Intelligent delivery decision-making method and system based on multi-dimensional index association

Abstract

The invention relates to an intelligent delivery decision-making method and system based on multi-dimensional index association, and the method comprises the following steps: S1, collecting user, advertisement and context multi-dimensional data, and carrying out the preprocessing of the data to generate standardized features; s2, performing fusion calculation on the multi-dimensional standardized features, extracting key features, and constructing and generating a user-advertisement-context joint feature set; s3, constructing a multi-task prediction model, and training according to a user-advertisement-context joint feature set; s4, according to a prediction result and real-time features of the trained multi-task prediction model, rapidly matching an optimal advertisement for a given user in a real-time bidding process; and S5, performing causal analysis according to the exposure / click log of the optimal advertisement, verifying the real effect of the advertisement, correcting the index, and feeding back to the feature engineering in the S2 and the model training step in the S3 according to the corrected index. According to the invention, the advertisement putting efficiency and effect are effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent advertising delivery, and in particular to an intelligent delivery decision-making method and system based on the correlation of multi-dimensional indicators. Background Art

[0002] With the rapid rise of digital content platforms like short dramas and online novels, advertising is becoming a crucial monetization tool within the content ecosystem. Unlike e-commerce and information platforms, advertising decisions for short drama and novel consumption scenarios are more complex, requiring consideration of both the volatile nature of user interests and the relevance of the content context and diverse creatives. Existing advertising systems primarily rely on simple single metrics like click-through rate or conversion rate for prediction and optimization, facing the following technical challenges: Traditional systems focus on short-term ad clicks or conversions, failing to capture the long-term impact of ads on deeper user behaviors (such as retention, paid unlocks, and social media engagement). This results in high customer acquisition costs for advertisers, weak user stickiness, and limited monetization efficiency. Ad placement decisions require simultaneous understanding and integration of user history (e.g., follow-up, reading preferences, and interaction frequency), ad creative attributes (genre, style, and creative), and content context (e.g., plot points, chapter progression, protagonist positioning, and setting atmosphere). Traditional methods often struggle to dynamically correlate and deeply mine these multi-dimensional features, resulting in poorly relevant ad recommendations and low traffic utilization. Summary of the Invention

[0003] In order to solve the above problems, the purpose of the present invention is to provide an intelligent delivery decision-making method and system based on the correlation of multi-dimensional indicators, so as to effectively improve the efficiency and effect of advertising delivery.

[0004] To achieve the above object, the present invention adopts the following technical solutions: An intelligent delivery decision-making method based on multi-dimensional indicator correlation includes the following steps: S1: Collect multi-dimensional data of users, ads, and contexts, and pre-process them to generate standardized features; S2: Fusion calculation of multi-dimensional standardized features, extraction of key features, and construction of a joint feature set for generating user-advertisement-context; S3: Build a multi-task prediction model and train it based on the joint feature set of user-advertisement-context; S4: Based on the prediction results and real-time features of the trained multi-task prediction model, the best ad is quickly matched for a given user during the real-time bidding process; S5: Perform causal analysis based on the exposure / click logs of the best ads to verify the real effect of the ads and correct the indicators. The corrected indicators are then fed back to the feature engineering steps of S2 and the model training steps of S3.

[0005] Furthermore, multi-dimensional data on users, advertisements, and context is collected, specifically as follows: the user data includes basic user characteristics, behavioral data, and interest tags; the advertisement data includes creative content, historical click-through rates, conversion rates, and delivery themes; the context data includes geographic location, device information, time period, and weather conditions; Kafka is used to implement real-time data collection, streaming multi-dimensional data on users, advertisements, and context to the computing platform.

