Agent system for automated marketing lead generation based on large language models

Abstract

The application relates to the technical field of intelligent marketing, and discloses an intelligent agent system for automatic marketing customer acquisition based on a large language model, which comprises a user forward Mogrifier LSTM module used for sequentially receiving user behavior output forward states, a user reverse Mogrifier LSTM module used for inversely receiving user behavior and necessity gate output reverse states, a commodity forward Mogrifier LSTM module used for sequentially receiving commodity state output forward states, a commodity reverse Mogrifier LSTM module used for inversely receiving commodity state and necessity gate output reverse states, a curvature coupling module used for calculating forward and reverse curvature coefficients, a joint loss module used for calculating contrast and manifold alignment losses, a decoder module used for receiving four-state output conversion probabilities, and a counterfactual intervention module used for outputting attribution heat maps. Through reverse Mogrifier and curvature coupling, the application realizes counterfactual attribution and bidirectional evolution, and is accurate in prediction and strong in interpretability.

Description

Technical Field

[0001] This invention relates to the field of intelligent marketing technology, and in particular to an intelligent agent system for automated marketing customer acquisition based on a large language model. Background Technology

[0002] With the rapid development of large language models, automated marketing and customer acquisition systems are beginning to incorporate intelligent agent architectures. By analyzing users' historical behavior sequences, product status changes, and multi-channel interaction records, these systems can predict user conversion intentions in real time and generate personalized marketing strategies. User behavior sequence modeling is the core component of such systems, and its accuracy directly impacts the conversion rate of marketing outreach.

[0003] Long Short-Term Memory (LSTM) networks and their variants are widely used in sequence modeling tasks. Mogrifier LSTM, as an improved structure of LSTM, enhances the fusion of contextual information by introducing an alternating gating interaction mechanism between the current input and the hidden state of the previous time step. However, in the specific application scenario of automated marketing customer acquisition, existing Mogrifier LSTMs have the following shortcomings: First, traditional Mogrifier LSTMs only propagate information forward in time, failing to use the final conversion result to infer the key preceding behaviors that led to the conversion, resulting in poor attribution interpretability; second, existing models only model the user's unilateral behavioral trajectory, ignoring the dynamic evolution of product-side features such as product inventory, coupon status, and content popularity, and their bidirectional coupling relationship with user behavior, while the user conversion probability in marketing customer acquisition is often determined by both states; third, the parameters of the gating interaction are fixed and cannot adapt to the information interaction needs of different users and different sequence lengths.

[0004] Therefore, there is an urgent need for an improved MogrifierLSTM solution that can integrate reverse causal inference with a user-product bidirectional coupling mechanism to enhance the predictive accuracy and attribution interpretability of automated marketing customer acquisition systems. Summary of the Invention

[0005] One objective of this invention is to propose an intelligent agent system for automated marketing customer acquisition based on a large language model. This invention fully integrates user forward and backward Mogrifier LSTM, product forward and backward Mogrifier LSTM, necessity gating mechanism, curvature coupling modulation module, joint contrast loss and manifold alignment loss, and counterfactual intervention solution algorithm. It describes in detail the entire process of automated marketing customer acquisition, from user behavior and product status data collection, dual-stream forward and backward temporal modeling, curvature coupling bidirectional evolution, necessity gating counterfactual backtracking, joint loss optimization training to attribution heatmap output. It has the advantages of high prediction accuracy, strong attribution interpretability, complete user-product coupling modeling, and adaptability to non-stationary marketing environments.

[0006] An intelligent agent system for automated marketing customer acquisition based on a large language model, according to an embodiment of the present invention, includes: The user forward Mogrifier LSTM module is used to receive user behavior sequences in chronological order and output the user's forward hidden state. The User Reverse Mogrifier LSTM module is used to receive the user behavior sequence in reverse order and output the user's reverse hidden state. This module uses a necessity gate instead of a standard forget gate, and the output value of the necessity gate is 1 minus the activation function value. The forward Mogrifier LSTM module is used to receive the product state sequence in chronological order and output the forward hidden state of the product. The Item Reverse Mogrifier LSTM module is used to receive the item state sequence in reverse order and output the item reverse hidden state. This module uses a necessity gate instead of a standard forget gate, and the output value of the necessity gate is 1 minus the activation function value. The curvature coupling modulation module is used to calculate the forward curvature coupling coefficient of the user's forward and the product's forward hidden states at each time step, and to calculate the reverse curvature coupling coefficient of the user's reverse and the product's reverse hidden states, and to modulate the reverse unit updates. The joint loss calculation module is used to calculate the contrast loss between the user's forward and reverse hidden states, the contrast loss between the product's forward and reverse hidden states, and the manifold alignment loss between the user's reverse hidden state and the product's reverse hidden state. The joint decoder module receives four hidden states and outputs the transformation prediction probability. The counterfactual intervention solution module is used to fix the transformation result label, it iteratively solves the minimum counterfactual intervention by reverse hidden state and curvature coupling coefficient, and outputs the attribution heatmap.

