Advertisement delivery adjustment system and method
By monitoring and converging user behavior data from multiple media platforms in real time in the advertising delivery system, combining third-party data sources, and using reinforcement learning and neural network models for advertising delivery adjustment, the inaccurate user portrait and delivery adjustment in the existing system are solved. The problem of not being smart improves the accuracy and user experience of advertising delivery.
Patent Information
- Application Number
- CN202410830924.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-06-26
- Publication Date
- 2024-11-05
- Estimated Expiration
- 2044-06-26
AI Technical Summary
The existing advertising delivery system lacks cross-platform data integration capabilities, and the generated user portraits are not accurate enough to fully reflect the user's multi-dimensional characteristics and real-time interests, resulting in a lack of intelligence and real-time nature of advertising delivery adjustments, which affects the effect and user experience.
A system for advertising delivery adjustment is designed, including user behavior monitoring module, data fusion module and advertising effect evaluation module. By monitoring user behavior data from multiple media platforms in real time, integrating third-party data sources, using reinforcement learning algorithms and hybrid neural networks The model makes advertising delivery adjustments, including changing creatives, adjusting delivery platforms and optimizing delivery time periods, personalized content recommendations and user interaction mode adjustments.
It realizes comprehensive monitoring and accurate image generation of cross-platform user behavior data, improves the accuracy and user experience of advertising delivery, can quickly respond to changes in user behavior and improve advertising effectiveness.
Smart Images

Figure CN118379095B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of computer technology, and in particular to an advertisement delivery adjustment system and method. Background Art
[0002] In the existing technology, advertising delivery systems usually rely on data collection and analysis of a single platform, and evaluate the effectiveness of advertising delivery by monitoring user behavior data on the platform (such as click-through rate, dwell time, etc.). These systems usually use basic data analysis methods to generate user portraits and adjust advertising delivery based on preset rules or simple algorithms. However, the control capabilities and accuracy of these existing systems are often limited by data sources and analysis methods, and cannot fully reflect user behavior and demand changes across platforms.
[0003] The main problems with existing technologies are: 1) lack of cross-platform data integration capabilities, making it difficult to fully obtain user behavior data; 2) the generated user portraits are not accurate enough to fully reflect the user's multi-dimensional characteristics and real-time interests; 3) advertising adjustments lack intelligence and real-time performance, and are unable to respond quickly to immediate changes in user behavior, thus affecting the effectiveness of advertising and user experience.
[0004] Therefore, there is an urgent need to develop a new advertisement delivery adjustment system and method. Summary of the invention
[0005] The present application provides an advertisement delivery adjustment system and method to improve the accuracy of advertisement delivery and user experience.
[0006] The present application provides an advertisement delivery adjustment system, comprising:
[0007] A user behavior monitoring module, which is used to monitor and collect user behavior data in real time on multiple media platforms, wherein the behavior data includes the user's click rate, dwell time, interaction frequency, browsing path and purchase history;
[0008] A data fusion module, used to fuse the collected behavioral data with third-party data sources to generate user portraits; wherein the third-party data sources include social media data and geographic location data;
[0009] An advertising effect evaluation module, which is used to perform a context-aware prediction evaluation on the advertising effect based on the user portrait and behavior data, wherein the evaluation results of the context-aware prediction evaluation include the advertising click-through rate, conversion rate, user retention rate and emotional response score;
[0010] A dynamic adjustment module is used to adjust the advertisement delivery using a reinforcement learning algorithm based on the evaluation results of the context-aware predictive evaluation, wherein the adjustment includes replacing the advertising material, adjusting the delivery platform and optimizing the delivery time period, personalized content recommendation and user interaction mode adjustment.
[0011] Furthermore, the user behavior monitoring module includes a data storage unit for storing user behavior data monitored and collected in real time.
[0012] Furthermore, the data fusion module includes a data cleaning unit for cleaning the collected behavior data and data from a third-party data source to remove noise and redundant data.
