A marketing strategy optimization processing method and system based on deep learning

Abstract

The application relates to the technical field of data processing, and discloses a marketing strategy optimization processing method and system based on deep learning, which comprises the following steps: obtaining historical marketing execution data and current marketing task data, generating historical strategy combination samples and an initial candidate marketing strategy combination; constructing an AutoInt model and mapping strategy composition embedding vectors; setting a strategy composition closed loop verification structure, generating a closed loop interaction representation and a closed loop breakpoint representation; selecting real strategy combinations with the same scene and different conversion labels, and generating replacement path samples; training an improved AutoInt model based on the historical strategy combination samples and the replacement path samples; generating a local rewriting candidate marketing strategy combination according to an interaction residual error; re-scoring and outputting a target marketing strategy combination with the maximum predicted effect value. The application realizes intelligent scoring and directional optimization of a marketing strategy combination, and improves strategy matching accuracy and conversion effect.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a marketing strategy optimization method and system based on deep learning. Background Technology

[0002] With the development of e-commerce and mobile internet marketing, businesses typically generate marketing strategies based on user behavior, product attributes, promotional content, and reach channels. They then utilize rule matching, manual segmentation, historical conversion rate ranking, or click-through rate and conversion rate prediction models to select the strategy with the highest expected performance from multiple candidate marketing strategy combinations for execution. While these methods can achieve basic user matching and strategy selection, they rely heavily on historical statistical results or overall prediction scores.

[0003] Existing methods typically treat candidate marketing strategy combinations as a whole input model, directly outputting predicted performance values. This lacks modeling of the relationships between strategy components, making it difficult to determine whether a consistent marketing execution chain exists between user groups, target audiences, promotional content, and outreach methods. Existing model training often relies on the relationship between individual historical strategy combinations and conversion tags, failing to fully utilize the substitution process between different real-world strategy combinations. This makes it difficult for the model to identify the impact of changes in a particular strategy component on the overall conversion result. For these reasons, existing methods easily overlook inconsistencies between strategy components during candidate strategy selection, resulting in marketing strategies with high predicted scores but inadequate matching and unstable conversion performance in actual implementation.

[0004] Therefore, how to provide a marketing strategy optimization method and system based on deep learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a marketing strategy optimization method and system based on deep learning. This invention utilizes an improved AutoInt model to perform feature interaction modeling on candidate marketing strategy combinations, and combines closed-loop verification, alternative path training, and interactive residual rewriting processes to achieve predictive scoring and optimized output of marketing strategies. It has the advantages of accurate strategy matching, traceable optimization process, and executable output results.

[0006] A marketing strategy optimization method based on deep learning according to an embodiment of the present invention includes the following steps: Acquire historical marketing execution data and current marketing task data; encode the historical marketing execution data to generate historical strategy combination samples; generate initial candidate marketing strategy combinations from the current marketing task data. Construct an AutoInt model to map the initial candidate marketing strategy combination into a strategy component embedding vector; In the self-attention feature interaction layer of the AutoInt model, a closed-loop verification structure for strategy composition is set up. Based on the strategy composition embedding vector, the interaction representation of adjacent strategy composition content is generated. The interaction representation is concatenated according to the marketing execution order to generate a closed-loop interaction representation and a closed-loop breakpoint representation. Select two real strategy combinations from the historical strategy combination sample that have the same marketing scenario identifier but different conversion tags; The differences between the two real strategy combinations are replaced item by item to generate replacement path samples. The historical strategy combination samples and replacement path samples are input into the AutoInt model for training. The AutoInt model learns the correspondence between the closed-loop interaction changes before and after the replacement and the changes in the conversion labels, thus obtaining the improved AutoInt model. The initial candidate marketing strategy combination is input into the improved AutoInt model. The interaction residual is calculated based on the closed-loop interaction representation and the closed-loop interaction benchmark with the conversion tag as the first tag. The low-match composition relationship is determined based on the interaction residual, and a locally rewritten candidate marketing strategy combination is generated. The candidate marketing strategy combination with partial rewrite is input into the improved AutoInt model for re-scoring, and the target marketing strategy combination with the highest predicted effect value is output.

[0007] Optionally, obtaining historical marketing execution data and current marketing task data specifically includes: Deduplication, missing field completion, abnormal reach record removal, and conversion time window marking are performed on historical marketing execution data to obtain standardized historical marketing records; The strategy components and conversion results in the standardized historical marketing records are encoded according to a unified field dictionary to generate a sample of historical strategy combinations with conversion tags. Perform task field validation and executable content filtering on the current marketing task data to obtain the components of the currently available strategies; The components of the currently available strategies are combined in the order of marketing execution to generate an initial candidate marketing strategy combination.

[0008] Optionally, the construction of the AutoInt model specifically includes: Construct an AutoInt model that includes an input layer, a feature embedding layer, a self-attention feature interaction layer, a residual connection layer, a policy-based closed-loop verification layer, and a prediction output layer; The input layer receives the strategy composition encoding sequence corresponding to the initial candidate marketing strategy combination; The feature embedding layer maps the discrete encoded values ​​in the policy composition encoding sequence to policy composition embedding vectors of the same dimension; The self-attention feature interaction layer performs interactive computation on the policy component embedding vectors based on a multi-head self-attention mechanism to generate a high-order interactive representation. The residual connection layer connects the policy composition embedding vector with the high-order interaction representation to generate the residual interaction representation; The strategy consists of a closed-loop verification layer that receives residual interaction representations and forms closed-loop interaction representations according to the marketing execution order. The prediction output layer outputs the prediction effect value based on closed-loop interaction representation.

[0009] Optionally, setting a strategy to form a closed-loop verification structure in the self-attention feature interaction layer of the AutoInt model specifically includes: The residual interaction representations are arranged according to the marketing execution order. Two adjacent residual interaction representations are selected as the preceding and following representations. The following representations are reconstructed in a forward manner based on the preceding representations. The preceding representations are then back-verified in a reverse manner based on the following representations. The first difference value between the forward reconstruction result and the following representation is calculated. The second difference value between the reverse verification result and the preceding representation is calculated. The first difference value and the second difference value are then fused to obtain the closed-loop acceptance deviation of the corresponding connection position. Perform the same processing on the residual interaction representation of the last position and the residual interaction representation of the first position to obtain the closed-loop connection deviation of the first and last positions. According to the marketing execution sequence, the closed-loop acceptance deviation of each connection position is merged with the corresponding residual interaction representation to generate a closed-loop interaction representation; The residual interactive representation corresponding to the connection position with the largest closed-loop acceptance deviation is merged with the closed-loop acceptance deviation to generate a closed-loop breakpoint representation.

