Digital marketing analysis and management system, method and device and storage medium

Abstract

The invention relates to the field of marketing analysis and management, and discloses a digital marketing analysis and management system, method and device and a storage medium, and the system comprises a general modeling module which is used for obtaining a general user behavior prediction model based on cross-domain user behavior data training, and outputting a corresponding model parameter set; the target domain data processing module is used for collecting and preprocessing user behavior data in a target marketing scene and completing structural alignment of input features and the universal user behavior prediction model; and the fine tuning control module is used for setting a parameter freezing and unfreezing strategy for the universal user behavior prediction model and defining a part of network layers needing to be updated. According to the method, the target domain user behavior data and the general behavior prediction model are combined, and the model network layer is selectively unfrozen and trained in the fine tuning process, so that the rapid adaptability of the model in a new scene is ensured, and the prediction accuracy of the model in a specific marketing scene is improved.

Description

Technical Field

[0001] The present invention relates to the field of marketing analysis and management, and particularly to digital marketing analysis and management systems, methods, devices, and storage media. Background Art

[0002] With the rapid development of Internet technology and big data analysis capabilities, the digital transformation of enterprises has become an important direction to promote business model innovation and improve operational efficiency. Especially in the field of marketing, achieving precise reach and personalized recommendations through data-driven approaches has gradually become the mainstream trend. Enterprises increasingly rely on in-depth mining and prediction of user behavior to shift from a "shotgun" marketing approach to an efficient "user-centered" marketing strategy.

[0003] Currently, mainstream digital marketing systems generally have the ability to collect multi-channel data and can integrate consumer behavior data from multiple sources such as e-commerce platforms, social media, and content distribution platforms. These systems usually model user browsing records, purchase behaviors, interaction feedback, etc. data to build user profiles and perform behavior predictions, thereby supporting core marketing tasks such as recommendation delivery, precise targeting, and personalized operations.

[0004] However, in practical applications, existing systems mostly rely on traditional machine learning models such as logistic regression, decision trees, and random forests. Although such models have certain interpretability and stability in static data analysis, they have obvious limitations in the face of increasingly complex user behavior patterns and rapidly changing market environments. Especially in the context of multi-scenario, multi-platform, and cross-domain data fusion, traditional models are difficult to effectively capture deep semantic information and lack end-to-end generalization capabilities. In addition, due to the lack of an efficient model update and migration mechanism, when market trends change or business scenarios switch, the model often cannot respond quickly, resulting in lagging marketing strategies, declining recommendation effects, and ultimately affecting conversion rates and user experiences.

[0005] To address the above problems, some studies have attempted to introduce deep learning models to improve behavior prediction capabilities and modeling accuracy. However, deep models generally suffer from high training costs and large migration difficulties. Especially in multi-scenario marketing systems, it is often necessary to train models separately for each business scenario, resulting in low model reuse efficiency and high operation and maintenance costs, making it difficult to meet the actual business needs of rapid deployment and continuous optimization. Summary of the Invention

[0006] To make up for the above deficiencies, the present invention provides a digital marketing analysis and management system, aiming to improve the technical problems of low model migration efficiency, poor scenario adaptability, and unstable prediction results in existing digital marketing systems.

[0007] In a first aspect, the present invention provides the following technical solution: a digital marketing analysis and management system, the system comprising: A general modeling module, configured to train a general user behavior prediction model based on cross-domain user behavior data and output a corresponding model parameter set; A target domain data processing module, configured to collect and preprocess user behavior data in a target marketing scenario and complete the structural alignment between input features and the general user behavior prediction model; A fine-tuning control module, configured to set parameter freezing and thawing strategies for the general user behavior prediction model and define partial network layers that need to be updated; A transfer training module, configured to perform fine-tuning training on the network layer parameters set to be trainable using the user behavior data in the target marketing scenario; A model evaluation module, configured to evaluate the behavior prediction performance and output stability of the fine-tuned model in the target domain. If the evaluation fails to meet the standard, adjust the fine-tuning strategy and retrain.

[0008] Preferably, the parameter freezing and thawing strategy set by the fine-tuning control module includes a thawing timing function set layer by layer, and this function is used to control whether each network layer participates in parameter update according to the number of training rounds.