[0013] Furthermore, the advertising effect evaluation module uses a trained hybrid neural network model to perform context-aware prediction and evaluation on the advertising effect based on the user portrait and behavior data; the hybrid neural network model includes an input layer, a conditional generative adversarial network layer, a dynamic routing layer, an attention mechanism layer and an output layer; wherein the input layer accepts user portraits and behavior data as input, and performs standardization and normalization preprocessing on the input data to obtain preprocessed data; the generative adversarial network layer includes a generator and a discriminator, and the generator generates a potential feature representation based on the preprocessed data; the discriminator distinguishes the true and false output of the generator, and feeds back the error to the generator for improvement; the dynamic routing layer is implemented using a capsule network, processes the potential feature representation, and obtains a feature vector after routing adjustment; the attention mechanism layer dynamically assigns importance weights of different features by calculating the feature weighted average of the feature vector after routing adjustment, and generates a weighted feature vector; the output layer is used to receive the weighted feature vector, and uses four independent neurons to output the predicted values of the advertising click-through rate, conversion rate, user retention rate and emotional response score.
[0014] The present application provides an advertisement delivery adjustment method, comprising:
[0015] Real-time monitoring and collection of user behavior data on multiple media platforms, wherein the behavior data includes user click-through rate, dwell time, interaction frequency, browsing path and purchase history;
[0016] Fusion of the collected behavioral data with third-party data sources to generate user profiles; wherein the third-party data sources include social media data and geographic location data;
[0017] Based on the user portrait and behavior data, a situation-aware prediction evaluation is performed on the advertising effect, wherein the evaluation results of the situation-aware prediction evaluation include the advertising click-through rate, conversion rate, user retention rate and emotional response score;
[0018] Based on the evaluation results of the context-aware predictive evaluation, a reinforcement learning algorithm is used to adjust the advertisement delivery, wherein the adjustment includes replacing the advertisement material, adjusting the delivery platform and optimizing the delivery time period, personalized content recommendation and user interaction mode adjustment.
[0019] The beneficial effects of this application mainly include: (1) Through the user behavior monitoring module, user behavior data can be monitored and collected in real time on multiple media platforms, including click-through rate, dwell time, interaction frequency, browsing path and purchase history, so as to obtain comprehensive user behavior information. This cross-platform monitoring capability can provide more complete and accurate user behavior data, which helps to fully understand the interests and needs of users. (2) The data fusion module integrates the collected behavior data with third-party data sources (such as social media data and geographic location data) to generate a more accurate user portrait. This multi-source data fusion can improve the accuracy and richness of user portraits, reflect the multi-dimensional characteristics of users, and provide more accurate basic data for advertising. BRIEF DESCRIPTION OF THE DRAWINGS
[0020] Figure 1 It is a schematic diagram of an advertisement delivery adjustment system provided in the first embodiment of the present application.
[0021] Figure 2 This is a flow chart of an advertisement delivery adjustment method provided in the second embodiment of the present application. DETAILED DESCRIPTION
[0022] Many specific details are described in the following description to facilitate a full understanding of the present application. However, the present application can be implemented in many other ways than those described herein, and those skilled in the art can make similar generalizations without violating the connotation of the present application, so the present application is not limited by the specific implementation disclosed below.
[0023] The first embodiment of the present application provides an advertisement delivery adjustment system. Figure 1 , which is a schematic diagram of the first embodiment of the present application. Figure 1 A first embodiment of the present application provides an advertisement delivery adjustment system for detailed description.
[0024] The advertisement delivery adjustment system includes a user behavior monitoring module 101 , a data fusion module 102 , an advertisement effect evaluation module 103 and a dynamic adjustment module 104 .
[0025] The user behavior monitoring module 101 is used to monitor and collect user behavior data in real time on multiple media platforms, wherein the behavior data includes the user's click rate, dwell time, interaction frequency, browsing path and purchase record.
[0026] The user behavior monitoring module 101 is used to monitor and collect user behavior data in real time on multiple media platforms. Specifically, the user behavior monitoring module 101 can interact with multiple different media platforms through API (application programming interface) to obtain a series of user behavior data on these platforms. These data include but are not limited to the user's click rate, dwell time, interaction frequency, browsing path and purchase record.
