A potential customer discovery method and system based on large model semantic understanding

Abstract

The application relates to the field of artificial intelligence and data mining technology, and discloses a potential customer discovery method and system based on large model semantic understanding, which comprises the following steps: extracting historical text representation and posterior state distribution, sampling to generate group-level semantic response prototypes, representing non-sampling nodes, outputting current representation and credibility state; message propagation is carried out in combination with a time sequence diagram, and a posterior state distribution is solved by optimal transmission distribution; a residual error consensus vector is calculated, a candidate parameter update vector is generated in combination with a historical dictionary, low-rank adapter parameters are aggregated and updated after direction consistency verification and local effectiveness verification; reverse residual error is calculated, and graph network parameters are calibrated by state feedback alignment loss to regenerate posterior data; and a conversion probability score is calculated in combination with the current representation to output a potential customer list. The application effectively reduces the calculation overhead of large model dynamic updating, guarantees the stability of parameter iteration, and improves the accuracy of potential customer prediction.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and data mining technology, specifically to a method and system for discovering potential customers based on large-scale model semantic understanding. Background Technology

[0002] In modern business activities, companies typically leverage natural language processing and graph network technologies to analyze customer interaction texts and behavioral patterns to identify customer groups with potential conversion rates. With the development of large language modeling technology, extracting semantic representations of customers from unstructured text using large language models can provide richer textual features for customer status analysis and potential customer discovery. However, in dynamic business scenarios, customer interaction data is continuously generated, and business contexts and customer relationship states change over time. Therefore, relevant analytical models usually need to be continuously updated based on new data and changes in business status.

[0003] Existing potential customer discovery methods still have certain limitations in adapting to the aforementioned dynamic changes. Regarding text representation updates, when model parameters change or the system continuously receives new customer data, re-encoding the entire large language model for all customer nodes will result in high computational resource consumption and processing latency. Regarding parameter updates, some dynamic parameter update methods lack verification of the parameter update direction and local validity, potentially leading to abnormal parameter changes or semantic representation shifts during continuous updates. Regarding customer state analysis, some graph network analysis methods lack differentiated processing for the reliability of text representations of different customer nodes during message propagation, allowing low-reliability node representations or noisy representations to participate in graph message aggregation and affect state inference results. Furthermore, there is often a lack of feedback correction paths based on actual semantic responses between the semantic representation update results of the large language model and the graph network state analysis process, making it difficult for the posterior state distribution calculation process to adapt in a timely manner to changes in the semantic representation space caused by parameter updates, thus affecting the stability of potential customer identification results in dynamic scenarios.

[0004] Therefore, this invention proposes a potential customer discovery method and system based on large model semantic understanding to address the shortcomings of existing technologies. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method and system for discovering potential customers based on large-scale model semantic understanding. It solves the problems of high computational overhead caused by frequent updates of large language models, semantic shifts easily caused by parameter iteration, and low accuracy of potential customer conversion prediction in complex temporal interaction networks in dynamic business interaction scenarios.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a potential customer discovery method based on large-model semantic understanding, comprising:

[0007] Extract the historical text representations of customer nodes and the posterior state distribution data of the previous processing cycle, sample the customer nodes, and generate a group-level semantic response prototype based on the sampled customer nodes.

[0008] The group-level semantic response prototype is used to perform text representation compensation on non-sampled client nodes, and the current text representation and node-level credibility status are output.

[0009] Based on the current text representation, node-level credibility state, and time sequence graph structure, time sequence graph message propagation is performed to generate spatiotemporal representation data and calculate the prior state probability distribution. The posterior state distribution data of the client node is generated by solving the constrained node state optimal transmission allocation matrix.

[0010] Calculate the state group-level residual consensus vector based on the posterior state distribution data and the prior state probability distribution, and generate the candidate parameter update vector by combining the historical verification parameter update dictionary.

[0011] The candidate parameter update vectors are validated, the validated candidate parameter update vectors are aggregated into a unified parameter update vector, and the unified parameter update vector is submitted to update the low-rank adapter parameter state.

[0012] The state group-level inverse residual vector is calculated based on the text representation changes before and after updating the low-rank adapter parameter state. The trainable parameters in the posterior state distribution calculation process are calibrated using the state feedback alignment loss, and the calibrated posterior state distribution data is regenerated based on the calibrated trainable parameters.

[0013] The customer conversion probability score is calculated based on the calibrated posterior state distribution data and the current text representation, and a list of potential customers is output.

[0014] Preferably, a group-level semantic response prototype is generated based on the sampled client nodes; and text representation compensation is performed on non-sampled client nodes using the group-level semantic response prototype, specifically including:

[0015] Based on the historical low-rank adapter parameter version identifier of the client node and the posterior state distribution data of the previous processing cycle, the client nodes are jointly grouped, and the parameter change direction between the current effective low-rank adapter parameter state and the historical low-rank adapter parameter state is calculated.

[0016] The semantic encoding function of the large language model is used to perform Jacobian vector product calculation for the direction of parameter change to obtain the approximate incremental vector of the representation corresponding to the sampled client node. The approximate incremental vector of the representation within the same joint group subset is then averaged to generate the group-level semantic response prototype.

[0017] Based on the posterior state distribution data of the previous processing cycle, the semantic response prototypes of each group corresponding to the historical parameter version group to which the non-sampled client node belongs are weighted and combined to obtain the response increment vector. The response increment vector is then added to the historical text representation to generate the current text representation of the non-sampled client node.

[0018] Preferably, the node-level credibility state is calculated in the following way:

[0019] The input text of the sampled client node is fully forward encoded using the current effective low-rank adapter parameter state to generate an actual text representation. The Euclidean distance between the actual text representation and the prototype compensated text representation generated after representation compensation is calculated, and the obtained Euclidean distance is used as the node approximation error.

[0020] Using the node approximation error as a supervision label, the credibility evaluation network is trained by utilizing the historical text representation corresponding to the sampled customer node, the posterior state distribution vector of the previous processing cycle, the low-rank adapter parameter change norm, the cache time interval, and the second norm of the response increment vector obtained by the weighted combination of the group-level semantic response prototype.

[0021] The credibility evaluation network is trained by inputting the historical text representation, the posterior state distribution vector of the previous processing cycle, the low-rank adapter parameter change norm, the cache time interval, and the second norm of the response increment vector obtained by weighted combination of the group-level semantic response prototypes, and outputs the node prediction approximation error of the customer node.

[0022] The node prediction approximation error is converted into a node-level confidence state based on the following formula:

[0023] ;

[0024] In the formula, Represents client node The node-level credibility status; This represents the approximate error of node predictions in the network output for credibility assessment. This represents the smoothing coefficient of the credibility mapping; It represents exponential operations with the natural constant as the base.

[0025] Preferably, time-series graph message propagation is performed based on the current text representation, node-level credibility state, and time-series graph structure to generate spatiotemporal representation data, specifically including:

[0026] Based on the time sequence graph structure, extract the client node to be updated and the neighboring client nodes that have interaction with the client node to be updated. Perform a linear transformation on the current text representation of the client node to be updated and the neighboring client nodes and concatenate them to calculate the original attention score.

[0027] Extract the normalized edge time intervals corresponding to the interaction events between the client node to be updated and its neighboring client nodes. Combine this with the node-level credibility state and the original attention score, and calculate the target attention coefficient using the following formula:

[0028] ;

[0029] In the formula, Indicates client nodes to be updated With the current neighboring client nodes to be calculated The target attention coefficient between them; This represents the index of the neighboring client node used to calculate the target attention coefficient. Indicates client nodes to be updated The set of neighboring customer node indexes; This represents the set of indexes for the neighboring client nodes. The summation index for the traversal; and These represent the indexes as follows: and The node-level trustworthiness status corresponding to the neighboring customer nodes; and These represent the client nodes to be updated. With index and The original attention scores between neighboring customer nodes; and These represent the client nodes to be updated. With index and The normalized edge time interval between neighboring client nodes; Indicates the time decay coefficient;

[0030] The target attention coefficient is used to perform weighted aggregation on the current text representation of each neighboring customer node after linear transformation, generating spatiotemporal representation data of the customer node to be updated.

[0031] Preferably, the posterior state distribution data of the client nodes is generated by solving the constrained node state optimal transmission allocation matrix, specifically including:

[0032] Calculate the cosine distance between the spatiotemporal representation data and the state codebook vector, generate the state matching cost matrix, and normalize the node-level credibility state corresponding to the client nodes participating in the optimal transmission calculation of node state to establish the source edge distribution.

[0033] The prior state probability distribution is weighted and summed using the source edge distribution to establish the target state edge distribution, and a kernel matrix is ​​constructed based on the state matching cost matrix and the preset entropy regularization coefficient.

[0034] Under the constraints of the source edge distribution and the target edge distribution, the kernel matrix is ​​subjected to alternating scaling iteration to solve the optimal transmission allocation matrix of the node state, and the optimal transmission allocation matrix of the node state is normalized row by row to generate the posterior state distribution data.

[0035] Preferably, the candidate parameter update vector is generated by combining the historical verification parameter update dictionary, specifically including:

[0036] Extract the parameter update vectors of the historical valid verification period to construct the historical parameter change matrix. Perform truncated singular value decomposition on the historical parameter change matrix, extract the right singular vectors to establish the initial valid dictionary basis vectors, and store the initial valid dictionary basis vectors into the valid dictionary slots of the historical verification parameter update dictionary.

[0037] The state group-level residual consensus vector is input into the dictionary coefficient mapping network, and the dictionary combination coefficients corresponding to each dictionary slot are output. Based on the valid identifier of the dictionary slot, the dictionary combination coefficients are weighted and summed with the corresponding valid dictionary basis vectors to generate the candidate parameter update vector.

[0038] Preferably, the process of verifying the candidate parameter update vector and forming a unified parameter update vector specifically includes:

[0039] The candidate parameter update vector is scaled proportionally according to the local linear effective domain radius parameter to generate a candidate update step vector. The product of the perturbation scalar and the candidate update step vector is temporarily superimposed on the current effective low-rank adapter parameter state. Finite difference directional response estimation is performed to obtain the state group-level prediction characterization increment vector.

[0040] Obtain the state group-level semantic change reference vector corresponding to the candidate parameter update vector, calculate the cosine similarity between the state group-level prediction representation increment vector and the state group-level semantic change reference vector, and determine that the direction consistency verification is passed when the cosine similarity is greater than or equal to the direction consistency threshold.

