Advertisement putting ROI real-time prediction and optimization method based on deep learning driving

Abstract

The invention discloses an advertisement putting ROI real-time prediction and optimization method based on deep learning driving, and relates to the technical field of advertisement prediction. After the uniform feature tensor and static features in the putting state abstract, known future features and historical observable features are coded at the same time, key driving factors are selected through an attention and gating structure, parallel prediction intermediate quantities are obtained through parallel prediction, the parallel prediction intermediate quantities are fused into a risk annotation ROI prediction result, and the risk annotation ROI prediction result is obtained. Generating an actionable prediction list, and correcting the distribution difference in the actionable prediction list by using an exposure click generation mechanism; by constructing the multi-time-scale depth predictor, the risk labeling ROI prediction result is generated, joint interactive modeling of three types of features is achieved, and the interpretability and stability of multi-time-scale profit and cost trend prediction are improved.

Description

Technical Field

[0001] This invention relates to the field of advertising prediction technology, and in particular to a method for real-time prediction and optimization of advertising ROI based on deep learning. Background Technology

[0002] With the rapid development of programmatic advertising and real-time bidding technologies, the advertising field has shifted from traditional human experience-driven to data-driven automated decision-making. Early ad optimization primarily relied on machine learning models such as logistic regression and gradient boosting decision trees to independently predict click-through rates and conversion rates, combined with linear or rule-based bidding strategies to allocate budgets. In recent years, deep learning models, due to their advantages in modeling high-dimensional sparse feature interactions, have been widely applied to ad performance prediction, further improving prediction accuracy. Simultaneously, reinforcement learning frameworks have begun to be used for dynamic bidding optimization, attempting to directly maximize long-term ROI.

[0003] However, existing methods also have room for improvement. First, most models rely on static features or short-term history, making it difficult to effectively capture the joint interaction between static user attributes, known future constraints, and historical observable sequences. This results in a lack of interpretability and stability in the prediction of multi-timescale revenue and cost trends, and is prone to distribution bias. Second, bidding strategies in real-time bidding scenarios often lead to sample selection bias, with high-bid slices being overexposed while low-potential slices have sparse data. The training distribution is inconsistent with the online prediction distribution, further amplifying the prediction error. In addition, the delayed feedback of conversion value signals leads to underestimation of value and deviation in ROI prediction. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a deep learning-driven method for real-time prediction and optimization of advertising ROI to solve the problems of lack of interpretability and stability in prediction, and inconsistency between the training distribution and the online prediction distribution.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: This invention provides a method for real-time prediction and optimization of advertising ROI based on deep learning, which includes: Real-time signals are collected, preprocessed, and written into an online feature library to generate a unified feature tensor and a delivery status summary. A multi-timescale deep predictor is constructed. After simultaneously encoding static features, known future features, and historical observable features in the unified feature tensor and the delivery state summary, key driving factors are selected through attention and gating structures. Parallel prediction intermediates are obtained through parallel prediction, and the parallel prediction intermediates are merged into risk-labeled ROI prediction results to generate an actionable prediction list. After correcting the distribution differences in the actionable prediction list using the exposure click generation mechanism, the delayed compensation layer is used to complete the actionable prediction list for any missing value signals, and the feasibility of the risk slices in the actionable prediction list is adjusted to obtain a corrected slice-level opportunity map. By using a layered collaborative optimization approach, the slice-level opportunity map is sequentially allocated with coarse-grained budget, adjusted with fine-grained bidding and frequency control, and selected and rotated in combination of creative materials. After generating a real-time delivery instruction package, it is sent to the advertising platform for execution, and the results are received back from the advertising platform.

[0007] As a preferred embodiment of the deep learning-driven real-time prediction and optimization method for advertising ROI, as described in this invention, the step of collecting real-time signals, preprocessing the real-time signals, and writing them into an online feature library specifically involves: Based on the configuration entry set by the advertising entity, it collects impressions, clicks, add-to-cart, order placement, and payment as real-time signals; The configuration entry includes the target ROI range, budget limit, allowed channels and allowed material sets, and cost constraint threshold; After deduplication, anti-fraud filtering, timestamp alignment and missing completion of the collected real-time signals at the edge nodes, user sequence behavior features, material multimodal representation features, channel and placement context features, competition intensity features and traffic fluctuation features are obtained and uniformly written into the online feature library.

[0008] As a preferred embodiment of the deep learning-driven real-time prediction and optimization method for advertising ROI, as described in this invention, the generation of a unified feature tensor and a delivery status summary specifically includes: Organize user sequence behavior features, multimodal representation features of creative materials, channel and placement context features, competition intensity features, and traffic fluctuation features written into the online feature library into a unified feature tensor by channel-audience-materials-time period slices; It aggregates campaign spending, impressions, clicks, and conversions to generate a campaign status summary.

[0009] As a preferred embodiment of the deep learning-driven real-time prediction and optimization method for advertising ROI, as described in this invention, the method involves simultaneously encoding static features, known future features, and historical observable features in a unified feature tensor and a delivery status summary, and selecting key driving factors through attention and gating structures. Specifically: Extract static features, known future features, and historical observable features from a unified feature tensor and a delivery state summary; After performing embedding encoding on static features, constant padding encoding on known future features, and sequence encoding on historical observable features, the encoding results of static features, known future features, and historical observable features are jointly organized to form a joint encoding representation. The correlation weights between feature components in the joint encoding representation are calculated through an attention mechanism, and key driving factors are selected by controlling the flow of information in the short, medium and long term through a gating structure.

[0010] As a preferred embodiment of the deep learning-driven real-time prediction and optimization method for advertising ROI, as described in this invention, the step of obtaining intermediate values ​​for parallel predictions through parallel predictions, merging these intermediate values ​​into a risk-labeled ROI prediction result, and generating an actionable prediction list specifically involves: Key drivers are used as shared inputs for multiple prediction branches, and short-term conversion tendency, expected value per unit conversion, and cost per unit exposure click intensity are output in parallel by multiple prediction branches as intermediate quantities for parallel prediction. The intermediate quantities of parallel predictions are fused at the output layer to generate risk-labeled ROI prediction results for short-term, medium-term, long-term, and cost trends. After calculating confidence weights and labeling risk levels, an actionable prediction list is obtained.

[0011] As a preferred embodiment of the deep learning-driven real-time prediction and optimization method for advertising ROI, the method described in this invention includes: correcting the distribution differences in the actionable prediction list by utilizing the exposure-click generation mechanism, which means recording the exposure-click generation mechanism information, calculating the distribution differences between the training distribution and the predicted distribution in the actionable prediction list, limiting the correction range according to the allowed channels and allowed material sets in the configuration entry, and correcting the distribution differences within the limited correction range.

[0012] As a preferred embodiment of the deep learning-driven real-time prediction and optimization method for advertising ROI, as described in this invention, the step of supplementing the actionable prediction list with unreturned value signals through a delay compensation layer and adjusting the feasibility of risk slices in the actionable prediction list refers to establishing a delay compensation layer, calculating the occurrence probability and expected value of unreturned value signals in the actionable prediction list, and after supplementing the actionable prediction list with value, reducing the feasibility weight of high-risk slices in the actionable prediction list and increasing the feasibility weight of low-exposure, high-potential slices, thereby generating a corrected slice-level opportunity map.

