AI-based short video batch generation and intelligent delivery closed-loop optimization system and method

Abstract

This invention discloses an AI-based closed-loop optimization method and system for batch generation and intelligent delivery of short videos, belonging to the field of intelligent short video marketing technology. This invention constructs a causal directed acyclic graph and combines it with a backdoor adjustment algorithm to remove confounding variables, calculates the CATE (Customer Aspect Ratio) of sub-groups, and divides them into heterogeneous differential user clusters. Relying on a generative model embedded with the MMD (Multi-Level Marketing) loss function, it produces differentiated short video materials, avoiding content homogenization and traffic involution problems. This invention splits the delivery budget into three independent quotas, achieves dynamic delivery through game optimization using three sets of reinforcement learning agents, reuses a causal model to decouple data, builds a causal reasoning knowledge base, and employs high- and low-frequency peak-shifting self-correction and adaptive noise injection mechanisms to solve the problems of model iteration oscillation and collapse. This invention achieves closed-loop optimization of the entire short video generation and delivery chain, effectively improving delivery accuracy and marketing ROI, and is suitable for large-scale intelligent operation scenarios of short videos.

Description

Technical Field

[0001] This invention belongs to the field of short video technology, specifically relating to an AI-based closed-loop optimization system and method for batch generation and intelligent delivery of short videos. Background Technology

[0002] With the rapid development of mobile internet and short video platforms, short video marketing has become a core channel for businesses to promote products and build brands. Currently, mainstream short video advertising models typically rely on a combination strategy of "content creation + feed delivery".

[0003] In existing technologies, the optimization of short video delivery mainly suffers from the following limitations: Audience targeting relies on correlation analysis, which suffers from label distortion issues. Existing ad targeting systems mostly rely on historical conversion data for correlation analysis when targeting audiences. For example, the system might identify and target audiences with "high click-through rates." However, this method fails to eliminate the influence of external confounding variables such as bid volume, timing of ad placement, and channel differences. This often leads the system to misclassify "high-conversion audiences" as "high-gain audiences," when in reality these audiences may already have a strong purchase intent, and the increase in value may not be due to the ad content itself. This results in wasted marketing budgets and distorted audience segmentation.

[0004] Mass-generated content suffers from severe homogenization, easily leading to traffic involution. To improve production efficiency, a large number of tools using AI to generate short videos in batches have emerged in the market. However, the materials generated by these tools are often highly templated and lack content differentiation. When a large number of similar video materials are placed in the same audience pool, they compete for traffic, causing the cost per mille (CPM) and cost per appeal (CPA) of a single video to rise sharply, forming "traffic involution" and severely reducing the overall return on investment (ROI) of the campaign.

[0005] The disconnect between delivery strategy and content generation makes model iteration prone to oscillations and collapse. Current automated ad delivery systems typically treat "content generation" and "ad delivery" as two separate processes, lacking an effective closed-loop feedback mechanism. Even when some systems incorporate reinforcement learning for budget allocation, they often rely on a single agent for global optimization. In complex, dynamic environments, a single agent struggles to balance exploring new content (exploration) with maintaining stable ad delivery volume (utilization), easily leading to drastic budget fluctuations across different campaigns. Furthermore, when the generation and delivery models iterate simultaneously at high frequencies, the lack of a collaborative correction mechanism can easily trigger systemic performance degradation or even model "collapse," resulting in a precipitous drop in ad delivery effectiveness.

[0006] Therefore, how to construct a short video delivery method that can accurately identify target audiences by removing confounding variables, generate differentiated content to avoid involution, and achieve full-link closed-loop self-balancing optimization has become a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] To address the shortcomings of the existing technologies, this invention proposes an AI-based closed-loop optimization method, system, computer equipment, and computer-readable storage medium for batch generation and intelligent delivery of short videos.

[0008] The technical solution adopted in this invention is as follows: In one implementation of the first aspect, this application provides an AI-based closed-loop optimization method for batch generation and intelligent delivery of short videos, comprising the following steps: Step S1: Collect full-domain user profile data, platform delivery strategy parameters and video conversion result data, construct a standardized causal directed acyclic graph (DAG), use a backdoor adjustment algorithm to remove confounding variables, calculate the conditional average treatment effect (CATE) for different subgroups, and divide them into multiple heterogeneous differential population clusters based on the CATE value. Step S2: Using the heterogeneous differential population clusters as the orientation benchmark, perform adversarial constraint generation operations using a generative model that embeds the maximum mean difference (MMD) loss function to generate short video materials in batches with significant differences in high-dimensional features between clusters and rich and diverse content features within clusters. Step S3: Divide the total budget into three independent quota resource pools for exploration, scaling, and risk control, and deploy three independent reinforcement learning agents to conduct budget allocation game optimization within their respective exclusive quotas. Based on the game results, dynamically distribute the short video materials to different channels. Step S4: Collect business data from the entire delivery chain, reuse the causal directed acyclic graph (DAG) to perform a second causal decoupling operation, separate non-content interference factors to output pure content dimension causal benefit values, and build a causal reasoning knowledge base based on the causal benefit values. Step S5: Retrieve feedback data from the causal reasoning knowledge base, execute a time-series peak-shifting bidirectional self-correction iterative process, perform high-frequency small-amplitude parameter real-time fine-tuning for the delivery-side intelligent agent, and perform low-frequency large-amplitude rule periodic updates for the video generation model to form a full-link dynamic self-balancing operation closed loop.