[0009] Preferably, the transfer training module constructs a loss function based on target domain samples, and the loss function includes a prediction error term and a regularization term, and the regularization term is used to limit the difference between the new parameters and the original parameters of the general model during the fine-tuning process.

[0010] Preferably, the user behavior prediction model constructed by the general modeling module is a multi-layer neural network model, with the input being a user behavior feature vector and the output being the probability value of the corresponding behavior occurring.

[0011] Preferably, the model evaluation module is used to calculate the difference in prediction probability distributions before and after fine-tuning, and determine whether the behavior prediction of the model in the target scenario is stable by setting an offset threshold.

[0012] Preferably, the behavior prediction difference calculated by the model evaluation module is implemented based on the KL divergence, and the KL divergence is used to measure the distance between the model output probability distributions before and after fine-tuning.

[0013] Preferably, when the evaluation result of the model evaluation module does not meet the requirements, it automatically rolls back to the fine-tuning control module and resets the freezing and thawing strategies.

[0014] In a second aspect, the present invention provides the following technical solution: a digital marketing analysis and management method, including the following steps: S1. Train a general behavior prediction model based on historical cross-domain user behavior data; S2. Collect user behavior data in the target marketing scenario, standardize it, and align it with the input structure. S3. Set the network layers with trainable parameters according to the preset freezing and thawing strategies. S4. Perform transfer fine-tuning training on the selected network layers in the general model using the target scenario data. S5. Evaluate the behavior prediction accuracy and output stability of the fine-tuned model in the target marketing scenario. If the standard is met, push it. S6. When the evaluation index does not reach the preset threshold, return to step S3.

[0015] In a third aspect, the present invention provides the following technical solution. A computer device includes a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, it implements the above digital marketing analysis and management method.

[0016] In a fourth aspect, the present invention provides the following technical solution. A readable storage medium stores a computer program, and when the computer program is executed by a processor, it implements the above digital marketing analysis and management method.

[0017] The present invention has the following beneficial effects: 1. In the present invention, by combining the user behavior data in the target domain with the general behavior prediction model, selectively unfreezing and training the model network layers during the fine-tuning process ensures the rapid adaptability of the model in the new scenario, improves the prediction accuracy of the model in the specific marketing scenario, and enables the system to make more accurate behavior predictions based on the target user behavior data.

[0018] 2. In the present invention, by setting the freezing and thawing strategies and combining the KL divergence to measure the difference in the output probability distribution of the model before and after fine-tuning, the present invention ensures that the fine-tuned model can maintain stability in the target scenario and at the same time has a certain degree of flexibility to adapt to changes in different market environments.

[0019] 3. In the present invention, by setting the network layers with trainable parameters and the thawing timing function set layer by layer, the fine-tuning process is made more efficient. During the calculation process, the training status of each network layer can be flexibly adjusted according to the training progress, thereby reducing unnecessary calculations, improving the efficiency of the fine-tuning process, and reducing the consumption of system computing resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is the system architecture diagram of the digital marketing analysis and management system proposed by the present invention; Figure 2 It is the method flow diagram of the digital marketing analysis and management method proposed by the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0021] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0022] Embodiment 1: Referring to Figure 1 , in the first embodiment of the present invention, the present invention provides a digital marketing analysis and management system, which includes: A general modeling module that trains a general user behavior prediction model based on cross-domain user behavior data and outputs a corresponding model parameter set; the user behavior prediction model constructed by the general modeling module is a multi-layer neural network model, with the input being the user behavior feature vector and the output being the probability value of the corresponding behavior occurrence.

[0023] Specifically, first, it receives and processes user behavior data from multiple business domains. These data usually include user click records, purchase behaviors, search keywords, stay time, etc. on e-commerce platforms, and may also include behavior characteristics in social platforms and content recommendation systems. Due to significant differences in sources and structures, unified preprocessing is required. Here, the feature sources and structures are not limited. For each type of feature, the module uses the standardized Z-score method to transform all input features to a unified scale.

[0024] The model body adopts a multi-layer perceptron (MLP) architecture, that is, a typical feedforward neural network. Each layer of neurons obtains the output through weighted summation and non-linear transformation. Assume that the input of the th layer is , and the output is , and its calculation expression is: ; where is the weight matrix, is the bias vector, is the non-linear activation function, and the ReLU function is selected. The input of the first layer is the user feature vector , and the final output is the behavior probability prediction value .

