An intelligent delivery decision method and system based on multi-dimensional index correlation

Abstract

The application relates to an intelligent putting decision method and system based on multi-dimensional index correlation, which comprises the following steps: S1: collecting multi-dimensional data of users, advertisements and contexts, and preprocessing to generate standardized features; S2: performing fusion calculation on the multi-dimensional standardized features, extracting key features, and constructing a joint feature set of users-advertisements-contexts; S3: constructing a multi-task prediction model and training according to the joint feature set of users-advertisements-contexts; S4: according to the prediction result of the trained multi-task prediction model and real-time features, quickly matching the best advertisement for a given user in the real-time bidding process; and S5: according to the exposure / click log of the best advertisement, performing causal analysis, verifying the real effect of the advertisement, correcting the index, and feeding back the corrected index to the feature engineering of S2 and the model training step of S3. The application effectively improves the efficiency and effect of advertisement putting.

Description

Technical Field

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

[0002] With the rapid rise of digital content platforms such as short dramas and online novels, advertising is becoming an important commercialization tool in the content ecosystem. Unlike e-commerce and news platforms, advertising decisions in the consumption scenarios of short dramas and novels are more complex, requiring consideration of both the instantaneous changes in user interests and the relevance of content context and diverse advertising materials. Existing advertising systems mainly rely on simple single metrics such as click-through rate or conversion rate for prediction and optimization, facing the following technical challenges:

[0003] Traditional systems focus on short-term clicks or conversions from ads, making it difficult to capture the long-term impact of ads on deeper user behaviors (such as follow-up retention, paid unlocks, and social sharing of series). This results in high customer acquisition costs, weak user stickiness, and limited monetization efficiency for advertisers. Effective ad placement decisions require a comprehensive understanding and integration of user history (following series, reading preferences, interaction frequency), ad creative attributes (type, style, creativity), and content context (plot points, chapter progress, character positioning, scene atmosphere, etc.). Traditional methods often struggle to dynamically correlate and deeply mine these multi-dimensional features, leading to poor ad recommendation relevance and low traffic utilization. Summary of the Invention

[0004] To address the aforementioned issues, the present invention aims to provide an intelligent advertising decision-making method and system based on multi-dimensional indicator correlation, effectively improving advertising efficiency and effectiveness.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A smart delivery decision-making method based on multi-dimensional indicator correlation includes the following steps:

[0007] S1: Collect multi-dimensional data on users, advertisements, and context, and preprocess it to generate standardized features;

[0008] S2: Perform fusion calculation on multi-dimensional standardized features, extract key features, and construct a joint feature set of user-advertisement-context;

[0009] S3: Construct a multi-task prediction model and train it based on the joint feature set of user-advertisement-context;

[0010] S4: Based on the prediction results of the trained multi-task prediction model and real-time features, quickly match the best advertisement for a given user during the real-time bidding process;

[0011] S5: Perform causal analysis based on the exposure / click logs of the best ads to verify the actual effect of the ads, and correct the metrics. Feedback the corrected metrics to the feature engineering in S2 and the model training steps in S3.

[0012] Furthermore, multi-dimensional data on users, advertisements, and context is collected, specifically as follows: User data includes basic user characteristics, behavioral data, and interest tags; Ad data includes creative content, historical click-through rates, conversion rates, and campaign themes; Context data includes geographic location, device information, time period, and weather conditions; Kafka is used to achieve real-time data collection, and multi-dimensional data on users, advertisements, and context is streamed to the computing platform.

[0013] Furthermore, preprocessing includes data cleaning, feature normalization, feature encoding, and real-time feature computation, as follows:

[0014] Data cleaning of multi-dimensional data including users, advertisements, and context includes:

[0015] User data cleaning: Deduplication: Remove duplicate or invalid behavior events and remove unusable or meaningless values;

[0016] Ad data cleaning: Deduplicating ad click logs and filtering out abnormal ads (such as fraudulent ads with abnormal click frequencies);

[0017] Contextual data cleaning: filling in missing values ​​in environmental data;

[0018] Feature normalization processing involves normalizing and standardizing different features based on their range and distribution.

