An advertisement putting intelligent optimization system and method based on a cloud platform

Abstract

The application discloses an advertisement putting intelligent optimization system and method based on a cloud platform, relates to the technical field of advertisement putting optimization, quantitatively disassembles a global putting optimization target, constructs a cloud edge end unified feature semantic space, generates a global feature distribution benchmark, collects and semantically maps real-time context features of a putting scene, completes real-time detection and quantification of feature distribution deviation based on the global feature distribution benchmark, completes dynamic matching and loading of an adaptive model based on the quantification result of the feature distribution deviation, inputs the quantitatively disassembled global putting optimization target into the adaptive model, performs checking and correction, generates and executes a cooperative putting decision, completes collection and isolated attribution of putting full-link data on the cloud platform side, realizes correlation tracing of feature distribution deviation and global putting effect loss, and iteratively adjusts the cloud edge end, and the application improves the precision and rationality of advertisement putting decisions through the global putting optimization target.

Description

Technical Field

[0001] This invention relates to the field of advertising optimization technology, specifically to an intelligent advertising optimization system and method based on a cloud platform. Background Technology

[0002] In the field of cloud platform advertising, existing technologies generally suffer from a core technical problem of imbalance between global optimization goals and local terminal scenarios, making it difficult to simultaneously achieve low latency and optimal performance. Traditional advertising methods often employ centralized inference in the cloud or independent edge-based advertising. While the former can ensure global optimization consistency, it suffers from high latency in terminal feature collection and decision response, making it unsuitable for the real-time advertising needs of terminals such as in-vehicle and OTT devices. The latter, while achieving low-latency inference, lacks global feature support from the cloud side, making it prone to distribution shifts between real-time terminal features and global cloud features. This leads to decreased model generalization ability, advertising performance deviating from the global goal, and potential advertising decision biases. Summary of the Invention

[0003] The purpose of this invention is to provide an intelligent optimization system and method for advertising delivery based on a cloud platform, so as to solve the problems mentioned in the background art.

[0004] To address the aforementioned technical problems, this invention provides the following technical solution: a cloud-based intelligent optimization method for ad placement, comprising the following steps: S1: On the cloud platform side, the global deployment optimization target is quantitatively decomposed, a unified feature semantic space of cloud, edge and terminal is constructed, a global feature distribution benchmark is generated, and the global feature distribution benchmark is distributed to edge nodes. S2: For the edge node collaborative terminal side, collect and semantically map the real-time context features of the delivery scenario, and complete the real-time detection and quantization of feature distribution offset based on the global feature distribution benchmark; S3: For edge nodes, based on the quantization results of feature distribution offset, complete the dynamic matching and loading of the adaptation model, and input the global deployment optimization target after quantization decomposition into the adaptation model; S4: Through collaboration between terminals and edge nodes, low-latency model inference is completed, and the delivery decision is verified and corrected in conjunction with the global delivery optimization target, generating and executing collaborative delivery decisions; S5: Complete the collection, isolation, and attribution of full-link data on the cloud platform side to achieve the correlation and source tracing between feature distribution offset and global delivery effect loss; S6: Based on the source tracing results, iteratively adjust the entire cloud-edge-device link, and perform closed-loop iterative optimization of feature semantic space, adaptation model and optimization target decomposition rules.

[0005] By constructing a unified latent semantic embedding space for cloud, edge, and device, the semantic dimension alignment between global user features on the cloud side and contextual features of the terminal scene is achieved. Based on the multi-branch lightweight model library obtained by knowledge distillation and pre-deployed on edge nodes, the optimal model branch is matched according to the offset level. The global delivery optimization objective is decomposed into two levels of quantifiable parameters, namely hard constraint parameters and soft optimization weights, and injected into the model to output collaborative delivery decisions.

[0006] Furthermore, step S1 includes the following: S101: On the cloud platform side, based on the advertiser's omnichannel delivery needs, the global delivery optimization target is broken down into two levels of quantifiable parameters. The first level consists of hard constraints that cannot be broken, including: the advertiser's daily delivery budget threshold, the upper limit of ad exposure frequency per user, and the highest conversion cost threshold per channel. The second level consists of soft optimization weights that can be dynamically adjusted, including: user lifetime value (LTV) weight, new customer acquisition priority weight, cross-channel conversion attribution weight, and long-term brand exposure weight. These two levels of quantifiable parameters are synchronized in real time to the accessed edge nodes. The accessed edge nodes represent edge advertising service nodes that have completed cloud platform identity authentication, are included in the cloud-edge collaborative scheduling system, and can receive strategies and model data issued by the cloud platform. S102: Based on the cloud platform's full-scale global delivery feature library and the terminal's full-scenario historical context feature library, a cross-domain contrastive learning algorithm is used to construct a unified latent semantic embedding space for the cloud, edge, and terminal. This maps the cloud-side global user features and the terminal-side context features to the same semantic dimension. The cloud platform's full-scale global delivery feature library is constructed by the cloud platform's data middle platform after aggregating, cleaning, normalizing, and tagging advertising delivery data from all channels and the entire link. The cloud platform's full-scale global delivery features include: basic user profile features, user historical advertising interaction behavior features, advertising delivery channel features, and advertising material features. The terminal's full-scenario historical context feature library is formed by real-time collection of various advertising delivery terminals (including mobile applications, web clients, offline smart advertising screens, in-vehicle terminals, OTT terminals, etc.) during historical advertising delivery, followed by local preprocessing and uploading to edge nodes for aggregation. The terminal's full-scenario historical context feature library includes: terminal real-time spatiotemporal location features, terminal device status features, and context features of the user's current browsing content. A method to map cloud-side global user features and terminal-side contextual features to the same semantic dimension is as follows: Construct a dual-tower encoding structure, set up a cloud-side feature encoder and a terminal-side contextual feature encoder respectively, input the cloud platform-side global delivery features into the cloud-side feature encoder, and input the terminal-side contextual features into the terminal-side feature encoder, and perform high-dimensional feature encoding respectively. Use cloud-side features and terminal-side features under the same user and the same delivery scenario as positive sample pairs, and use features of different users and different scenarios as negative sample pairs. Train the two encoders by contrastive loss function. After training converges, the feature vectors output by the two encoders are in the same latent semantic metric space, that is, complete the construction of a unified latent semantic embedding space between cloud, edge and terminal, and realize the alignment mapping of cloud-side global user features and terminal-side contextual features under the same semantic dimension. S103: Based on the unified implicit semantic embedding space, generate a cloud-side global feature distribution benchmark library, and pre-distribute the global feature distribution benchmark library to all edge nodes as a reference benchmark for feature distribution offset detection. This system transforms advertisers' omnichannel campaign needs into a clearly structured and actionable optimization framework. By differentiating between rigid constraints and flexible optimization weights, it maintains the core principles of cost control, user experience assurance, and channel efficiency compliance, avoiding resource waste and experience degradation caused by blind campaigning. Simultaneously, it addresses diverse campaign demands such as long-term user value, new customer acquisition, cross-channel conversion, and brand communication, ensuring optimization aligns more closely with actual business needs. Synchronizing relevant parameters to edge nodes unifies optimization goals between the cloud platform and the edge, preventing a disconnect between cloud and edge strategies. This establishes standardized execution guidelines for intelligent ad delivery, enhancing the overall synergy and controllability of the campaign system and facilitating a shift from extensive management to refined intelligent scheduling.