[0025] During the model training process, cross-entropy is used as the loss function to measure the difference between the predicted probability and the actual label. For a sample, if the true behavior label is , and the predicted value is , the cross-entropy loss is: ; The loss of the entire training set is aggregated by sample averaging or weighted averaging and used as the object to be minimized in the objective function. The model parameters are updated using the Adam optimizer, which has the ability to adaptively adjust the learning rate. Its core update rules are as follows: ; where is the first moment estimate of the gradient, is the second moment estimate, is the base learning rate, is a constant to prevent division by zero. This algorithm can converge quickly in the early stage of training and maintain stable parameter updates in the later stage.

[0026] To prevent overfitting, L2 regularization is introduced during training. The sum of the squares of all weights is added as an additional loss term to the total loss function: ; where is the regularization factor used to control the model complexity. In addition, the Dropout technique is applied to some layers, randomly discarding a certain proportion of neuron outputs during the training phase to enhance the generalization ability of the model. After the model training is completed, the obtained parameter set (i.e., the of each layer in the network) is saved as the general modeling output, that is, the pre-trained model.

[0027] This model can be directly applied to the target task or fine-tuned on the target task. During the fine-tuning phase, generally only the parameters of the last several layers are adjusted, and the previous layers are frozen to retain the general representation ability. The process still uses the cross-entropy loss and the Adam optimizer, except that the training samples come from the specific marketing business domain.

[0028] The target domain data processing module is used to collect and preprocess the user behavior data in the target marketing scenario and complete the structural alignment between the input features and the general user behavior prediction model; Specifically, the target domain data processing module is a key link in connecting the general modeling model and the actual marketing scenario in the present invention. The tasks it undertakes are not only data collection and preliminary cleaning, but also the responsibility of model structure alignment, enabling the data from specific scenarios to be seamlessly connected to the general user behavior prediction model and achieving high adaptability for cross-scenario applications.

[0029] First, during the actual deployment process, the system pulls the latest user behavior data from the front-end system of the target domain. These data sources include, but are not limited to: user browsing product records, adding to cart behavior, click advertisement timestamps, page stay duration, keyword search records, etc. All these data may come from different sources, have different formats, and inconsistent field names. The target domain data processing module first performs field mapping, corresponding the original fields in the target scenario with the standard fields required for the input of the general model one by one. For example, the field named item_click_time in the target domain will be mapped to the standard click_timestamp field in the general model.

[0030] Next, it enters the cleaning process. The module automatically eliminates samples with severe missing values and clips extreme values. For example, if the page stay time exceeds 10 hours, it is very likely to be abnormal and will be directly set to the maximum threshold or excluded. At the same time, it also processes data with incorrect formats, such as converting the string-formatted timestamp "2025-04-0712:05:00" to a unified Unix time. During this process, it also performs re-encoding of categorical features. For example, "ios", "android", "web" are unified into three category indices of 0, 1, 2. The encoding method is kept consistent with the general model to avoid dimensional shift when the model reads the input vector.

[0031] A more core step is feature alignment. Since the input of the general model is based on a certain fixed structure, such as a floating-point feature vector with a length of 128. Each behavior data in the target domain must be mapped to this structure, and the order, data type, and position are exactly the same. This part is completed by the feature construction sub-module. The module internally maintains a set of template files, recording the position and value-taking rules of each input feature in the vector. For example, the position 5 corresponds to "the number of clicks in the past 7 days". The module will pull the click behavior records of this user in the past 7 days, count them, and write them into this position. If the data is missing, the default value is filled in or mean imputation is used. Each user sample goes through this set of processes, and finally a structured and aligned feature vector is formed. 。

[0032] During this process, the module also has a lightweight rule engine built in. It is used to dynamically detect the data drift situation in the target domain. For example, if the distribution of a certain type of behavior shows an obvious shift (such as the product click ratio drops from 70% to 20%), the module will automatically issue a warning and record the importance weight of this feature for reference during subsequent model fine-tuning. This mechanism avoids a sharp drop in the performance of the general model when facing new scenarios that deviate from the training data distribution.