[0013] As a preferred embodiment of the deep learning-driven real-time prediction and optimization method for advertising ROI, the method of sequentially allocating budget in a coarse-grained manner through hierarchical collaborative optimization refers to identifying high-potential slices in the corrected slice-level opportunity map, prioritizing the movement of budget shares to high-potential slices at the channel and placement levels, obtaining updated budget shares at each channel and placement level, and completing the coarse-grained allocation.

[0014] As a preferred embodiment of the deep learning-driven real-time prediction and optimization method for advertising ROI, as described in this invention, the fine-grained adjustment of bidding and frequency control, as well as the combination and rotation of creative materials, generate a real-time delivery instruction package, specifically as follows: Within each slice of the calibrated slice-level opportunity map, the bidding range and frequency control parameters are adjusted according to the cost constraint threshold to obtain the adjusted bidding range and frequency control parameters. The adjusted bid range and frequency control parameters are used as the basis for material allocation. Material combination selection and rotation are performed within the allowed material set. High-feasibility materials in the allowed material set are prioritized for allocation to high-budget slices that have undergone a coarse-grained budget process, resulting in material allocation results. The results of material allocation, the adjusted bid range and frequency control parameters, as well as the updated budget shares of each channel and the placement level are linked to generate a real-time placement instruction package.

[0015] As a preferred embodiment of the deep learning-driven real-time prediction and optimization method for advertising ROI, as described in this invention, the step of sending the execution to the advertising platform and receiving the feedback results from the advertising platform refers to sending a real-time delivery instruction package through the interface of the advertising platform, and after setting the time granularity and event triggering conditions, executing the budget share, bid range, frequency control parameters and material allocation results in the real-time delivery instruction package through the advertising platform, and collecting the consumption, exposure, click and conversion results fed back by the advertising platform in the same window.

[0016] The beneficial effects of this invention are as follows: By constructing a multi-timescale deep predictor, risk-labeled ROI prediction results are generated, realizing joint interactive modeling of three types of features, improving the interpretability and stability of multi-timescale revenue and cost trend prediction, and effectively alleviating the distribution offset problem; at the same time, by correcting the distribution differences of the actionable prediction list through the exposure click generation mechanism and supplementing the value signals that have not yet been returned through the delay compensation layer, and by combining feasibility adjustment to generate a corrected slice-level opportunity map, the value underestimation caused by sample selection bias and conversion delay under real-time bidding is solved, further reducing prediction errors and improving the accuracy and reliability of ROI prediction. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a deep learning-driven method for real-time prediction and optimization of advertising ROI.

[0019] Figure 2 A flowchart for generating an actionable prediction list.

[0020] Figure 3 This is a flowchart for distribution difference correction and delay compensation.

[0021] Figure 4 Optimize the flowchart for hierarchical collaboration. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0025] Reference Figures 1-4 This is one embodiment of the present invention, which provides a method for real-time prediction and optimization of advertising ROI based on deep learning, including the following steps: S1: Collect real-time signals, preprocess the real-time signals and write them into the online feature library to generate a unified feature tensor and a delivery status summary; S1.1: Based on the target ROI range, budget limit, allowed channels, allowed creative sets, cost constraint threshold, and exposure and short-term revenue trend thresholds for low-exposure and high-potential ranges set by the advertising entity through the configuration entry, continuous data collection is performed on exposures, clicks, add-to-cart, orders, and payments generated during the advertising campaign. The advertising campaign log stream serves as one of the main sources of real-time signals, recording the number of exposures, click events, spending, and campaign parameter logs during the advertising campaign. The bidding environment stream serves as another main source of real-time signals, recording the bidding records, bid changes, and traffic allocation results for each exposure impression. Among these, exposures reflect the display of the advertisement on different channels and placements, clicks depict the user's immediate response to the advertisement content, and add-to-cart, orders, and payments describe the gradual evolution of conversion behavior over time.

[0026] After the collected exposures, clicks, add-to-cart, order placements, and payments enter the edge node, duplicate identification is performed based on the behavioral record information carried by each exposure, click, add-to-cart, order placement, and payment. Duplicate exposure records and duplicate click records corresponding to the same campaign slice within the same time window (e.g., a continuous interval of several minutes) are merged or removed to avoid the same campaign behavior being counted multiple times. After completing duplicate identification, anomaly identification is performed by combining exposure and click behaviors with behavior trigger frequency, behavior occurrence rhythm, and conversion path integrity. High-frequency clicks, abnormal jumps, and non-continuous conversion behaviors that deviate significantly from the normal campaign rhythm are filtered out, thereby reducing the interference of abnormal behaviors on subsequent feature construction. After anomaly filtering, exposures, clicks, add-to-cart, order placements, and payments from different collection sources are timestamped according to a unified time base, so that behaviors within the same campaign slice can be arranged continuously on the same time axis. When there are missing behavior nodes in the user sequence behavior, the missing nodes are filled in according to the interval relationship of adjacent behaviors on the time axis to ensure that exposures, clicks, add-to-cart, order placements, and payments form continuous and traceable user sequence behavior features.

[0027] It should also be noted that high-frequency clicks, abnormal redirects, and discontinuous conversion behaviors that clearly deviate from the normal campaign rhythm refer to behaviors that should normally occur in the order of exposure, clicks, add-to-cart, order placement, and payment within the same campaign segment. When a large number of click behaviors occur consecutively in a very short period of time (less than 1 second) but there is a lack of corresponding exposure behaviors, or when a continuous conversion path of add-to-cart, order placement, and payment is not formed after the click behaviors, or even when payment behaviors occur before click behaviors, the above behaviors are considered high-frequency clicks, abnormal redirects, or discontinuous conversion behaviors that clearly deviate from the normal campaign rhythm.

[0028] After forming user sequence behavior characteristics, for the ad creatives themselves, visual structure information is extracted from the image content and semantic information is extracted from the text content. The visual and semantic information are then combined to form multimodal representation features of the creatives, used to characterize the differences between different creatives at the visual and content levels. Simultaneously, information such as the type of ad placement, ad placement location, and display method recorded during ad delivery is used to form channel and placement context features, used to describe the display conditions of the creatives in different ad delivery environments. Furthermore, within the same time window (within a consecutive five minutes), bidding changes and traffic changes under the same channel and placement are aggregated. From the continuously collected bidding environment stream, the bidding records of each exposure impression under the same channel and placement are arranged chronologically to form a bidding amount sequence. From the continuously collected ad delivery day... In ZhiLiu, the number of exposures for each exposure impression under the same channel and placement is counted in chronological order to form an exposure volume sequence. After aligning the bidding amount sequence and the exposure volume sequence within the same time window, the bidding intensity change value is obtained by performing the first difference on the bidding amount sequence within a unit time (1 minute), and the exposure density change value is obtained by performing the first difference on the exposure volume sequence. The bidding intensity change value sequence and the exposure density change value sequence obtained by the first difference are normalized on the time axis and used as the competition intensity feature and traffic fluctuation feature. The user sequence behavior feature, multimodal representation feature of creative material, channel and placement context feature, competition intensity feature and traffic fluctuation feature are organized under the same feature dimension system and uniformly written into the online feature library, providing a complete and consistent feature source for subsequent organization of unified feature tensors by channel, audience, creative material and time period slice.