[0007] Furthermore, step S2 includes the following: S201: The terminal side captures the context features of the current advertising scenario in real time, maps the context features to the unified implicit semantic embedding space in real time, generates a standardized feature embedding vector, and uploads it to the corresponding bound edge node. S202: After receiving the feature embedding vector, the edge node, based on the pre-loaded global feature distribution benchmark library, uses a dual-index joint calculation method to obtain the distribution offset of the current real-time feature relative to the global benchmark. The specific method for obtaining the distribution offset of the current real-time feature relative to the global benchmark using the dual-index joint calculation method is as follows: the overall distribution distance between the current real-time feature embedding vector and the global feature benchmark is calculated using MMD to characterize the global distribution difference between the two; the probability distribution difference between the current real-time feature and the global feature benchmark in the core feature dimension is calculated using KL divergence to characterize the local offset degree of the key dimension; then the calculation results of the two indices are weighted and fused (the weights can be dynamically adjusted according to the deployment scenario) to finally obtain the distribution offset of the current real-time feature relative to the global benchmark. S203: Based on the preset offset threshold range, the distribution offset is divided into multiple levels to generate a distribution offset level corresponding to several offset degrees. At the same time, the core feature dimensions and offset magnitude that cause the distribution offset are marked. In the process of calculating the probability distribution difference of each feature dimension through KL divergence, the feature dimension with KL divergence value greater than the preset dimension difference threshold is determined as the core feature dimension. The offset magnitude is calculated based on the KL divergence value corresponding to each core feature dimension and the vector distance normalized with the unified latent semantic embedding space, and synchronized to the model adaptation module of the edge node. By breaking down semantic barriers between cloud-based global features and terminal scene features, unified alignment and standardized expression of multi-source advertising features are achieved, resulting in more complete and consistent feature perception for ad delivery. Real-time collection and standardized mapping of terminal context features enable sensitive capture of dynamic changes in ad delivery scenarios. Detection of real-time feature distribution shifts based on global feature benchmarks accurately identifies scene feature anomalies, pinpointing key dimensions and degrees of shift, providing precise data for subsequent model adaptive adjustments. This effectively avoids inaccurate delivery decisions due to sudden changes in scene features, improves the adaptability of ad delivery to real-time scenarios, ensures stable and reliable delivery decisions, and continuously optimizes the accuracy and overall effectiveness of ad delivery. S301: A multi-branch lightweight adaptation library for the cloud-side global optimization master model is pre-deployed at edge nodes. In this library, model branches retain the core conversion prediction logic and underlying global optimization capabilities of the master model through knowledge distillation. Each branch corresponds to a specific interval's distribution offset and distribution offset level, and lightweight fine-tuning is performed for corresponding feature distribution scenarios. The cloud-side global optimization master model is a core advertising optimization model deployed in the cloud platform data center and trained based on full historical advertising data. The cloud-side global optimization master model aims to achieve optimal global advertising performance, integrating full user features and multi-channel advertising. Features, full-scene context features, and global deployment constraint rules; the method for obtaining model branches is as follows: using the cloud-side global optimization master model as the teacher model, constructing a lightweight neural network structure as the initial student model; through knowledge distillation, transferring the global deployment reasoning logic, transformation prediction distribution, and constraint decision rules of the teacher model to the student model, completing model simplification and core capability inheritance; then, based on the scenario-based sample datasets corresponding to different feature distribution offset degrees, the distilled lightweight student model is specifically fine-tuned to adapt the model to the feature distribution rules of the corresponding offset scenarios, and finally obtaining lightweight model branches adapted to different distribution offset scenarios; S302: Based on the distribution offset degree and distribution offset level, the edge node matches the model branch with the highest adaptability to the current real-time feature distribution and completes the lightweight loading of the corresponding model branch. Specifically, the edge node compares the calculated feature distribution offset degree, distribution offset level and core feature dimension with the preset adaptation scene labels of each model branch in the multi-branch lightweight adaptation library, and matches the model branch with the highest similarity to the current real-time feature distribution according to the offset degree, thus completing the automatic selection of the optimal model branch. S303: Edge nodes inject two levels of quantifiable parameters of the pre-synchronized global deployment optimization target into the matched model branches. Hard constraint parameters are written into the constraint layer of the model inference as hard boundary masks for the inference output, and soft optimization weights are written into the loss function and output layer of the model to adjust the inference optimization direction, so that the adapted model can simultaneously adapt to the real-time feature distribution of the terminal and the global optimization target of the cloud side. To address the challenge of efficiently running large-scale cloud-based advertising models at the edge, a lightweight model branching approach is used to retain the core prediction and optimization capabilities of global advertising while adapting to various feature distribution offset scenarios. By automatically matching the optimal model branch based on scene feature offsets, the responsiveness of edge-side advertising decisions is significantly improved, ensuring smooth model inference. Integrating global advertising constraints and optimization weights into the adapted model allows edge-side inference to both align with real-time terminal scene characteristics and strictly adhere to the overall cloud-based advertising goals, achieving a unification of local scene adaptation and global optimization orientation. This effectively improves the accuracy and stability of advertising decisions, avoids inefficiencies caused by scene adaptation biases, and enhances the adaptive and collaborative optimization capabilities of the entire advertising system.