[0033] After alignment with the general model, these feature vectors can be directly input into the general user behavior prediction model to obtain the behavior probability prediction result without retraining the model. A significant advantage of this structure alignment mechanism is to reduce the cold start cost. In the traditional approach, the model structure needs to be redesigned and retrained every time a new business scenario is entered. However, with the target domain data processing module of the present invention, the model structure can be reused, and only the data form on the input side needs to be processed, greatly accelerating the model deployment speed.

[0034] The fine-tuning control module is used to set the parameter freezing and thawing strategies for the general user behavior prediction model and define the partial network layers that need to be updated; the parameter freezing and thawing strategies set by the fine-tuning control module include the thawing timing function set layer by layer, and this function is used to control whether each network layer participates in parameter update according to the number of training rounds.

[0035] Specifically, the fine-tuning control module is the core link connecting the general model and the target scenario adaptability in the present invention. Its main task is to flexibly control the parameter update status of each layer of the neural network during the model migration process. Specifically, it is responsible for setting which layers are frozen, which layers are thawed at specific times, and which layers remain in training. This process is not statically preset, but is driven by a set of dynamic policy functions, and can be judged and adjusted according to the training progress, loss change, and even gradient fluctuation.

[0036] In the module initialization stage, first perform a structure mapping on the network structure of the general model.

[0037] The freezing operation is implemented through the underlying interface of the framework, such as in PyTorch: “for param in layer.parameters(): param.requires_grad = False”; The freezing here is not simply skipped, but the gradient flow to the optimizer is prohibited to ensure that the model retains the cross-domain feature expression learned in the general stage.

[0038] The key to the thawing strategy lies in the design of the “timing function”. This function is not a fixed value, nor a linear increment, but a layer-by-layer thawing timing function with non-linear response . Among them, represents the layer index of the network, represents the current number of training rounds. The function output is a boolean value, indicating whether the th layer is allowed to thaw in the current round. For example: ; Among them is a hyperparameter for adjusting curvature and determines the overall thawing rhythm. For shallow layers, it may be able to participate in training in the 3rd to 5th rounds; for deep layers, such as the bottom Embedding layer or feature extraction layer, it may not be thawed until the 10th round or even frozen throughout the process. This function supports hot plugging, and developers can dynamically insert new control strategies, such as linking with the validation set loss curve or triggering emergency thawing when the training loss oscillates for a long time.

[0039] Furthermore, the fine-tuning control module supports thawing by group. That is to say, instead of thawing layer by layer, the network is divided into multiple logical groups, such as "front feature layer", "middle abstraction layer", and "output discrimination layer". The layers within each group share the thawing rhythm. Define a group thawing function: ; where is the group index, is the starting round of thawing for this group. The advantage of this design is that it is convenient to regulate medium and large models, and it will not cause global instability due to local layer perturbations.

[0040] During the operation of the module, before the start of each round of training, a hierarchical scanning function will be called to determine whether the current state of each layer allows updates. If it is thawed, the requires_grad attribute of the corresponding layer's parameters will be reset to True. The module can also record the loss reduction effect brought by each thawing and automatically learn the thawing strategy for the next round. Combined with the early stopping strategy, when the validation set accuracy tends to be stable, the thawing process can be automatically terminated to prevent the model from overfitting the target scenario data.

[0041] In the specific implementation, developers can set the initial frozen state for each layer based on the modular API of the framework, call the thawing function with the dynamic scheduler, and bind the optimizer to update the parameters of the thawed layers. All control logics can be encapsulated into an independent Python class or module, which has good reusability and maintainability.

[0042] The transfer training module is used to fine-tune the trainable network layer parameters on the user behavior data in the target marketing scenario; the transfer training module constructs a loss function based on the target domain samples, and the loss function includes a prediction error term and a regularization term, and the regularization term is used to limit the difference between the new parameters and the original parameters of the general model during the fine-tuning process.

[0043] Specifically, the transfer training module is the executor for the general model to adapt to the target marketing scenario. It focuses on fine-tuning the thawed network layers to make the model better understand the scenario and be more in line with users. This module is not responsible for the decision-making freezing strategy, nor for the feature input, but it determines the accuracy of the final output of the user behavior prediction.

[0044] Everything starts with data, samples from the target domain, which are input into the model one by one. Each sample contains a set of structured behavioral feature vectors, usually in the standard input format output from the target domain data processing module. The module calls the training engine and passes the neural network in batches. The frozen layers remain unchanged, and only the unfrozen layers participate in the gradient update.