[0029] It should also be noted that the online feature library is pre-established based on the advertising entity's business scope before the advertising campaign begins. The establishment process includes determining the feature field structure for storing user sequence behavior features, multimodal representation features of creative materials, channel and placement context features, competition intensity features, and traffic fluctuation features, and configuring unified time stamps and slice identification rules for each feature field, so that features from different sources can be matched according to channel, audience, creative material, and time period. After the feature field structure and identification rules are configured, the online feature library is initialized to an empty state, which is used to receive subsequently collected and preprocessed user sequence behavior features, multimodal representation features of creative materials, channel and placement context features, competition intensity features, and traffic fluctuation features, thereby ensuring that when written to the online feature library, each type of feature has a clear storage location and organization method, and can be continuously updated and read under the same feature dimension system.

[0030] S1.2: After completing the online feature library writing, based on the set channel, audience, creative, and time slot slicing rules, read the corresponding user sequence behavior features, creative multimodal representation features, channel and placement context features, competition intensity features, and traffic fluctuation features from the online feature library in slice index order. Among these, user sequence behavior features reflect the evolution of user behavior in exposure, clicks, add-to-cart, order placement, and payment within the campaign slice; creative multimodal representation features characterize the visual and textual content presentation of creatives within the campaign slice; channel and placement context features describe the campaign environment in which the campaign slice is located; and competition intensity and traffic fluctuation features reflect market competition and traffic changes within the corresponding time period of the campaign slice. After reading, the features are arranged in a unified order (user sequence behavior features, creative multimodal representation features, channel and placement context features, channel and placement context features, competition intensity features, and traffic fluctuation features). The following features (including competitive intensity features and traffic fluctuation features) are spliced ​​and aligned to ensure that various features under the same channel, audience, creative, and time period are consistent in the time and slice dimensions. This forms a unified feature tensor that can fully reflect the comprehensive characteristics of a single campaign slice in terms of user behavior, creative attributes, campaign environment, and market status. At the same time, within the same time window (a continuous ten minutes), the campaign spending, impressions, clicks, and conversions corresponding to the unified feature tensor are aggregated to characterize the current campaign operation status. Among them, campaign spending, impressions, clicks, and conversions together constitute the campaign status summary, ensuring that the unified feature tensor and the campaign status summary are consistent in the time and slice dimensions. This provides a continuous, complete, and traceable input foundation for the multi-timescale deep predictor to simultaneously encode static features, known future features, and historical observable features in subsequent steps.

[0031] S2: Construct a multi-timescale deep predictor. After simultaneously encoding the static features, known future features, and historical observable features in the unified feature tensor and the delivery state summary, key driving factors are selected through attention and gating structures. Parallel prediction intermediates are obtained through parallel prediction. The parallel prediction intermediates are then merged into risk-labeled ROI prediction results to generate an actionable prediction list. S2.1: The construction of the multi-timescale deep predictor uses a unified feature tensor and a delivery state summary as input interfaces, and risk-labeled ROI prediction results based on short-term, medium-term, long-term, and cost trends as output interfaces. Structurally, it incorporates simultaneous encoding paths, an attention mechanism, a gating structure, and multiple prediction branches. The construction process first determines the input layer to receive the unified feature tensor and the delivery state summary. Following the input layer, feature extraction branches are connected to process static features, known future features, and historical observable features. The static feature branch uses an embedding encoding layer to convert discrete features into a dense representation. The known future feature branch uses a constant padding encoding layer to repeatedly pad the feature vector within the prediction time range. The historical observable feature branch uses a long short-term memory (LSTM) encoding layer. The network sequence coding layer retains the chronological order to form the sequence coding result. Simultaneously, a concatenation layer is set after the coding layer to concatenate the embedded coding result of static features, the constant filling coding result of known future features, and the sequence coding result of historical observable features in a fixed order to form a joint coding representation. The joint coding representation is connected to the attention mechanism layer, which uses multi-head attention to calculate the correlation weight of feature components. The weighted representation is connected to the gating structure layer, which uses gating linear units to adjust the information flow at short-term, medium-term, and long-term time scales to form a multi-scale feature representation. The multi-scale feature representation is connected to multiple prediction branches. Each prediction branch sets a fully connected layer and an output mapping layer to generate the corresponding prediction target value. The outputs of multiple prediction branches are fused at the output layer to form the risk-labeled ROI prediction result.

[0032] The training process of the multi-timescale deep predictor is as follows: training samples are generated using historical advertising delivery records. The generation method involves performing the same S1 steps as real-time delivery on historical advertising delivery records: data collection, preprocessing, writing to the online feature library, and generating a unified feature tensor and delivery status summary. This yields the unified feature tensor and delivery status summary corresponding to the historical period as the input of the training samples. Simultaneously, the actual short-term conversion tendency, actual expected value per unit conversion, and actual cost-per-click intensity per unit exposure are extracted from the historical advertising delivery records as the output of the training samples. The output of the training samples corresponds to the supervision label of the intermediate quantity of the parallel prediction. During the training phase, the input of the training samples is fed into the multi-timescale deep predictor to obtain the prediction results of the intermediate quantity of the parallel prediction and the risk-labeled ROI. After measuring the error between the intermediate quantity of the parallel prediction and the output of the training samples, the parameters of the simultaneous encoding path, attention mechanism, gating structure, and multiple prediction branches are updated until the preset convergence condition is reached, thus completing the training.

[0033] It should also be noted that the preset convergence condition is determined by the change in the deviation between the prediction result and the corresponding training sample output during the training process. When the prediction deviations corresponding to the short-term conversion tendency, the expected value per unit conversion, and the cost per unit exposure per click remain within a stable range (0.001, since 0.001 is sufficient to indicate high stability) in multiple consecutive training cycles, and the change amplitude of the risk label ROI prediction result no longer fluctuates in the short, medium and long time scales, the multi-time scale deep predictor is determined to have reached the preset convergence condition.

[0034] S2.2: After obtaining the trained multi-timescale deep predictor, the unified feature tensor and delivery state summary formed in step S1 are fed into the multi-timescale deep predictor as prediction inputs. The multi-timescale deep predictor extracts static features, known future features, and historical observable features from the unified feature tensor and delivery state summary according to the pre-determined feature field mapping relationship. Static features correspond to the slice attribute information that remains unchanged during the prediction period. Known future features correspond to the slice attribute information that has been determined at the prediction time and can be known in advance during the prediction period. Historical observable features correspond to the user behavior, delivery effect, and environmental state information that change continuously over time before the prediction time. The static features, known future features, and historical observable features are respectively fed into the corresponding feature processing paths inside the multi-timescale deep predictor to provide structured inputs for subsequent encoding operations.

[0035] It should also be noted that the pre-determined feature field mapping relationship is based on the correspondence rules set according to the business attributes and time attributes of each feature field in the unified feature tensor and the delivery status summary. This is used to clarify which feature fields in the unified feature tensor and the delivery status summary are classified as static features, which feature fields are classified as known future features, and which feature fields are classified as historical observable features. Among them, the slice attributes that remain unchanged during the prediction period correspond to static feature fields, the delivery-related information that is determined at the prediction time and will not change during the prediction period corresponds to known future feature fields, and the user behavior, delivery effect, and environmental status that change continuously over time and are recorded in chronological order before the prediction time point correspond to historical observable feature fields. Through the feature field mapping relationship, each feature field in the unified feature tensor and the delivery status summary can be accurately sent to the corresponding feature processing path in the multi-timescale deep predictor.

