Cross-platform advertisement automatic putting system

Abstract

The invention discloses an automatic cross-platform advertisement putting system and method, and belongs to the technical field of advertisement putting, and the system comprises a multi-modal content analysis module which is used for carrying out the cross-modal theme aggregation of the graph, text and video contents of a social platform, and generating a unified content vector; the correlation scoring module is used for calculating correlation scores of the content vectors and the commodities in the advertisement library; the dynamic delivery strategy module executes a hierarchical strategy according to the score: high-time-sharing accurate delivery, middle-time-sharing associated extension delivery, and low-time-sharing exploratory delivery decision algorithm is started to perform opportunity exploration; and the strategy self-adaptive optimization module is used for automatically adjusting various parameters in the putting strategy by periodically backtesting the actual putting return efficiency of the advertisement. Through combination of deep content awareness and strategy closed-loop optimization, the accuracy and the return rate of advertisement putting and the utilization efficiency of advertisement inventory are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of advertising delivery, and in particular to a cross-platform automated advertising delivery system. Background Technology

[0002] As social media becomes the primary medium for information dissemination, online advertising systems strive to connect content with monetization. However, existing systems suffer from significant shortcomings in the depth and breadth of content understanding. Most rely on superficial analysis of single modalities, particularly text keywords, failing to achieve a comprehensive, context-aware understanding of increasingly complex combinations of text, images, and videos. This directly leads to bottlenecks in ad matching accuracy. More critically, existing ad delivery strategies are often rigid and pre-defined, operating in an open-loop model. They lack the ability to optimize in a closed loop based on actual return on investment and to learn independently, resulting in poor system adaptability and an inability to maximize the value of the advertising budget. This deficiency is particularly pronounced in weak-match scenarios. When the system cannot find strongly relevant ads, it typically delivers inefficient generic ads or abandons the system altogether, leading to a significant waste of ad inventory and exposing a widespread lack of intelligent exploration mechanisms. Therefore, there is an urgent need in this field for an intelligent ad delivery solution capable of achieving deep multimodal understanding and adaptive strategy optimization. Summary of the Invention

[0003] To address the aforementioned problems in existing technologies, the present invention aims to provide a cross-platform automated advertising delivery system, comprising: A multimodal content acquisition module is used to acquire content data containing at least one modality, including text, images, or videos, from at least one social platform.

[0004] The multimodal content analysis module, connected to the multimodal content acquisition module, is used to perform cross-modal topic aggregation on multimodal content data to generate a unified content vector that represents the core topic of the content data.

[0005] The relevance scoring module is used to calculate the relevance score between a unified content vector and multiple product vectors pre-stored in the ad library.

[0006] The dynamic delivery strategy module is used to select an advertising product to be delivered from the ad library based on the relevant product score and a tiered delivery strategy that includes multiple trigger conditions.

[0007] The advertising performance backtesting module is used to track the return on investment efficiency of the ads that have been placed and to perform backtesting analysis on the tiered advertising strategy based on historical data.

[0008] The strategy adaptive optimization module is used to automatically adjust the triggering conditions in the hierarchical delivery strategy based on the results of the backtesting analysis.

[0009] Furthermore, the multimodal content analysis module includes a multimodal saliency fusion network, which further includes: a set of modality-specific encoders for converting data of the corresponding modality into initial feature vectors; a cross-modal attention module for receiving the initial feature vectors and calculating a dynamic saliency weight for each modality by evaluating the interrelationship between features of each modality; and a fusion layer for weighted aggregation of the initial feature vectors based on the dynamic saliency weights to generate the final unified content vector.

[0010] Furthermore, the relevance scoring module generates the final relevance product score through a two-stage process, which includes: a semantic matching stage, used to obtain a basic content-advertisement similarity score by calculating the cosine similarity between the unified content vector and the product vector; and a business value adjustment stage, used to adjust the basic content-advertisement similarity score by utilizing business value factors associated with the advertised product, wherein the business value factors depend on at least one of the following: the historical performance of the advertised product, profit margin, and inventory status.

