Advertisement putting strategy generation system based on artificial intelligence

Abstract

The invention provides an advertisement putting strategy generation system based on artificial intelligence, and belongs to the technical field of artificial intelligence, and the system comprises a vector construction module which is used for collecting multi-source data of advertisement putting to be played, carrying out the feature extraction of the multi-source data according to dimension indexes, and constructing a feature vector of each dimension index; the strategy determination module is used for comprehensively analyzing the feature vectors under all dimension indexes according to the putting target of the to-be-put advertisement, and determining a plurality of candidate putting strategies in combination with the own playing strategy of the to-be-played advertisement; the effect evaluation module is used for performing pre-simulation delivery evaluation on each candidate delivery strategy, predicting the delivery effect of each candidate delivery strategy, determining strategy adjustment parameters according to the delivery effect, and screening an optimal delivery strategy; and the updating module is used for dynamically updating and re-putting the optimal putting strategy in combination with the newly collected multi-source data. And the putting pertinence, the effect stability and the environmental adaptability are obviously improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to an artificial intelligence-based advertising strategy generation system. Background Technology

[0002] With the rapid development of the digital advertising industry, the accuracy and dynamic adaptability of ad placement have become core requirements for improving campaign effectiveness. Current ad placement strategy generation faces significant technical bottlenecks in areas such as multi-source data processing and strategy generation, as detailed below:

[0003] Existing technologies for advertising delivery typically employ simple concatenation or single-dimensional segmentation of multi-source data (such as user behavior data, platform traffic data, and ad content data), lacking refined breakdown of data dimensions. For example, simply categorizing user data into user attribute dimensions without further subdividing sub-indicators such as interest tags and spending power leads to the loss of key details during feature extraction, creating potential problems for subsequent strategy generation. Furthermore, existing methods often rely on a single delivery objective (such as maximizing exposure) for feature vector analysis when generating candidate delivery strategies, and fail to adequately utilize the ad's own playback strategy, resulting in poor targeting of candidate strategies and significantly reduced delivery effectiveness.

[0004] Therefore, this invention proposes an advertising strategy generation system based on artificial intelligence. Summary of the Invention

[0005] This invention provides an artificial intelligence-based advertising strategy generation system to solve the aforementioned technical problems.

[0006] This invention provides an artificial intelligence-based advertising strategy generation system, comprising:

[0007] The vector construction module is used to collect multi-source data of the advertisement to be played, and to extract features from the multi-source data according to the dimensional indicators to construct the feature vector of each dimensional indicator. There are several dimensional indicators, and each dimensional indicator involves at least one sub-indicator.

[0008] The strategy determination module is used to comprehensively analyze the feature vectors under all dimensions of the advertising to be delivered based on the delivery target of the advertising to be delivered, and to determine multiple candidate delivery strategies in combination with the playback strategy of the advertising to be played. The candidate delivery strategies include several delivery setting parameters.

[0009] The performance evaluation module is used to perform pre-simulated delivery evaluation for each candidate delivery strategy, predict the delivery performance of each candidate delivery strategy, determine strategy adjustment parameters based on the delivery performance, and select the optimal delivery strategy.

[0010] The update module is used to dynamically update and re-deploy the optimal delivery strategy by combining newly collected multi-source data.

[0011] Preferably, the vector construction module includes:

[0012] A partitioning unit is used to partition the multi-source data according to a dimensional index, and to extract features from the partitioned data according to the sub-indicators under the dimensional index, so as to obtain the basic sample features and feature weights of the corresponding sub-indicators.

[0013] The fusion unit is used to fuse the weighted sub-features to obtain the single-dimensional initial feature set of the corresponding dimension index based on the feature weights, sample features and the correlation strength between sub-indicators.

[0014] The representation unit is used to input the initial single-dimensional feature set of all dimensions into the cross-dimensional global attention network and output the enhanced feature representation of each dimension index.

[0015] The building unit is used to map the enhanced feature representations of each dimension indicator to a unified high-dimensional feature space, and to construct the feature vector of each dimension indicator.

[0016] Preferred options also include:

[0017] The frame extraction module is used to extract the text content, canvas content and voice-over content of each frame in the advertisement to be delivered, and obtain the corresponding frame feature vector. At the same time, the core feature vector is obtained according to the core theme of the advertisement to be delivered.

[0018] The calculation module is used to calculate the first comprehensive distance between each frame feature vector and the core feature vector, as well as the individual element distance between each frame feature vector and the core feature vector, to obtain the analysis array of the corresponding frame.

[0019] The association acquisition module is used to input the feature vector of each frame into the feature analysis model to obtain the association between elements in the corresponding frame and obtain the association array;

[0020] The probability determination module is used to determine the display probability of the corresponding frame based on the analysis array and the association array;

[0021] The continuous segmentation module is used to construct a probability sequence of the advertisement to be delivered based on the display probability of each frame, and to divide the advertisement to be delivered into continuous segments based on the probability sequence.

[0022] The identifier setting module is used to assign a placeholder playback identifier to the corresponding individual continuous segment according to the output meaning and average probability of each individual continuous segment. The placeholder playback identifier is unique and is related to the playback position and playback timing logic.

[0023] The condition addition module is used to add trigger conditions to each individual continuous segment of sub-video. When the corresponding trigger condition is met, the sub-video of the individual continuous segment that matches the trigger condition is displayed in original color in a designated small window at the set playback position of the overall canvas, and the sub-video of the individual continuous segment that does not meet the trigger condition is displayed transparently at the set playback position, until the corresponding individual continuous segment finishes playing according to the set playback time, thereby obtaining its own playback strategy.

[0024] Preferably, the identifier setting module includes:

[0025] The multidimensional extraction unit is used to extract the multidimensional set Wi = {Gi, Pi, Ti, Ki, Xi} for each individual continuous segment, where Gi, Pi, Ti, Ki, and Xi are, respectively, the semantic encoding of the output meaning of the i-th individual continuous segment, the normalized value of the average probability, the dependency relationship with the predecessor segment, the triggering relationship with the successor segment, and the spatial requirements based on the required canvas area ratio and window shape.

[0026] The matching retrieval unit is used to retrieve the tuple {Fui,Pui,Tui} with the highest matching degree from the identifier pool when the corresponding individual continuous segment meets the preset conditions, and automatically associate the end event of the predecessor segment with the start condition of the successor segment. Here, Fui, Pui, and Tui represent the function identifier, playback position rule, and playback timing logic of the output meaning mapping of the i-th individual continuous segment, respectively.

[0027] The tuple update unit is used to collect interactive feedback test data on the advertisement to be played in real time, and update the priority weight of the playback position of the tuple and the periodic parameters of the timing logic.

[0028] Preferably, the strategy determination module includes:

[0029] The semantic enhancement unit is used to perform multi-scale attention fusion on feature vectors under all dimensions, calculate the importance weights of feature vectors in each dimension, and generate a weighted fused feature vector.

[0030] The matrix construction unit is used to semantically enhance the weighted fusion feature vector based on the advertising domain knowledge graph, and construct an enhanced feature matrix containing explicit features and implicit semantic associations;

[0031] The input unit is used to extract the interaction features and temporal features of the playback strategy itself, and input them into the playback strategy temporal model to generate a playback strategy feature tensor.

[0032] Define a unit, used to define the set of deployment constraint domains D = {D j1,j1＝1,2,3,...,Nk}, where D j1Let Nk represent the j1-th constraint domain, and let Nk represent the number of constraint domains. For each constraint domain, a bidirectional group {Chj1, Cgj1} is constructed, where Chj1 is the constraint verification identifier of the j1-th constraint domain, and Cgj1 is the policy generation trigger identifier of the j1-th constraint domain.

[0033] The decomposition unit is used to decompose the target of the advertisement to be played into a set of subtasks ∈={∈j2,j2=1,2,3,..,Nq}, and construct a directed acyclic dependency graph according to the domain dependency relationship, where ∈j2 represents the j2-th subtask; Nq represents the number of subtasks; and the subtasks in the directed acyclic dependency graph ∈ j3 Dependence on D j1 Chj1 must be completed before startup can begin, and ∈ j3 and ∈ j3+1 Dependency edges;

[0034] Binding unit, used to bind adjacent subtasks in the directed acyclic dependency graph ∈ j3 and ∈ j3+1 Binding to the same constraint domain D j1 Establish a driver execution chain for the bidirectional group {Chj1,Cgj1}:

[0035] ∈ j3+1 Continuous monitoring of D using an adaptive monitoring window j1 The Chj1 state, when ∈ j3 Once completed, send a message to D. j1 Write to the active state Chj1;

[0036] When ∈ j3+1 After Chj1 activation is detected, from D j1 Extract Cgj1 from it, trigger its own execution, and output the domain policy parameters;

[0037] The clustering fusion unit is used to perform clustering analysis on the enhanced feature matrix and the policy feature tensor, and fuse them with the domain policy parameters to generate multiple candidate deployment strategies.

[0038] Preferably, the effect evaluation module includes:

[0039] The model building unit is used to build a response surface model based on historical deployment data and fit the response function f(M) through Gaussian process regression.

[0040] The factor determination unit is used to map the candidate delivery strategy to the response function f(M) to generate a perturbation sequence. At the same time, it extracts the influencing factors related to user behavior from the candidate delivery strategy, and combines the user group behavior model to characterize the uncertainty of user behavior to determine the probability distribution of the potential interaction path of individual users to the advertisement.

[0041] The comparison unit is used to perform a first comparison between the perturbation sequence and the first evaluation table, and at the same time, to perform a second comparison between the probability distribution of the potential interaction path and the second evaluation table.

[0042] The prediction unit is used to predict the delivery effect of the corresponding candidate delivery strategy based on the first control result and the second control result.

