AI-based cross-platform advertisement material life cycle management and optimization method and system

Abstract

The invention discloses an AI-based cross-platform advertisement material life cycle management and optimization method and system. The method comprises the steps of collecting original data and metadata to construct a heterogeneous data graph; carrying out hierarchical graph convolution, and extracting cross-modal features; carrying out splicing and dimension alignment on the cross-modal features, and carrying out feature dimension reduction to obtain unified features; gene individuals are constructed based on the unified features, and cross-platform materials are generated through a multi-target constraint coevolution algorithm; generating a material pushing strategy according to cross-platform material mapping, optimizing budget allocation and a pushing strategy by constructing a hierarchical game framework, and combining federal constraint synchronization to obtain a pushing strategy set of multi-platform collaborative optimization; and by modeling a material recession rule and calculating a survival function of the material, the life cycle of material recession is warned in real time, and adjustment of a material pushing strategy is triggered. The technical problems of insufficient dynamic strategy collaborative optimization, material life cycle decline prediction lag and the like in cross-platform advertisement material management are solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of artificial intelligence, and particularly relates to a cross-platform advertisement material life cycle management and optimization method and system based on AI. Background Art

[0002] With the digital process of the advertising industry, the management and optimization of advertisement materials have increasingly become a key link for advertisers to improve advertising effects and reduce costs. However, the existing advertisement material life cycle management methods have the following problems: Cross-platform data fragmentation: The placement data of advertisement materials on multiple platforms (such as social media, search engines, short video platforms) cannot be analyzed uniformly, resulting in a lag in strategy adjustment; High manual dependence: The generation of materials depends on the experience of designers, and the placement strategy depends on manual parameter adjustment, with low efficiency and difficulty in scaling up; Insufficient dynamic adaptability: The traditional A / B test has a long cycle, cannot respond to market changes in real time, and lacks an automated optimization mechanism based on user behavior feedback.

[0003] Chinese Patent with Publication No. CN117391776A discloses an advertisement placement method, device, equipment and medium. The method includes: selecting corresponding advertisement materials according to the type of advertisement to be placed, wherein the advertisement materials are stored in an advertisement material library constructed according to a historical advertisement data set; determining the element components and promotional copy of the advertisement to be placed based on the advertisement materials; generating a target advertisement page and a target placement plan corresponding to the advertisement to be placed based on the element components of the advertisement to be placed and the promotional copy of the advertisement to be placed; placing the target advertisement page according to the target placement plan, and recording the placement log of the advertisement to be placed, so as to optimize the target advertisement page and the target placement plan based on the placement log. This invention lacks cross-platform dynamic collaborative optimization, cannot achieve global optimization of the material push strategy, and may lead to an imbalance in cross-platform resource competition; moreover, this invention relies on post-optimization of historical logs, is difficult to trigger material iteration and strategy adjustment in a timely manner, and has a placement lag. Summary of the Invention

[0004] The purpose of the present invention is to provide a cross-platform advertisement material life cycle management and optimization method and system based on AI to solve the technical problems of difficult integration of multi-source heterogeneous data, insufficient dynamic strategy collaborative optimization, and lag in prediction of the decline of the material life cycle in cross-platform advertisement material management.

[0005] The technical solution of the present invention is as follows: On the one hand, the present invention provides a cross-platform advertisement material life cycle management and optimization method based on AI, including the following steps: For the acquisition of raw data and metadata on different platforms, a heterogeneous data graph is constructed based on the raw data and metadata. In the heterogeneous data graph, the node types are defined as user nodes, material nodes, and context nodes, and the edge types are defined as click relationship edges, browsing relationship edges, modal association edges, and user-context edges.

[0006] Based on the heterogeneous data graph structure, perform hierarchical graph convolution of the corresponding relationship types on each node to extract cross-modal features; perform splicing and dimension alignment processing on the cross-modal features of different relationship types to obtain high-dimensional features, and perform feature dimensionality reduction on the high-dimensional features through an attention mechanism to obtain the unified features after dimensionality reduction.

[0007] Construct core semantic items based on the unified features, and construct gene individuals composed of core semantic items, physical constraint items, and creative performance items, and generate cross-platform materials through a multi-objective constraint co-evolution algorithm.

[0008] Generate a material push strategy according to the generated cross-platform materials, optimize the budget allocation and push strategy by constructing a hierarchical game framework including offline Nash equilibrium pre-computation and online No-regret learning, and combine federal constraints synchronization to obtain a set of push strategies for multi-platform collaborative optimization.

[0009] By modeling the material decay law and calculating the survival function of the material, real-time warning of the life cycle of material decay is carried out and the material push strategy is triggered to be adjusted.

[0010] Preferably, based on the heterogeneous data graph structure, perform hierarchical graph convolution of the corresponding relationship types on each node to extract cross-modal features, which is expressed as:

[0011] In the formula, , , is a pre-defined set of relationship types; is the node in the layer of graph convolution features, represents dimensional real vector space; is a non-linear activation function; is the node in the relationship under the neighbor node set; is the relationship in the layer of graph convolution learnable weight matrix, obtained through backpropagation training; is the node in the layer of graph convolution features; , , is the pre-defined number of graph convolution layers.