[0011] Furthermore, the tiered delivery strategy is based on a comparison between the highest relevant product score and two preset thresholds, namely the first preset threshold and the second preset threshold. Specifically, it includes: a high-score precision delivery strategy triggered when the highest score is higher than the first preset threshold; an association expansion delivery strategy triggered when the highest score is not higher than the first preset threshold but higher than the second preset threshold; and an opportunity exploration delivery strategy triggered when the highest score is not higher than the second preset threshold.

[0012] Furthermore, the opportunity exploration delivery strategy employs a reward-driven exploratory delivery decision algorithm, which determines the delivery of each ad in the exploratory ad pool. Calculate a potential return score And select the highest-scoring ads for display, among which The calculation formula is: ,in It is the historical average return on investment for advertising. This represents the total number of opportunity exploration deployments that have been executed. It's an advertisement. Number of times it has been deployed It is a preset constant used to balance the exploratory and exploitative aspects of deployment.

[0013] Furthermore, the system executes a cross-platform advertising delivery method, including the following steps: Step S1: Collect multimodal content data from the designated social media platform according to the set information; Step S2: Perform a multimodal content parsing and aggregation method on the multimodal content data to generate a content vector that can uniformly represent its core theme; Step S3: Calculate the similarity between the content vector and multiple product vectors in the advertising library to obtain a set of related product scores; Step S4: Based on the relevant product scores, compare them with preset thresholds to trigger corresponding tiered delivery strategies to determine the final ads to be delivered; Step S5: Track the return on investment (ROI) of the deployed ads, and periodically backtest and adaptively optimize the preset thresholds of the tiered deployment strategy based on the ROI data.

[0014] Furthermore, the multimodal content parsing and aggregation method described in step S2 is a context-aware multimodal fusion process. The multimodal fusion process further includes performing cross-modal association analysis on the initial feature vectors extracted from each modality to determine the relative importance of each modality in expressing the core theme in the specific context of the current content data, thereby generating a dynamic saliency weight for each modality; and performing weighted aggregation on the initial feature vectors based on the dynamic saliency weight to generate the final content vector.

[0015] Furthermore, the triggering of the tiered delivery strategy in step S4 includes: if the highest relevant product score is greater than a first threshold, then selecting a specific product advertisement corresponding to the highest score for delivery; if the highest score is not greater than the first threshold but greater than a second threshold, then selecting a category-level advertisement or alternative product advertisement related to the content data theme from a pool of alternative advertisements for delivery; if the highest score is not greater than the second threshold, then initiating an opportunity exploration process to select an advertisement from a dedicated exploration advertisement pool for delivery.

[0016] Furthermore, the backtesting and adaptive optimization described in step S5 further includes: defining a strategy parameter vector that includes the first threshold, the second threshold, and the exploration constant in the opportunity exploration delivery algorithm; periodically simulating multiple different sets of the strategy parameter vectors on a historical dataset to evaluate the simulated total return on delivery corresponding to each set of vectors; and automatically updating the strategy parameter vector currently used by the online system to the parameter vector that produces the highest total return on delivery in the simulation.

[0017] Furthermore, in executing the opportunity exploration process, this step further includes a complete learning and updating loop: calculating a potential return score for each ad in the exploration ad pool that combines its known historical average return with an exploration reward item; selecting and delivering the ad with the highest potential return score; and after delivery, capturing the actual return data of this delivery and using this data to update the historical average return of the ad and the statistical record of the number of times it has been delivered for future score calculations.

[0018] Compared to existing technologies, the advantages of this invention are as follows: By constructing a complete closed-loop system from multimodal content deep perception to strategy backtesting and optimization, this invention significantly improves the accuracy and return on investment (ROI) of ad placement. The system surpasses traditional shallow matching based on keywords or single modalities. Through cross-modal topic aggregation of various information such as text, images, and videos, it generates a precise profile of the core intent of the content, thereby achieving a high degree of relevance between ads and content. More importantly, this invention introduces an adaptive optimization mechanism based on actual placement ROI efficiency, enabling the placement strategy to continuously learn and evolve, dynamically adjusting matching thresholds and algorithm parameters to ensure that the advertising budget is always utilized in a near-optimal manner, thus fundamentally solving the problems of rigid strategies and inability to self-iterate in existing technologies.

