E-commerce marketing roi prediction and delivery strategy generation method based on ai learning

Abstract

The application discloses an e-commerce marketing ROI prediction and delivery strategy generation method based on AI learning, and relates to the technical field of business marketing. It comprises the following steps: S1: merging the obtained aggregated conversion verification data, aggregated delivery logs and user behavior event streams to construct a corresponding cross-channel behavior embedding tensor; S2: constructing an end-to-end joint prediction model, taking the cross-channel behavior embedding tensor as the input of the end-to-end joint prediction model, and outputting the corresponding user conversion probability and ROI prediction value; S3: generating a generative adversarial network, setting a dynamic resource allocation optimizer, taking the user conversion probability and ROI prediction value as the input of the dynamic resource allocation optimizer, and outputting the corresponding real-time strategy instruction set. The application does not depend on complete user identification data of a single platform, and can more comprehensively and accurately depict user behavior paths even in the face of incomplete, high-dimensional and nonlinear e-commerce marketing data.

Description

Technical Field

[0001] This invention relates to the field of commercial marketing technology, specifically to a method for predicting e-commerce marketing ROI and generating advertising strategies based on AI learning. Background Technology

[0002] Currently, the industry mainly relies on the following technical means to evaluate and optimize the effectiveness of e-commerce marketing:

[0003] Traditional statistical models employ classic statistical methods such as linear regression, logistic regression, or time series analysis. Based on shallow features such as historical sales data, ad exposure, and click-through rates, they construct simple mathematical relationships to predict performance.

[0004] Rule engine based on human experience: Relies on the subjective experience of senior operations personnel to set fixed "if-then" rules (e.g., suspend campaigns when the cost per click exceeds a threshold X, or allocate the budget according to a fixed ratio).

[0005] Isolated A / B testing mechanism: Test the short-term effects (such as click-through rate and conversion rate) of different ad creatives, landing pages or bidding strategies with small traffic, and make manual decisions based on the test results.

[0006] While the above methods have assisted marketing decisions to some extent, they still face technical barriers such as insufficient prediction accuracy, slow response speed, strong strategy subjectivity, and lack of multi-objective optimization when dealing with the inherent characteristics of e-commerce data, which are high-dimensional, non-linear, and highly dynamic.

[0007] Chinese invention patent CN120952882A discloses a method and system for predicting and optimizing multi-channel advertising effectiveness based on artificial intelligence. The method includes the following steps: acquiring and preprocessing advertising delivery data, user behavior data, and external environment data; extracting features from the preprocessed data, mapping different types of content feature vectors to a unified semantic space using a cross-channel feature projection matrix to generate an advertising feature set; integrating a prediction model using Transformer and graph neural networks to simultaneously predict the conversion rate, return on investment, and user interaction metrics of the target advertisement on each channel based on the advertising feature set; constructing a reinforcement learning-based strategy optimization network, using the prediction results output by the prediction model as state input, and dynamically generating a budget allocation matrix and channel selection weights through a strategy gradient algorithm; and pushing the adjusted advertising delivery strategy to the advertising management platform according to the optimization strategy to execute advertising delivery. This invention can improve the accuracy and stability of advertising effectiveness prediction.

[0008] However, the aforementioned and similar technical solutions still have the following shortcomings: Due to the continuous strengthening of mobile internet privacy protection policies (such as Apple's ATT framework restricting the acquisition of device identifiers such as IDFA) and increasingly stringent data barriers between various commercial platforms, existing advertising monitoring and analysis systems struggle to obtain complete and accurate user cross-touchpoint behavioral data, thus preventing traditional attribution models from effectively constructing complete conversion paths. Simultaneously, existing privacy compliance solutions (such as probabilistic attribution and aggregated data) cannot meet the requirements of real-time dynamic delivery strategies for high-precision, low-latency data feedback, thereby hindering the accurate allocation of advertising budgets and the scientific evaluation of marketing ROI. Summary of the Invention

[0009] The purpose of this invention is to provide a method for predicting e-commerce marketing ROI and generating advertising strategies based on AI learning, so as to solve the problems mentioned in the background art.

[0010] To achieve the above objectives, the present invention provides the following technical solution: a method for predicting e-commerce marketing ROI and generating advertising strategies based on AI learning, comprising:

[0011] S1: Multi-source heterogeneous data fusion: By setting up a data fusion engine, the acquired aggregated conversion verification data, aggregated delivery logs and user behavior event streams are fused together to construct the corresponding cross-channel behavior embedding tensor;

[0012] S2: End-to-end model prediction: Combine the multi-task learning model and the Transformer encoder to construct an end-to-end joint prediction model, and use the cross-channel behavior embedding tensor as the input of the end-to-end joint prediction model, and output the corresponding user conversion probability and ROI prediction value.