[0010] Optionally, selecting two real strategy combinations from the historical strategy combination sample that have the same marketing scenario identifier but different conversion tags specifically includes: Based on the target user group, marketing task type, and execution cycle of the historical strategy combination samples, a marketing scenario identifier is generated. The historical strategy combination samples are then grouped according to the marketing scenario identifier. Within the same group, the historical strategy combination samples with the first conversion tag are selected as the first real strategy combination, and the historical strategy combination samples with the second conversion tag and different strategy composition are selected as the second real strategy combination.

[0011] Optionally, the step of replacing each component of the differences between the two real strategy combinations specifically includes: The policy composition encoding sequences of the first real policy combination and the second real policy combination are compared bit by bit to determine the positions of the differences in the encoding values; Starting with the first real strategy combination, select the differentiating components in sequence according to the marketing execution order, and replace the coded values ​​of the starting strategy combination at the selected differentiating components with the coded values ​​of the corresponding positions in the second real strategy combination to generate intermediate strategy combinations. Each time a replacement is completed, the strategy combination before replacement, the strategy combination after replacement, the position of the replaced difference component, and the corresponding change in conversion tag are bound together to form a single-step replacement sample; All single-step replacement samples are concatenated according to the replacement execution order to generate replacement path samples.

[0012] Optionally, the step of inputting historical strategy combination samples and replacement path samples into the AutoInt model for training specifically includes: Input historical strategy combination samples into the AutoInt model to obtain the corresponding prediction effect value, and calculate the conversion prediction loss based on the prediction effect value and conversion label; The single-step replacement samples in the replacement path samples are sequentially input into the AutoInt model to obtain the closed-loop interaction representations corresponding to the policy combination before and after replacement, respectively. Calculate the closed-loop change between the closed-loop interaction representation before and after the replacement, bind the closed-loop change with the corresponding conversion tag change, and generate a replacement monitoring signal; The AutoInt model is jointly trained based on the transformation prediction loss and the replacement supervision signal. The parameters of the feature embedding layer, the self-attention feature interaction layer, the residual connection layer, the policy composition closed-loop verification layer, and the prediction output layer are updated to obtain the improved AutoInt model.

[0013] Optionally, the step of determining low-match composition relationships based on interaction residuals and generating locally rewritten candidate marketing strategy combinations specifically includes: The initial candidate marketing strategy combination is input into the improved AutoInt model to obtain a closed-loop interaction representation; Generate a closed-loop interaction benchmark from historical strategy combination samples with the same marketing scenario identifier and conversion tag as the first tag; Generate interaction residuals based on the difference between the closed-loop interaction representation and the closed-loop interaction benchmark, and select at least two low-matching relationships based on the interaction residuals; Among the historical strategy combination samples with the same marketing scenario identification and the first conversion tag, the historical strategy combination sample with the smallest difference from the closed-loop interaction benchmark is selected as the replacement source sample; Extract the strategy component codes corresponding to low-matching composition relationships from the source samples, replace the strategy component codes in the same positions in the initial candidate marketing strategy combination, and generate a partially rewritten candidate marketing strategy combination.

[0014] Optionally, the target marketing strategy combination that maximizes the output predicted effect value specifically includes: Input the partially rewritten candidate marketing strategy combination and the initial candidate marketing strategy combination into the improved AutoInt model to obtain the corresponding predicted performance values; The candidate marketing strategy combination with the highest predicted performance value is selected as the target marketing strategy combination. Output a structured result of the target marketing strategy mix. The structured result includes the target user group, recommended marketing targets, discount methods and discount levels, reach channels and reach time and predicted effect values.

[0015] A marketing strategy optimization processing system based on deep learning according to an embodiment of the present invention includes: The data processing module is used to acquire historical marketing execution data and current marketing task data, encode historical marketing execution data to generate historical strategy combination samples, and generate initial candidate marketing strategy combinations based on current marketing task data. The AutoInt modeling module is used to build an AutoInt model that includes a feature embedding layer, a self-attention feature interaction layer, a residual connection layer, a strategy composition closed-loop verification layer, and a prediction output layer, mapping the initial candidate marketing strategy combination to a strategy composition embedding vector. The closed-loop verification module is used to generate residual interaction representations based on policy composition embedding vectors, perform forward reconstruction and reverse verification on the residual interaction representations, and generate closed-loop interaction representations and closed-loop breakpoint representations. The path training module is used to select two real strategy combinations with the same marketing scenario identifier and different conversion labels from the historical strategy combination samples, generate replacement path samples, and train the AutoInt model based on the historical strategy combination samples and replacement path samples to obtain the improved AutoInt model. The strategy rewriting module is used to input the initial candidate marketing strategy combination into the improved AutoInt model, calculate the interaction residual based on the closed-loop interaction representation and the closed-loop interaction benchmark, and generate a locally rewritten candidate marketing strategy combination based on the interaction residual. The strategy output module is used to input the partially rewritten candidate marketing strategy combination into the improved AutoInt model for re-scoring and output the target marketing strategy combination with the highest predicted effect value.

[0016] The beneficial effects of this invention are: (1) The present invention sets up a closed-loop verification structure for strategy composition in the AutoInt model, and performs closed-loop interactive modeling of the strategy composition in the candidate marketing strategy combination according to the marketing execution order. This can reduce the problem of insufficient consideration of the connection relationship of the composition when selecting strategies based solely on the overall prediction score, and make the output target marketing strategy combination more in line with the actual marketing execution logic.

[0017] (2) By constructing replacement path samples, the present invention uses historical strategy combination samples and replacement path samples together for model training, so that the model not only learns the correspondence between historical strategy combinations and conversion labels, but also learns the closed-loop interaction changes before and after the replacement of strategy components, which is conducive to improving the model's ability to identify the impact of changes in strategy components.

[0018] (3) The present invention calculates the interaction residual based on the closed-loop interaction representation and the closed-loop interaction benchmark, generates a locally rewritten candidate marketing strategy combination based on the interaction residual, and then outputs the target marketing strategy combination by re-scoring through the improved AutoInt model. This can reduce the processing volume of aimless enumeration of candidate strategies and improve the pertinence and executability of the marketing strategy optimization results. Attached Figure Description

[0019] 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 a marketing strategy optimization method based on deep learning proposed in this invention; Figure 2 This is a flowchart of a marketing strategy optimization processing system based on deep learning proposed in this invention; Figure 3 This is a schematic diagram of the target strategy output of a marketing strategy optimization processing method based on deep learning proposed in this invention. Detailed Implementation

[0020] 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.