[0007] Optionally, the output user-forward hidden state of the user-forward Mogrifier LSTM module specifically includes: The user behavior sequence is input into the forward dynamic rounds submodule in chronological order. This submodule calculates the gate control rounds based on the sequence length and entropy value. The gate control rounds are positive integers, ranging from 1 to a preset maximum value. The input vector of the current time step and the hidden state of the previous time step are input into the user forward Mogrifier LSTM unit, along with the calculated gate control rounds. The hidden state of the previous time step is used as interaction object A, and the current input vector is used as interaction object B. The gated interaction process is repeated for the number of times equal to the number of gate control rounds. In each interaction, the result of interaction object A after linear transformation and activation function is multiplied by interaction object B, and then the result of interaction object B after linear transformation and activation function is multiplied by interaction object A. Interaction objects A and B are updated alternately. The final interaction object B obtained after all gate control round iterations is used as a candidate input and, together with the hidden state of the previous time step, is entered into the standard update formula of the Long Short-Term Memory unit to output the user forward hidden state of the current time step.

[0008] Optionally, the output user-inverse hidden state of the user-inverse Mogrifier LSTM module specifically includes: The user behavior sequence is input into the reverse dynamic round number submodule in reverse time order. The reverse dynamic round number submodule calculates the gate control round number based on the length and entropy value of the reverse sequence. The gate control round number is a positive integer, ranging from 1 to a preset maximum value. The input vector of the current reverse time step and the hidden state of the previous reverse time step are input into the user reverse MogrifierLSTM unit, along with the calculated gate control round number. The hidden state of the previous reverse time step is used as interaction object A, and the input vector of the current reverse time step is used as interaction object B. The gated interaction process is repeated for the number of times equal to the gate control round number. In each interaction, the result of interaction object A after linear transformation and activation function is multiplied by interaction object B, and then the result of interaction object B after linear transformation and activation function is multiplied by interaction object A. Interaction object A and interaction object B are updated alternately. The final interactive object B obtained after all gate control rounds of iteration is used as a candidate input. Together with the hidden state of the previous reverse time step, it enters the necessity gate update formula of the long short-term memory unit. The output value of the necessity gate is 1 minus the activation function value, which outputs the user's reverse hidden state at the current reverse time step.

[0009] Optionally, the output of the forward product hiding state of the forward product Mogrifier LSTM module is specifically: The product status time series is input into the forward dynamic wheel counting submodule in chronological order. The forward dynamic wheel counting submodule calculates the gate control wheel number based on the length of the sequence and the entropy value. The product input vector at the current time step and the product hidden state at the previous time step are input into the product forward Mogrifier LSTM unit. After the alternating gating interaction process, the product forward hidden state at the current time step is output.

[0010] Optionally, the output of the product reverse Mogrifier LSTM module is the same as the output of the user reverse Mogrifier LSTM module, except that the input data is replaced with a product state sequence.

[0011] Optionally, the use of the necessity gate to replace the standard forget gate in the commodity reverse Mogrifier LSTM module specifically includes: The output of the necessity gate is equal to 1 minus the output of the activation function, where the input of the activation function is a linear combination of the hidden state of the previous time step and the current input vector. This necessity gate controls the proportion of the cell state from the previous time step that is passed to the current time step; a larger output value indicates less forgetting. In reverse time propagation, when the final transformed label is positive, the necessity gate automatically learns to increase the output value; when the final transformed label is negative, the necessity gate automatically learns to decrease the output value.

[0012] Optionally, the curvature coupling modulation module calculates the forward curvature coupling coefficient, and the process of modulating the update of the two forward units includes: acquiring the user's forward hidden state vector and the product's forward hidden state vector at each time step; concatenating the two vectors and inputting them into a bilinear mapping layer, which outputs a scalar value as the forward curvature coupling coefficient; multiplying the forward curvature coupling coefficient by the candidate memory value in the standard update equation of the user's forward Mogrifier LSTM unit to obtain the corrected user's forward candidate memory value; simultaneously multiplying the forward curvature coupling coefficient by the candidate memory value in the standard update equation of the product's forward Mogrifier LSTM unit to obtain the corrected product's forward candidate memory value; and having each of the two forward units use their respective corrected candidate memory values ​​to complete the state update at the current time step.

[0013] Optionally, the curvature coupling modulation module calculates the inverse curvature coupling coefficient and modulates the update of two inverse units in the same way as it calculates the forward curvature coupling coefficient and modulates the update of two forward units. The user forward and product forward are replaced with user inverse and product inverse, and the standard update equation is replaced with the necessity gate update equation.

