Artificial intelligence-based multi-channel advertisement effect prediction and optimization method and system

Abstract

The application discloses a multi-channel advertisement effect prediction and optimization method and system based on artificial intelligence, and the method comprises the following steps: acquiring advertisement delivery data, user behavior data and external environment data and preprocessing; feature extraction is performed on the preprocessed data, different types of content feature vectors are mapped to a unified semantic space through a cross-channel feature projection matrix, and an advertisement feature set is generated; a prediction model of fusing a Transformer and a graph neural network is used to synchronously predict the conversion rate, delivery return rate and user interaction index of a target advertisement in each channel based on the advertisement feature set; a strategy optimization network based on reinforcement learning is constructed, the prediction result output by the prediction model is taken as a state input, and a budget allocation matrix and a channel selection weight are dynamically generated through a strategy gradient algorithm; and the adjusted advertisement delivery strategy is pushed to an advertisement management platform according to the optimization strategy, and advertisement delivery is performed. The application can improve the accuracy and stability of advertisement effect prediction.

Description

Technical Field

[0001] This invention belongs to the field of advertising placement analysis technology, and specifically relates to a method and system for predicting and optimizing the effectiveness of multi-channel advertising based on artificial intelligence. Background Technology

[0002] With the rapid development of internet technology and the rise of digital marketing, advertising has evolved from the traditional single-channel model to a multi-channel, multi-terminal, and multi-scenario intelligent delivery approach. Businesses are increasingly reliant on programmatic advertising platforms, leveraging precise user profiles and behavioral analysis to deliver personalized ads. However, multi-channel advertising still faces numerous challenges, including the uncertainty of ad performance, the difficulty of optimizing delivery strategies, consistency issues across channels, and the intelligent optimization of ad creatives.

[0003] Chinese invention patent application CN113657926A discloses a method, system, electronic device, and storage medium for predicting advertising effectiveness. The method includes the following steps: determining a timeline for an advertising campaign and dividing the timeline into multiple discrete time periods based on the stages of the campaign; determining the form of a loss function, training an online learning model using sample information from the current time period, and calculating the loss function for the current time period; when the advertising campaign enters the next time period, retraining the online learning model and calculating the current loss function, further calculating the average loss function; updating the parameters of the online learning model based on the average loss function, and predicting advertising effectiveness based on the updated parameters.

[0004] This approach relies on a manually defined timeline divided into multiple discrete time periods. However, the effectiveness of advertising campaigns is often influenced by various factors, such as market conditions, changes in user interests, and competitor strategies. Fixed time period divisions may not be suitable for real-world scenarios, leading to insufficient model learning in some time periods and excessively frequent updates in others, thus affecting prediction accuracy. This approach primarily focuses on changes in the time dimension, but advertising effectiveness is not only time-dependent but also influenced by the attractiveness of the ad content itself and users' personalized interests. Relying solely on online learning models without adequately modeling user behavior sequences and ad-user relationship networks may result in lower prediction accuracy. Summary of the Invention

[0005] This invention provides a method and system for predicting and optimizing the effectiveness of multi-channel advertising based on artificial intelligence, aiming to solve the problems of insufficient prediction accuracy and stability in the field of advertising effectiveness prediction in existing technologies.

[0006] To address the aforementioned technical problems, this invention proposes a multi-channel advertising effectiveness prediction and optimization method based on artificial intelligence, comprising the following steps:

[0007] Acquire and preprocess advertising data, user behavior data, and external environment data;

[0008] Feature extraction is performed on the preprocessed data, and the feature vectors of different types of content are mapped to a unified semantic space through a cross-channel feature projection matrix to generate an advertising feature set;

[0009] A predictive model that integrates Transformer temporal encoder and graph neural network channel relationship modeling can simultaneously predict the conversion rate, return on investment and user interaction metrics of target ads on each channel based on the ad feature set.

[0010] A policy optimization network based on reinforcement learning is constructed, with the prediction results output by the prediction model as the state input, and the budget allocation matrix and channel selection weights are dynamically generated through the policy gradient algorithm.

[0011] Based on the optimization strategy, the adjusted advertising strategy is pushed to the advertising management platform for advertising execution.

[0012] Preferably, the content feature vectors include text, images, and videos, and the specific method for unified representation is as follows:

[0013]

[0014]

[0015]

[0016]

[0017] In the formula, As a fusion feature of the advertising feature set, These are text features, image features, and video features extracted using different methods. These are the query vector, key vector, and value vector, respectively. Let be the projection matrix, satisfying , and , They represent A real matrix of dimension This indicates that the similarity between the query vector and the key vector is calculated. It is a scaling factor used to prevent the value from becoming too large. This indicates Softmax normalization.