[0006] Furthermore, preprocessing includes data cleaning, feature normalization, feature encoding, and real-time feature calculation, as follows: Data cleaning of multi-dimensional data on users, ads, and context includes: User data cleaning: Deduplication: Remove duplicate or invalid behavior events, and remove unusable or meaningless values; Ad data cleaning: Deduplication of ad click logs and filtering of abnormal ads (such as cheating ads with abnormal click frequencies); Context data cleaning: filling missing values in environmental data; Feature normalization processing, normalization and standardization of the range and distribution of different features; Feature encoding: One-hot encoding is used for categorical features to convert discrete values into sparse vectors; numerical features are directly used as input after normalization; Real-time features are calculated in Flink using sliding windows, including recent user click-through rates, ad dynamic features, contextual feature associations, and user behavior sequence features: User U i Click-through rate: ; Advertisement dynamic features: ; ; in, For Ad A j In a specific time period T k CTR; For Ad A j In a specific time period T k conversion rate; Contextual feature association: ; in, For device type D m Click-through rate; User behavior sequence features,use a model to model the behavior sequence and generate the user's dynamic,interest vector.

[0007] Furthermore, we perform fusion calculations on the multi-dimensional standardized features, extract key features, and construct a joint feature set for generating user-advertisement-context, as follows: Construct user, ad, and context features as joint feature input: Fraw=Concat(FU,FA,FC); Among them, FU is user feature; FA is advertising feature; FC is context feature; Use polynomial feature cross or embedding similarity to generate high-order cross features to capture multi-dimensional data interaction relationships; Use the feature selection PCA method to reduce the interference of non-critical features on the model and obtain the final feature input; Finally, a joint feature set of user-advertisement-context is constructed: FeatureVector=[CTRU,CTRA,CVRA,CTRD,IU,FA,FC,HighOrderFeatures]; Among them, CTRU is the user's historical click-through rate; CTRA is the advertising historical click-through rate, CVRA is the advertising conversion rate, CTRD is the context-related click-through rate, IU is the user's real-time interest vector, and HighOrderFeatures is the high-order interaction feature.

[0008] Furthermore, the multi-task prediction model includes an input layer, a dynamic parameter sharing layer, a deep cross network, a task-specific layer, and a multi-task prediction output layer, as follows: The input layer inputs the joint features of user-advertisement-context, and the shared feature extraction layer passes through the DNN layer; The dynamic parameter sharing layer uses a cross-stitch network to perform unified underlying feature extraction on the input user, advertisement and context features to obtain the shared feature h shared : h shared =ReLU(W shared *FeatureVector+b shared ); Among them, ReLU is the activation function; W shared、 b shared are weights and biases respectively; Dynamically adjust the shared features between tasks, for each task k, , where CTR is click-through rate, CVR is click-through conversion rate, and CLV is lifetime value. Dynamically integrate shared features with task-specific features x: ; in, The fusion features of task k; is the cross-stitch weight matrix of task k; is the independent weight of task k; The deep cross network automatically learns high-order feature crosses based on the fusion features to obtain the final splicing features: Cross-layer: ; Among them, x0 is the initial input feature, ⊙ represents element-by-element multiplication; are the output features of the 1st, 2nd and 3rd cross layers respectively; are the weights of the 1st, 2nd and 3rd cross layers respectively; Then the final feature of task k is : ; The task-specific layers include CTR prediction head, CVR prediction head, and CLV prediction head: ; in, 、 and are the weights of the CTR prediction head, CVR prediction head, and CLV prediction head respectively; 、 and They are the biases of CTR prediction head, CVR prediction head and CLV prediction head respectively; 、 and They are the outputs of the CTR prediction head, CVR prediction head, and CLV prediction head, respectively, and are output through the multi-task prediction output layer.