[0019] For feature encoding, one-hot encoding is used for categorical features to convert discrete values ​​into sparse vectors; numerical features are normalized and directly used as input.

[0020] Real-time features are computed in Flink using a sliding window, including recent user click-through rate, ad dynamic features, contextual feature correlation, and user behavior sequence features:

[0021] User U i Click-through rate:

[0022] ;

[0023] Dynamic features of advertisements:

[0024] ;

[0025] ;

[0026] in, For Advertisement Aj During a specific time period T k Click-through rate; For Advertisement A j During a specific time period T k Conversion rate;

[0027] Contextual feature association:

[0028] ;

[0029] in, For equipment type D m Click-through rate;

[0030] User behavior sequence features are used to model the behavior sequence and generate dynamic interest vectors for users.

[0031] Furthermore, the multi-dimensional standardized features are fused and calculated to extract key features, and a joint feature set of user-advertisement-context is constructed, as follows:

[0032] Combine user, ad, and contextual features into a joint feature input:

[0033] Fraw = Concat(FU, FA, FC);

[0034] Wherein, FU represents user features; FA represents advertising features; and FC represents contextual features.

[0035] Use polynomial feature cross or embedding similarity to generate high-order cross features to capture multidimensional data interaction relationships;

[0036] The PCA feature selection method is used to reduce the interference of non-critical features on the model and obtain the final feature input.

[0037] Finally, a joint feature set of user-advertisement-context is constructed:

[0038] FeatureVector=[CTRU,CTRA,CVRA,CTRD,IU,FA,FC,HighOrderFeatures];

[0039] Wherein, CTRU is the user's historical click-through rate; CTRA is the ad's historical click-through rate; CVRA is the ad conversion rate; CTRD is the context-dependent click-through rate; IU is the user's real-time interest vector; and HighOrderFeatures are high-order interaction features.

[0040] 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 detailed below:

[0041] The input layer takes in the joint features of user-advertisement-context, and the shared feature extraction layer is obtained through a DNN layer;

[0042] The dynamic parameter sharing layer employs a cross-stitch network to perform unified low-level feature extraction on the input user, advertisement, and context features, obtaining the shared feature h. shared :

[0043] h shared =ReLU(W shared *FeatureVector+b shared );

[0044] Where ReLU is the activation function; W shared、 b shared These are weights and biases, respectively.

[0045] Dynamically adjust shared features between tasks, for each task k. Where CTR is click-through rate, CVR is click-to-conversion rate, and CLV is lifetime value, dynamically integrating shared features and task-specific features.

[0046] ;

[0047] in, The fusion features of task k; Let K be the cross-stitch weight matrix for task k; These are the independent weights for task k;

[0048] The deep cross-connect network automatically learns higher-order feature crosses based on the fusion features to obtain the final concatenated features:

[0049] Cross layer:

[0050] ;

[0051] Where x0 is the initial input feature, and ⊙ represents element-wise multiplication; These are the output features of the cross-layers 1, 2, and 3, respectively; These are the weights of the intersection layers of layers 1, 2, and 3, respectively.

[0052] Then the final feature of task k is :

[0053] ;

[0054] The task-specific layer includes the CTR prediction head, CVR prediction head, and CLV prediction head:

[0055] ;

[0056] in, , and These are the weights of the CTR prediction head, CVR prediction head, and CLV prediction head, respectively. , and These are the biases for the CTR prediction head, CVR prediction head, and CLV prediction head, respectively. , and The outputs of the CTR prediction head, CVR prediction head, and CLV prediction head are respectively output through a multi-task prediction output layer.