[0009] In one implementation of the second aspect, this application provides an AI-based short video batch generation and intelligent delivery closed-loop optimization system, which serves as a computer carrier to execute the method provided by the first aspect or any possible implementation of the first aspect.

[0010] The architecture specifically includes: The population modeling module is used to construct a standardized causal directed acyclic graph (DAG), calculate the conditional average treatment effect (CATE) using a backdoor adjustment algorithm, and divide the population into multiple heterogeneous differential population clusters. The content generation module is used to generate short video materials in batches based on the heterogeneous differential population cluster by utilizing a generation model that embeds the maximum mean difference (MMD) loss function. The game-theoretic delivery module is used to split the total delivery budget into three independent quota resource pools, deploy three independent reinforcement learning agents to carry out budget allocation game optimization and execute dynamic delivery by channel; The reasoning and detection module is used to reuse the causal directed acyclic graph (DAG) for secondary causal decoupling operations and build a causal reasoning knowledge base. The bidirectional self-correction module is used to retrieve feedback data from the causal reasoning knowledge base and execute the time-series peak-shifting bidirectional self-correction iterative process. Each module completes real-time data transmission, interaction, and synchronization through a dedicated causal bidirectional data bus, forming a complete closed-loop data interaction system.

[0011] In one implementation of the third aspect, this application provides a computer device including a processor, a memory, and a communication bus, wherein the processor and the memory communicate with each other through the communication bus; The memory stores program instructions that can be executed by the processor, which can invoke the program instructions to perform the method provided by the first aspect or any possible implementation thereof.

[0012] In one implementation of the fourth aspect, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method provided by the first aspect or any possible implementation of the first aspect.

[0013] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: This invention offers significant technological advantages and application benefits compared to existing technologies. Its core lies in using causal DAG modeling and a backdoor adjustment algorithm to remove confounding variables and accurately calculate the true causal benefit of CATE (Customer Targeting), thus solving the pain points of traditional audience targeting labels being distorted and prone to incremental misjudgments. Simultaneously, it embeds the MMD (Multi-Level Marketing) loss function to achieve differentiated generation of short videos between clusters and diversified generation within clusters, preventing content homogenization and traffic involution. Through a three-layer quota resource pool and three sets of reinforcement learning agents, it optimizes dynamic delivery, balancing innovation and risk control. Combined with a time-series staggered peak bidirectional self-correction mechanism and an adaptive noise injection mechanism, it avoids model iteration oscillation and collapse issues, ultimately constructing a fully intelligent self-balancing closed-loop system that significantly improves the ROI and scalable operational efficiency of short video delivery. Attached Figure Description

[0014] The present invention will be described by way of example and with reference to the accompanying drawings, wherein: Figure 1 This is a flowchart of the method in this invention; Figure 2 This is a block diagram of the system architecture in this invention.

[0015] Figure 3 This is a hardware block diagram of the computer device of the present invention. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0017] Example 1:

[0018] This embodiment provides an AI-based closed-loop optimization method for batch generation and intelligent delivery of short videos. (See attached document.) Figure 1 This includes the following steps: Step S1: Collect comprehensive user profile data, platform placement strategy parameters, and video conversion result data. Construct a standardized causal directed acyclic graph (DAG). Use a backdoor adjustment algorithm to remove confounding variables, calculate the conditional average treatment effect (CATE) for different sub-groups, and use the CATE value as the core clustering criterion to form multiple heterogeneous differential user clusters. Specifically, the system comprehensively acquires user profile data such as historical behavior and consumption preferences, as well as placement strategies and conversion results from various channels, by connecting to multi-source data interfaces. Based on this, a standardized causal directed acyclic graph (DAG) containing multiple variables is constructed. By introducing a backdoor adjustment algorithm, the system can effectively identify and remove confounding variables such as placement time and bid level that interfere with conversion results, thereby accurately calculating the true causal gain of specific short video content on different sub-groups, i.e., the conditional average treatment effect (CATE). Subsequently, using the calculated CATE value as the core clustering criterion, a clustering algorithm is used to divide users into multiple heterogeneous differential user clusters with significant differences in content preferences, providing data support for subsequent precise targeting.