[0014] Optionally, the calculation of the comparison loss between the user's forward and reverse hidden states, the comparison loss between the product's forward and reverse hidden states, and the manifold alignment loss between the user's reverse and product's reverse hidden states by the joint loss calculation module specifically includes: The comparison loss calculation unit receives the user's positive hidden state and the user's negative hidden state, uses the user's positive hidden state as the anchor point, the user's negative hidden state as the positive example, and other hidden states in the same batch as negative examples, and outputs the user's comparison loss value. The comparison loss calculation unit receives the product's positive hidden state and the product's negative hidden state, uses the product's positive hidden state as the anchor point, the product's negative hidden state as the positive example, and other hidden states in the same batch as negative examples, and outputs the product's comparison loss value. The manifold alignment loss calculation unit receives the user's negative hidden state and the product's negative hidden state, projects both onto a unit hypersphere, calculates the negative logarithm of the dot product of the two projected vectors, and outputs the manifold alignment loss value. The total loss generation unit adds the user's comparison loss value, the product's comparison loss value, and the manifold alignment loss value, using this sum as the optimization objective for backpropagation.

[0015] Optionally, the joint decoder module receives four hidden states and outputs the transformation prediction probability, specifically including: The user's forward hidden state, user's backward hidden state, product's forward hidden state, and product's backward hidden state are concatenated end-to-end along the feature dimension to form a combined feature vector. Dimensionality reduction is performed using a two-layer fully connected network: the first layer compresses the dimension of the combined feature vector to half of the original dimension, and the second layer maps the compressed vector to scalar values. The sigmoid activation function is applied to the scalar values, outputting a conversion prediction probability between 0 and 1.

[0016] The beneficial effects of this invention are: First, by setting up user-inverse Mogrifier LSTM modules and product-inverse Mogrifier LSTM modules, and replacing the standard forget gate with a necessity gate, counterfactual backtracking reasoning capability is achieved. Given the final conversion result label, the model can infer the necessary preceding user behavior and necessary product state changes that led to the conversion, solving the problem that traditional sequence models can only predict forward but cannot attribute backward.

[0017] Second, by introducing a curvature coupling modulation module, a two-way coupling mechanism is established between the user manifold and the product manifold. The forward curvature coupling coefficient and the reverse curvature coupling coefficient at each time step modulate the state updates of the two manifolds respectively, enabling the model to simultaneously capture the interaction between user behavior and product dynamics, such as the stimulating effect of inventory decline on user conversion intention, thus improving the accuracy of conversion prediction.

[0018] Third, by using the contrast loss and manifold alignment loss in the joint loss calculation module, the consistency of the user's forward and backward states, the consistency of the product's forward and backward states, and the alignment of the user's backward and product's backward states in the manifold space are constrained, thereby enhancing the model's representational robustness across multiple time scales.

[0019] Fourth, the attribution heatmap output by the counterfactual intervention solution module presents key events on the timeline, user behavior dimension, and product attribute dimension corresponding to the minimum counterfactual intervention in a visual form, providing an intuitive and actionable decision-making basis for marketing strategy optimization. Attached Figure Description

[0020] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of the module of the intelligent agent system for automated marketing customer acquisition based on a large language model proposed in this invention. Figure 2 This is a schematic diagram illustrating the implementation process of the anti-geometric Mogrifier in the intelligent agent system for automated marketing customer acquisition based on a large language model proposed in this invention.

[0021] Figure 3 This is a schematic diagram illustrating the implementation process of the dual manifold Mogrifier of the intelligent agent system for automated marketing customer acquisition based on a large language model proposed in this invention. Detailed Implementation

[0022] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0023] refer to Figures 1-3 An intelligent agent system for automated marketing and customer acquisition based on a large language model includes: The user forward Mogrifier LSTM module is used to receive user behavior sequences in chronological order and output the user's forward hidden state. The User Reverse Mogrifier LSTM module is used to receive the user behavior sequence in reverse order and output the user's reverse hidden state. This module uses a necessity gate instead of a standard forget gate, and the output value of the necessity gate is 1 minus the activation function value. The forward Mogrifier LSTM module is used to receive the product state sequence in chronological order and output the forward hidden state of the product. The Item Reverse Mogrifier LSTM module is used to receive the item state sequence in reverse order and output the item reverse hidden state. This module uses a necessity gate instead of a standard forget gate, and the output value of the necessity gate is 1 minus the activation function value. The curvature coupling modulation module is used to calculate the forward curvature coupling coefficient of the user's forward and the product's forward hidden states at each time step, and to calculate the reverse curvature coupling coefficient of the user's reverse and the product's reverse hidden states, and to modulate the reverse unit updates. The joint loss calculation module is used to calculate the contrast loss between the user's forward and reverse hidden states, the contrast loss between the product's forward and reverse hidden states, and the manifold alignment loss between the user's reverse hidden state and the product's reverse hidden state. The joint decoder module receives four hidden states and outputs the transformation prediction probability. The counterfactual intervention solution module is used to fix the transformation result label, it iteratively solves the minimum counterfactual intervention by reverse hidden state and curvature coupling coefficient, and outputs the attribution heatmap.