[0018] Preferably, the method for generating the advertising feature set is as follows:

[0019] An independent feature encoder is pre-trained for each advertising channel to generate an original content feature vector of dimension n;

[0020] Construct a trainable channel-specific projection matrix , where d is the dimension of the unified semantic space;

[0021] Through linear transformation The characteristics of each channel are projected onto a d-dimensional space, where Characteristics of the original channel The channel features are projected, and then layer normalization is used to eliminate the distribution differences between channels;

[0022] The cross-channel similarity loss function is calculated to optimize the projection matrix. The loss function is expressed as follows:

[0023]

[0024] in, For loss function, Indicates the channel after projection and layer normalization. Channel characteristics, Indicates the channel after projection and layer normalization. Channel characteristics, This represents the dot product operation. and Indicates correspondence and The Euclidean norm;

[0025] The optimized projection matrix is ​​used to map the original features to a unified semantic space, generating a complete set of advertising features.

[0026] Preferably, the prediction model uses Transformer to process user behavior sequences and outputs a representation of users' potential interest in advertisements; it uses graph neural network to process the advertisement-user-market relationship and outputs an advertisement feature vector; and it constructs a fusion layer to fuse the user interest vector extracted by Transformer with the advertisement feature vector extracted by graph neural network.

[0027] Preferably, the fusion method of the fusion layer is as follows:

[0028]

[0029]

[0030] In the formula, The control vector is used to control the latent interest representation. The retention ratio, Indicates reservation Feature section, Used to control advertising feature vectors The retention ratio, Indicates reservation Feature section, This indicates an element-wise multiplication operation. Given a learnable linear transformation matrix, Representation of features and splicing in the last dimension, This represents the Sigmoid activation function. These are the fused features in the prediction model.

[0031] Preferably, the fusion method of the fusion layer is as follows:

[0032] Feature importance gating is generated for user interest vectors and advertising feature vectors respectively;

[0033] Construct a bidirectional cross-attention mechanism, wherein the bidirectional cross-attention includes user→ad attention and ad→user attention;

[0034] The cross-attention results are then weighted and fused with the original features:

[0035]

[0036] In the formula, and These are gating mechanisms that calculate the importance of user and ad features. This represents user → ad attention. This indicates that advertising equates to user attention. This indicates an element-wise multiplication operation. It is a regularization parameter. This represents the class-based tagging features of users and ads after pooling. This indicates the class-labeled features used in processing users and advertisements using a multilayer perceptron.

[0037] Preferably, the strategy optimization network specifically comprises: using the output of the prediction model as the state input, constructing a neural network model, defining budget allocation and channel selection as actions, and a reward function to measure the optimization objective; wherein, the decision strategy of the neural network model is optimized using a policy gradient algorithm.

[0038] Preferably, the optimization steps of the policy gradient algorithm are as follows:

[0039] Sample a set of state-action pairs from the policy optimization network;

[0040] Calculate the corresponding reward based on the current state and actions;

[0041] The policy update is calculated using the policy gradient:

[0042]

[0043] In the formula, The result of the policy gradient calculation represents the parameters of the network at the current policy optimization stage. Below, the gradient of the policy, It is the objective function. This indicates that the objective function is relative to the parameters. The gradient; The expectation operator indicates that the current policy will be applied. To calculate the expectation; Indicates the state Select action The probability of; Logarithmic probability function relative to policy parameters The gradient; The immediate reward generated by the environment based on the current state and actions is used to measure the effect of the agent taking actions in the current state;

[0044] Use a gradient update strategy to optimize the network parameters.

[0045] Preferably, the parameter updates of the policy optimization network are performed using the Adam optimizer.

[0046] In another aspect, the present invention also proposes an artificial intelligence-based multi-channel advertising effectiveness prediction and optimization system, the system being used to implement the above-described multi-channel advertising effectiveness prediction and optimization method, comprising:

[0047] The data acquisition module is used to acquire advertising data, user behavior data, and external environment data, and to preprocess them to provide a cleaned dataset for subsequent feature extraction and modeling.

[0048] The feature extraction module is used to extract advertising content features, user behavior features, and external environment features from the preprocessed data. It uses a cross-channel feature projection matrix to map different types of content feature vectors to a unified semantic space to generate an advertising feature set.