[0057] Furthermore, the training of the multi-task prediction model includes shared layer pre-training, task head fine-tuning, and end-to-end joint training, as detailed below:

[0058] Shared layer pre-training: The shared feature extraction layer and DCN are pre-trained using historical data. The shared layer and DCN are trained using only CTR task data. The loss function is binary cross-entropy loss L. CTR ;

[0059] Fine-tuning of the task heads to initially adapt the parameters of each task head, freezing the parameters of the shared layer and DCN, and training the CTR, CVR, and CLV task heads respectively: the loss function for CTR and CVR adopts the binary cross-entropy loss L. CTR and L CVR CLV uses MSE loss L CLV ;

[0060] End-to-end joint training and joint optimization of multi-task loss first involves dynamic loss weighting and calculation of the gradient norm for each task.

[0061] ;

[0062] in, Let K be the normalized gradient weights for task k at step t. Let be the gradient of the loss function for task k with respect to the model parameters θ; Let be the gradient norm; K be the total number of tasks, and i be the task guide;

[0063] Update weights:

[0064] ;

[0065] in, Let be the final dynamic weight of task k at step t; η For smoothing coefficients;

[0066] Solving for the multi-objective gradient direction using Pareto optimization:

[0067] ;

[0068] in, For the Pareto optimization of task k; The square of the L2 norm;

[0069] Update model parameters:

[0070] ;

[0071] Where γ is the learning rate.

[0072] Furthermore, based on the prediction results of the trained multi-task prediction model and real-time features, the best advertisement is quickly matched for a given user during the real-time bidding process, as follows:

[0073] Acquire real-time user requests and an ad candidate pool; use a trained multi-task prediction model to predict relevant metrics, including click-through rate (CTR), click-to-conversion rate (CVR), and lifetime value (CLV).

[0074] Estimate the expected revenue from ad impressions based on business objectives and forecast results;

[0075] Based on the estimated expected revenue from ad impressions, calculate a comprehensive score for each candidate ad, sort them, and select the ad with the highest score for delivery.

[0076] Furthermore, causal analysis is performed based on the exposure / click logs of the best ads to verify the true effectiveness of the ads and to correct the metrics. The corrected metrics are then fed back to the feature engineering in S2 and the model training in S3. The specific steps are as follows:

[0077] Based on the exposure / click logs, obtain the user ID, ad ID, whether it clicked, whether it converted, and contextual features. Based on the user behavior sequence, obtain the user's historical click, purchase, and browsing behavior.

[0078] For each user-ad exposure event, record whether it was clicked (T=1 indicates clicked, T=0 indicates not clicked) and whether it was converted (Y=1 indicates conversion, Y=0 indicates no conversion).

[0079] Propensity score matching was used to estimate the true causal effect (ATE) of the advertisement.

[0080] Based on the true causal effect (ATE), the click-through rate (CTR) and conversion rate (CVR) are adjusted as follows:

[0081] 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.

[0082] In the model training steps 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 modified 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.

[0083] Furthermore, propensity score matching is used to estimate the true causal effect (ATE) of the advertisement, as follows:

[0084] Predict the probability of a user clicking an ad using a logistic regression model. :

[0085] ;

[0086] Where X represents the input features, including user features, advertising features, and contextual features; β0 is the transpose of the coefficient vector; β0 is the intercept term.

[0087] For each user who clicks (T=1), find the sample with the closest tendency score among the non-clicking users (T=0) and construct a matching group;

[0088] Calculate the average treatment effect (ATE):

[0089] ;

[0090] in, Let it be the expected function; For users i’ Conversion results after clicking; N represents the conversion results for the matched users who did not click, and the number of users.

[0091] A smart delivery decision-making system based on multi-dimensional indicator correlation 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 smart delivery decision-making method based on multi-dimensional indicator correlation as described above.

[0092] The present invention has the following beneficial effects:

[0093] 1. This invention constructs 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 prediction.

[0094] 2. This invention calculates a comprehensive score for candidate advertisements by combining the matching results of a multi-task prediction model with real-time features, sorts them by priority and delivers them, ensuring accurate and rapid advertisement selection. Furthermore, the introduction of expected returns optimizes the advertisement sorting, achieving the best balance between accurately matching user interests and enhancing the true value of advertisements.