[0024] In this embodiment, the output of the user-forward Mogrifier LSTM module specifically includes the user-forward hidden state: The user behavior sequence is input into the forward dynamic rounds submodule in chronological order. This submodule calculates the gate control rounds based on the sequence length and entropy value. The gate control rounds are positive integers, ranging from 1 to a preset maximum value. The input vector of the current time step and the hidden state of the previous time step are input into the user forward Mogrifier LSTM unit, along with the calculated gate control rounds. The hidden state of the previous time step is used as interaction object A, and the current input vector is used as interaction object B. The gated interaction process is repeated for the number of times equal to the number of gate control rounds. In each interaction, the result of interaction object A after linear transformation and activation function is multiplied by interaction object B, and then the result of interaction object B after linear transformation and activation function is multiplied by interaction object A. Interaction objects A and B are updated alternately. The final interaction object B obtained after all gate control round iterations is used as a candidate input and, together with the hidden state of the previous time step, is entered into the standard update formula of the Long Short-Term Memory unit to output the user forward hidden state of the current time step.

[0025] In this embodiment, the output user-inverse hidden state of the user-inverse Mogrifier LSTM module specifically includes: The user behavior sequence is input into the reverse dynamic round number submodule in reverse time order. The reverse dynamic round number submodule calculates the gate control round number based on the length and entropy value of the reverse sequence. The gate control round number is a positive integer, ranging from 1 to a preset maximum value. The input vector of the current reverse time step and the hidden state of the previous reverse time step are input into the user reverse MogrifierLSTM unit, along with the calculated gate control round number. The hidden state of the previous reverse time step is used as interaction object A, and the input vector of the current reverse time step is used as interaction object B. The gated interaction process is repeated for the number of times equal to the gate control round number. In each interaction, the result of interaction object A after linear transformation and activation function is multiplied by interaction object B, and then the result of interaction object B after linear transformation and activation function is multiplied by interaction object A. Interaction object A and interaction object B are updated alternately. The final interactive object B obtained after all gate control rounds of iteration is used as a candidate input. Together with the hidden state of the previous reverse time step, it enters the necessity gate update formula of the long short-term memory unit. The output value of the necessity gate is 1 minus the activation function value, which outputs the user's reverse hidden state at the current reverse time step.

[0026] Specifically, the reverse dynamic round number submodule calculates the gate control round number based on the length and entropy value of the reverse sequence as follows: Obtain the total length of the reverse sequence; calculate the frequency of different behavior types in the reverse sequence, and calculate the Shannon entropy of the sequence based on the frequency distribution; concatenate the length and entropy values ​​into a two-dimensional feature vector; input this feature vector into a two-layer fully connected network, where the first layer maps the two-dimensional features to 16-dimensional features and activates them through linear rectification, and the second layer maps the 16-dimensional features to a preset maximum number of output nodes; apply a differentiable Gibberian softened maximum sampling operation to the output nodes, and sample an integer from it as the number of gated rounds; the number of gated rounds ranges from 1 to the preset maximum value.

[0027] In this embodiment, the output of the product forward Mogrifier LSTM module, specifically the product forward hidden state, is as follows: The product status time series is input into the forward dynamic wheel counting submodule in chronological order. The forward dynamic wheel counting submodule calculates the gate control wheel number based on the length of the sequence and the entropy value. The product input vector at the current time step and the product hidden state at the previous time step are input into the product forward Mogrifier LSTM unit. After the alternating gating interaction process, the product forward hidden state at the current time step is output.

[0028] In this embodiment, the output of the product reverse Mogrifier LSTM module is the same as the output of the user reverse Mogrifier LSTM module, except that the input data is replaced with a product state sequence.

[0029] In this embodiment, the necessity gate used in the product reverse Mogrifier LSTM module to replace the standard forget gate specifically includes: The output of the necessity gate is equal to 1 minus the output of the activation function, where the input of the activation function is a linear combination of the hidden state of the previous time step and the current input vector. This necessity gate controls the proportion of the cell state from the previous time step that is passed to the current time step; a larger output value indicates less forgetting. In reverse time propagation, when the final transformed label is positive, the necessity gate automatically learns to increase the output value; when the final transformed label is negative, the necessity gate automatically learns to decrease the output value.

[0030] Specifically, the proportion of the unit state from the previous time step that is passed to the current time step controlled by the necessity gate is: The necessity gate receives a linear combination of the hidden state vector from the previous reverse time step and the input vector from the current reverse time step as input. This linear combination is passed through a sigmoid activation function to obtain an intermediate value between 0 and 1. The final output value of the necessity gate is equal to 1 minus this intermediate value. The final output value is multiplied element-wise with the unit state vector from the previous reverse time step to obtain the retained historical state component. The retained historical state component is added to the candidate state vector from the current reverse time step to obtain the updated unit state vector from the current reverse time step. The closer the final output value is to 1, the more historical information is retained; the closer it is to 0, the more historical information is forgotten.