[0043] Preferably, the effect evaluation module further includes:

[0044] The comparison unit is used to mine the causal chain of the deployment parameters from the deployment effect, and analyze the causal chain based on the structural causal model to obtain the simulated direct causal and simulated indirect causal of the corresponding deployment parameters on the effect index, and compare them with the ideal direct causal and ideal indirect causal of the set ideal effect.

[0045] The perturbation definition unit is used to analyze the distribution drift of the delivery data for the corresponding candidate delivery strategy and define the data distribution perturbation set;

[0046] The re-analysis unit is used to determine the strategy adjustment parameters based on the comparison results and the data distribution disturbance set, and to re-analyze the pre-simulation deployment effect;

[0047] The filtering unit is used to select the strategy with the best performance from the re-analyzed pre-simulation delivery results as the optimal delivery strategy.

[0048] Preferably, the update module includes:

[0049] The mapping unit is used to map newly acquired multi-source data to the corresponding spatiotemporal scene unit. At the same time, the credibility weight of each newly acquired data source is calculated based on the deep evidence network.

[0050] The trajectory evolution unit is used to characterize the core parameters of the optimal deployment strategy as quantum superposition vectors and determine the quantum evolution trajectory of each quantum parameter.

[0051] The comparison and analysis unit is used to compare and analyze the quantum evolution trajectory with the mapping results and the confidence weight, and update and re-deploy the corresponding core parameters.

[0052] Compared with the prior art, the beneficial effects of this application are as follows:

[0053] By refining multi-source data, deeply integrating the target audience with the characteristics of the ads themselves to generate candidate strategies, and combining scientific evaluation and dynamic optimization mechanisms to screen and iterate the optimal strategy, the entire process of advertising delivery from data collection to performance optimization has been made intelligent, significantly improving the targeting, performance stability and environmental adaptability of the campaign.

[0054] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0055] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0056] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0057] Figure 1 This is a structural diagram of an artificial intelligence-based advertising strategy generation system according to an embodiment of the present invention. Detailed Implementation

[0058] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0059] This invention provides an artificial intelligence-based advertising strategy generation system, such as... Figure 1 As shown, it includes:

[0060] The vector construction module is used to collect multi-source data of the advertisement to be played, and to extract features from the multi-source data according to the dimensional indicators to construct the feature vector of each dimensional indicator. There are several dimensional indicators, and each dimensional indicator involves at least one sub-indicator.

[0061] The strategy determination module is used to comprehensively analyze the feature vectors under all dimensions of the advertising to be delivered based on the delivery target of the advertising to be delivered, and to determine multiple candidate delivery strategies in combination with the playback strategy of the advertising to be played. The candidate delivery strategies include several delivery setting parameters.

[0062] The performance evaluation module is used to perform pre-simulated delivery evaluation for each candidate delivery strategy, predict the delivery performance of each candidate delivery strategy, determine strategy adjustment parameters based on the delivery performance, and select the optimal delivery strategy.

[0063] The update module is used to dynamically update and re-deploy the optimal delivery strategy by combining newly collected multi-source data.

[0064] In this embodiment, multi-source data refers to a collection of data related to advertising placement collected from different channels and different types of data sources. For example, user behavior data (such as click records, browsing time, and search keywords), platform data (such as peak traffic periods and user geographic distribution), advertising content data (such as advertising video frame information and copy keywords), and third-party data (such as industry reports and competitor placement data). Specifically, data is obtained by connecting to advertising platforms (such as Douyin and Taobao), user behavior tracking tools (such as event tracking systems), and data service providers (such as Aurora Data) through API interfaces and storing the data in a data warehouse (such as Hadoop HDFS).

[0065] Dimensional metrics are the core dimensions for classifying multi-source data, used to characterize advertising-related features from different perspectives, including: user dimension (characterizing target user characteristics), platform dimension (characterizing advertising platform attributes), advertising dimension (characterizing advertising characteristics), and scenario dimension (characterizing advertising scenario features).

[0066] Sub-indicators are specific sub-indicators under dimensional indicators, used to quantify dimensional characteristics. User dimension sub-indicators include: age, gender, interest tags (such as digital products, beauty products), and consumption level (such as high net worth, ordinary).

[0067] The platform-level sub-metrics are daily active users (DAU), peak traffic periods (e.g., 19:00-21:00), and user retention rate. The advertising-level sub-metrics are video duration, keyword density in the copy, and color saturation in the image.

[0068] A feature vector is a numerical array representing the extracted features for subsequent algorithm processing. For example, a user dimension feature vector can be represented as [25 (age), 0.8 (interest matching degree), 0.3 (consumption potential)], where each element corresponds to the quantified value of a sub-indicator.

[0069] The campaign objective is the business goal that advertisers hope to achieve through the campaign, which guides the direction of strategy generation. These objectives include: exposure (e.g., reaching 1 million users), conversion rate (e.g., click-to-purchase conversion rate ≥ 5%), brand awareness (e.g., a 30% increase in search volume), and ROI (e.g., an investment of 1 yuan yields an output of 3 yuan).

[0070] Feature vector comprehensive analysis involves fusing and interpreting feature vectors from various dimensions to uncover the correlations between data in order to support strategy generation. For example, by combining the feature vectors of the user dimension "women aged 25-30" with the feature vectors of the advertising dimension "beauty products", the matching degree between the two can be analyzed to determine the targeting strategy.

[0071] The self-play strategy is the display rule of the advertising content itself, including screen segmentation, display logic, trigger conditions, etc. For example, the self-play strategy of a car advertisement is "to show the exterior for the first 5 seconds (transparent display), and when the user clicks the screen, switch to the interior details (true color display), and automatically play the price information after 10 seconds". Specifically, the advertisement is cut into continuous segments by video editing tools (such as FFmpeg) and trigger conditions are set by JavaScript scripts (such as listening for user click events).

[0072] The candidate targeting strategies are multiple potential targeting plans generated based on data analysis, which are then selected for further screening. For example, for the goal of "exposure of beauty products", three candidate strategies are generated: Strategy A (targeting women aged 25-35, targeting from 19:00 to 22:00), Strategy B (targeting users in first-tier cities, targeting from 12:00 to 14:00), and Strategy C (targeting high-spending users, targeting all-day targeting).

[0073] The campaign settings parameters are the specific configuration items that constitute the candidate strategy and determine the execution details of the strategy. For example, the campaign platform (Douyin / Xiaohongshu), the bidding method (CPM / CPC), the budget allocation (5,000 yuan per day), the targeting tags (beauty interests, 25-35 years old), and the creative display order (showing products first and then discounts).

[0074] Pre-simulation campaign evaluation involves using a simulation model to simulate the strategy execution process and predict the results before the actual campaign. For example, a simulation model trained with historical data can simulate the campaign process of strategy A targeting women aged 25-35 from 19:00 to 22:00, and output the predicted exposure and click-through rate. Specifically, AnyLogic is used to build a campaign environment simulation model, candidate strategy parameters are input, and 100 Monte Carlo simulations are run to reduce prediction errors.

[0075] In this embodiment, the campaign performance refers to the performance indicators of the candidate strategy in the pre-simulation, reflecting the quality of the strategy. Examples include click-through rate (CTR = clicks / impressions), conversion rate (CVR = conversions / clicks), cost per thousand impressions (CPM = total cost / impressions × 1000), and ROI (revenue / cost).

[0076] In this embodiment, the strategy adjustment parameters are correction values ​​for candidate strategy parameters based on the campaign performance, used to optimize the strategy. For example, if the predicted click-through rate of strategy A is lower than the target, the adjustment parameters are "increase bid by 10%" and "add 'skincare habits' to the targeting tags".

[0077] The optimal delivery strategy is the candidate strategy that achieves the best overall performance after evaluation and adjustment. For example, after adjusting strategy A, the predicted CTR increased from 3% to 5%, and the ROI reached 4.2, which is better than other strategies and was selected as the optimal strategy.

[0078] The newly collected multi-source data is new data generated in real time during the strategy execution process, used to monitor the strategy's adaptability. For example, new data collected 1 hour after the campaign: real-time click-through rate (2.8%), the addition of "third-tier cities" in user regions, and competitors suddenly increasing their campaign efforts.

[0079] Dynamic updates are based on new data to adjust the parameters of the optimal delivery strategy in real time, ensuring that the strategy adapts to changes. For example, based on new data that "users in third-tier cities have high click-through rates", the target area of ​​the optimal strategy is expanded from "first-tier" to "first-tier + third-tier", and the bid is reduced by 5% to control costs.

[0080] Re-launching involves re-implementing the updated optimal strategy in the actual campaign process, replacing the old strategy. For example, in the API interface of an advertising platform (such as ByteDance's advertising platform), the original strategy parameters are overridden to execute the updated campaign plan.

[0081] The beneficial effects of the above technical solution are: by refining the processing of multi-source data, deeply integrating the target of the campaign with the characteristics of the advertisement itself to generate candidate strategies, and combining scientific evaluation and dynamic optimization mechanisms to screen and iterate the optimal strategy, the intelligentization of the entire link of advertising campaign from data collection to effect optimization is realized, which significantly improves the targeting, effect stability and environmental adaptability of the campaign.

[0082] This invention provides an artificial intelligence-based advertising strategy generation system, wherein the vector construction module includes:

[0083] A partitioning unit is used to partition the multi-source data according to a dimensional index, and to extract features from the partitioned data according to the sub-indicators under the dimensional index, so as to obtain the basic sample features and feature weights of the corresponding sub-indicators.

[0084] The fusion unit is used to fuse the weighted sub-features to obtain the single-dimensional initial feature set of the corresponding dimension index based on the feature weights, sample features and the correlation strength between sub-indicators.

[0085] The representation unit is used to input the initial single-dimensional feature set of all dimensions into the cross-dimensional global attention network and output the enhanced feature representation of each dimension index.

[0086] The building unit is used to map the enhanced feature representations of each dimension indicator to a unified high-dimensional feature space, and to construct the feature vector of each dimension indicator.