[0041] The candidate update step vector is temporarily and completely loaded into the current effective low-rank adapter parameter state. The actual representation increment vector of the state group-level candidate update step is calculated, and the approximation error is calculated based on the difference between the actual representation increment vector of the state group-level candidate update step and the predicted representation increment vector of the state group. When the approximation error is less than the local validity threshold, the local validity verification is determined to be passed. The candidate update step vectors that pass both the direction consistency verification and the local validity verification are aggregated to obtain the unified parameter update vector.

[0042] Preferably, after determining that the directional consistency verification and local validity verification have passed, the method further includes:

[0043] The larger value between the cosine similarity and zero is determined as the non-negative directional consistency value, and the gating control weight is calculated based on the non-negative directional consistency value.

[0044] Extract the state classification cross-entropy loss term, and use the gate control weight to weight and adjust the state classification cross-entropy loss term. Calculate the gradient vector of the adjusted state classification cross-entropy loss term with respect to the state group level parameters of the low-rank adapter parameter state.

[0045] Calculate the least-squares projection of the state group-level parameter gradient vector into the parameter update subspace spanned by the effective dictionary basis matrices contained in the historical verification parameter update dictionary. Subtract the least-squares projection from the state group-level parameter gradient vector to obtain the dictionary out-of-dictionary parameter direction residual vector. Normalize the dictionary out-of-dictionary parameter direction residual vector whose L2 norm is greater than the dictionary update judgment threshold to obtain candidate new parameter directions. Write the candidate new parameter directions into the set of parameter directions to be verified. After the candidate new parameter directions pass the direction consistency verification and local validity verification, update the historical verification parameter update dictionary using the candidate new parameter directions.

[0046] Preferably, the state group-level inverse residual vector is calculated based on the text representation changes before and after updating the low-rank adapter parameter states, and the trainable parameters in the posterior state distribution calculation process are calibrated using state feedback alignment loss, specifically including:

[0047] For each valid business relationship state group, the customer node used when performing directional consistency verification and local validity verification on the candidate parameter update vector is obtained as the verification customer node; the text representation of the verification customer node before and after the unified parameter update vector is subtracted to obtain the node-level parameter update submission response increment vector of the corresponding verification customer node.

[0048] For each valid business relationship state group, calculate the mean vector and variance vector of the node-level parameter update submission response increment vector corresponding to each verified customer node in the business relationship state group. Concatenate the mean vector and variance vector to generate the corresponding state group-level true semantic response statistics vector. Input the state group-level true semantic response statistics vector into the inverse mapping network in the parameter frozen state and output the corresponding state group-level inverse residual vector.

[0049] When the set of valid business relationship states is not empty, with the large language model body, low-rank adapter parameters, and inverse mapping network parameters frozen, the state feedback alignment loss is calculated based on the following formula, and the trainable parameters in the posterior state distribution calculation process are updated in reverse using the state feedback alignment loss:

[0050] ;

[0051] In the formula, This represents the state feedback alignment loss; Represents the set of valid business relationship states; Represents a set The number of effective business relationship status groups; Indicates the business relationship status group State group-level residual consensus vector; Indicates the business relationship status group The state group-level inverse residual vector; This indicates that the gradient operation has been stopped. This represents the square operation of the 2-norm of a vector.

[0052] This invention also provides a potential customer discovery system based on large model semantic understanding, comprising:

[0053] The semantic extraction module is used to extract the historical text representations of customer nodes and the posterior state distribution data of the previous processing cycle, sample the customer nodes, generate group-level semantic response prototypes based on the sampled customer nodes, and use the group-level semantic response prototypes to perform text representation compensation on non-sampled customer nodes, and output the current text representation and node-level credibility status.

[0054] The graph network analysis module is used to perform time-series graph message propagation based on the current text representation, node-level credibility state, and time-series graph structure, generate spatiotemporal representation data and calculate the prior state probability distribution, and generate the posterior state distribution data of the client node by solving the constrained node state optimal transmission allocation matrix.

[0055] The optimization control module is used to calculate the state group-level residual consensus vector based on the posterior state distribution data and the prior state probability distribution, generate candidate parameter update vectors by combining them with the historical verification parameter update dictionary, and perform verification on the candidate parameter update vectors. The verified candidate parameter update vectors are aggregated into a unified parameter update vector. The unified parameter update vector is submitted to update the low-rank adapter parameter state. The state group-level inverse residual vector is calculated based on the text representation changes before and after updating the low-rank adapter parameter state, and the state feedback alignment loss is calculated to calibrate the trainable parameters in the graph network analysis module.

[0056] The graph network analysis module is also used to regenerate the calibrated posterior state distribution data based on the calibrated trainable parameters.

[0057] The customer output module is used to calculate the customer conversion probability score based on the calibrated posterior state distribution data and the current text representation, and output a list of potential customers.

[0058] This invention provides a method and system for discovering potential customers based on large-model semantic understanding. It has the following beneficial effects:

[0059] 1. This invention compensates for the text representation of non-sampled client nodes by generating a group-level semantic response prototype based on sampled client nodes, and completes subsequent state analysis by combining node-level credibility status. This reduces the computational overhead of repeatedly executing the complete forward encoding of the large language model on all client nodes, while improving the stability of text representation updates under dynamic parameter changes.

[0060] 2. This invention combines current text representation, node-level credibility status, and time sequence graph structure to generate spatiotemporal representation data, and generates posterior state distribution data through optimal transmission allocation of node states constrained by edge distribution. This enables customer interaction relationships, node representation credibility, and state probability allocation constraints to collaboratively participate in customer state inference, reducing the interference of low-credibility representations on state analysis results.

[0061] 3. This invention generates candidate parameter update vectors based on historical verification parameter update dictionaries, and performs directional consistency verification and local validity verification before parameter update submission, thereby restricting unverified parameter updates from being directly submitted and reducing the risk of abnormal parameter changes and semantic representation shifts during continuous parameter updates.

[0062] 4. This invention constructs a state feedback correction mechanism based on the actual text representation changes before and after the low-rank adapter parameter update. The trainable parameters in the posterior state distribution calculation process are calibrated by the state feedback alignment loss, so that the customer state inference process can adapt to the semantic representation space changes caused by the continuous update of the low-rank adapter parameters, thereby improving the stability and reliability of the potential customer identification results. Attached Figure Description

[0063] Figure 1 This is a system logic architecture and module interaction diagram according to an embodiment of the present invention;

[0064] Figure 2 This is a schematic diagram of the method flow according to an embodiment of the present invention;

[0065] Figure 3 This is a schematic diagram illustrating the semantic extraction and representation compensation process according to an embodiment of the present invention.

[0066] Figure 4 This is a block diagram illustrating the principle of the timing diagram message propagation and optimal transmission allocation process in an embodiment of the present invention.

[0067] Figure 5 This is a flowchart illustrating the parallel generation mechanism of candidate parameter update vectors and semantic change references in an embodiment of the present invention.

[0068] Figure 6 This is a logic block diagram of the dual-gated verification and parameter submission update mechanism according to an embodiment of the present invention;

[0069] Figure 7 This is an integrated flowchart of the status feedback correction and customer conversion scoring evaluation loop provided in an embodiment of the present invention. Detailed Implementation

[0070] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0071] Reference Figure 1 This invention provides a potential customer discovery system based on large model semantic understanding, comprising: a semantic extraction module, a graph network analysis module, an optimization control module, and a customer output module.

[0072] The semantic extraction module is used to receive unstructured text data corresponding to client nodes, generate the current text representation of client nodes through the main body of the large language model and the parameters of the low-rank adapter, and calculate the node-level credibility state.

[0073] The graph network analysis module receives the current text representation and node-level credibility state, combines it with customer interaction behavior data to construct a time series graph structure, and calculates the prior state probability distribution and posterior state distribution data of customer nodes.

[0074] The optimization control module is used to calculate the state group-level residual consensus vector based on the prior state probability distribution and posterior state distribution data, and generate candidate parameter update vectors based on the historical verification parameter update dictionary; after verifying the candidate parameter updates, it executes the low-rank adapter parameter update submission, and performs state feedback correction on the graph network analysis module based on the real semantic response.

[0075] The customer output module receives calibrated posterior state distribution data and current text representation, calculates customer conversion probability scores, and generates a list of potential customers based on the customer conversion probability scores.

[0076] Reference Figure 2 This invention provides a potential customer discovery method based on large-model semantic understanding, comprising the following steps:

[0077] S100: Extract the historical text representation, historical low-rank adapter parameter version identifier, and posterior state distribution data of the client nodes, and perform joint grouping and sampling of the client nodes; generate group-level semantic response prototypes based on sampled client nodes, perform text representation compensation on non-sampled client nodes, and output the current text representation and node-level credibility status.

[0078] S200 generates spatiotemporal representation data based on the current text representation, node-level credibility status, and temporal graph structure, calculates the prior state probability distribution, and obtains the posterior state distribution data of the client nodes through optimal transmission allocation.

[0079] S300 calculates the state group-level residual consensus vector based on the posterior state distribution data and the prior state probability distribution, and generates a candidate parameter update vector based on the dictionary update based on the historical verification parameters, while generating a state group-level semantic change reference vector.

[0080] S400 performs directional consistency verification and local validity verification on candidate parameter updates, aggregates the verified state group-level candidate update step vectors, generates a unified parameter update vector and submits it to the current valid low-rank adapter parameter state, and updates the historical verified parameter update dictionary at the same time.

[0081] S500 calculates the state group-level true semantic response statistics vector based on the changes in text representation of the verified customer nodes corresponding to the effective business relationship state group before and after the unified parameter update vector is submitted, and generates the state group-level inverse residual vector; using the relevant parameters of the state feedback alignment loss update graph network analysis module, it calculates the customer conversion probability score based on the calibrated posterior state distribution data and the current text representation, and generates a list of potential customers.

[0082] To further clarify the implementation of each technical aspect of the present invention, the following will provide a detailed description of the implementation of each functional module involved above and its internal processing flow.

[0083] Reference Figure 3 S100, to obtain the text representation of the current processing cycle while reducing the computational cost of forward encoding repetition of client node text, the semantic extraction module jointly groups client nodes based on the historical low-rank adapter parameter versions and the posterior state distribution data of the previous processing cycle. Sampled client nodes are selected from each joint group subset, and group-level semantic response prototypes are calculated based on the parameter direction responses of the sampled client nodes. The semantic extraction module uses the group-level semantic response prototypes to compensate for the historical text representations of non-sampled client nodes, and generates node-level credibility states based on the difference between the actual text representations of the sampled client nodes and the prototype-compensated text representations. This includes the following steps:

[0084] S101, the semantic extraction module reads the historical text representation cached by the target client node, as well as the version identifier of the historical low-rank adapter parameter corresponding to the generation of the historical text representation.