[0045] During training, the module constructs a composite loss function, which consists of two parts: the prediction error term and the parameter constraint term: ; Prediction error term It is used to measure the prediction accuracy of the model for user behavior in the target domain. This term usually takes the form of cross-entropy loss and is applicable to binary classification or multi-classification tasks: ; Among them, represents the true behavior label of the target sample, represents the predicted value of the model for its behavior, is the number of batch samples. The regularization term is used to limit the change amplitude of the model parameters during the fine-tuning process. The goal is to maintain the structural consistency of the fine-tuned model and the original general model at the performance layer and prevent performance degradation caused by excessive deviation. It is specifically defined as: ; Among them, represents the weight parameter of the th layer in the current network, is the initial parameter copy of this layer in the general model, is the set of network layers currently participating in fine-tuning. This term is controlled by an adjustable coefficient to control its proportion in the total loss. The larger the coefficient, the more the model tends to retain the original knowledge structure.

[0046] The model parameter optimization uses the Adam optimizer with a small learning rate. During training, only the parameters of the unfrozen layers participate in the gradient calculation, and the remaining layers remain frozen. Before each round of training iteration, the system will call the strategy set by the fine-tuning control module to determine whether the current network layer is unfrozen. If it is unfrozen, the gradient is allowed to flow and participate in the update.

[0047] The training module also integrates an early stopping mechanism to terminate training early after the performance of the validation set stabilizes to prevent overfitting. In addition, an optional monitoring mechanism will track the change trend of the parameters of the unfrozen layers in real time. When a large deviation occurs, it can prompt the fine-tuning control module to dynamically adjust the strategy or reset the set of trainable layers.

[0048] The design and implementation of this module significantly improve the adaptation efficiency of the general model in the target domain scenario, ensuring both the stability of the original model structure and enhancing the modeling ability for specific marketing tasks.

[0049] The model evaluation module is used to evaluate the behavior prediction performance and output stability of the fine-tuned model in the target domain. If the evaluation fails to meet the standard, the fine-tuning strategy is adjusted and retraining is carried out. The model evaluation module is used to calculate the difference in the predicted probability distributions before and after fine-tuning, and determines whether the behavior prediction of the model in the target scenario is stable by setting an offset threshold. The calculated behavior prediction difference is implemented based on the KL divergence method. KL divergence is used to measure the distance between the output probability distributions of the model before and after fine-tuning. If the evaluation result does not meet the requirements, it will automatically fallback to the fine-tuning control module and reset the freezing and thawing strategies.

[0050] Specifically, the model evaluation module is used to systematically evaluate the behavior prediction performance and output stability of the model in the target marketing scenario after the fine-tuning training. This module not only focuses on the final prediction accuracy but also focuses on detecting whether there is an abnormal drift in the model output distribution to ensure the generalization effect and stability of the fine-tuned model in the target domain.

[0051] In the evaluation process, the module performs forward inference on the models before and after fine-tuning based on the target domain validation set, and records the predicted probability distribution corresponding to each sample. The model before fine-tuning is the original state of the general user behavior prediction model, and the output result can be regarded as the reference distribution; the model after fine-tuning is the version with parameters adjusted in the target domain, and its output result needs to be subjected to stability detection.

[0052] To quantify the difference between the model output probability distributions before and after fine-tuning, the module uses the Kullback-Leibler divergence (KL divergence) as the core metric. KL divergence measures the information deviation degree between two distributions, and the calculation formula is as follows: ; where represents the predicted probability of the general model for the -th class, represents the corresponding output of the fine-tuned model, is the number of classes. This metric is averaged over multiple samples to obtain the overall KL divergence mean: ; where is the number of evaluation samples, , is the predicted distribution of the -th sample.

[0053] To ensure that the model output fluctuates within a reasonable range, the evaluation module sets a stability offset threshold δ. When this condition is met, it is considered that the model output is stable and the fine-tuning result is valid; when it exceeds this threshold, it indicates that the model's prediction behavior deviates too much from the original distribution, and there may be problems such as overfitting and improper drift.

[0054] In addition, the evaluation module also synchronously calculates traditional behavior prediction performance metrics, such as AUC, LogLoss, Accuracy, etc., to ensure that the model not only meets the standard in terms of distribution stability but also has acceptability in terms of prediction accuracy.

