matrix is ​​verified, checking if the matrix dimensions match, if element values ​​are within the range of zero to one, and if the sum of all element values ​​equals 1. If the verification fails, the corrected parameters are discarded, and the previous feature weight matrix is ​​used. If the verification passes, it is applied to cosine similarity calculation, ensuring the accuracy and effectiveness of the matching function operation.

[0162] Adjusting the weight generation strategy in the multi-dimensional demand-ability matching model based on cosine similarity and dynamic weights refers to updating the network connection weights of the long short-term memory network in the dynamic weight fusion layer using the parameters output by the actor network.

[0163] Furthermore, when performing structured extraction and standardized preprocessing on the original connection weights of the Long Short-Term Memory (LSTM) network in the dynamic weight fusion layer, the LSM network in the dynamic weight fusion layer adopts an architecture of two LSM layers plus two fully connected layers. The core related part of the weight generation strategy is the two LSM layers, whose connection weights include the weight matrix and bias terms of the input gate, forget gate, output gate, and cell state candidate values. The weights of the fully connected layers do not participate in the adjustment of the weight generation strategy in this case; only the connection weights of the two LSM layers are extracted and updated. The first long short-term memory (LSM) layer has 512 hidden neurons and an input feature dimension of 608. Its input gate, forget gate, and output gate each have a 608×512 weight matrix and a 512-dimensional bias term. The cell state candidate value also has a 608×512 weight matrix and a 512-dimensional bias term. The total number of weight parameters for this layer is 1,247,232, or (608×512+512)×4=1,247,232. Here, 4 refers to the three gating structures and one candidate cell state in the LSM layer, i.e., the input gate, forget gate, and output gate. The second long short-term memory layer has 256 hidden neurons and an input feature dimension of 512. Its input gate, forget gate, and output gate are each configured with a 512×256 weight matrix and a 256-dimensional bias term. The cell state candidate value is configured with a 512×256 weight matrix and a 256-dimensional bias term. The total number of weight parameters in this layer is 525312. The total number of connection weight parameters between the two long short-term memory layers is 1772544, which is completely consistent with the dimension of the update parameter vector output by the actor network.

[0164] Furthermore, the connection weights of the extracted two long short-term memory layers are structurally decomposed. According to the order of input gate, forget gate, output gate, and cell state candidate value, they are decomposed into corresponding weight matrices and bias terms. Then, all the decomposed weight matrices and bias terms are flattened into one-dimensional vectors in dimensional order to obtain the original one-dimensional vector of connection weights with dimensions of 1,772,544. Subsequently, layer normalization is performed on this vector to eliminate the differences in numerical distribution of weights of different gates and different layers, so as to ensure the numerical stability of subsequent parameter fusion operations.

[0165] Furthermore, when performing the fusion operation between the updated parameters output by the actor network and the original connection weights of the Long Short-Term Memory network to obtain the updated connection weights, a linear weighted fusion method is adopted to balance the adjustment effect of the updated parameters with the learning pattern of the original weights, ensuring the smoothness of the updated weights and avoiding abrupt changes in the network operation. First, the 1,772,544-dimensional updated parameter vector output by the actor network is subjected to min-max normalization, mapping all values ​​in the vector to the range of -0.1 to 0.1, to prevent the Long Short-Term Memory network from losing its original feature learning ability due to excessively large update amplitude. The normalized updated parameter vector is added element-wise to the preprocessed original one-dimensional connection weight vector to obtain the initial updated one-dimensional connection weight vector. This initial updated vector is then subjected to a layer normalization process, with the mean set to zero and the variance set to one, to eliminate the numerical distribution shift caused by the addition operation, resulting in the final updated one-dimensional connection weight vector. This final updated one-dimensional connection weight vector is then reshaped in the reverse order of the original structured decomposition, sequentially restoring the bias terms and weight matrices of the second long short-term memory layer, the cell state candidate values ​​of the first long short-term memory layer, the output gate, the forget gate, and the input gate. This yields the corresponding updated weight matrices and bias terms for each part. All reshaped weight matrices and... The bias term dimensions remain consistent with the original dimensions, completing the structured update of the Long Short-Term Memory (LSTM) network connection weights. The updated connection weights are then loaded into the LTM network of the dynamic weight fusion layer, completing the dynamic adjustment of weight replacement and weight generation strategies. Simultaneously, the validity of the updated network is verified to ensure its normal forward propagation operation. The reshaped weight matrix and bias terms are then used to replace the original corresponding connection weights one by one according to the hierarchy and gating structure of the LTM network. The input gate, forget gate, output gate, and cell state candidate values ​​of the first LTM layer are all loaded with the updated weight matrix and bias terms, and the same applies to the second LTM layer. The connection weights of the fully connected layers remain unchanged, completing the weight update of the entire LTM network.