[0085] Unstructured text data corresponding to customer nodes is input from the customer relationship management system, business communication terminal, or business log interface. Each piece of unstructured text data includes at least a customer node identifier, text content, text generation timestamp, and data source type field. Historical text representations are stored in association according to the customer node identifier, text generation timestamp, and historical low-rank adapter parameter version identifier.

[0086] The semantic extraction module reads the historical low-rank adapter parameters used when generating the corresponding historical text representation from the parameter version storage area based on the historical low-rank adapter parameter version identifier. For multiple parameter tensors contained in the low-rank adapter, the semantic extraction module calculates the difference between the currently valid low-rank adapter parameter tensor and the corresponding historical low-rank adapter parameter tensor according to the preset low-rank adapter layer order, and calculates the square root of the sum of the squares of the Frobenius norms of the differences of each parameter tensor to obtain the low-rank adapter parameter change norm of the corresponding client node.

[0087] The semantic extraction module also calculates the caching time interval for the corresponding client node based on the time difference between the end timestamp of the current processing cycle and the generation timestamp of the historical text representation. The low-rank adapter parameter variation norm and the caching time interval are not involved in the joint grouping of client nodes, but are used as input variables for the credibility evaluation network in S103.

[0088] The semantic extraction module reads the posterior state distribution data output and stored by the graph network analysis module in the previous processing cycle. For client nodes... The posterior state distribution vector of the previous processing cycle is denoted as ,in: ;in, Represents client node In the previous processing cycle, it belonged to the business relationship status group. The posterior state probability, This indicates the number of business relationship status groups.

[0089] The semantic extraction module divides historical parameter version groups according to the historical low-rank adapter parameter version identifier, and determines the business relationship state group to which the customer node belongs based on the state number corresponding to the maximum posterior state probability of the previous processing cycle. ;

[0090] A joint group subset is determined by the same historical parameter version group and the same business relationship status group. For each joint group subset, the semantic extraction module performs uniform sampling, selecting 5% to 15% of the total number of customer nodes in the corresponding joint group subset to form a sampled customer node set, and the customer nodes not selected form a non-sampled customer node set.

[0091] S102, for the customer nodes in the sampled customer node set The semantic extraction module segments the corresponding text data and inputs it into the main body of the large language model. It then reads the unfilled token hidden states output from the last Transformer layer and performs average pooling on these unfilled token hidden states to obtain a fixed-dimensional text representation. Here, Transformer represents a self-attention-based transformer network, and Token represents the word units obtained after text segmentation. The semantic encoding function corresponding to the above text representation generation process is denoted as... ,in, Indicates low-rank adapter parameters; [This refers to the client node...] Corresponding input text The text representation obtained after inputting the semantic encoding function is denoted as .

[0092] For those belonging to the historical parameter version group Sampling client nodes The semantic extraction module flattens and concatenates the tensor differences between the current valid low-rank adapter parameters and the corresponding historical low-rank adapter parameters according to a fixed low-rank adapter layer order to obtain the direction of parameter change. The direction of parameter change is represented as... ;in, Indicates the current valid low-rank adapter parameter status. Indicates the historical parameter version group The corresponding historical low-rank adapter parameter status.

[0093] The semantic extraction module targets the direction of parameter change. Perform the Jacobian vector product calculation to obtain the sampled client nodes. The approximate increment vector of the representation :

[0094] ;

[0095] In the formula, Indicates sampling client node The approximate incremental vector is represented by . Semantic encoding function Regarding low-rank adapter parameters The Jacobian matrix in the historical low-rank adapter parameter state The value at that location.

[0096] The semantic extraction module uses forward mode to automatically differentiate and calculate the Jacobian vector product. During the calculation, the direction of parameter changes is... As the tangent vector is input into the automatic differentiation calculation process, the semantic encoding function is directly obtained along the direction of parameter variation. The directional derivative is not explicitly generated in memory to form the complete Jacobian matrix.

[0097] After calculating the Jacobian vector product of the sampled client nodes, the semantic extraction module performs average pooling on the approximate incremental vectors representing the joint grouping subsets to obtain the group-level semantic response prototypes for the corresponding joint grouping subsets:

[0098] ;

[0099] In the formula, Indicates the historical parameter version group Business Relationship Status Group Commonly corresponding group-level semantic response prototypes; This represents the set of sampled client node indices in the corresponding joint group subset; Represents a set The number of sampled client nodes.

[0100] when At this time, the semantic extraction module does not perform average pooling operation, and will convert the corresponding group-level semantic response prototypes. Set as the approximate increment vector of the representation A zero vector with the same dimension.

[0101] S103, for non-sampling client nodes The semantic extraction module is based on the client node. The posterior state distribution vector of the previous processing cycle, relative to its corresponding historical parameter version group. The corresponding semantic response prototypes at each group level are weighted and combined to obtain the response increment vector. This response increment vector is then added to the historical text representation to obtain the current text representation.

[0102] ;

[0103] In the formula, Indicates non-sampled client nodes The current text representation obtained after representation compensation; Indicates non-sampled client nodes Cached historical text representation; Indicates the number of business relationship status groups; Indicates non-sampled client nodes In the previous processing cycle, it belonged to the business relationship status group. The posterior state probability; Indicates non-sampled client nodes Belongs to the historical parameter version group Business Relationship Status Group The common group-level semantic response prototype.

[0104] For sampling client nodes The semantic extraction module uses the same computational method as the non-sampled client nodes, based on the client nodes. The posterior state distribution vector of the previous processing cycle is weighted and combined with the semantic response prototypes of each group corresponding to its historical parameter version group. The resulting response increment vector is then added to the historical text representation to obtain the prototype compensation text representation.

[0105] Meanwhile, the semantic extraction module uses the currently valid low-rank adapter parameter state. For sampling client nodes Perform full forward encoding on the same input text to obtain the actual text representation. The semantic extraction module calculates the Euclidean distance between the actual text representation and the prototype compensated text representation, and determines the obtained distance as the sampling client node. The node approximation error.

[0106] The semantic extraction module incorporates a credibility evaluation network. This network employs a multilayer perceptron structure, consisting of an input layer, a first hidden layer, a second hidden layer, and an output layer connected sequentially.

[0107] Let the dimension of historical text representation be . The number of neurons in the input layer is The input data includes historical text representations, the posterior state distribution vector of the previous processing cycle, and three scalar variables: the low-rank adapter parameter variation norm, the cache interval, and the L2 norm of the response increment vector obtained by weighting the group-level semantic response prototypes according to the posterior state distribution vector.

[0108] The number of neurons in the first hidden layer is set to half the sum of the number of neurons in the input layer and 1, rounded up. The number of neurons in the second hidden layer is set to half the number of neurons in the first hidden layer, rounded up. Both the first and second hidden layers use Gaussian error linear units as activation functions. The output layer has one neuron and uses the Softplus (smoothing rectification) function to output a non-negative node prediction approximation error.

[0109] During the training of the credibility assessment network, the sampled customer nodes are divided into training samples and validation samples in a ratio of 80% and 20%. For each sampled customer node, the historical text representation, the posterior state distribution vector of the previous processing cycle, the low-rank adapter parameter change norm, the cache time interval, and the L2 norm of the response increment vector are concatenated along the feature dimension to form the credibility assessment network input vector; the node approximation error of the corresponding sampled customer node is used as the supervision label.

[0110] The credibility evaluation network is trained using the mean squared error loss function and the Adam (adaptive moment estimation) optimizer, with a learning rate set at 10. -4 Up to 10 -3 Within this range, the training batch size is set between 32 and 256. When the relative change in the mean square error of the validation samples is less than 10 for 10 consecutive training epochs... -4 Alternatively, training can be terminated when the training cycle reaches 200.

[0111] After training the credibility assessment network, the semantic extraction module generates credibility assessment network input vectors for each client node in the current processing cycle using the same data organization method as in the training phase. These vectors are then input into the credibility assessment network to obtain the node prediction approximation error. .

[0112] The semantic extraction module converts the node prediction approximation error into a node-level confidence state according to the exponential decay mapping relationship:

[0113] ;

[0114] In the formula, Represents client node The node-level credibility state has a value range of (0,1]. This represents the approximate error of node predictions in the network output for credibility assessment. This represents the confidence mapping smoothing coefficient, with a value ranging from 0.1 to 1.0; It represents exponential operations with the natural constant as the base.

[0115] The semantic extraction module outputs the current text representation and node-level credibility state. The current text representation serves as the node input feature for time-series graph message propagation in S200, while the node-level credibility state is used for attention weight adjustment and optimal transmission edge distribution calculation in S200, and state group-level residual consensus vector calculation in S300.

[0116] The semantic extraction module will extract client nodes The final text representation output in the current processing cycle is uniformly denoted as... For non-sampling client nodes, Take the current text representation obtained after group-level semantic response prototype compensation; for sampled client nodes, The actual text representation is obtained by fully forward encoding under the current valid low-rank adapter parameter state. This will be used uniformly in subsequent S200 implementations. This represents the current text representation output by S100.

[0117] Reference Figure 4 S200, the graph network analysis module, based on the current text representation and node-level credibility state output by S100, performs time-series graph message propagation processing on the customer interaction behavior data within the current processing cycle to obtain the spatiotemporal representation data of the customer nodes; calculates the prior state probability distribution of the current processing cycle based on the historical posterior state distribution vector, and establishes a state matching cost matrix based on the spatiotemporal representation data; and solves the optimal transmission allocation matrix of node states under the constraints of the source-end edge distribution and the target-end state edge distribution to obtain the posterior state distribution data of the current processing cycle, including the following steps:

[0118] S201, the graph network analysis module receives the current text representation and node-level credibility status of the customer node output by S100, and reads the customer interaction behavior records within the current processing cycle time window.

[0119] Customer interaction records include at least the source customer node identifier, the target customer node identifier, the interaction event type, and the interaction timestamp. The graph network analysis module uses customer nodes as graph nodes, and uses the source customer node identifier and the target customer node identifier to jointly determine the directed edges between customer nodes. Within the same processing cycle, when there are multiple interaction records between the same source customer node and the target customer node, the interaction records are merged into the same edge, and the interaction timestamp of the most recent interaction record is used as the most recent interaction timestamp of that edge, thus forming the time sequence graph structure corresponding to the current processing cycle.

[0120] The graph network analysis module performs a linear transformation on the current text representation of the client node. Let the client node... The current text representation is The trainable linear transformation matrix is The transformed node is represented as Linear transformation matrix The input dimension is the same as the current text representation dimension, and the output dimension is the preset spatiotemporal representation dimension. In this embodiment Set it to 256.