[0055] Once the evaluation result does not meet the set criteria (such as a simultaneous decrease in prediction accuracy and a too large KL divergence), the module will automatically trigger a fallback mechanism. This mechanism will notify the fine-tuning control module to reset the freezing and thawing strategies, such as raising the thawing threshold, adjusting the thawing rhythm function, restricting the range of deep parameter updates, etc. After the new strategy is set, the system will restart the transfer training process and re-enter the evaluation stage, forming a closed-loop optimization path.

[0056] During the operation of the evaluation module, it will record the logs of multiple rounds of fine-tuning and evaluation, which supports analyzing the impact of different strategies on the stability of the model output and provides data basis for subsequent strategy optimization. The evaluation process can run automatically without manual intervention and is applicable to the model management scenario under the continuous integration training system (CI / CT).

[0057] Embodiment 2: Referring to Figure 2 , in the second embodiment of the present invention, the present invention provides a digital marketing analysis and management method, including the following steps: S1. Train a general behavior prediction model based on historical cross-domain user behavior data; S2. Collect user behavior data in the target marketing scenario and perform standardization and input structure alignment; S3. Set the network layer of the trainable parameters according to the preset freezing and thawing strategies; S4. Perform transfer fine-tuning training on the selected network layer in the general model using the target scenario data; S5. Evaluate the behavior prediction accuracy and output stability of the fine-tuned model in the target marketing scenario, and push it if it meets the standard; S6. When the evaluation index does not reach the preset threshold, return to step S3.

[0058] Specifically, the system constructs a basic model using historical behavior data from multiple business scenarios in the early stage. The model structure has a high degree of abstraction ability and can uniformly model user interaction patterns under different platforms and different content types. At this stage, the system focuses on the consistent encoding of cross-domain features and the general expression of behavior relationships, and finally forms a general behavior prediction model with good transferability in multiple marketing applications. This model, as the basis for subsequent fine-tuning and deployment, has strong structural generality and expression ability.

[0059] When the system is deployed to specific marketing tasks, such as content recommendation optimization, advertising targeted delivery, or user conversion prediction scenarios, it first collects and structurally organizes the user behavior data of the target platform. In this process, the system uniformly regularizes problems such as structural differences, missing fields, and inconsistent encoding methods in the original data, and performs preprocessing operations such as feature normalization, category mapping, and time series filling to ensure that the generated samples can be semantically consistent with the data structure in the general model training stage. The input features are finally constructed in the form of standardized vectors to ensure an exact match with the input layer structure of the general model.

[0060] After completing the data preparation, the system enters the fine-tuning strategy setting stage. At this time, each layer inside the model sets its training state according to the preset parameter unfreezing plan. Different network layers are divided into frozen or updatable states according to their functional attributes. Generally speaking, the representation layer and the general feature extraction module remain frozen in the initial stage, and only the upper-layer specialization decision module is opened to adapt to the target domain feature distribution. The strategy supports dynamically adjusting the degree of unfreezing according to the training rounds, enabling the model to maintain structural stability and inherit the original capabilities while learning the target scenario features.

[0061] Initiate the transfer training stage based on the target domain samples. The system performs forward propagation and parameter update on the new data through the existing network structure. Due to the limited range of trainable parameters, the overall parameter adjustment of the model is small, the update process is more concentrated, and the convergence speed is fast. During the training process, the system continuously monitors the change in the prediction ability of the model on the target task, and at the same time evaluates the learning trend of the model in each training stage to judge whether the model has effective generalization ability.

[0062] To ensure that the model training result not only has high prediction accuracy but also maintains good output stability, the system introduces a unified evaluation mechanism after the fine-tuning. This evaluation process not only examines traditional performance indicators such as accuracy and recall rate of the model on the validation set, but also focuses on the change situation of the model at the level of behavior output distribution. The system compares the prediction probability outputs of the model before and after fine-tuning under the same input to identify whether there is a significant deviation, so as to judge whether the model has the risk of overfitting, structural drift, or over-reliance on target scenario features in learning.