[0027] Click-through rate refers to the ratio of the number of times a user clicks on an advertisement or content to the number of times the advertisement or content is displayed. The user behavior monitoring module 101 can record each user click event and calculate the corresponding click-through rate. The dwell time is the length of time a user stays on an advertisement or content page. The user behavior monitoring module 101 accurately calculates the dwell time by monitoring the timestamps of the user entering and leaving the page. The interaction frequency refers to the frequency with which a user interacts with page elements on an advertisement or content page, such as liking, commenting, sharing, etc. The user behavior monitoring module 101 can record the number of occurrences of these interactive behaviors and perform statistical analysis.
[0028] The browsing path refers to the navigation path of the user between various pages. The user behavior monitoring module 101 can track and record the various pages and links clicked by the user from entering a certain website or application, forming a complete browsing path record. The purchase record refers to the purchase behavior data of the user on the platform, including detailed information such as the purchased goods, purchase time, transaction amount, etc. The user behavior monitoring module 101 can obtain this data from the e-commerce platform or social media with transaction functions to form the user's purchase behavior record.
[0029] In order to achieve the above functions, the user behavior monitoring module 101 includes multiple submodules. These submodules include a data collection submodule, a data processing submodule and a data storage submodule. The data collection submodule is connected to the API interface of each platform to obtain user behavior data in real time. The data processing submodule pre-processes the collected data, including data cleaning, format conversion and deduplication processing to ensure the accuracy and consistency of the data. The data storage submodule stores the processed data in a database for subsequent analysis and processing.
[0030] The design and implementation of the user behavior monitoring module 101 needs to consider the real-time and accuracy of the data. To this end, the system adopts efficient data collection and processing algorithms, and uses distributed storage technology to ensure efficient storage and fast access to large-scale data. In addition, the user behavior monitoring module 101 also needs to work closely with other modules, such as the data fusion module 102, the advertising effect evaluation module 103 and the dynamic adjustment module 104, to ensure the efficient operation of the entire advertising delivery adjustment system.
[0031] The user behavior monitoring module 101 collects and processes the user's behavior data on multiple media platforms, providing reliable data support for the optimization of advertising delivery.
[0032] Furthermore, the user behavior monitoring module includes a data storage unit for storing user behavior data monitored and collected in real time.
[0033] Specifically, the user behavior monitoring module continuously monitors various user behavior data on multiple media platforms, including but not limited to the user's click-through rate, dwell time, interaction frequency, browsing path and purchase history. In order to effectively manage and utilize this large amount of behavior data, the system has designed a data storage unit.
[0034] The data storage unit uses high-performance storage technology and can quickly write and read large amounts of data. This unit can be a cloud storage-based database or a local high-performance database system. The specific choice depends on the actual needs and design of the system. The data storage unit must not only be able to store the original user behavior data, but also have certain data processing capabilities, such as data compression, deduplication, and indexing, to ensure data integrity and accessibility.
[0035] During the operation of the user behavior monitoring module, when the system detects new user behavior data, these data will be written into the data storage unit in real time. The system uses stream processing technology to process high-frequency data writing through message queues or data stream platforms (such as Apache Kafka), ensuring that the data storage unit can cope with large-scale concurrent data writing requests.
[0036] The stored data includes detailed user behavior records, such as the timestamp of each ad click, the time the user stays on the ad page, the user's interactive behavior (such as likes, comments, and shares), and the user's purchase behavior. After preprocessing and cleaning, these data will be stored in a structured manner for subsequent data fusion and analysis.
[0037] In addition, the data storage unit also needs to have high reliability and security. To this end, the system uses multiple copies of storage and automatic backup mechanisms to prevent data loss. At the same time, the data storage unit needs to comply with relevant data privacy and protection laws and regulations to ensure that user data is encrypted and protected during storage and transmission.
[0038] In short, through the data storage unit, the user behavior monitoring module can efficiently and securely store and manage user behavior data collected in real time, providing a solid data foundation for subsequent data fusion, advertising effect evaluation and dynamic adjustment.