[0121] For the client nodes to be updated in the sequence diagram structure and customer nodes to be updated Neighboring client nodes with interactive relationships The graph network analysis module concatenates the node representations after linear transformation of the two and utilizes trainable attention parameter vectors. Calculate the raw attention score:

[0122] ;

[0123] In the formula, Indicates the client nodes to be updated within the current processing cycle. With neighboring customer nodes The original attention scores between them; This represents a vector concatenation operation; The negative half-axis slope coefficient of the (leakage linear rectifier) ​​function is set to 0.2. Linear transformation matrix. and attention parameter vector These are all trainable parameters of the graph network analysis module.

[0124] The graph network analysis module calculates the normalized edge time interval based on the most recent interaction timestamp of each edge. For client nodes to be updated... With neighboring customer nodes For edges between processes, the normalized edge time interval is obtained by dividing the time difference between the end time of the current processing cycle and the timestamp of the most recent interaction with that edge by the duration of a single processing cycle. The graph network analysis module sets self-loop edges for each client node and connects the client nodes... Normalized edge time interval corresponding to self-loop edges Set to 0.

[0125] The graph network analysis module uses the node-level credibility state of neighboring customer nodes and the normalized edge time interval together to calculate attention weights. (Customer nodes to be updated) With neighboring customer nodes The target attention coefficient between them is calculated as follows:

[0126] ;

[0127] In the formula, Indicates client nodes to be updated With the current neighboring client nodes to be calculated The target attention coefficient between them; This represents the index of the neighboring client node used to calculate the target attention coefficient. Indicates client nodes to be updated The set of neighboring customer node indexes; Represents the set of indexes for neighboring client nodes. The summation index for the traversal; and These represent the indexes as follows: and The node-level trustworthiness status corresponding to the neighboring customer nodes; and These represent the client nodes to be updated. With index and The original attention scores between neighboring customer nodes; and These represent the client nodes to be updated. With index and The normalized edge time interval between neighboring client nodes; This represents the time decay coefficient.

[0128] Time decay coefficient The initialization parameters for the graph network analysis module are set to a range of 0.1 to 2.0. Before training, multiple candidate values ​​are set within this range. The state classification cross-entropy loss is calculated based on historical validation samples, and the candidate value with the smallest corresponding validation loss is selected as the time decay coefficient. The value of .

[0129] The graph network analysis module performs weighted aggregation on the linearly transformed representations of neighboring customer nodes based on the target attention coefficient, thus obtaining the customer node representations. Spatiotemporal representation data:

[0130] ;

[0131] In the formula, Represents client node The spatiotemporal representation vector in the current processing cycle has the following dimensions: Spatiotemporal representation data is used for calculating the state matching cost matrix in S202.

[0132] S202, the graph network analysis module establishes a state transition probability matrix based on the posterior state distribution data of continuous historical processing cycles, and calculates the prior state probability distribution of the current processing cycle accordingly.

[0133] The posterior state distribution data of historical processing cycles is read from historical operation data; when the system is deployed for the first time and online historical state data has not yet been formed, the posterior state distribution data corresponding to each historical processing cycle is generated by using offline historical customer data according to the S200 state distribution calculation process.

[0134] The graph network analysis module reads posterior state distribution data from the past 10 to 30 consecutive processing cycles. For data derived from business relationship states... To business relationship status State transition statistics Calculate according to the following formula:

[0135] ;

[0136] In the formula, Represents client node In the Each historical processing cycle falls under the category of business relationship status. The posterior state probability; Represents client node In the Each historical processing cycle falls under the category of business relationship status. The posterior state probability; 1 represents the smoothing term.

[0137] The graph network analysis module initializes the trainable state transition parameter matrix based on the state transition statistics. The state transition parameter matrix of the first... Line 1 The initial value of the column is: ;

[0138] For the state transition parameter matrix Perform the Softmax (normalization exponent) operation row by row to obtain the state transition probability matrix. :

[0139] ;

[0140] In the formula, Indicates the number of business relationship status groups; This indicates that the customer node has a business relationship status. Transfer to business relationship status The state transition probabilities. State transition parameter matrix. As trainable parameters in the graph network analysis module, they are updated during the state feedback correction process of the S500.

[0141] For customer nodes The posterior state distribution vector of the previous processing cycle With the state transition probability matrix Perform matrix multiplication to obtain the prior state probability distribution for the current processing cycle: ;in, , Represents client node Without incorporating spatiotemporal representation data from the current processing cycle, it belongs to the business relationship state group. The prior state probability.

[0142] The graph network analysis module also includes a status code book. The status code book includes... Each of the state code vectors corresponds to... A business relationship status group. The state code vector is denoted as Its dimension and spatiotemporal representation vector Dimensions The same applies. The status code vector is initialized at the start of training according to a normal distribution with a mean of 0 and a standard deviation of 1, and is updated as a trainable parameter in the graph network analysis module.

[0143] The business relationship status label is derived from the business stage field and historical conversion event records in the customer relationship management system, according to a pre-stored status mapping table. Each customer node corresponds to one status label in each training cycle. Business relationship state labels. During training, state labels are used as supervision labels, and the posterior state distribution data obtained from S203 is used as the prediction probability distribution to calculate the state classification cross-entropy loss. The state classification cross-entropy loss is then used to update the state codebook vector and the trainable parameters of the graph network analysis module.

[0144] The graph network analysis module calculates the spatiotemporal representation vector of the client node. With each state codebook vector The cosine distance between them yields the state matching cost matrix. The state matching cost matrix of the th Line 1 The state matching cost of the column is:

[0145] ;

[0146] In the formula, Represents client node Business Relationship Status Group The cost of matching states between them; To prevent division by zero in the cosine distance calculation, this embodiment sets it to 10. -8 State matching cost matrix Used for calculating the optimal transmission allocation matrix for node states in S203.

[0147] S203, the graph network analysis module establishes the source edge distribution and the target edge distribution based on the node-level credibility state and the prior state probability distribution, and solves the optimal transmission allocation matrix of the node state under the edge distribution constraint.

[0148] For calculations involved in the current processing cycle There are 1 client node, and the source edge distribution vector is denoted as . ,in:

[0149] ;

[0150] In the formula, Represents client node Source-end probability quality in the calculation of the optimal transmission allocation matrix for node state; Represents client node The node-level credibility status; Indicates the calculations involved in the current processing cycle. The summation index is calculated by traversing each client node.

[0151] The target end state edge distribution vector is denoted as The target end state edge distribution is as follows: Each element is calculated by weighting it according to the source edge distribution and the prior state probability distribution:

[0152] ;

[0153] In the formula, Indicates the business relationship status group The target end probability quality in the calculation of the optimal transmission allocation matrix in the current node state.

[0154] The graph network analysis module matches the cost matrix based on the state. Source edge distribution vector and the target end state edge distribution vector Solve for the optimal transfer allocation matrix of node states with entropy regularization. .

[0155] First, based on the state matching cost matrix Calculate the Sinkhorn kernel matrix :

[0156] ;

[0157] In the formula, Represents the client nodes in the Sinkhorn kernel matrix. Business Relationship Status Group The corresponding element; This represents the entropy regularization coefficient, which is set to 0.1 in this embodiment.

[0158] The graph network analysis module initializes a left scaling vector with all elements having a value of 1. and right scaling vector And update alternately according to the following relationship:

[0159] ;

[0160] ;

[0161] In the formula, division represents element-by-element division between corresponding elements; Indicates the number of iterations; To prevent division by zero in the Sinkhorn alternating scaling calculation, this embodiment sets it to 10. -9 .

[0162] After each update of the right scaling vector, calculate the maximum value of the element-wise absolute difference between the current right scaling vector and the right scaling vector of the previous iteration. When the maximum value is less than 10... -5 The iteration stops when the stopping condition is met; it continues if the stopping condition is not met, with the maximum number of iterations set between 50 and 100. After meeting either the stopping condition or reaching the maximum number of iterations, the final left-scaled vector is obtained. and right scaling vector .

[0163] The optimal transmission allocation matrix for node states is calculated as follows:

[0164] ;

[0165] In the formula, ,element Indicates allocation to client nodes Business Relationship Status Group The probabilistic quality of a combination; This means converting the corresponding vector into a diagonal matrix with the elements of that vector as the main diagonal elements.

[0166] The graph network analysis module normalizes the optimal transmission allocation matrix of node states row by row to obtain the posterior state distribution data of client nodes. Belongs to the business relationship status group The posterior state probability is calculated as follows:

[0167] ;

[0168] In the formula, Represents client node The current processing cycle belongs to the business relationship status group. The posterior state probability.

[0169] The posterior state distribution vector of customer node i in the current processing cycle is composed of the posterior state probabilities corresponding to each business relationship state group. The graph network analysis module assembles the posterior state distribution vectors of each client node into posterior state distribution data and outputs it to S300 for calculating node-level state deviation vectors and state group-level residual consensus vectors. Simultaneously, it stores the posterior state distribution data of the current processing cycle for use as historical state input for the next processing cycle.

[0170] Reference Figure 5 S300, the optimization control module calculates the node-level state deviation vector based on the prior state probability distribution and posterior state distribution data output by S200, and obtains the state group-level residual consensus vector by combining the joint weight formed by the node-level credibility state and the posterior state probability; it calculates the dictionary combination coefficients of the historical verification parameter update dictionary based on the state group-level residual consensus vector, generates the candidate parameter update vector, and generates the state group-level semantic change reference vector based on the semantic representation changes of the newly added text in the current processing cycle, including the following steps:

[0171] S301, the optimization control module receives the prior state probability distribution and posterior state distribution data output by the graph network analysis module. For client nodes... The optimization control module calculates the difference between the posterior state distribution and the prior state probability distribution to obtain the node-level state deviation vector:

[0172] ;

[0173] In the formula, Represents client node The node-level state deviation vector in the current processing cycle; Represents client node The posterior state distribution vector; Represents client node The prior state probability distribution vector. The dimension of the node-level state deviation vector is... , where each element represents the change in the posterior state probability relative to the prior state probability of the corresponding business relationship state.

[0174] For business relationship status groups The optimized control module will optimize the client node. Node-level credibility status Posterior state probability of the client node The product of these is used as the joint weight, i.e.: ;

[0175] The optimization control module uses joint weights to perform weighted aggregation of node-level state deviation vectors to obtain business relationship state groups. The corresponding state group-level residual consensus vector:

[0176] ;

[0177] In the formula, Indicates the business relationship status group within the current processing cycle. The state group-level residual consensus vector has the following dimensions: ; To prevent division by zero, this embodiment sets the constant to 10. -8 The state group-level residual consensus vector is used to represent the average state deviation direction of customer nodes within the corresponding business relationship state group.