[0009] Furthermore, the multi-task prediction model training includes shared layer pre-training, task head fine-tuning, and end-to-end joint training, as follows: Shared layer pre-training, using historical data to pre-train the shared feature extraction layer and DCN, only using CTR task data to train the shared layer and DCN, the loss function uses binary cross entropy loss L CTR ; Fine-tune the task head, preliminarily adapt the parameters of each task head, freeze the shared layer and DCN parameters, and train the CTR, CVR, and CLV task heads separately: the loss function CTR and CVR use binary cross entropy loss L CTR and L CVR , CLV uses MSE loss L CLV ; End-to-end joint training, joint optimization of multi-task loss, first dynamic loss weighting, calculate the gradient norm of each task: ; in, is the normalized gradient weight of task k at step t; is the gradient of the loss function of task k with respect to the model parameters θ; is the gradient norm; K is the total number of tasks, i is the task guide; Update weights: ; in, is the final dynamic weight of task k at step t; η is the smoothing coefficient; Solve the multi-objective gradient direction through Pareto optimization: ; in, is the weight variable in the Pareto optimization of task k; is the L2 norm squared; Update model parameters: ; Among them, γ is the learning rate.

[0010] Furthermore, based on the prediction results and real-time features of the trained multi-task prediction model, the best ad is quickly matched for a given user during the real-time bidding process, as follows: Obtain real-time user requests and ad candidate pools; predict relevant indicators, including click-through rate (CTR), click-through conversion rate (CVR), and lifetime value (CLV), through the trained multi-task prediction model. Estimate the expected revenue from ad display based on business goals and forecast results; Combined with the estimated expected revenue of ad display, a comprehensive score is calculated for each candidate ad, and the ad with the highest score is sorted and selected for delivery.

[0011] Furthermore, we conduct causal analysis based on the exposure / click logs of the best ads to verify the real effects of the ads and correct the indicators. The corrected indicators are then fed back to the feature engineering of S2 and the model training of S3. The specific steps are as follows: Obtain user ID, ad ID, click status, conversion status, and contextual features based on exposure / click logs. Also, obtain user history of clicks, purchases, and browsing behaviors based on user behavior sequences. For each user-ad exposure event, record whether it was clicked (T=1 for clicked, T=0 for not clicked), and whether it was converted (Y=1 for conversion, Y=0 for not converted). Propensity score matching was used to estimate the true causal effect ATE of advertising; According to the true causal effect ATE, the click-through rate CTR and conversion rate CVR are modified: The causal effect estimate ATE is added as a new feature to the joint feature set, and the feature importance is adjusted according to the causal effect; In the model training step of S3, causal effect weights are introduced into the original loss function to prioritize the optimization of samples with high true causal effects; and the corrected click-through rate (CTR) and conversion rate (CVR) are used as pseudo-labels to retrain the model and optimize the multi-task prediction model in real time.

[0012] Furthermore, propensity score matching is used to estimate the true causal effect ATE of advertising, as follows: Use a logistic regression model to predict the probability of a user clicking an ad : ; Among them, X is the input feature, including user features, advertising features, and context features; is the transpose of the coefficient vector; β0 is the intercept term For each click user T=1, find the sample with the closest propensity score among the non-click users T=0 to construct a matching group; Calculate the average treatment effect ATE: ; in, is the expected function; For users i’ Conversion results after clicking; is the conversion result of the matched non-clicking users, N is the number of users.

[0013] An intelligent delivery decision-making system based on the association of multi-dimensional indicators includes a processor, a memory, and a computer program stored in the memory. When the processor executes the computer program, it specifically performs the steps of the intelligent delivery decision-making method based on the association of multi-dimensional indicators as described above.

[0014] The present invention has the following beneficial effects: 1. This paper builds a multi-task prediction model for CTR, CVR, and CLV. Through shared layers, dynamic parameter sharing (cross-stitch network), and deep cross-network, it automatically learns high-order interactions between features to improve the comprehensiveness of predictions. 2. This invention uses the matching results of a multi-task prediction model, combined with real-time features, to calculate a comprehensive score for candidate ads, prioritize them, and deliver them, ensuring accurate ad selection and rapid response. It also optimizes ad ranking by introducing expected revenue, achieving the optimal balance between accurately matching user interests and enhancing the true value of ads. 3. This invention uses propensity score matching to estimate the average treatment effect (ATE) of advertisements, removing false correlations caused by data bias, making prediction indicators such as click-through rate and conversion rate more consistent with the actual effect of advertisements. ATE is fed back into feature engineering and model training as a feature, and model predictions are dynamically adjusted based on the corrected indicators, forming an automated real-time closed-loop optimization mechanism, making advertising delivery more flexible and data updates more real-time. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 Flow chart of the method of the present invention. DETAILED DESCRIPTION