[0049] The predictive model employs a combination of Transformer temporal encoder and graph neural network channel relationship modeling. Based on the advertising feature set, it predicts the conversion rate, return on investment, and user interaction metrics of the target advertisement on each channel.

[0050] The optimization module constructs a policy optimization network based on reinforcement learning, uses the results of the prediction model as state input, and dynamically generates a budget allocation matrix and channel selection weights through a policy gradient algorithm to optimize the advertising strategy.

[0051] The advertising management platform module receives optimized advertising strategies and pushes the adjusted strategies to the advertising management platform to realize actual advertising delivery.

[0052] Compared with the prior art, the present invention has the following technical effects:

[0053] 1. The advertising effectiveness prediction and optimization method proposed in this invention combines a Transformer temporal encoder and a graph neural network for channel relationship modeling. This model can better capture the complex temporal relationships between user behavior and advertising, as well as the multi-dimensional interactions between advertising, users, and the market. It can improve the accuracy of predicting conversion rates, ROI, and user interaction metrics for target ads across various channels, thereby helping advertisers make more accurate placement decisions, reduce ineffective placements, and increase advertising ROI.

[0054] 2. The advertising effectiveness prediction and optimization method proposed in this invention utilizes a cross-channel feature projection matrix to map different types of advertising content features (such as text, images, and videos) to a unified semantic space, generating a unified advertising feature set. This unified feature representation across channels allows for comparison and fusion of advertising performance across multiple channels within the same semantic space, avoiding prediction errors caused by feature inconsistencies between different channels, thereby improving model performance.

[0055] 3. The advertising effectiveness prediction and optimization method proposed in this invention constructs a policy optimization network based on reinforcement learning, uses the prediction results output by the prediction model as state input, and dynamically generates a budget allocation matrix and channel selection weights through a policy gradient algorithm to optimize the advertising delivery strategy. It can dynamically adjust the advertising budget allocation and channel selection weights based on real-time data, achieving a personalized and optimal advertising delivery strategy, and improving the overall advertising effectiveness and delivery efficiency.

[0056] 4. The advertising effectiveness prediction and optimization method proposed in this invention uses a multimodal feature fusion method to uniformly process the text, image, and video features of advertising content, and then fuses them through a bidirectional cross-attention mechanism. This fully utilizes various features within the advertising content, enhances the understanding of the advertising content, and thus improves the accuracy of advertising effectiveness prediction.

[0057] 5. The advertising effectiveness prediction and optimization method proposed in this invention optimizes advertising channel features through a cross-channel similarity loss function, mapping advertising features from different channels to a unified semantic space and reducing distribution differences between channels. This eliminates biases caused by data distribution differences between different channels, improving the uniformity and stability of multi-channel advertising strategies. Attached Figure Description

[0058] Figure 1 This is a flowchart illustrating the advertising effect prediction and optimization method described in this invention. Detailed Implementation

[0059] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with specific embodiments of the present application and with reference to the accompanying drawings.

[0060] In the field of advertising, channels refer to the platforms or media environments on which advertisements are placed, including but not limited to: advertising media platforms, such as Tencent Ads, Baidu Information Flow, Douyin Information Flow, etc.; advertising networks, such as DSP (Demand Side Platform), advertising alliances, RTB (Real-Time Bidding Platform), etc.; owned media or third-party media, such as corporate websites, mobile app ad slots, e-commerce platforms (such as Taobao, JD.com ads), etc.

[0061] Example 1

[0062] This embodiment describes a method for predicting and optimizing the effectiveness of multi-channel advertising based on artificial intelligence. Figure 1 As shown, it includes the following steps one through five:

[0063] Step 1: Acquire and preprocess advertising data, user behavior data, and external environment data.

[0064] Advertising data can be obtained from advertising platforms (such as Baidu, Tencent, etc.). The data includes basic information about the advertisement, such as the advertisement ID, advertisement type, campaign time, advertisement content, budget, bid, impressions, and clicks.

[0065] As an example, advertising data from this embodiment is shown in Table 1 below.

[0066] Table 1 Advertising Placement Data

[0067]

[0068] User behavior data comes from website analytics tools or user behavior tracking systems, and includes user browsing history, page dwell time, click events, purchase behavior, etc.

[0069] As an example, user behavior data from this embodiment is shown in Table 2 below.

[0070] Table 2 User Behavior Data

[0071]

[0072] External environmental data can be obtained from public data sources such as meteorological data, social events, and holidays. This data helps in understanding the external factors influencing advertising effectiveness and is typically obtained through APIs from weather services, social media analytics platforms, and the like.