[0021] refer to Figures 1-3 A marketing strategy optimization method based on deep learning includes the following steps: Acquire historical marketing execution data and current marketing task data; encode the historical marketing execution data to generate historical strategy combination samples; generate initial candidate marketing strategy combinations from the current marketing task data. Construct an AutoInt model to map the initial candidate marketing strategy combination into a strategy component embedding vector; In the self-attention feature interaction layer of the AutoInt model, a closed-loop verification structure for strategy composition is set up. Based on the strategy composition embedding vector, the interaction representation of adjacent strategy composition content is generated. The interaction representation is concatenated according to the marketing execution order to generate a closed-loop interaction representation and a closed-loop breakpoint representation. Select two real strategy combinations from the historical strategy combination sample that have the same marketing scenario identifier but different conversion tags; The differences between the two real strategy combinations are replaced item by item to generate replacement path samples. The historical strategy combination samples and replacement path samples are input into the AutoInt model for training. The AutoInt model learns the correspondence between the closed-loop interaction changes before and after the replacement and the changes in the conversion labels, thus obtaining the improved AutoInt model. The initial candidate marketing strategy combination is input into the improved AutoInt model. The interaction residual is calculated based on the closed-loop interaction representation and the closed-loop interaction benchmark with the conversion tag as the first tag. The low-match composition relationship is determined based on the interaction residual, and a locally rewritten candidate marketing strategy combination is generated. The candidate marketing strategy combination with partial rewrite is input into the improved AutoInt model for re-scoring, and the target marketing strategy combination with the highest predicted effect value is output.

[0022] In this embodiment, obtaining historical marketing execution data and current marketing task data specifically includes: The historical marketing execution data is deduplicated, missing fields are filled in, abnormal reach records are removed, and conversion time windows are defined to obtain standardized historical marketing records. Historical marketing execution data refers to marketing records that have been completed. Reach refers to the act of sending marketing content to the user's terminal through App push, SMS, email, or in-app message. During deduplication, user identifier, marketing task identifier, reach time, and marketing content code are used as joint judgment criteria. Records with the same joint judgment criteria and a reach time interval of less than 5 minutes are retained. When filling in missing fields, numerical fields are filled in using the median under the same marketing task, and category fields are filled in using a preset default code. Abnormal reach records include records of failed delivery, user unsubscription, missing channel receipts, and reach times earlier than the task generation time. The conversion time window is the period of time for observing results calculated from the time of reach completion, where App push corresponds to 72 hours, SMS corresponds to 48 hours, and email corresponds to 7 days. Records of payment behavior occurring within the conversion time window are marked as conversion records. The strategy components and conversion results in standardized historical marketing records are encoded using a unified field dictionary to generate historical strategy combination samples with conversion tags. Strategy components refer to the data required to form an executable marketing action, including at least user group content, target audience content, promotional content, and outreach content. A strategy combination is a complete marketing action record formed by arranging the above strategy components in the order of marketing execution. The unified field dictionary is used to convert strategy components in text or category form into discrete encoded values, and conversion results are converted into conversion tags based on whether payment was completed within the conversion time window. The discrete encoded values ​​corresponding to the same standardized historical marketing record are bound to conversion tags. The current marketing task data undergoes task field validation and executable content filtering to obtain the components of the currently available strategies. Current marketing task data refers to the data set corresponding to the marketing tasks to be executed. Task field validation determines whether the current marketing task data meets the data conditions required to generate marketing actions. Validation content includes task validity period, user group scope, target audience scope, discount content scope, and reach content scope. Executable content filtering removes content that cannot be executed from the validated data, including target audiences with insufficient inventory, discount content with insufficient remaining quota, and reach content that cannot be sent at the current time. The data retained after filtering is used as the components of the currently available strategies. The currently available strategy components are combined according to the marketing execution order to generate initial candidate marketing strategy combinations. The marketing execution order refers to the processing order of determining user group content, matching marketing target content, matching promotional content, and matching reach content. Initial candidate marketing strategy combinations refer to the records of candidate marketing actions that have not yet been scored by the improved AutoInt model. During generation, the same user group content is sequentially concatenated with marketing target content, promotional content, and reach content that meet the conditions of the current marketing task to obtain multiple initial candidate marketing strategy combinations that can be input into the model. Each initial candidate marketing strategy combination corresponds to one executable marketing action.

[0023] In this embodiment, constructing the AutoInt model specifically includes: Construct an AutoInt model that includes an input layer, a feature embedding layer, a self-attention feature interaction layer, a residual connection layer, a policy-based closed-loop verification layer, and a prediction output layer; The input layer receives the strategy component encoding sequence corresponding to the initial candidate marketing strategy combination. The strategy component encoding sequence is a fixed-length encoding sequence formed by converting the various strategy components in the initial candidate marketing strategy combination through a unified field dictionary. The sequence length is consistent with the number of categories of strategy components. The length of the strategy component encoding sequence is set to 4, which corresponds to user group content, marketing target content, promotional content, and reach content, respectively. The input layer receives the encoding sequence in this fixed order and passes the encoding sequence to the feature embedding layer. The feature embedding layer maps the discrete encoded values ​​in the policy composition encoding sequence to policy composition embedding vectors of the same dimension; the feature embedding layer configures an embedding table for each type of policy composition content; when the policy composition encoding sequence is input into the feature embedding layer, the feature embedding layer searches for vectors in the corresponding embedding table according to the discrete encoded values, and arranges the searched vectors in the input order; The self-attention feature interaction layer performs interactive computation on the policy composition embedding vectors based on a multi-head self-attention mechanism to generate a high-order interactive representation. The self-attention feature interaction layer adopts the multi-head self-attention structure in AutoInt, where each attention head maps the policy composition embedding vector to a query vector, a key vector, and a value vector. For any policy composition embedding vector, the query vector corresponding to that embedding vector is multiplied by the key vectors corresponding to the other policy composition embedding vectors. The dot product results are then Softmax normalized to obtain attention coefficients. These attention coefficients determine the proportion of the corresponding value vectors participating in the current attention head's output computation. The value vectors are summed according to their corresponding attention coefficients to obtain the current attention head's output vector. The output vectors of multiple attention heads are concatenated to generate a high-order interactive representation, which represents the combination relationship between the policy composition embedding vectors after self-attention computation. The residual connection layer connects the policy composition embedding vector with the higher-order interaction representation to generate the residual interaction representation. The residual connection layer adds the policy composition embedding vector and the higher-order interaction representation bit by bit. When the dimensions of the two are inconsistent, the higher-order interaction representation is adjusted to the same dimension as the policy composition embedding vector through linear mapping, and bit by bit addition is performed to obtain the residual interaction representation. The residual interaction representation retains both the input policy composition content information and the interaction information generated by the self-attention feature interaction layer. The strategy composition closed-loop verification layer receives residual interaction representations and forms closed-loop interaction representations according to the marketing execution order. The strategy composition closed-loop verification layer connects the residual interaction representations sequentially according to the marketing execution order to form a closed-loop path composed of adjacent strategy components. Each segment of the closed-loop path corresponds to an interaction relationship between adjacent strategy components. The strategy composition closed-loop verification layer performs consistency calculation on the interaction relationship between two adjacent segments and marks the connection position where the consistency calculation result is lower than a set threshold as a closed-loop breakpoint. The set threshold is 0.5. All closed-loop path interaction results and closed-loop breakpoint markings together form a closed-loop interaction representation. The prediction output layer outputs the prediction effect value based on the closed-loop interaction representation; the prediction output layer inputs the closed-loop interaction representation into the fully connected layer and outputs the prediction effect value with a value range of 0 to 1 through the Sigmoid function; the prediction effect value is used to represent the predicted conversion probability corresponding to the initial candidate marketing strategy combination. When the prediction effect value is closer to 1, it means that the corresponding candidate marketing strategy combination scores higher in the direction where the conversion label is the first label.