[0063] If the evaluation results indicate that the model prediction performance fails to meet the preset indicators, or the behavioral output shows an unstable trend, the system will automatically roll back to the parameter setting stage, reconfigure the freezing and thawing strategies, and restart the transfer training process. During the rollback process, a new strategy can be automatically generated based on the previous training record, or an external manual intervention plan can be introduced by the configuration center to achieve flexible adjustment.

[0064] Embodiment 3: In the third embodiment of the present invention, based on the same inventive concept, a computer-readable storage medium is proposed. The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, it implements the steps of the digital marketing analysis and management method in the above embodiment.

[0065] Embodiment 4: In the fourth embodiment of the present invention, based on the same inventive concept, a computer is proposed. The computer includes: a processor and a memory; the processor and the memory communicate with each other; the memory is used to store instructions; the processor is used to execute the instructions in the memory to execute the digital marketing analysis and management method in the above embodiment.

[0066] It should be understood that each part of the present invention can be implemented by hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, any one or a combination of the following well-known technologies in the art can be used: discrete logic circuits with logic gate circuits for implementing logical functions on data signals, application-specific integrated circuits with appropriate combinational logic gate circuits, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0067] Finally, it should be noted that the above are only the preferred embodiments of the present invention and are not used to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, for those skilled in the art, they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A digital marketing analysis and management system, characterized in that, The system includes: A general modeling module, which is used to train a general user behavior prediction model based on cross-domain user behavior data and output a corresponding model parameter set; A target domain data processing module, which is used to collect and preprocess user behavior data in the target marketing scenario and complete the structural alignment between the input features and the general user behavior prediction model; A fine-tuning control module, which is used to set parameter freezing and thawing strategies for the general user behavior prediction model and define the network layers that need to be updated; A transfer training module, which is used to perform fine-tuning training on the network layer parameters set as trainable on the user behavior data in the target marketing scenario; A model evaluation module, which is used to evaluate the behavior prediction performance and output stability of the fine-tuned model in the target domain. If the evaluation fails to meet the standard, the fine-tuning strategy is adjusted and retraining is performed.

2. The digital marketing analysis and management system according to claim 1, wherein The parameter freezing and thawing strategy set by the fine-tuning control module includes a thawing timing function set layer by layer, which is used to control whether each network layer participates in parameter update according to the number of training rounds.

3. The digital marketing analysis and management system according to claim 1, characterized in that The transfer training module constructs a loss function based on target domain samples. The loss function includes a prediction error term and a regularization term, and the regularization term is used to limit the difference between the new parameters and the original parameters of the general model during the fine-tuning process.

4. The digital marketing analysis and management system according to claim 1, characterized in that The user behavior prediction model constructed by the general modeling module is a multi-layer neural network model, with the input being the user behavior feature vector and the output being the probability value of the corresponding behavior occurring.

5. The digital marketing analysis and management system according to claim 1, wherein The model evaluation module is used to calculate the difference in the prediction probability distribution before and after fine-tuning, and determine whether the behavior prediction of the model in the target scenario is stable by setting an offset threshold.

6. The digital marketing analysis and management system according to claim 1, characterized in that, The behavior prediction difference calculated by the model evaluation module is implemented based on the KL divergence method, and the KL divergence is used to measure the distance between the model output probability distributions before and after fine-tuning.

7. The digital marketing analysis and management system according to claim 1, characterized in that, When the evaluation result of the model evaluation module does not meet the requirements, it automatically returns to the fine-tuning control module and resets the freezing and thawing strategy.

8. A digital marketing analysis and management method, characterized in that, For the digital marketing analysis and management system according to any one of claims 1-7, the following steps are included: S1. Train a general behavior prediction model based on historical cross-domain user behavior data; S2. Collect user behavior data in the target marketing scenario and perform standardization and input structure alignment; S3. Set the network layers of the trainable parameters according to the preset freezing and thawing strategy; S4. Perform transfer fine-tuning training on the selected network layers in the general model on the target scenario data; S5. Evaluate the behavior prediction accuracy and output stability of the fine-tuned model in the target marketing scenario. If the standard is met, push it; S6. When the evaluation index does not reach the preset threshold, return to step S3.

9. A computer device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the digital marketing analysis and management method according to claim 8.

10. A readable storage medium, characterized in that, A computer program is stored on the readable storage medium. When the computer program is executed by the processor, it implements the digital marketing analysis and management method according to claim 8.

Images