[0013] S3: Generative simulation optimization: By using a generative adversarial network, a dynamic resource allocation optimizer is set up, and the user conversion probability and ROI prediction value are used as inputs to the dynamic resource allocation optimizer, and the corresponding real-time policy instruction set is obtained as output.

[0014] Furthermore, the corresponding cross-channel behavior embedding tensor is constructed, including:

[0015] S1.1: Data processing: Collect and obtain aggregate reports, aggregate delivery logs and user behavior event streams through the preset application interfaces of each platform, and perform format normalization processing on the aggregate reports, aggregate delivery logs and user behavior event streams to obtain the corresponding formatted aggregate reports, formatted aggregate delivery logs and formatted user behavior event streams. At the same time, determine the corresponding reliable conversion sequence through the preset rule base.

[0016] S1.2: Feature extraction: An enhanced ad sequence is constructed using the aggregated delivery logs and user behavior event stream. Both the enhanced ad sequence and the reliable conversion sequence are used as inputs to the time-series behavior embedding model, and the corresponding behavior context vector is output.

[0017] S1.3: Feature Fusion: Based on the temporal stability of data sources from each platform, set corresponding scoring weights, and combine the scoring weights with formatted aggregated reports, formatted aggregated delivery logs, and formatted user behavior event streams to obtain corresponding enhanced features. At the same time, through an attention-based neural network model, determine the correlation weights between the enhanced features and the behavior context vectors, and fuse the enhanced features according to the correlation weights. Through all the fused enhanced features, construct the corresponding cross-channel behavior embedding tensor.

[0018] Furthermore, the statistical time window corresponding to the formatted aggregation report is aligned and calibrated using the preset rule base to obtain the corresponding preprocessed aggregation report. Simultaneously, based on the acquisition channel and acquisition date corresponding to the preprocessed aggregation report, the click volume sequence of the media platform corresponding to the aggregation delivery log is determined. Based on the click volume sequence of the media platform, the estimated conversion number corresponding to the click volume of each time unit is determined. At the same time, the estimated conversion number corresponding to the click volume of each time unit is arranged and combined in an orderly manner according to the time sequence to obtain the corresponding reliable conversion sequence.

[0019] Furthermore, based on the click volume sequence of the media platform, the corresponding total click volume for the whole day and the click volume for each time unit are determined. At the same time, based on the percentage between the click volume for each time unit and the total click volume for the whole day, a basic weight corresponding to the click volume for each time unit is set. The basic weight is then combined with the set total daily conversions to determine the estimated conversions corresponding to the click volume for each time unit.

[0020] Furthermore, the formatted user behavior event stream is arranged in order by timestamp and segmented according to a preset segmentation time to obtain a corresponding ordered event list. At the same time, the corresponding context features are extracted from the formatted aggregated delivery logs, and the formatted user behavior event stream within the corresponding time period is determined based on the periodic features corresponding to the context features. The context features and the ordered event list corresponding to the corresponding formatted user behavior event stream are combined to construct the corresponding enhanced advertising sequence.

[0021] Furthermore, the fusion of the enhanced features includes:

[0022] S1.3.1: Score determination: Combine the authority level, expected data delay and expected data coverage obtained from each platform to obtain the corresponding static base score, and obtain the corresponding normalized static base score through data normalization processing. At the same time, determine the corresponding score weight based on the percentage between the normalized static base scores.

[0023] S1.3.2: Fusion Processing: The scoring weights are combined with the formatted aggregated report, formatted aggregated delivery log, and formatted user behavior event stream to obtain the corresponding enhanced features. At the same time, the enhanced features and the behavior context vector are used as inputs to the set attention-based neural network model, and the output obtains the correlation weights between the enhanced features and the behavior context vector.

[0024] S1.3.3: Dynamic Fusion: The relevant weights and enhanced features are combined to obtain the corresponding initial fusion features, and the initial fusion features are combined to construct the corresponding dynamic fusion features, specifically:

[0025]

[0026] in: As a dynamic fusion feature, For the i-th enhanced feature, Let be the correlation weight between the i-th enhanced feature and the behavioral context vector.

[0027] Furthermore, using a multi-task learning model as the basic model architecture, and setting the Transformer encoder as the bottom layer of the model, a corresponding shared feature encoder is constructed. At the same time, the shared feature encoder is connected with the constructed implicit transformation tendency prediction head and causal ROI estimation and attribution head to construct a corresponding end-to-end joint prediction model.

[0028] Furthermore, the multilayer perceptron is connected to the shared feature encoder, and the corresponding user conversion probability is obtained by outputting the sigmoid activation function, thus constructing the corresponding implicit conversion tendency prediction head;

[0029] By using small neural networks, regression networks, and differentiable deterministic computational units, and setting up corresponding propensity score networks, potential outcome prediction networks, and dual robust estimation computational layers, the corresponding causal ROI estimates and attribution heads are constructed.