[0012] Preferably, the feature dimensionality reduction of the high-dimensional features through the attention mechanism is specifically as follows: Global feature extraction is performed on the high-dimensional features of all nodes after hierarchical graph convolution in the heterogeneous data graph, expressed as:

[0013] In the formula, is the global average feature, represents a real vector space of dimension; is the total number of nodes in the heterogeneous data graph; is the node 's high-dimensional feature.

[0014] Calculate the attention weight of each node:

[0015] In the formula, is the attention weight, representing the importance of each target dimension; is the normalization function; is a learnable matrix that maps the concatenated features to dimensions, ; is the concatenation operation.

[0016] Based on the concatenated high-dimensional features and the global average feature, weighted summation is performed on the obtained attention weights to obtain the unified feature after dimensionality reduction.

[0017] Preferably, constructing gene individuals composed of core semantic items, physical constraint items, and creative expression items, and generating cross-platform materials through a multi-objective constrained co-evolution algorithm is specifically as follows: S1: Based on the unified features, construct core semantic items, sample physical constraint items from the multi-platform constraint rule library, and perform probability sampling based on the historical material library and the predefined style probability matrix to construct creative expression items. The gene individuals encoded by the core semantic items, physical constraint items, and creative expression items form the main population for global search, and the TOPP gene individuals are selected from the main population to form the local population for single-platform search.

[0018] S2: In single-generation evolution, calculate the multi-objective fitness function for each gene individual in the main population and the local population, and perform evolution using elitist retention selection operations, cross-recombination operations, and directed mutation operations. Adjust the weights of each objective in the multi-objective fitness function through an adaptive optimization strategy to generate a new generation of the main population and the local population.

[0019] S3: Repeat step S2, and perform cross-population collaboration operations to optimize the local population. Terminate the multi-objective constrained co-evolution algorithm after meeting the stop condition of reaching the maximum number of iterations or convergence of the multi-objective fitness function, and output the set of gene individuals with the highest fitness in the main population and the local population, and generate cross-platform materials.

[0020] Preferably, the calculation of the multi-objective fitness function is expressed as:

[0021] In the formula, is the multi-objective fitness function; is a gene individual; is the parameter vector of the gene individual on platform ; is the total number of platforms; is the semantic standard vector of platform ; is the semantic weight of platform ; is the semantic matching degree; is the satisfaction degree of the th constraint; is the total number of constraints; is the constraint penalty coefficient, which is adaptively adjusted through each generation of evolution; is the creativity diversity score; is the diversity reward coefficient.

[0022] Preferably, according to the generated cross-platform material mapping, generate a material push strategy, optimize the budget allocation and push strategy by constructing a hierarchical game framework including offline Nash equilibrium pre-computation and online No-regret learning, and combine the federal constraint synchronization to obtain a set of push strategies for multi-platform collaborative optimization, specifically: Model each platform as an agent, and the strategies of the agent include: the budget allocation ratio of the platform and the material push strategy generated according to the generated cross-platform material mapping, and determine the initial budget allocation ratio of each platform through offline Nash equilibrium pre-computation using the Lemke-Howson algorithm.

[0023] Combine the real-time ROI and user behavior data, optimize the strategies of each platform through the No-regret learning mechanism, and exchange the budget boundary information of each platform through the federal constraint synchronization to ensure meeting the global budget constraint, and output the optimized set of push strategies. The strategy update rule of the No-regret learning mechanism is:

[0024] In the formula, is the optional strategy adjustment action of platform ; The probability of the platform selecting an action at time ; The probability of the platform selecting an action at time ; Let be all other platforms except the platform ; Let be the joint strategy of other platforms at time Let be the immediate reward of action ; Let be the dynamic learning rate, adjusted according to data freshness; Let be any candidate policy adjustment action; The probability of the platform selecting an action at time ;

[0025] Preferably, define the material decay law modeling as:

[0026] In the formula, Let be the material click intensity of the platform at time Let be the baseline decay rate; Let be the platforms except the platform ; Let be the cross-platform propagation coefficient; Let be the cross-platform influence decay rate; Let be the time; Let be the platform at time Let be the external event type number; Let be the event impact intensity; Let be the event occurrence indicator function.

[0027] Define the survival function of the material as:

[0028] In the formula, Let be the survival function, indicating the probability that the material of the platform has not decayed at time Let be the integration variable.

[0029] Based on the survival function values, a multi-level decline level response mechanism is preset. The decline level of the material is real-time warned according to the calculated survival function values, and the adjustment method corresponding to the material decline level is triggered according to the response mechanism to adjust the material push strategy.

[0030] On the other hand, the present invention provides an AI-based cross-platform advertising material life cycle management and optimization system, including a heterogeneous data graph construction module, a cross-modal feature extraction and dimensionality reduction module, a cross-platform material generation module, a material push strategy generation and optimization module, and a material life cycle warning and push strategy adjustment module.