[0019] Step S2: Using the heterogeneous differential population clusters as the orientation benchmark, the system performs adversarial constraint generation operations using a generative model embedded with the Maximum Mean Difference (MMD) loss function to batch-produce short video materials with significant differences in high-dimensional features between clusters and rich and diverse content features within clusters. During the generation operation, the system uses the generative model to batch-create video materials based on the specific preferences of each heterogeneous differential population cluster. To overcome the content homogenization problem caused by traditional batch generation, the system innovatively embeds a Maximum Mean Difference (MMD) loss term into the loss function of the generation model. During the operation, this loss term, on the one hand, maximizes the high-dimensional feature distribution distance between different content clusters to ensure significant differences in the content produced between clusters; on the other hand, it minimizes the feature distribution distance within the same content cluster to ensure that the content generated within the cluster maintains diversity while not deviating from the core preferences of that population cluster. Through this adversarial constraint generation operation, the final batch-produced short video materials not only meet the personalized needs of different groups but also ensure the richness of the overall content matrix.

[0020] Step S3: The total budget is divided into three independent quota resource pools: exploration, scaling, and risk control. Three independent reinforcement learning agents are deployed to conduct budget allocation game optimization within their respective quotas, and the short video materials are dynamically distributed to different channels based on the game results. In the execution phase, the system abandons the traditional model of globally unified budget allocation. Instead, the total budget is scientifically divided into an exploration quota for testing new materials, a scaling quota for expanding the exposure of high-quality materials, and a risk control and retention quota to ensure the overall ROI baseline. These three quota resource pools are independent and do not interfere with each other. For each of these three resource pools, the system deploys three independent reinforcement learning agents. Each agent can only perform budget allocation game optimization calculations within its own quota based on real-time feedback on the delivery data. Finally, the system integrates the game optimization results of the three agents and accurately distributes the short video materials produced in Step S2 to different traffic channels, achieving dynamic and efficient delivery execution.

[0021] Step S4: Collect business data from the entire delivery chain, reuse the causal directed acyclic graph (DAG) for a second causal decoupling operation, separate non-content interference factors to output pure content-dimensional causal benefit values, and build a causal inference knowledge base based on these causal benefit values. When short video materials enter the delivery chain, the system collects full-chain business data in real time, including exposure, click-through rate, and conversion rate. To accurately assess the attractiveness of the video content itself, the system reuses the causal directed acyclic graph (DAG) model built in Step S1 to perform a second causal decoupling operation on this business data. This operation aims to remove non-content interference factors such as platform traffic fluctuations and holiday effects, thereby accurately outputting pure content-dimensional causal benefit values ​​solely derived from the video content itself. The system stores these high-value causal benefit values ​​in a structured manner, gradually building a self-improving causal inference knowledge base to provide clean and reliable feedback signals for subsequent model iterations.

[0022] Step S5: Retrieve feedback data from the causal reasoning knowledge base and execute a time-series staggered bidirectional self-correcting iterative process. For the delivery-side agent, perform high-frequency, small-amplitude parameter fine-tuning in real time; for the video generation model, perform low-frequency, large-amplitude rule periodic updates, forming a dynamic self-balancing operational closed loop across the entire chain. To achieve long-term stable operation, the system continuously retrieves the latest feedback data from the causal reasoning knowledge base in Step S4 and executes a time-series staggered bidirectional self-correcting iterative process. In this process, for the delivery-side agent, which has extremely high timeliness requirements, the system performs high-frequency, small-amplitude parameter fine-tuning in real time, enabling it to quickly respond to the rapidly changing market traffic environment; while for the video generation model, it performs low-frequency, large-amplitude rule periodic updates, using a large amount of accumulated causal revenue data to deeply optimize its generation logic. Through this high- and low-frequency decoupled iterative approach, the delivery model and the generation model can mutually correct and co-evolve, ultimately forming a self-optimizing, dynamically self-balancing short video generation and delivery operation closed loop.

[0023] In one specific implementation, step S1 involves using a backdoor adjustment algorithm to remove confounding variables and calculate the conditional average treatment effect (CATE) for each subgroup. This includes defining confounding nodes Z, processing nodes X, covariate nodes U, and result nodes Y. The confounding nodes Z include bid price, delivery time, and distribution channel; the processing nodes X correspond to the core content cluster features of the short video; the covariate nodes U include user age, consumption level, and interest preferences; and the result nodes Y include completion rate and conversion rate. All confounding nodes Z are fixed, and the conditional causal effect (CATE) for each subgroup is calculated for each type of independent covariate U. The calculation formula is: CATE(u) = E[Y|do (X=1), Z=z, U=u] - E[Y|do (X=0), Z=z, U=u]; where E[Y|do (X=1), Z=z, U=u] represents the expected conversion of subgroup u after exposure to the target content feature X. U=u] ​​represents the baseline conversion expectation of subgroup u before it has been exposed to the target content features. The normalized CATE value is used as the core clustering index, and a K-means clustering algorithm is executed in conjunction with the user covariate U. A fixed inter-cluster CATE difference threshold Δ is preset to ensure that the content preference difference between any two subgroups is quantifiable, outputting standardized heterogeneous difference adversarial clusters. In this implementation, to accurately remove confounding variables, the system first clearly defines four types of core nodes in the causal directed acyclic graph.