[0166] The updated Long Short-Term Memory (LSTM) network learns the association features between the context identifier code and the historical weight sequence based on the new connection weights. Its weight generation strategy adapts to service performance feedback data. When service performance feedback data shows that a specific weight allocation strategy in any context better matches the matching requirements, the actor network outputs corresponding updated parameters, adjusting the connection weights of the LTM network to more accurately output adaptive weight coefficients in that context. Conversely, it adjusts the network connection weights to weaken ineffective weight allocation patterns, achieving dynamic iterative optimization of the weight generation strategy based on service feedback. A weight generation strategy update cycle is also set: every thirty sets of service performance feedback data, matching decision tracing graphs, and service provider priority ranking list training samples are collected, triggering an update operation on the actor network's parameter output and the LTM network's connection weights. This ensures that the weight generation strategy can adapt to changes in service scenarios and the matching requirements of the demand side in a timely manner.

[0167] Example 3:

[0168] Please see Figure 3 Another embodiment of the present invention provides: a demand adaptation system based on personalized profiles, comprising:

[0169] The multi-source data acquisition module 10, through a multi-source heterogeneous sensor array and interactive probe, collects the physiological sensing time-series stream, interactive behavior event stream and environmental physical parameter stream of the demand side in real time, forming native multimodal asynchronous data. After being timestamped and labeled by a multimodal data time-series alignment and context labeling algorithm based on a unified time reference, an enhanced synchronous sensing time-series sequence is generated.

[0170] The demand profile generation module 20 inputs the enhanced synchronous perception time sequence into the pre-trained cross-modal feature fusion network, and generates a structured dynamic demand profile consisting of a state quantization vector cluster and an interaction preference probability transition matrix through multimodal feature extraction, cross attention fusion, and sparse attention decoupling.

[0171] The capability profile construction module 30 retrieves the historical service records of the service provider, decomposes them into the smallest capability units with semantic labels through a service capability decomposition method based on semantic analysis, and then performs incremental learning through a graph node feature incremental update neural network with temporal dependence to construct and update the dynamic capability profile of the service provider, and updates it in real time according to the new historical service records.

[0172] The multi-objective matching optimization module 40 inputs the dynamic demand profile and dynamic capability profile into the multi-dimensional demand-capability matching model based on cosine similarity and dynamic weights, calculates the original matching vector in the preset dimensions, obtains the comprehensive suitability score through dynamic weight fusion, generates a matching decision traceability graph, and outputs a service provider priority sorting list according to the comprehensive suitability score.

[0173] The iterative optimization module 50 collects service effect feedback data after the demand side and the service provider complete the matching service. The feedback data, the matching decision traceability graph, and the service provider priority ranking list are used as training samples and input into the agent. By using service feedback as a reward signal to adaptively adjust the matching parameters, the matching function and weight generation strategy of the multi-dimensional demand-capability matching model based on cosine similarity and dynamic weights are dynamically adjusted to achieve self-optimization of the adaptation system.

[0174] Furthermore, the multi-source data acquisition module 10 includes a data acquisition unit and a context fusion unit; the data acquisition unit consists of physiological sensors, interactive probes, and environmental sensors, supporting real-time acquisition of multi-frequency and multi-type asynchronous data; the context fusion unit has a built-in unified time series benchmark generator and context semantic rule library, which completes the time alignment of data and the assignment of context identification codes, and outputs an enhanced synchronous sensing time series sequence.

[0175] For example, taking the matching of students with dedicated tutors on an online education platform as an example, when a student initiates a math tutoring request on the platform, the data collection unit collects the student's physiological data, behavioral data, and environmental data through the learning terminal. Among them, physiological data includes concentration and fatigue level, behavioral data includes clicks on the question bank, marking of wrong questions, and answering speed, and environmental data includes lighting and noise in the learning environment. The context fusion unit identifies that the student is in a scenario of tackling a difficult math geometry problem, with the identifier code L005, concentration level 0.65, fatigue level 0.70, repeated clicking on geometry wrong questions, and environmental noise less than or equal to 20 decibels, and generates the corresponding enhanced synchronous perception time sequence.