[0031] The heterogeneous data graph construction module is used to collect raw data and metadata for different platforms, and construct a heterogeneous data graph according to the raw data and metadata. The node types in the heterogeneous data graph are defined as user nodes, material nodes, and context nodes, and the edge types are defined as click relationship edges, tour relationship edges, modal association edges, and user-context edges.

[0032] The cross-modal feature extraction and dimensionality reduction module is used to perform hierarchical graph convolution of corresponding relationship types on each node based on the heterogeneous data graph structure to extract cross-modal features; splice and dimension-align the cross-modal features of different relationship types to obtain high-dimensional features, and perform feature dimensionality reduction on the high-dimensional features through an attention mechanism to obtain unified features after dimensionality reduction.

[0033] The cross-platform material generation module is used to construct core semantic items based on the unified features, and construct gene individuals composed of core semantic items, physical constraint items, and creative expression items, and generate cross-platform materials through a multi-objective constraint co-evolution algorithm.

[0034] The material push strategy generation and optimization module is used to map and generate a material push strategy according to the generated cross-platform materials, optimize the budget allocation and push strategy by constructing a hierarchical game framework including offline Nash equilibrium pre-computation and online No-regret learning, and combine federal constraints synchronization to obtain a set of push strategies for multi-platform collaborative optimization.

[0035] The material life cycle warning and push strategy adjustment module is used to model the material decline law and calculate the survival function of the material, and real-time warn the life cycle of the material decline and trigger the adjustment of the material push strategy.

[0036] On yet another aspect, the present invention further provides an electronic device, which includes: a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, it implements the AI-based cross-platform advertising material life cycle management and optimization method according to any embodiment of the present invention.

[0037] On the other hand, the present invention also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the AI-based cross-platform advertisement material life cycle management and optimization method as described in any embodiment of the present invention.

[0038] Compared with the prior art, the present invention has the following technical effects: 1. By modeling multi-type relationships among users, materials, and context nodes, the present invention integrates cross-platform behaviors, content attributes, and environmental factors to solve the data island problem; and by extracting cross-modal features and dimensionality reduction, it improves the representation ability of complex user-material interaction relationships and supports the generation of precise push strategies.

[0039] 2. The present invention combines core semantics (user preferences), physical constraints (platform rules), and creative expressions (content forms) to design gene individuals, and conducts multi-objective co-evolution to generate differentiated materials that conform to the characteristics of multiple platforms, avoiding creative homogenization caused by single-objective optimization.

[0040] 3. By pre-computing the initial strategy combination through offline Nash equilibrium, the present invention ensures the balance of interests among multiple platforms, and dynamically adjusts the strategy according to real-time ROI and user retention data to adapt to market changes.

[0041] 4. By quantifying the cross-platform communication effect and external event impact, the present invention realizes real-time early warning and precise management of the material life cycle, ensuring that the placement strategy always matches the actual benefits of the material, thereby avoiding ineffective placement and resource waste, and improving the accuracy and effect of the cross-platform material push strategy. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 is the overall flowchart of the AI-based cross-platform advertisement material life cycle management and optimization method described in the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0043] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with specific embodiments of the present application and with reference to the accompanying drawings.

[0044] Embodiment 1 This embodiment provides an AI-based cross-platform advertisement material life cycle management and optimization method. Referring to Figure 1 as shown, it includes the following steps: For the collection of raw data and metadata on different platforms, a heterogeneous data graph is constructed based on the raw data and metadata. In the heterogeneous data graph, the node types are defined as user nodes, material nodes, and context nodes, and the edge types are defined as click relationship edges, browsing relationship edges, modal association edges, and user-context edges. Specifically, the raw data includes but is not limited to: user behaviors (such as the number of user clicks, stay duration, etc.), material content (such as image pixel matrices, video frame sequences, etc.), and context (such as timestamps, geographical locations); the metadata includes but is not limited to: material ID (unique identifier), platform type, user tags (such as age, gender, interests). The user node is a feature vector containing user portrait features, the material node is a feature vector containing material metadata, and the context node is a feature vector containing platform environment features. The click relationship edge is defined as: a user clicks on a material, representing the user's behavior preference; the browsing relationship edge is defined as: a user browses a material, representing the depth of the user's interest. The modal association edge is defined as: a material is associated with video frames, linking the material with content details; the user-context edge is defined as: a user is associated with the context, representing the temporal pattern of the user's behavior.

[0045] Based on the heterogeneous data graph structure, perform hierarchical graph convolution of corresponding relationship types on each node to extract cross-modal features; perform splicing and dimension alignment processing on the cross-modal features of different relationship types to obtain high-dimensional features, and perform feature dimensionality reduction on the high-dimensional features through an attention mechanism to obtain the unified features after dimensionality reduction.