[0073] As an example, one set of external environment data for this embodiment is shown in Table 3 below.

[0074] Table 3 External Environment Data

[0075]

[0076] The data in Tables 1-3 above can be recorded in JSON format. Corresponding to Table 1, the following two JSON data entries are generated:

[0077] {"Ad ID":"AD001","Ad Type":"Banner Ad","Placement Time":"2025 / 4 / 1","Ad Description":"Summer Sale","Ad Budget":1000 yuan","Bidding":2 yuan","Impressions":50000","Clicks":1200};

[0078] {"Ad ID":"AD002","Ad Type":"Video Ad","Placement Time":"2025 / 4 / 1","Ad Description":"New Smartphone Launch","Ad Budget":1500 yuan","Bid":3 yuan","Impressions":60000","Clicks":1500}

[0079] The data in Tables 2 and 3 are in the same JSON format as in Table 1.

[0080] For ad placement data, user behavior data, and external environment data, the first step is data cleaning, including missing value imputation, outlier detection, and duplicate data removal. Numerical features (such as ad budget and impressions) are normalized to ensure effective integration of data from different units. Categorical data (such as ad type and user ID) are encoded using one-hot encoding or label encoding methods. For ad placement time and user behavior time information, time features (such as hour, day of the week, and month) can be extracted as additional features to enhance the model's understanding of time-series data.

[0081] Specifically, missing values ​​are imputed using either mean imputation or median imputation for numerical data; for missing categorical data, mode imputation is used. Duplicate records are removed by checking the uniqueness of ad IDs and user IDs. Additionally, information such as hour, day of the week, and month is extracted from timestamps and input as extra features into the model.

[0082] As an example, the JSON data corresponding to ad ID AD001 in Table 1 above, after preprocessing, is represented as follows:

[0083] {"Ad ID":"AD001","Ad Type":"[1,0,0]","Placement Time (hours)":8,"Ad Budget":0.2,"Bid":0.1,"Impressions":0.8,"Clicks":0.15}.

[0084] After cleaning, normalization, and encoding, advertising data, user behavior data, and external environment data can provide high-quality input data for subsequent feature extraction, modeling, and optimization, ensuring the accuracy and efficiency of the entire advertising effect prediction and optimization system.

[0085] Step two involves feature extraction from the preprocessed data. A cross-channel feature projection matrix maps feature vectors of different content types to a unified semantic space, generating an advertising feature set. Since each advertising channel has its own feature representation, these features may originate from different data sources (such as text, images, click records, etc.). In different channels, advertising features may have different distributions and dimensions. The role of the cross-channel feature projection matrix is ​​to map these features from different channels into a unified semantic space, allowing advertising features from different channels to be compared, integrated, and further processed in the same space. Because the data and feature types of different channels are different, their feature spaces are also independent and differ. To enable analysis and comparison within a unified semantic space, a cross-channel feature projection matrix is ​​required.

[0086] In this step, the content feature vectors include text, images, and videos, and the specific method for unified representation is as follows:

[0087]

[0088]

[0089]

[0090]

[0091] In the formula, As a fusion feature of the advertising feature set, These are text features, image features, and video features extracted using different methods. These are the query vector, key vector, and value vector, respectively. Let be the projection matrix, satisfying , and , They represent A real matrix of dimension This indicates that the similarity between the query vector and the key vector is calculated. It is a scaling factor used to prevent the value from becoming too large. This indicates Softmax normalization. The projection matrix is ​​used for this process. The model weights, which are naturally learned during the joint training of advertising delivery data, user behavior data, and advertising content data, are used to map the advertising features of different modalities to a unified representation space in order to achieve feature fusion.

[0092] Specifically, extracting textual information from advertising content can be achieved through Natural Language Processing (NLP) techniques, such as using models like BERT to extract semantic representations of the advertisement. Feature extraction from images within the advertisement can be performed using convolutional neural networks to obtain image feature representations. Finally, video processing networks such as 3D-CNN or Transformer-based networks can be used to extract features from the video within the advertisement, resulting in video feature representations.

[0093] The method for generating the advertising feature set specifically includes the following steps S21-S25:

[0094] S21: Pre-train an independent feature encoder for each advertising channel to generate an original content feature vector of dimension n.

[0095] S22: Constructing a trainable channel-specific projection matrix , where d is the dimension of the unified semantic space. For each advertising channel i, there is a dedicated trainable projection matrix. .