[0095] 3. This invention uses propensity score matching to estimate the average treatment effect (ATE) of advertisements, which removes false correlations caused by data bias, making the predicted indicators such as click-through rate and conversion rate more consistent with the actual effect of the advertisement. The ATE is used as a feature to feed back into feature engineering and model training. The model prediction is dynamically adjusted through the corrected indicators, forming an automated real-time closed-loop optimization mechanism, which makes the ad placement more flexible and the data updates more real-time. Attached Figure Description

[0096] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0097] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0098] refer to Figure 1 In this embodiment, an intelligent delivery decision-making method based on multi-dimensional indicator correlation is provided, including the following steps:

[0099] S1: Collect multi-dimensional data on users, advertisements, and context, and preprocess it to generate standardized features;

[0100] S2: Perform fusion calculation on multi-dimensional standardized features, extract key features, and construct a joint feature set of user-advertisement-context;

[0101] S3: Construct a multi-task prediction model and train it based on the joint feature set of user-advertisement-context;

[0102] S4: Based on the prediction results of the trained multi-task prediction model and real-time features, quickly match the best advertisement for a given user during the real-time bidding process;

[0103] S5: Perform causal analysis based on the exposure / click logs of the best ads to verify the actual effect of the ads, and correct the metrics. Feedback the corrected metrics to the feature engineering in S2 and the model training steps in S3.

[0104] In this implementation, multi-dimensional data of users, advertisements, and context are collected, specifically as follows: The user data includes basic user characteristics (age, gender), behavioral data (browsing, clicking, purchasing), and interest tags; the advertisement data includes creative content, historical click-through rate, conversion rate, and campaign theme; the context data includes geographical 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 context are streamed to the computing platform.

[0105] In this implementation, preprocessing includes data cleaning, feature normalization, feature encoding, and real-time feature computation, as follows:

[0106] Data cleaning of multi-dimensional data including users, advertisements, and context includes:

[0107] User data cleaning: Deduplication: Remove duplicate or invalid behavioral events (such as multiple identical clicks in a short period of time), and remove unusable or meaningless values ​​(such as age that is outside the reasonable range).

[0108] Ad data cleaning: Deduplicating ad click logs and filtering out abnormal ads (such as fraudulent ads with abnormal click frequencies);

[0109] Contextual data cleaning: Fill in missing values ​​in environmental data (e.g., fill in with historical data when weather is missing).

[0110] Feature normalization processing involves normalizing and standardizing different features based on their range and distribution.

[0111] Feature encoding processing uses one-hot encoding for categorical features to convert discrete values ​​(such as "male" and "female") into sparse vectors; numerical features are normalized and directly used as input; for example, click-through rate (CTR) and conversion rate (CVR).

[0112] Real-time features are computed in Flink using a sliding window, including recent user click-through rate, ad dynamic features, contextual feature correlation, and user behavior sequence features:

[0113] User U i Click-through rate:

[0114] ;

[0115] Dynamic features of advertisements:

[0116] ;

[0117] ;

[0118] in, For Advertisement Aj During a specific time period T k Click-through rate; For Advertisement A j During a specific time period T k Conversion rate;

[0119] Contextual feature association:

[0120] ;

[0121] in, For equipment type D m Click-through rate;

[0122] User behavior sequence features are modeled using models (such as GRU and Transformer) to generate dynamic interest vectors for users.

[0123] In this implementation, multi-dimensional standardized features are fused and calculated to extract key features and construct a joint feature set of user-advertisement-context, as follows:

[0124] Combine user, ad, and contextual features into a joint feature input:

[0125] Fraw = Concat(FU, FA, FC);

[0126] Among them, FU represents user characteristics (including click-through rate, age, gender, interest vector, etc.); FA represents advertising characteristics (including historical CTR, CVR, creative embedding vector, etc.); and FC represents contextual characteristics (including time, geographic location, device type).