[0024] Specifically, the hybrid node Z, which may simultaneously affect content delivery and conversion results, is defined as including bid price, delivery time, and distribution channel; the processing node X, which serves as an intervention variable, is defined as the core content cluster feature of the short video; the covariate node U, which describes the user's own attributes, is defined as including user age, consumption level, and interest preferences; and the result node Y, which measures marketing effectiveness, is defined as including completion rate and conversion rate. After defining the nodes, the system fixes all hybrid nodes Z and calculates the specific conditional causal effect CATE for each subgroup for each independent covariate U. The calculation formula is: CATE (u) = E [Y|do (X=1), Z=z, U=u] - E [Y|do (X=0), Z=z, U=u]. In this formula, E [Y|do (X=1), Z=z, U=u] represents the conversion expectation of subgroup u after exposure to the target content feature X, while E [Y|do (X=0), Z=z, U=u] represents the baseline conversion expectation of subgroup u before exposure to the target content feature X. The difference between the two values ​​represents the true causal gain of content feature X on subgroup u after eliminating external interference. After calculating the CATE value, the system normalizes it and uses it as the core clustering indicator, which is then used in conjunction with the user covariate U to execute the K-means clustering algorithm. Simultaneously, the system presets a fixed inter-cluster CATE difference threshold Δ. This threshold ensures that the content preference difference between any two group clusters is significant and quantifiable, ultimately outputting standardized heterogeneous differential adversarial clusters, which form the basis for subsequent differentiated content generation.

[0025] In one specific implementation, step S2 involves using a generative model with an embedded maximum mean difference (MMD) loss function to perform adversarial constraint generation operations. Specifically, this includes adding an MMD-specific loss term to the loss function of the generative model, maximizing the high-dimensional feature distribution distance between different content clusters while simultaneously minimizing the feature distribution distance within the same content cluster, thus achieving controllable batch generation of short videos with strong inter-cluster differentiation and micro-diversification within clusters. Step S2 also includes a shot temporal causal coding screening process: retrieving the CATE causal contribution of historical shots to complete scoring and sorting, automatically eliminating low-gain invalid shots, and assigning higher generation weights to the golden 3-second opening shot. Step S2 is further supported by a dynamic iterative platform knowledge graph content distillation step: pre-building a knowledge graph covering multi-platform review rules, automatically performing content distillation processing on original materials from the same source according to the differentiated rules of the target platform, and dynamically updating the weights of various rules within the knowledge graph based on platform-pushed traffic limiting and violation feedback data.

[0026] In this implementation, to achieve controllable batch generation of short video materials, the system adds an MMD-specific loss term to the loss function of the generation model. During model training and generation operations, this loss term guides the model to maximize the high-dimensional feature distribution distance between different content clusters, thereby ensuring that the generated videos have significant inter-cluster differentiation in visual, narrative, and other dimensions; simultaneously, it minimizes the feature distribution distance within the same content cluster, ensuring that videos within the same cluster maintain micro-diversity while staying true to the core theme of that cluster. Furthermore, this step innovatively introduces a shot temporal causal coding screening process. The system retrieves the CATE causal contribution of each shot from historical delivery data and scores and ranks them. Based on the ranking results, the system can automatically remove invalid shots with extremely low conversion gains and assign higher generation weights to the crucial "golden 3 seconds" opening shot at the beginning of the video, thereby improving the overall video completion rate and attractiveness. Moreover, this step is also accompanied by a dynamic iterative platform knowledge graph content distillation step. The system pre-builds a knowledge graph covering the review rules of major short video platforms. When generating videos, the system automatically performs content distillation processing on original source materials according to the differentiated rules of the target platform. This includes automatically blurring sensitive scenes or replacing prohibited text. More importantly, the system can dynamically update the weights of various rules within the knowledge graph based on feedback data such as traffic restrictions and violations from the platform. This allows the content generation model to continuously adapt to changes in platform rules, ensuring the compliance of batch-generated content.

[0027] In one specific implementation, step S3 involves deploying three independent reinforcement learning agents to conduct budget allocation game optimization within their respective exclusive quotas. Specifically, this includes dividing the quota into three independent resource pools: exploration quota, volume expansion quota, and risk control retention quota, each corresponding to an exploration agent, a volume expansion agent, and a risk control agent, respectively. Each agent can only conduct budget allocation game calculations within its own exclusive quota, and the flow of quota resources is subject to unidirectional constraints, supporting only the flow from exploration quota to volume expansion quota or from volume expansion quota to risk control quota, and prohibiting reverse resource allocation. Step S3 also includes a quantitative involution penalty mechanism for traffic within the same cluster: real-time monitoring of the total number of parallel delivery plans and the overlap of video exposures in scenarios where the same user group is paired with the same content group. When both exceed a preset threshold, the involution penalty logic is automatically triggered, and the penalty coefficient is positively correlated with the high-dimensional feature similarity between videos.