[0096] S23: Through linear transformation The characteristics of each channel are projected onto a d-dimensional space, where Characteristics of the original channel The channel features are projected, and then layer normalization is used to eliminate distribution differences between channels. To eliminate distribution differences between different channels, layer normalization is performed on the projection results of each channel to normalize the mean and variance of each dimension, making the scale of each feature consistent, thereby reducing distribution differences between channels.

[0097] S24: Calculate the cross-channel similarity loss function to optimize the projection matrix. The loss function is expressed as follows:

[0098]

[0099] in, For loss function, Indicates the channel after projection and layer normalization. Channel characteristics, Indicates the channel after projection and layer normalization. Channel characteristics, This represents the dot product operation. and Indicates correspondence and The Euclidean norm.

[0100] S25: The optimized projection matrix is ​​used to map the original features to a unified semantic space, generating a complete advertising feature set. Optimized projection matrix It can be used to generate ad feature sets. Specifically, ad feature sets include not only static features of each ad, such as ad content and keywords, but also time-series dynamic features, such as ad click count, impression count, and user interaction behavior.

[0101] Step 3: The prediction model, which integrates the Transformer temporal encoder and the graph neural network channel relationship modeling, is used to simultaneously predict the conversion rate, return on investment, and user interaction metrics of the target advertisement on each channel based on the advertised feature set.

[0102] The prediction model uses a Transformer to process user behavior sequences and outputs a representation of the user's potential interest in the advertisement, i.e., a user interest vector; it uses a graph neural network to process the advertisement-user-market relationship and outputs an advertisement feature vector; a fusion layer is constructed to fuse the user interest vector extracted by the Transformer and the advertisement feature vector extracted by the graph neural network. As an example of this embodiment, a user interest vector with a dimension of 256 is represented as follows: An ad feature vector with dimension 256 is represented as: .

[0103] To capture temporal patterns in user behavior sequences, the model employs a Transformer temporal encoder. The Transformer, through its self-attention mechanism, captures dependencies in long sequences and processes user behavior patterns at different times. The user behavior sequence is input into the Transformer encoder, and the behavior data at each time step undergoes positional encoding and multi-head self-attention processing. The output, a representation of the user's latent interest, is a vector containing temporal information, reflecting the user's behavioral trends and interest changes at different points in time. This vector represents the user's latent interest in the advertisement, encompassing temporal dependencies and dynamic changes.

[0104] Graph neural networks are used to model the relationships between ads, users, and channels. Ads and users form a graph structure through historical interactions, while the relationship between ads and the market environment is established through dependencies between channels. Graph neural networks can propagate information through adjacency matrices, capturing the complex interactions and relationships between ads, users, and channels.

[0105] In one embodiment of the present invention, the fusion method of the fusion layer is as follows:

[0106]

[0107]

[0108] In the formula, The control vector is used to control the latent interest representation. The retention ratio, Indicates reservation Feature section, Used to control advertising feature vectors The retention ratio, Indicates reservation Feature section, This indicates an element-wise multiplication operation. Given a learnable linear transformation matrix, Representation of features and splicing in the last dimension, This represents the Sigmoid activation function. These are the fused features in the prediction model.

[0109] In another embodiment of the present invention, the fusion method of the fusion layer is as follows:

[0110] Feature importance gating is generated for user interest vectors and advertising feature vectors respectively;

[0111] Construct a bidirectional cross-attention mechanism, wherein the bidirectional cross-attention includes user→ad attention and ad→user attention;

[0112] The cross-attention results are then weighted and fused with the original features:

[0113]

[0114] In the formula, and These are gating mechanisms that calculate the importance of user and ad features. This represents user → ad attention. This indicates that advertising equates to user attention. This indicates an element-wise multiplication operation. It is a regularization parameter. This represents the class-based tagging features of users and ads after pooling. This indicates the class-labeled features used in processing users and advertisements using a multilayer perceptron.

[0115] By leveraging the synergistic effect of gating mechanisms and cross-attention, the problems of fixed feature dominance and coarse interaction granularity in traditional fusion methods are solved.

[0116] In steps one and two, the preprocessing of ad delivery data, user behavior data, and external environment data has been completed, and ad features and potential user interest representations have been processed and fused using Transformer and Graph Neural Network. Based on this workflow, an example of a prediction result output by the prediction model is shown below:

[0117] {"Channel":"Channel 1","Conversion Rate":0.12,"Return on Investment":3.5,"Click-Through Rate":0.15};

[0118] {"Channel":"Channel 2","Conversion Rate":0.08,"Return on Investment":2.8,"Click-Through Rate":0.12};

[0119] {"Channel":"Channel 3","Conversion Rate":0.20,"Return on Investment":4.2,"Click-Through Rate":0.18}.