[0127] Use polynomial feature cross or embedding similarity to generate high-order cross features to capture multidimensional data interaction relationships;

[0128] The PCA feature selection method is used to reduce the interference of non-critical features on the model and obtain the final feature input.

[0129] Finally, a joint feature set of user-advertisement-context is constructed:

[0130] FeatureVector=[CTRU,CTRA,CVRA,CTRD,IU,FA,FC,HighOrderFeatures];

[0131] Wherein, CTRU is the user's historical click-through rate; CTRA is the ad's historical click-through rate; CVRA is the ad conversion rate; CTRD is the context-dependent click-through rate; IU is the user's real-time interest vector; and HighOrderFeatures are high-order interaction features.

[0132] 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 detailed below:

[0133] The input layer takes in the joint features of user-advertisement-context, and the shared feature extraction layer is obtained through a DNN (deep neural network) layer;

[0134] The dynamic parameter sharing layer employs Cross-stitch Networks to perform unified low-level feature extraction on the input user, advertisement, and context features, obtaining shared features h. shared :

[0135] h shared =ReLU(W shared *FeatureVector+b shared );

[0136] Where ReLU is the activation function; W shared、 b shared These are weights and biases, respectively.

[0137] Dynamically adjust shared features between tasks, for each task k. Where CTR is click-through rate, CVR is click-to-conversion rate, and CLV is lifetime value, dynamically integrating shared features and task-specific features.

[0138] ;

[0139] in, The fusion features of task k; Let K be the cross-stitch weight matrix for task k; These are the independent weights for task k; For activation functions;

[0140] The deep cross-connect network automatically learns higher-order feature crosses based on the fusion features to obtain the final concatenated features:

[0141] Cross layer:

[0142] ;

[0143] Where x0 is the initial input feature, and ⊙ represents element-wise multiplication; These are the output features of the cross-layers 1, 2, and 3, respectively; These are the weights of the intersection layers of layers 1, 2, and 3, respectively.

[0144] Then the final feature of task k is :

[0145] ;

[0146] The task-specific layer includes the CTR prediction head, CVR prediction head, and CLV prediction head:

[0147] ;

[0148] in, , and These are the weights of the CTR prediction head, CVR prediction head, and CLV prediction head, respectively. , and These are the biases for the CTR prediction head, CVR prediction head, and CLV prediction head, respectively. , and The outputs of the CTR prediction head, CVR prediction head, and CLV prediction head are respectively output through a multi-task prediction output layer.

[0149] In this implementation, the training of the multi-task prediction model includes shared layer pre-training, task head fine-tuning, and end-to-end joint training, as detailed below:

[0150] Shared layer pre-training: The shared feature extraction layer and DCN are pre-trained using historical data. The shared layer and DCN are trained using only CTR task data. The loss function is binary cross-entropy loss L. CTR ;

[0151] Optimizer: Adam, learning rate 1e−3; Regularization: feature-level Dropout (scale 0.2).

[0152] Fine-tuning of the task heads to initially adapt the parameters of each task head, freezing the parameters of the shared layer and DCN, and training the CTR, CVR, and CLV task heads respectively: the loss function for CTR and CVR adopts the binary cross-entropy loss L. CTR and L CVR CLV uses MSE loss L CLV Optimizer: SGD, learning rate 1e−4;

[0153] End-to-end joint training and joint optimization of multi-task loss first involves dynamic loss weighting (GradNorm) and calculating the gradient norm for each task.

[0154] ;

[0155] in, Let K be the normalized gradient weights for task k at step t. Let be the gradient of the loss function for task k with respect to the model parameters θ; Let be the gradient norm; K be the total number of tasks, and i be the task guide;

[0156] Update weights:

[0157] ;

[0158] in, Let be the final dynamic weight of task k at step t; η For smoothing coefficients;

[0159] Solving for the multi-objective gradient direction using Pareto optimization (MGDA):

[0160] ;

[0161] in, For the Pareto optimization of task k; The square of the L2 norm;

[0162] Update model parameters:

[0163] ;

[0164] Where γ is the learning rate.