[0039] The data fusion module 102 is used to fuse the collected behavioral data with a third-party data source to generate a user portrait; wherein the third-party data source includes social media data and geographic location data.
[0040] The data fusion module 102 plays a key role in the advertising delivery adjustment system, and is intended to fuse the user behavior data collected from the user behavior monitoring module 101 with the data from the third-party data source to generate an accurate user portrait. The implementation of this module requires multiple technical means and steps to ensure the accuracy and effectiveness of data fusion.
[0041] First, the data fusion module 102 receives data from the user behavior monitoring module 101, including the user's click rate, dwell time, interaction frequency, browsing path and purchase record on multiple media platforms. These data are collected and transmitted to the data fusion module 102 through the API interface.
[0042] At the same time, the data fusion module 102 also obtains supplementary data from various third-party data sources. These third-party data sources include social media data and geographic location data. Social media data may include user interaction information, friend relationships, published content and comments on different social platforms. Geographic location data records user location information, such as frequently visited places, movement trajectories and geographic preferences. These data are connected and obtained through the API interface of the third-party data source.
[0043] In the data fusion module 102, data preprocessing is first performed. The preprocessing steps include data cleaning, deduplication and format conversion. Data cleaning refers to removing noise data and incomplete data to ensure data accuracy. Deduplication is to eliminate duplicate records and ensure data consistency. Format conversion converts data from different sources into a standard format that can be processed by the system.
[0044] After the preprocessing is completed, the data fusion module 102 starts data matching and fusion. The user behavior data is matched with the data from the third-party data source through the unique user identifier (such as user ID, device ID, etc.). This step ensures that data from different sources can accurately correspond to the same user. Then, the user information from different data sources is integrated using the data fusion algorithm to generate a comprehensive user profile.
[0045] The generated user portrait not only includes the user's behavior data on various platforms, but also includes the user's interaction information on social media and geographic location data. Through this multi-dimensional data fusion, we can have a more comprehensive understanding of the user's interests, behavior habits and needs.
[0046] In order to ensure the accuracy and real-time performance of data fusion, the data fusion module 102 also uses machine learning algorithms and real-time data processing technology. Machine learning algorithms can be used to predict and fill missing values in data to improve data integrity. Real-time data processing technology ensures that the data fusion process can quickly respond to and process large-scale data to meet the system's real-time requirements.
[0047] Finally, the data fusion module 102 stores the generated user portraits in the database for use by the advertising effect evaluation module 103 and the dynamic adjustment module 104. These user portraits will serve as an important basis for advertising effect evaluation and strategy adjustment, helping the system to achieve accurate advertising delivery.
[0048] Furthermore, the data fusion module includes a data cleaning unit for cleaning the collected behavior data and data from a third-party data source to remove noise and redundant data.
[0049] The data cleaning unit is designed to ensure that the data input into the system is accurate, high quality, and suitable for subsequent analysis and processing.
[0050] In the process of collecting behavioral data and data from third-party data sources, the data may contain various noise and redundant information. For example, user behavior data may contain duplicate click records, invalid dwell time records, or erroneous data caused by network delays. Similarly, data provided by third-party data sources may also have inconsistent formats, missing values, or be duplicated with existing data. The role of the data cleaning unit is to identify and handle these problems to ensure the integrity and accuracy of the data.
[0051] The data cleaning unit first preprocesses the collected raw data. This step includes format conversion, data type verification, and basic error detection. The preprocessed data will be further analyzed, and the cleaning unit uses a series of rules and algorithms to identify noise and redundant data. For example, for duplicate records in behavioral data, the cleaning unit will identify and delete these duplicates by matching timestamps and user IDs. For invalid residence time records, the cleaning unit will set a reasonable time range, and data exceeding or below this range will be regarded as abnormal and processed.
[0052] When dealing with missing values, the data cleaning unit can use a variety of methods, such as filling missing values, deleting records with missing values, or using interpolation to estimate missing values. The method to choose depends on the specific data characteristics and application requirements. For data from different data sources, the cleaning unit also needs to align and standardize the data to ensure that data from different sources can be processed and integrated consistently.