[0178] S302 optimizes the internal storage of the historical verification parameter update dictionary in the control module. The historical verification parameter update dictionary uses a fixed-slot structure, with the maximum number of dictionary slots denoted as [missing information]. In this embodiment Set to 50. Each dictionary slot corresponds to a dictionary base vector and a valid identifier for the dictionary slot; the valid identifier for the dictionary slot is denoted as... ,in, Indicates the first Each dictionary slot contains a valid dictionary basis vector. Indicates the first One dictionary slot is not enabled.

[0179] When historical parameter update records have been accumulated during online operation, the optimization control module reads the low-rank adapter parameter update vectors that have undergone directional consistency verification and local validity verification in the past 50 to 100 consecutive processing cycles, and uses them as historical verification parameter update samples.

[0180] When the system is deployed for the first time and no online verification records have been formed, the optimization control module uses offline historical customer data to train the low-rank adapter parameters and saves the changes in the low-rank adapter parameters at fixed training intervals; the same directional consistency verification and local validity verification process as S401 and S402 is used for screening, and the parameter update vectors that pass the verification are used as the initial dictionary to construct samples.

[0181] For each historical validation parameter update sample, the optimization control module expands and concatenates the parameter update tensors of each low-rank adapter layer according to a fixed low-rank adapter layer order to form a parameter update vector of uniform dimension; and arranges multiple parameter update vectors in rows to form a historical parameter change matrix.

[0182] The optimization control module performs truncated singular value decomposition on the historical parameter change matrix and selects the top-ranked singular values ​​in descending order. The first set of right singular vectors are used as the initial valid dictionary basis vectors. The number of initial valid dictionary basis vectors... Set within the range of 10 to 50; when the effective rank of the historical parameter change matrix is ​​less than the set value. At that time, the number of initial valid dictionary basis vectors is determined according to the actual valid rank. The valid dictionary slot identifier for the remaining dictionary slots is set to 0.

[0183] The control module internally incorporates a dictionary coefficient mapping network. This network employs a multilayer perceptron structure, with the input layer containing [number of neurons]. Used to receive state group-level residual consensus vectors; the number of hidden layer neurons is set to... A linear rectified function with leakage is used as the activation function; the number of neurons in the output layer is fixed. The hyperbolic tangent function is used as the activation function to output the dictionary combination coefficients corresponding to each dictionary slot.

[0184] During the supervised training phase of the dictionary coefficient mapping network, the optimization control module uses the state group-level residual consensus vector of the historical processing cycle as the training sample input. Upon initial system deployment, for each historical processing cycle and each business relationship state group, the main parameters of the large language model are frozen, and the low-rank adapter parameters are independently trained using the historical training samples of the corresponding business relationship state groups to obtain the corresponding state group-level historical parameter update vectors. Following the directional consistency verification and local validity verification processes specified in S401 and S402, the obtained state group-level historical parameter update vectors are then filtered.

[0185] For the verified state group-level historical parameter update vector, the optimization control module orthogonally projects it onto the parameter update subspace spanned by the current valid dictionary basis vectors. Based on the projection results, the target dictionary combination coefficients corresponding to each valid dictionary basis vector are obtained, and the target dictionary combination coefficients are used as supervision labels. The target dictionary combination coefficients corresponding to the unenabled dictionary slots are set to 0.

[0186] During the initial training phase, the historical training samples, consisting of the state group-level residual consensus vector and the corresponding target dictionary combination coefficients, are divided into training samples and validation samples in a ratio of 80% and 20%, respectively. The training samples are used to update the dictionary coefficients to map network parameters, while the validation samples are used to calculate the mean squared error of the validation samples and determine the training stopping condition.

[0187] After the system is online, the candidate update step vectors of the state group level that have passed the S402 verification but have not yet been aggregated between state groups, and their corresponding state group level residual consensus vectors are used to form new training samples for incremental training of the dictionary coefficient mapping network.

[0188] The dictionary coefficient mapping network is trained using the mean squared error loss function and the Adam optimizer, with a learning rate set to 10. -4 The training batch size is set between 16 and 128. When the relative change in the mean square error of the validation samples is less than 10 for 10 consecutive training epochs... -4 Alternatively, training can be terminated when the training cycle reaches 200.

[0189] Within the current processing cycle, the optimization control module will respectively convert the state group-level residual consensus vectors corresponding to each business relationship state group. Input the trained dictionary coefficient mapping network to obtain the business relationship state group. The corresponding set of dictionary combination coefficients. The dictionary combination coefficients corresponding to each dictionary slot are denoted as . Its value range is [-1, 1].

[0190] S303, the optimization control module generates candidate parameter update vectors for each business relationship state group based on the valid identifiers of dictionary slots, dictionary combination coefficients, and dictionary basis vectors. (Business Relationship State Group) The candidate parameter update vector is calculated as follows:

[0191] ;

[0192] In the formula, Indicates the business relationship status group within the current processing cycle. The corresponding candidate parameter update vector; Indicates the maximum number of dictionary slots; Indicates the first A valid identifier for each dictionary slot; This represents the dictionary coefficients that map the dictionary combination coefficients output by the network. Indicates the first Each dictionary slot corresponds to a dictionary basis vector. When the valid identifier of a dictionary slot is 0, the corresponding dictionary basis vector does not participate in the calculation of the candidate parameter update vector.

[0193] The dimension of the candidate parameter update vector is consistent with the dimension of the parameter update vector formed by unfolding and concatenating the vectors according to the fixed low-rank adapter layer order in S302. When it is necessary to load the candidate parameter update vector into the low-rank adapter parameter state, the optimization control module performs segmentation and shape restoration on the candidate parameter update vector according to the pre-recorded low-rank adapter layer order, the original shape of each parameter tensor, and the positional relationship of each element in the concatenated vector, to obtain the candidate parameter update tensor that corresponds one-to-one with each low-rank adapter parameter tensor.

[0194] While generating candidate parameter update vectors, the optimization control module generates state group-level semantic change reference vectors for the corresponding business relationship state groups.

[0195] The optimization control module filters client nodes containing newly added unstructured text data from the current processing cycle based on the interaction event timestamps. When multiple newly added unstructured text data entries exist for the same client node, the semantic extraction module performs complete forward encoding using the currently valid low-rank adapter parameter state, and performs average pooling on the newly added text representations corresponding to the same client node to obtain the node-level newly added text representation for that client node.

[0196] For business relationship status groups The optimization control module uses the product of the client node-level credibility state and the posterior state probability as the aggregation weight to perform a weighted summation on the newly added text representation at the node level. The weighted summation result is then divided by the sum of the weights participating in the aggregation to obtain the currently added semantic representation at the state group level. .

[0197] The corresponding historical semantic baseline is denoted as The optimization control module reads the business relationship status group from the previous processing cycle. The text representation of the corresponding client node and its associated historical low-rank adapter parameter version identifier are used. For text representations whose historical low-rank adapter parameter versions are inconsistent with the current valid low-rank adapter parameter versions, representation compensation is performed according to the group-level semantic response prototype compensation method in S100, so that the text representations participating in subsequent aggregation correspond to the current valid low-rank adapter parameter versions. After completing the representation compensation, the aggregation weight is determined according to the product of the node-level credibility state and the posterior state probability corresponding to the previous processing cycle. The text representations of each client node are then weighted and normalized and aggregated, and the result is used as the historical semantic baseline. .

[0198] Business Relationship Status Group The corresponding state group-level semantic change reference vector is calculated as follows:

[0199] ;

[0200] In the formula, Indicates the business relationship status group within the current processing cycle. State group-level semantic change reference vector; This indicates the newly added semantic representation at the state group level; This represents the historical semantic baseline corresponding to the currently valid low-rank adapter parameter version.

[0201] When the business relationship status group If there are no customer nodes containing newly added unstructured text data in the current processing cycle, the semantic change reference validity identifier corresponding to the business relationship state group is set to 0, and the directional consistency verification and low-rank adapter parameter update commit in S400 are not performed on the business relationship state group; if there is newly added text data that meets the conditions, the semantic change reference validity identifier is set to 1.

[0202] The optimization control module associates and stores the candidate parameter update vector, state group-level semantic change reference vector, and semantic change reference validity identifier corresponding to each business relationship state group, and transmits them to S400 for candidate parameter update direction consistency verification and local validity verification.

[0203] Reference Figure 6 In step S400, the optimization control module determines the candidate update step vector based on the candidate parameter update vector generated in step S300, and obtains the state group-level prediction representation increment vector through finite difference directional response estimation. The semantic extraction module performs directional consistency verification and local validity verification respectively. The optimization control module adjusts the local linear effective domain radius parameter according to the verification results, and aggregates the verified state group-level candidate update step vectors to form a unified parameter update vector. After completing the submission of the low-rank adapter parameter update, the optimization control module updates the historical verification parameter update dictionary according to the direction of the new parameter, including the following steps:

[0204] S401, the optimization control module receives the candidate parameter update vectors corresponding to each business relationship state group output by S300. For each business relationship state group... Its candidate parameter update vector is denoted as .

[0205] The optimization control module maintains local linear effective domain radius parameters for each business relationship state group. The local linear effective domain radius parameter corresponding to the current processing cycle is denoted as . The initial value is set in the range of 0.01 to 0.1.

[0206] The optimization control module calculates the candidate parameter update vector. The second norm, and then the second norm is compared with 10. -9 Compare them. When the L2 norm is less than 10...-9 When the candidate parameter update vector remains unchanged; when the L2 norm is not less than 10 -9 And greater than the local linear effective domain radius parameter At that time, according to The candidate parameter update vector is scaled proportionally by the ratio to the L2 norm; otherwise, the candidate parameter update vector remains unchanged. The vector obtained after the above processing is denoted as the candidate update step vector. .

[0207] The L2 norm of the candidate update step vector does not exceed the radius parameter of the corresponding local linear effective domain. Whether the candidate update step vector satisfies the first-order approximation condition near the current parameter state is determined by the local validity verification in S402.

[0208] For semantic changes in S300, refer to the business relationship state group with a valid identifier of 1. The semantic extraction module selects the state number corresponding to the highest posterior state probability from the client nodes with newly added unstructured text data in the current processing cycle. Customer nodes form a business relationship status group. The candidate verification node set is selected; the candidate verification nodes are sorted from high to low according to the number of new interaction records in the current processing cycle, and the top 1 to 5 customer nodes are selected to form the verification customer node set for the corresponding business relationship status group; if there are less than 5 candidate verification nodes, all candidate verification nodes are used.