[0036] S2.3: After extracting static features, known future features, and historical observable features, the multi-timescale deep predictor performs corresponding encoding processing on the three types of features. Static features typically exist in discrete values ​​or high-dimensional sparse forms. To enable static features to participate in subsequent modeling processing within a unified feature space, the multi-timescale deep predictor performs embedding encoding on static features, converting them into a dense representation with fixed dimensions, so that they can be jointly organized with other features at the same scale. Since known future features are determined at the prediction time and remain unchanged within the prediction period, the multi-timescale deep predictor performs constant padding encoding on known future features, ensuring that known future features maintain a consistent expression within the prediction time range, thereby guaranteeing the continuous availability of known future information at different prediction time steps. Historical observable features record user behavior, delivery effects, and environmental states in time series form. The multi-timescale deep predictor performs sequence encoding on historical observable features, preserving the chronological order of historical observable features in the time dimension, so that the evolution process of behavior and state can be fully expressed.

[0037] After performing embedding encoding on static features, constant padding encoding on known future features, and sequence encoding on historical observable features, the multi-timescale deep predictor jointly organizes the three types of encoding results according to a pre-determined feature field mapping relationship. The joint organization process uses the time dimension and slice dimension as the alignment basis, so that the embedding encoding results corresponding to static features, the constant padding encoding results corresponding to known future features, and the sequence encoding results corresponding to historical observable features form a complete feature representation at the same time step. Among them, the embedding encoding results corresponding to static features remain unchanged throughout the entire prediction time range, the constant padding encoding results corresponding to known future features are repeatedly padded at time steps within the prediction time range, and the sequence encoding results corresponding to historical observable features are gradually expanded in chronological order, thus forming a feature combination that simultaneously contains static information, future information, and historical information at each prediction time step.

[0038] After time-dimensional alignment, the multi-timescale deep predictor concatenates the embedding encoding results of static features, the constant padding encoding results of known future features, and the sequence encoding results of historical observable features in a unified feature arrangement order. This allows features from different sources and with different time attributes to form a structurally consistent joint representation in the same feature space. The feature arrangement order remains fixed during the prediction process to ensure that the jointly organized features are comparable across different prediction times and different delivery slices. Through time alignment and feature concatenation, the jointly organized encoding (i.e., joint encoding) can simultaneously reflect the stable attributes of the delivery slice, known future constraints, and historical evolution state in a single representation structure. This provides a unified, complete, and continuous input foundation for subsequent selection of key driving factors through attention mechanisms and gating structures.

[0039] After the joint encoding representation is formed, the multi-timescale deep predictor generates corresponding relevance weights for each feature component in the joint encoding representation through an attention mechanism. The relevance weights are generated based on the degree of correlation between the joint encoding representation and three prediction targets: short-term conversion tendency, expected value per unit conversion, and cost per exposure per click. The attention mechanism matches the feature components in the joint encoding representation with the prediction targets one by one and outputs a set of one-to-one relevance weights. Specifically, the attention mechanism adopts a multi-head attention approach, where each attention head calculates the point integral of the query vector and the key vector. The query vector is formed by combining the embedding representations of the three prediction targets, and both the key vector and the value vector originate from the feature components in the joint encoding representation. The point integral is scaled and normalized using the Softmax function to obtain the attention weights. After being concatenated dimensionally, a linear projection is used to form the relevance weights that correspond one-to-one with each feature component. The multi-timescale deep predictor forms a weight sequence according to the normalized relevance weights from largest to smallest. The weight sequence is used to determine the key. The selection scope of driving factors and the rules for selecting key driving factors are expressed using fixed rules. These fixed rules include selecting the first few feature components according to the weight sequence or selecting feature components whose normalized relevance weights meet preset conditions. The preset conditions are given in configuration form, for example, the preset condition is that the normalized relevance weight is greater than the example value of 0.1. After the feature component selection is completed, the selected feature components in the joint encoding representation are used as inputs to the gating structure of key driving factors. The gating structure adopts the form of a gated linear unit. The sigmoid activation calculation of the forget gate and update gate is performed on the key driving factors. The forget gate determines the proportion of short-term information retention, and the update gate determines the proportion of medium-term and long-term information integration, thereby adjusting the information flow on short-term, medium-term and long-term time scales to form the selected key driving factor representation. Feature components that are not selected in the joint encoding representation do not enter the gating structure, so that the key driving factor selection process has a clear weight generation, weight comparison, feature selection and subsequent transmission path, avoiding the key driving factor selection process remaining at the level of abstract description.

[0040] S2.4: After completing the selection of key driving factors, the multi-timescale deep predictor simultaneously inputs the key driving factors into multiple independently set prediction branches. Each prediction branch performs layer-by-layer transformation processing on the key driving factors according to a fixed processing order. The layer-by-layer transformation processing includes three consecutive steps: linear mapping, nonlinear transformation, and output mapping, which are used to convert the key driving factors into numerical expressions of the corresponding prediction targets. Among them, the prediction branch used to output short-term transformation tendency receives the key driving factors, performs linear mapping on the key driving factors to compress or reorganize the feature dimensions, obtains the linear mapping result, performs ReLU nonlinear transformation on the linear mapping result to introduce nonlinear activation, and then generates short-term transformation tendency through output mapping. Short-term transformation tendency represents the probability that the delivery slice will undergo transformation behavior in a short time range.

[0041] To further explain, the fixed processing order is as follows: first, the linear mapping layer is executed to compress the dimensions of the key driving factors or reorganize the features to obtain the linear mapping result; then, the ReLU nonlinear transformation layer is executed to introduce nonlinear activation into the linear mapping result to capture the complex relationship between features; finally, the output mapping layer is executed to perform numerical transformation on the nonlinear transformation result to generate the output value of the corresponding prediction target.

[0042] The prediction branch for outputting the expected value per unit conversion, under the same key driver input conditions, follows the same processing order as the short-term conversion tendency prediction branch, sequentially performing linear mapping, ReLU nonlinear transformation, and output mapping steps on the key driver. However, the parameter configuration of the output mapping step differs from that of the short-term conversion tendency prediction branch. The difference lies in that the weight matrix and bias vector of the output mapping layer are specifically optimized for the expected value per unit conversion, so that the output result is the expected value per unit conversion, which represents the expected value corresponding to the conversion behavior of the delivery slice. The prediction branch for outputting the cost-per-click intensity per unit exposure also follows the same processing order as the short-term conversion tendency prediction branch, namely, sequentially performing linear mapping, ReLU nonlinear transformation, and output mapping steps, and receiving key driver inputs. However, the parameter configuration of the output mapping step is specifically optimized for the cost-per-click intensity per unit exposure, so that the output result is the cost-per-click intensity per unit exposure.