[0053] The cleaned data will be stored in an intermediate database for subsequent integration and analysis. During the entire data cleaning process, the data cleaning unit needs to record the processing logs of each step for subsequent data tracing and problem troubleshooting. These log information includes which data was deleted or modified, the reasons for the modification, and the rules based on which it was modified.
[0054] In addition, the data cleaning unit should have efficient processing capabilities to cope with the cleaning needs of large-scale data. To this end, a distributed computing framework (such as Apache Spark) can be used to accelerate the data cleaning process. At the same time, the cleaning unit needs to have high reliability to ensure that no data is lost during the data processing process and can automatically recover the failed tasks.
[0055] Through the above process, the data cleaning unit can effectively remove noise and redundant data, and provide high-quality input data for the data fusion module. This not only improves the accuracy of data fusion and analysis, but also provides a reliable data basis for the evaluation and adjustment of advertising effects.
[0056] The advertising effect evaluation module 103 is used to perform a context-aware predictive evaluation on the advertising effect based on the user portrait and behavior data. The evaluation results of the context-aware predictive evaluation include the advertising click-through rate, conversion rate, user retention rate and emotional response score.
[0057] The advertising effect evaluation module 103 plays a vital role in the advertising delivery adjustment system. Its function is to perform a situation-aware prediction and evaluation of the advertising delivery effect based on user portraits and behavior data. The implementation of this module requires a variety of technologies and methods to ensure the accuracy and real-time nature of the evaluation and provide a reliable basis for subsequent advertising delivery adjustments.
[0058] The advertising effect evaluation module 103 first receives the user portrait and user behavior data from the data fusion module 102. These user portraits include the user's behavior data on multiple media platforms, social media data, and geographic location data. These comprehensive data enable the system to fully understand the user's interests and behavior patterns.
[0059] When evaluating the effectiveness of advertising, the advertising effectiveness evaluation module 103 uses a variety of indicators to measure the effectiveness of advertising. One of the key indicators is the click-through rate of advertising, which reflects the frequency of the advertisement being clicked by users. By recording the number of times users click on the advertisement and the number of times the advertisement is displayed, the system can calculate the click-through rate. Conversion rate is another important indicator, which refers to the proportion of users who actually complete a purchase or other target behavior after clicking on the advertisement. The user retention rate is used to evaluate the impact of advertising on the continued retention of users, and is measured by calculating the proportion of users who continue to interact with the platform within a certain period of time. The emotional response score judges the user's emotional response to the advertisement by analyzing the user's interactive behaviors such as comments, likes, and sharing on the advertisement.
[0060] In order to achieve context-aware prediction and evaluation, the advertising effect evaluation module 103 also combines the user's context data, such as the user's current location, time and weather conditions. These context data can be extracted from the geographic location data and timestamps, and combined with the user behavior data, so that the module can more accurately evaluate the effectiveness of advertisements in different scenarios. For example, it can be evaluated whether advertisements placed at a specific time and place can more effectively attract the user's attention and generate interaction.
[0061] The advertising effectiveness evaluation module 103 uses advanced data analysis techniques and machine learning algorithms to process and analyze these data. First, the module cleans and normalizes the raw data through data preprocessing steps to ensure the accuracy and consistency of the data. Then, the machine learning algorithm is used to model user behavior patterns and identify which factors have a significant impact on advertising effectiveness. Based on these models, the module can predict and evaluate the effectiveness of the current advertising and generate a detailed evaluation report.
[0062] In addition, the advertising effect evaluation module 103 also has prediction capabilities. By accumulating and analyzing historical data, the module can predict future advertising effects. These prediction results can provide valuable references for advertisers to help them optimize advertising delivery strategies. For example, based on past user behavior and scenario data, the module can predict advertising effects in similar scenarios, thereby helping advertisers better plan and execute advertising delivery strategies.
[0063] The evaluation results of the advertisement effect evaluation module 103 will be passed to the dynamic adjustment module 104 for subsequent advertisement delivery adjustment. These evaluation results include not only quantitative indicators, but also qualitative user feedback information to help the dynamic adjustment module 104 make more accurate and effective adjustment decisions.