[0087] In this embodiment, the segmented data is a subset of multi-source data classified according to dimensional indicators, containing only data related to that dimension. For example, the segmented data for the user dimension only includes user behavior, profile, and other data, excluding platform traffic data.

[0088] Feature extraction involves extracting quantifiable information that reflects the essence of sub-indicators from segmented data, transforming the raw data into feature forms. For example, for text data (such as user search terms) of the user dimension "interest tags" sub-indicator, TF-IDF is used to extract keyword features such as "games" and "digital"; for image data of the advertising content dimension "visual" sub-indicator, ResNet is used to extract visual features such as "product outline" and "red main color". Specifically, text features are extracted using Python's Scikit-learn library (TF-IDF), and image features are extracted using a pre-trained model (ResNet50) from PyTorch.

[0089] Basic sample features are the original quantitative features of sub-indicators obtained after feature extraction. They directly reflect the attributes of sub-indicators. For example, the basic sample feature of the user dimension "age segmentation" sub-indicator is "25-30 years old" (quantified as 0.6, corresponding to the standardized value of this range); the basic sample feature of the advertising content dimension "copy keyword density" sub-indicator is "0.2" (keyword proportion of 20%). The sample features, i.e. the basic sample features, are the original quantitative features of the sub-indicators.

[0090] Feature weights measure the importance of the basic sample features of a sub-indicator in a dimensional indicator. The higher the weight, the greater the impact on the dimensional feature. For example, in the user dimension, the feature weight of the "interest matching degree" sub-indicator (0.7) is higher than that of "device type" (0.3) because the former is more relevant to the advertisement. Specifically, the weights are determined by the AHP (analytic hierarchy process) method (a combination of subjective and objective methods).

[0091] In this embodiment, the correlation strength between sub-indicators is the correlation between different sub-indicator features under the same dimension index, reflecting the mutual influence between sub-indicators. For example, in the user dimension, the correlation strength between the sub-indicators "monthly spending" and "interest tag: luxury goods" is 0.8 (highly positive correlation), calculated using the Pearson correlation coefficient. Furthermore, the weighted sub-feature is the product of the basic sample feature and the feature weight, highlighting the influence of important sub-features. For example, the basic feature of "interest matching degree" in the user dimension is 0.8, with a weight of 0.7, and the weighted sub-feature is 0.8 × 0.7 = 0.56; the basic feature of "age segmentation" is 0.6, with a weight of 0.3, and the weighted sub-feature is 0.18.

[0092] A single-dimensional initial feature set is a feature set that integrates weighted sub-features and the correlation strength of sub-indicators under the same dimension indicator, reflecting the overall characteristics of that dimension. For example, the initial feature set of the user dimension is {0.56 (interest matching weighted feature), 0.18 (age segment weighted feature), 0.05 (correlation strength correction value)}, where the correlation strength correction value is calculated from the correlation between sub-indicators (such as the correlation between "interest" and "consumption" to supplement the overall characteristics).

[0093] Cross-dimensional global attention networks are neural network models based on attention mechanisms that can learn the correlations between indicators of different dimensions (such as the interaction between user dimensions and platform dimensions) and enhance the global representativeness of features. For example, this network can capture the correlation between "user interests (user dimension)" and "platform traffic peak hours (platform dimension)" (such as "game-interested users" being active from 20:00 to 22:00), thereby strengthening the interactive information in the features. Specifically, it adopts the Transformer model architecture, calculates the attention weights between dimensions through a multi-head self-attention mechanism, and outputs features that fuse cross-dimensional correlations.

[0094] Enhanced feature representations are features obtained after processing by a cross-dimensional global attention network. They not only contain single-dimensional information but also integrate implicit relationships between dimensions, making them more expressive. For example, the enhanced feature representation of the user dimension not only includes the user's own attributes but also implicitly contains the association information of "the user group during peak traffic periods on platform A"; the enhanced feature representation of the platform dimension integrates "the matching tendency of users on the platform with advertising content".

[0095] A unified high-dimensional feature space is a standardized high-dimensional mathematical space that maps enhanced feature representations of different dimensions to the same space, eliminating differences in dimensionality and distribution between dimensions, and facilitating subsequent comprehensive analysis. For example, the enhanced feature representations of user dimension (originally 3-dimensional), platform dimension (originally 4-dimensional), and advertising content dimension (originally 5-dimensional) can be mapped to a 10-dimensional unified space through embedding, enabling direct comparison and computation of features across dimensions.

[0096] Feature vectors are the final quantized representation after mapping to a unified high-dimensional feature space. They are presented as numerical arrays and are the core input for subsequent strategy generation. For example, the user dimension feature vector is [0.82, 0.35, 0.61, 0.12, 0.48, 0.73, 0.29, 0.55, 0.91, 0.37] (10 dimensions), where each element corresponds to one dimension of the high-dimensional space, comprehensively reflecting user attributes and cross-dimensional relationships. The platform dimension feature vector is another 10-dimensional array, reflecting platform characteristics and their relationship with users and content.

[0097] The beneficial effects of the above technical solution are as follows: through the entire process of multi-source data partitioning, sub-feature extraction and weighting, cross-dimensional association enhancement, and unified spatial mapping, the data is processed in a refined manner. The implicit association between dimensions is captured through the cross-dimensional attention mechanism. The final generated feature vector has both accuracy and richness, providing a high-quality feature foundation for the subsequent generation of delivery strategies and improving the targeted generation of strategies.

[0098] This invention provides an artificial intelligence-based advertising strategy generation system, which further includes:

[0099] The system comprises the following modules: a frame extraction module, a core feature vector, and a continuous partitioning module. The frame extraction module extracts the text content, canvas content, and voice-over content of each frame in the advertisement to be displayed, obtaining corresponding frame feature vectors. Simultaneously, it obtains a core feature vector based on the core theme of the advertisement. The calculation module calculates the first comprehensive distance between each frame feature vector and the core feature vector, as well as the individual element distance between each frame feature vector and the core feature vector, obtaining an analysis array for the corresponding frame. The association acquisition module inputs each frame feature vector into a feature analysis model to obtain the association between elements in the corresponding frame, obtaining an association array. The probability determination module determines the display probability of the corresponding frame based on the analysis array and the association array. The continuous partitioning module constructs a continuous partitioning module based on the display probability of each frame in the advertisement to be displayed. A probability sequence is used to divide the advertisement to be displayed into continuous segments based on the probability sequence; an identifier setting module is used to assign a placeholder playback identifier to the corresponding individual continuous segment according to the output meaning and average probability of each individual continuous segment, wherein the placeholder playback identifier is unique and related to the playback position and playback timing logic; a condition adding module is used to add trigger conditions to the sub-videos of each individual continuous segment. When the corresponding trigger condition is met, the sub-videos of the individual continuous segment that match the trigger condition are controlled to be displayed in original color in a designated small window at the set playback position of the overall canvas, and the sub-videos of the individual continuous segment that do not meet the trigger condition are controlled to be displayed transparently at the set playback position, until the corresponding individual continuous segment finishes playing according to the set playback time, thereby obtaining its own playback strategy.

[0100] In this embodiment, the advertisement to be delivered refers to the advertisement content to be delivered, which may include videos, images and text, etc., and is the object of subsequent feature extraction and strategy generation.

[0101] Each frame is a single static image that makes up a video advertisement. It is the basic unit for extracting visual features. For example, a 60-second video, calculated at 25 frames per second, has a total of 1,500 frames. The 300th frame may show a product usage scenario.

[0102] Text content refers to the text information contained in each frame, such as titles, advertising slogans, and product parameters.

[0103] The canvas content is the layout and visual features of each frame, including color, composition, and object outlines. For example, the canvas content of a certain frame might be a blue background with a close-up of the product in the center and the brand logo in the lower right corner.

[0104] The voiceover content is the text-to-text conversion information (through speech recognition) corresponding to the voice narration, background music, or sound effects in the video advertisement.

[0105] A frame feature vector is a numerical array formed by quantifying the features of the text content, canvas content, and voice-over content of a single frame. It is used to represent the comprehensive features of the frame. For example, the feature vector of a certain frame is [0.8 (matching degree between text and the theme of "moisturizing"), 0.6 (product proportion in the canvas), 0.9 (relevance of voice-over keywords)]. Each element corresponds to the standardized value of a type of feature. Specifically, the text features are converted into vectors using Word2Vec, the visual features are output using CNN, and the voice-over features are output using TF-IDF. They are then concatenated into a vector of a unified dimension (e.g., 128-dimensional).

[0106] The core theme is the core message or marketing focus of the advertisement. It is the benchmark for measuring the relevance of each frame and is pre-set. For example, the core theme of a skincare product advertisement is "moisturizing and anti-allergy + limited-time offer".

[0107] The core feature vector is a quantitative representation of the core theme of the advertisement, used to compare similarity with the frame feature vector. For example, the feature vector of the core theme "moisturizing and anti-allergy + limited-time offer" is [0.9 (moisturizing weight), 0.8 (anti-allergy weight), 0.7 (offer weight)].

[0108] The first comprehensive distance is a measure of the overall difference between the frame feature vector and the core feature vector, reflecting the overall matching degree between the frame image and the core theme. Specifically, it is calculated by calculating the cosine similarity between the vectors.

[0109] The individual element distance is the difference between the corresponding elements (sub-features) in the frame feature vector and the core feature vector, reflecting the matching degree of a single type of feature. For example, if the value of the discount information element in the frame feature vector is 0.2, and the corresponding element in the core feature vector is 0.7, the individual element distance is 0.5 (the difference is large, indicating that the discount information in this frame is weak).

[0110] The analysis array is an array composed of the first overall distance and the distances of all individual elements, comprehensively depicting the differences between the frame and the core theme. For example, the analysis array of a certain frame is [0.3 (first overall distance), 0.1 (moisturizing feature distance), 0.5 (discount feature distance)], which intuitively reflects the shortcomings of this frame in terms of discount features.