[0165] In this implementation, based on the prediction results of the trained multi-task prediction model and real-time features, the best advertisement is quickly matched for a given user during the real-time bidding process, as follows:

[0166] Acquire real-time user requests (such as user behavior, context, etc.) and the ad candidate pool; predict relevant metrics, including click-through rate (CTR), click-to-conversion rate (CVR), and lifetime value (CLV), using a trained multi-task prediction model.

[0167] Estimate the expected revenue from ad impressions based on business objectives (such as projected revenue and ROI) and the projected results.

[0168] Based on the estimated expected revenue from ad impressions, calculate a comprehensive score for each candidate ad, sort them, and select the ad with the highest score for delivery.

[0169] In this implementation, causal analysis is performed based on the exposure / click logs of the best advertisement to verify the true effect of the advertisement, and the metrics are corrected. The corrected metrics are then fed back to the feature engineering in S2 and the model training in S3. The specific steps are as follows:

[0170] Based on the exposure / click logs, obtain the user ID, ad ID, whether it clicked, whether it converted, and contextual features. Based on the user behavior sequence, obtain the user's historical click, purchase, and browsing behavior.

[0171] For each user-ad exposure event, record whether it was clicked (T=1 indicates clicked, T=0 indicates not clicked) and whether it was converted (Y=1 indicates conversion, Y=0 indicates no conversion).

[0172] Propensity score matching was used to estimate the true causal effect (ATE) of the advertisement.

[0173] Based on the true causal effect (ATE), the click-through rate (CTR) and conversion rate (CVR) are adjusted as follows:

[0174] 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.

[0175] In the model training steps 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 modified 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.

[0176] In this implementation, propensity score matching is used to estimate the true causal effect (ATE) of the advertisement, as follows:

[0177] Predict the probability of a user clicking an ad using a logistic regression model. :

[0178] ;

[0179] Where X represents the input features, including user features, advertising features, and contextual features; β is the transpose of the coefficient vector; β0 is the intercept term;

[0180] For each user who clicks (T=1), find the sample with the closest tendency score among the non-clicking users (T=0) and construct a matching group;

[0181] Calculate the average treatment effect (ATE):

[0182] ;

[0183] in, Let it be the expected function; For users i’ Conversion results after clicking; N represents the conversion results for the matched users who did not click, and the number of users.

[0184] A smart delivery decision-making system based on multi-dimensional indicator correlation 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 smart delivery decision-making method based on multi-dimensional indicator correlation as described above.