[0046] As a preferred implementation manner of this embodiment, based on the heterogeneous data graph structure, performing hierarchical graph convolution of corresponding relationship types on each node to extract cross-modal features is expressed as:

[0047] In the formula, , , is a predefined set of relationship types; is the feature of node in the -th layer of graph convolution, represents a real vector space of dimension is a non-linear activation function, such as the LeakyReLU function, to enhance the non-linear expression ability; is the set of neighbor nodes of node under the relationship ; is the learnable weight matrix of relationship in the -th layer of graph convolution, randomly initialized and trained through backpropagation; is the feature of node in the The features of layer graph convolution, initially the original features of nodes, are updated layer by layer through graph convolution; , is the predefined number of graph convolution layers. The selection of the number of layers is not restricted here and can be set according to actual needs. For example, if is set, the neighbor nodes are directly learned; if is set, the second-order neighbor nodes can be captured to expand the perception field (such as "the video frames associated with the materials clicked by user A").

[0048] The cross-modal features of different relationship types are concatenated and dimensionally aligned to obtain high-dimensional features. Specifically: for different types of nodes, zero-padding is performed on the relationship types not participated in to align the feature dimensions. For example, for user nodes, dimensionality completion of relationship types is required.

[0049] As a preferred implementation manner of this embodiment, the high-dimensional features are dimensionally reduced through an attention mechanism. Specifically: Global feature extraction is performed on the high-dimensional features of all nodes after hierarchical graph convolution in the heterogeneous data graph, expressed as:

[0050] In the formula, is the global average feature, represents d-dimensional real vector space; is the total number of nodes in the heterogeneous data graph; is the th node's high-dimensional feature.

[0051] Calculate the attention weight of each node:

[0052] In the formula, is the attention weight, representing the importance of each target dimension; is the normalization function, used to normalize the weights into a probability distribution to ensure that the sum of the weights of each dimension is 1; is a learnable matrix that maps the concatenated features to dimensions, randomly initialized and optimized through training, ; is the concatenation operation; is the feature vector of node .

[0053] Based on the concatenated high-dimensional features and the global average feature, the obtained attention weights are weighted and summed to obtain the unified feature after dimensional reduction. Specifically:

[0054] In the formula, is the projection matrix, which maps the 8d-dimensional feature after stitching to dimensions.

[0055] The output unified feature captures the comprehensive semantics of cross-platform user-media-context from feature dimensions such as user behavior including click patterns (such as click frequency and time period distribution of users on different platforms), dwell time (such as the average dwell time of users on materials such as pictures, texts, and videos), cross-platform conversion paths, etc., media content including visual features (such as HSV values of image color matching, dynamic elements of videos), text keywords, multi-modal associations (such as semantic matching degree between pictures, texts and videos), etc., and context features including time context (such as delivery time period, holiday markers), geographical context (such as the adaptability between user geographical location and material region), device context (such as display differences between mobile and PC terminals), etc.

[0056] Based on the unified feature, core semantic items are constructed, and gene individuals composed of core semantic items, physical constraint items, and creative performance items are constructed. Cross-platform materials are generated through a multi-objective constraint co-evolution algorithm.

[0057] As a preferred implementation manner of this embodiment, constructing gene individuals composed of core semantic items, physical constraint items, and creative performance items, and generating cross-platform materials through a multi-objective constraint co-evolution algorithm specifically includes: S1: Based on the unified feature, core semantic items are constructed, physical constraint items are constructed by sampling from a multi-platform constraint rule library, and creative performance items are constructed by probability sampling based on a historical material library and a predefined style probability matrix. Gene individuals encoded by core semantic items, physical constraint items, and creative performance items form the main population for global search, and the TOPP gene individuals are selected from the main population to form a local population for single-platform search (by restricting the search space of the local population, that is, only allowing parameter combinations of the same platform, the platform adaptation accuracy is improved). Specifically, the core semantic items directly inherit the unified feature and contain abstract semantic information such as user behavior and cross-modal relationships. The physical constraint items are encoded by dynamically loading and obtaining technical specification parameters of each platform from a multi-platform rule library. The creative performance items contain encoding of specific creative elements such as copywriting, vision, and sound effects.

[0058] S2: During single-generation evolution, calculate the multi-objective fitness function for each gene individual in the main population and the local population. Use elitist retention selection operation, crossover recombination operation, and directional mutation operation for evolution. Adjust the weights of each objective in the multi-objective fitness function through an adaptive optimization strategy to generate a new generation of the main population and the local population. Further, the elitist retention operation is specifically as follows: Retain the individuals with the highest fitness (such as setting 20% of the individuals) directly into the next generation. The crossover recombination operation is specifically as follows: The main population uses a high crossover rate to mix different platform parameters, and the local population focuses on optimizing the same platform parameters. The directional mutation operation is specifically as follows: Adjust the mutation direction according to real-time feedback. Further, the multi-constraint co-evolution algorithm also performs real-time data fusion, including: Receiving real-time data streams such as user click heatmaps and platform rule changes by collecting user behaviors to achieve semantic weight update, gene mutation direction adjustment, and dynamic update of the constraint library.