[0028] In this implementation, to ensure the stability and efficiency of the deployment strategy, the system strictly divides the total deployment budget into three independent resource pools: exploration quota, volume expansion quota, and risk control retention quota. Exploration agents, volume expansion agents, and risk control agents are deployed correspondingly for each pool. During the game-theoretic optimization process, each agent is strictly restricted to performing budget allocation calculations only within its own dedicated quota to prevent resource misallocation. Simultaneously, the flow of quota resources is subject to a clear one-way constraint mechanism: flow is only supported from exploration quota to volume expansion quota, or from volume expansion quota to risk control retention quota. The system strictly prohibits any form of reverse resource transfer, thereby ensuring the hierarchy and security of budget allocation.

[0029] Furthermore, to address the issue of traffic involution caused by competition among mass-generated content within the same audience, this step also includes a quantitative penalty mechanism for involution within the same traffic cluster. The system monitors in real time the total number of parallel campaigns running in the same audience cluster with the same content cluster, as well as the overlap in exposure between videos. When both metrics exceed the system's preset safety thresholds, the system automatically triggers the involution penalty logic, penalizingly reducing the budget allocation for the relevant campaigns. Moreover, this penalty mechanism has dynamic adjustment capabilities; its penalty coefficient is positively correlated with the high-dimensional feature similarity between related videos. That is, the more similar the video content, the heavier the involution penalty, thus forcing the system to generate more differentiated content.

[0030] In a further embodiment, step S3, which involves dynamically distributing the short video materials across different channels based on the game results, further includes: each agent automatically transferring the video between the test warehouse, the potential volume-boosting warehouse, and the stop-loss warehouse based on the real-time CATE gain value of the short video; calculating the CATE causal benefit corresponding to the unit budget of each distribution channel in real time, and automatically migrating idle quotas from inefficient channels to high-yield channels to complete cross-channel budget dynamic migration compensation. In this further embodiment, in order to maximize budget utilization efficiency, the system establishes a video transfer and budget migration mechanism based on real-time data when performing dynamic distribution across different channels.

[0031] Specifically, each reinforcement learning AI automatically categorizes videos based on their real-time CATE gain values ​​generated during the campaign, dynamically managing their movement between the test, potential growth, and stop-loss zones. High-performing videos are quickly promoted to the potential growth zone, while poorly performing videos are downgraded to the stop-loss zone or even suspended. Simultaneously, the system calculates the CATE causal return for each campaign channel per unit budget. If inefficient budget utilization is detected on certain channels, the system automatically migrates idle quotas from these inefficient channels to high-yield channels, achieving dynamic budget migration compensation across channels and ensuring that every penny of the budget is invested where the return on investment is highest.

[0032] In one specific implementation, step S5 involves executing a time-series staggered bidirectional self-correcting iterative process, specifically employing a high- and low-frequency decoupled independent operation mechanism: the game-playing agent performs high-frequency small-amplitude iterative calculations, fine-tuning the quota allocation ratio, bidding price, and targeted audience parameters in real time; the multimodal generation model performs low-frequency large-amplitude iterative calculations, periodically updating the basic rules for content generation, MMD adversarial constraint parameters, and shot generation weights; the time-series staggered bidirectional self-correcting iterative process includes downlink correction and uplink correction: downlink correction fine-tunes the generation model parameters based on high CATE causal features; uplink correction corrects the quota game weights of the three agents based on real-time content cluster distribution data.

[0033] In this implementation, to avoid system oscillations during synchronous iteration between the generation model and the delivery model, step S5 employs a decoupled high- and low-frequency independent operation mechanism. For the delivery game agent that needs to respond quickly to market changes, the system performs high-frequency, small-amplitude iterative calculations, enabling it to fine-tune the quota allocation ratios, bidding prices, and targeted audience parameters of each channel in real time. For the multimodal generation model, the system performs low-frequency, large-amplitude iterative calculations to periodically update its basic content generation rules, MMD adversarial constraint parameters, and generation weights of various shots. Furthermore, this time-series staggered bidirectional self-correcting iterative process also includes a collaborative mechanism with two dimensions: downlink correction and uplink correction. Downlink correction refers to fine-tuning the parameters of the generation model based on the high CATE causal characteristics fed back from the delivery end, guiding it to evolve towards higher conversion rates. Uplink correction refers to adjusting the quota game weights of the three groups of agents based on real-time content cluster distribution data from the generation end, ensuring that the delivery strategy matches the current content matrix distribution.

[0034] In a further embodiment, step S5 is also equipped with a Gini coefficient adaptive noise injection anti-collapse step: the Gini coefficient corresponding to the current distribution of all content clusters is calculated in real time, and the bias degree generated during the model iteration is quantified by the Gini coefficient value; the injection intensity of the anti-noise is adaptively adjusted according to the Gini coefficient value: when the Gini coefficient is determined to be in a slightly biased state, a small-intensity perturbation noise is injected; when the Gini coefficient is determined to be in a heavily biased state, a high-intensity perturbation noise is injected, and the overall content diversity is forcibly expanded by relying on strong perturbation.