[0019] This invention significantly improves the utilization efficiency and value mining capabilities of ad inventory by designing a hierarchical dynamic delivery strategy, especially in scenarios with weak or no matches. When a high-scoring match cannot be found, the system does not simply abandon or deliver inefficient ads, but intelligently initiates association expansion or opportunity exploration strategies. In particular, the introduced reward-driven exploratory delivery decision algorithm can conduct delivery tests in a scientific manner that balances exploration and utilization under uncertain conditions, continuously discovering new high-value "content-ad" combinations. This transforms potentially wasted ad slots into opportunities for data collection and value discovery, significantly enhancing the robustness and long-term profitability of the entire advertising system. Attached Figure Description

[0020] Figure 1 This is a schematic block diagram illustrating the system module configuration of the present invention.

[0021] Figure 2 This is an exemplary step in the advertising delivery method of the present invention. Detailed Implementation

[0022] The present invention will be further described below with reference to specific embodiments.

[0023] like Figure 1 The diagram shown is a schematic representation of the module structure of a cross-platform automated advertising delivery system provided in this embodiment, including: A multimodal content acquisition module is used to acquire content data containing at least one modality, including text, images, or videos, from at least one social platform.

[0024] The multimodal content analysis module, connected to the multimodal content acquisition module, is used to perform cross-modal topic aggregation on multimodal content data to generate a unified content vector that represents the core topic of the content data.

[0025] The relevance scoring module is used to calculate the relevance score between a unified content vector and multiple product vectors pre-stored in the ad library.

[0026] The dynamic delivery strategy module is used to select an advertising product to be delivered from the ad library based on the relevant product score and a tiered delivery strategy that includes multiple trigger conditions.

[0027] The advertising performance backtesting module is used to track the return on investment (ROI) of ads that have been placed and to perform backtesting analysis on tiered advertising strategies based on historical data.

[0028] The strategy adaptive optimization module is used to automatically adjust the triggering conditions in the tiered delivery strategy based on the results of backtesting analysis.

[0029] In one embodiment, the system described in claim 1 is illustrated. This cross-platform automated advertising delivery system can be deployed on a cloud server cluster, with each module implemented as an independent microservice communicating via an application programming interface (API). The multimodal content acquisition module can be configured as a set of API clients that pull data from platforms such as WeChat, Weibo, and Douyin in real time through officially authorized interfaces, along with a supplementary web crawler, responsible for cleaning and standardizing the collected heterogeneous data. The multimodal content analysis engine, as the system's computational core, can be deployed on a server equipped with a graphics processing unit (GPU) to perform intensive neural network calculations, performing cross-modal topic aggregation on the content data to generate a unified content vector in high-dimensional floating-point form. The relevance scoring module can be implemented as a high-speed query service, internally employing an efficient vector retrieval engine such as FAISS or Milvus to calculate the relevance score between the unified content vector and the product vector. The dynamic delivery strategy module can be considered a lightweight decision engine, selecting advertising products to be delivered based on the relevance score and a hierarchical delivery strategy. The advertising performance backtesting module and the strategy adaptive optimization module are usually run as background batch processing tasks to track the return on investment efficiency and periodically backtest the tiered delivery strategy. Finally, based on the results of the backtesting analysis, the triggering conditions in the strategy are automatically adjusted to form a complete and automated optimization loop.

[0030] The multimodal content analysis module includes a multimodal saliency fusion network, which further includes: a set of modality-specific encoders for transforming data of the corresponding modality into initial feature vectors; a cross-modal attention module for receiving the initial feature vectors and calculating a dynamic saliency weight for each modality by evaluating the interrelationship between features of each modality; and a fusion layer for weighted aggregation of the initial feature vectors based on the dynamic saliency weights to generate the final unified content vector.