[0030] Furthermore, the output retrieves the corresponding real-time policy instruction set, including:

[0031] S3.1: Generative Simulation: A generator and discriminator of a generative adversarial network are constructed through a deep neural network. The cross-channel behavior embedding tensor, ROI prediction value and user conversion probability are used as inputs to the generator, and the output is to obtain the corresponding synthetic strategy vector. At the same time, the corresponding real strategy and generation strategy are set through the historical database and synthetic strategy vector. The real strategy and generation strategy are used as inputs to the discriminator, and the output is to obtain the accuracy of the generation strategy. Based on the accuracy, a corresponding high-quality candidate strategy pool is constructed.

[0032] S3.2: Balance Determination: Combine the predicted ROI value with the corresponding budget cost to determine the incremental conversion value of each channel. Based on the incremental conversion value of each channel and the advertising expenditure of each channel, set the corresponding objective function. At the same time, set the corresponding mathematical constraints according to the actual business requirements. Solve the objective function through an online convex optimization algorithm and mathematical constraints, and obtain the corresponding real-time strategy instruction set from the high-quality candidate strategy pool.

[0033] Furthermore, the accuracy is compared with a preset accuracy threshold, and the synthesis strategy vector is filtered based on the comparison result, specifically as follows:

[0034] When the accuracy is less than a preset accuracy threshold, the synthetic strategy vector corresponding to the accuracy is deleted; otherwise, the synthetic strategy vector corresponding to the accuracy is retained, and a corresponding high-quality candidate strategy pool is constructed based on the retained synthetic strategy vector.

[0035] Compared with the prior art, the beneficial effects of the present invention are:

[0036] Firstly, this invention integrates data from advertising platforms, media platforms, and e-commerce platforms through a multi-source heterogeneous data fusion engine, and constructs a cross-channel behavior embedding tensor. This allows it to address the challenge of incomplete user behavior data caused by privacy protection policies and platform data barriers without relying on complete user identification data from a single platform. Furthermore, even when faced with incomplete, high-dimensional, and non-linear e-commerce marketing data, it can more comprehensively and accurately depict user behavior paths, thus providing a higher-quality data foundation for subsequent ROI prediction and strategy generation.

[0037] Secondly, this invention combines the Transformer encoder with a multi-task learning model to construct an end-to-end joint prediction model. The task heads of the model predict the user conversion probability and the ROI prediction value, respectively. This achieves joint learning of implicit conversion tendency prediction and causal ROI estimation. It can not only predict whether users will convert, but also evaluate the real incremental value of ad exposure to conversion from the perspective of causal inference. This surpasses the traditional correlation model, making the ROI prediction results more scientific and the attribution more reliable, providing a solid quantitative basis for budget allocation.

[0038] Thirdly, this invention constructs a dynamic resource allocation optimizer through generative adversarial networks. The generator generates candidate strategies based on real-time data, while the discriminator selects a pool of high-quality candidate strategies based on historical data. Combined with an online convex optimization algorithm, it solves the objective function that maximizes total marketing revenue while satisfying business constraints such as total budget, budgets for each channel, and minimum ROI. The output is a real-time strategy instruction set, which enables the dynamic and real-time generation and optimization of advertising strategies. It automatically adapts to market changes and avoids the lag and subjectivity of relying on manual experience rules and static A / B testing, thereby improving the efficiency of marketing resource allocation and return on investment. Attached Figure Description

[0039] Figure 1 This is a flowchart illustrating the e-commerce marketing ROI prediction and campaign strategy generation method of the present invention.

[0040] Figure 2 This is a schematic diagram of the process for obtaining the behavior context vector in this invention;

[0041] Figure 3 This is a schematic diagram of the process for obtaining dynamic fusion features in this invention;

[0042] Figure 4 This is a schematic diagram illustrating the construction process of the end-to-end joint prediction model in this invention;

[0043] Figure 5 This is a flowchart illustrating the generative simulation optimization method of the present invention. Detailed Implementation

[0044] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0045] refer to Figure 1This embodiment provides a method for predicting e-commerce marketing ROI and generating advertising strategies based on AI learning. The method specifically includes the following steps:

[0046] Step S1: Multi-source heterogeneous data fusion. This involves obtaining aggregated conversion verification data from advertising platforms, aggregated delivery logs such as ad impressions and clicks from media platforms, and user behavior event streams (such as browsing and adding to cart) from e-commerce platforms. The acquired aggregated conversion verification data, aggregated delivery logs, and user behavior event streams are then fused together using a data fusion engine to construct the corresponding cross-channel behavior embedding tensor.