[0008] Furthermore, step S4 includes the following: S401: The terminal side performs lightweight fine-tuning, loads the model branches and global delivery optimization target parameters issued by the edge nodes, inputs standardized feature embedding vectors to complete local inference, and generates an initial delivery decision. The initial delivery decision includes the advertising candidate ranking result, the estimated delivery bid value, the user exposure matching result, the delivery channel allocation suggestion, and the delivery conversion effect estimate. S402: The terminal synchronizes the initial deployment decision to the global constraint verification module of the edge node in real time. The verification module verifies whether the initial deployment decision exceeds the preset threshold based on hard constraint parameters. S403: If the initial delivery decision does not exceed the hard constraint threshold, the edge node directly sends the initial delivery decision as a collaborative delivery decision to the terminal for execution; if the initial delivery decision exceeds the hard constraint threshold, the edge node performs gradient correction on the initial delivery decision based on soft optimization weights, generates a collaborative delivery decision that takes into account both the local scene conversion potential and the cloud-side global optimization goal, sends it to the terminal for execution, and records the decision correction trajectory and corresponding parameters at the same time. Lightweight inference enables rapid generation of initial ad delivery decisions, significantly improving ad response speed and real-time scenario adaptation. Global constraint verification via edge nodes promptly intercepts unauthorized ad delivery behavior, firmly upholding the core bottom lines of cost control and user experience. Gradual correction of decisions exceeding constraints effectively balances local scenario conversion potential with global cloud optimization goals, ensuring that delivery decisions align with both real-time scenario needs and the overall delivery strategy. Simultaneous retention of decision correction trajectories provides detailed evidence for subsequent optimization iterations, significantly improving the compliance and scientific rigor of ad delivery decisions. This enables efficient implementation of cloud-edge-device collaborative decision-making, continuously optimizing delivery effectiveness and user interaction experience.

[0009] Furthermore, step S5 includes the following: S501: Cloud platform-side full-link data middleware, synchronously collects real-time feature distribution data, distribution offset level, delivery decision data and decision correction trajectory from the terminal side, corresponding full-link effect data of delivery, and global delivery operation data from the cloud side. S502: Employing a dual-difference model, a benchmark control group with the same user segment, delivery environment, delivery time, and no feature distribution shift is used to isolate and attribute the performance data of this delivery. Specifically, this isolation and attribution involves: based on the terminal real-time context feature embedding vector, core feature dimensions, and distribution shift level corresponding to this delivery, historical delivery data, cross-channel interference factors, and external environmental variables not related to this decision are eliminated. Combining the conversion prediction logic of the lightweight model branch with the two-level quantifiable parameter constraints of the global delivery, the actual delivery performance data is decomposed into effective results directly contributed to this delivery decision and irrelevant interference items, yielding local conversion increments and distribution shifts. The resulting global performance loss and the core quantitative indicators brought about by global constraint correction are specifically as follows: Based on the feature vector distance, distribution offset, and core feature dimension offset magnitude in the unified latent semantic embedding space, and taking the non-offset baseline inference result of the cloud-side global optimization main model as a reference, irrelevant interference variables are first removed and the independent inference contribution of the current adapted model branch is extracted, and the local transformation increment is obtained by splitting. Then, the feature distribution offset deviation quantified by MMD and KL is compared with the global baseline effect and the actual edge inference effect to calculate the global performance loss caused by distribution offset. Finally, the difference between the delivery decision output before and after injecting the two-level quantifiable parameters of the global system into the model is extracted, and the core quantitative indicators brought about by global constraint correction are obtained by normalization calculation. S503: Based on the attribution results, locate the core offset feature dimensions, model adaptation defects and constraint parameter adaptation problems that cause global effect loss, generate an effect fluctuation source tracing report, and clarify the optimization direction and priority; It aggregates end-to-end campaign data, comprehensively retaining information on feature distribution, decision execution, and performance feedback. Through scientific attribution methods, it eliminates external interference, accurately breaking down the true components of campaign performance, clearly distinguishing the actual contribution of decisions, performance loss due to feature shifts, and the optimization value brought by constraint adjustments. Based on the attribution results, it quickly locates the root cause of performance fluctuations, pinpointing problems in feature dimensions, model adaptation, and parameter configuration, generating a clear source-tracing report. This makes campaign performance evaluation more objective and reliable, pointing the way and prioritizing subsequent optimizations, helping the advertising campaign system achieve precise iteration, and continuously improving overall campaign efficiency.

[0010] Furthermore, step S6 includes the following: S601: Based on the effect fluctuation traceability report, supplement cross-domain comparative learning training samples for the core offset feature dimension, iteratively optimize the unified latent semantic embedding space of cloud, edge and terminal, and reduce the feature distribution offset probability of similar scenarios. S602: For model branches corresponding to distribution offset scenarios where the global effect loss exceeds the preset threshold, call the offline computing power of the cloud platform, perform incremental fine-tuning and knowledge distillation based on the newly added full-link delivery data, and after optimization, synchronously update the multi-branch lightweight adaptation library to the edge node. S603: Based on the delivery effect data of terminal scenarios and distribution offset levels, dynamically adjust the distribution threshold of hard constraint parameters and the allocation ratio of soft optimization weights. S604: The optimized latent semantic embedding space, model branches, optimized target decomposition rules, and distribution offset detection parameters are synchronously updated to the corresponding cloud-side, edge node, and terminal-side modules to complete the full-link closed-loop iteration.