[0016] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments: refer to Figure 1 In this embodiment, a smart delivery decision-making method based on multi-dimensional indicator correlation is provided, including the following steps: S1: Collect multi-dimensional data of users, ads, and contexts, and pre-process them to generate standardized features; S2: Fusion calculation of multi-dimensional standardized features, extraction of key features, and construction of a joint feature set for generating user-advertisement-context; S3: Build a multi-task prediction model and train it based on the joint feature set of user-advertisement-context; S4: Based on the prediction results and real-time features of the trained multi-task prediction model, the best ad is quickly matched for a given user during the real-time bidding process; S5: Perform causal analysis based on the exposure / click logs of the best ads to verify the real effect of the ads and correct the indicators. The corrected indicators are then fed back to the feature engineering steps of S2 and the model training steps of S3.

[0017] In this implementation, multi-dimensional data on users, advertisements, and context is collected, specifically as follows: user data includes basic user characteristics (age, gender), behavioral data (browsing, clicking, purchasing), and interest tags; advertising data includes creative content, historical click-through rates, conversion rates, and delivery themes; context data includes geographic location, device information, time period, and weather conditions; Kafka is used to implement real-time data collection, streaming multi-dimensional data on users, advertisements, and context to the computing platform.

[0018] In this implementation, preprocessing includes data cleaning, feature normalization, feature encoding, and real-time feature calculation, as follows: Data cleaning of multi-dimensional data on users, ads, and context includes: User data cleaning: Deduplication: Remove repeated or invalid behavior events (such as multiple identical clicks in a short period of time) and remove unavailable or meaningless values (such as age outside a reasonable range); Ad data cleaning: Deduplication of ad click logs and filtering of abnormal ads (such as cheating ads with abnormal click frequencies); Contextual data cleaning: Fill missing values in environmental data (e.g., when weather is empty, use historical data to fill it); Feature normalization processing, normalization and standardization of the range and distribution of different features; Feature encoding: One-hot encoding is used for categorical features, converting discrete values (such as "male" and "female") into sparse vectors. Numerical features are directly used as input after normalization, such as ad click-through rate (CTR) and conversion rate (CVR). Real-time features are calculated in Flink using sliding windows, including recent user click-through rates, ad dynamic features, contextual feature associations, and user behavior sequence features: User U i Click-through rate: ; Advertisement dynamic features: ; ; in, For Ad A j In a specific time period T k CTR; For Ad A j In a specific time period T k conversion rate; Contextual feature association: ; in, For device type D m Click-through rate; User behavior sequence features, use models (such as GRU, Transformer) to model the behavior sequence and generate the user's dynamic interest vector.

[0019] In this implementation, multi-dimensional standardized features are fused and calculated to extract key features and construct a joint feature set for user-advertisement-context, as follows: Construct user, ad, and context features as joint feature input: Fraw=Concat(FU,FA,FC); Among them, FU is user features (including click-through rate, age, gender, interest vector, etc.); FA is advertising features (including historical CTR, CVR, creative embedding vector, etc.); FC is context features (including time, geographic location, device type); Use polynomial feature cross or embedding similarity to generate high-order cross features to capture multi-dimensional data interaction relationships; Use the feature selection PCA method to reduce the interference of non-critical features on the model and obtain the final feature input; Finally, a joint feature set of user-advertisement-context is constructed: FeatureVector=[CTRU,CTRA,CVRA,CTRD,IU,FA,FC,HighOrderFeatures]; Among them, CTRU is the user's historical click-through rate; CTRA is the advertising historical click-through rate, CVRA is the advertising conversion rate, CTRD is the context-related click-through rate, IU is the user's real-time interest vector, and HighOrderFeatures is the high-order interaction feature.