[0185] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0186] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations 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, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0187] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0188] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0189] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for intelligent delivery decision based on multi-dimension index correlation, characterized in that, The method comprises the following steps: S1: Collecting user, advertisement and context multi-dimensional data, and preprocessing to generate standardized features; S2: Fusion calculation is performed on the multi-dimensional standardized features, key features are extracted, and a joint feature set of user-advertisement-context is constructed; S3: A multi-task prediction model is constructed, and training is performed according to the joint feature set of user-advertisement-context; S4: According to the prediction result of the trained multi-task prediction model and real-time features, the best advertisement is quickly matched for a given user in the real-time bidding process; S5: According to the exposure / click log of the best advertisement, a causal analysis is performed to verify the real effect of the advertisement, and the index is corrected, and the corrected index is fed back to the feature engineering of S2 and the model training step of S3; The multi-task prediction model comprises an input layer, a dynamic parameter sharing layer, a deep cross network, a task-specific layer and a multi-task prediction output layer, and specifically as follows: The input layer inputs the joint features of user-advertisement-context, the dynamic parameter sharing layer adopts cross-stitch network, uniformly extracts bottom features of the input user, advertisement and context features, and obtains shared features h shared : h shared = ReLU(W shared *FeatureVector + b shared ); where ReLU is an activation function; W shared、 b shared are weight and bias, respectively; and FeatureVector is the joint feature set of user-advertisement-context. dynamically adjust inter-task shared features, each task k, where CTR is the click rate, CVR is the click conversion rate, CLV is the lifetime value, and dynamically fuse shared features and task-specific features ； wherein, fusion features for task k; cross-stitch weight matrix for task k; independent weight for task k; activation function; The deep cross network automatically learns high-order feature cross according to the fusion features to obtain the final splicing features: The cross layer: ； wherein x0is an initial input feature, and represents element-wise multiplication; are output features of the 1st, 2nd and 3rd cross layers, respectively; are weights of the 1st, 2nd and 3rd cross layers, respectively. The final feature for task k is then : ； Wherein, Concat is a splicing function; The task-specific layer comprises a CTR prediction head, a CVR prediction head and a CLV prediction head: ; wherein, , and are weights of the CTR prediction head, the CVR prediction head and the CLV prediction head, respectively; , and are biases of the CTR prediction head, the CVR prediction head and the CLV prediction head, respectively; , and are outputs of the CTR prediction head, the CVR prediction head and the CLV prediction head, respectively, output by the multi-task prediction output layer; The multi-task prediction model training comprises shared layer pre-training, task head fine-tuning and end-to-end joint training, and specifically as follows: The shared layer is pre-trained, the shared feature extraction layer and the DCN are pre-trained using historical data, the shared layer and the DCN are trained only using CTR task data, and a binary cross-entropy loss L is used as a loss function CTR ; Task head fine-tuning, preliminary adaptation of each task head parameter, freezing of shared layer and DCN parameter, training of CTR, CVR and CLV task head respectively: loss function CTR, CVR adopts binary cross-entropy loss L CTR and L CVR , CLV adopts MSE loss L CLV ; In the end-to-end joint training, the multi-task loss is jointly optimized, first the dynamic loss is weighted, and the gradient norm of each task is calculated: ； wherein, is the normalized gradient weight for task k at step t; is the gradient of the loss function for task k with respect to the model parameters θ; is the gradient norm; K is the total number of tasks, and i is the task index. The weight is updated as follows: ； wherein, is the final dynamic weight for task k at step t; η is a smoothing coefficient; The multi-objective gradient direction is solved through Pareto optimization: ； wherein, is a weight variable in the Pareto optimization for task k; is the L2 norm square; The model parameters are updated as follows: ； Wherein, γ is a learning rate; The causal analysis according to the exposure / click log of the best advertisement, the verification of the real effect of the advertisement, and the correction of the index, and the corrected index is fed back to the feature engineering of S2 and the model training step of S3, specifically as follows: According to the exposure / click log, user ID, advertisement ID, whether to click, whether to convert, context features are obtained, and according to the user behavior sequence, user historical click, purchase and browsing behaviors are obtained; For each user-advertisement exposure event, whether it is clicked, T=1 indicates that it is clicked, T=0 indicates that it is not clicked, and whether it is converted, Y=1 indicates conversion, and Y=0 indicates non-conversion; The propensity score matching is used to estimate the real causal effect ATE of the advertisement; According to the real causal effect ATE, the click rate CTR and the click conversion rate CVR are corrected: The causal effect estimation value ATE is added to the joint feature set as a new feature, and the feature importance is adjusted according to the causal effect; In the model training step of S3, the causal effect weight is introduced into the original loss function, and the sample with high real causal