[0031] In one embodiment, the multimodal saliency fusion network of claim 2 is described. The network includes a set of modality-specific encoders. For example, for the text modality, the encoder may employ a pre-trained bidirectional language model based on a Transformer architecture; for the image modality, a visual Transformer or a deep convolutional neural network may be used; and for the video modality, visual features from its keyframe image sequence and semantic features from the audio-transcribed text may be combined, each used to transform the data of the corresponding modality into an initial feature vector. The network also includes a cross-modal attention module for receiving the initial feature vector and calculating a dynamic saliency weight for each modality by evaluating the interrelationships between the features of each modality. Finally, the network includes a fusion layer for weighted aggregation of the initial feature vector based on the dynamic saliency weight, for example, performing a weighted summation operation to generate the final unified content vector.

[0032] The relevance scoring module generates the final relevance product score through a two-stage process. The process includes a semantic matching stage, which calculates a basic content-ad similarity score by measuring the cosine similarity between the unified content vector and the product vector; and a business value adjustment stage, which adjusts the basic content-ad similarity score by using business value factors associated with the advertised product. These business value factors depend on at least one of the following: the historical performance of the advertised product, its profit margin, or its inventory status.

[0033] The tiered targeting strategy is based on a comparison between the highest relevant product score and two preset thresholds, namely the first preset threshold and the second preset threshold. Specifically, it includes: a high-score precision targeting strategy triggered when the highest score is higher than the first preset threshold; an association expansion targeting strategy triggered when the highest score is not higher than the first preset threshold but higher than the second preset threshold; and an opportunity exploration targeting strategy triggered when the highest score is not higher than the second preset threshold.

[0034] The opportunity exploration delivery strategy employs a reward-driven exploratory delivery decision algorithm, which evaluates each ad in the exploratory ad pool. Calculate a potential return score And select the highest-scoring ads for display, among which The calculation formula is: ,in It's an advertisement. Historical average return efficiency, This represents the total number of opportunity exploration deployments that have been executed. It represents the number of times the ad has been displayed. It is a preset constant used to balance the exploratory and exploitative aspects of deployment.

[0035] In one embodiment, the reward-driven exploratory delivery decision algorithm of claim 5 is described. This algorithm applies a decision to each ad in the exploratory ad pool. Calculate a potential return score The calculation strictly follows the formula, whereby the system reads the historical average ROI of ad j from the database in real time as... Read the number of times it has been delivered as And obtain the total number of explorations globally. Exploring constants This setting can be preset by the system administrator, for example, set to 2.0. The algorithm then selects the ad with the highest score for delivery and updates the ad's performance using the actual return data from this delivery. and Statistical values ​​form a complete learning and updating cycle.

[0036] like Figure 2 As shown, the system in this embodiment executes a cross-platform advertising delivery method, which includes the following steps: Step S1: Collect multimodal content data from designated social platforms according to the set information. This step can be performed in near real-time by calling the official APIs of each social platform, supplemented by web crawlers. In addition to the content itself, the collected data can also include metadata such as publication time and user interaction data. After collection, the data is cleaned and formatted in a unified manner.

[0037] Step S2 involves performing multimodal content parsing and aggregation on the multimodal content data to generate a content vector that can uniformly represent its core theme. This step is performed on a server equipped with a GPU. By performing semantic analysis on the text, visual element recognition on the image, keyframe extraction on the video, and audio transcription, a high-dimensional floating-point content vector of, for example, 768 dimensions is generated using a context-aware fusion mechanism.

[0038] Step S3 calculates the similarity between the content vector and multiple product vectors in the advertising database to obtain a set of related product scores. This step utilizes an efficient vector retrieval engine to perform a highly efficient approximate nearest neighbor search, and can combine business value factors to adjust the scores, resulting in a set of related product scores sorted from high to low.

[0039] Step S4: Based on the relevant product scores, a tiered delivery strategy is triggered by comparing them with preset thresholds to determine the final ads to be delivered. This step is completed by a lightweight decision engine, which receives a sorted list of scores and compares them sequentially with a first threshold and a second threshold, outputting a unique delivery decision instruction.

[0040] Step S5: Track the return on investment (ROI) of the ads already placed, and periodically backtest and adaptively optimize the preset thresholds of the tiered placement strategy based on the ROI data.