[0031] In this embodiment, the curvature coupling modulation module calculates the forward curvature coupling coefficient, and the process of modulating the update of the two forward units includes: acquiring the user's forward hidden state vector and the product's forward hidden state vector at each time step; concatenating the two vectors and inputting them into a bilinear mapping layer, which outputs a scalar value as the forward curvature coupling coefficient; multiplying the forward curvature coupling coefficient by the candidate memory value in the standard update equation of the user's forward Mogrifier LSTM unit to obtain the corrected user's forward candidate memory value; simultaneously multiplying the forward curvature coupling coefficient by the candidate memory value in the standard update equation of the product's forward Mogrifier LSTM unit to obtain the corrected product's forward candidate memory value; and having each of the two forward units use their respective corrected candidate memory values ​​to complete the state update at the current time step.

[0032] Specifically, the concatenation of two vectors is input into a bilinear mapping layer, and the bilinear mapping layer outputs a scalar value, which serves as the positive curvature coupling coefficient. The user's positive hidden state vector and the product's positive hidden state vector are concatenated end to end to form a joint vector; the dimension of the joint vector is twice the dimension of the hidden state; the bilinear mapping layer contains a learnable weight matrix and a bias scalar; the joint vector is first subjected to a linear transformation of the weight matrix, the size of which is 1 times the dimension of the joint vector, and the output is a scalar; this scalar is then added to the bias scalar and input into a sigmoid activation function to obtain an output value between 0 and 1; this output value is the positive curvature coupling coefficient.

[0033] In this embodiment, the curvature coupling modulation module calculates the inverse curvature coupling coefficient and modulates the update of the two inverse units in the same way as calculating the forward curvature coupling coefficient and modulating the update of the two forward units. The user forward and product forward are replaced with user inverse and product inverse, and the standard update equation is replaced with the necessity gate update equation.

[0034] In this embodiment, the joint loss calculation module calculates the comparison loss between the user's forward and reverse hidden states, the comparison loss between the product's forward and reverse hidden states, and the manifold alignment loss between the user's reverse hidden state and the product's reverse hidden state, specifically including: The comparison loss calculation unit receives the user's positive hidden state and the user's negative hidden state, uses the user's positive hidden state as the anchor point, the user's negative hidden state as the positive example, and other hidden states in the same batch as negative examples, and outputs the user's comparison loss value. The comparison loss calculation unit receives the product's positive hidden state and the product's negative hidden state, uses the product's positive hidden state as the anchor point, the product's negative hidden state as the positive example, and other hidden states in the same batch as negative examples, and outputs the product's comparison loss value. The manifold alignment loss calculation unit receives the user's negative hidden state and the product's negative hidden state, projects both onto a unit hypersphere, calculates the negative logarithm of the dot product of the two projected vectors, and outputs the manifold alignment loss value. The total loss generation unit adds the user's comparison loss value, the product's comparison loss value, and the manifold alignment loss value, using this sum as the optimization objective for backpropagation.

[0035] The process of outputting the user's comparative loss value, using the user's positive hidden state as the anchor point, the user's negative hidden state as a positive example, and other hidden states within the same batch as negative examples, specifically includes the following steps: Calculate the cosine similarity between the user's forward hidden state and the user's backward hidden state, and use this cosine similarity as the numerator; calculate the sum of the cosine similarities between the user's forward hidden state and the backward hidden states of all other users in the batch, and use this sum as the denominator; divide the numerator by the denominator, take the natural logarithm, and then take the negative value to obtain the contrast loss of a single sample; calculate the arithmetic mean of the contrast losses of all samples in the batch, and output the user's contrast loss value.

[0036] The specific steps for calculating the negative logarithm of the dot product of the two projected vectors and outputting the manifold alignment loss value include the following: Divide the user's inverse hidden state vector by its magnitude and project it onto the unit hypersphere to obtain the first projection vector; divide the product's inverse hidden state vector by its magnitude and project it onto the unit hypersphere to obtain the second projection vector; calculate the dot product of the first and second projection vectors, with the dot product value ranging from -1 to +1; take the logarithm of the dot product value to the base of the natural constant and then take the negative value to obtain the manifold alignment loss value of a single sample; calculate the arithmetic mean of the manifold alignment loss values ​​of all samples in the batch and output the final manifold alignment loss value.

[0037] In this embodiment, the joint decoder module receives four hidden states and outputs the transformation prediction probability, specifically including: The user's forward hidden state, user's backward hidden state, product's forward hidden state, and product's backward hidden state are concatenated end-to-end along the feature dimension to form a combined feature vector. Dimensionality reduction is performed using a two-layer fully connected network: the first layer compresses the dimension of the combined feature vector to half of the original dimension, and the second layer maps the compressed vector to scalar values. The sigmoid activation function is applied to the scalar values, outputting a conversion prediction probability between 0 and 1.