[0011] A cloud-based intelligent optimization system for advertising delivery includes: an advertising target benchmark module, an edge feature detection module, an adaptation model matching module, a decision verification and correction module, an effect attribution and tracing module, and a full-chain iterative optimization module. The advertising target benchmark module is used to quantify and decompose the global delivery optimization target on the cloud platform side, construct a unified feature semantic space for cloud, edge and terminal, generate a global feature distribution benchmark, and distribute the global feature distribution benchmark to edge nodes. The edge feature detection module is used to collect and semantically map real-time context features of the delivery scenario for the edge node collaborative terminal side, and to complete the real-time detection and quantization of feature distribution offset based on the global feature distribution benchmark. The adaptation model matching module is used to dynamically match and load the adaptation model for edge nodes based on the quantization results of feature distribution offset, and input the global deployment optimization target after quantization decomposition into the adaptation model. The decision verification and correction module is used to complete low-latency model inference through collaboration between the terminal and edge nodes, and to verify and correct the delivery decision in conjunction with the global delivery optimization objective, thereby generating and executing collaborative delivery decisions. The effect attribution and tracing module is used to complete the collection and isolation attribution of the entire delivery chain data on the cloud platform side, and realize the correlation tracing between feature distribution offset and global delivery effect loss; The full-chain iterative optimization module is used to iteratively adjust the entire cloud-edge-device link based on the source tracing results, and to perform closed-loop iterative optimization of the feature semantic space, adaptation model and optimization target decomposition rules.

[0012] Compared with existing technologies, the beneficial effects achieved by this invention are as follows: On the one hand, it solves the technical problem of the offset between terminal features and global features in cloud-edge advertising. By constructing a unified feature semantic space between cloud, edge, and terminal, it achieves semantic alignment between global features in the cloud and features in the terminal scene. Relying on the accurate detection and quantification of feature distribution offset, coupled with a multi-branch lightweight adaptation model, the edge model can quickly match the feature distribution patterns of different scenarios. This not only retains the low latency advantage of edge inference but also eliminates the model adaptation deviation caused by scene differences, significantly improving the generalization ability of the advertising model in various terminal scenarios and ensuring the coordinated unity of global advertising optimization goals and local scene conversion needs.

[0013] On the one hand, by breaking down the global campaign optimization goals into tiered and quantifiable parameters, the system enables refined control and flexible correction of campaign decisions. Relying on the rigid verification of hard constraints and the gradient adjustment of soft optimization weights, it not only prevents violations of campaign decisions that exceed operational thresholds, but also fully taps the conversion potential of terminal scenarios. This allows the final campaign decisions to balance global compliance with local effectiveness, changing the problem of disconnect between global control and edge execution in traditional campaigns and improving the accuracy and rationality of advertising campaign decisions.

[0014] On the other hand, a mechanism for isolating and attributing the effectiveness of ad placements and a closed-loop iteration mechanism for the entire process have been constructed. This mechanism can accurately locate the performance loss caused by factors such as feature offset and model defects. After clarifying the optimization direction, the feature semantic space, adaptation model and parameter rules are continuously iterated and optimized to form a complete closed loop from ad placement execution to performance tracing and system upgrade. This fundamentally improves the problem of ad placement performance fluctuations, continuously enhances the stability and optimization efficiency of the cloud-edge collaborative ad placement system, and enables the ad placement system to adapt to various scenario changes, thus ensuring the overall effectiveness of ad placement in the long term. Attached Figure Description

[0015] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a structural diagram of an intelligent optimization system for advertising delivery based on a cloud platform, according to the present invention. Figure 2 This is a flowchart of an intelligent optimization method for advertising delivery based on a cloud platform, according to the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Please see Figure 1 and Figure 2 This invention provides a technical solution: a cloud-based intelligent optimization method for advertising delivery, comprising the following steps: S1: On the cloud platform side, the global deployment optimization target is quantitatively decomposed, a unified feature semantic space of cloud, edge and terminal is constructed, a global feature distribution benchmark is generated, and the global feature distribution benchmark is distributed to edge nodes. S2: For the edge node collaborative terminal side, collect and semantically map the real-time context features of the delivery scenario, and complete the real-time detection and quantization of feature distribution offset based on the global feature distribution benchmark; S3: For edge nodes, based on the quantization results of feature distribution offset, complete the dynamic matching and loading of the adaptation model, and input the global deployment optimization target after quantization decomposition into the adaptation model; S4: Through collaboration between terminals and edge nodes, low-latency model inference is completed, and the delivery decision is verified and corrected in conjunction with the global delivery optimization target, generating and executing collaborative delivery decisions; S5: Complete the collection, isolation, and attribution of full-link data on the cloud platform side to achieve the correlation and source tracing between feature distribution offset and global delivery effect loss; S6: Based on the source tracing results, iteratively adjust the entire cloud-edge-device link, and perform closed-loop iterative optimization of feature semantic space, adaptation model and optimization target decomposition rules.

[0018] By constructing a unified latent semantic embedding space for cloud, edge, and device, the semantic dimension alignment between global user features on the cloud side and contextual features of the terminal scene is achieved. Based on the multi-branch lightweight model library obtained by knowledge distillation and pre-deployed on edge nodes, the optimal model branch is matched according to the offset level. The global delivery optimization objective is decomposed into two levels of quantifiable parameters, namely hard constraint parameters and soft optimization weights, and injected into the model to output collaborative delivery decisions.