[0035] In this further implementation, to prevent the generative model from getting stuck in local optima during long-term iterations, leading to content homogenization or even model "collapse," step S5 also includes a Gini coefficient adaptive noise injection anti-collapse step. The system calculates the Gini coefficient corresponding to the current distribution of all content clusters in real time and uses this Gini coefficient value to accurately quantify the bias generated by the model during iteration. Subsequently, the system adaptively adjusts the injection intensity of adversarial noise based on the magnitude of the Gini coefficient value.

[0036] Specifically, when the Gini coefficient determination model is in a slightly biased state, the system will inject low-intensity perturbation noise to gently correct the model; while when the Gini coefficient determination model is in a heavily biased state, the system will inject high-intensity perturbation noise. This strong perturbation forces the expansion of the diversity of the overall content, enabling the model to escape the trap of local optima, thereby ensuring that the entire closed-loop system can operate in a long-term, healthy and stable manner.

[0037] Example 2:

[0038] This embodiment provides an AI-based closed-loop optimization system for batch generation and intelligent delivery of short videos. This system aims to achieve intelligent and self-consistent operation across the entire short video marketing chain through the collaborative work of its various functional modules. (See also...) Figure 2 Specifically, the system includes the following core modules: The audience modeling module, serving as the foundation for the entire system's targeting, is primarily used to construct a standardized causal directed acyclic graph (DAG). This module deeply mines the entire dataset, employing a backdoor adjustment algorithm to precisely remove external confounding variables such as bids and time periods, thereby scientifically calculating the Conditional Average Treatment Effect (CATE) for different sub-groups. Based on this, the module uses CATE values ​​as the core clustering criterion, accurately dividing massive user data into multiple heterogeneous differential audience clusters with significant differences in content preferences, providing solid data support for subsequent differentiated marketing.

[0039] The content generation module, closely integrated with the crowd modeling module, is primarily used for creating content using a generative model that incorporates the Maximum Mean Difference (MMD) loss function. This module uses the heterogeneous, differentially distributed crowd clusters as a directional benchmark. During the generation process, it employs adversarial constraints on the MMD loss term to forcibly widen the high-dimensional feature distance between different clusters while maintaining the micro-diversity of features within each cluster. Through this intelligent computation, the module can batch-produce short video content with significant differences in high-dimensional features between clusters and rich, diverse content features within each cluster, effectively solving the content homogenization problem caused by traditional batch generation.

[0040] The game-theoretic delivery module, serving as the system's execution hub, is primarily responsible for scientifically dividing the total delivery budget into three independent quota resource pools: exploration, scaling, and risk control. For each of these three resource pools, the module deploys three independent reinforcement learning agents, enabling them to conduct budget allocation game-theoretic optimization within their respective exclusive quotas. Through this isolated game-theoretic mechanism, the module can dynamically deliver short video materials across different channels based on the game results, ensuring both the security of budget allocation and maximizing traffic utilization.

[0041] The reasoning and detection module, serving as the system's feedback and evaluation hub, primarily reuses the previously constructed causal directed acyclic graph (DAG) to perform secondary causal decoupling operations on the entire business data chain. This module effectively isolates non-content-related interference factors such as platform traffic fluctuations, accurately outputting causal benefit values ​​in the pure content dimension. Based on these high-value causal benefit values, this module gradually builds and improves the causal reasoning knowledge base, providing clean and reliable feedback signals for the system's continuous evolution.

[0042] The bidirectional self-calibration module, as the core component ensuring the long-term stable operation of the system, is mainly used to retrieve the latest feedback data from the causal reasoning knowledge base and execute a time-series staggered bidirectional self-calibration iterative process. In this process, the module performs high-frequency, small-amplitude parameter fine-tuning on the delivery-side agent in real time to quickly respond to market changes; simultaneously, it performs low-frequency, large-amplitude rule periodic updates on the video generation model to deeply optimize the generation logic. Through this decoupled high- and low-frequency iterative approach, the module successfully forms a closed-loop dynamic self-balancing operation across the entire chain, avoiding system oscillations caused by synchronous model iteration.

[0043] A dedicated causal bidirectional data bus serves as the underlying data channel connecting the aforementioned modules. All modules communicate with each other via this dedicated causal bidirectional data bus for real-time data transmission, interaction, and synchronization. This bus ensures seamless data flow throughout the entire process, from audience targeting, content generation, game-based deployment to feedback and correction, thus forming a complete, efficient, and self-evolving closed-loop data interaction system.

[0044] Example 3:

[0045] This embodiment provides a computer device as the physical hardware carrier for carrying and running the aforementioned AI-based short video batch generation and intelligent delivery closed-loop optimization method. (See also...) Figure 3 Specifically, the computer device includes a processor, a memory, and a communication bus. The processor and the memory communicate with each other via the communication bus to ensure efficient and stable transmission of internal data streams and control signals.

[0046] The memory stores program instructions executable by the processor. This memory may include volatile memory (such as random access memory, RAM) or non-volatile memory (such as read-only memory, flash memory, or hard disk) for persistently or temporarily storing program code and various types of data generated during execution. The processor (e.g., a central processing unit (CPU), a graphics processing unit (GPU), or a dedicated AI acceleration chip) can invoke the program instructions to execute the methods provided by the first aspect or any possible implementation thereof.