[0120] Step four involves constructing a reinforcement learning-based policy optimization network. Using the prediction results from the prediction model as state input, the network dynamically generates the budget allocation matrix and channel selection weights through a policy gradient algorithm. In this embodiment, we will use a reinforcement learning-based method to dynamically optimize the advertising delivery strategy. Specifically, the policy optimization network uses a policy gradient algorithm to generate the advertising budget allocation matrix and channel selection weights to optimize the advertising delivery effect.

[0121] The strategy optimization network specifically comprises: using the output of the prediction model as the state input to construct a neural network model; defining budget allocation and channel selection as actions; and using a reward function to measure the optimization objective. The decision-making strategy of the neural network model is optimized using a policy gradient algorithm. The training of the strategy optimization network is based on historical advertising data and the output results of the prediction model. The training data includes budget configurations for different advertising channels and corresponding advertising performance metrics (such as click-through rate, conversion rate, and return on investment). Those skilled in the art can train this network using existing machine learning frameworks.

[0122] Specifically, the state is first defined, which includes the output results from the prediction model, i.e., advertising performance prediction information, such as conversion rate, return on investment, click-through rate, etc. These outputs constitute the state input of the model; the state is composed of the advertising performance information output by the prediction model. State information can include conversion rate predictions, return on investment predictions, and click-through rate predictions for different channels.

[0123] Based on the prediction results, the actions include a budget allocation matrix and channel selection weights, where: the budget allocation matrix represents the proportion of budget that should be allocated to each advertising channel, and the channel selection weights represent the probability or importance of each advertising channel being selected.

[0124] Define a reward function to measure the effectiveness of ad placement under the current strategy. The goal of the reward function is to maximize the return on ad placement. The reward function can be defined based on metrics such as total ad return and ad effectiveness.

[0125] By updating the parameters of the policy optimization network using the policy gradient algorithm, the network can optimize advertising budget allocation and channel selection strategies in each training iteration.

[0126] The optimization of the policy gradient algorithm includes the following steps S51-S54:

[0127] S51: Sample a set of state-action pairs from the policy optimization network.

[0128] S52: Calculate the corresponding reward based on the current state and action to evaluate the effect of the current action.

[0129] S53: Calculate policy updates using policy gradients:

[0130]

[0131] In the formula, The result of the policy gradient calculation represents the parameters of the network at the current policy optimization stage. Below, the gradient of the policy, It is the objective function. This indicates that the objective function is relative to the parameters. The gradient; The expectation operator indicates that the current policy will be applied. To calculate the expectation; Indicates the state Select action The probability of; Logarithmic probability function relative to policy parameters The gradient; The immediate reward generated by the environment based on the current state and actions is used to measure the effect of the agent taking actions in the current state.

[0132] S54: Optimize the network parameters using a gradient update strategy.

[0133] In one embodiment of the present invention, in order to optimize the parameters of the policy network and make the optimization of the policy gradient algorithm more efficient and stable, the parameter update of the policy optimization network uses the Adam optimizer.

[0134] After multiple training iterations and policy gradient optimization, the policy optimization network can eventually generate the optimal budget allocation matrix and channel selection weights, maximizing the effectiveness of ad delivery. For example, the final generated budget allocation and channel selection weights are as follows:

[0135] {"Channel":"Channel 1","Budget Allocation Weight":0.45};

[0136] {"Channel":"Channel 2","Budget Allocation Weight":0.30};

[0137] {"Channel":"Channel 3","Budget Allocation Weight":0.25}.

[0138] Step 5: Based on the optimization strategy, push the adjusted ad delivery strategy to the ad management platform and execute ad delivery. In the previous step, the optimal ad budget allocation and channel selection weights were obtained through reinforcement learning algorithms. This optimization strategy includes the budget allocation weights for each ad channel and the selection of ad delivery strategies.

[0139] Once the final optimization strategies are generated, the next step is to push these strategies to the ad management platform. The ad management platform is responsible for receiving the campaign plans from the optimization strategy system and executing the corresponding ad deliveries. The ad management platform provides an API interface for receiving and processing ad delivery strategies from the optimization strategy system. This interface receives information such as budget allocation weights for each channel, ad delivery time periods, ad types, and ad objectives.