[0038] The process of concatenating the user's forward hidden state, user's reverse hidden state, product's forward hidden state, and product's reverse hidden state along the feature dimension to form a combined feature vector includes the following steps: Obtain the user's forward hidden state vector, whose dimension is the same as the hidden layer dimension; obtain the user's backward hidden state vector, with the same dimension; obtain the product's forward hidden state vector, with the same dimension; obtain the product's backward hidden state vector, with the same dimension; in a fixed order of user forward, user backward, product forward, product backward, concatenate the four vectors end to end. The dimension of the new vector is four times the hidden layer dimension. This new vector is the combined feature vector.

[0039] In this embodiment, the specific steps for outputting the attribution heatmap by the counterfactual intervention solution module are as follows: The conversion result label is fixed as positive; the user reverse hidden state sequence and the product reverse hidden state sequence are used as initial variables, and the curvature coupling coefficient is used as a constraint condition; The values ​​in the inverse hidden state sequence are iteratively adjusted using gradient descent to find a set of inverse hidden states with minimum modification. The minimum modification satisfies the condition that the conversion prediction probability output by the joint decoder is still positive under the modified inverse hidden state. The modification is mapped back to the original user behavior space and product attribute space according to time step and feature dimension to obtain the counterfactual contribution score of each user behavior feature and each product attribute feature at each time step. The counterfactual contribution scores are plotted on a two-dimensional plane using varying shades of color, with the horizontal axis representing time steps and the vertical axis representing user behavior and product attribute dimensions, generating an attribution heatmap.

[0040] Example 1: This invention was deployed in an automated marketing intelligence system on an e-commerce platform to address the problems of traditional marketing models failing to accurately pinpoint key user conversion behaviors and neglecting dynamic changes in product status, leading to attribution bias. This e-commerce platform generates massive amounts of user browsing, click, and add-to-cart behavior data daily, while inventory changes and promotional activity statuses are updated in real time. Traditional models rely solely on positive user behavior sequences; when a user ultimately completes a purchase, they cannot trace back to determine which click or change in product attribute played a decisive role.

[0041] In this scenario, the system collects the user's behavior sequence over the past 30 days in real time, including events such as browsing product detail pages, adding items to favorites, adding items to the shopping cart, and claiming coupons. Simultaneously, it collects the corresponding product status sequence, including inventory quantity, price changes, remaining coupons, and sentiment ratings in the review section. The user behavior sequence is input into the user's forward Mogrifier LSTM module in chronological order, and simultaneously into the user's backward Mogrifier LSTM module in reverse order. The product status sequence is also input bidirectionally into the product's forward and backward modules. In the backward module, the necessity gate replaces the standard forget gate; the gate's output value is equal to 1 minus the sigmoid activation function value. This allows the model to backward retain essential prior information for conversion, even when the purchase outcome is known.

[0042] The curvature coupling modulation module calculates the curvature coefficient between the user's forward and the product's forward hidden states at each time step. This coefficient, a real number between 0 and 1, is used to amplify or reduce the candidate memory values ​​of the two manifolds. When a user browses a product and its inventory is less than 10 units, the curvature coefficient automatically increases, strengthening the mutual influence between the two states. The inverse curvature coefficient is used for coupling modulation during the reverse propagation process. The joint loss calculation module simultaneously optimizes three objectives: consistency comparison loss between the user's forward and inverse states, consistency comparison loss between the product's forward and inverse states, and manifold alignment loss between the user's inverse and product inverse states on a unit hypersphere. The joint decoder concatenates the four hidden states and outputs the transformation probability through a two-layer fully connected network. The counterfactual intervention solution module fixes the transformation label at 1 and iteratively searches for the minimum modification amount through gradient descent, ensuring that the modified input produces the same result as the original positive sample in the inverse flow, outputting an attribution heatmap.

[0043] To verify the effectiveness of this invention, 60 consecutive days of real marketing data from the platform were selected, and time-series cross-validation was used to compare it with three mainstream baseline models. The baseline models included standard LSTM, standard MogrifierLSTM, and an ablation model without backflow but preserving user-product coupling. Evaluation metrics included AUC, precision, recall, F1 score, and attribution accuracy. Attribution accuracy was validated using key events from 1000 manually labeled conversion samples. The labeling team consisted of marketing experts who independently judged the top three key behaviors or product status changes that led to the conversion. The model's average inference time (milliseconds / sample) and the number of training epochs required for convergence were also recorded. The experimental results are shown in the table below.