[0111] In this embodiment, the feature analysis model is an algorithm model used to mine the correlation between elements (text, objects, colors, etc.) in a frame. It is constructed using a graph neural network (GNN). The model takes "product bottle", "moisturizing text" and "blue background" in the frame as nodes and calculates the connection weight between nodes (e.g., the correlation weight between "product bottle" and "moisturizing text" is 0.8). Its input is the frame feature vector and its output is the element correlation matrix.

[0112] The relationship between elements refers to the mutual influence or synergy between different elements (such as text and objects, color and theme) in a frame. For example, in a certain frame, the "red promotional label" is strongly related to the "limited-time offer copy" (together they reinforce the promotional information), but weakly related to the "product ingredient description".

[0113] An associative array is an array that quantifies the correlation between elements in a frame. Each element represents the correlation strength between a pair of elements. For example, an associative array of [0.8(promotional label - discount copy), 0.2(promotional label - ingredient description)] reflects the degree of synergy between elements.

[0114] The display probability is the probability that a frame will be prioritized for display (or displayed clearly) during an advertisement playback. It is jointly determined by the analysis array (theme matching degree) and the association array (element synergy). Specifically, it is: Display Probability = 0.4 × (1 - First Comprehensive Distance) + 0.3 × (1 - Average Distance Between Individual Elements) + 0.3 × (Average Association Degree), where 0.4, 0.3, and 0.3 are randomly varied according to the actual scenario.

[0115] A probability sequence is a sequence of the display probabilities of each frame arranged in chronological order, reflecting the overall display priority distribution of an advertisement. For example, the probability sequence of a 60-second video is [0.3, 0.5, 0.8, ..., 0.6] (a total of 1500 elements), with the first 100 frames having lower probabilities and the middle 500 frames having higher probabilities.

[0116] Continuous segmentation divides an advertisement into several consecutive video segments based on the continuity of the probability sequence (such as consecutive high probabilities or consecutive low probabilities). For example, if the display probability of frames 300-800 in the probability sequence is ≥0.7, it is divided into the "core information segment"; if the probability of frames 1-299 is ≤0.4, it is divided into the "prelude segment". Specifically, a sliding window is used to detect the continuous threshold region of the probability sequence. When the probability of 10 consecutive frames is ≥0.6, they are merged into a continuous segment.

[0117] A single continuous segment is an independent video clip after being divided. Each segment has a relatively uniform display probability and content theme. For example, the "core information segment" is a single continuous segment that includes the display of the product's core selling points and lasts for 20 seconds.

[0118] The output meaning is the core information or function conveyed by a single continuous segment, such as product introduction, promotion guidance, brand display, etc.

[0119] The average probability is the average of the display probabilities of all frames within a single continuous segment, reflecting the overall importance of that segment.

[0120] The placeholder playback identifier is a unique identifier assigned to a single continuous segment, which includes playback rules such as associated playback position and timing logic. For example, the identifier "Seg-001" corresponds to the "core information segment", and the associated rule is "playback position: center of the screen, timing logic: play from the 10th to the 30th second".

[0121] The playback position is the display area of ​​a continuous segment within the overall advertisement canvas, such as full screen or a small window in the upper right corner. For example, the core information segment is played in the center of the full screen, while the prelude segment is played in the small window in the lower right corner of the screen.

[0122] Playback timing logic: Playback time rules for consecutive segments, including absolute time (e.g., starting at the 10th second) and relative triggering (e.g., starting after the previous segment ends). For example, the timing logic for the core information segment is to automatically start after the preceding prelude segment is played, lasting for 20 seconds.

[0123] Triggering conditions are user behaviors or time conditions that control the switching of the visible and hidden states of continuous segments, such as user clicks, dwell time targets, specific time points, etc. For example, the triggering conditions for the core information segment are "user clicks any position on the screen" or "video plays to the 10th second".

[0124] Sub-videos are video files corresponding to separate consecutive segments, and their visibility can be controlled independently. For example, the sub-video of the "core information segment" is a 20-second MP4 file that has been extracted.

[0125] The overall canvas is the entire screen area for ad playback, and each sub-video is displayed within the canvas according to rules.

[0126] Setting the playback position is the predefined display coordinates and size of the sub-video in the overall canvas. For example, the core information segment is set to "coordinates (0,0) to (1080,1920)" (full screen), and the prelude segment is set to "coordinates (720,1440) to (1080,1920)" (small window in the lower right corner).

[0127] Specify small window: A local area within the overall canvas used to display a specific sub-video. Its size is smaller than the full screen. For example, a small window of 100×100 pixels in the upper right corner is used to display the "Brand LOGO segment" sub-video.

[0128] Original Color Display: Sub-videos are displayed with normal brightness and clarity, highlighting the content of that segment.

[0129] Transparent display: Sub-videos are displayed with low transparency (e.g., 20%), weakening the content of that segment and not interfering with the core information.

[0130] The playback time setting is a preset playback duration for continuous segments, ensuring that the content is displayed completely.

[0131] The beneficial effects of the above technical solution are as follows: through refined analysis of the characteristics of the advertising frame, continuous segment division based on the core theme, dynamic identifier allocation and trigger-based display control, the solution enables the advertisement to dynamically adjust its display logic according to the importance of its content and user behavior, which not only ensures the effective transmission of core information, but also reduces interference to users through transparent display, ultimately improving the information transmission efficiency and user acceptance of the advertisement.

[0132] This invention provides an artificial intelligence-based advertising strategy generation system, wherein the identifier setting module includes:

[0133] The multidimensional extraction unit is used to extract the multidimensional set Wi = {Gi, Pi, Ti, Ki, Xi} for each individual continuous segment, where Gi, Pi, Ti, Ki, and Xi are, respectively, the semantic encoding of the output meaning of the i-th individual continuous segment, the normalized value of the average probability, the dependency relationship with the predecessor segment, the triggering relationship with the successor segment, and the spatial requirements based on the required canvas area ratio and window shape.

[0134] The matching retrieval unit is used to retrieve the tuple {Fui,Pui,Tui} with the highest matching degree from the identifier pool when the corresponding individual continuous segment meets the preset conditions, and automatically associate the end event of the predecessor segment with the start condition of the successor segment. Here, Fui, Pui, and Tui represent the function identifier, playback position rule, and playback timing logic of the output meaning mapping of the i-th individual continuous segment, respectively.

[0135] The tuple update unit is used to collect interactive feedback test data on the advertisement to be played in real time, and update the priority weight of the playback position of the tuple and the periodic parameters of the timing logic.

[0136] In this embodiment, the multidimensional set is a five-dimensional data set used to comprehensively describe the characteristics of a single continuous segment, covering semantics, probability, dependency, triggering relationship and spatial requirements. For example, the multidimensional set of the limited-time discount segment is [0.85 (semantic encoding), 0.9 (normalized probability), end of dependent product display segment (precursor dependency), trigger payment guidance segment (subsequent trigger), full screen occupancy 80%, rectangular window (spatial requirements)].

[0137] Semantic encoding of output meaning transforms the core information (output meaning) of a single continuous segment into a numerical vector that can be recognized by a computer, quantifying semantic features. For example, the output meaning of a limited-time discount segment is promotional offer, which is encoded into a 128-dimensional vector by the BERT model, where the dimension value related to "offer" and "discount" is relatively high (e.g., 0.85).

[0138] The normalized value of the average probability is a value in the 0-1 range obtained by normalizing the average display probability of all frames in a single continuous segment through Min-Max normalization (eliminating the difference in probability scale between different segments).

[0139] The dependency relationship with the preceding segment is the playback dependency of the current individual continuous segment on the preceding continuous segment (such as it must be started after the preceding segment ends). For example, the preceding dependency relationship of the "payment guidance segment" is that it must be started after the limited-time discount segment has finished playing, so as to avoid users being guided to pay before they see the discount.

[0140] The triggering relationship with subsequent segments is the activation rule for subsequent segments after the current single continuous segment ends (such as active triggering or conditional triggering). For example, the subsequent triggering relationship of a limited-time discount segment is that the payment guidance segment is automatically triggered after the playback ends, forming a coherent process of promotion → conversion.

[0141] Space requirements refer to the canvas occupancy requirements when playing individual continuous segments, including area proportion (e.g., 60% of the full screen) and window shape (e.g., rectangle, circle, floating window). For example, the "product detail segment" needs to highlight the product close-up, and the space requirement is 70% of the canvas in the center, with a rounded rectangular window; the brand logo segment only needs 10% of the upper right corner, with a square window.

[0142] The preset conditions are the trigger thresholds for a single continuous segment to enter the identifier allocation process. They are usually related to the average probability and semantic importance. For example, the preset conditions are "average probability normalization value ≥ 0.7 and semantic encoding matching degree with the core theme of the advertisement ≥ 0.8". Only segments that meet these conditions (such as "limited-time discount segment") will enter the retrieval stage.

[0143] The identifier pool is a collection that stores all available placeholder playback identifiers. Each identifier exists in the form of a tuple, which covers function, location, and timing information. For example, the identifier pool contains a set of tuples: {("Promotion", "Full Screen", "Start after prequel ends"), ("Brand", "Top Right Window", "Force Start at 30 Seconds"), ...}, for a total of 50 tuples.

[0144] A tuple is a triplet of data in the identifier pool used to describe the placeholder playback identifier. It contains the function identifier, playback position rules, and playback timing logic. For example, the tuple (Fui = "Promotion", Pui = "Full Screen Center", Tui = "Start 5 seconds after the preceding segment ends") corresponds to the playback rules of the continuous segments of the promotion category.

[0145] Matching degree is the degree of fit between a single continuous segment's multidimensional set and the identifier pool tuple. It is calculated through multidimensional feature comparison. For example, the matching degree between the limited-time discount segment and the tuple ("promotion", "full screen", "start after the previous sequence ends") is 0.92 (semantic matching 0.95, position matching 0.9, temporal matching 0.9), which is higher than other tuples (average 0.7). It is calculated using weighted cosine similarity.