[0043] To further explain, multiple prediction branches execute linear mapping, nonlinear transformation, and output mapping steps in parallel at the same prediction time, and the prediction branches do not share output parameters. This allows short-term conversion tendency, expected value per unit conversion, and cost intensity per unit exposure click to be generated simultaneously in the same prediction process and output as intermediate quantities for parallel prediction.

[0044] S2.5: At the output layer, the intermediate values ​​of parallel predictions are uniformly received and combined according to a predetermined fusion order. The fusion order includes: associating the short-term conversion tendency with the expected value per unit conversion through multiplication to form the expected revenue; associating it with the cost intensity per unit exposure click through subtraction to form the short-term revenue trend; associating the expected value per unit conversion with the medium-term and long-term time span information reflected in historical observable features through weighted average to form the medium-term revenue trend and the long-term revenue trend; and associating the cost intensity per unit exposure click with the consumption information in the delivery status summary to form the cost trend reflecting cost changes. Thus, at the output layer, an ROI prediction result that simultaneously includes the short-term revenue trend, medium-term revenue trend, long-term revenue trend, and cost trend is obtained.

[0045] After obtaining the risk-labeled ROI prediction results for short-term, medium-term, long-term, and cost trends, the output layer evaluates the stability of the risk-labeled ROI prediction results over consecutive prediction periods to form confidence weights that characterize the reliability of the predictions. The stability is assessed by quantifying the degree of prediction volatility by calculating the magnitude of change in the risk-labeled ROI prediction results between adjacent prediction periods. Specifically, the expression for calculating the magnitude of change in the ROI prediction results between adjacent prediction periods is as follows: ; in, Indicates the first The magnitude of change in ROI forecast results between each forecast period and the previous forecast period Indicates the first Risk-labeled ROI prediction results for each prediction period Indicates the adjacent number The first forecast period before the Risk-labeled ROI prediction results for each prediction period.

[0046] And the magnitude of all adjacent changes within the continuous prediction period. By summarizing, the average change over the continuous forecast period is obtained, expressed as: ; in, This represents the average magnitude of change within a continuous forecast period. This indicates the magnitude of change in ROI prediction results between adjacent prediction periods. This indicates the total number of consecutive forecast periods participating in the evaluation. Indicates in The number of changes in the magnitude of adjacent forecast periods within a series of consecutive forecast periods.

[0047] After obtaining the average change magnitude, the output layer forms confidence weights based on the average change magnitude, expressed as: ; in, This represents the confidence weight, used to characterize the reliability of the ROI prediction results for risk labeling. The value ranges between 0 and 1. The smaller the average change, the closer the confidence weight is to 1; the larger the average change, the closer the confidence weight is to 0.

[0048] Based on the obtained confidence weights, the output layer labels the risk level of the risk-labeled ROI prediction results according to the risk range set by the configuration method, and organizes the risk-labeled ROI prediction results, corresponding confidence weights and risk levels in a unified manner to form an actionable prediction list, which is used as the direct input for subsequent distribution difference correction, delay compensation and hierarchical collaborative optimization steps.

[0049] To further explain, the risk level labeling process is as follows: Multiple non-overlapping confidence weight intervals are set in the configuration entry, and each confidence weight interval corresponds to a specific risk level. When the confidence weight corresponding to a risk-labeled ROI prediction result falls within a certain confidence weight interval, the current risk-labeled ROI prediction result is labeled with the risk level corresponding to the current confidence weight interval. Confidence weight intervals are divided based on the statistical distribution characteristics of confidence weights and business risk tolerance. The division method involves statistically analyzing the confidence weights within the historical prediction period to obtain distribution quantiles. High confidence intervals correspond to confidence weights greater than the 75th percentile of the historical distribution, medium confidence intervals correspond to confidence weights between the 25th and 75th percentiles of the historical distribution, and low confidence intervals correspond to confidence weights less than the 25th percentile of the historical distribution. For example, when the confidence weight falls into a high confidence interval, it is labeled as low risk; when it falls into a medium confidence interval, it is labeled as medium risk; and when it falls into a low confidence interval, it is labeled as high risk. This ensures that each risk-labeled ROI prediction result receives a unique and definite risk level label.

[0050] Preferably, compared to traditional prediction methods, this invention constructs a multi-timescale deep predictor to achieve classification and joint simultaneous encoding of static features, known future features, and historical observable features in a unified feature tensor and delivery status summary, thereby improving the comprehensiveness and temporal continuity of feature representation. The selection of attention mechanisms and gating structures as key driving factors enhances the ability to dynamically focus on factors influencing delivery. Parallel prediction intermediates are fused to generate risk-labeled ROI prediction results with confidence weights and risk levels, enhancing the actionability and reliability of the prediction. Overall, an end-to-end process from feature input to actionable prediction list is formed, improving the accuracy and interpretability of advertising delivery ROI prediction.

[0051] S3: After correcting the distribution differences in the actionable prediction list using the exposure click generation mechanism, the delay compensation layer is used to complete the actionable prediction list for the value signals that have not yet been returned, and the feasibility of the risk slices in the actionable prediction list is adjusted to obtain the corrected slice-level opportunity map. S3.1: When the actionable prediction list enters the distribution difference correction process, it records the exposure click generation mechanism information around the same campaign slice. The exposure click generation mechanism information includes at least the channel and ad placement corresponding to the campaign slice, the display ranking position or display level, the traffic allocation result under bidding and budget constraints, exposure trigger conditions, click trigger conditions, and the bidding intensity characteristics and traffic fluctuation characteristics related to exposure clicks. The exposure click generation mechanism information is established in a one-to-one correspondence with the short-term conversion tendency, expected value per unit conversion, cost per unit exposure click, confidence weight, and risk level in the actionable prediction list. When forming the training distribution and prediction distribution used to measure distribution differences, the training distribution is based on the historical training samples used in the multi-timescale deep predictor training process in step S2.1. The historical training samples come from the unified feature tensor and campaign status summary corresponding to the historical ad campaign records. The training distribution is based on "channel-audience-materials-time slot". At the granularity of "channel-audience-materials-time slot", the key fields of the unified feature tensor, the key fields of the delivery status summary, and the key fields of the exposure and click generation mechanism information are statistically summarized to form the statistical representation of the training stage; the prediction distribution is obtained from the current online prediction, which comes from the risk-labeled ROI prediction result sequence obtained by the output layer in step S2.5 within the continuous prediction period. The prediction distribution is granular with "channel-audience-materials-time slot" and forms the statistical representation of the prediction stage based on the statistical characteristics of the risk-labeled ROI prediction result sequence within the continuous prediction period. The statistical characteristics include the change magnitude of adjacent prediction periods and the average change magnitude obtained by summing the change magnitudes of adjacent prediction periods; when measuring the distribution difference, the statistical representation corresponding to the training distribution and the statistical representation corresponding to the prediction distribution are aligned according to the same delivery slot index, so that the statistical results of the training stage and the statistical results of the prediction stage can be compared under the same delivery slot dimension.

[0052] After alignment, the difference between the training and prediction distributions for each projection slice is measured. The difference can be measured using KL divergence, expressed as: ; in, The Kullback-Leibler divergence is a measure of the divergence between two probability distributions. and The asymmetric measure of the difference between them express Compared to KL divergence, Represents probability distribution In taking values The probability mass function value on, Represents probability distribution In taking values The probability mass function value on.