[0209] The semantic extraction module performs finite-difference directional response estimation on each input text in the set of validation client nodes. For the validation client nodes... In the current valid low-rank adapter parameter state Based on this, the candidate update steps after perturbation Temporarily loaded into the corresponding low-rank adapter parameters, where the perturbation scalar Set to 10 -5 .

[0210] Verify client node The corresponding predicted characterization increment is calculated as follows:

[0211] ;

[0212] In the formula, Indicates the business relationship status group Corresponding verification client node The predicted incremental vector; This represents the semantic encoding function defined in S100; Indicates the business relationship status group The corresponding candidate update step vector; This represents a finite difference perturbation scalar.

[0213] The semantic extraction module performs an arithmetic average of the predicted representation increment vectors in the set of customer nodes within the same business relationship state group to obtain the state group-level predicted representation increment vector, denoted as . The state group-level prediction representation increment vector is used for directional consistency verification and local validity verification in S402.

[0214] S402, the semantic extraction module performs directional consistency verification and local validity verification respectively.

[0215] For directional consistency verification, the semantic extraction module reads the state group-level semantic change reference vector output by S300. Calculate the state group-level prediction representation increment vector The cosine similarity between the vector and the state group-level semantic change reference vector is used as the directional similarity.

[0216] ;

[0217] In the formula, Indicates the business relationship status group directional similarity; Indicates the business relationship status group State group-level semantic change reference vector; To prevent division by zero, this embodiment sets the constant to 10. -8 .

[0218] The orientation consistency threshold is set to 0.5. When the orientation similarity... When the similarity is greater than or equal to 0.5, the corresponding business relationship status group is deemed to have passed the directional consistency verification; when the directional similarity is less than 0.5, the directional consistency verification is deemed to have failed.

[0219] For local validity verification, the semantic extraction module continues to use the same business relationship state group already determined in S401 to verify the set of customer nodes and the same text input, without reselecting the customer nodes to verify.

[0220] The semantic extraction module extracts candidate update step vectors according to the low-rank adapter layer order and parameter tensor shape recorded by S300. Restore the candidate update step tensor to the tensor corresponding to each low-rank adapter parameter tensor, and temporarily superimpose it onto the current valid low-rank adapter parameter state.

[0221] For verifying client nodes The semantic extraction module calculates the text representation before and after the candidate update step is fully loaded, and uses the difference between the two as the actual representation increment vector of the candidate update step. Subsequently, the arithmetic mean of the actual representation increment vectors of the candidate update steps for each verified customer node within the same business relationship state group is calculated to obtain the state group-level candidate update step actual representation increment vector. .

[0222] The semantic extraction module calculates the approximation error based on the difference between the state group-level predicted representation increment vector and the actual representation increment vector of the state group-level candidate update step.

[0223] ;

[0224] In the formula, Indicates the business relationship status group Approximate error; To prevent division by zero, this embodiment sets the constant to 10. -8 .

[0225] The local validity threshold is set to 0.2. When the approximation error... If the value is less than 0.2, the corresponding business relationship state group is deemed to have passed the local validity verification; otherwise, the local validity verification is deemed to have failed.

[0226] S403, the optimization control module adjusts the corresponding local linear effective domain radius parameter based on the approximate error of each business relationship state group.

[0227] When the approximation error is less than 0.1, the local linear effective domain radius parameter of the corresponding business relationship state group is multiplied by 1.2; when the approximation error is greater than 0.25, the corresponding local linear effective domain radius parameter is multiplied by 0.8; when the approximation error is within the closed interval of 0.1 to 0.25, the corresponding local linear effective domain radius parameter remains unchanged.

[0228] The lower limit of the local linear effective domain radius parameter is set to 10. -4 The upper limit is set to 0.5. When the feedback adjustment result is lower than the lower limit, the local linear effective domain radius parameter for the next processing cycle is set to 10. -4 When the feedback adjustment result is higher than the upper limit, it is set to 0.5; otherwise, the value after feedback adjustment is directly used as the local linear effective domain radius parameter of the business relationship state group corresponding to the next processing cycle.

[0229] During training, the optimization control module calculates the gating control weights based on orientation similarity. First, the negative values ​​of the orientation similarity are truncated to 0 to obtain the non-negative orientation consistency values. The gating control weights are then calculated according to the following relationship:

[0230] ;

[0231] In the formula, Indicates the business relationship status group The gating control weights range from 0 to 1. The calculation rule is when Take when greater than 0 Otherwise, take 0; This represents the gating sensitivity coefficient, which is set to 5.0 in this embodiment.

[0232] In the subsequent calculation of the state group-level parameter gradient in S403, the optimization control module uses the gating control weights to adjust the state classification cross-entropy loss term of the corresponding business relationship state group. When the directional consistency is low, the contribution of the corresponding state group loss term to the state group-level parameter gradient is reduced; when the directional consistency is high, a relatively large gradient contribution is retained.

[0233] The optimization control module will combine the business relationship state group numbers that have passed both directional consistency verification and local validity verification into a set of valid business relationship state groups. .

[0234] When the effective business relationship status group set When the set is empty, the optimization control module does not perform state group update weight normalization and unified parameter update vector generation, keeps the current effective low-rank adapter parameter state unchanged, and does not perform historical verification parameter update dictionary addition or replacement operations in this processing cycle.

[0235] For the effective business relationship status group The optimization control module calculates the total effective weight of the state group based on the node-level trustworthiness status and the posterior state probability. The total effective weight of the state group is the sum of the node-level trustworthiness status of each customer node and its business relationship state group within the current processing cycle. The sum of the products of the posterior state probabilities is denoted as... .

[0236] The optimization control module uses the product of the total effective weights of the state group and the gating control weights as the unnormalized value of the state group update weights, and then applies this value to the set of effective business relationship state groups. Normalization is performed internally. Using the normalized state group update weights, the candidate update step vectors at each state group level are weighted and summed to obtain the unified parameter update vector for the current processing cycle.

[0237] The optimization control module restores the unified parameter update vector to the corresponding low-rank adapter parameter update tensor according to the low-rank adapter layer order and parameter tensor shape recorded by S300, and accumulates it to the current valid low-rank adapter parameter state.

[0238] After submitting the low-rank adapter parameter update, the optimization control module generates a new low-rank adapter parameter version identifier. The parameter version identifier consists of the current processing cycle number and an incrementing version number, and is associated with the set of valid business relationship status groups for this processing cycle. Associated storage.

[0239] After submitting the low-rank adapter parameter update, the optimization control module updates the historical verification parameter update dictionary.

[0240] For the effective business relationship status group The optimization control module extracts the state classification cross-entropy loss term after adjustment by the corresponding gated control weights, and calculates the gradient of this loss term with respect to the low-rank adapter parameters. Following the fixed low-rank adapter layer order specified in S300, the optimization control module expands and concatenates the gradients corresponding to each parameter tensor sequentially to obtain the state group-level parameter gradient vector. .

[0241] The optimization control module extracts historical verification parameters to update the dictionary basis vectors whose valid identifiers in the dictionary slots are 1, and arranges the dictionary basis vectors column-wise to form a valid dictionary basis matrix. .

[0242] The optimization control module calculates the gradient vector of state group-level parameters. In the effective dictionary basis matrix Zhang Cheng's parameter update subspace uses least-squares projection, and the corresponding projection component is subtracted from the state group-level parameter gradient vector to obtain the dictionary-external parameter direction residual vector. In the least squares projection calculation, a value of 10 is added to the diagonal of the Gram matrix. -8 Regularization terms are used to reduce numerical instability that occurs when a matrix approaches a singular state.

[0243] The optimization control module calculates the directional residual vector of the dictionary extrinsic parameters. The L2 norm is calculated and compared with the dictionary update decision threshold. The dictionary update decision threshold is set to 0.1. When the dictionary extrinsic parameter direction residual vector... When the L2 norm is greater than 0.1, the vector is normalized to obtain the candidate new parameter direction; when its L2 norm is not greater than 0.1, no candidate new parameter direction is generated.

[0244] The candidate new parameter direction is first written into the set of parameter directions to be verified. The optimization control module scales the candidate new parameter direction according to the business relationship state group to which it belongs, using the local linear effective domain radius parameter of the corresponding business relationship state group, to form an update step to be verified, and verifies it according to the direction consistency verification and local effectiveness verification methods specified in S402.

[0245] When the candidate new parameter direction passes the verification and there is an empty slot with a valid dictionary slot identifier of 0 in the historical verification parameter update dictionary, the candidate new parameter direction is written into an empty dictionary slot and the valid dictionary slot identifier is set to 1.

[0246] The optimization control module maintains the historical call count for each valid dictionary slot and sets a dictionary slot call determination threshold. When generating the state group-level candidate parameter update vector, if the valid identifier of a dictionary slot is 1 and the absolute value of the dictionary combination coefficient corresponding to that dictionary slot is greater than or equal to the dictionary slot call determination threshold, the historical call count of the corresponding dictionary slot is increased by 1; in this embodiment, the dictionary slot call determination threshold is set to 0.05.

[0247] When all historical verification parameters in the dictionary are updated When all dictionary slots are in a valid state, the optimization control module compares the historical call counts of each dictionary slot, selects the slot containing the dictionary base vector with the fewest historical call counts, writes the verified candidate new parameter direction into that slot to replace the original dictionary base vector, and resets the historical call count of the replaced dictionary slot to 0.

[0248] After the historical validation parameter update dictionary is updated with new or replaced data, the optimization control module uses the new training samples and historical training samples to perform incremental training on the dictionary coefficient mapping network. Before the incremental training is completed, the output coefficients corresponding to the updated dictionary slots are fixed at 0; after the incremental training is completed, the corresponding dictionary slots are restored to participate in the generation of candidate parameter update vectors.

[0249] Reference Figure 7 In step S500, the optimization control module calculates the text representation change before and after submitting the unified parameter update vector based on the effective business relationship state group set determined in S403 and the corresponding verification customer node set determined in S401, obtaining the state group-level true semantic response statistics vector. This vector is then input into the inverse mapping network to generate the corresponding state group-level inverse residual vector. The state feedback alignment loss is calculated based on the state group-level residual consensus vector and the state group-level inverse residual vector. The trainable parameters in the graph network analysis module are updated using the state feedback alignment loss, and the posterior state distribution data for the current processing cycle is recalculated. The customer output module calculates the customer conversion probability score based on the calibrated posterior state distribution data and the current text representation, and generates a potential customer list, including the following steps:

[0250] The set of valid business relationship states in the current processing cycle When the set is empty, the real semantic response statistics and state feedback correction process in S501 and S502 are not executed, and the posterior state distribution data output by S200 is directly used as the posterior state distribution data for customer conversion probability scoring in the current processing cycle.