[0024] In this embodiment, setting a strategy to form a closed-loop verification structure in the self-attention feature interaction layer of the AutoInt model specifically includes: The residual interaction representations are arranged according to the marketing execution order. Two adjacent residual interaction representations are selected sequentially as the preorder and postorder representations. The postorder representation is reconstructed forward based on the preorder representation, and the preorder representation is backtested backward based on the postorder representation. The first difference value between the forward reconstruction result and the postorder representation is calculated, as is the second difference value between the backtesting result and the preorder representation. The first and second difference values ​​are then fused to obtain the closed-loop acceptance bias at the corresponding connection position. The residual interaction representations are vectors of the same dimension. Forward reconstruction refers to inputting the preorder representation into the first fully connected variable... The first fully connected transform unit outputs a reconstructed vector with the same dimension as the subsequent representation. The input and output dimensions of the first fully connected transform unit are consistent. Reverse proof refers to inputting the subsequent representation into the second fully connected transform unit, which outputs a proof vector with the same dimension as the preceding representation. The input and output dimensions of the second fully connected transform unit are consistent. The first difference value is the mean of the absolute values ​​of the forward reconstruction result and the subsequent representation subtracted bit by bit. The second difference value is the mean of the absolute values ​​of the reverse proof result and the preceding representation subtracted bit by bit. The closed-loop connection deviation is the arithmetic mean of the first and second difference values. The residual interaction representation at the end of the sequence is processed in the same way as the residual interaction representation at the beginning of the sequence to obtain the closed-loop connection deviation at the beginning and end connection position. The residual interaction representation at the end of the sequence is used as the preceding representation, and the residual interaction representation at the beginning of the sequence is used as the following representation. The closed-loop connection deviation at the beginning and end connection position is calculated in the same way as forward reconstruction, reverse verification and difference value calculation, so that the last strategic component in the marketing execution sequence forms a closed connection with the first strategic component. According to the marketing execution order, the closed-loop acceptance deviation of each connection position is fused with the corresponding residual interaction representation to generate a closed-loop interaction representation. During fusion, the closed-loop acceptance deviation of each connection position is expanded into a deviation vector with the same dimension as the residual interaction representation, and concatenated with the preceding and following representations of that connection position. The concatenation result is input into the third fully connected transformation unit, and the closed-loop local representation of that connection position is output. All closed-loop local representations are concatenated according to the marketing execution order. The residual interaction representation corresponding to the connection position with the largest closed-loop acceptance deviation is fused with the closed-loop acceptance deviation to generate a closed-loop breakpoint representation. The closed-loop acceptance deviations of all connection positions are sorted according to their numerical values, and the connection position with the largest value is selected as the breakpoint position. The preceding and following representations of the breakpoint position and the extended vector of the closed-loop acceptance deviation are concatenated and input into the fourth fully connected transformation unit to output the closed-loop breakpoint representation. The closed-loop breakpoint representation is used to record the composition relationship of the adjacent strategies with the largest acceptance deviation in the current candidate marketing strategy combination.

[0025] The first fully connected transformation unit, the second fully connected transformation unit, the third fully connected transformation unit, and the fourth fully connected transformation unit are all trained parameter units in the improved AutoInt model. They are updated synchronously with the feature embedding layer, the self-attention feature interaction layer, the residual connection layer, and the prediction output layer during the same training process. The update is based on the changes in the transformation labels corresponding to the historical strategy combination samples and the replacement path samples.

[0026] In this embodiment, selecting two real strategy combinations from the historical strategy combination sample that have the same marketing scenario identifier but different conversion tags specifically includes: Marketing scenario identifiers are generated based on the target user group, marketing task type, and execution cycle in the historical strategy combination samples. The historical strategy combination samples are then grouped according to these marketing scenario identifiers. Within each group, historical strategy combination samples with the first conversion tag are selected as the first real strategy combination, and historical strategy combination samples with the second conversion tag and different strategy components are selected as the second real strategy combination. The marketing scenario identifier is a coded identifier used to distinguish the marketing environment to which the historical strategy combination sample belongs. During generation, the target user group code, marketing task type code, and execution cycle code are concatenated in a fixed order, and a hash operation is performed on the concatenation result to obtain the marketing scenario identifier. The execution cycle is categorized by natural day and natural week. Alternatively, marketing campaign cycles can be divided. When grouping, historical strategy combination samples with the same marketing scenario identifier are grouped together. Within the same group, historical strategy combination samples with a conversion tag as the first tag and whose payment time is no more than the time between reaching the target audience do not exceed the conversion time window are selected as the first real strategy combination. Historical strategy combination samples with a conversion tag as the second tag and whose strategy composition coding sequence differs from the strategy composition coding sequence of the first real strategy combination by at least one coding bit are selected as the second real strategy combination. The first tag indicates that the user corresponding to the historical strategy combination sample completed the payment behavior within the conversion time window, and the second tag indicates that the user corresponding to the historical strategy combination sample did not complete the payment behavior within the conversion time window.