[0146] The end event of the preceding segment is a signal that the preceding segment has finished playing (such as time expired or content ended). It serves as the trigger source for starting the current segment. For example, when the product display segment ends at 15 seconds, the segment A end event is generated, triggering the start of the limited-time discount segment.

[0147] The start condition for a subsequent segment is the rule that the subsequent segment begins playing after the current segment ends. It is usually bound to the end event of the current segment. For example, the start condition for the subsequent segment of a limited-time discount segment is the end event of segment B + the user has not exited the page. The payment guidance segment is only displayed when the user is still watching.

[0148] Interaction feedback test data is behavioral data generated when users watch ads. It is used to evaluate the effectiveness of playback indicators. For example, interaction data for limited-time discount periods includes user dwell time ≥ 8 seconds (positive feedback), clicking skip (negative feedback), and clicking the discount button (strong positive feedback).

[0149] The priority weight of the playback position is the importance coefficient of different playback positions in the identifier pool. The higher the weight, the higher the priority of matching. For example, the initial priority weight of the full-screen position is 0.8. After feedback, it was found that full-screen is easily disliked by users on mobile devices, so it was adjusted to 0.6. The weight of half-screen was increased from 0.5 to 0.7.

[0150] The period parameter of the timing logic is a value that controls the playback timing interval (such as X in "start after X seconds after the prequel"). It can be adjusted according to feedback. For example, the initial period parameter of the limited-time discount segment is to start 2 seconds after the prequel ends. Feedback shows that the user needs 1 second of buffering, so it is adjusted to start after 1 second.

[0151] The beneficial effects of the above technical solution are as follows: by extracting multi-dimensional features (semantics, probability, dependency, etc.) of individual continuous segments, the optimal playback identifier is accurately matched from the identifier pool, and the position weight and timing parameters are dynamically optimized based on user interaction feedback. This solution achieves a unified "accuracy-coherence-adaptability" in the allocation of playback identifiers, ensuring that the playback logic of each segment of the advertisement is smooth and that adjustments can be made in real time according to user behavior, ultimately improving the efficiency of advertising information delivery and user acceptance.

[0152] This invention provides an artificial intelligence-based advertising strategy generation system, wherein the strategy determination module includes:

[0153] The semantic enhancement unit is used to perform multi-scale attention fusion on feature vectors under all dimensions, calculate the importance weights of feature vectors in each dimension, and generate a weighted fused feature vector.

[0154] The matrix construction unit is used to semantically enhance the weighted fusion feature vector based on the advertising domain knowledge graph, and construct an enhanced feature matrix containing explicit features and implicit semantic associations;

[0155] The input unit is used to extract the interaction features and temporal features of the playback strategy itself, and input them into the playback strategy temporal model to generate a playback strategy feature tensor.

[0156] Define a unit, used to define the set of deployment constraint domains D = {D j1,j1＝1,2,3,...,Nk ], where D j1 Let Nk represent the j1-th constraint domain, and let Nk represent the number of constraint domains. For each constraint domain, a bidirectional group {Chj1, Cgj1} is constructed, where Chj1 is the constraint verification identifier of the j1-th constraint domain, and Cgj1 is the policy generation trigger identifier of the j1-th constraint domain.

[0157] The decomposition unit is used to decompose the target of the advertisement to be played into a set of subtasks ∈={∈j2,j2=1,2,3,..,Nq}, and construct a directed acyclic dependency graph according to the domain dependency relationship, where ∈j2 represents the j2-th subtask; Nq represents the number of subtasks; and the subtasks in the directed acyclic dependency graph ∈ j3 Dependence on D j1 Chj1 must be completed before startup can begin, and ∈ j3 and ∈ j3+1 Dependency edges;

[0158] Binding unit, used to bind adjacent subtasks in the directed acyclic dependency graph ∈ j3 and ∈ j3+1 Binding to the same constraint domain D j1 Establish a driver execution chain for the bidirectional group {Chj1,Cgj1}:

[0159] ∈ j3+1 Continuous monitoring of D using an adaptive monitoring window j1 The Chj1 state, when ∈ j3 Once completed, send a message to D. j1 Write to the active state Chj1;

[0160] When ∈ j3+1 After Chj1 activation is detected, from D j1 Extract Cgj1 from it, trigger its own execution, and output the domain policy parameters;

[0161] The clustering fusion unit is used to perform clustering analysis on the enhanced feature matrix and the policy feature tensor, and fuse them with the domain policy parameters to generate multiple candidate deployment strategies.

[0162] In this embodiment, multi-scale attention fusion is a fusion method that assigns attention weights to feature vectors of different dimensions at multiple scales (such as time granularity and feature granularity). It considers both the importance of features within a single dimension and the correlation between dimensions. For feature vectors of user dimension (fine granularity: age / interest), platform dimension (medium granularity: DAU / time period), and advertising dimension (coarse granularity: type / duration), attention weights are calculated at two time scales: "hour" and "day", highlighting features with high matching degree.

[0163] The feature vectors for each dimension are the feature vectors for the user, platform, and advertising dimensions in step 1 (e.g., the user dimension vector [25 years old, 0.8 interest matching degree]).

[0164] Importance weight: A coefficient (between 0 and 1) that measures the contribution of each dimension feature vector to the target of the campaign. The higher the weight, the greater the proportion of that dimension in the fusion. For example, when the target of the campaign is "conversion rate", the user dimension has a weight of 0.6 (great impact on user accuracy), the platform dimension has a weight of 0.3 (impact on traffic quality), and the advertising dimension has a weight of 0.1 (content suitability).

[0165] The weighted fusion feature vector is a comprehensive vector obtained by multiplying the feature vectors of each dimension with their corresponding importance weights and then summing them. It centrally reflects the core information of multiple dimensions. For example: user vector [0.8, 0.3] (weight 0.6) + platform vector [0.5, 0.7] (weight 0.3) + advertising vector [0.2, 0.9] (weight 0.1) = fusion vector [0.65, 0.48].

[0166] The knowledge graph in the advertising domain is a structured graph that stores advertising-related entities (such as "skincare products" and "sensitive skin"), attributes (such as "moisturizing" and "promotion"), and relationships (such as "skincare products are suitable for sensitive skin"). Specifically, it involves building the knowledge graph using Neo4j and querying implicit relationships between entities using SPARQL.

[0167] Semantic enhancement involves injecting implicit associations from the knowledge graph (such as the synonymy between "promotion" and "limited-time discount") into the feature vector to enrich the semantic expression of the features. For example, the "promotion" feature in the feature vector can be linked to implicit features such as "discount for purchases over a certain amount" and "gifts" through the knowledge graph, thereby enhancing support for the conversion goal.

[0168] The enhanced feature matrix is ​​a two-dimensional matrix containing explicit features (original feature vector elements) and implicit semantic associations (supplementary features from the knowledge graph). Rows represent samples and columns represent features. For example, the matrix rows correspond to different user groups, and the columns contain features such as "age", "interests" (explicit), and "potential purchase time period" (implicit, from the knowledge graph).

[0169] The interactive features of the self-playing strategy are the interaction rules between users and content when the advertisement is played, such as "click to trigger original color display" and "stay for 3 seconds to display details". For example, the interactive features are quantified as [0.8 (click sensitivity), 0.3 (stay threshold)], and the higher the value, the easier it is to trigger the interaction.

[0170] Temporal features are the time-dimensional features of the playback order, duration, and interval of each consecutive segment of an advertisement, such as "segment 1 (5 seconds) → segment 2 (10 seconds, interval 1 second)". For example, the temporal features are represented as [5,10,1] (segment 1 duration, segment 2 duration, interval).

[0171] Playback strategy time series models are models that process time series features and capture the temporal dependencies between consecutive segments (such as LSTM and Transformer). Specifically, they involve building an LSTM model, taking the time series sequence as input, and outputting the hidden states as features.

[0172] The playback strategy feature tensor is a three-dimensional feature array (samples × time steps × features), which integrates interaction features, temporal features, and temporal dependencies. For example, the tensor shape is [10 (number of samples) × 3 (time steps: segment 1 / segment 2 / segment 3) × 5 (features: duration / interaction sensitivity / dependency strength, etc.)].

[0173] The set of ad placement constraint domains is a set of restriction domains that ad placement must meet, such as budget, compliance, and traffic. Each domain is a constraint domain. The constraint verification flag is used to verify whether the constraint domain conditions are met. The strategy generation trigger flag is the flag that allows the generation of corresponding domain strategy parameters after the constraint conditions are met.

[0174] The campaign objective is broken down into a set of sub-tasks: the overall objective (such as "ROI≥3") is broken down into actionable sub-tasks, such as "impressions≥100,000" or "conversion rate≥5%". For example, the sub-task set is: {audience targeting, bid optimization, creative scheduling}. Domain dependency is the sub-task's dependence on the constraint domain (such as "bid optimization" depending on the constraint of the "budget domain").

[0175] A directed acyclic dependency graph is a graph that uses directed edges to represent the dependencies between subtasks (without circular dependencies). For example, in the graph "Audience Targeting → Bidding Optimization", it means that bidding optimization depends on audience targeting.

[0176] Adjacent subtasks are subtasks that are connected one after the other in the dependency graph and are bound to a bidirectional group with the same constraint domain: adjacent subtasks share a constraint domain, ensuring that the dependency relationship is constrained and managed.

[0177] The adaptive monitoring window is a time window that dynamically adjusts the monitoring frequency (e.g., monitoring once every 10 minutes when the budget is sufficient, and once every 2 minutes when the budget is insufficient).

[0178] Activation state: The effective state of Chj1 (e.g., activation of budget domain Chj1 indicates "sufficient budget").