[0020] In this implementation, the multi-task prediction model includes an input layer, a dynamic parameter sharing layer, a deep cross network, a task-specific layer, and a multi-task prediction output layer, as follows: The input layer inputs the joint features of user-advertisement-context, and the shared feature extraction layer passes through the DNN (deep neural network) layer; The dynamic parameter sharing layer uses a cross-stitch network to perform unified underlying feature extraction on the input user, advertisement, and context features to obtain the shared feature h shared : h shared =ReLU(W shared *FeatureVector+b shared ); Among them, ReLU is the activation function; W shared、 b shared are weights and biases respectively; Dynamically adjust the shared features between tasks, for each task k, , where CTR is click-through rate, CVR is click-through conversion rate, and CLV is lifetime value. Dynamically integrate shared features with task-specific features x: ; in, The fusion features of task k; is the cross-stitch weight matrix of task k; is the independent weight of task k; is the activation function; The deep cross network automatically learns high-order feature crosses based on the fusion features to obtain the final splicing features: Cross-layer: ; Among them, x0 is the initial input feature, ⊙ represents element-by-element multiplication; are the output features of the 1st, 2nd and 3rd cross layers respectively; are the weights of the 1st, 2nd and 3rd cross layers respectively; Then the final feature of task k is : ; The task-specific layers include CTR prediction head, CVR prediction head, and CLV prediction head: ; in, 、 and are the weights of the CTR prediction head, CVR prediction head, and CLV prediction head respectively; 、 and They are the biases of CTR prediction head, CVR prediction head and CLV prediction head respectively; 、 and They are the outputs of the CTR prediction head, CVR prediction head, and CLV prediction head, respectively, and are output through the multi-task prediction output layer.

[0021] In this implementation, multi-task prediction model training includes shared layer pre-training, task head fine-tuning, and end-to-end joint training, as follows: Shared layer pre-training, using historical data to pre-train the shared feature extraction layer and DCN, only using CTR task data to train the shared layer and DCN, the loss function uses binary cross entropy loss L CTR ; Optimizer: Adam, learning rate 1e−3; Regularization: feature-level Dropout (ratio 0.2); Fine-tune the task head, preliminarily adapt the parameters of each task head, freeze the shared layer and DCN parameters, and train the CTR, CVR, and CLV task heads separately: the loss function CTR and CVR use binary cross entropy loss L CTR and L CVR , CLV uses MSE loss L CLV Optimizer: SGD, learning rate 1e−4; End-to-end joint training, joint optimization of multi-task loss, first dynamic loss weighting (GradNorm), calculate the gradient norm of each task: ; in, is the normalized gradient weight of task k at step t; is the gradient of the loss function of task k with respect to the model parameters θ; is the gradient norm; K is the total number of tasks, i is the task guide; Update weights: ; in, is the final dynamic weight of task k at step t; η is the smoothing coefficient; Solve the multi-objective gradient direction through Pareto optimization (MGDA): ; in, is the weight variable in the Pareto optimization of task k; is the L2 norm squared; Update model parameters: ; Among them, γ is the learning rate.

[0022] In this implementation, the best ad is quickly matched to a given user during the real-time bidding process based on the prediction results of the trained multi-task prediction model and real-time features, as follows: Obtain real-time user requests (such as user behavior, context, etc.) and ad candidate pools; predict relevant indicators, including click-through rate (CTR), click-through conversion rate (CVR), and lifetime value (CLV), through the trained multi-task prediction model. Based on business objectives (such as predicted revenue and ROI), estimate the expected revenue from ad display based on the forecast results; Combined with the estimated expected revenue of ad display, a comprehensive score is calculated for each candidate ad, and the ad with the highest score is sorted and selected for delivery.