[0059] S3: Repeat step S2 and perform cross-population cooperation to optimize the local population. Terminate the multi-objective constraint co-evolution algorithm after meeting the stop conditions of reaching the maximum number of iterations or convergence of the multi-objective fitness function. Output the set of gene individuals with the highest fitness in the main population and the local population to generate cross-platform materials. Specifically, the cross-population cooperation operation is as follows: Perform elitist immigration every X generations (such as setting it to 5 generations), and inject the best gene individuals in the main population (such as setting it to Top 5%) into the local population to replace the same number of the worst gene individuals in the local population. The output of the main population is the set of gene individuals with the highest global fitness (such as setting it to Top 10%), covering different platform combination schemes. The output of the local population is the set of the best candidate gene individuals retained separately for each platform (such as setting it to Top 5).

[0060] As a preferred implementation manner of this embodiment, the multi-objective fitness function is evaluated through semantic matching degree, platform constraint satisfaction degree (checking whether parameters such as resolution and duration meet the requirements of each platform), and creative diversity score (comparing the difference degree between the current gene and historical materials), and is expressed as:

[0061] In the formula, is the multi-objective fitness function; is the gene individual; is the parameter vector of the gene individual on platform ; is the total number of platforms; is the semantic standard vector of platform , which is the unified feature obtained by dimensionality reduction; is the semantic weight of platform ; is the semantic matching degree, generally calculated through cosine similarity; is the satisfaction degree of the th constraint; is the total number of constraints; is the constraint penalty coefficient, which is adjusted by the adaptive optimization strategy of each generation of evolution; is the creativity diversity score; is the diversity reward coefficient.

[0062] The semantic weight is specifically calculated as:

[0063] The constraint penalty coefficient is adjusted by the adaptive optimization strategy of each generation of evolution:

[0064] In the formula, is the th generation constraint penalty coefficient; is th generation constraint penalty coefficient; is the total number of gene individuals in the current main / local population.

[0065] The satisfaction degree is specifically calculated as:

[0066] In the formula, is the parameter deviation; is the hard limit threshold of the material parameters for each platform (such as the duration constraint threshold, resolution constraint threshold).

[0067] Furthermore, the parameter deviation is specifically calculated as:

[0068] In the formula, is the total number of parameters of the gene individual, ; is the importance weight of the parameter which can preset the basic weight in the platform rule library and dynamically fine-tune according to the real-time click data of users; is the th parameter in the gene individual; is the platform for the parameter hard limit threshold.

[0069] The diversity reward coefficient is specifically calculated as:

[0070] In the formula, is the initial value of the diversity reward coefficient; the repeated batches are counted as 1 repeated batch when the creative similarity of consecutive preset generations detected is greater than the set percentage.

[0071] The creative diversity score is specifically calculated as:

[0072] According to the generated cross-platform material mapping, generate a material push strategy. Optimize the budget allocation and push strategy by constructing a hierarchical game framework including offline Nash equilibrium pre-computation and online No-regret learning, and combine the federated constraint synchronization to obtain a set of push strategies for multi-platform collaborative optimization.

[0073] As a preferred implementation manner of this embodiment, this step is specifically: Model each platform as an agent, and the strategies of the agent include: the platform 's budget allocation ratio and generate a material push strategy according to the generated cross-platform material mapping, and determine the initial budget allocation ratio of each platform through offline Nash equilibrium pre-computation by the Lemke-Howson algorithm.

[0074] Furthermore, determining the initial budget allocation ratio of each platform through offline Nash equilibrium pre-computation by the Lemke-Howson algorithm is specifically: Define the global optimization goal as the revenue function :

[0075] In the formula, is the total number of platforms, ; is the ROI revenue of platform , is the fairness adjustment factor, which is used to balance the total revenue and the fairness among platforms; the Gini coefficient is used to measure the inequality of platform revenues.

[0076] Discretize the initial continuous budget allocation ratio in the agent strategy into a finite action set, calculate the corresponding revenue function for each strategy combination to construct a payoff matrix, and the elements of the payoff matrix are the revenue function values of the corresponding strategy combinations.

[0077] Through complementary pivot iteration, find the equilibrium point that satisfies , that is, the marginal utility of each platform's strategy is zero, and the overall revenue value cannot be improved by slightly adjusting the budget allocation, and output the initial strategy combination of each platform.

[0078] Combine real-time ROI with user behavior data, optimize the strategies of each platform through the No-regret learning mechanism, and synchronize and exchange the budget boundary information of each platform through federal constraints to ensure compliance with the global budget constraint, and output an optimized set of push strategies. The strategy update rule of the No-regret learning mechanism is as follows:

[0079] In the formula, is the platform Optional strategy adjustment actions (including budget allocation adjustment and material push strategy selection); is the platform At The probability of selecting action at time; is the platform At The probability of selecting action at time; is all other platforms except platform ; is the joint strategy of other platforms at time; is the immediate reward of action ; is the dynamic learning rate, adjusted according to data freshness; is any candidate strategy adjustment action; is the platform At The probability of selecting action at time.

[0080] Furthermore, the No-regret learning mechanism can be updated in a time series (updated at fixed time intervals) or an event (triggered when the ROI fluctuation exceeds the threshold).