[0064] In summary, the advertising effectiveness evaluation module 103 combines user portraits, behavioral data and situational data, and uses advanced data analysis and machine learning technologies to achieve a comprehensive, real-time and situation-aware evaluation of the effectiveness of advertising, providing reliable data support and decision-making basis for advertising adjustments.
[0065] Furthermore, the advertising effect evaluation module uses a trained hybrid neural network model to perform context-aware prediction and evaluation on the advertising effect based on the user portrait and behavior data; the hybrid neural network model includes an input layer, a conditional generative adversarial network layer, a dynamic routing layer, an attention mechanism layer and an output layer; wherein the input layer accepts user portraits and behavior data as input, and performs standardization and normalization preprocessing on the input data to obtain preprocessed data; the generative adversarial network layer includes a generator and a discriminator, and the generator generates a potential feature representation based on the preprocessed data; the discriminator distinguishes the true and false output of the generator, and feeds back the error to the generator for improvement; the dynamic routing layer is implemented using a capsule network, processes the potential feature representation, and obtains a feature vector after routing adjustment; the attention mechanism layer dynamically assigns importance weights of different features by calculating the feature weighted average of the feature vector after routing adjustment, and generates a weighted feature vector; the output layer is used to receive the weighted feature vector, and uses four independent neurons to output the predicted values of the advertising click-through rate, conversion rate, user retention rate and emotional response score.
[0066] First, the input layer accepts user portrait and behavior data as input and performs standardization and normalization preprocessing on the input data. This step is necessary to ensure that the data has a consistent scale and distribution in the subsequent processing stage, thereby improving the training effect and prediction accuracy of the model. After standardization and normalization preprocessing, the data will be passed to the next network layer in an appropriate format.
[0067] Next, the preprocessed data is input into the conditional generative adversarial network layer. This layer includes two sub-networks: the generator and the discriminator. The generator generates potential feature representations based on the preprocessed data, which can capture the high-order features of the data. The discriminator receives the output of the generator and distinguishes whether it is true or false. The discriminator feeds back the error to the generator, guiding the generator to improve its generated feature representation so that it is closer to the distribution of real data. This process continuously optimizes the generator and the discriminator through adversarial training, making the feature representation generated by the generator more and more real and useful.
[0068] The latent feature representations generated by the generator are then passed to the dynamic routing layer. The dynamic routing layer is implemented using a capsule network, which aims to process the latent feature representations to capture the complex relationships between features. The capsule network dynamically adjusts the routing coefficients so that the feature representations can more accurately reflect the nature of the input data. In the dynamic routing layer, the initial capsule is mapped to the capsule space of the next layer through a series of transformation matrices. Through multiple iterations, the routing coefficients are continuously updated according to the similarity between the initial capsule and the target capsule, and finally the capsule vectors after routing adjustment are output. These vectors can more accurately convey the structural information of the input data.
[0069] The feature vector processed by the dynamic routing layer is input to the attention mechanism layer. The task of the attention mechanism layer is to dynamically assign importance weights to different features by calculating the weighted average of the feature vectors. This process weights the input feature vectors so that the model can focus on the features that contribute most to the prediction results. Finally, the attention mechanism layer outputs weighted feature vectors that combine the information of all important features and provide more accurate input for the final prediction.
[0070] Finally, the weighted feature vector is input to the output layer. The output layer includes four independent neurons, corresponding to the ad click-through rate, conversion rate, user retention rate, and emotional response score. Each neuron uses the appropriate activation function and loss function to calculate according to the input weighted feature vector, and finally outputs the predicted value. These predicted values can reflect all aspects of the advertising effect and help the advertising system make more accurate adjustments and optimizations.
[0071] Furthermore, the generator adopts a deep convolution generator network, including a first input layer, a first convolution layer, a first deconvolution layer and a first output layer; the first input layer is used to receive the preprocessed data; the first convolution layer is implemented by multiple stacked convolution layers, and is used to process the output of the first input layer to obtain a high-order feature map; the first deconvolution layer includes multiple deconvolution layers, and is used to upsample the high-order feature map extracted by the convolution layer to the size of the original input data to obtain a deconvolution feature representation; the first output layer is used to generate a potential feature representation based on the deconvolution feature representation.