[0027] In this embodiment, the item-by-item replacement of the differences between the two real strategy combinations specifically includes: The coding sequences of the strategy components of the first and second real strategy combinations are compared bit by bit to determine the positions of the components with different coding values. The coding sequences of the strategy components correspond to the target user group, the recommended marketing target, the discount method and discount level, the reach channel and the reach time in sequence. During the bit-by-bit comparison, the coding values ​​of the first and second real strategy combinations at the same sequence position are compared. The sequence positions with different coding values ​​are recorded as the positions of the components with differences. When the number of positions of the components with differences is zero, no replacement path sample is generated. Starting with the first true strategy combination, select the difference component positions sequentially according to the marketing execution order. Replace the coded values ​​of the selected difference component positions of the starting strategy combination with the coded values ​​of the corresponding positions in the second true strategy combination to generate intermediate strategy combinations. The marketing execution order is determined according to the order of the strategy component coding sequence. Only one difference component position is replaced at a time, and the unselected difference component positions retain their current coded values. After one replacement is completed, the replaced strategy combination is used as the starting strategy combination for the next replacement, until all difference component positions have been replaced with the corresponding coded values ​​in the second true strategy combination. Each time a replacement is completed, the strategy combination before replacement, the strategy combination after replacement, the position of the replaced difference component, and the corresponding change in conversion label are bound together to form a single-step replacement sample. The corresponding change in conversion label is determined by the conversion label of the first real strategy combination and the conversion label of the second real strategy combination. When the conversion label of the first real strategy combination is the first label and the conversion label of the second real strategy combination is the second label, the corresponding change in conversion label is recorded as changing from the first label to the second label. The single-step replacement sample is used to record the encoding changes and conversion label changes before and after a strategy component content replacement. All single-step replacement samples are concatenated according to the replacement execution order to generate replacement path samples; the replacement execution order is the order in which the difference component positions appear in the policy component encoding sequence; during concatenation, each single-step replacement sample is arranged according to the replacement execution order, and the pre-replacement policy combination corresponding to the first single-step replacement sample and the post-replacement policy combination corresponding to the last single-step replacement sample are recorded, so that the replacement path sample represents the complete replacement process of the first real policy combination gradually transitioning to the second real policy combination.

[0028] In this embodiment, inputting historical strategy combination samples and replacement path samples into the AutoInt model for training specifically includes: Input historical strategy combination samples into the AutoInt model to obtain the corresponding predicted performance values. Calculate the conversion prediction loss based on the predicted performance values ​​and conversion labels. The predicted performance value is the conversion probability ranging from 0 to 1. The first label in the conversion label is encoded as 1, and the second label is encoded as 0. The conversion prediction loss is calculated using binary cross-entropy loss. The conversion prediction loss of a single sample is the binary cross-entropy between the conversion label and the predicted performance value. During batch training, the average conversion prediction loss of all historical strategy combination samples in the same batch is calculated. The single-step replacement samples in the replacement path samples are sequentially input into the AutoInt model to obtain the closed-loop interaction representations corresponding to the policy combination before and after replacement, respectively. The single-step replacement sample includes the policy combination before and after replacement. When inputting, the policy composition encoding sequence corresponding to the policy combination before replacement and the policy composition encoding sequence corresponding to the policy combination after replacement are respectively input into the AutoInt model. After passing through the feature embedding layer, the self-attention feature interaction layer, the residual connection layer, and the policy composition closed-loop verification layer, two closed-loop interaction representations are obtained. The two inputs in the same single-step replacement sample share the same set of AutoInt model parameters. The closed-loop change between the closed-loop interaction representations before and after replacement is calculated, and the closed-loop change is bound to the corresponding transformation label change to generate a replacement supervision signal. The closed-loop change is the vector difference obtained by subtracting the closed-loop interaction representation of the strategy combination before replacement from the closed-loop interaction representation of the strategy combination after replacement. The corresponding transformation label change is determined by the transformation labels of the first real strategy combination and the second real strategy combination. When the first real strategy combination is labeled as the first label and the second real strategy combination is labeled as the second label, the corresponding transformation label change is recorded as a negative change. When the first real strategy combination is labeled as the second label and the second real strategy combination is labeled as the first label, the corresponding transformation label change is recorded as a positive change. The replacement supervision signal consists of the closed-loop change and the corresponding transformation label change. The AutoInt model is jointly trained based on the conversion prediction loss and the replacement supervision signal. The parameters of the feature embedding layer, self-attention feature interaction layer, residual connection layer, policy composition closed-loop verification layer, and prediction output layer are updated to obtain an improved AutoInt model. During joint training, the conversion prediction loss corresponding to the historical policy combination samples is used as the first loss. For each single-step replacement sample in the replacement path samples, the closed-loop change between the closed-loop interaction representation before and after replacement and the prediction change between the prediction effect values ​​are calculated. The change contribution value of each single-step replacement sample is determined based on the closed-loop change and the prediction change. The single-step replacement sample with the largest change contribution value is selected as the key replacement sample. The parameters of the key replacement sample are then used to determine the key replacement sample's contribution value. The second loss is calculated based on the difference in location and the change in transformation label; the first and second losses are added together to obtain the total loss, and the trainable parameters in the AutoInt model are updated through backpropagation; for the i-th single-step replacement sample in the same replacement path sample, the Euclidean distance between the closed-loop interaction representation after replacement and the closed-loop interaction representation before replacement is calculated to obtain the closed-loop change, and the absolute difference between the predicted effect value after replacement and the predicted effect value before replacement is calculated to obtain the predicted change; the closed-loop change and predicted change of all single-step replacement samples in the same replacement path sample are normalized respectively, and the normalized closed-loop change is multiplied by the normalized predicted change to obtain the change contribution value of the i-th single-step replacement sample.