effect is preferentially optimized; and the corrected click rate CTR and conversion rate CVR are used as pseudo-labels to retrain the model, and the multi-task prediction model is optimized in real time; The propensity score matching is used to estimate the real causal effect ATE of the advertisement, specifically as follows: Predicting the probability of a user clicking on an advertisement using a logistic regression model : ； Wherein, X is an input feature, including user features, advertisement features, context features; is the transpose of the coefficient vector; β0is the intercept term; For each clicked user T=1, find the sample with the closest propensity score in the non-clicked user T=0 to construct a matching group; The average treatment effect ATE is calculated as follows: ； wherein, is the desired function; is the user i’ conversion result after click; is the conversion result for the matched non-clicked users, N is the number of users. 2.The intelligent delivery decision-making method based on multi-dimensional index correlation of claim 1, characterized in that, The multi-dimensional data of users, advertisements and contexts are collected, and the collection of the multi-dimensional data of users, advertisements and contexts is specifically as follows: the user data includes basic characteristics of users, behavior data and interest labels; the advertisement data includes creative content, historical click rate, conversion rate and delivery theme; and the context data includes geographical position, device information, time period and weather condition; real-time data collection is realized by using Kafka, and the multi-dimensional data of users, advertisements and contexts are streamed to a computing platform. 3.The intelligent delivery decision-making method based on multi-dimensional index correlation of claim 2, characterized in that, The preprocessing includes data cleaning, feature normalization, feature coding and real-time feature calculation, and the preprocessing is specifically as follows: Data cleaning of the multi-dimensional data of users, advertisements and contexts includes the following steps: User data cleaning: removing duplicates or invalid behavior events, and removing unusable or meaningless values; Advertisement data cleaning: removing duplicates of advertisement click logs, and filtering abnormal advertisements; Context data cleaning: filling in missing values in environmental data; Feature normalization processing: normalization and standardization processing are performed on the range and distribution of different features; Feature coding processing: One-hot coding is used for category type features, discrete values are converted into sparse vectors, and normalized numerical type features are directly used as inputs; Real-time features are calculated in Flink through a sliding window, including recent click rate of users, dynamic features of advertisements, context feature association and user behavior sequence features: User U i Click rate : ； Dynamic features of advertisements: ； ； wherein, is the click rate of the advertisement A j in the specific time period T k ; is the conversion rate of the advertisement A j in the specific time period T k ; Context feature association: ; wherein, is a device type D m click rate; User behavior sequence features: a model is used to model the behavior sequence to generate a dynamic interest vector of a user. 4.The intelligent delivery decision method based on multi-dimension index correlation of claim 3, characterized in that, The multi-dimensional standardized features are fused and calculated, key features are extracted, and a joint feature set of users, advertisements and contexts is constructed, and the joint feature set is specifically as follows: User, advertisement and context features are constructed into a joint feature input Fraw: Fraw=Concat(FU,FA,FC); Wherein, FU is a user feature; FA is an advertisement feature; FC is a context feature; and Concat represents a concatenation function; High-order cross features are generated by using polynomial feature crossing or embedding similarity to capture the interaction relationship of multi-dimensional data; A feature selection PCA method is used to reduce the interference of non-key features on a model to obtain a final feature input; Finally, a joint feature set of users, advertisements and contexts is constructed: FeatureVector=[CTRU,CTRA,CVRA,CTRD,IU,FA,FC,HighOrderFeatures]; Wherein, CTRU is a historical click rate of a user; CTRA is a historical click rate of an advertisement, CVRA is an advertisement conversion rate, CTRD is a context related click rate, IU is a real-time interest vector of a user, and HighOrderFeatures are high-order interaction features.

5. The intelligent delivery decision-making method based on multi-dimensional index association according to claim 1, characterized in that, The best advertisement is quickly matched for a given user in a real-time bidding process according to a prediction result of a trained multi-task prediction model and real-time features, and the matching is specifically as follows: A user real-time request and an advertisement candidate pool are obtained; relevant indexes including a click rate CTR, a click conversion rate CVR and a life cycle value CLV are predicted by using the trained multi-task prediction model. According to the business target, the expected income of the advertisement display is estimated in combination with the prediction result; In combination with the expected income of the advertisement display, the comprehensive score of each advertisement is calculated for the candidate advertisements, the advertisements are ranked and the highest-scored advertisement is selected for delivery.

6. An intelligent delivery decision system based on multi-dimension index correlation, characterized in that, The computer program product comprises a processor, a memory and a computer program stored on the memory, and the processor executes the computer program to specifically execute the steps in the intelligent delivery decision-making method based on multi-dimensional index association according to any one of claims 1-5.

Images