[0072] First, the first input layer is used to receive preprocessed data, which includes user portraits and behavior data. The preprocessed data is standardized and normalized to ensure that it has a consistent scale and distribution, so that the data can be input in the best state in the subsequent neural network processing.
[0073] Next, the output of the first input layer is passed to the first convolutional layer. This convolutional layer consists of multiple stacked convolutional layers, each of which is responsible for extracting different features of the input data. In these convolutional layers, the multi-layer perceptron method is used to gradually extract high-order features through layer-by-layer convolution operations. Each convolutional layer is followed by batch normalization and leaky ReLU activation functions, which can stabilize the training process and prevent the gradient vanishing problem while maintaining the sparsity and nonlinear changes of the features. Finally, the first convolutional layer outputs a set of high-order feature maps, which contain deep information of the input data.
[0074] These high-order feature maps are then passed to the first deconvolution layer. The first deconvolution layer includes multiple deconvolution layers, which upsample the high-order feature maps extracted by the convolution layer. The upsampling process gradually restores the low-resolution feature maps to the same size as the original input data through deconvolution operations. In each deconvolution layer, batch normalization and ReLU activation functions are applied to ensure that the features remain stable during the upsampling process and can capture more spatial information. After a series of deconvolution processes, deconvolution feature representations are output, which contain high-quality feature maps of the same size as the original input data.
[0075] Finally, the deconvolution feature representations are passed to the first output layer. In the first output layer, these feature representations are further processed to generate latent feature representations. The first output layer converts the deconvolution feature representations into latent feature representations through a series of weighted summations of neurons and nonlinear transformations of activation functions. These latent feature representations have high-dimensional feature information and can fully characterize the essential attributes of the input data.
[0076] Through the design of this deep convolutional generator network, the generator can efficiently process the preprocessed data, extract deep features from it, and generate high-quality latent feature representations. These latent feature representations provide rich feature information for the subsequent discriminator and other network layers, ensuring that the effect evaluation and optimization process of the entire advertising delivery adjustment system can be carried out accurately and stably.
[0077] Furthermore, the discriminator is implemented using a multilayer perceptron structure, which includes a second input layer, a second hidden layer, and a second output layer; wherein the second input layer is used to accept potential feature representations and real data generated by the generator; the second hidden layer includes multiple neurons for processing the data of the second input layer; and the output layer is used to process the output of the second hidden layer and output a true or false discrimination result.
[0078] Specifically, the second input layer is designed to accept the latent feature representations and real data generated by the generator. These input data enter the network structure of the discriminator through the input layer. The latent feature representations come from the generator, which are high-quality features generated by the generator based on preprocessed user portraits and behavior data, while the real data is directly extracted from the actual data. The second input layer normalizes these data to make them suitable for subsequent neural network calculations.
[0079] After entering the second hidden layer, the data will undergo a series of complex nonlinear transformations and processing. The second hidden layer contains multiple neurons, each of which performs a weighted summation of the input data and performs a nonlinear transformation through an activation function. The role of the hidden layer is to extract deep features in the input data and combine these features in a highly nonlinear way to capture the complex patterns and relationships of the data. Common activation functions such as Leaky ReLU can be used between each hidden layer to ensure that the network can capture the nonlinear relationship of the data during processing and that the parameters can be effectively updated during the back-propagation process.
[0080] Finally, the features processed by the hidden layer are passed to the second output layer. The purpose of the output layer is to map the complex features of the hidden layer to a simpler space for the final true and false discrimination. Specifically, the output layer performs a weighted summation on the output of the hidden layer and compresses the output value to a range between 0 and 1 through the Sigmoid activation function. This output value represents the discriminator's confidence in whether the input data is real data or generated data.
[0081] During the training process, the discriminator continuously improves its ability to distinguish generated data from real data through adversarial training with the generator. The generator and the discriminator compete with each other. Through this adversarial method, the potential feature representation generated by the generator will become more and more realistic, and the discriminator's discriminative ability will continue to increase.