[0140] The system encapsulates the optimized advertising budget allocation and channel selection weights into a delivery strategy. This strategy data is then sent to the advertising management platform via its API. The platform parses the received data, updates the advertising delivery settings, and begins preparing for ad delivery.

[0141] Example 2

[0142] This embodiment is an artificial intelligence-based multi-channel advertising effectiveness prediction and optimization system. The system is used to implement the multi-channel advertising effectiveness prediction and optimization method as described in Embodiment 1, including:

[0143] The data acquisition module is used to acquire advertising data, user behavior data, and external environment data, and preprocess them to provide a cleaned dataset for subsequent feature extraction and modeling.

[0144] The feature extraction module is used to extract advertising content features, user behavior features, and external environment features from the preprocessed data. It uses a cross-channel feature projection matrix to map different types of content feature vectors to a unified semantic space to generate an advertising feature set.

[0145] The predictive model employs a combination of Transformer temporal encoder and graph neural network channel relationship modeling. Based on the advertising feature set, it predicts the conversion rate, return on investment, and user interaction metrics of target ads across various channels.

[0146] The optimization module constructs a policy optimization network based on reinforcement learning, uses the results of the prediction model as state input, and dynamically generates a budget allocation matrix and channel selection weights through a policy gradient algorithm to optimize the advertising strategy.

[0147] The advertising management platform module receives optimized advertising strategies and pushes the adjusted strategies to the advertising management platform to realize actual advertising delivery.

[0148] In other embodiments of the present invention, the multi-channel advertising effectiveness prediction and optimization system further includes the following modules:

[0149] The strategy evaluation module monitors the effectiveness of advertising in real time, evaluates the results of advertising campaigns and the implementation of optimization strategies, and provides feedback to the system for further strategy optimization.

[0150] The computing resource module, based on efficient computing frameworks such as GPUs / TPUs, supports the training and inference of deep learning and reinforcement learning models, ensuring the efficiency of large-scale advertising data processing and real-time prediction.

[0151] The interface module provides API interfaces, enabling the advertising management platform to flexibly interact and integrate with other advertising delivery systems, data storage systems, user behavior analysis platforms, and so on.

[0152] The above description is only a preferred embodiment of the present invention. It should be noted that those skilled in the art can make several modifications and improvements without departing from the inventive concept of the present invention, and these all fall within the protection scope of the present invention.

Claims

1. A method for predicting and optimizing multi-channel advertising effects based on artificial intelligence, characterized in that, Includes the following steps: Acquire and preprocess advertising data, user behavior data, and external environment data; Feature extraction is performed on the preprocessed data, and the feature vectors of different types of content are mapped to a unified semantic space through a cross-channel feature projection matrix to generate an advertising feature set; A predictive model that integrates Transformer temporal encoder and graph neural network channel relationship modeling can simultaneously predict the conversion rate, return on investment and user interaction metrics of target ads on each channel based on the ad feature set. A policy optimization network based on reinforcement learning is constructed, with the prediction results output by the prediction model as the state input, and the budget allocation matrix and channel selection weights are dynamically generated through the policy gradient algorithm. Based on the optimization strategy, the adjusted advertising strategy is pushed to the advertising management platform for ad delivery; The method for generating the advertising feature set is as follows: An independent feature encoder is pre-trained for each advertising channel to generate an original content feature vector of dimension n; Constructing trainable channel-specific projection matrices where d is the unified semantic space dimensionality; Through linear transformation The characteristics of each channel are projected onto a d-dimensional space, where Characteristics of the original channel The channel features are projected, and then layer normalization is used to eliminate the distribution differences between channels; The cross-channel similarity loss function is calculated to optimize the projection matrix. The loss function is expressed as follows: in, For loss function, Indicates the channel after projection and layer normalization. Channel characteristics, Indicates the channel after projection and layer normalization. Channel characteristics, This represents the dot product operation. and Indicates correspondence and The Euclidean norm; The optimized projection matrix is ​​used to map the original features to a unified semantic space, generating a complete set of advertising features.

2. The method for predicting and optimizing the effectiveness of multi-channel advertising based on artificial intelligence according to claim 1, characterized in that, The content feature vectors include text, images, and videos, and the specific method for unified representation is as follows: In the formula, As a fusion feature of the advertising feature set, These are text features, image features, and video features extracted using different methods. These are the query vector, key vector, and value vector, respectively. Let be the projection matrix, satisfying , and , They represent A real matrix of dimension This indicates that the similarity between the query vector and the key vector is calculated. It is a scaling factor used to prevent the value from becoming too large. This indicates Softmax normalization.