[0176] Furthermore, the cross-modal feature fusion network of the demand profile generation module 20 adopts a three-encoder structure of visual, temporal, and textual components. The cross-attention mechanism supports dynamic weight fusion of multiple feature vectors, and the sparse attention feature decoupler can adaptively adjust the sparse weights of feature dimensions according to different demand scenarios, ensuring that the parsed state quantization vector clusters and interaction preference probability transition matrices accurately represent the real-time state and core needs of the demander. In this example, the generated dynamic demand profile of the student is as follows: weak geometry knowledge (0.88), low patience for problem-solving (0.72), preference for step-by-step explanations with graphics and text (0.90), aversion to live audio explanations (0.15), a need for tutors with expertise in teaching geometry problem-solving techniques, and a slow pace of explanation.

[0177] Furthermore, the service capability decomposition method based on semantic analysis in the capability profile construction module 30 incorporates an atomized capability classification dictionary for the online education industry. This allows for the precise extraction of the tutor's knowledge point explanation-teaching object-teaching effect triple, mapping it to the smallest capability units such as geometry problem-solving skills K006, step-by-step text and image explanation T003, and slow-paced teaching M008. The time-dependent graph node feature incremental update neural network updates the teacher's capability map in real time through incremental learning, ensuring that the dynamic capability profile remains consistent with the teacher's latest teaching level and style. In this example, after processing, Teacher A's dynamic capability profile core is a geometry problem-solving skills teaching proficiency of 0.92, a step-by-step text and image explanation matching degree of 0.95, and a slow-paced teaching matching degree of 0.88; Teacher B's dynamic capability profile core is an algebra explanation proficiency of 0.90 and a voice live-stream explanation matching degree of 0.93; and Teacher C's dynamic capability profile core is a geometry problem-solving skills proficiency of 0.85, a text and image explanation matching degree of 0.80, and a fast-paced teaching matching degree of 0.90.

[0178] Furthermore, the multi-objective matching optimization module 40, based on cosine similarity and dynamic weights, pre-sets four dimensions for knowledge point expertise matching, teaching method matching, teaching rhythm matching, and teaching effect matching. The LSTM of the dynamic weight fusion layer outputs adaptive weight coefficients for the mathematical geometry problem-solving scenario L005: knowledge point expertise 0.40, teaching method 0.30, teaching rhythm 0.20, and teaching effect 0.10. After weighted calculation, Teacher A's comprehensive suitability score is 0.912, Teacher C's comprehensive suitability score is 0.825, and Teacher B's comprehensive suitability score is 0.380. The framework outputs a priority list of Teacher A-Teacher C-Teacher B and generates a tree-like matching decision tracing graph to visually display the contribution of each dimension to the score.

[0179] Furthermore, the agent in the iterative optimization module 50 adopts a proximal strategy optimization algorithm. If, after the student and teacher A complete tutoring, the student reports an improvement of 0.7 in their mastery of geometric knowledge and a satisfaction score of 9.8 for the step-by-step explanation with illustrations, the agent calculates a positive reward value of +0.90, and then adjusts the framework parameters: increasing the feature weights of "proficiency in geometric problem-solving skills" and "matching degree of step-by-step explanation with illustrations," and strengthening the weight ratio of the teaching method matching dimension in the "mathematical geometry problem-solving scenario." If the matching service effect is not good, a negative reward value is generated to optimize the parameters in reverse, thereby realizing the self-iterative upgrade of the system.

[0180] This invention discloses a demand matching method and system based on personalized profiles. Through precise collection and fusion of multimodal data, deep feature analysis of cross-modal networks, construction of capability profiles using temporal graph networks, multi-objective matching within a differentiable framework, and iterative optimization via reinforcement learning, it achieves personalized and precise matching between demanders and service providers. This effectively solves the problems of static matching, single-dimensionality, and insufficient precision in traditional matching methods. It has broad application value in fields such as psychological counseling, online education, and personalized service recommendations. It can flexibly adjust preset matching dimensions, atomized capability dictionaries, and contextual semantic rule bases according to the demand scenarios of different industries, improving matching efficiency and service effectiveness, and achieving dynamic and intelligent matching of needs and capabilities.