[0179] Domain strategy parameters are the specific delivery parameters corresponding to the constraint domain, such as the budget domain outputting "time period bid cap" and the traffic domain outputting "platform selection weight".

[0180] Cluster analysis groups similar samples in the enhancement feature matrix (multi-dimensional features) and the strategy feature tensor (playback strategy features) into a class. Each class corresponds to a set of potential strategy directions. K-means is used to cluster user features into two classes: "young women" and "middle-aged men". Each class corresponds to different delivery parameters.

[0181] Fusion (enhanced feature matrix + policy feature tensor + domain policy parameters): Combine the three types of features according to their weights to generate a set of policy parameters containing multi-dimensional information. For example, fuse the "young women" cluster feature + "click to trigger playback" policy feature + "budget limit of 5000 yuan" domain parameter to form a candidate policy.

[0182] Multiple candidate targeting strategies are generated by merging different parameter combinations, such as: target audience: women aged 25-30, channel: Xiaohongshu, bid: 0.8 yuan / click; target audience: women aged 30-35, channel: Douyin, bid: 1.2 yuan / click, etc.

[0183] The beneficial effects of the above technical solution are: by fusing features from various dimensions through multi-scale attention, enhancing semantic association through knowledge graphs, capturing playback strategy patterns through temporal models, and combining the collaborative execution of sub-tasks under constraint domain control, the final generated candidate delivery strategy integrates both explicit and implicit features of multi-source data, laying a high-quality foundation for subsequent selection of the optimal strategy.

[0184] This invention provides an artificial intelligence-based advertising strategy generation system, wherein the effect evaluation module includes:

[0185] The model building unit is used to build a response surface model based on historical deployment data and fit the response function f(M) through Gaussian process regression.

[0186] The factor determination unit is used to map the candidate delivery strategy to the response function f(M) to generate a perturbation sequence. At the same time, it extracts the influencing factors related to user behavior from the candidate delivery strategy, and combines the user group behavior model to characterize the uncertainty of user behavior to determine the probability distribution of the potential interaction path of individual users to the advertisement.

[0187] The comparison unit is used to perform a first comparison between the perturbation sequence and the first evaluation table, and at the same time, to perform a second comparison between the probability distribution of the potential interaction path and the second evaluation table.

[0188] The prediction unit is used to predict the delivery effect of the corresponding candidate delivery strategy based on the first control result and the second control result.

[0189] In this embodiment, historical campaign data is structured data accumulated during past advertising campaigns, including campaign parameters (such as audience targeting, bid, and time period) and corresponding performance metrics (such as click-through rate, conversion rate, and ROI). For example, it includes 100,000 historical records of 25-30 year old women + bid of 0.8 yuan + campaign at 19:00 → click-through rate of 5% and ROI of 3.2.

[0190] The response surface model is a mathematical model that describes the nonlinear relationship between input parameters (targeting strategy) and output results (targeting performance). It is used to fit the surface relationship between parameter combinations and effects. The input of this model is "bid (x1) and targeting precision (x2)", and the output is "click-through rate (y)". The fitting surface formula is y = 0.2x1 + 0.5x2 - 0.1x1x2 + a0 (a0 is the error).

[0191] Gaussian process regression is a nonparametric regression method that assumes the data follows a Gaussian process and uses a kernel function to capture the nonlinear relationship between variables, outputting a prediction result with a confidence interval. For example, the RBF kernel function (radial basis function) can be used to fit the "bid-click-through rate" relationship and output the predicted click-through rate and a confidence interval of ±0.5%.

[0192] The response function f(M) is a function obtained by fitting a Gaussian process regression. The input is the delivery parameter vector M (such as [bid, time period, audience tags]), and the output is the predicted delivery effect (such as expected conversion rate), for example, bid audience matching degree and time period deviation, which quantifies the comprehensive impact of the combination of parameters on the effect.

[0193] Candidate campaign strategies are potential campaign plans to be evaluated, including specific parameter combinations (such as "target audience: women aged 25-35, bid: 1.2 yuan, platform: Douyin, time period: 20:00-22:00").

[0194] Mapping transforms the parameters of candidate delivery strategies into an input format (such as vector form) acceptable to the response function f(M), establishing a correspondence between strategies and functions. For example, a candidate strategy with a bid of 1.2 yuan and a target audience matching degree of 0.8 is mapped to an input vector M = [1.2, 0.8], which is then substituted into f(M) to calculate the prediction effect.

[0195] The perturbation sequence is a parameter sequence generated by adding small fluctuations (such as ±5%) to the candidate strategy parameters. It is used to simulate possible parameter deviations (such as bid fluctuations and audience targeting errors) in actual deployment. For example, a perturbation sequence [1.14, 1.2, 1.26] is generated for a bid of 1.2 yuan to simulate the fluctuation of the bid within the allowable range and evaluate the stability of the strategy.

[0196] Influencing factors (related to user behavior): Key parameters in candidate strategies that affect user interaction behavior, such as creative attractiveness, the exposure position corresponding to the bid, and the user activity corresponding to the time period. For example, the influencing factors of a certain strategy are [0.9 (creative click-through rate), 0.7 (exposure position weight), 0.8 (time period activity)], which quantify their driving effect on user behavior.

[0197] User group behavior models are mathematical models that characterize the behavioral patterns of a specific group of people. They include typical user response patterns to advertisements (such as the conversion path probability of "browse → click → add to cart → purchase"). For example, a model based on a Bayesian network outputs a probability of 0.3 for browsing and 0.2 for purchasing among women aged 25-30.

[0198] The uncertainty of user behavior is the randomness of user behavior deviating from typical patterns (such as some users in the same group making direct purchases while others only browse). For example, the average probability of purchasing after clicking is 0.2, but individual differences cause the probability to fluctuate between 0.05 and 0.35.

[0199] The probability distribution of potential interaction paths is the probability distribution of an individual user's possible interaction paths with an advertisement (such as "ignore → close", "browse → click → exit", "browse → click → purchase") and their occurrence. For example, a user's interaction path distribution is: "ignore" 30%, "browse → exit" 50%, "browse → click → purchase" 20%, quantifying the probability of each path.

[0200] The first evaluation table is a mapping table that stores historical perturbation sequences and their corresponding actual effects. It is used to convert the current perturbation sequence into predicted effects. For example, the table records "bid perturbation +5% → click-through rate fluctuation +2%" and "audience targeting perturbation -3% → conversion rate fluctuation -1%", for a total of 1,000 mapping relationships.

[0201] The first comparison involves matching the perturbation sequence of the current candidate strategy with the first evaluation table to find the effect fluctuations corresponding to similar perturbations and predict the stability of the strategy's effect under parameter fluctuations. For example, the perturbation sequence [1.14, 1.2, 1.26] is compared with the first evaluation table to obtain the predicted click-through rate fluctuation range [4.8%, 5.0%, 5.2%].

[0202] The second evaluation table stores the correspondence between the probability distribution of interaction paths and their actual effects, such as "20% probability of purchase path → corresponding conversion rate of 4%". For example, the table contains mappings such as "10% of 'purchase' in the interaction path distribution → conversion rate of 2%" and "20% → conversion rate of 4%", built based on 100,000 user behavior data. The second comparison matches the current potential interaction path probability distribution with the second evaluation table to predict the conversion effect corresponding to the strategy. For example, if the probability of the "purchase" path for a certain strategy is 20%, the predicted conversion rate is 4% according to the second evaluation table.

[0203] The first control result is the effect fluctuation range obtained after matching the perturbation sequence with the first evaluation table (e.g., click-through rate 4.8%-5.2%). The second control result is the effect prediction value obtained after matching the interaction path probability distribution with the second evaluation table (e.g., conversion rate 4%). The campaign effect prediction combines the first and second control results to output a comprehensive effect index of the candidate strategy (e.g., "click-through rate 5.0% ± 0.2%, conversion rate 4%, ROI 3.5"). For example, by combining the click-through rate fluctuation range and the conversion rate prediction value, the expected value and fluctuation range of ROI are calculated as the final prediction result.

[0204] The beneficial effects of the above technical solution are: by constructing a response surface model based on historical data to capture the nonlinear relationship between strategy and effect, by combining perturbation sequence to simulate parameter fluctuations and user behavior model to quantify interaction uncertainty, and by comparing two evaluation tables to achieve dual prediction of macro-effect trends and micro-behavioral details, the final output of the delivery effect includes both expected value and fluctuation range, which significantly improves the accuracy and robustness of candidate strategy evaluation and provides a reliable basis for subsequent optimal strategy selection.

[0205] This invention provides an artificial intelligence-based advertising strategy generation system, wherein the effect evaluation module further includes:

[0206] The comparison unit is used to mine the causal chain of the deployment parameters from the deployment effect, and analyze the causal chain based on the structural causal model to obtain the simulated direct causal and simulated indirect causal of the corresponding deployment parameters on the effect index, and compare them with the ideal direct causal and ideal indirect causal of the set ideal effect.

[0207] The perturbation definition unit is used to analyze the distribution drift of the delivery data for the corresponding candidate delivery strategy and define the data distribution perturbation set;

[0208] The re-analysis unit is used to determine the strategy adjustment parameters based on the comparison results and the data distribution disturbance set, and to re-analyze the pre-simulation deployment effect;

[0209] The filtering unit is used to select the strategy with the best performance from the re-analyzed pre-simulation delivery results as the optimal delivery strategy.

[0210] In this embodiment, the delivery effect is a pre-simulated performance indicator of the candidate delivery strategy, such as click-through rate (CTR), conversion rate (CVR), ROI, and impressions. For example, the delivery effect of a certain strategy is CTR = 5%, CVR = 3%, and ROI = 4.2.

[0211] The campaign parameters are the specific settings that constitute the campaign strategy, such as bid (0.8 yuan / click), audience targeting (women aged 25-30), time period (19:00-21:00), and creative type (video / image / text).