[0044] Table 1: Performance Comparison of Various Models in Automated Marketing Customer Acquisition Tasks

[0045] As shown in the table above, the standard LSTM has an AUC of only 0.723, precision of 0.684, recall of 0.657, F1 score of 0.670, and attribution accuracy of only 0.512, indicating that traditional LSTM performs poorly in high-noise sequence tasks such as marketing and customer acquisition. While the standard Mogrifier LSTM introduces alternating gating interactions between input and hidden states, improving the AUC to 0.751, the F1 score to 0.705, and the attribution accuracy to 0.568, it still cannot solve the problem of reverse attribution and user-product coupling. The ablation model removes the reverse flow but retains curvature coupling, achieving an AUC of 0.784, an F1 score of 0.737, and an attribution accuracy of 0.623. This shows that simple bidirectional coupling helps with prediction, but the lack of necessary gates in counterfactual backtracking limits its attribution ability. This invention achieved optimal values ​​across all metrics, with an AUC of 0.831, 0.047 higher than the ablation model and 0.080 higher than the standard Mogrifier; precision broke the 0.8 mark for the first time; recall was 0.793, 0.064 higher than the ablation model; and F1 score was 0.797, 0.060 higher than the ablation model. The most significant improvement was in attribution accuracy, reaching 0.761, an increase of 0.138 compared to the ablation model and 0.193 compared to the standard Mogrifier. This means that in 1000 manually labeled samples, this invention correctly identified 761 key events, while the standard Mogrifier could only identify 568. Regarding inference time, this invention achieved 18.9 milliseconds per sample, which, while higher than the standard LSTM's 8.3 milliseconds, is still far below the 50 millisecond limit typically required for real-time marketing decisions, demonstrating its practical value. Training convergence requires 52 rounds, 14 more than the standard Mogrifier's 38 rounds. This is due to the increased optimization complexity caused by the reverse flow and curvature coupling, but it results in a significant improvement in prediction accuracy and attribution ability. Further analysis revealed that on a subset of inventory-sensitive products, this invention achieved an attribution recall of 0.84 for events such as "inventory decline triggering conversion," indicating that the curvature coupling mechanism effectively captures the dynamic interaction between users and products. On samples where users remained silent for more than 7 days, this invention maintained a recall of 0.71 thanks to the reverse necessity gate, while the standard Mogrifier only achieved 0.53, a decrease of 0.18. This demonstrates that counterfactual backtracking capability has a significant advantage in modeling sparse behavioral sequences. The marketing team evaluated the operability of the attribution heatmap output by the counterfactual intervention solution module. This invention received an average score of 4.2 out of 5, while the heatmap given by the traditional attention mechanism only scored 2.8. This indicates that the attribution results output by this invention are more consistent with actual business practices and can directly guide operations personnel in formulating intervention strategies. Based on the above data, the present invention outperforms existing technologies in terms of prediction accuracy, attribution interpretability, and robustness to sparse sequences.

[0046] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An intelligent agent system for automated marketing customer acquisition based on a large language model, characterized in that: include: The user forward Mogrifier LSTM module is used to receive user behavior sequences in chronological order and output the user forward hidden state. The User Reverse Mogrifier LSTM module is used to receive the user behavior sequence in reverse order and output the user's reverse hidden state. This module uses the necessity gate instead of the standard forget gate, and the output value of the necessity gate is 1 minus the activation function value. The forward Mogrifier LSTM module is used to receive the product state sequence in chronological order and output the forward hidden state of the product. The Item Reverse Mogrifier LSTM module is used to receive the item state sequence in reverse order and output the item reverse hidden state. This module uses the necessity gate instead of the standard forget gate. The output value of the necessity gate is 1 minus the activation function value. The curvature coupling modulation module is used to calculate the forward curvature coupling coefficient of the user's forward and the product's forward hidden states at each time step, and to calculate the reverse curvature coupling coefficient of the user's reverse and the product's reverse hidden states, and to modulate the reverse unit updates. The joint loss calculation module is used to calculate the contrast loss between the user's forward and reverse hidden states, the contrast loss between the product's forward and reverse hidden states, and the manifold alignment loss between the user's reverse hidden state and the product's reverse hidden state. The joint decoder module receives four hidden states and outputs the transformation prediction probability. The counterfactual intervention solution module is used to fix the transformation result label, it iteratively solves the minimum counterfactual intervention by reverse hidden state and curvature coupling coefficient, and outputs the attribution heatmap.

2. The intelligent agent system for automated marketing customer acquisition based on a large language model according to claim 1, characterized in that, The user-forward hidden state output by the user-forward Mogrifier LSTM module specifically includes: The user behavior sequence is input into the forward dynamic rounds submodule in chronological order. This submodule calculates the gate control rounds based on the sequence length and entropy value. The gate control rounds are positive integers, ranging from 1 to a preset maximum value. The input vector of the current time step and the hidden state of the previous time step are input into the user forward Mogrifier LSTM unit, along with the calculated gate control rounds. The hidden state of the previous time step is used as interaction object A, and the current input vector is used as interaction object B. The gate control interaction process is repeated for a number of times equal to the gate control rounds. The final interaction object B obtained after all gate control round iterations is used as a candidate input and, together with the hidden state of the previous time step, is entered into the standard update formula of the Long Short-Term Memory unit. The user forward hidden state of the current time step is then output.