[0023] In this implementation, we conduct causal analysis based on the exposure / click logs of the best ads to verify the actual effectiveness of the ads and correct the indicators. The corrected indicators are then fed back to the feature engineering of S2 and the model training of S3. The specific steps are as follows: Obtain user ID, ad ID, click status, conversion status, and contextual features based on exposure / click logs. Also, obtain user history of clicks, purchases, and browsing behaviors based on user behavior sequences. For each user-ad exposure event, record whether it was clicked (T=1 for clicked, T=0 for not clicked), and whether it was converted (Y=1 for conversion, Y=0 for not converted). Propensity score matching was used to estimate the true causal effect ATE of advertising; According to the true causal effect ATE, the click-through rate CTR and conversion rate CVR are modified: The causal effect estimate ATE is added as a new feature to the joint feature set, and the feature importance is adjusted according to the causal effect; In the model training step of S3, causal effect weights are introduced into the original loss function to prioritize the optimization of samples with high true causal effects; and the corrected click-through rate (CTR) and conversion rate (CVR) are used as pseudo-labels to retrain the model and optimize the multi-task prediction model in real time.

[0024] In this implementation, propensity score matching is used to estimate the true causal effect ATE of advertising, as follows: Use a logistic regression model to predict the probability of a user clicking an ad : ; Among them, X is the input feature, including user features, advertising features, and context features; is the transpose of the coefficient vector; β0 is the intercept term; For each click user T=1, find the sample with the closest propensity score among the non-click users T=0 to construct a matching group; Calculate the average treatment effect ATE: ; in, is the expected function; For users i’ Conversion results after clicking; is the conversion result of the matched non-clicking users, N is the number of users.

[0025] An intelligent delivery decision-making system based on the association of multi-dimensional indicators includes a processor, a memory, and a computer program stored in the memory. When the processor executes the computer program, it specifically performs the steps of the intelligent delivery decision-making method based on the association of multi-dimensional indicators as described above.

[0026] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0027] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0028] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0029] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0030] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other manner. Any person skilled in the art may utilize the above-disclosed technical content to modify or modify the present invention into equivalent embodiments. However, any simple modifications, equivalent variations, and modifications to the above embodiments that do not depart from the technical content of the present invention and are based on the technical essence of the present invention remain within the scope of protection of the present invention.

Claims

1. An intelligent placement decision-making method based on multi-dimensional indicator correlation, characterized in that: The following steps are involved: S1: Collect multi-dimensional data of users, ads, and contexts, and pre-process them to generate standardized features; S2: Fusion calculation of multi-dimensional standardized features, extraction of key features, and construction of a joint feature set for generating user-advertisement-context; S3: Build a multi-task prediction model and train it based on the joint feature set of user-advertisement-context; S4: Based on the prediction results and real-time features of the trained multi-task prediction model, the best ad is quickly matched for a given user during the real-time bidding process; S5: Perform causal analysis based on the exposure / click logs of the best ads to verify the real effect of the ads and correct the indicators. The corrected indicators are then fed back to the feature engineering steps of S2 and the model training steps of S3.

2. The intelligent placement decision-making method based on multi-dimensional indicator correlation according to claim 1 is characterized in that: The collection of multi-dimensional data of users, advertisements and contexts is as follows: the user data includes basic user characteristics, behavioral data, and interest tags; the advertisement data includes creative content, historical click-through rate, conversion rate, and delivery theme; the context data includes geographic location, device information, time period and weather conditions; Kafka is used to achieve real-time data collection, and the multi-dimensional data of users, advertisements and contexts is streamed to the computing platform.