[0053] A larger KL divergence value indicates a greater difference between the training distribution and the predicted distribution.

[0054] The magnitude of the difference is mapped to a correction factor, which is used to correct the prediction offset related to the delivery slice in the actionable prediction list. The scope of the correction factor is limited by the allowed channels and allowed material sets in the configuration entry. Allowed channels filter delivery slices by channel dimension, and allowed material sets filter delivery slices by material dimension. The filtered delivery slices constitute the limited correction range. Delivery slices within the limited correction range are consistent with the short-term conversion tendency, expected value per unit conversion, cost per unit exposure click intensity, and risk-labeled ROI prediction results obtained by the output layer fusion according to the corresponding correction factor. Delivery slices outside the limited correction range keep the original value of the actionable prediction list unchanged, so that the distribution difference correction result can be consistent with the configuration entry constraints and can be directly passed to the subsequent delay compensation layer.

[0055] To further explain, the key fields of the unified feature tensor refer to the feature fields that are divided into static features, known future features, or historically observable features in the feature field mapping relationship, and are used as inputs to the multi-timescale deep predictor in step S2 to participate in the prediction calculation; the key fields of the campaign status summary refer to the campaign consumption, exposure, clicks, and conversions fields used to characterize the campaign slice's operating status and participate in the statistical and correction processing during the formation and distribution difference correction of the risk labeling ROI prediction results; the key fields of the exposure and click generation mechanism information refer to the fields used to describe the display ranking position, bidding intensity status, traffic allocation results, and exposure triggering conditions and click triggering conditions of the campaign slice under the corresponding channel and ad placement, and have a direct impact on the exposure and click generation process.

[0056] S3.2: The delay compensation layer uses the corrected actionable prediction list as input and establishes a delay status mark for the delivery slices that "have not yet returned value signals". The lack of returned value signals includes at least the following situations: payment has not yet returned, order has not yet returned, or value amount has not yet returned. The delay compensation layer reads the add-to-cart, order, and payment path status in the user sequence behavior characteristics at the same delivery slice granularity, reads the exposure, click, and conversion in the delivery status summary, and combines them with the short-term conversion tendency, expected value per unit conversion, and cost per unit exposure click intensity in the actionable prediction list to form the delay compensation input.

[0057] The delay compensation layer, based on the delay compensation input, filters historical ad delivery records to identify similar ad segments that are identical to the current segment in terms of channel, audience, creative, and time period. It then statistically analyzes the delay distribution of value signals from generation to feedback within these similar ad segments, calculating the proportion of value signals that have been fed back within a given time window (e.g., 1 minute). This proportion serves as the probability that a value signal that has not yet been fed back will be fed back within the time window. After obtaining the probability, the layer further summarizes the actual fed-back value corresponding to similar ad segments that have been fed back within the time window, using the statistical results as the expected value corresponding to the value signal that has not yet been fed back, thus determining the probability and expected value. After the probability and expected value are established, the delay compensation layer performs value completion on the actionable prediction list, writing the expected value corresponding to the value signal that has not yet been fed back into the value-side field of the actionable prediction list, and simultaneously updating the ROI prediction results of the value-related risk labels. This ensures that the actionable prediction list has completeness in the value dimension, allowing for subsequent optimization and preventing bias caused by relying solely on the fed-back portion in subsequent decisions.

[0058] S3.3: The delay compensation layer performs feasibility adjustments on risk slices in the actionable prediction list. Risk slices are determined by the risk level and confidence weight in the actionable prediction list. Slices with a risk level in the high-risk range and a confidence weight in the low-confidence range are identified as high-risk slices. The delay compensation layer reduces the feasibility weight of high-risk slices (for each slice, the initial feasibility weight is usually 1.0, indicating that before any risk adjustment, the resource allocation priority of each slice is equal). The reduction of feasibility weight is achieved using a preset weight scaling rule, so that high-risk slices obtain a lower resource allocation priority in subsequent hierarchical collaborative optimization. The delay compensation layer also identifies low-exposure, high-potential slices. Low-exposure, high-potential slices are determined by the exposure in the delivery status summary being in the low-exposure range and the short-term or medium-term revenue trend in the actionable prediction list being in the high-potential range.

[0059] To further explain, the low exposure range and high potential range are defined by the exposure threshold and short-term revenue trend threshold in the S1.1 configuration entry. The exposure threshold for the low exposure range is set to 50% of the historical average exposure. This value is chosen because slices with exposure below 50% of the historical average exposure usually suffer from insufficient exposure due to bidding or traffic limitations, requiring enhancement through feasibility weighting to uncover potential opportunities. The short-term revenue trend threshold for the high potential range is 1.2 times the target ROI range. This value is based on the fact that slices with a short-term revenue trend exceeding 1.2 times the target ROI range show higher-than-expected revenue potential, and even if the current exposure is low, it is worthwhile to prioritize resource allocation through feasibility weighting.

[0060] The preset weight scaling rules are configured in the settings entry, including the attenuation coefficient for high-risk slices and the boost coefficient for low-exposure, high-potential slices. The attenuation coefficient for high-risk slices is preset to 0.8, determined by multiplying the feasibility weight of the high-risk slice by 0.8. This moderately reduces the priority of slices with a confidence weight below 0.6 or a high-risk level in subsequent optimizations, preventing excessive resource allocation to slices with high uncertainty. The boost coefficient for low-exposure, high-potential slices is preset to 1.2, determined by multiplying the feasibility weight of the low-exposure, high-potential slice by 1.2. This prioritizes slices with exposure below the historical average but a risk-labeled ROI prediction result higher than the target ROI range after value completion, in order to uncover potential high-return opportunities. The attenuation and boost coefficients can be adjusted according to business needs in different deployment scenarios.

[0061] The delay compensation layer increases the feasibility weight of low-exposure, high-potential slices. The increase is also achieved using the preset weight scaling rules, so that low-exposure, high-potential slices can obtain higher exploration and allocation priority in subsequent hierarchical collaborative optimization compared to before the adjustment. After the feasibility weight adjustment is completed, the delay compensation layer organizes the corrected risk labeling ROI prediction results, value completion results and feasibility weights corresponding to each deployment slice in a unified manner to form a corrected slice-level opportunity map indexed by the deployment slice.

[0062] Preferably, compared with traditional methods, this invention achieves accurate distributional difference correction for the actionable prediction list by recording exposure and click generation mechanism information and constructing statistical representations of training and predicted distributions, effectively mitigating selection bias caused by bidding strategies; the delay compensation layer completes the data based on the historical statistical probability and expected value of similar slices, compensating for the underestimation of value caused by conversion delays; by combining confidence weights and risk levels to reduce the feasibility weight of high-risk slices and increase the feasibility weight of low-exposure, high-potential slices, the risk control and opportunity mining are further balanced, forming a corrected slice-level opportunity map, improving the reliability and action value of prediction results, and providing a more accurate and reliable decision-making basis for subsequent hierarchical collaborative optimization, thereby improving the overall stability and optimization efficiency of advertising ROI.

[0063] S4: Through layered collaborative optimization, the slice-level opportunity map is sequentially allocated with coarse-grained budget, adjusted with fine-grained bidding and frequency control, and selected and rotated in combination of creative materials. After generating a real-time delivery instruction package, it is sent to the advertising platform for execution and the results are received back from the advertising platform.