[0251] In step S501, the optimization control module reads the set of valid business relationship state groups determined in step S403, and the set of verification customer nodes corresponding to each valid business relationship state group determined in step S401. For the same verification customer node and the same text input, the semantic extraction module performs complete forward encoding using the low-rank adapter parameter version corresponding to the unified parameter update vector before submission, and the current valid low-rank adapter parameter state after the low-rank adapter parameter update submission. The module then subtracts the text representation before the parameter update submission from the text representation after the parameter update submission to obtain the node-level parameter update submission response increment vector for the corresponding customer node.

[0252] Targeting the effective business relationship status group The optimization control module calculates the mean and variance vectors of the node-level parameter update submission response increment vectors for each verified customer node within the business relationship state group. When the set of verified customer nodes contains only one customer node, the corresponding variance vector is set to a zero vector with the same dimension as the node-level parameter update submission response increment vector; when the number of verified customer nodes is greater than 1, the sample variance is calculated separately according to each feature dimension.

[0253] The optimization control module concatenates the mean vector and variance vector along the feature dimensions to obtain the business relationship state group. The corresponding state group-level true semantic response statistics vector. Let the text representation dimension be... Then the dimension of the state group-level true semantic response statistics vector is .

[0254] The control module internally incorporates an inverse mapping network. This network employs a multilayer perceptron structure, consisting of a sequentially connected input layer, hidden layer, and output layer. The input layer contains [number of neurons]. The dimension is consistent with the dimension of the state group-level true semantic response statistics vector; the number of hidden layer neurons is set to... and The sum is rounded up to half its original value, and a linear rectified function is used as the activation function; the number of neurons in the output layer is... Each output neuron corresponds to a business relationship state group, therefore the dimension of the output vector of the output layer is... No activation function is set.

[0255] During the training of the inverse mapping network, training samples are jointly determined by historical processing cycles and business relationship state groups. Upon initial system deployment, the parameter update process for the corresponding historical processing cycle is replayed using offline historical customer data. For parameter updates that pass directional consistency verification and local validity verification, the same processing method as S501 is adopted: complete forward encoding before and after parameter updates is performed on the same verified customer nodes and the same text input. Based on the changes in text representation, the state group-level true semantic response statistics vector for the corresponding business relationship state group is calculated, and the state group-level residual consensus vector corresponding to the same historical processing cycle and the same business relationship state group is used as the supervision label.

[0256] After the system is running online, the state group-level true semantic response statistics vector generated by S501, and the state group-level residual consensus vector corresponding to the same processing cycle and the same business relationship state group, will be added to the inverse mapping network training sample set. Before training, the training samples will be divided into training samples and validation samples in a ratio of 80% and 20%.

[0257] The inverse mapping network is trained using the mean squared error loss function and the Adam optimizer, with a learning rate set to 10. -4 The training batch size is set between 16 and 128. When the relative change in the mean square error of the validation samples is less than 10 for 10 consecutive training epochs... -5 Alternatively, training can be terminated when the training cycle reaches 200.

[0258] After training is complete, the optimization control module freezes the parameters of the inverse mapping network. This applies to the valid business relationship state group for the current processing cycle. The corresponding state group-level true semantic response statistics vector is input into the inverse mapping network to obtain the state group-level inverse residual vector, denoted as... .

[0259] State group level inverse residual vector The dimension is , and the consensus vector of the state group level residual The dimensions are consistent, and it is used for state feedback alignment calculation in S502.

[0260] S502, the optimization control module reads the set of valid business relationship status groups determined in S403. For sets For each business relationship state group, read the state group-level residual consensus vector obtained from S301. And the state group-level inverse residual vector obtained by S501 .

[0261] During the state feedback correction phase, the inverse residual vector at the state group level is... Perform a stopping gradient operation, making it only a target for state feedback correction, and do not update the inverse mapping network parameters through the state feedback alignment loss at this stage.

[0262] The state feedback alignment loss is calculated as the average of the squared errors of each state group within the set of valid business relationship state groups:

[0263] ;

[0264] In the formula, This represents the state feedback alignment loss; Represents a set The number of effective business relationship status groups; This indicates that the gradient operation has been stopped. This represents the square operation of the 2-norm of a vector.

[0265] The optimization control module updates the trainable parameters in the graph network analysis module based on the state feedback and alignment loss. During this feedback update phase, the large language model body, low-rank adapter parameters, and inverse mapping network parameters remain frozen.

[0266] The parameters involved in feedback updates in the graph network analysis module include the trainable state transition parameter matrix in S202. Status code vector, and linear transformation matrix in S201 and attention parameter vector .

[0267] During training, the computational graph formed by the time-series graph message propagation process in S201, the prior state probability distribution and state matching cost matrix calculation process in S202, the node state optimal transmission allocation matrix calculation and posterior state distribution normalization process in S203, and the node-level state deviation vector and state group-level residual consensus vector calculation process in S301 is retained. This allows the gradient of the state feedback alignment loss to be transmitted to the trainable parameters in the graph network analysis module along the state group-level residual consensus vector, posterior state distribution data, node state optimal transmission allocation matrix, state matching cost matrix, and time-space graph message propagation path.

[0268] The optimization control module uses the Adam optimizer to update the trainable parameters in the graph network analysis module, with the learning rate set to 10. -4 After completing the feedback update, the graph network analysis module reads the updated parameters and re-executes the processing steps S201 to S203 based on the current text representation, node-level credibility status, and temporal graph structure of the current processing cycle to obtain the calibrated posterior state distribution data.

[0269] In step S503, the customer output module reads the posterior state distribution data used for customer conversion probability scoring in the current processing cycle, and the current text representation of the corresponding customer node output in S100. It then concatenates these two data points along the feature dimension to obtain the joint feature vector of the customer node. When the set of effective business relationship states is not empty, the posterior state distribution data used for customer conversion probability scoring is the calibrated posterior state distribution data output in S502; when the set of effective business relationship states is empty, the posterior state distribution data used for customer conversion probability scoring is the posterior state distribution data output in S200.

[0270] The client output module internally sets the probability mapping weight vector. and bias terms Client Node The corresponding output layer scalar calculation is as follows:

[0271] ;

[0272] In the formula, Represents client node The corresponding output layer scalar; Represents client node The joint eigenvectors.

[0273] The probability mapping weight vector and bias term are trained under supervision using historical customer-processing cycle samples. For each historical customer-processing cycle sample, the current text representation of the customer node in the corresponding historical processing cycle is concatenated with the posterior state distribution data after state feedback correction to form a joint feature vector consistent with the data structure of the online computing stage.

[0274] The end time of the processing cycle is used as the starting point for observation. When a customer node generates a preset conversion event within 30 calendar days after the starting point of observation, the corresponding training label is set to 1; otherwise, it is set to 0. The preset conversion event is determined by the transaction status field, contract generation record, or order generation record in the customer relationship management system.

[0275] Before training, historical customer-processing cycle samples with joint feature vectors and corresponding training labels are divided into training samples and validation samples in a ratio of 80% and 20%, respectively. The training samples are used to update the probability mapping weight vector and bias term, while the validation samples are used to calculate the validation set loss and determine the training stopping condition.

[0276] During training, a binary cross-entropy loss function and the Adam optimizer were used, with the learning rate set at 10. -4 Up to 10 -3 Within this range, the training batch size is set between 32 and 256. When the relative change in validation set loss is less than 10 for 10 consecutive training epochs... -4Alternatively, training can be terminated when the training cycle reaches 200.

[0277] The client output module will output layer scalar Input the Sigmoid (S-shaped) function to obtain the client node. Customer conversion probability score:

[0278] ;

[0279] In the formula, Represents client node The customer conversion probability score ranges from 0 to 1. This represents the output layer scalar of the corresponding customer node. The customer conversion probability score is used to represent the probability estimate of a customer node generating a preset conversion event within the preset conversion event observation window.

[0280] The system pre-sets conversion score thresholds and activity count thresholds. In this embodiment, the conversion score threshold is set to 0.8, and the activity count threshold is set to 3.

[0281] The customer output module reads customer interaction records and counts the number of historical interactions of a customer node within the 30 calendar days prior to the end of the current processing cycle. When the number of historical interactions of a customer node is greater than or equal to the active number threshold, and the customer conversion probability score is greater than or equal to the conversion score threshold, the customer node's customer node identifier is added to the potential customer list; otherwise, the customer node is not added to the potential customer list.

[0282] After completing the judgment of all customer nodes within the current processing cycle, the customer output module sends the list of potential customers to the customer relationship management system or business execution terminal.

[0283] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for discovering potential customers based on large-scale model semantic understanding, characterized in that, include: Extract the historical text representations of customer nodes and the posterior state distribution data of the previous processing cycle, sample the customer nodes, and generate a group-level semantic response prototype based on the sampled customer nodes. The group-level semantic response prototype is used to perform text representation compensation on non-sampled client nodes, and the current text representation and node-level credibility status are output. Based on the current text representation, node-level credibility state, and time sequence graph structure, time sequence graph message propagation is performed to generate spatiotemporal representation data and calculate the prior state probability distribution. The posterior state distribution data of the client node is generated by solving the constrained node state optimal transmission allocation matrix. Calculate the state group-level residual consensus vector based on the posterior state distribution data and the prior state probability distribution, and generate the candidate parameter update vector by combining the historical verification parameter update dictionary. The candidate parameter update vectors are validated, the validated candidate parameter update vectors are aggregated into a unified parameter update vector, and the unified parameter update vector is submitted to update the low-rank adapter parameter state. The state group-level inverse residual vector is calculated based on the text representation changes before and after updating the low-rank adapter parameter state. The trainable parameters in the posterior state distribution calculation process are calibrated using the state feedback alignment loss, and the calibrated posterior state distribution data is regenerated based on the calibrated trainable parameters. The customer conversion probability score is calculated based on the calibrated posterior state distribution data and the current text representation, and a list of potential customers is output.