3. The intelligent placement decision-making method based on multi-dimensional indicator correlation according to claim 2 is characterized in that: The preprocessing includes data cleaning, feature normalization, feature encoding and real-time feature calculation, as follows: Data cleaning of multi-dimensional data on users, ads, and context includes: User data cleaning: deduplication, removing duplicate or invalid behavior events, and removing unusable or meaningless values; Ad data cleaning: deduplication of ad click logs and filtering of abnormal ads; Context data cleaning: filling missing values in environmental data; Feature normalization processing, normalization and standardization of the range and distribution of different features; Feature encoding: One-hot encoding is used for categorical features to convert discrete values into sparse vectors; numerical features are directly used as input after normalization; Real-time features are calculated in Flink using sliding windows, including recent user click-through rates, ad dynamic features, contextual feature associations, and user behavior sequence features: User U i Click-through rate : ； Advertisement dynamic features: ； ； in, For Ad A j In a specific time period T k CTR; For Ad A j In a specific time period T k conversion rate; Contextual feature association: ; in, For device type D m Click-through rate; User behavior sequence features: A model is used to model the behavior sequence and generate the user's dynamic interest vector.

4. The intelligent placement decision-making method based on multi-dimensional indicator correlation according to claim 3 is characterized in that: The multi-dimensional standardized features are fused and calculated to extract key features and construct a joint feature set of user-advertisement-context, as follows: Construct user, ad, and context features into joint features and input Fraw: Fraw=Concat(FU,FA,FC); Among them, FU is user feature; FA is advertising feature; FC is context feature; Concat represents concatenation function; Use polynomial feature cross or embedding similarity to generate high-order cross features to capture multi-dimensional data interaction relationships; Use the feature selection PCA method to reduce the interference of non-critical features on the model and obtain the final feature input; Finally, a joint feature set of user-advertisement-context is constructed: FeatureVector=[CTRU,CTRA,CVRA,CTRD,IU,FA,FC,HighOrderFeatures]; Among them, CTRU is the user's historical click-through rate; CTRA is the advertising historical click-through rate, CVRA is the advertising conversion rate, CTRD is the context-related click-through rate, IU is the user's real-time interest vector, and HighOrderFeatures is the high-order interaction feature.

5. The intelligent placement decision-making method based on multi-dimensional indicator correlation according to claim 4 is characterized in that: The multi-task prediction model includes an input layer, a dynamic parameter sharing layer, a deep cross network, a task-specific layer, and a multi-task prediction output layer, as follows: The input layer inputs the joint features of user, advertisement and context. The dynamic parameter sharing layer uses a cross-stitch network to perform unified underlying feature extraction on the input user, advertisement and context features to obtain the shared features h shared : h shared =ReLU(W shared *FeatureVector+b shared ); Among them, ReLU is the activation function; W shared、 b shared are weights and biases respectively; Dynamically adjust the shared features between tasks, for each task k, , where CTR is click-through rate, CVR is click-through conversion rate, and CLV is lifetime value. Dynamically integrate shared features with task-specific features x: ； in, The fusion features of task k; is the cross-stitch weight matrix of task k; is the independent weight of task k; is the activation function; The deep cross network automatically learns high-order feature crosses based on the fusion features to obtain the final splicing features: Cross-layer: ； Among them, x0 is the initial input feature, ⊙ represents element-by-element multiplication; are the output features of the 1st, 2nd and 3rd cross layers respectively; are the weights of the 1st, 2nd and 3rd cross layers respectively; Then the final feature of task k is : ； The task-specific layers include CTR prediction head, CVR prediction head, and CLV prediction head: ; in, 、 and are the weights of the CTR prediction head, CVR prediction head, and CLV prediction head respectively; 、 and They are the biases of CTR prediction head, CVR prediction head and CLV prediction head respectively; 、 and They are the outputs of the CTR prediction head, CVR prediction head, and CLV prediction head, respectively, and are output through the multi-task prediction output layer.