[0064] S4.1: The corrected slice-level opportunity map serves as the input for hierarchical collaborative optimization. Each slice in the corrected slice-level opportunity map corresponds to the corrected risk label ROI prediction result, feasibility weight, suggested range of available budget increment, available bid adjustment range, and available creative replacement priority. Hierarchical collaborative optimization performs coarse-grained budget allocation at the channel and ad placement unit levels. Slices in the corrected slice-level opportunity map with feasibility weights higher than the configured threshold and corrected risk label ROI prediction results that conform to the target ROI range are identified as high-potential slices. The remaining budget share within the budget limit set in the S1.1 configuration entry is preferentially moved to the channels and ad placement centers where the high-potential slices are located. The movement method is achieved through proportional allocation, and the proportion is determined by the product of the feasibility weight of the high-potential slice and the corrected risk label ROI prediction result. After completing the coarse-grained budget allocation, the updated budget share for each channel and ad placement unit is obtained.

[0065] To further explain, the purpose of setting the configuration threshold is to balance resource allocation efficiency and risk control, and to avoid over-distributing resources to slices with high uncertainty. Through historical deployment backtesting, the configuration threshold is set to 0.7 to 0.9. This is because a value below 0.7 will over-restrict high-potential slices and cause opportunities to be lost, while a value above 0.9 will include too many uncertain slices and increase risk.

[0066] S4.2: After completing the coarse-grained budget allocation, the layered collaborative optimization process moves into each slice for fine-grained adjustments to bidding and frequency control. Based on the cost constraint threshold in the configuration entry, an upper limit is imposed on the bidding range of each slice, ensuring that the adjusted bidding range does not exceed the cost per click or impression corresponding to the cost constraint threshold. Simultaneously, the adjustment magnitude and direction of the bidding range are determined based on the slice's feasibility weight and the ROI prediction result after value completion and risk labeling. For every 0.1 increase in the feasibility weight, the upper limit of the bidding range is increased by 5%; for every 0.1 decrease in the feasibility weight, the lower limit of the bidding range is decreased by 3%. The direction of the bidding range adjustment leans towards higher ROI prediction results after value completion and risk labeling. The bidding range adjustment range is controlled between 80% and 120% of the cost constraint threshold. Furthermore, based on the slice... After the feasibility weight and value are completed, the risk labeling ROI prediction results determine the lower and upper limits of the bidding range. The lower limit of the bidding range is the minimum value to maintain exposure competitiveness, and the upper limit of the bidding range is the maximum value to meet the cost constraint threshold. The frequency control parameters of each slice are adjusted. The frequency control parameters include the maximum number of exposures per day or per hour. The tightness of the frequency control parameters is set according to the feasibility weight of the slice. For slices with a feasibility weight higher than 0.8, the upper limit of the frequency control parameters is relaxed by 20% to increase exposure opportunities. For slices with a feasibility weight lower than 0.6, the upper limit of the frequency control parameters is tightened by 30% to control risk. Slices with a feasibility weight between 0.6 and 0.8 keep the standard frequency control parameters unchanged. After completing the fine-grained adjustment of bidding and frequency control, the updated bidding range and frequency control parameters of each slice are obtained.

[0067] To further explain, for every 0.1 increase in feasibility weight, the upper limit of the bid range is increased by 5%: the rationale is that an increased feasibility weight indicates higher slice reliability, and moderately increasing the upper limit can increase exposure competitiveness while avoiding excessive bidding that leads to cost overruns; for every 0.1 decrease in feasibility weight, the lower limit of the bid range is decreased by 3%: the rationale is that a decreased feasibility weight indicates higher risk, and lowering the lower limit (3% is more conservative) reduces the minimum bid to control risk while retaining basic exposure opportunities; the bid range adjustment is controlled between 80% and 120% of the cost constraint threshold: the rationale is that the 80% lower limit ensures basic competitiveness and avoids zero exposure, while the 120% upper limit prevents exceeding cost constraints and resulting in losses, which is also a common safety range in advertising optimization practice.

[0068] The reason for adjusting feasibility weights above 0.8 and below 0.6 is as follows: A weight above 0.8 indicates that the slice has highly stable prediction results across multiple time scales, a confidence weight significantly higher than average, and a low risk level, possessing strong certainty and high return potential, making it worthwhile to proactively relax frequency control to amplify exposure opportunities; setting the threshold higher (e.g., 0.9) would miss some reliable slices, resulting in lost opportunities; a weight below 0.6 indicates that the slice's confidence weight is too low, the risk level is too high, or the prediction stability is insufficient, and continuing high exposure could easily lead to budget waste and ROI fluctuations, therefore, frequency control needs to be proactively tightened for risk control; setting the threshold lower (e.g., 0.5) would... Excessive tolerance for uncertain slices increases overall risk exposure. The reason for extending the limit by 20% and tightening it by 30% is that high-feasibility slices (>0.8) have clear potential, but excessive leniency can easily lead to rapid budget depletion or creative fatigue. The 20% reduction is an empirical balance value in advertising optimization practice, which can significantly increase exposure opportunities to obtain more conversions while keeping it within a safe range to avoid overexposure risks. The 30% reduction is because low-feasibility slices (<0.6) have higher risks and require stronger protection measures. The larger 30% reduction prioritizes risk control, effectively reducing exposure and thus reducing budget waste, while retaining basic exposure to maintain data feedback.

[0069] S4.3: The updated bid range and frequency control parameters are used as the basis for material allocation. Material combination selection and rotation are performed within the allowed material set in the configuration entry. Multiple material combinations with the highest feasibility weight are selected from the allowed material set. Materials with high feasibility weight are prioritized for allocation to segments with high updated budget share and high bid range. Materials with low feasibility weight are allocated to segments with low budget share or high risk. At the same time, the material rotation rhythm is set. The rotation rhythm is dynamically determined based on the exposure and click feedback of the segment. After the exposure reaches a preset number (5000), the next set of materials is rotated. The material allocation results, bid range, frequency control parameters, budget share and rotation rhythm are linked and organized into a real-time delivery instruction package. The real-time delivery instruction package includes the budget share of each channel, the bid range and frequency control parameters of each delivery unit, and the material listing list and rotation rhythm of each segment.

[0070] S4.4: After the real-time delivery instruction package is generated, it is sent to the interface of the advertising platform. Under the event trigger condition that the risk label ROI prediction result falls below the target ROI range after the set time granularity is reached or corrected, the advertising platform executes the budget share allocation, bid range adjustment, frequency control parameter setting, and material listing and rotation in the real-time delivery instruction package. Within the same time window, the consumption, impression, click and conversion results returned by the advertising platform are collected. The returned results are re-aggregated according to the channel-audience-material-time segment dimension to update the delivery status summary and a new round of unified feature tensor, forming a closed loop feedback.

[0071] This embodiment also provides a computer device applicable to the real-time prediction and optimization method of advertising ROI based on deep learning, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the real-time prediction and optimization method of advertising ROI based on deep learning proposed in the above embodiment.