[0047] Step S2: End-to-end model prediction. This involves using a multi-task learning model as the basic model architecture and combining it with a Transformer encoder to construct the corresponding end-to-end joint prediction model. Simultaneously, the cross-channel behavior embedding tensor obtained in Step S1 is used as the input to the end-to-end joint prediction model, outputting the corresponding user conversion probability and ROI prediction value.

[0048] refer to Figure 4 Specifically, a multi-task learning model is used as the basic model architecture, and the Transformer encoder is set as the bottom layer of the model to construct the corresponding shared feature encoder. Simultaneously, a multilayer perceptron (composed of multiple fully connected layers) is connected to the shared feature encoder. That is, through the Transformer encoder, the corresponding general features are extracted from the cross-channel behavior embedding tensor constructed in step S1.3.3. These extracted general features are then used as input to the multilayer perceptron, and the corresponding user conversion probability is obtained through the Sigmoid activation function output. This probability represents the likelihood of a user converting within a preset time window (which can be specifically set according to actual needs, and is therefore not specifically described in this embodiment). In other words, through the configured multilayer perceptron and Sigmoid activation function, a corresponding implicit conversion tendency prediction head is constructed.

[0049] Furthermore, a propensity score network is set up using a small neural network to obtain the estimated probability of a user being reached by a specific advertisement through the general features extracted by the Transformer encoder. Simultaneously, a latent outcome prediction network is set up using a regression network to obtain the observed results (i.e., processing results) of user exposure to advertisements and the counterfactual results (i.e., control results) of user non-exposure to advertisements, using the general features extracted by the Transformer encoder. A corresponding dual robust estimation computation layer is also set up using a differentiable deterministic computation unit to combine the obtained user conversion probability, the estimated probability of a user being reached by a specific advertisement, the observed results of user exposure to advertisements, and the counterfactual results of user non-exposure to advertisements. The corresponding ROI prediction value is obtained through the mathematical formula of the dual robust estimator (which is a conventional mathematical formula in this field and is not specifically described in this embodiment). In other words, by constructing the propensity score network, the latent outcome prediction network, and the dual robust estimation computation layer, the corresponding causal ROI estimation and attribution head are constructed.

[0050] Specifically, a multi-task learning model is used as the basic model architecture, with the Transformer encoder set as the bottom layer. The constructed implicit conversion tendency prediction head and causal ROI estimation and attribution heads are connected to the Transformer encoder, and the implicit conversion tendency prediction head is connected to the causal ROI estimation and attribution heads, thereby constructing the corresponding end-to-end joint prediction model. That is, the cross-channel behavior embedding tensor constructed in step S1.3.3 is used as the input of the end-to-end joint prediction model, and the corresponding user conversion probability and ROI prediction value are obtained as the output.

[0051] Step S3: Generative simulation optimization. This involves setting up a corresponding dynamic resource allocation optimizer using a generative adversarial network (GAN). The user conversion probability and ROI prediction obtained from the output of Step S2 are used as inputs to the dynamic resource allocation optimizer to obtain the corresponding real-time budget allocation, audience targeting, and bidding strategies. These strategies are then automatically deployed and executed via the advertising platform's API.

[0052] In this embodiment, a data fusion engine is used to fuse the acquired aggregated conversion verification data, aggregated delivery logs, and user behavior event streams to construct a corresponding cross-channel behavior embedding tensor. (See reference...) Figure 2 and Figure 3 This embodiment provides a method for fusing multi-source heterogeneous data, which specifically includes the following steps:

[0053] Step S1.1: Data Processing. This involves connecting to the data warehouses of advertising platforms (such as Meta Ads, Google Ads, and Apple Search Ads), media platforms (such as major traffic platforms), and e-commerce platforms through pre-defined application programming interfaces (APIs) conforming to the specifications of each platform. This allows for information exchange and the acquisition of aggregated reports from the advertising platform's privacy framework (such as SKAdNetwork), aggregated delivery logs (including ad impressions and clicks) from the media platform, and user behavior event streams from the e-commerce platform. Specifically, a unified data model is used to normalize the format of the acquired aggregated reports, aggregated delivery logs, and user behavior event streams to obtain corresponding formatted aggregated reports, formatted aggregated delivery logs, and formatted user behavior event streams. In this embodiment, the unified data model includes core fields corresponding to each event, including an identifier ID, timestamp, event type, channel identifier, and contextual information (such as ad placement and device type). Notably, during the acquisition of the aggregated reports from the advertising platform's privacy framework, a key set by the advertising platform is used to decrypt the acquired aggregated reports to obtain the corresponding structured aggregated data.

[0054] Furthermore, based on the attribution logic set by each advertising platform, a corresponding preset rule base is constructed (which can be specifically set according to actual needs, so it is not specifically described in this embodiment). Based on the constructed preset rule base, the statistical time window corresponding to the obtained formatted aggregation report is aligned and calibrated to obtain the corresponding preprocessed aggregation report, thereby ensuring that the time window corresponding to the preprocessed aggregation report and the aggregation delivery log obtained by the media platform is comparable.