[0041] The multimodal content parsing and aggregation method in step S2 is a context-aware multimodal fusion process. The multimodal fusion process further includes performing cross-modal correlation analysis on the initial feature vectors extracted from each modality to determine the relative importance of each modality in expressing the core theme in the specific context of the current content data, thereby generating a dynamic saliency weight for each modality; and performing weighted aggregation on the initial feature vectors based on the dynamic saliency weight to generate the final content vector.

[0042] Step S4 triggers a tiered delivery strategy, including: if the highest relevant product score is greater than the first threshold, then select the specific product ad corresponding to the highest score for delivery; if the highest score is not greater than the first threshold but greater than the second threshold, then select a category-level ad or alternative product ad related to the content data theme from a pool of alternative ads for delivery; if the highest score is not greater than the second threshold, then initiate an opportunity exploration process to select an ad from a dedicated exploration ad pool for delivery.

[0043] Step S5, backtesting and adaptive optimization, further includes: defining a strategy parameter vector containing a first threshold, a second threshold, and an exploration constant in the opportunity exploration delivery algorithm; periodically simulating multiple different sets of strategy parameter vectors on a historical dataset to evaluate the simulated total return on delivery for each set of vectors; and automatically updating the strategy parameter vector currently used by the online system to the parameter vector that produces the highest total return on delivery in the simulation.

[0044] When performing the opportunity exploration process, this step further includes a complete learning and updating loop: calculating a potential return score for each ad in the exploration pool that combines its known historical average return with an exploration reward; selecting and delivering the ad with the highest potential return score; and after delivery, capturing the actual return data for this delivery and using this data to update the historical average return and delivery frequency statistics of the ad for future score calculations.

[0045] In executing the opportunity exploration process, this step first calculates a potential return score for each ad in the exploration ad pool. This score consists of two parts: one part is its known historical average return retrieved from the database, and the other part is an exploration bonus item. The size of this bonus item is carefully designed to be positively correlated with the total number of opportunity exploration runs and negatively correlated with the number of times the ad itself has been run. Next, the ad with the highest potential return score is selected and run. Finally, after the run is completed, which is crucial for forming a closed loop, the system captures the actual return data of this run and immediately updates the historical average return and the number of times the ad has been run with this new data. These updated values ​​will be directly used in future potential return score calculations.

[0046] The above description is merely an example and illustration of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the scope defined by the invention, and all such modifications and additions should fall within the protection scope of the present invention.

Claims

1. A cross-platform automated advertising delivery system, characterized in that, include: A multimodal content acquisition module is used to acquire content data containing at least one modality of text, images, or videos from at least one social platform; The multimodal content analysis module, connected to the multimodal content acquisition module, is used to perform cross-modal topic aggregation on multimodal content data to generate a unified content vector that represents the core topic of the content data. The relevance scoring module is used to calculate the relevance score between a unified content vector and multiple product vectors pre-stored in the ad library. The dynamic delivery strategy module is used to select an advertising product to be delivered from the ad library based on the relevant product score and a hierarchical delivery strategy containing multiple trigger conditions. The advertising performance backtesting module is used to track the return on investment efficiency of the ads that have been placed and to perform backtesting analysis on the tiered advertising strategy based on historical data. The strategy adaptive optimization module is used to automatically adjust the triggering conditions in the hierarchical delivery strategy based on the results of the backtesting analysis.

2. The cross-platform automated advertising delivery system according to claim 1, characterized in that: The multimodal content analysis module includes a multimodal saliency fusion network, which further includes a set of modality-specific encoders for converting data of the corresponding modality into initial feature vectors; and a cross-modal attention module for receiving the initial feature vectors and calculating a dynamic saliency weight for each modality by evaluating the interrelationship between the features of each modality. A fusion layer is used to perform weighted aggregation of the initial feature vector based on the dynamic saliency weights to generate the final unified content vector.