[0019] Step S1 includes the following: S101: On the cloud platform side, based on the advertiser's omnichannel delivery needs, the global delivery optimization target is broken down into two levels of quantifiable parameters. The first level consists of hard constraints that cannot be broken, including: the advertiser's daily delivery budget threshold, the upper limit of ad exposure frequency per user, and the highest conversion cost threshold per channel. The second level consists of soft optimization weights that can be dynamically adjusted, including: user lifetime value (LTV) weight, new customer acquisition priority weight, cross-channel conversion attribution weight, and long-term brand exposure weight. These two levels of quantifiable parameters are synchronized in real time to the accessed edge nodes. The accessed edge nodes represent edge advertising service nodes that have completed cloud platform identity authentication, are included in the cloud-edge collaborative scheduling system, and can receive strategies and model data issued by the cloud platform. S102: Based on the cloud platform's full-scale global delivery feature library and the terminal's full-scenario historical context feature library, a cross-domain contrastive learning algorithm is used to construct a unified latent semantic embedding space for the cloud, edge, and terminal. This maps the cloud-side global user features and the terminal-side context features to the same semantic dimension. The cloud platform's full-scale global delivery feature library is constructed by the cloud platform's data middle platform after aggregating, cleaning, normalizing, and tagging advertising delivery data from all channels and the entire link. The cloud platform's full-scale global delivery features include: basic user profile features, user historical advertising interaction behavior features, advertising delivery channel features, and advertising material features. The terminal's full-scenario historical context feature library is formed by real-time collection of various advertising delivery terminals (including mobile applications, web clients, offline smart advertising screens, in-vehicle terminals, OTT terminals, etc.) during historical advertising delivery, followed by local preprocessing and uploading to edge nodes for aggregation. The terminal's full-scenario historical context feature library includes: terminal real-time spatiotemporal location features, terminal device status features, and context features of the user's current browsing content. A method for mapping cloud-side global user features and terminal-side contextual features to the same semantic dimension is as follows: Construct a dual-tower encoding structure, setting up a cloud-side feature encoder and a terminal-side contextual feature encoder respectively. Input the cloud platform-side global delivery features into the cloud-side feature encoder and the terminal-side contextual features into the terminal-side feature encoder, and perform high-dimensional feature encoding respectively. Use cloud-side features and terminal-side features under the same user and the same delivery scenario as positive sample pairs, and use features of different users and different scenarios as negative sample pairs. Train the two encoders through a contrastive loss function. After training converges, the feature vectors output by the two encoders are in the same latent semantic metric space, that is, complete the construction of a unified latent semantic embedding space between cloud, edge, and terminal, and realize the alignment mapping of cloud-side global user features and terminal-side contextual features under the same semantic dimension.

[0020] Step S2 includes the following: S201: The terminal side captures the context features of the current advertising scenario in real time, maps the context features to the unified implicit semantic embedding space in real time, generates a standardized feature embedding vector, and uploads it to the corresponding bound edge node. S202: After receiving the feature embedding vector, the edge node, based on the pre-loaded global feature distribution benchmark library, uses a dual-index joint calculation method to obtain the distribution offset of the current real-time feature relative to the global benchmark. The specific method for obtaining the distribution offset of the current real-time feature relative to the global benchmark using the dual-index joint calculation method is as follows: the overall distribution distance between the current real-time feature embedding vector and the global feature benchmark is calculated using MMD to characterize the global distribution difference between the two; the probability distribution difference between the current real-time feature and the global feature benchmark in the core feature dimension is calculated using KL divergence to characterize the local offset degree of the key dimension; then the calculation results of the two indices are weighted and fused (the weights can be dynamically adjusted according to the deployment scenario) to finally obtain the distribution offset of the current real-time feature relative to the global benchmark. S203: Based on a preset offset threshold range, the distribution offset is divided into multiple levels to generate several distribution offset levels. At the same time, the core feature dimensions and offset magnitudes that cause the distribution offset are marked. In the process of calculating the probability distribution difference of each feature dimension through KL divergence, the feature dimensions whose KL divergence value is greater than the preset dimension difference threshold are determined as core feature dimensions. The offset magnitude is calculated based on the KL divergence value corresponding to each core feature dimension and the vector distance normalized with the unified latent semantic embedding space, and synchronized to the model adaptation module of the edge node.

[0021] Step S3 includes the following: S301: A multi-branch lightweight adaptation library for the cloud-side global optimization master model is pre-deployed at edge nodes. In this library, model branches retain the core conversion prediction logic and underlying global optimization capabilities of the master model through knowledge distillation. Each branch corresponds to a specific interval's distribution offset and distribution offset level, and lightweight fine-tuning is performed for corresponding feature distribution scenarios. The cloud-side global optimization master model is a core advertising optimization model deployed in the cloud platform data center and trained based on full historical advertising data. The cloud-side global optimization master model aims to achieve optimal global advertising performance, integrating full user features and multi-channel advertising. Features, full-scene context features, and global deployment constraint rules; the method for obtaining model branches is as follows: using the cloud-side global optimization master model as the teacher model, constructing a lightweight neural network structure as the initial student model; through knowledge distillation, transferring the global deployment reasoning logic, transformation prediction distribution, and constraint decision rules of the teacher model to the student model, completing model simplification and core capability inheritance; then, based on the scenario-based sample datasets corresponding to different feature distribution offset degrees, the distilled lightweight student model is specifically fine-tuned to adapt the model to the feature distribution rules of the corresponding offset scenarios, and finally obtaining lightweight model branches adapted to different distribution offset scenarios; S302: Based on the distribution offset degree and distribution offset level, the edge node matches the model branch with the highest adaptability to the current real-time feature distribution and completes the lightweight loading of the corresponding model branch. Specifically, the edge node compares the calculated feature distribution offset degree, distribution offset level and core feature dimension with the preset adaptation scene labels of each model branch in the multi-branch lightweight adaptation library, and matches the model branch with the highest similarity to the current real-time feature distribution according to the offset degree, thus completing the automatic selection of the optimal model branch. S303: Edge nodes inject two levels of quantifiable parameters of the pre-synchronized global deployment optimization target into the matched model branches. Hard constraint parameters are written into the constraint layer of the model inference as hard boundary masks for the inference output, and soft optimization weights are written into the loss function and output layer of the model to adjust the inference optimization direction, so that the adapted model can simultaneously adapt to the real-time feature distribution of the terminal and the global optimization target of the cloud side.

[0022] Step S4 includes the following: S401: The terminal side performs lightweight fine-tuning, loads the model branches and global delivery optimization target parameters issued by the edge nodes, inputs standardized feature embedding vectors to complete local inference, and generates an initial delivery decision. The initial delivery decision includes the advertising candidate ranking result, the estimated delivery bid value, the user exposure matching result, the delivery channel allocation suggestion, and the delivery conversion effect estimate. S402: The terminal synchronizes the initial deployment decision to the global constraint verification module of the edge node in real time. The verification module verifies whether the initial deployment decision exceeds the preset threshold based on hard constraint parameters. S403: If the initial delivery decision does not exceed the hard constraint threshold, the edge node directly sends the initial delivery decision as a collaborative delivery decision to the terminal for execution; if the initial delivery decision exceeds the hard constraint threshold, the edge node performs gradient correction on the initial delivery decision based on soft optimization weights, generates a collaborative delivery decision that takes into account both the local scene conversion potential and the cloud-side global optimization goal, sends it to the terminal for execution, and records the decision correction trajectory and corresponding parameters.