[0055] Furthermore, based on the acquisition channel corresponding to the preprocessed aggregated report, the corresponding media platform is determined, and the corresponding formatted aggregated delivery logs are obtained. Simultaneously, based on the acquisition date corresponding to the preprocessed aggregated report, the corresponding click volume sequence is extracted from the obtained formatted aggregated delivery logs (e.g., 0-1 AM: 500 clicks; 1-2 AM: 300 clicks). Specifically, based on the extracted click volume sequence, the corresponding total daily clicks and time unit clicks are determined, and the percentage between the clicks in each time unit and the total daily clicks is obtained (e.g., if the total daily clicks are 10,000, then the percentage for 0-1 AM is 5%, and the percentage for 1-2 AM is 3%). Based on the percentage corresponding to the clicks in each time unit, a base weight is set for the clicks in each time unit (i.e., the base weight for 0-1 AM is 5%, and the base weight for 1-2 AM is 3%). Simultaneously, based on the base weight corresponding to the click volume of each time unit and the corresponding set total daily conversions (which can be specifically set according to actual needs, so it is not specifically described in this embodiment, such as 100 total conversions), the estimated conversions corresponding to the click volume of each time unit are determined (i.e., the estimated conversions for points 0-1 are: 100 * 5% = 5 times, and the estimated conversions for points 1-2 are: 100 * 3% = 3 times). In other words, the estimated conversions corresponding to the click volume of each time unit are arranged and combined in an orderly manner according to the time sequence to obtain the corresponding reliable conversion sequence.

[0056] It is worth noting that when determining the estimated conversion number corresponding to the clicks for each time unit, the estimated conversion number can be rounded down using the rounding and maximum remainder methods. Simultaneously, the sum of the estimated conversion numbers corresponding to clicks for all time units should be equal to the set total daily conversion number.

[0057] Step S1.2: Feature Extraction. The formatted user behavior event stream obtained in Step S1.1 is arranged in order by timestamps. A preset segmentation time (which can be set according to actual needs, but is not specifically described in this embodiment, e.g., 30 minutes) is used to segment the ordered formatted user behavior event stream (e.g., if the time difference between two consecutive events exceeds 30 minutes, they belong to two different sessions) to obtain the corresponding ordered event list. Simultaneously, based on the formatted aggregated delivery logs obtained in Step S1.1, corresponding contextual features are extracted, including category features (e.g., creative ID, media type, ad placement, and device system) and periodic features (e.g., the specific week number (Monday to Sunday) and the specific time of day (0:00-24:00)). Specifically, based on the periodic features in the contextual features extracted from the formatted aggregated delivery logs, the formatted user behavior event stream within the corresponding time period is determined. The ordered event list corresponding to the determined formatted user behavior event stream is combined with the contextual features corresponding to the corresponding formatted aggregated delivery logs to construct the corresponding enhanced ad sequence.

[0058] Furthermore, by acquiring historical aggregation reports, historical aggregation delivery logs, and historical user behavior event streams, the set sequence model (such as a lightweight Transformer Encoder) is trained to obtain the corresponding temporal behavior embedding model. At the same time, the reliable conversion sequence obtained in step S1.1 and the constructed enhanced advertising sequence are both used as inputs to the temporal behavior embedding model to output the corresponding behavior context vector.

[0059] Step S1.3: Feature Fusion. This involves setting corresponding scoring weights based on the temporal stability of the data sources from advertising platforms, media platforms, and e-commerce platforms. Based on these weights, the formatted aggregated report, formatted aggregated delivery log, and formatted user behavior event stream obtained in Step S1.1 are augmented to obtain corresponding enhanced features. Simultaneously, an attention-based neural network model is used to determine the correlation weights between the obtained enhanced features and the behavioral context vectors output in Step S1.2. Based on these determined correlation weights, all enhanced features are fused to construct a corresponding cross-channel behavioral embedding tensor. Details are as follows:

[0060] Step S1.3.1: Score Determination. This involves determining the score based on the authority level (e.g., HIGH (1.0), MEDIUM (0.9), LOW (0.7) of the advertising platform, media platform, and e-commerce platform, which can be specifically set according to actual needs and is not detailed in this embodiment), expected data latency, and expected data coverage (e.g., the estimated proportion of platform traffic to the overall market). The authority level, expected data latency, and expected data coverage are combined to obtain the corresponding static base score, including the static base score for the advertising platform, the static base score for the media platform, and the static base score for the e-commerce platform.