3. The cross-platform automated advertising delivery system according to claim 1, characterized in that: The relevance scoring module generates the final relevance product score through a two-stage process, which includes a semantic matching stage, used to obtain a basic content-advertisement similarity score by calculating the cosine similarity between the unified content vector and the product vector. In the business value adjustment phase, the basic content-advertisement similarity score is adjusted using business value factors associated with the advertised product, wherein the business value factors depend on at least one of the following: the historical performance of the advertised product, profit margin, and inventory status.

4. The cross-platform automated advertising delivery system according to claim 1, characterized in that: The tiered targeting strategy is based on a comparison between the highest relevant product score and two preset thresholds, namely a first preset threshold and a second preset threshold. Specifically, it includes: a high-score precision targeting strategy triggered when the highest score is higher than the first preset threshold; an association expansion targeting strategy triggered when the highest score is not higher than the first preset threshold but higher than the second preset threshold; and an opportunity exploration targeting strategy triggered when the highest score is not higher than the second preset threshold.

5. The cross-platform automated advertising delivery system according to claim 4, characterized in that: The opportunity exploration delivery strategy employs a reward-driven exploratory delivery decision algorithm, which performs a decision on each ad in the exploratory ad pool. Calculate a potential return score And select the highest-scoring ads for display, among which The calculation formula is: ,in It's an advertisement. Historical average return efficiency, This represents the total number of opportunity exploration campaigns executed, which is the number of times the ad has been placed. It is a preset constant used to balance the exploratory and exploitative aspects of deployment.

6. The cross-platform automated advertising delivery system according to claim 1, comprising the method of executing cross-platform advertising delivery, characterized in that: Includes the following steps, Step S1: Collect multimodal content data from the designated social media platform according to the set information; Step S2: Perform a multimodal content parsing and aggregation method on the multimodal content data to generate a content vector that can uniformly represent its core theme; Step S3: Calculate the similarity between the content vector and multiple product vectors in the advertising library to obtain a set of related product scores; Step S4: Based on the relevant product scores, compare them with preset thresholds to trigger corresponding tiered delivery strategies to determine the final ads to be delivered; Step S5: Track the return on investment (ROI) of the deployed ads, and periodically backtest and adaptively optimize the preset thresholds of the tiered deployment strategy based on the ROI data.

7. The cross-platform automated advertising delivery system according to claim 6, characterized in that: The multimodal content parsing and aggregation method described in step S2 is a context-aware multimodal fusion process. The multimodal fusion process further includes performing cross-modal association analysis on the initial feature vectors extracted from each modality to determine the relative importance of each modality in expressing the core theme in the specific context of the current content data, thereby generating a dynamic saliency weight for each modality; and performing weighted aggregation on the initial feature vectors based on the dynamic saliency weight to generate the final content vector.

8. The cross-platform automated advertising delivery system according to claim 6, characterized in that: The triggering of the tiered delivery strategy in step S4 includes: if the highest relevant product score is greater than a first threshold, then selecting a specific product advertisement corresponding to the highest score for delivery; if the highest score is not greater than the first threshold but greater than a second threshold, then selecting a category-level advertisement or alternative product advertisement related to the content data theme from a pool of alternative advertisements for delivery; if the highest score is not greater than the second threshold, then initiating an opportunity exploration process to select an advertisement from a dedicated exploration advertisement pool for delivery.

9. A cross-platform automated advertising delivery system according to claim 6, characterized in that: The backtesting and adaptive optimization described in step S5 further includes: defining a strategy parameter vector that includes the first threshold, the second threshold, and the exploration constant in the opportunity exploration delivery algorithm; periodically simulating multiple different sets of the strategy parameter vectors on a historical dataset to evaluate the simulated total return on delivery corresponding to each set of vectors; and automatically updating the strategy parameter vector currently used by the online system to the parameter vector that produces the highest total return on delivery in the simulation.

10. A cross-platform automated advertising delivery system according to claim 6, characterized in that: In performing the opportunity exploration process, this step further includes a complete learning and updating loop: calculating a potential return score for each ad in the exploration ad pool that combines its known historical average return with an exploration reward; selecting and delivering the ad with the highest potential return score; and after delivery, capturing the actual return data for this delivery and using this data to update the historical average return and delivery frequency statistics of the ad for future score calculations.

Images