[0029] In this embodiment, determining low-match composition relationships and generating locally rewritten candidate marketing strategy combinations based on interaction residuals specifically includes: The initial candidate marketing strategy combination is input into the improved AutoInt model to obtain a closed-loop interaction representation; A closed-loop interaction benchmark is generated from historical strategy combination samples with the same marketing scenario identifier and the conversion tag as the first tag. The closed-loop interaction benchmark refers to the closed-loop representation reference value used to represent the converted strategy combination under the same marketing scenario. During generation, historical strategy combination samples with the same marketing scenario identifier and the conversion tag as the first tag are input into the improved AutoInt model, and the corresponding closed-loop interaction representation is extracted. The mean of the representation vectors at the same connection position is calculated to obtain the closed-loop interaction benchmark for each connection position. When the number of first tag samples under the same marketing scenario identifier is less than 10, the first tag samples under the same target user group are used to generate the closed-loop interaction benchmark. Interaction residuals are generated based on the difference between the closed-loop interaction representation and the closed-loop interaction benchmark. At least two low-matching composition relationships are selected according to the interaction residuals. The interaction residual refers to the difference between the connection position representation of the initial candidate marketing strategy combination and the corresponding closed-loop interaction benchmark. The interaction residual is calculated using Euclidean distance. After calculating the interaction residuals of each connection position, the interaction residuals are sorted from largest to smallest. The top two connection positions in the sorted order are selected as low-matching composition relationships. When the number of connection positions is greater than 4, the top 30% of the connection positions in the sorted order are selected as low-matching composition relationships. Among historical strategy combination samples with the same marketing scenario identification and conversion tag as the first tag, the historical strategy combination sample with the smallest difference from the closed-loop interaction benchmark is selected as the replacement source sample; the replacement source sample refers to the historical strategy combination sample used to provide the code of replaceable strategy components; when selecting, the average Euclidean distance between the closed-loop interaction representation corresponding to each historical strategy combination sample with the conversion tag as the first tag and the closed-loop interaction benchmark is calculated, and the historical strategy combination sample with the smallest average Euclidean distance is determined as the replacement source sample; when there are multiple historical strategy combination samples with the same average Euclidean distance, the historical strategy combination sample with the closest reach time is selected as the replacement source sample; Extract the strategy component codes corresponding to low-matching composition relationships from the replacement source samples, and replace the strategy component codes at the same positions in the initial candidate marketing strategy combinations to generate locally rewritten candidate marketing strategy combinations. For low-matching composition relationships, which correspond to two adjacent strategy component code positions, extract the strategy component codes at the same coding positions from the replacement source samples during replacement and write them into the corresponding positions in the initial candidate marketing strategy combinations. Strategy component codes that do not correspond to low-matching composition relationships remain unchanged. After replacement, use the updated strategy component code sequence as the locally rewritten candidate marketing strategy combination and input it into the improved AutoInt model for re-scoring.

[0030] In this embodiment, the target marketing strategy combination that maximizes the output predicted effect value specifically includes: Input the partially rewritten candidate marketing strategy combination and the initial candidate marketing strategy combination into the improved AutoInt model to obtain the corresponding predicted performance values; The candidate marketing strategy combination with the highest predicted performance value is selected as the target marketing strategy combination. Output a structured result of the target marketing strategy mix. The structured result includes the target user group, recommended marketing targets, discount methods and discount levels, reach channels and reach time and predicted effect values.

[0031] A marketing strategy optimization processing system based on deep learning, comprising: The data processing module is used to acquire historical marketing execution data and current marketing task data, encode historical marketing execution data to generate historical strategy combination samples, and generate initial candidate marketing strategy combinations based on current marketing task data. The AutoInt modeling module is used to build an AutoInt model that includes a feature embedding layer, a self-attention feature interaction layer, a residual connection layer, a strategy composition closed-loop verification layer, and a prediction output layer, mapping the initial candidate marketing strategy combination to a strategy composition embedding vector. The closed-loop verification module is used to generate residual interaction representations based on policy composition embedding vectors, perform forward reconstruction and reverse verification on the residual interaction representations, and generate closed-loop interaction representations and closed-loop breakpoint representations. The path training module is used to select two real strategy combinations with the same marketing scenario identifier and different conversion labels from the historical strategy combination samples, generate replacement path samples, and train the AutoInt model based on the historical strategy combination samples and replacement path samples to obtain the improved AutoInt model. The strategy rewriting module is used to input the initial candidate marketing strategy combination into the improved AutoInt model, calculate the interaction residual based on the closed-loop interaction representation and the closed-loop interaction benchmark, and generate a locally rewritten candidate marketing strategy combination based on the interaction residual. The strategy output module is used to input the partially rewritten candidate marketing strategy combination into the improved AutoInt model for re-scoring and output the target marketing strategy combination with the highest predicted effect value.

[0032] Example 1: To verify the feasibility of this invention in practice, it was applied to an online retail marketing scenario. In this scenario, users engage in behaviors such as browsing, adding to favorites, adding to cart, placing orders, and making payments. The marketing task requires generating executable marketing strategies based on different user states. The candidate marketing strategy combination mainly consists of the target user group, recommended marketing objects, discount methods and discount levels, and reach channels and reach times. Existing processing methods typically rank candidate marketing strategy combinations based on historical average conversion rates or ordinary prediction models, easily focusing only on the overall score while ignoring the interrelationships between the strategy components. For example, while some user groups may be suitable for receiving coupons, if the recommended marketing objects are inconsistent with the users' historical interests, or if the reach channels and reach times do not conform to user activity patterns, the final output strategy combination will have unstable conversion results in actual execution.

[0033] In this scenario, historical marketing execution data and current marketing task data are first acquired. The historical marketing execution data undergoes deduplication, missing data completion, abnormal reach record removal, and conversion time window calibration to generate historical strategy combination samples. The current marketing task data, after task field validation and executable content filtering, generates initial candidate marketing strategy combinations. These initial candidate marketing strategy combinations are input into the AutoInt model, where a feature embedding layer converts the strategy component encoding sequence into strategy component embedding vectors. Then, a self-attention feature interaction layer learns the high-order interaction relationships between the strategy component content.

[0034] Unlike ordinary AutoInt models, this invention incorporates a closed-loop verification structure for strategy components. This structure arranges residual interaction representations according to the marketing execution order and calculates the closed-loop continuity deviation through forward reconstruction and backward verification, generating closed-loop interaction representations and closed-loop breakpoint representations. Through this process, the model not only determines whether a candidate marketing strategy combination as a whole is likely to convert, but also characterizes whether there is a continuous continuity relationship between the strategy components. Two real strategy combinations with the same marketing scenario identifier but different conversion labels are selected from historical strategy combination samples. The differing components between the two real strategy combinations are replaced item by item, generating replacement path samples. These replacement path samples and historical strategy combination samples are input into the AutoInt model for training, enabling the model to learn the correspondence between changes in closed-loop interactions and changes in conversion labels before and after replacement, resulting in an improved AutoInt model.