[0061] Furthermore, based on the obtained static base scores, data normalization processing is performed (this is a conventional data processing method, so it is not specifically described in this embodiment) to obtain the corresponding normalized static base scores, namely, the normalized static base scores for the advertising platform, media platform, and e-commerce platform. Simultaneously, the obtained normalized static base scores for the advertising platform, media platform, and e-commerce platform are combined to determine the sum of the corresponding base scores. The corresponding scoring weights are then determined based on the percentage between the normalized static base scores and the sum of the base scores. In other words, the advertising platform scoring weight is obtained based on the percentage between the normalized static base score for the advertising platform and the sum of the base scores. The media platform scoring weight is obtained based on the percentage between the normalized static base score for the media platform and the sum of the base scores. The e-commerce platform scoring weight is obtained based on the percentage between the normalized static base score for the e-commerce platform and the sum of the base scores.

[0062] Step S1.3.2: Fusion Processing. Based on the advertising platform rating weights, media platform rating weights, and e-commerce platform rating weights obtained in Step S1.3.1, the formatted aggregated report, formatted aggregated delivery logs, and formatted user behavior event streams obtained in Step S1.1 are enhanced to obtain corresponding enhanced features. Specifically, the advertising platform rating weights and the formatted aggregated report obtained in Step S1.3.1 are combined to obtain the corresponding enhanced aggregated report. The media platform rating weights and the formatted aggregated delivery logs obtained in Step S1.3.1 are combined to obtain the corresponding enhanced aggregated delivery logs. The e-commerce platform rating weights and the formatted user behavior event streams obtained in Step S1.3.1 are combined to obtain the corresponding enhanced user behavior event streams.

[0063] Furthermore, the obtained enhanced aggregation report, enhanced aggregation delivery log, enhanced user behavior event stream, and behavior context vector obtained in step S1.2 are all used as inputs to the attention-based neural network model to output the correlation weights between the enhanced features and the behavior context vector, specifically:

[0064]

[0065] in: The weights are the correlation weights between the i-th augmented feature and the behavioral context vector. This represents the query vector corresponding to the behavioral context vector in an attention-based neural network model. Let be the transpose of the key vector corresponding to the i-th enhanced feature in the attention-based neural network model. This represents the scaling factor in the attention-based neural network model. To enhance the vector dimension of features and query vectors, It is a normalized exponential function.

[0066] Step S1.3.3: Dynamic Fusion. Based on the relevant weights obtained from the output of Step S1.3.2, including the weights related to the enhanced aggregated report and behavior context vector, the weights related to the enhanced aggregated delivery log and behavior context vector, and the weights related to the enhanced user behavior event stream and behavior context vector, these weights are combined with the corresponding enhanced features (i.e., enhanced aggregated report, enhanced aggregated delivery log, and enhanced user behavior event stream) to obtain the corresponding initial fusion features, including initial aggregated report fusion features, initial aggregated delivery log fusion features, and initial user behavior event stream fusion features. Simultaneously, the obtained initial aggregated report fusion features, initial aggregated delivery log fusion features, and initial user behavior event stream fusion features are combined to construct the corresponding dynamic fusion features, specifically:

[0067]

[0068] in: As a dynamic fusion feature, For the i-th enhanced feature, Let be the correlation weight between the i-th enhanced feature and the behavioral context vector.

[0069] Furthermore, based on the constructed dynamic fusion features, the dynamic fusion features are concatenated and combined with the formatted aggregate report, formatted aggregate delivery log and formatted user behavior event stream obtained in step S1.1 and the behavior context vector obtained in step S1.2 to construct the corresponding cross-channel behavior embedding tensor.

[0070] In this embodiment, the user conversion probability and ROI prediction value obtained in step S2 are used as inputs to the dynamic resource allocation optimizer to obtain the corresponding real-time budget allocation, audience targeting, and bidding strategies. (See reference) Figure 5 This embodiment provides a generative simulation optimization method, which specifically includes the following steps:

[0071] Step S3.1: Generative Simulation. This involves constructing a generator and discriminator for a generative adversarial network using a deep neural network (such as a recurrent neural network). Specifically, the current available total budget, the cross-channel behavior embedding tensor constructed in step S1.1.3, and the ROI prediction and user conversion probability obtained from the output in step S2 are used as inputs to the generator to output a corresponding synthetic strategy vector, including the corresponding budget allocation ratio, target audience package combination, and base bidding coefficient. It is worth noting that the synthetic strategy vector obtained by the generator output in this embodiment is not limited to one; its specific number can be set according to the actual data, so it is not specifically described in this embodiment.