[0081] The implementation process of the federal constraint synchronization is specifically as follows: Each platform encrypts the local budget ratio ; Collaboratively decrypt , verify whether it meets the global budget constraint; if it exceeds the limit, compress the budgets of each platform proportionally to ensure .

[0082] As a preferred implementation manner of this embodiment, it is also ensured through revenue improvement verification that the adjustment of the material push strategy conforms to the overall interests of the agent system composed of each platform. The revenue improvement verification is specifically as follows: Calculate the revenue changes of each platform:

[0083] In the formula, is the platform The revenue change for the platform after changing to the strategy is the revenue; for the platform when implementing the strategy is the revenue.

[0084] If there exists and for other platforms , then accept the strategy change.

[0085] By modeling the decay law of the material and calculating the survival function of the material, the life cycle of the material decay is warned in real time and the material push strategy is triggered to be adjusted. As a preferred implementation mode of this embodiment, the above steps are specifically as follows: Define the material decay law modeling as:

[0086] In the formula, is the material click intensity of the platform at time, which can be obtained by fitting the historical click feature in the unified feature after dimensionality reduction; is the baseline decay rate, which is fitted by the click decay curve without external intervention; is the platform except the platform ; is the cross-platform propagation coefficient; is the cross-platform influence decay rate; is the time; is the platform at the event increment at time; is the external event type number; is the event impact intensity, which is verified by the difference-in-differences method, with a positive value indicating an increase and a negative value indicating a suppression; is the event occurrence indicator function, which is 1 when the event takes effect and 0 when it does not take effect.

[0087] Furthermore, the cross-platform propagation coefficient and the cross-platform influence decay rate are estimated and calculated by the EM algorithm: Assume the current parameters and calculate the number of potential cross-platform influence events:

[0088] In the formula, is the number of potential cross-platform influence events, indicating the platform at time on the platform at Click contribution ratio at a certain moment.

[0089] Update the cross-platform propagation coefficient and the cross-platform influence decay rate to maximize the log-likelihood function until the parameters converge, and obtain and .

[0090] Define the survival function of the material as:

[0091] In the formula, is the survival function, indicating the probability that the material on platform has not decayed at moment; is the integration variable, representing a temporary variable in the integration process.

[0092] Based on the survival function value, preset a multi-level decay level response mechanism, real-time warning of the decay level of the material according to the calculated survival function value, and trigger the adjustment method corresponding to the material decay level according to the response mechanism to guide the dynamic adjustment of the budget allocation and the material push strategy in the previous step, avoiding ineffective delivery. The total number of level divisions is not limited here and can be divided according to actual needs. The following is a simple example of strategy adjustment: when the survival function value of a certain material reaches the first-level decay level, reduce the delivery frequency of this material, dynamically adjust the initial budget allocation ratio of each platform, and at the same time generate an iterative instruction such as "replace the promotion slogan with 'Limited-time discount'".

[0093] Embodiment 2 Correspondingly, this embodiment provides an AI-based cross-platform advertising material life cycle management and optimization system, which is used to implement the AI-based cross-platform advertising material life cycle management and optimization method as described in Embodiment 1 of the present invention, including a heterogeneous data graph construction module, a cross-modal feature extraction and dimensionality reduction module, a cross-platform material generation module, a material push strategy generation and optimization module, and a material life cycle warning and push strategy adjustment module.

[0094] The heterogeneous data graph construction module is used to collect raw data and metadata for different platforms, and construct a heterogeneous data graph according to the raw data and metadata. The node types in the heterogeneous data graph are defined as user nodes, material nodes, and context nodes, and the edge types are defined as click relationship edges, tour relationship edges, modal association edges, and user-context edges.

[0095] The cross-modal feature extraction and dimensionality reduction module is used to perform hierarchical graph convolution of corresponding relationship types on each node based on the heterogeneous data graph structure to extract cross-modal features; splice and dimension-align the cross-modal features of different relationship types to obtain high-dimensional features, and perform feature dimensionality reduction on the high-dimensional features through an attention mechanism to obtain unified features after dimensionality reduction.

[0096] A cross-platform material generation module, which is used to construct core semantic items based on unified features, and construct gene individuals composed of core semantic items, physical constraint items, and creative performance items, and generate cross-platform materials through a multi-objective constrained co-evolution algorithm.

[0097] A material push strategy generation and optimization module, which is used to generate a material push strategy according to the generated cross-platform materials, optimize the budget allocation and push strategy by constructing a hierarchical game framework including offline Nash equilibrium pre-computation and online No-regret learning, and combine with federated constraint synchronization to obtain a set of push strategies for multi-platform collaborative optimization.

[0098] Embodiment 3 This embodiment provides an electronic device, which includes: a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, it implements the AI-based cross-platform advertising material life cycle management and optimization method as described in Embodiment 1 of the present invention.

[0099] Embodiment 4 This embodiment provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the AI-based cross-platform advertising material life cycle management and optimization method as described in Embodiment 1 of the present invention.