6. The intelligent placement decision-making method based on multi-dimensional indicator correlation according to claim 5 is characterized in that: The multi-task prediction model training includes shared layer pre-training, task head fine-tuning and end-to-end joint training, as follows: Shared layer pre-training, using historical data to pre-train the shared feature extraction layer and DCN, only using CTR task data to train the shared layer and DCN, the loss function uses binary cross entropy loss L CTR ; Fine-tune the task head, preliminarily adapt the parameters of each task head, freeze the shared layer and DCN parameters, and train the CTR, CVR, and CLV task heads separately: the loss function CTR and CVR use binary cross entropy loss L CTR and L CVR , CLV uses MSE loss L CLV ; End-to-end joint training, joint optimization of multi-task loss, first dynamic loss weighting, calculate the gradient norm of each task: ； in, is the normalized gradient weight of task k at step t; is the gradient of the loss function of task k with respect to the model parameters θ; is the gradient norm; K is the total number of tasks, i is the task guide; Update weights: ； in, is the final dynamic weight of task k at step t; η is the smoothing coefficient; Solve the multi-objective gradient direction through Pareto optimization: ； in, is the weight variable in the Pareto optimization of task k; is the L2 norm squared; Update model parameters: ； Among them, γ is the learning rate.

7. The intelligent placement decision-making method based on multi-dimensional indicator correlation according to claim 1 is characterized in that: The method uses the trained multi-task prediction model prediction results and real-time features to quickly match the best ad for a given user during the real-time bidding process, specifically as follows: Obtain real-time user requests and ad candidate pools; use the trained multi-task prediction model to predict relevant indicators, including click-through rate (CTR), click-through conversion rate (CVR), and lifetime value (CLV); Estimate the expected revenue from ad display based on business goals and forecast results; Combined with the estimated expected revenue of ad display, a comprehensive score is calculated for each candidate ad, and the ad with the highest score is sorted and selected for delivery.

8. The intelligent placement decision-making method based on multi-dimensional indicator correlation according to claim 1 is characterized in that: The causal analysis is performed based on the exposure / click logs of the best ads to verify the real effect of the ads and correct the indicators. The corrected indicators are fed back to the feature engineering of S2 and the model training of S3. The specific steps are as follows: Obtain user ID, ad ID, click status, conversion status, and contextual features based on exposure / click logs. Also, obtain user history of clicks, purchases, and browsing behaviors based on user behavior sequences. For each user-ad exposure event, record whether it was clicked (T=1 for clicked, T=0 for not clicked), and whether it was converted (Y=1 for conversion, Y=0 for not converted). Propensity score matching was used to estimate the true causal effect ATE of advertising; According to the true causal effect ATE, the click-through rate CTR and click-through conversion rate CVR are modified: The causal effect estimate ATE is added as a new feature to the joint feature set, and the feature importance is adjusted according to the causal effect; In the model training step of S3, causal effect weights are introduced into the original loss function to prioritize the optimization of samples with high true causal effects; and the corrected click-through rate (CTR) and conversion rate (CVR) are used as pseudo-labels to retrain the model and optimize the multi-task prediction model in real time.

9. The intelligent placement decision-making method based on multi-dimensional indicator correlation according to claim 8 is characterized in that: The propensity score matching is used to estimate the true causal effect ATE of advertising, as follows: Use a logistic regression model to predict the probability of a user clicking an ad : ； Among them, X is the input feature, including user features, advertising features, and context features; is the transpose of the coefficient vector; β0 is the intercept term; For each click user T=1, find the sample with the closest propensity score among the non-click users T=0 to construct a matching group; Calculate the average treatment effect ATE: ； in, is the expected function; For users i’ Conversion results after clicking; is the conversion result of the matched non-clicking users, N is the number of users.

10. An intelligent placement decision-making system based on multi-dimensional indicator correlation, characterized in that: It includes a processor, a memory, and a computer program stored in the memory. When the processor executes the computer program, it specifically performs the steps in the intelligent delivery decision-making method based on multi-dimensional indicator association as described in any one of claims 1 to 9.

Images