3. The intelligent agent system for automated marketing customer acquisition based on a large language model according to claim 1, characterized in that, The user-inverse hidden state output by the user-inverse Mogrifier LSTM module specifically includes: The user behavior sequence is input into the reverse dynamic round number submodule in reverse time order. The reverse dynamic round number submodule calculates the gate control round number based on the length and entropy value of the reverse sequence. The gate control round number is a positive integer, ranging from 1 to a preset maximum value. The input vector of the current reverse time step and the hidden state of the previous reverse time step are input into the user reverse MogrifierLSTM unit, along with the calculated gate control round number. The hidden state of the previous reverse time step is used as interaction object A, and the input vector of the current reverse time step is used as interaction object B. The gate control interaction process is repeated. The final interaction object B obtained after all gate control round iterations is used as a candidate input, which, together with the hidden state of the previous reverse time step, enters the necessary gate update formula of the Long Short-Term Memory unit.

4. The intelligent agent system for automated marketing customer acquisition based on a large language model according to claim 1, characterized in that, The output of the product forward Mogrifier LSTM module is specifically the product forward hidden state: The product status time series is input into the forward dynamic wheel counting submodule in chronological order. The forward dynamic wheel counting submodule calculates the gate control wheel number based on the length of the sequence and the entropy value. The product input vector at the current time step and the product hidden state at the previous time step are input into the product forward MogrifierLSTM unit. After the alternating gating interaction process, the product forward hidden state at the current time step is output.

5. The intelligent agent system for automated marketing customer acquisition based on a large language model according to claim 1, characterized in that, The output of the product reverse Mogrifier LSTM module is the same as the output of the user reverse Mogrifier LSTM module, except that the input data is replaced with a product state sequence.

6. The intelligent agent system for automated marketing customer acquisition based on a large language model according to claim 1, characterized in that, The use of the necessity gate to replace the standard forget gate in the product reverse Mogrifier LSTM module specifically includes: the necessity gate controls the proportion of the cell state from the previous time step that is passed to the current time step, and the larger the output value, the less is forgotten.

7. The intelligent agent system for automated marketing customer acquisition based on a large language model according to claim 1, characterized in that, The curvature coupling modulation module calculates the forward curvature coupling coefficient. The process of modulating the update of the two forward units includes: acquiring the user's forward hidden state vector and the product's forward hidden state vector at each time step; concatenating the two vectors and inputting them into a bilinear mapping layer, which outputs a scalar value as the forward curvature coupling coefficient; multiplying the forward curvature coupling coefficient by the candidate memory value in the standard update equation of the user's forward Mogrifier LSTM unit to obtain the corrected user's forward candidate memory value; simultaneously multiplying the forward curvature coupling coefficient by the candidate memory value in the standard update equation of the product's forward Mogrifier LSTM unit to obtain the corrected product's forward candidate memory value; and the two forward units respectively use their corrected candidate memory values ​​to complete the state update of the current time step.

8. The intelligent agent system for automated marketing customer acquisition based on a large language model according to claim 1, characterized in that, The curvature coupling modulation module calculates the inverse curvature coupling coefficient and modulates the update of two inverse units in the same way as it calculates the forward curvature coupling coefficient and modulates the update of two forward units. The user forward and product forward are replaced with user inverse and product inverse, and the standard update equation is replaced with the necessity gate update equation.

9. The intelligent agent system for automated marketing customer acquisition based on a large language model according to claim 1, characterized in that, The joint loss calculation module calculates the comparison loss between the user's forward and reverse hidden states, the comparison loss between the product's forward and reverse hidden states, and the manifold alignment loss between the user's reverse hidden state and the product's reverse hidden state, specifically including: The comparison loss calculation unit receives the user's positive hidden state and the user's negative hidden state, uses the user's positive hidden state as the anchor point, the user's negative hidden state as the positive example, and other hidden states in the same batch as negative examples, and outputs the user's comparison loss value. The comparison loss calculation unit receives the product's positive hidden state and the product's negative hidden state, uses the product's positive hidden state as the anchor point, the product's negative hidden state as the positive example, and other hidden states in the same batch as negative examples, and outputs the product's comparison loss value. The manifold alignment loss calculation unit receives the user's negative hidden state and the product's negative hidden state, projects both onto a unit hypersphere, calculates the negative logarithm of the dot product of the two projected vectors, and outputs the manifold alignment loss value. The total loss generation unit adds the user's comparison loss value, the product's comparison loss value, and the manifold alignment loss value as the optimization target for backpropagation.

10. The intelligent agent system for automated marketing customer acquisition based on a large language model according to claim 1, characterized in that, The joint decoder module receives four hidden states and outputs the transformation prediction probability, specifically including: The user's forward hidden state, user's backward hidden state, product's forward hidden state, and product's backward hidden state are concatenated end-to-end along the feature dimension to form a combined feature vector. A two-layer fully connected network is used for dimensionality reduction. The first layer compresses the dimension of the combined feature vector to half of the original dimension, and the second layer maps the compressed vector to scalar values. The sigmoid activation function is applied to the scalar values, and the output is the conversion prediction probability between 0 and 1.

Images