2. The potential customer discovery method based on large-model semantic understanding according to claim 1, characterized in that, Generate a group-level semantic response prototype based on sampled client nodes; and use the group-level semantic response prototype to perform text representation compensation on non-sampled client nodes, specifically including: Based on the historical low-rank adapter parameter version identifier of the client node and the posterior state distribution data of the previous processing cycle, the client nodes are jointly grouped, and the parameter change direction between the current effective low-rank adapter parameter state and the historical low-rank adapter parameter state is calculated. The semantic encoding function of the large language model is used to perform Jacobian vector product calculation for the direction of parameter change to obtain the approximate incremental vector of the representation corresponding to the sampled client node. The approximate incremental vector of the representation within the same joint group subset is then averaged to generate the group-level semantic response prototype. Based on the posterior state distribution data of the previous processing cycle, the semantic response prototypes of each group corresponding to the historical parameter version group to which the non-sampled client node belongs are weighted and combined to obtain the response increment vector. The response increment vector is then added to the historical text representation to generate the current text representation of the non-sampled client node.

3. The potential customer discovery method based on large-model semantic understanding according to claim 2, characterized in that, The node-level credibility status is calculated in the following way: The input text of the sampled client node is fully forward encoded using the current effective low-rank adapter parameter state to generate an actual text representation. The Euclidean distance between the actual text representation and the prototype compensated text representation generated after representation compensation is calculated, and the obtained Euclidean distance is used as the node approximation error. Using the node approximation error as a supervision label, the credibility evaluation network is trained by utilizing the historical text representation corresponding to the sampled customer node, the posterior state distribution vector of the previous processing cycle, the low-rank adapter parameter change norm, the cache time interval, and the second norm of the response increment vector obtained by the weighted combination of the group-level semantic response prototype. The credibility evaluation network is trained by inputting the historical text representation, the posterior state distribution vector of the previous processing cycle, the low-rank adapter parameter change norm, the cache time interval, and the second norm of the response increment vector obtained by weighted combination of the group-level semantic response prototypes, and outputs the node prediction approximation error of the customer node. The node prediction approximation error is converted into a node-level confidence state based on the following formula: ； In the formula, Represents client node The node-level credibility status; This represents the approximate error of node predictions in the network output for credibility assessment. This represents the smoothing coefficient of the credibility mapping; It represents exponential operations with the natural constant as the base.

4. The potential customer discovery method based on large-model semantic understanding according to claim 1, characterized in that, Based on the current text representation, node-level credibility state, and temporal graph structure, temporal graph message propagation is performed to generate spatiotemporal representation data, specifically including: Based on the time sequence graph structure, extract the client node to be updated and the neighboring client nodes that have interaction with the client node to be updated. Perform a linear transformation on the current text representation of the client node to be updated and the neighboring client nodes and concatenate them to calculate the original attention score. Extract the normalized edge time intervals corresponding to the interaction events between the client node to be updated and its neighboring client nodes. Combine this with the node-level credibility state and the original attention score, and calculate the target attention coefficient using the following formula: ； In the formula, Indicates client nodes to be updated With the current neighboring client nodes to be calculated The target attention coefficient between them; This represents the index of the neighboring client node used to calculate the target attention coefficient. Indicates client nodes to be updated The set of neighboring customer node indexes; This represents the set of indexes for the neighboring client nodes. The summation index for the traversal; and These represent the indexes as follows: and The node-level trustworthiness status corresponding to the neighboring customer nodes; and These represent the client nodes to be updated. With index and The original attention scores between neighboring customer nodes; and These represent the client nodes to be updated. With index and The normalized edge time interval between neighboring client nodes; Indicates the time decay coefficient; The target attention coefficient is used to perform weighted aggregation on the current text representation of each neighboring customer node after linear transformation, generating spatiotemporal representation data of the customer node to be updated.

5. The potential customer discovery method based on large-model semantic understanding according to claim 1, characterized in that, The posterior state distribution data of the client nodes is generated by solving the constrained node state optimal transmission allocation matrix, specifically including: Calculate the cosine distance between the spatiotemporal representation data and the state codebook vector, generate the state matching cost matrix, and normalize the node-level credibility state corresponding to the client nodes participating in the optimal transmission calculation of node state to establish the source edge distribution. The prior state probability distribution is weighted and summed using the source edge distribution to establish the target state edge distribution, and a kernel matrix is ​​constructed based on the state matching cost matrix and the preset entropy regularization coefficient. Under the constraints of the source edge distribution and the target edge distribution, the kernel matrix is ​​subjected to alternating scaling iteration to solve the optimal transmission allocation matrix of the node state, and the optimal transmission allocation matrix of the node state is normalized row by row to generate the posterior state distribution data.

6. The potential customer discovery method based on large-model semantic understanding according to claim 1, characterized in that, The candidate parameter update vector is generated by combining the historical validation parameter update dictionary, specifically including: Extract the parameter update vectors of the historical valid verification period to construct the historical parameter change matrix. Perform truncated singular value decomposition on the historical parameter change matrix, extract the right singular vectors to establish the initial valid dictionary basis vectors, and store the initial valid dictionary basis vectors into the valid dictionary slots of the historical verification parameter update dictionary. The state group-level residual consensus vector is input into the dictionary coefficient mapping network, and the dictionary combination coefficients corresponding to each dictionary slot are output. Based on the valid identifier of the dictionary slot, the dictionary combination coefficients are weighted and summed with the corresponding valid dictionary basis vectors to generate the candidate parameter update vector.

7. The potential customer discovery method based on large-model semantic understanding according to claim 1, characterized in that, The process of verifying the candidate parameter update vectors and forming a unified parameter update vector specifically includes: The candidate parameter update vector is scaled proportionally according to the local linear effective domain radius parameter to generate a candidate update step vector. The product of the perturbation scalar and the candidate update step vector is temporarily superimposed on the current effective low-rank adapter parameter state. Finite difference directional response estimation is performed to obtain the state group-level prediction characterization increment vector. Obtain the state group-level semantic change reference vector corresponding to the candidate parameter update vector, calculate the cosine similarity between the state group-level prediction representation increment vector and the state group-level semantic change reference vector, and determine that the direction consistency verification is passed when the cosine similarity is greater than or equal to the direction consistency threshold. The candidate update step vector is temporarily and completely loaded into the current effective low-rank adapter parameter state. The actual representation increment vector of the state group-level candidate update step is calculated, and the approximation error is calculated based on the difference between the actual representation increment vector of the state group-level candidate update step and the predicted representation increment vector of the state group. When the approximation error is less than the local validity threshold, the local validity verification is determined to be passed. The candidate update step vectors that pass both the direction consistency verification and the local validity verification are aggregated to obtain the unified parameter update vector.

8. The potential customer discovery method based on large-model semantic understanding according to claim 7, characterized in that, After determining that the directional consistency verification and local validity verification have passed, the method further includes: The larger value between the cosine similarity and zero is determined as the non-negative directional consistency value, and the gating control weight is calculated based on the non-negative directional consistency value. Extract the state classification cross-entropy loss term, and use the gate control weight to weight and adjust the state classification cross-entropy loss term. Calculate the gradient vector of the adjusted state classification cross-entropy loss term with respect to the state group level parameters of the low-rank adapter parameter state. Calculate the least-squares projection of the state group-level parameter gradient vector into the parameter update subspace spanned by the effective dictionary basis matrices contained in the historical verification parameter update dictionary. Subtract the least-squares projection from the state group-level parameter gradient vector to obtain the dictionary out-of-dictionary parameter direction residual vector. Normalize the dictionary out-of-dictionary parameter direction residual vector whose L2 norm is greater than the dictionary update judgment threshold to obtain candidate new parameter directions. Write the candidate new parameter directions into the set of parameter directions to be verified. After the candidate new parameter directions pass the direction consistency verification and local validity verification, update the historical verification parameter update dictionary using the candidate new parameter directions.

9. The potential customer discovery method based on large-model semantic understanding according to claim 1, characterized in that, The state group-level inverse residual vector is calculated based on the text representation changes before and after updating the low-rank adapter parameter states. The trainable parameters in the posterior state distribution calculation process are calibrated using the state feedback alignment loss, specifically including: For each valid business relationship state group, the customer node used when performing directional consistency verification and local validity verification on the candidate parameter update vector is obtained as the verification customer node; the text representation of the verification customer node before and after the unified parameter update vector is subtracted to obtain the node-level parameter update submission response increment vector of the corresponding verification customer node. For each valid business relationship state group, calculate the mean vector and variance vector of the node-level parameter update submission response increment vector corresponding to each verified customer node in the business relationship state group. Concatenate the mean vector and variance vector to generate the corresponding state group-level true semantic response statistics vector. Input the state group-level true semantic response statistics vector into the inverse mapping network in the parameter frozen state and output the corresponding state group-level inverse residual vector. When the set of valid business relationship states is not empty, with the large language model body, low-rank adapter parameters, and inverse mapping network parameters frozen, the state feedback alignment loss is calculated based on the following formula, and the trainable parameters in the posterior state distribution calculation process are updated in reverse using the state feedback alignment loss: ； In the formula, This represents the state feedback alignment loss; Represents the set of valid business relationship states; Represents a set The number of effective business relationship status groups; Indicates the business relationship status group State group-level residual consensus vector; Indicates the business relationship status group The state group-level inverse residual vector; This indicates that the gradient operation has been stopped. This represents the square operation of the 2-norm of a vector.

10. A potential customer discovery system based on large-model semantic understanding, applied to the method described in any one of claims 1-9, characterized in that, include: The semantic extraction module is used to extract the historical text representations of customer nodes and the posterior state distribution data of the previous processing cycle, sample the customer nodes, generate group-level semantic response prototypes based on the sampled customer nodes, and use the group-level semantic response prototypes to perform text representation compensation on non-sampled customer nodes, and output the current text representation and node-level credibility status. The graph network analysis module is used to perform time-series graph message propagation based on the current text representation, node-level credibility state, and time-series graph structure, generate spatiotemporal representation data and calculate the prior state probability distribution, and generate the posterior state distribution data of the client node by solving the constrained node state optimal transmission allocation matrix. The optimization control module is used to calculate the state group-level residual consensus vector based on the posterior state distribution data and the prior state probability distribution, generate candidate parameter update vectors by combining them with the historical verification parameter update dictionary, and perform verification on the candidate parameter update vectors. The verified candidate parameter update vectors are aggregated into a unified parameter update vector. The unified parameter update vector is submitted to update the low-rank adapter parameter state. The state group-level inverse residual vector is calculated based on the text representation changes before and after updating the low-rank adapter parameter state, and the state feedback alignment loss is calculated to calibrate the trainable parameters in the graph network analysis module. The graph network analysis module is also used to regenerate the calibrated posterior state distribution data based on the calibrated trainable parameters. The customer output module is used to calculate the customer conversion probability score based on the calibrated posterior state distribution data and the current text representation, and output a list of potential customers.

Images