[0212] A causal chain is a path of causal dependence between campaign parameters and performance metrics, namely a chain relationship of "parameter A → intermediate variable → parameter B → performance metric". This is different from a simple correlation. For example, by discovering the causal chain "increased bid → increased exposure → increased clicks → increased CVR", we can clarify that "bid" indirectly affects the conversion rate through "exposure" and "clicks", rather than directly.

[0213] Structural causal model (SCM) is a model that uses functional relationships and noise terms to represent causal relationships. It consists of three parts: "variable set, function set, and noise distribution". It can quantify the direct / indirect effects between variables. For example, in the model, the function of click-through rate (Y) is Y = f(bid X1, creative quality X2, noise ∈), where X1 indirectly affects Y through "exposure (M)" (M = g(X1), Y = h(M,X2)).

[0214] Simulating direct causality refers to the direct impact of campaign parameters on performance metrics (without intermediate variables). For example, creative quality directly affects click-through rate (CTR) (without relying on impressions). For instance, structural causal model analysis shows that the direct causal effect of creative quality on CTR is 0.3 (i.e., a 1-unit increase in creative quality directly increases CTR by 0.3 units). Simulating indirect causality refers to the impact of campaign parameters on performance metrics through intermediate variables. For instance, bid indirectly affects CTR through impressions. For instance, the indirect causal effect of bid on CTR is 0.2 (a 1-unit increase in bid → a 0.5-unit increase in impressions → a 0.2-unit increase in CTR).

[0215] Setting the ideal effect refers to the advertiser's expected target performance metrics and the corresponding causal effect distribution, such as a click-through rate of ≥5% and a direct causal ratio of creative quality of ≥60% (avoiding over-reliance on bid-driven performance).

[0216] Ideal direct causality and ideal indirect causality: When the desired effect is achieved, the direct / indirect causal effect of the placement parameters should be strong. For example, ideally, the indirect causal effect of "bid" on click-through rate should be ≤0.3 (to avoid excessive costs), and the direct causal effect of "creative quality" should be ≥0.4 (to highlight the value of the content).

[0217] The comparison involves comparing the simulated direct / indirect causality with the ideal value to identify biases, such as "actual bid indirect effect 0.5 > ideal 0.3", which requires adjusting parameters to reduce bid dependence.

[0218] The corresponding candidate targeting data: The dataset related to the strategy generated during the pre-simulation process, including user characteristics (age, interests), scene characteristics (time period, device), interaction characteristics (clicks, dwell time), etc. For example, the targeting data of a certain strategy includes distribution characteristics such as "70% of users are female aged 25-30 and 60% of clicks are from 19:00 to 21:00".

[0219] Distribution drift is a change in the statistical distribution of campaign data over time or in a given context (difference from historical data or expected distribution), such as "user age distribution changes from 25-30 years old to 30-35 years old" or "peak click time changes from 7 pm to 9 pm". For example, detecting "the proportion of 'beauty' in actual user interest tags drops from 40% to 20%" is an example of interest distribution drift.

[0220] The data distribution perturbation set is a collection of possible distribution drift types and degrees, used to simulate the stability of strategies under different drift scenarios. For example, the perturbation set includes 10 typical perturbations such as "age drift of the population (±5 years)", "decline in interest tag coverage (-20%)", and "time period traffic fluctuation (±30%)".

[0221] Strategy adjustment parameters are the delivery parameters that need to be corrected to eliminate bias in comparison results and adapt to data distribution disturbances. For example, "reduce the bid by 0.2 yuan", "improve the creative quality score (by optimizing the copy)" and "expand the target audience to 30-35 years old (to deal with age drift)". For instance, to address "excessive indirect effect of bidding", the parameter is adjusted to "reduce the bid cap from 1.2 yuan to 1.0 yuan"; to address "age distribution drift", the parameter is adjusted to "expand the target age to 25-35 years old".

[0222] The re-analysis of the pre-simulation campaign effect involves substituting the adjusted parameters into the pre-simulation evaluation process (such as the previous response surface model and user behavior model), recalculating the campaign effect, and verifying whether the adjustment is effective.

[0223] The best effect is the optimal combination of the following: "ideal causal effect", "resistance to distribution drift" and "meeting of core indicators (such as ROI≥4)".

[0224] The beneficial effects of the above technical solution are: by mining the causal chain between the deployment parameters and the effect and combining it with the structural causal model, the deviation of the parameter influence can be accurately located. At the same time, considering the disturbance scenario of data distribution drift, the determined strategy adjustment parameters can not only correct the causal structure to meet the ideal goal, but also enhance the strategy's anti-interference ability. Finally, the optimal deployment strategy selected not only meets the core effect indicators, but also significantly improves the sustainability and stability of the deployment.

[0225] This invention provides an AI-based advertising strategy generation system. The update module includes: a mapping unit, used to map newly collected multi-source data to corresponding spatiotemporal scene units, and simultaneously calculate the credibility weights of each newly collected data source based on a deep evidence network; a trajectory evolution unit, used to characterize the core parameters of the optimal advertising strategy as quantum superposition vectors and determine the quantum evolution trajectory of each quantum parameter; and a comparison and analysis unit, used to compare and analyze the quantum evolution trajectory with the mapping results and credibility weights, update the corresponding core parameters, and re-advertise.

[0226] In this embodiment, the newly collected multi-source data are various types of data generated in real time during the strategy execution process, covering dimensions such as users, platforms, and environments, such as real-time user click logs, platform traffic fluctuation data, competitor sudden launch information, and weather scene data (such as rainy days).

[0227] Spatiotemporal scene units are the smallest data units divided according to spatial dimensions (e.g., geographical location: first-tier city / second-tier city), time dimensions (e.g., time period: 10:00-12:00), and scene dimensions (e.g., user status: commuting / leisure). They are used to accurately match the context of data. For example, a spatiotemporal scene unit can be represented as "space = Beijing, time = 10:00-12:00, scene = commuting". Corresponding to the user behavior data within this unit, newly collected multi-source data is allocated to the corresponding spatiotemporal scene unit according to its spatial, time, and scene attributes, achieving precise binding between data and scene. For example, "click data of Beijing users at 10:30 (commuting time)" is mapped to the unit "space = Beijing, time = 10:00-12:00, scene = commuting".

[0228] Deep Evidence Network (DEN) is a model that integrates deep learning and evidence theory. It extracts data features through multi-layer neural networks and evaluates the credibility of data sources by combining the Dempster-Shafer evidence combination rule, outputting a weight value between 0 and 1 (1 being completely credible). For example, if the input features of "user click data" (such as click frequency, device consistency, and whether the IP is abnormal) are input, the network outputs a credibility weight of 0.9 (high credibility); if the input features of "third-party competitor data" (such as update frequency and historical accuracy) are input, the network outputs a credibility weight of 0.6 (medium credibility). Specifically, the deep evidence network is built using TensorFlow, and the training data consists of historical "data source feature-credibility label" pairs (human-labeled credible / uncredible cases).

[0229] Credibility weight is the credibility coefficient of each data source in the corresponding spatiotemporal scenario unit. The higher the weight, the greater the impact of the data on strategy updates. For example, the credibility weight of "official platform traffic data" is 0.9, the weight of "user self-reported interest data" is 0.3, and the weight of "third-party weather data" is 0.7, ensuring that strategy updates rely on high-quality data.

[0230] The core parameters of the optimal campaign strategy are the key settings that determine the effectiveness of the strategy, such as "bid coefficient (0.8-1.5)", "audience targeting tag weight (e.g., 'beauty interest' accounts for 0.6)", "time period distribution (e.g., 19:00-21:00 accounts for 50%)", and "creative rotation cycle (30 minutes)".

[0231] The quantum superposition vector is based on the concept of "superposition" in quantum mechanics. It represents the core parameter as a probability superposition (non-deterministic value) of multiple possible values, expressed in vector form. For example, the quantum superposition vector of the bidding parameter is 0.6|1.0>+0.4|1.2>, which means that the bid has a 60% probability of being 1.0 yuan and a 40% probability of being 1.2 yuan, reflecting the uncertainty and potential adjustment space of the parameter.

[0232] Quantum parameters are core parameters characterized by quantum superposition vectors. Unlike traditional deterministic parameters, they can simultaneously contain multiple possible values ​​and probability distributions. The quantum evolution trajectory is the path of change of quantum parameters with new data input (such as spatiotemporal scene data and credibility weights). It is simulated by the Schrödinger equation in quantum mechanics to reflect the dynamic migration of the probability of parameter values. For example, the initial evolution trajectory of the bidding parameter shows that "the probability of 1.0 increases from 60% to 80%" (because the new data shows that users are more sensitive to high prices), and finally converges to 0.8|1.0>+0.2|1.2>.

[0233] The quantum evolution trajectory is the dynamic change path of the core parameters, reflecting the potential value trend of the parameters under the influence of new data. The mapping result is the binding result of newly collected data → spatiotemporal scene unit, such as "user click data during Beijing commuting hours are concentrated in women aged 25-30".

[0234] The comparative analysis involves comparing the quantum evolution trajectory with the mapping results and credibility weights in multiple dimensions to determine the direction of parameter adjustment: if the trajectory shows that the bid needs to be reduced, and the click-through rate decreases during the high-bid period in the mapping results (supported by high-credibility data), then the bid should be reduced; if the trajectory conflicts with the high-credibility data (e.g., the trajectory shows an expanded audience but the high-credibility data shows that the audience targeting is too broad, resulting in a decrease in conversion rate), then the trajectory should be corrected, prioritizing the high-credibility data.

[0235] The core parameter update is based on the comparative analysis results, which collapses the quantum superposition state vector into a definite value (taking the ground state with the highest probability) to generate the updated core parameters. For example, the bidding parameter collapses from the superposition state 0.8|1.0>+0.2|1.2> to the definite value of 1.0 yuan; the target audience is narrowed from "25-35 years old" to "25-30 years old" (because high-reliability data shows that the conversion rate of 30-35 years old is low).