[0181] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments under the guidance of the present invention without departing from the spirit and scope of the present invention. All of these variations are within the protection scope of the present invention.

Claims

1. A demand adaptation method based on personalized profiles, characterized in that, include: By using a multi-source heterogeneous sensor array and interactive probes, native multimodal asynchronous data from the demand side is collected in real time to generate an enhanced synchronous sensing time sequence. Feature extraction and modal interaction analysis are performed on enhanced synchronous sensing time-series sequences to generate structured dynamic demand profiles; Retrieve the service provider's historical service records, break them down into the smallest capability units, and perform incremental learning to build and update the service provider's dynamic capability profile. Input the dynamic demand profile and dynamic capability profile into the pre-built multi-dimensional demand-capability matching model, calculate the native matching vector of the preset dimensions, and output the comprehensive suitability score, the matching decision traceability diagram and the service provider priority ranking list. After completing the service according to the service provider priority ranking list, feedback data on the service effect is collected. The feedback data, the matching decision traceability graph, and the service provider priority ranking list are used as training samples to input the agent, and the matching function and weight generation strategy in the multi-dimensional demand-capability matching model are dynamically adjusted.

2. The demand adaptation method based on personalized profiles as described in claim 1, characterized in that, The native multimodal asynchronous data includes physiological sensing time-series streams, interactive behavior event streams, and environmental physical parameter streams; The generation process of the enhanced synchronous sensing time sequence includes: establishing a unified time reference, performing timestamp alignment and interpolation processing on the physiological sensing time stream, interactive behavior event stream and environmental physical parameter stream, and generating time-aligned data frames; Each time-aligned data frame is assigned a unique context identifier code by a pre-defined context semantic rule base, and then the data frames are arranged in chronological order to generate an enhanced synchronous sensing time sequence.

3. The demand adaptation method based on personalized profiles as described in claim 2, characterized in that, The feature extraction and modal interaction analysis of the enhanced synchronous sensing time series are performed using a cross-modal feature fusion network, which is a three-layer encoder-decoder structure, specifically as follows: The enhanced synchronous sensing temporal sequence is input into the visual encoder, temporal encoder and text encoder respectively for feature extraction, generating visual feature vector, temporal feature vector and text feature vector; The interaction weights between visual feature vectors, temporal feature vectors, and text feature vectors are calculated through a cross-attention mechanism. Based on the interaction weights, the visual feature vectors, temporal feature vectors, and text feature vectors are weighted and fused to generate a fused high-dimensional dynamic feature vector, which is then input into a sparse attention feature decoupling unit. The sparse attention feature decoupling unit uses a multi-head attention mechanism to calculate the sparse weights between each feature dimension within the fused high-dimensional dynamic feature vector. Based on the sparse weights, the fused high-dimensional dynamic feature vector is parsed into a state quantization vector cluster and an interaction preference probability transition matrix, which together constitute a structured dynamic demand profile. The state quantization vector cluster is a set of multiple numerical vectors representing real-time physiological and cognitive states; the interaction preference probability transition matrix is ​​a two-dimensional probability matrix used to describe the dynamic transition probability between different interaction methods. The two-dimensional probability matrix is ​​obtained by the decoder performing a normalized exponential function operation on the fused high-dimensional dynamic feature vector.

4. The demand adaptation method based on personalized profiles as described in claim 3, characterized in that, The process of retrieving the service provider's historical service records and decomposing them into the smallest capability units includes: extracting service action, operation object, and result evaluation triples by performing semantic and dependency parsing analysis on the historical service records; and mapping the triples to the smallest capability units with unified semantic identifiers based on a preset capability atomization classification dictionary.

5. The demand adaptation method based on personalized profiles as described in claim 4, characterized in that, The incremental learning process, which constructs and updates the dynamic capability profile of the service provider, includes: The minimum capability unit is used as the node of the graph, and the temporal association and semantic co-occurrence relationship between the minimum capability units are used as the edge of the graph to construct an initial capability graph. Then, a neural network with temporal dependency graph node feature incremental update is used to aggregate and update the node features of the initial capability graph. When a new historical service record is input, the minimum capability unit, the edge weights between nodes, and the node feature vectors in the neural network with time-dependent graph node features are updated through an incremental learning algorithm to form an updated capability graph. The updated capability graph serves as a dynamic capability profile of the service provider.