[0035] In the application phase, the initial candidate marketing strategy combinations are input into the improved AutoInt model to obtain closed-loop interaction representations. Then, a closed-loop interaction baseline is generated from historical strategy combination samples with the same marketing scenario identifier and a conversion tag of 1. Interaction residuals are calculated based on the differences between the closed-loop interaction representation and the baseline, thereby determining low-match composition relationships. Next, strategy composition codes at corresponding positions are extracted from historical strategy combination samples with a conversion tag of 1 and minimal differences from the baseline, and the initial candidate marketing strategy combinations are partially rewritten to generate partially rewritten candidate marketing strategy combinations. Finally, the partially rewritten candidate marketing strategy combinations are re-input into the improved AutoInt model for scoring, and the target marketing strategy combination with the highest predicted performance value is output.

[0036] To verify the effectiveness of this invention, training and validation samples were divided within the same batch of historical marketing execution data, and three methods were used for comparison. The historical conversion rate ranking method ranked candidates based on their historical average conversion rates under the same marketing scenario. The original AutoInt model used a standard AutoInt structure, directly outputting predicted performance values ​​for candidate marketing strategy combinations. The method of this invention uses an improved AutoInt model with closed-loop validation of strategy components, alternative path training, and interactive residual rewriting. AUC was calculated based on the predicted performance values ​​and actual conversion labels in the validation samples; LogLoss was calculated based on the predicted conversion probability and actual conversion labels; the actual conversion rate was calculated as the ratio of users who completed payment to the actual number of users reached; ROI was calculated as the ratio of transaction amount to marketing cost; and the optimal strategy hit rate was statistically analyzed based on whether the candidate marketing strategy combination selected by the model was the strategy combination with the highest actual conversion rate in the validation samples of the same marketing scenario.

[0037] Table 1: Comparison of the Implementation Effects of Different Marketing Strategy Optimization Methods

[0038] As shown in Table 1, the historical conversion rate ranking method mainly relies on existing conversion performance and has a weak ability to identify newly emerging or low-frequency candidate marketing strategy combinations, resulting in low AUC and optimal strategy hit rate. The original AutoInt model learns the feature interaction relationships between strategy components through a self-attention mechanism, improving AUC from 0.681 to 0.742 and reducing LogLoss from 0.438 to 0.391, indicating improved consistency between predicted performance and actual conversion labels. However, the original AutoInt model still primarily relies on overall interaction representation and cannot further utilize the closed-loop continuity in the marketing execution sequence or the substitution path information between actual strategy combinations.

[0039] The method of this invention achieves an AUC of 0.796 and a LogLoss of 0.354, indicating that the improved AutoInt model outperforms the comparative methods in both the ability to distinguish and calibrate conversion probabilities. The actual conversion rate increased from 5.72% (based on historical conversion rate ranking) to 8.03%, demonstrating that the target marketing strategy combination output by this invention achieved a higher proportion of payment conversions after actual outreach. The ROI increased from 2.18 to 3.04, indicating that this invention generates a higher return on investment for the same marketing cost. The optimal strategy hit rate increased from 41.6% to 66.7%, indicating that this invention is more likely to select the best-performing strategy combination from multiple candidate marketing strategy combinations. In summary, this invention, through closed-loop verification, alternative path training, and local rewriting, enables the marketing strategy optimization process to not only rely on the overall prediction score but also utilize the relationships and substitution information between the strategy components, thereby improving the accuracy and actual execution effect of the target marketing strategy combination.

[0040] 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. A marketing strategy optimization method based on deep learning, characterized in that, Includes the following steps: Acquire historical marketing execution data and current marketing task data; encode the historical marketing execution data to generate historical strategy combination samples; generate initial candidate marketing strategy combinations from the current marketing task data. Construct an AutoInt model to map the initial candidate marketing strategy combination into a strategy component embedding vector; In the self-attention feature interaction layer of the AutoInt model, a closed-loop verification structure for strategy composition is set up. Based on the strategy composition embedding vector, the interaction representation of adjacent strategy composition content is generated. The interaction representation is concatenated according to the marketing execution order to generate a closed-loop interaction representation and a closed-loop breakpoint representation. Select two real strategy combinations from the historical strategy combination sample that have the same marketing scenario identifier but different conversion tags; The differences between the two real strategy combinations are replaced item by item to generate replacement path samples. The historical strategy combination samples and replacement path samples are input into the AutoInt model for training. The AutoInt model learns the correspondence between the closed-loop interaction changes before and after the replacement and the changes in the conversion labels, thus obtaining the improved AutoInt model. The initial candidate marketing strategy combination is input into the improved AutoInt model. The interaction residual is calculated based on the closed-loop interaction representation and the closed-loop interaction benchmark with the conversion tag as the first tag. The low-match composition relationship is determined based on the interaction residual, and a locally rewritten candidate marketing strategy combination is generated. The candidate marketing strategy combination with partial rewrite is input into the improved AutoInt model for re-scoring, and the target marketing strategy combination with the highest predicted effect value is output.

2. The marketing strategy optimization method based on deep learning according to claim 1, characterized in that, The acquisition of historical marketing execution data and current marketing task data specifically includes: Deduplication, missing field completion, abnormal reach record removal, and conversion time window marking are performed on historical marketing execution data to obtain standardized historical marketing records; The strategy components and conversion results in the standardized historical marketing records are encoded according to a unified field dictionary to generate a sample of historical strategy combinations with conversion tags. Perform task field validation and executable content filtering on the current marketing task data to obtain the components of the currently available strategies; The components of the currently available strategies are combined in the order of marketing execution to generate an initial candidate marketing strategy combination.

3. The marketing strategy optimization method based on deep learning according to claim 2, characterized in that, The construction of the AutoInt model specifically includes: Construct an AutoInt model that includes an input layer, a feature embedding layer, a self-attention feature interaction layer, a residual connection layer, a policy-based closed-loop verification layer, and a prediction output layer; The input layer receives the strategy composition encoding sequence corresponding to the initial candidate marketing strategy combination; The feature embedding layer maps the discrete encoded values ​​in the policy composition encoding sequence to policy composition embedding vectors of the same dimension; The self-attention feature interaction layer performs interactive computation on the policy component embedding vectors based on a multi-head self-attention mechanism to generate a high-order interactive representation. The residual connection layer connects the policy composition embedding vector with the high-order interaction representation to generate the residual interaction representation; The strategy consists of a closed-loop verification layer that receives residual interaction representations and forms closed-loop interaction representations according to the marketing execution order. The prediction output layer outputs the prediction effect value based on closed-loop interaction representation.