[0236] Re-launching involves applying the updated core parameters to the actual campaign process, overriding the original strategy parameters, and ensuring that the new strategy takes effect in real time. For example, by using the API interface of the advertising platform (such as the real-time strategy update interface of ByteDance's advertising platform), the adjusted bid, audience, and time period parameters are written into the campaign system, and full-channel synchronization is completed within 5 minutes.

[0237] The beneficial effects of the above technical solution are as follows: by accurately mapping new data to spatiotemporal scene units and quantifying their credibility, and by combining quantum superposition states and evolutionary trajectories to capture the dynamic change trend of parameters, the core parameters are accurately updated through multi-dimensional comparative analysis. This solution not only ensures the rapid response of the strategy to real-time data (spatiotemporal scene binding + priority of high-credibility data), but also takes into account the uncertainty and scientific nature of parameter adjustment through quantum models, so that the re-deployment strategy can still maintain high adaptability and effect stability in complex dynamic environments.

[0238] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. An artificial intelligence-based advertising strategy generation system, characterized in that, include: The vector construction module is used to collect multi-source data of the advertisement to be played, and to extract features from the multi-source data according to the dimensional indicators to construct the feature vector of each dimensional indicator. There are several dimensional indicators, and each dimensional indicator involves at least one sub-indicator. The strategy determination module is used to comprehensively analyze the feature vectors under all dimensions of the advertising to be delivered based on the delivery target of the advertising to be delivered, and to determine multiple candidate delivery strategies in combination with the playback strategy of the advertising to be played. The candidate delivery strategies include several delivery setting parameters. The performance evaluation module is used to perform pre-simulated delivery evaluation for each candidate delivery strategy, predict the delivery performance of each candidate delivery strategy, determine strategy adjustment parameters based on the delivery performance, and select the optimal delivery strategy. The update module is used to dynamically update and re-deploy the optimal delivery strategy by combining newly collected multi-source data.

2. The AI-based advertising strategy generation system according to claim 1, characterized in that, The vector construction module includes: A partitioning unit is used to partition the multi-source data according to a dimensional index, and to extract features from the partitioned data according to the sub-indicators under the dimensional index, so as to obtain the basic sample features and feature weights of the corresponding sub-indicators. The fusion unit is used to fuse the weighted sub-features to obtain the single-dimensional initial feature set of the corresponding dimension index based on the feature weights, sample features and the correlation strength between sub-indicators. The representation unit is used to input the initial single-dimensional feature set of all dimensions into the cross-dimensional global attention network and output the enhanced feature representation of each dimension index. The building unit is used to map the enhanced feature representations of each dimension indicator to a unified high-dimensional feature space, and to construct the feature vector of each dimension indicator.

3. The AI-based advertising strategy generation system according to claim 1, characterized in that, Also includes: The frame extraction module is used to extract the text content, canvas content and voice-over content of each frame in the advertisement to be delivered, and obtain the corresponding frame feature vector. At the same time, the core feature vector is obtained according to the core theme of the advertisement to be delivered. The calculation module is used to calculate the first comprehensive distance between each frame feature vector and the core feature vector, as well as the individual element distance between each frame feature vector and the core feature vector, to obtain the analysis array of the corresponding frame. The association acquisition module is used to input the feature vector of each frame into the feature analysis model to obtain the association between elements in the corresponding frame and obtain the association array; The probability determination module is used to determine the display probability of the corresponding frame based on the analysis array and the association array; The continuous segmentation module is used to construct a probability sequence of the advertisement to be delivered based on the display probability of each frame, and to divide the advertisement to be delivered into continuous segments based on the probability sequence. The identifier setting module is used to assign a placeholder playback identifier to the corresponding individual continuous segment according to the output meaning and average probability of each individual continuous segment. The placeholder playback identifier is unique and is related to the playback position and playback timing logic. The condition addition module is used to add trigger conditions to each individual continuous segment of sub-video. When the corresponding trigger condition is met, the sub-video of the individual continuous segment that matches the trigger condition is displayed in original color in a designated small window at the set playback position of the overall canvas, and the sub-video of the individual continuous segment that does not meet the trigger condition is displayed transparently at the set playback position, until the corresponding individual continuous segment finishes playing according to the set playback time, thereby obtaining its own playback strategy.

4. The AI-based advertising strategy generation system according to claim 3, characterized in that, The identifier setting module includes: The multidimensional extraction unit is used to extract the multidimensional set Wi = {Gi, Pi, Ti, Ki, Xi} for each individual continuous segment, where Gi, Pi, Ti, Ki, and Xi are, respectively, the semantic encoding of the output meaning of the i-th individual continuous segment, the normalized value of the average probability, the dependency relationship with the predecessor segment, the triggering relationship with the successor segment, and the spatial requirements based on the required canvas area ratio and window shape. The matching retrieval unit is used to retrieve the tuple {Fui,Pui,Tui} with the highest matching degree from the identifier pool when the corresponding individual continuous segment meets the preset conditions, and automatically associate the end event of the predecessor segment with the start condition of the successor segment. Here, Fui, Pui, and Tui represent the function identifier, playback position rule, and playback timing logic of the output meaning mapping of the i-th individual continuous segment, respectively. The tuple update unit is used to collect interactive feedback test data on the advertisement to be played in real time, and update the priority weight of the playback position of the tuple and the periodic parameters of the timing logic.

5. The AI-based advertising strategy generation system according to claim 1, characterized in that, The strategy determination module includes: The semantic enhancement unit is used to perform multi-scale attention fusion on feature vectors under all dimensions, calculate the importance weights of feature vectors in each dimension, and generate a weighted fused feature vector. The matrix construction unit is used to semantically enhance the weighted fusion feature vector based on the advertising domain knowledge graph, and construct an enhanced feature matrix containing explicit features and implicit semantic associations; The input unit is used to extract the interaction features and temporal features of the playback strategy itself, and input them into the playback strategy temporal model to generate a playback strategy feature tensor. Define a unit, used to define the set of deployment constraint domains D = {D j1,j1＝1,2,3,...,Nk }, where D j1 Let Nk represent the j1-th constraint domain, and let Nk represent the number of constraint domains. For each constraint domain, a bidirectional group {Chj1, Cgj1} is constructed, where Chj1 is the constraint verification identifier of the j1-th constraint domain, and Cgj1 is the policy generation trigger identifier of the j1-th constraint domain. The decomposition unit is used to decompose the target of the advertisement to be played into a set of subtasks ∈={∈j2,j2=1,2,3,..,Nq}, and construct a directed acyclic dependency graph according to the domain dependency relationship, where ∈j2 represents the j2-th subtask; Nq represents the number of subtasks; and the subtasks in the directed acyclic dependency graph ∈ j3 Dependence on D j1 Chj1 must be completed before startup can begin, and ∈ j3 and ∈ j3+1 Dependency edges; Binding unit, used to bind adjacent subtasks in the directed acyclic dependency graph ∈ j3 and ∈ j3+1 Binding to the same constraint domain D j1 Establish a driver execution chain for the bidirectional group {Chj1,Cgj1}: ∈ j3+1 Continuous monitoring of D using an adaptive monitoring window j1 The Chj1 state, when ∈ j3 Once completed, send a message to D. j1 Write to the active state Chj1; When ∈ j3+1 After Chj1 activation is detected, from D j1 Extract Cgj1 from it, trigger its own execution, and output the domain policy parameters; The clustering fusion unit is used to perform clustering analysis on the enhanced feature matrix and the policy feature tensor, and fuse them with the domain policy parameters to generate multiple candidate deployment strategies.

6. The AI-based advertising strategy generation system according to claim 1, characterized in that, The effect evaluation module includes: The model building unit is used to build a response surface model based on historical deployment data and fit the response function f(M) through Gaussian process regression. The factor determination unit is used to map the candidate delivery strategy to the response function f(M) to generate a perturbation sequence. At the same time, it extracts the influencing factors related to user behavior from the candidate delivery strategy, and combines the user group behavior model to characterize the uncertainty of user behavior to determine the probability distribution of the potential interaction path of individual users to the advertisement. The comparison unit is used to perform a first comparison between the perturbation sequence and the first evaluation table, and at the same time, to perform a second comparison between the probability distribution of the potential interaction path and the second evaluation table. The prediction unit is used to predict the delivery effect of the corresponding candidate delivery strategy based on the first control result and the second control result.

7. The method for generating advertising delivery strategies based on artificial intelligence according to claim 6, characterized in that, The effect evaluation module also includes: The comparison unit is used to mine the causal chain of the deployment parameters from the deployment effect, and analyze the causal chain based on the structural causal model to obtain the simulated direct causal and simulated indirect causal of the corresponding deployment parameters on the effect index, and compare them with the ideal direct causal and ideal indirect causal of the set ideal effect. The perturbation definition unit is used to analyze the distribution drift of the delivery data for the corresponding candidate delivery strategy and define the data distribution perturbation set; The re-analysis unit is used to determine the strategy adjustment parameters based on the comparison results and the data distribution disturbance set, and to re-analyze the pre-simulation deployment effect; The filtering unit is used to select the strategy with the best performance from the re-analyzed pre-simulation delivery results as the optimal delivery strategy.

8. The method for generating advertising delivery strategies based on artificial intelligence according to claim 1, characterized in that, The update module includes: The mapping unit is used to map newly acquired multi-source data to the corresponding spatiotemporal scene unit. At the same time, the credibility weight of each newly acquired data source is calculated based on the deep evidence network. The trajectory evolution unit is used to characterize the core parameters of the optimal deployment strategy as quantum superposition vectors and determine the quantum evolution trajectory of each quantum parameter. The comparison and analysis unit is used to compare and analyze the quantum evolution trajectory with the mapping results and the confidence weight, and update and re-deploy the corresponding core parameters.

Images