6. The demand adaptation method based on personalized profiles as described in claim 5, characterized in that, The dynamic capability profile is specifically a domain-pattern dual embedding network, which includes a domain subgraph and a pattern subgraph. The domain subgraph contains multiple first nodes, each first node representing a specialty domain. The first nodes are connected by edges, and the edges between the first nodes represent the semantic association strength between the specialty domains. The pattern subgraph contains multiple second nodes, each of which represents an operational habit pattern. The second nodes are connected by edges, and the edges between the second nodes represent the state transition probabilities between operational habit patterns. The first node of the domain subgraph and the second node of the pattern subgraph are associated and bridged through the smallest capability unit obtained by decomposing from the historical service records of the same service provider.

7. The demand adaptation method based on personalized profiles as described in claim 6, characterized in that, The multi-dimensional demand-capability matching model includes a matching degree calculation layer, a dynamic weight fusion layer, and a decision backtracking layer. The matching degree calculation layer receives the dynamic demand profile and the dynamic capability profile, and calculates the cosine similarity value on multiple preset dimensions to obtain the original matching vector; each component of the original matching vector corresponds to a cosine similarity value of a preset dimension. The dynamic weight fusion layer is implemented based on a long short-term memory network. The long short-term memory network takes the context identifier code corresponding to the enhanced synchronous perception time sequence and the historical weight sequence as input, outputs the adaptive weight coefficients of each preset dimension, and performs a weighted summation of the original matching vector and the adaptive weight coefficients to generate a comprehensive fit score. The decision backtracking layer records the cosine similarity value of each preset dimension in the matching degree calculation layer, as well as the corresponding adaptive weight coefficient in the dynamic weight fusion layer. Based on the recorded cosine similarity value and adaptive weight coefficient, a tree-structured matching decision backtracking graph is generated to visualize the contribution path of each dimension to the comprehensive fit score.

8. The demand adaptation method based on personalized profiles as described in claim 7, characterized in that, The process of generating the matching decision tracing graph includes: The overall fit score is used as the root node; Each preset dimension is connected to the root node as a child node; Configure two leaf nodes for each child node. The first leaf node stores the cosine similarity value corresponding to the preset dimension, and the second leaf node stores the adaptive weight coefficient corresponding to the preset dimension. Based on the magnitude of the adaptive weight coefficients, all child nodes connected to the root node are sorted and visualized to generate a tree-structured matching decision tracing graph.

9. The demand adaptation method based on personalized profiles as described in claim 8, characterized in that, The agent is implemented using a proximal policy optimization algorithm, including an actor network and a critic network; Adjusting the matching function in the multi-dimensional demand-ability matching model means dynamically modifying the feature weights of the dynamic demand profile and dynamic ability profile used in calculating cosine similarity in the matching degree calculation layer through the parameters output by the actor network. Adjusting the weight generation strategy in the multi-dimensional demand-ability matching model means updating the network connection weights of the long short-term memory network in the dynamic weight fusion layer using the parameters output by the actor network.

10. A demand adaptation system based on personalized profiles, used to implement the demand adaptation method based on personalized profiles as described in any one of claims 1-9, characterized in that, include: The multi-source data acquisition module is used to collect physiological sensor time-series streams, interactive behavior event streams, and environmental physical parameter streams from the demand side in real time, forming native multimodal asynchronous data. After timestamp alignment and context identification, it generates an enhanced synchronous sensing time-series sequence. The demand profile generation module inputs the enhanced synchronous perception time sequence into a pre-trained cross-modal feature fusion network to generate a structured dynamic demand profile. The capability profile building module retrieves the service provider's historical service records, decomposes them into the smallest capability units using a semantic analysis-based service capability decomposition method, builds and updates the service provider's dynamic capability profile through incremental learning, and updates it in real time based on new historical service records. The multi-objective matching optimization module inputs dynamic demand profiles and dynamic capability profiles into a multi-dimensional demand-capability matching model, generates a comprehensive fit score and a matching decision tracing diagram, and outputs a service provider priority ranking list. The iterative optimization module is used to collect feedback data on service effectiveness. It uses the feedback data, the matching decision traceability graph, and the service provider priority ranking list as training samples to input the agent, and dynamically adjusts the matching function and weight generation strategy of the multi-dimensional demand-capability matching model.

Images