[0023] Step S5 includes the following: S501: Cloud platform-side full-link data middleware, synchronously collects real-time feature distribution data, distribution offset level, delivery decision data and decision correction trajectory from the terminal side, corresponding full-link effect data of delivery, and global delivery operation data from the cloud side. S502: Employing a dual-difference model, a benchmark control group with the same user segment, delivery environment, delivery time, and no feature distribution shift is used to isolate and attribute the performance data of this delivery. Specifically, this isolation and attribution involves: based on the terminal real-time context feature embedding vector, core feature dimensions, and distribution shift level corresponding to this delivery, historical delivery data, cross-channel interference factors, and external environmental variables not related to this decision are eliminated. Combining the conversion prediction logic of the lightweight model branch with the two-level quantifiable parameter constraints of the global delivery, the actual delivery performance data is decomposed into effective results directly contributed to this delivery decision and irrelevant interference items, yielding local conversion increments and distribution shifts. The resulting global performance loss and the core quantitative indicators brought about by global constraint correction are specifically as follows: Based on the feature vector distance, distribution offset, and core feature dimension offset magnitude in the unified latent semantic embedding space, and taking the non-offset baseline inference result of the cloud-side global optimization main model as a reference, irrelevant interference variables are first removed and the independent inference contribution of the current adapted model branch is extracted, and the local transformation increment is obtained by splitting. Then, the feature distribution offset deviation quantified by MMD and KL is compared with the global baseline effect and the actual edge inference effect to calculate the global performance loss caused by distribution offset. Finally, the difference between the delivery decision output before and after injecting the two-level quantifiable parameters of the global system into the model is extracted, and the core quantitative indicators brought about by global constraint correction are obtained by normalization calculation. S503: Based on the attribution results, locate the core offset feature dimensions, model adaptation defects, and constraint parameter adaptation problems that cause global effect loss, generate an effect fluctuation tracing report, and clarify the optimization direction and priority.

[0024] Step S6 includes the following: S601: Based on the effect fluctuation traceability report, supplement cross-domain comparative learning training samples for the core offset feature dimension, iteratively optimize the unified latent semantic embedding space of cloud, edge and terminal, and reduce the feature distribution offset probability of similar scenarios. S602: For model branches corresponding to distribution offset scenarios where the global effect loss exceeds the preset threshold, call the offline computing power of the cloud platform, perform incremental fine-tuning and knowledge distillation based on the newly added full-link delivery data, and after optimization, synchronously update the multi-branch lightweight adaptation library to the edge node. S603: Based on the delivery effect data of terminal scenarios and distribution offset levels, dynamically adjust the distribution threshold of hard constraint parameters and the allocation ratio of soft optimization weights. S604: The optimized latent semantic embedding space, model branches, optimized target decomposition rules, and distribution offset detection parameters are synchronously updated to the corresponding cloud-side, edge node, and terminal-side modules to complete the full-link closed-loop iteration.

[0025] A cloud-based intelligent optimization system for advertising delivery includes: an advertising target benchmark module, an edge feature detection module, an adaptation model matching module, a decision verification and correction module, an effect attribution and tracing module, and a full-chain iterative optimization module. The advertising target benchmark module is used to quantify and decompose the global delivery optimization target on the cloud platform side, construct a unified feature semantic space for cloud, edge and terminal, generate a global feature distribution benchmark, and distribute the global feature distribution benchmark to edge nodes. The edge feature detection module is used to collect and semantically map real-time contextual features of the deployment scenario for edge nodes and collaborative terminals, and to complete the real-time detection and quantization of feature distribution offset based on the global feature distribution benchmark. The adaptation model matching module is used to dynamically match and load the adaptation model for edge nodes based on the quantization results of feature distribution offset, and input the global deployment optimization target after quantization decomposition into the adaptation model. The decision verification and correction module is used to complete low-latency model inference through collaboration between the terminal and edge nodes, and to verify and correct the delivery decision in conjunction with the global delivery optimization target, thereby generating and executing collaborative delivery decisions. The effect attribution and tracing module is used to complete the collection and isolation attribution of data across the entire delivery process on the cloud platform side, and to realize the correlation and tracing of feature distribution offset and global delivery effect loss. The full-chain iterative optimization module is used to iteratively adjust the entire cloud-edge-device link based on the source tracing results, and to perform closed-loop iterative optimization of the feature semantic space, adaptation model and optimization target decomposition rules.