[0100] In the embodiments of the present application, "at least one" means one or more, and "a plurality" means two or more. "And / or" describes the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B may represent the situation where A exists alone, A and B exist simultaneously, or B exists alone. Where A and B may be singular or plural. The character " / " generally represents an "or" relationship between the associated objects before and after. "At least one of the following" and its similar expressions refer to any combination of these items, including any combination of single items or plural items. For example, at least one of a, b, and c may represent: a, b, c, a and b, a and c, b and c, or a and b and c, where a, b, and c may be single or multiple.

[0101] Those of ordinary skill in the art can realize that the units and algorithm steps described in the embodiments disclosed herein can be implemented by a combination of electronic hardware, computer software, and electronic hardware. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. Professional technicians can use different methods to implement the described functions for each specific application, but this implementation should not be considered to exceed the scope of the present application.

[0102] Those skilled in the art can clearly understand that for the convenience and conciseness of description, the specific working processes of the systems, devices, and units described above can refer to the corresponding processes in the foregoing method embodiments and will not be elaborated herein.

[0103] In several embodiments provided by the present application, if any function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present application. The foregoing storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memories (hereinafter referred to as ROMs), random access memories (hereinafter referred to as RAMs), magnetic disks, or optical discs that can store program codes.

[0104] The above are only the embodiments of the present invention, and do not limit the patent scope of the present invention accordingly. Any equivalent structure or equivalent process transformation made by using the content of the specification and drawings of the present invention, or directly or indirectly applied in other related technical fields, shall be equally included in the patent protection scope of the present invention.

Claims

1. A cross-platform advertising material lifecycle management and optimization method based on AI, characterized in that: The following steps are involved: Collect raw data and metadata from different platforms, and construct a heterogeneous data graph based on the raw data and metadata. The node types in the heterogeneous data graph are defined as user nodes, material nodes, and context nodes, and the edge types are defined as click relationship edges, browsing relationship edges, modality association edges, and user-context edges; Based on the heterogeneous data graph structure, layered graph convolution of corresponding relationship types is performed on each node to extract cross-modal features; The cross-modal features of different relationship types are concatenated and dimensionally aligned to obtain high-dimensional features. The high-dimensional features are then dimensionally reduced through the attention mechanism to obtain unified features after dimension reduction. Construct core semantic items based on unified features, and construct gene individuals composed of core semantic items, physical constraint items, and creative expression items, and generate cross-platform materials through a multi-objective constraint co-evolution algorithm; Generate a material push strategy based on the generated cross-platform material mapping, optimize budget allocation and push strategy by building a hierarchical game framework including offline Nash equilibrium pre-calculation and online No-regret learning, and obtain a set of push strategies for multi-platform collaborative optimization by combining federated constraint synchronization; By modeling the material decay rules and calculating the material survival function, real-time warnings of material decay life cycle can be provided and adjustments to the material push strategy can be triggered.

2. The AI-based cross-platform advertising material lifecycle management and optimization method according to claim 1, characterized in that: Based on the heterogeneous data graph structure, a layered graph convolution of the corresponding relationship type is performed on each node to extract cross-modal features represented as: In the formula, , , is a set of predefined relationship types; For Node In the The features of layer graph convolution, express dimensional real vector space; is a nonlinear activation function; For Node In relationship The set of neighbor nodes under For relationship In the The learnable weight matrix of the layer graph convolution is obtained through back-propagation training; For Node In the Features of layer graph convolution; , is the predefined number of graph convolution layers.

3. The AI-based cross-platform advertising material lifecycle management and optimization method according to claim 1, characterized in that: The feature dimensionality reduction of high-dimensional features through the attention mechanism is as follows: The global feature extraction of high-dimensional features of all nodes after layered graph convolution in the heterogeneous data graph is expressed as: In the formula, is the global average feature, express dimensional real vector space; is the total number of nodes in the heterogeneous data graph; For Node High-dimensional features; Calculate the attention weight of each node: In the formula, is the attention weight, indicating the importance of each target dimension; is the normalization function; is a learnable matrix, mapping the concatenated features to dimension, ; For splicing operation; Based on the concatenated high-dimensional features and the global average features, the obtained attention weights are weighted summed to obtain the unified features after dimensionality reduction.

4. The AI-based cross-platform advertising material lifecycle management and optimization method according to claim 1, characterized in that: Construct a gene individual consisting of core semantic items, physical constraint items, and creative expression items, and generate cross-platform materials through a multi-objective constraint co-evolution algorithm. Specifically: S1: Construct core semantic items based on unified features, construct physical constraint items by sampling from multi-platform constraint rule libraries, and construct creative expression items by probability sampling based on historical material libraries and predefined style probability matrices. Gene individuals are generated by encoding core semantic items, physical constraint items, and creative expression items to form a main population for global search. Top P gene individuals are selected from the main population to form a local population for single-platform search. S2: In single-generation evolution, the multi-objective fitness function is calculated for each gene individual in the main population and the local population, and the evolution is carried out using elite retention selection operations, crossover recombination operations, and directed mutation operations. The weights of each objective in the multi-objective fitness function are adjusted through an adaptive optimization strategy to generate a new generation of the main population and local population; S3: Repeat step S2 and perform cross-population collaborative operations to optimize the local population. After the stopping conditions of reaching the maximum number of iterations or the convergence of the multi-objective fitness function are met, the multi-objective constrained collaborative evolution algorithm is terminated, and the gene individual set with the highest fitness of the main population and the local population is output to generate cross-platform materials.