[0072] Furthermore, by using a historical database, corresponding historical strategies and results are obtained, and a corresponding real strategy is set based on these historical strategies and results. Simultaneously, a corresponding generation strategy is set based on the synthesized strategy vector obtained from the generator's output. Both the set real strategy and the generation strategy are used as input to a discriminator, and the discriminator outputs the accuracy corresponding to the generation strategy. The obtained accuracy is then compared with a preset accuracy threshold (which can be specifically set according to actual needs, and is not specifically described in this embodiment). Based on the comparison result, the synthesized strategy vector obtained from the generator's output is filtered, specifically as follows:

[0073] If the accuracy of the output is less than a preset accuracy threshold, the synthesized policy vector corresponding to that accuracy is deleted. Conversely, if the accuracy of the output is not less than the preset accuracy threshold, the synthesized policy vector corresponding to that accuracy is retained, and a pool of high-quality candidate policies is constructed based on the retained synthesized policy vectors.

[0074] Step S3.2: Balance Determination. Based on the ROI prediction value and corresponding budget cost obtained in Step S2, the ROI prediction value and budget cost are combined to determine the incremental conversion value for each channel. Simultaneously, based on the incremental conversion value and advertising expenditure for each channel, the sum of the incremental conversion value and the sum of the advertising expenditure are obtained. The difference between the sum of the incremental conversion value and the sum of the advertising expenditure is used to determine the corresponding total marketing revenue. The objective function is set based on maximizing the total marketing revenue. In this embodiment, corresponding mathematical constraints are set according to the actual business requirements, including a total budget upper limit, upper and lower limits for each channel's budget, a minimum ROI value, and user coverage breadth. In other words, through the set mathematical constraints and objective function, a mathematical model for the corresponding optimization problem is constructed.

[0075] Furthermore, by using an online convex optimization algorithm (which is a conventional technique and therefore not specifically described in this embodiment), the mathematical model of the optimization problem constructed in step S3.2 is solved based on the high-quality candidate policy pool obtained in step S3.1, so as to obtain the corresponding optimal solution from the high-quality candidate policy pool, and the obtained optimal solution is the corresponding real-time policy instruction set.

[0076] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended embodiments and their equivalents.

Claims

1. A method for predicting e-commerce marketing ROI and generating advertising strategies based on AI learning, characterized in that, Including: S1: Multi-source heterogeneous data fusion: By setting up a data fusion engine, the acquired aggregated conversion verification data, aggregated delivery logs and user behavior event streams are fused together to construct the corresponding cross-channel behavior embedding tensor; S2: End-to-end model prediction: Combine the multi-task learning model and the Transformer encoder to construct an end-to-end joint prediction model, and use the cross-channel behavior embedding tensor as the input of the end-to-end joint prediction model to output the corresponding user conversion probability and ROI prediction value. S3: Generative simulation optimization: By using a generative adversarial network, a dynamic resource allocation optimizer is set up, and the user conversion probability and ROI prediction value are used as inputs to the dynamic resource allocation optimizer, and the corresponding real-time policy instruction set is obtained as output.

2. The method for predicting e-commerce marketing ROI and generating advertising strategies based on AI learning according to claim 1, characterized in that, The corresponding cross-channel behavior embedding tensor is constructed, including: S1.1: Data processing: Collect and obtain aggregate reports, aggregate delivery logs and user behavior event streams through the preset application interfaces of each platform, and perform format normalization processing on the aggregate reports, aggregate delivery logs and user behavior event streams to obtain the corresponding formatted aggregate reports, formatted aggregate delivery logs and formatted user behavior event streams. At the same time, determine the corresponding reliable conversion sequence through the preset rule base. S1.2: Feature extraction: An enhanced ad sequence is constructed using the aggregated delivery logs and user behavior event stream. Both the enhanced ad sequence and the reliable conversion sequence are used as inputs to the time-series behavior embedding model, and the corresponding behavior context vector is output. S1.3 Feature Fusion: Based on the temporal stability of data sources from each platform, corresponding scoring weights are set, and these scoring weights are combined with formatted aggregated reports, formatted aggregated delivery logs, and formatted user behavior event streams to obtain corresponding enhanced features. At the same time, through an attention-based neural network model, the enhanced features and the behavior context vector are determined as the correlation weights between them. Based on the correlation weights, the enhanced features are fused, and through all the fused enhanced features, a corresponding cross-channel behavior embedding tensor is constructed.

3. The method for predicting e-commerce marketing ROI and generating advertising strategies based on AI learning according to claim 2, characterized in that, Using the preset rule base, the statistical time window corresponding to the formatted aggregation report is aligned and calibrated to obtain the corresponding preprocessed aggregation report. At the same time, based on the acquisition channel and acquisition date corresponding to the preprocessed aggregation report, the click volume sequence of the media platform corresponding to the aggregation delivery log is determined. Based on the click volume sequence of the media platform, the estimated conversion number corresponding to the click volume of each time unit is determined. Furthermore, according to the time sequence, the estimated conversion number corresponding to the click volume of each time unit is arranged and combined in an orderly manner to obtain the corresponding reliable conversion sequence.