[0026] Example 1: The cloud platform first breaks down the global campaign optimization objective into two types of parameters based on the advertiser's omnichannel campaign requirements: rigid constraints and flexible optimization weights. Rigid constraints control campaign costs, user exposure experience, and channel conversion cost limits, while flexible optimization weights consider diverse needs such as long-term user value, new customer acquisition, cross-channel conversion, and brand communication. Both types of parameters are synchronized in real-time to the edge advertising service nodes integrated into the collaborative system. Subsequently, the platform integrates the full-scale campaign features from the cloud side with multi-scenario contextual features from the terminal. Through cross-domain comparative learning, a unified feature semantic space is constructed, achieving alignment and mapping between global user features and real-time terminal scenario features, laying a unified foundation for subsequent feature detection and decision-making inference. Terminal devices collect the contextual features of the current advertising scenario in real-time, map them to the unified semantic space to generate standardized feature vectors, and upload them to the bound edge nodes. The edge nodes, based on the pre-loaded global feature benchmark, detect the distribution offset of real-time features, classify the offset levels, and locate the core feature dimensions and offset magnitudes that cause the offset, providing accurate basis for model adaptation. Edge nodes pre-store a multi-branch lightweight model library that has undergone knowledge distillation. Each model branch retains the core reasoning capabilities of the globally optimized model and adapts to different feature offset scenarios. Nodes automatically match the optimal model branch based on the feature offset level, injecting two levels of delivery optimization parameters into the model. This ensures that the model adapts to the real-time scenario features of the terminal while adhering to the global optimization goals of the cloud platform, guaranteeing the synergy of the reasoning logic. The terminal loads the adapted model, completes local reasoning, and generates initial delivery decisions such as ad ranking, bid estimation, and user matching, which are then uploaded to the edge nodes for constraint verification. Decisions that do not break rigid constraints are directly executed, while decisions that exceed constraints are gradient-corrected using flexible optimization weights to form a final delivery decision that balances local scenarios and global goals, while simultaneously recording the entire decision correction process. The cloud platform's data platform synchronously collects feature distribution, delivery decisions, correction trajectories, and end-to-end delivery performance data. Through scientific isolation and attribution methods, external interference is eliminated, the true contribution and loss of delivery performance are broken down, and issues related to feature offset, model adaptation, and parameter configuration are identified, generating a performance fluctuation tracing report. Finally, the system optimizes the feature semantic space based on the source tracing report, incrementally fine-tunes the model branches with poor performance, dynamically adjusts the placement constraint thresholds and optimizes the weight ratios, and synchronously updates the optimized feature space, model branches, and parameter rules to all modules in the cloud, edge, and terminal, forming a closed-loop iteration across the entire chain. Through the collaborative cooperation of cloud, edge, and terminal, the system achieves real-time scenario adaptation, global target control, and continuous iterative optimization for ad placement, effectively improving the accuracy and conversion efficiency of ad placement and driving the transformation of ad placement from extensive operation to refined intelligent optimization.

[0027] Example 2: The terminal side collects contextual features such as spatiotemporal location, device status, and browsing content of the current deployment scene in real time, maps them to a unified latent semantic embedding space, generates a standardized feature embedding vector with dimension 256, and uploads it to the bound edge node. The edge node initiates a dual-indicator joint calculation based on a pre-loaded global feature distribution benchmark library: using the maximum mean difference algorithm, the overall distribution distance between the real-time feature embedding vector and the global benchmark is found to be 0.28; the probability distribution difference of each feature dimension is calculated using KL divergence, where the KL divergence of the user's real-time location feature is 0.51, the KL divergence of the scene content preference feature is 0.47, and the KL divergence of the other dimensions is below a preset threshold. Subsequently, a weighted fusion is performed with an MMD weight of 0.4 and a core dimension average KL divergence weight of 0.6, resulting in a distribution offset of 0.406, which is determined to be a moderate offset level. The normalized offset amplitudes of the core dimensions are 0.48 (location feature) and 0.36 (content preference feature), with an average offset amplitude of 0.42. This result is synchronized to the model adaptation module.

[0028] In the multi-branch lightweight model library pre-deployed at edge nodes, each branch inherits the core capabilities of the global main model through knowledge distillation. Nodes match the lightweight model branch with the highest suitability based on the moderate offset level and core dimensions. Subsequently, hard constraint parameters such as the advertiser's daily budget cap and the single-user daily exposure frequency cap are written into the model constraint layer. The user lifecycle value weight, new customer acquisition priority weight, cross-channel attribution weight, and long-term brand exposure weight are injected into the model loss function and output layer in a ratio of 0.3:0.4:0.2:0.1.

[0029] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary sensing device embodiments described above, and that the invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A cloud-based intelligent optimization method for ad delivery, characterized in that: The method includes the following steps: S1: On the cloud platform side, the global deployment optimization target is quantitatively decomposed, a unified feature semantic space of cloud, edge and terminal is constructed, a global feature distribution benchmark is generated, and the global feature distribution benchmark is distributed to edge nodes. S2: For the edge node collaborative terminal side, collect and semantically map the real-time context features of the delivery scenario, and complete the real-time detection and quantization of feature distribution offset based on the global feature distribution benchmark; S3: For edge nodes, based on the quantization results of feature distribution offset, complete the dynamic matching and loading of the adaptation model, and input the global deployment optimization target after quantization decomposition into the adaptation model; S4: Through collaboration between terminals and edge nodes, low-latency model inference is completed, and the delivery decision is verified and corrected in conjunction with the global delivery optimization target, generating and executing collaborative delivery decisions; S5: Complete the collection, isolation, and attribution of full-link data on the cloud platform side to achieve the correlation and source tracing between feature distribution offset and global delivery effect loss; S6: Based on the source tracing results, iteratively adjust the entire cloud-edge-device link, and perform closed-loop iterative optimization of feature semantic space, adaptation model and optimization target decomposition rules.

2. The intelligent optimization method for advertising delivery based on a cloud platform according to claim 1, characterized in that: By constructing a unified latent semantic embedding space for cloud, edge, and device, the semantic dimension alignment between global user features on the cloud side and contextual features of the terminal scene is achieved. Based on the multi-branch lightweight model library obtained by knowledge distillation and pre-deployed on edge nodes, the optimal model branch is matched according to the offset level. The global delivery optimization objective is decomposed into two levels of quantifiable parameters, namely hard constraint parameters and soft optimization weights, and injected into the model to output collaborative delivery decisions.

3. The intelligent optimization method for advertising delivery based on a cloud platform according to claim 2, characterized in that: In step S1, Includes the following: S101: On the cloud platform side, based on the advertiser's omnichannel delivery needs, the global delivery optimization target is decomposed into two levels of quantifiable parameters. The first level is the hard constraint parameter that cannot be broken by the advertisement, and the second level is the soft optimization weight that can be dynamically adjusted by the advertisement. The two levels of quantifiable parameters are synchronized to the access edge nodes in real time. S102: Based on the full global deployment feature library on the cloud platform and the full-scenario historical context feature library on the terminal, a cross-domain contrastive learning algorithm is used to construct a unified latent semantic embedding space for cloud, edge and terminal, mapping the global user features on the cloud side and the context features on the terminal side to the same semantic dimension. S103: Based on the unified implicit semantic embedding space, generate a global feature distribution benchmark library for the cloud side, and pre-distribute the global feature distribution benchmark library to all edge nodes.