5. The AI-based cross-platform advertising material lifecycle management and optimization method according to claim 4, characterized in that: The multi-objective fitness function calculation is expressed as: In the formula, is the multi-objective fitness function; For genetic individuals; For genetic individuals on the platform The parameter vector of is the total number of platforms; For the platform The semantic standard vector of For the platform The semantic weight of is the semantic matching degree; For the The satisfaction of the item constraints; is the total number of constraints; To constrain the penalty coefficient, it is adaptively adjusted through each generation of evolution; score creative diversity; is the diversity bonus coefficient.

6. The AI-based cross-platform advertising material lifecycle management and optimization method according to claim 1, characterized in that: The material push strategy is generated based on the generated cross-platform material mapping. The budget allocation and push strategy are optimized by building a hierarchical game framework including offline Nash equilibrium pre-calculation and online No-regret learning. Combined with the federated constraint synchronization, the push strategy set for multi-platform collaborative optimization is obtained as follows: Each platform is modeled as an intelligent agent, and the strategy of the intelligent agent includes: budget allocation ratio of the platform and material push strategy generated according to the generated cross-platform material mapping, and the initial budget allocation ratio of each platform is determined by offline Nash equilibrium pre-calculation through Lemke-Howson algorithm; Combining real-time ROI and user behavior data, the No-regret learning mechanism is used to optimize the strategies of each platform, and the budget boundary information of each platform is synchronously exchanged through federated constraints to ensure that the global budget constraints are met and output the optimized push strategy set. The strategy update rules of the No-regret learning mechanism are as follows: In the formula, For the platform Optional policy adjustment actions; For the platform exist Always choose your action probability; For the platform exist Always choose your action probability; To remove the platform All other platforms except For other platforms Joint strategies at all times; For Action immediate benefits; It is a dynamic learning rate, which is adjusted according to the freshness of the data; Adjust actions for any candidate strategy; For the platform exist Always choose your action probability.

7. The AI-based cross-platform advertising material lifecycle management and optimization method according to claim 1, characterized in that: The material decay law is defined as: In the formula, For the platform exist The intensity of the material click at the moment; is the baseline decay rate; To remove the platform External platforms; is the cross-platform propagation coefficient; Affects the decay rate across platforms; for the moment; For the platform exist The event increment at the moment; Number the external event type; For events Impact strength; For events The indicator function that occurs; The survival function of the material is defined as: In the formula, is the survival function, indicating the platform The material is The probability that the moment has not decayed; is the integration variable; A multi-level decay level response mechanism is preset based on the survival function value. The decay level of the material is warned in real time according to the calculated survival function value, and the material push strategy is adjusted according to the adjustment method corresponding to the material decay level triggered by the response mechanism.

8. An AI-based cross-platform advertising material lifecycle management and optimization system, characterized in that: The system is used to implement the AI-based cross-platform advertising material lifecycle management and optimization method according to any one of claims 1 to 7, including a heterogeneous data graph construction module, a cross-modal feature extraction and dimensionality reduction module, a cross-platform material generation module, a material push strategy generation and optimization module, and a material lifecycle warning and push strategy adjustment module; A heterogeneous data graph construction module is used to collect raw data and metadata from different platforms and construct a heterogeneous data graph based on the raw data and metadata. The node types in the heterogeneous data graph are defined as user nodes, material nodes, and context nodes, and the edge types are defined as click relationship edges, browsing relationship edges, modality association edges, and user-context edges. The cross-modal feature extraction and dimensionality reduction module is used to extract cross-modal features by performing hierarchical graph convolution on each node based on the heterogeneous data graph structure and corresponding relationship type; The cross-modal features of different relationship types are concatenated and dimensionally aligned to obtain high-dimensional features. The high-dimensional features are then dimensionally reduced through the attention mechanism to obtain unified features after dimension reduction. The cross-platform material generation module is used to construct core semantic items based on unified features, and to construct gene individuals composed of core semantic items, physical constraint items, and creative expression items, and to generate cross-platform materials through a multi-objective constraint co-evolution algorithm; The material push strategy generation and optimization module is used to generate material push strategies based on the generated cross-platform material mapping, optimize budget allocation and push strategies by building a hierarchical game framework including offline Nash equilibrium pre-calculation and online No-regret learning, and combine federal constraint synchronization to obtain a set of push strategies for multi-platform collaborative optimization; The material life cycle warning and push strategy adjustment module is used to provide real-time warning of material decay life cycle and trigger material push strategy adjustment by modeling material decay rules and calculating material survival functions.

9. An electronic device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the AI-based cross-platform advertising material lifecycle management and optimization method as described in any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the AI-based cross-platform advertising material lifecycle management and optimization method according to any one of claims 1 to 7 is implemented.

Images