[0047] In actual operation, when the processor executes the aforementioned program instructions, it can drive the computer device to achieve the core functions described in the previous embodiments. For example, the processor can control data acquisition and processing, construct a standardized causal directed acyclic graph (DAG) and calculate the conditional average treatment effect (CATE); it can drive a generative model with an embedded maximum mean difference (MMD) loss function to perform adversarial constraint calculations and batch-produce differentiated short video materials; it can schedule three independent reinforcement learning agents to conduct budget allocation game optimization within three layers of independent quota resource pools; and it can execute a time-series peak-shifting bidirectional self-correcting iterative process. Through the close integration of hardware and software, the computer device can efficiently and stably achieve a dynamic self-balancing operation closed loop for the entire chain of short video generation and distribution.

[0048] Example 4:

[0049] This embodiment provides a computer-readable storage medium that can be applied to various hardware devices with data storage and program execution capabilities, such as computers, servers, smart terminals, and cloud computing devices. The medium stably stores a computer program that can be recognized, loaded, and executed by the device's processor. When the device is powered on, the computer program can be actively invoked and executed by the processor. When the program is running in its entirety, it can accurately implement all steps of the AI-based short video batch generation and intelligent delivery closed-loop optimization method described in Embodiment 1. It can completely replicate the entire technical solution of intelligent batch production of short videos, multi-channel intelligent delivery, real-time data collection, closed-loop analysis of delivery effects, and dynamic strategy optimization in Embodiment 1. It can stably implement the core functions and optimization logic corresponding to this method, providing a reliable program carrier support for the large-scale and intelligent operation of short videos. The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A closed-loop optimization method for batch generation and intelligent delivery of short videos based on AI, characterized in that, Includes the following steps: Step S1: Collect full-domain user profile data, platform delivery strategy parameters and video conversion result data, construct a standardized causal directed acyclic graph (DAG), use a backdoor adjustment algorithm to remove confounding variables, calculate the conditional average treatment effect (CATE) for different subgroups, and divide them into multiple heterogeneous differential population clusters based on the CATE value. Step S2: Using the heterogeneous differential population clusters as the orientation benchmark, perform adversarial constraint generation operations using a generative model that embeds the maximum mean difference (MMD) loss function to generate short video materials in batches with significant differences in high-dimensional features between clusters and rich and diverse content features within clusters. Step S3: Divide the total budget into three independent quota resource pools for exploration, scaling, and risk control, and deploy three independent reinforcement learning agents to conduct budget allocation game optimization within their respective exclusive quotas. Based on the game results, dynamically distribute the short video materials to different channels. Step S4: Collect business data from the entire delivery chain, reuse the causal directed acyclic graph (DAG) to perform a second causal decoupling operation, separate non-content interference factors to output pure content dimension causal benefit values, and build a causal reasoning knowledge base based on the causal benefit values. Step S5: Retrieve feedback data from the causal reasoning knowledge base, execute a time-series peak-shifting bidirectional self-correction iterative process, perform high-frequency small-amplitude parameter real-time fine-tuning for the delivery-side intelligent agent, and perform low-frequency large-amplitude rule periodic updates for the video generation model to form a full-link dynamic self-balancing operation closed loop.

2. The method according to claim 1, characterized in that, In step S1, the backdoor adjustment algorithm is used to remove confounding variables and calculate the conditionally average treatment effect (CATE) for the subgroup, specifically including: Define the hybrid node Z, processing node X, covariate node U, and result node Y; wherein, the hybrid node Z includes the bid price, the time period for the ad placement, and the distribution channel; the processing node X corresponds to the core content cluster features of the short video; the covariate node U includes user age, consumption level, and interest preferences; and the result node Y includes the completion rate and the conversion rate. With all heterogeneous nodes Z fixed, calculate the population-specific conditional causal effect CATE for each class of independent covariates U. The calculation formula is as follows: CATE (u) = E [Y|do (X=1), Z=z, U=u] - E [Y|do (X=0), Z=z, U=u]; In the formula, E[Y|do (X=1), Z=z, U=u] represents the conversion expectation of subgroup u after it comes into contact with the target content feature X, and E[Y|do (X=0), Z=z, U=u] represents the baseline conversion expectation of subgroup u before it comes into contact with the target content feature; The normalized CATE value is used as the core clustering index. K-means clustering algorithm is executed in combination with user covariate U. A fixed CATE difference threshold Δ between clusters is preset to ensure that the content preference difference between any two population clusters can be quantified, and standardized heterogeneous difference adversarial clusters are output.