4. The intelligent optimization method for advertising delivery based on a cloud platform according to claim 3, characterized in that: In step S2, Includes the following: S201: The terminal side captures the context features of the current advertising scenario in real time, maps the context features to the unified implicit semantic embedding space in real time, generates a standardized feature embedding vector, and uploads it to the corresponding bound edge node. S202: After receiving the feature embedding vector, the edge node uses a dual-index joint calculation method to obtain the distribution offset of the current real-time feature relative to the global benchmark, based on the preloaded global feature distribution benchmark library. S203: Based on the preset offset threshold range, the distribution offset is divided into multiple levels, generating a distribution offset level corresponding to several offset degrees. At the same time, the core feature dimension and offset magnitude that cause the distribution offset are marked and synchronized to the model adaptation module of the edge node.

5. The intelligent optimization method for advertising delivery based on a cloud platform according to claim 4, characterized in that: In step S3, Includes the following: S301: A multi-branch lightweight adaptation library for pre-deploying the cloud-side global optimization main model at edge nodes. In the multi-branch lightweight adaptation library, the model branches retain the core transformation prediction logic and global optimization underlying capabilities of the global optimization main model through knowledge distillation. The model branches correspond to the distribution offset degree and distribution offset level of the interval respectively, and lightweight fine-tuning is completed for the corresponding feature distribution scenarios. S302: Edge nodes are matched with the model branch that best fits the current real-time feature distribution based on the distribution offset degree and distribution offset level, and the corresponding model branch is loaded in a lightweight manner. S303: Edge nodes inject the pre-synchronized global deployment optimization target's two-level quantifiable parameters into the matched model branches. Hard constraint parameters are written into the model's inference constraint layer as hard boundary masks for inference output, while soft optimization weights are written into the model's loss function and output layer to adjust the inference optimization direction.

6. The intelligent optimization method for advertising delivery based on a cloud platform according to claim 5, characterized in that: Step S4 includes the following: S401: The terminal side performs lightweight fine-tuning, loads the model branches and global deployment optimization target parameters issued by the edge nodes, inputs standardized feature embedding vectors to complete local inference, and generates the initial deployment decision; S402: The terminal synchronizes the initial deployment decision to the global constraint verification module of the edge node in real time. The verification module verifies the relationship between the initial deployment decision and the preset threshold based on hard constraint parameters. S403: If the initial deployment decision does not exceed the hard constraint threshold, the edge node directly sends the initial deployment decision as a collaborative deployment decision to the terminal for execution; if the initial deployment decision exceeds the hard constraint threshold, the edge node performs gradient correction on the initial deployment decision based on soft optimization weights, generates a collaborative deployment decision, sends it to the terminal for execution, and records the decision correction trajectory and corresponding parameters.

7. The intelligent optimization method for advertising delivery based on a cloud platform according to claim 6, characterized in that: In step S5, Includes the following: S501: Cloud platform-side full-link data middleware, synchronously collects real-time feature distribution data, distribution offset level, delivery decision data and decision correction trajectory from the terminal side, corresponding full-link effect data of delivery, and global delivery operation data from the cloud side. S502: Using a dual difference model, the same user group, same delivery environment, same delivery time, and no characteristic distribution shift are used as the benchmark control group to isolate and attribute the performance data of this delivery, and break it down to obtain the core quantitative indicators of local conversion increment, global performance loss caused by distribution shift, and global constraint correction. S503: Based on the attribution results, locate the core offset feature dimensions, model adaptation defects, and constraint parameter adaptation problems that lead to the loss of global effect, and generate an effect fluctuation source tracing report.

8. The intelligent optimization method for advertising delivery based on a cloud platform according to claim 7, characterized in that: In step S6, Includes the following: S601: Based on the effect fluctuation traceability report, supplement cross-domain comparative learning training samples for the core offset feature dimension, and iteratively optimize the unified latent semantic embedding space of cloud, edge and terminal. S602: For model branches corresponding to distribution offset scenarios where the global effect loss exceeds the preset threshold, call the offline computing power of the cloud platform, perform incremental fine-tuning and knowledge distillation based on the newly added full-link delivery data, and after optimization, synchronously update the multi-branch lightweight adaptation library to the edge node. S603: Based on the delivery effect data of terminal scenarios and distribution offset levels, dynamically adjust the distribution threshold of hard constraint parameters and the allocation ratio of soft optimization weights. S604: The optimized latent semantic embedding space, model branches, optimized target decomposition rules, and distribution offset detection parameters are synchronously updated to the corresponding cloud-side, edge node, and terminal-side modules to complete the full-link closed-loop iteration.

9. A cloud-based intelligent advertising optimization system, wherein the system is applied to the cloud-based intelligent advertising optimization method according to any one of claims 1-8, characterized in that: The system includes: an advertising target benchmark module, an edge feature detection module, an adaptation model matching module, a decision verification and correction module, an effect attribution and tracing module, and a full-chain iterative optimization module; The advertising target benchmark module is used to quantify and decompose the global delivery optimization target on the cloud platform side, construct a unified feature semantic space for cloud, edge and terminal, generate a global feature distribution benchmark, and distribute the global feature distribution benchmark to edge nodes. The edge feature detection module is used to collect and semantically map real-time context features of the delivery scenario for the edge node collaborative terminal side, and to complete the real-time detection and quantization of feature distribution offset based on the global feature distribution benchmark. The adaptation model matching module is used to dynamically match and load the adaptation model for edge nodes based on the quantization results of feature distribution offset, and input the global deployment optimization target after quantization decomposition into the adaptation model. The decision verification and correction module is used to complete low-latency model inference through collaboration between the terminal and edge nodes, and to verify and correct the delivery decision in conjunction with the global delivery optimization objective, thereby generating and executing collaborative delivery decisions. The effect attribution and tracing module is used to complete the collection and isolation attribution of the entire delivery chain data on the cloud platform side, and realize the correlation tracing between feature distribution offset and global delivery effect loss; The full-chain iterative optimization module is used to iteratively adjust the entire cloud-edge-device link based on the source tracing results, and to perform closed-loop iterative optimization of the feature semantic space, adaptation model, and optimization target decomposition rules.

Images