3. The method for predicting and optimizing multi-channel advertising effectiveness based on artificial intelligence according to claim 1, characterized in that, The prediction model uses Transformer to process user behavior sequences and outputs a representation of users' potential interest in advertisements; it uses graph neural network to process the advertisement-user-market relationship and outputs an advertisement feature vector; and it constructs a fusion layer to fuse the user interest vector extracted by Transformer with the advertisement feature vector extracted by graph neural network.

4. The method for predicting and optimizing multi-channel advertising effectiveness based on artificial intelligence according to claim 3, characterized in that, The fusion method of the fusion layer is as follows: In the formula, The control vector is used to control the latent interest representation. The retention ratio, Indicates reservation Feature section, Used to control advertising feature vectors The retention ratio, Indicates reservation Feature section, This indicates an element-wise multiplication operation. Given a learnable linear transformation matrix, Representation of features and splicing in the last dimension, This represents the Sigmoid activation function. These are the fused features in the prediction model.

5. The method for predicting and optimizing the effectiveness of multi-channel advertising based on artificial intelligence according to claim 3, characterized in that, The fusion method of the fusion layer is as follows: Feature importance gating is generated for user interest vectors and advertising feature vectors respectively; Construct a bidirectional cross-attention mechanism, wherein the bidirectional cross-attention includes user→ad attention and ad→user attention; The cross-attention results are then weighted and fused with the original features: In the formula, and These are gating mechanisms that calculate the importance of user and ad features. This represents user → ad attention. This indicates that advertising equates to user attention. This indicates an element-wise multiplication operation. It is a regularization parameter. This represents the class-based tagging features of users and ads after pooling. This indicates the class-labeled features used in processing users and advertisements using a multilayer perceptron.

6. The method for predicting and optimizing the effectiveness of multi-channel advertising based on artificial intelligence according to claim 1, characterized in that, The policy optimization network specifically involves: using the output of the prediction model as the state input to construct a neural network model, defining budget allocation and channel selection as actions, and using a reward function to measure the optimization objective; wherein, the decision-making strategy of the neural network model is optimized using a policy gradient algorithm.

7. The method for predicting and optimizing multi-channel advertising effectiveness based on artificial intelligence according to claim 6, characterized in that, The optimization steps of the policy gradient algorithm are as follows: Sample a set of state-action pairs from the policy optimization network; Calculate the corresponding reward based on the current state and actions; The policy update is calculated using the policy gradient: In the formula, The result of the policy gradient calculation represents the optimization of network parameters under the current policy. Lower objective function Regarding parameters gradient, It is the objective function. This indicates that the objective function is relative to the parameters. The gradient; The expectation operator indicates that the current policy will be applied. To calculate the expectation; Indicates the state Select action The probability of; Logarithmic probability function relative to policy parameters The gradient; The immediate reward generated by the environment based on the current state and actions is used to measure the effect of the agent taking actions in the current state; Use a gradient update strategy to optimize the network parameters.

8. The method for predicting and optimizing the effectiveness of multi-channel advertising based on artificial intelligence according to claim 7, characterized in that, The policy optimizes the network's parameter updates using the Adam optimizer.

9. A multi-channel advertising effectiveness prediction and optimization system based on artificial intelligence, characterized in that, The system is used to implement the multi-channel advertising effectiveness prediction and optimization method as described in any one of claims 1-8, including: The data acquisition module is used to acquire advertising data, user behavior data, and external environment data, and to preprocess them to provide a cleaned dataset for subsequent feature extraction and modeling. The feature extraction module is used to extract advertising content features, user behavior features, and external environment features from the preprocessed data. It uses a cross-channel feature projection matrix to map different types of content feature vectors to a unified semantic space to generate an advertising feature set. The predictive model employs a combination of Transformer temporal encoder and graph neural network channel relationship modeling. Based on the advertising feature set, it predicts the conversion rate, return on investment, and user interaction metrics of the target advertisement on each channel. The optimization module constructs a policy optimization network based on reinforcement learning, uses the results of the prediction model as state input, and dynamically generates a budget allocation matrix and channel selection weights through a policy gradient algorithm to optimize the advertising strategy. The advertising management platform module receives optimized advertising strategies and pushes the adjusted strategies to the advertising management platform to realize actual advertising delivery.

Images