4. The method for predicting e-commerce marketing ROI and generating advertising strategies based on AI learning according to claim 3, characterized in that, Based on the click volume sequence of the media platform, the corresponding total click volume for the whole day and the click volume for each time unit are determined. At the same time, based on the percentage between the click volume for each time unit and the total click volume for the whole day, a basic weight corresponding to the click volume for each time unit is set. The basic weight is then combined with the set total daily conversions to determine the estimated conversions corresponding to the click volume for each time unit.

5. The method for predicting e-commerce marketing ROI and generating advertising strategies based on AI learning according to claim 2, characterized in that, The formatted user behavior event stream is arranged in order by timestamp and then segmented according to a preset segmentation time to obtain the corresponding ordered event list. At the same time, the corresponding context features are extracted from the formatted aggregated delivery logs, and the formatted user behavior event stream within the corresponding time period is determined based on the periodic features corresponding to the context features. The context features and the ordered event list corresponding to the corresponding formatted user behavior event stream are combined to construct the corresponding enhanced advertising sequence.

6. The method for predicting e-commerce marketing ROI and generating advertising strategies based on AI learning according to claim 2, characterized in that, The fusion of the enhanced features includes: S1.3.1: Score determination: Combine the authority level, expected data delay and expected data coverage obtained from each platform to obtain the corresponding static base score, and obtain the corresponding normalized static base score through data normalization processing. At the same time, determine the corresponding score weight based on the percentage between the normalized static base scores. S1.3.2: Fusion Processing: The scoring weights are combined with the formatted aggregated report, formatted aggregated delivery log, and formatted user behavior event stream to obtain the corresponding enhanced features. At the same time, the enhanced features and the behavior context vector are used as inputs to the set attention-based neural network model, and the output obtains the correlation weights between the enhanced features and the behavior context vector. S1.3.3: Dynamic Fusion: The relevant weights and enhanced features are combined to obtain the corresponding initial fusion features, and the initial fusion features are combined to construct the corresponding dynamic fusion features, specifically: in: As a dynamic fusion feature, For the i-th enhanced feature, Let be the correlation weight between the i-th enhanced feature and the behavioral context vector.

7. The method for predicting e-commerce marketing ROI and generating advertising strategies based on AI learning according to claim 1, characterized in that, Using a multi-task learning model as the basic model architecture, and setting the Transformer encoder as the bottom layer of the model, a corresponding shared feature encoder is constructed. At the same time, the shared feature encoder is connected with the constructed implicit transformation tendency prediction head and causal ROI estimation and attribution head to construct a corresponding end-to-end joint prediction model.

8. The method for predicting e-commerce marketing ROI and generating advertising strategies based on AI learning according to claim 7, characterized in that, The multilayer perceptron is connected to the shared feature encoder, and the corresponding user conversion probability is obtained by outputting the sigmoid activation function to construct the corresponding implicit conversion tendency prediction head; By using small neural networks, regression networks, and differentiable deterministic computational units, and setting up corresponding propensity score networks, potential outcome prediction networks, and dual robust estimation computational layers, the corresponding causal ROI estimates and attribution heads are constructed.

9. The method for predicting e-commerce marketing ROI and generating advertising strategies based on AI learning according to claim 1, characterized in that, The output retrieves the corresponding real-time policy instruction set, including: S3.1: Generative Simulation: A generator and discriminator of a generative adversarial network are constructed through a deep neural network. The cross-channel behavior embedding tensor, ROI prediction value and user conversion probability are used as inputs to the generator, and the output is to obtain the corresponding synthetic strategy vector. At the same time, the corresponding real strategy and generation strategy are set through the historical database and synthetic strategy vector. The real strategy and generation strategy are used as inputs to the discriminator, and the output is to obtain the accuracy of the generation strategy. Based on the accuracy, a corresponding high-quality candidate strategy pool is constructed. S3.2: Balance Determination: Combine the predicted ROI value with the corresponding budget cost to determine the incremental conversion value of each channel. Based on the incremental conversion value of each channel and the advertising expenditure of each channel, set the corresponding objective function. At the same time, set the corresponding mathematical constraints according to the actual business requirements. Solve the objective function through an online convex optimization algorithm and mathematical constraints, and obtain the corresponding real-time strategy instruction set from the high-quality candidate strategy pool.

10. The method for predicting e-commerce marketing ROI and generating advertising strategies based on AI learning according to claim 9, characterized in that, The accuracy is compared with a preset accuracy threshold, and the synthesis strategy vector is filtered based on the comparison result, specifically as follows: When the accuracy is less than a preset accuracy threshold, the synthetic strategy vector corresponding to the accuracy is deleted; otherwise, the synthetic strategy vector corresponding to the accuracy is retained, and a corresponding high-quality candidate strategy pool is constructed based on the retained synthetic strategy vector.

Images