[0072] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0073] This embodiment also provides a storage medium storing a computer program. When executed by a processor, the program implements the deep learning-driven real-time prediction and optimization method for advertising ROI, as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0074] In summary, this invention: by constructing a multi-timescale deep predictor to generate risk-labeled ROI prediction results, it achieves joint interactive modeling of three types of features, improving the interpretability and stability of multi-timescale revenue and cost trend predictions, and effectively alleviating the distribution bias problem; at the same time, by correcting the distribution differences of the actionable prediction list through the exposure click generation mechanism and supplementing the value signals that have not yet been returned through the delay compensation layer, and by combining feasibility adjustment to generate a corrected slice-level opportunity map, it solves the value underestimation caused by sample selection bias and conversion delay under real-time bidding, further reducing prediction errors and improving the accuracy and reliability of ROI prediction.

[0075] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for real-time prediction and optimization of advertising ROI driven by deep learning, characterized by: include, Real-time signals are collected, preprocessed, and written into an online feature library to generate a unified feature tensor and a delivery status summary. A multi-timescale deep predictor is constructed. After simultaneously encoding static features, known future features, and historical observable features in the unified feature tensor and the delivery state summary, key driving factors are selected through attention and gating structures. Parallel prediction intermediates are obtained through parallel prediction, and the parallel prediction intermediates are merged into risk-labeled ROI prediction results to generate an actionable prediction list. After correcting the distribution differences in the actionable prediction list using the exposure click generation mechanism, the delayed compensation layer is used to complete the actionable prediction list for any missing value signals, and the feasibility of the risk slices in the actionable prediction list is adjusted to obtain a corrected slice-level opportunity map. By using a layered collaborative optimization approach, the slice-level opportunity map is sequentially allocated with coarse-grained budget, adjusted with fine-grained bidding and frequency control, and selected and rotated in combination of creative materials. After generating a real-time delivery instruction package, it is sent to the advertising platform for execution, and the results are received back from the advertising platform.

2. The method for real-time prediction and optimization of advertising ROI based on deep learning as described in claim 1, characterized in that: The process of acquiring real-time signals, preprocessing them, and then writing them into an online feature library involves the following steps: Based on the configuration entry set by the advertising entity, it collects impressions, clicks, add-to-cart, order placement, and payment as real-time signals; The configuration entry includes the target ROI range, budget limit, allowed channels and allowed material sets, and cost constraint threshold; After deduplication, anti-fraud filtering, timestamp alignment and missing completion of the collected real-time signals at the edge nodes, user sequence behavior features, material multimodal representation features, channel and placement context features, competition intensity features and traffic fluctuation features are obtained and uniformly written into the online feature library.

3. The method for real-time prediction and optimization of advertising ROI based on deep learning as described in claim 1, characterized in that: The generation of the unified feature tensor and the delivery state summary specifically involves: Organize user sequence behavior features, multimodal representation features of creative materials, channel and placement context features, competition intensity features, and traffic fluctuation features written into the online feature library into a unified feature tensor by channel-audience-materials-time period slices; It aggregates campaign spending, impressions, clicks, and conversions to generate a campaign status summary.

4. The method for real-time prediction and optimization of advertising ROI based on deep learning as described in claim 1, characterized in that: The simultaneous encoding of static features, known future features, and historical observable features in the unified feature tensor and delivery state summary, and the selection of key driving factors through attention and gating structures, specifically involves: Extract static features, known future features, and historical observable features from a unified feature tensor and a delivery state summary; After performing embedding encoding on static features, constant padding encoding on known future features, and sequence encoding on historical observable features, the encoding results of static features, known future features, and historical observable features are jointly organized to form a joint encoding representation. The correlation weights between feature components in the joint encoding representation are calculated through an attention mechanism, and key driving factors are selected by controlling the flow of information in the short, medium and long term through a gating structure.

5. The method for real-time prediction and optimization of advertising ROI based on deep learning as described in claim 1, characterized in that: The process involves obtaining intermediate values ​​for parallel predictions, merging these intermediate values ​​into a risk-labeled ROI prediction result, and generating an actionable prediction list. Specifically: Key drivers are used as shared inputs for multiple prediction branches, and short-term conversion tendency, expected value per unit conversion, and cost per unit exposure click intensity are output in parallel by multiple prediction branches as intermediate quantities for parallel prediction. The intermediate quantities of parallel predictions are fused at the output layer to generate risk-labeled ROI prediction results for short-term, medium-term, long-term, and cost trends. After calculating confidence weights and labeling risk levels, an actionable prediction list is obtained.

6. The method for real-time prediction and optimization of advertising ROI based on deep learning as described in claim 1, characterized in that: The method of using the exposure-click generation mechanism to correct the distribution differences in the action prediction list refers to recording the exposure-click generation mechanism information, calculating the distribution differences between the training distribution and the predicted distribution in the action prediction list, limiting the correction range according to the allowed channels and allowed material sets in the configuration entry, and correcting the distribution differences within the limited correction range.

7. The method for real-time prediction and optimization of advertising ROI based on deep learning as described in claim 1, characterized in that: The process of supplementing the actionable prediction list with missing value signals through a delay compensation layer and adjusting the feasibility of risk slices in the actionable prediction list involves establishing a delay compensation layer, calculating the probability of occurrence and expected value of missing value signals in the actionable prediction list, supplementing the actionable prediction list with value signals, reducing the feasibility weight of high-risk slices in the actionable prediction list, increasing the feasibility weight of low-exposure, high-potential slices, and generating a corrected slice-level opportunity map.

8. The method for real-time prediction and optimization of advertising ROI based on deep learning as described in claim 1, characterized in that: The method of sequentially allocating budget in coarse-grained manner to slice-level opportunity maps through hierarchical collaborative optimization refers to identifying high-potential slices in the corrected slice-level opportunity map, prioritizing the movement of budget shares to high-potential slices at the channel and placement levels, obtaining updated budget shares at each channel and placement level, and completing the coarse-grained allocation.

9. The method for real-time prediction and optimization of advertising ROI based on deep learning as described in claim 1, characterized in that: The fine-grained adjustment of bidding and frequency control, as well as the combination and rotation of creative materials, generate real-time delivery instruction packages, specifically as follows: Within each slice of the calibrated slice-level opportunity map, the bidding range and frequency control parameters are adjusted according to the cost constraint threshold to obtain the adjusted bidding range and frequency control parameters. The adjusted bid range and frequency control parameters are used as the basis for material allocation. Material combination selection and rotation are performed within the allowed material set. High-feasibility materials in the allowed material set are prioritized for allocation to high-budget slices that have undergone a coarse-grained budget process, resulting in material allocation results. The results of material allocation, the adjusted bid range and frequency control parameters, as well as the updated budget shares of each channel and the placement level are linked to generate a real-time placement instruction package.

10. The method for real-time prediction and optimization of advertising ROI based on deep learning as described in claim 1, characterized in that: The process of sending the command to the advertising platform for execution and receiving the feedback results from the advertising platform refers to sending a real-time delivery instruction package through the interface of the advertising platform. After setting the time granularity and event triggering conditions, the budget share, bid range, frequency control parameters and material allocation results in the real-time delivery instruction package are executed through the advertising platform. The consumption, exposure, click and conversion results fed back by the advertising platform are collected in the same window.

Images