4. The marketing strategy optimization method based on deep learning according to claim 3, characterized in that, The specific steps of setting a strategy to form a closed-loop verification structure in the self-attention feature interaction layer of the AutoInt model include: The residual interaction representations are arranged according to the marketing execution order. Two adjacent residual interaction representations are selected as the preceding and following representations. The following representations are reconstructed in a forward manner based on the preceding representations. The preceding representations are then back-verified in a reverse manner based on the following representations. The first difference value between the forward reconstruction result and the following representation is calculated. The second difference value between the reverse verification result and the preceding representation is calculated. The first difference value and the second difference value are then fused to obtain the closed-loop acceptance deviation of the corresponding connection position. Perform the same processing on the residual interaction representation of the last position and the residual interaction representation of the first position to obtain the closed-loop connection deviation of the first and last positions. According to the marketing execution sequence, the closed-loop acceptance deviation of each connection position is merged with the corresponding residual interaction representation to generate a closed-loop interaction representation; The residual interactive representation corresponding to the connection position with the largest closed-loop acceptance deviation is merged with the closed-loop acceptance deviation to generate a closed-loop breakpoint representation.

5. The marketing strategy optimization method based on deep learning according to claim 4, characterized in that, The selection of two real strategy combinations from the historical strategy combination sample that have the same marketing scenario identifier but different conversion tags specifically includes: Based on the target user group, marketing task type, and execution cycle of the historical strategy combination samples, a marketing scenario identifier is generated. The historical strategy combination samples are then grouped according to the marketing scenario identifier. Within the same group, the historical strategy combination samples with the first conversion tag are selected as the first real strategy combination, and the historical strategy combination samples with the second conversion tag and different strategy composition are selected as the second real strategy combination.

6. The marketing strategy optimization method based on deep learning according to claim 5, characterized in that, The specific steps of replacing each component of the differences between the two real strategy combinations include: The policy composition encoding sequences of the first real policy combination and the second real policy combination are compared bit by bit to determine the positions of the differences in the encoding values; Starting with the first real strategy combination, select the differentiating components in sequence according to the marketing execution order, and replace the coded values ​​of the starting strategy combination at the selected differentiating components with the coded values ​​of the corresponding positions in the second real strategy combination to generate intermediate strategy combinations. Each time a replacement is completed, the strategy combination before replacement, the strategy combination after replacement, the position of the replaced difference component, and the corresponding change in conversion tag are bound together to form a single-step replacement sample; All single-step replacement samples are concatenated according to the replacement execution order to generate replacement path samples.

7. The marketing strategy optimization method based on deep learning according to claim 6, characterized in that, The specific steps of inputting historical strategy combination samples and replacement path samples into the AutoInt model for training include: Input historical strategy combination samples into the AutoInt model to obtain the corresponding prediction effect value, and calculate the conversion prediction loss based on the prediction effect value and conversion label; The single-step replacement samples in the replacement path samples are sequentially input into the AutoInt model to obtain the closed-loop interaction representations corresponding to the policy combination before and after replacement, respectively. Calculate the closed-loop change between the closed-loop interaction representation before and after the replacement, bind the closed-loop change with the corresponding conversion tag change, and generate a replacement monitoring signal; The AutoInt model is jointly trained based on the transformation prediction loss and the replacement supervision signal. The parameters of the feature embedding layer, the self-attention feature interaction layer, the residual connection layer, the policy composition closed-loop verification layer, and the prediction output layer are updated to obtain the improved AutoInt model.

8. The marketing strategy optimization method based on deep learning according to claim 7, characterized in that, The step of determining low-match composition relationships based on interaction residuals and generating locally rewritten candidate marketing strategy combinations specifically includes: The initial candidate marketing strategy combination is input into the improved AutoInt model to obtain a closed-loop interaction representation; Generate a closed-loop interaction benchmark from historical strategy combination samples with the same marketing scenario identifier and conversion tag as the first tag; Generate interaction residuals based on the difference between the closed-loop interaction representation and the closed-loop interaction benchmark, and select at least two low-matching relationships based on the interaction residuals; Among the historical strategy combination samples with the same marketing scenario identification and the first conversion tag, the historical strategy combination sample with the smallest difference from the closed-loop interaction benchmark is selected as the replacement source sample; Extract the strategy component codes corresponding to low-matching composition relationships from the source samples, replace the strategy component codes in the same positions in the initial candidate marketing strategy combination, and generate a partially rewritten candidate marketing strategy combination.

9. The marketing strategy optimization method based on deep learning according to claim 8, characterized in that, The target marketing strategy combination that maximizes the predicted output effect value specifically includes: Input the partially rewritten candidate marketing strategy combination and the initial candidate marketing strategy combination into the improved AutoInt model to obtain the corresponding predicted performance values; The candidate marketing strategy combination with the highest predicted performance value is selected as the target marketing strategy combination. Output a structured result of the target marketing strategy mix. The structured result includes the target user group, recommended marketing targets, discount methods and discount levels, reach channels and reach time and predicted effect values.

10. A marketing strategy optimization processing system based on deep learning, applied to the marketing strategy optimization processing method based on deep learning as described in any one of claims 1 to 9, characterized in that, include: The data processing module is used to acquire historical marketing execution data and current marketing task data, encode historical marketing execution data to generate historical strategy combination samples, and generate initial candidate marketing strategy combinations based on current marketing task data. The AutoInt modeling module is used to build an AutoInt model that includes a feature embedding layer, a self-attention feature interaction layer, a residual connection layer, a strategy composition closed-loop verification layer, and a prediction output layer, mapping the initial candidate marketing strategy combination to a strategy composition embedding vector. The closed-loop verification module is used to generate residual interaction representations based on policy composition embedding vectors, perform forward reconstruction and reverse verification on the residual interaction representations, and generate closed-loop interaction representations and closed-loop breakpoint representations. The path training module is used to select two real strategy combinations with the same marketing scenario identifier and different conversion labels from the historical strategy combination samples, generate replacement path samples, and train the AutoInt model based on the historical strategy combination samples and replacement path samples to obtain the improved AutoInt model. The strategy rewriting module is used to input the initial candidate marketing strategy combination into the improved AutoInt model, calculate the interaction residual based on the closed-loop interaction representation and the closed-loop interaction benchmark, and generate a locally rewritten candidate marketing strategy combination based on the interaction residual. The strategy output module is used to input the partially rewritten candidate marketing strategy combination into the improved AutoInt model for re-scoring and output the target marketing strategy combination with the highest predicted effect value.

Images