3. The method according to claim 1, characterized in that, In step S2, adversarial constraint generation is performed using a generative model that embeds the maximum mean difference (MMD) loss function, specifically including: A new MMD-specific loss term is added inside the loss function of the generative model. By maximizing the high-dimensional feature distribution distance between different content clusters and simultaneously minimizing the feature distribution distance within the same content cluster, controllable batch generation of short videos with strong inter-cluster differentiation and micro-diversity within clusters is achieved. Step S2 also includes a shot temporal causal coding screening process: retrieve the CATE causal contribution of historical shots to complete the scoring and sorting, automatically remove low-gain invalid shots, and assign higher generation weight to the golden 3-second opening shot. Step S2 is also equipped with a dynamic iterative platform knowledge graph content distillation step: a knowledge graph covering the review rules of multiple platforms is built in advance, and content distillation is automatically performed on the original materials from the same source according to the differentiated rules of the target platform. The weights of various rules inside the knowledge graph are dynamically updated in reverse based on the traffic restriction and violation feedback data pushed by the platform.

4. The method according to claim 1, characterized in that, In step S3, three independent reinforcement learning agents are deployed to conduct budget allocation game optimization within their respective exclusive quotas, specifically including: The resource pools are divided into three independent categories: exploration quota, volume release quota, and risk control retention quota, each corresponding to an exploration agent, a volume release agent, and a risk control agent, respectively. Each intelligent agent can only conduct budget allocation game calculations within its own exclusive quota, and the flow of quota resources is subject to one-way constraints, supporting only the flow from exploration quota to release quota, or from release quota to risk control quota, and prohibiting reverse resource allocation; Step S3 also includes a quantitative involution penalty mechanism for traffic within the same cluster: real-time monitoring of the total number of parallel delivery plans and the overlap of video exposures in scenarios where the same user group is paired with the same content group. When both exceed a preset threshold, the involution penalty logic is automatically triggered, and the penalty coefficient is positively correlated with the high-dimensional feature similarity between videos.

5. The method according to claim 4, characterized in that, In step S3, the dynamic distribution of the short video materials across different channels based on the game result further includes: Each intelligent agent automatically transfers video between the test chamber, potential volume expansion chamber, and stop-loss chamber based on the real-time CATE gain value of the short video. It calculates the CATE causal revenue corresponding to the unit budget of each distribution channel in real time, automatically migrates idle quotas from inefficient channels to high-yield channels, and completes cross-channel budget dynamic migration compensation.

6. The method according to claim 1, characterized in that, In step S5, a bidirectional self-correction iterative process for timing peak shifting is executed, specifically employing a high- and low-frequency decoupling independent operation mechanism: The intelligent agent performs high-frequency, small-amplitude iterative calculations to fine-tune the quota allocation ratio, bidding price, and target audience parameters in real time. The multimodal generation model performs low-frequency, large-amplitude iterative calculations to periodically update the basic rules for content generation, MMD adversarial constraint parameters, and shot generation weights. The time-series peak-shifting bidirectional self-correction iterative process includes downlink correction and uplink correction: downlink correction is based on high CATE causal features to fine-tune the generated model parameters; uplink correction is based on real-time content cluster distribution data to correct the three-agent quota game weights.

7. The method according to claim 6, characterized in that, Step S5 is also accompanied by a Gini coefficient adaptive noise injection anti-collapse step: Calculate the Gini coefficient corresponding to the current distribution of all content clusters in real time, and use the Gini coefficient numerical value to quantify the degree of bias generated during the model iteration process; The injection intensity of the anti-noise is adaptively adjusted according to the value of the Gini coefficient: when the Gini coefficient is determined to be slightly biased, a small intensity of perturbation noise is injected; when the Gini coefficient is determined to be heavily biased, a high intensity of perturbation noise is injected, relying on strong perturbation to forcibly expand the overall content diversity.

8. A closed-loop optimization system for batch generation and intelligent delivery of short videos based on AI, characterized in that, include: The population modeling module is used to construct a standardized causal directed acyclic graph (DAG), calculate the conditional average treatment effect (CATE) using a backdoor adjustment algorithm, and divide the population into multiple heterogeneous differential population clusters. The content generation module is used to generate short video materials in batches based on the heterogeneous differential population cluster by utilizing a generation model that embeds the maximum mean difference (MMD) loss function. The game-theoretic delivery module is used to split the total delivery budget into three independent quota resource pools, deploy three independent reinforcement learning agents to carry out budget allocation game optimization and execute dynamic delivery by channel; The reasoning and detection module is used to reuse the causal directed acyclic graph (DAG) for secondary causal decoupling operations and build a causal reasoning knowledge base. The bidirectional self-correction module is used to retrieve feedback data from the causal reasoning knowledge base and execute the time-series peak-shifting bidirectional self-correction iterative process. Each module completes real-time data transmission, interaction, and synchronization through a dedicated causal bidirectional data bus, forming a complete closed-loop data interaction system.

9. A computer device, comprising: A processor, a memory, and a communication bus, wherein the processor and the memory communicate with each other via the communication bus; The memory stores program instructions that can be executed by the processor, which can invoke the program instructions to perform the method provided by the first aspect or any possible implementation thereof.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the closed-loop optimization method for batch generation and intelligent delivery of short videos based on AI as described in any one of claims 1 to 7.

Images