[0082] Through the discriminator of this multi-layer perceptron structure, the advertising adjustment system can effectively distinguish the true and false between the potential feature representations generated by the generator and the real data. This structural design ensures that the system can accurately distinguish between high-quality generated features and actual data, thus providing a reliable basis for evaluating the effectiveness of advertising.
[0083] Furthermore, the dynamic routing layer is implemented using a capsule network, which includes an initial capsule layer, a capsule transformation layer and a dynamic routing layer; the initial capsule layer is used to receive the potential feature representation generated by the conditional generative adversarial network layer and decompose it into multiple initial capsules, each capsule containing a vector; the capsule transformation layer is used to apply a transformation matrix to the initial capsule to obtain a transformed capsule vector; the dynamic routing layer uses a dynamic routing algorithm to transfer information between capsules to generate a capsule vector after routing adjustment.
[0084] The initial capsule layer receives the latent feature representations generated by the conditional generative adversarial network layer and decomposes these feature representations into multiple initial capsules. Each capsule contains a vector that represents the attributes of a specific feature dimension. Through this decomposition, the initial capsule layer is able to structure complex feature representations into multiple vectors, providing more refined and actionable units for subsequent processing.
[0085] After the initial capsule layer, the generated capsule vectors are passed to the capsule transformation layer. In this layer, a specific transformation matrix is applied to each initial capsule vector. These transformation matrices are learned through model training to capture and represent complex relationships between features. By applying these transformation matrices, the initial capsule vectors are mapped to a new capsule space to generate transformed capsule vectors. These transformed vectors retain the core information of the original data while enhancing the high-order relationships represented by the features.
[0086] The transformed capsule vectors then enter the dynamic routing layer. In this layer, a dynamic routing algorithm is used to transfer information between capsules. The dynamic routing algorithm works by iteratively calculating and updating the routing coefficients between capsules. First, the algorithm initializes the routing coefficients based on the similarity between the initial capsule and the target capsule. Then, the outputs of all initial capsules are weighted summed to obtain the input of the target capsule. Next, a nonlinear activation function (such as the Squash function) is applied to generate the output vector of the target capsule. Based on these output vectors, the dynamic routing algorithm adjusts and updates the routing coefficients to make the transfer of information between capsules more accurate and efficient. After multiple iterative calculations, the capsule vector with adjusted routing is finally generated.
[0087] Through this design, the dynamic routing layer uses capsule networks and dynamic routing algorithms to effectively process potential feature representations, and ultimately generates high-quality capsule vectors after routing adjustments. These vectors can more accurately reflect the complex features and relationships of the input data, thereby providing accurate and reliable input data for the subsequent advertising effect evaluation module. This design ensures that the advertising delivery adjustment system can efficiently analyze and optimize advertising delivery effects, improving the overall performance of the system and the accuracy of effect prediction.
[0088] The following is the implementation code of the dynamic routing layer, including the detailed dynamic routing algorithm.
[0089] import numpy as np
[0090] # Define the initial capsule layer of the capsule network
[0091] class InitialCapsuleLayer:
[0092] def __init__(self):
[0093] # Initialize the layer parameters
[0094] pass
[0095] def forward(self, input_data):
[0096] # Decompose the input latent feature representation into multiple initial capsules
[0097] capsules = self.create_capsules(input_data)
[0098] return capsules
[0099] def create_capsules(self, data):
[0100] # Create initial capsule vector
[0101] # Here we assume that each capsule is a vector
[0102] capsules = [self.create_capsule_vector(data, i) for i inrange(len(data))]
[0103] return capsules
[0104] def create_capsule_vector(self, data, index):
[0105] # Generate capsule vector based on index
[0106] return data[index] # This is just a hint
[0107] # Define the capsule transformation layer of the capsule network
[0108] class CapsuleTransformLayer:
[0109] def __init__(self, transform_matrix):
[0110] # Initialize the layer parameters, including the transformation matrix
[0111] self.transform_matrix = transform_matrix
[0112] def forward(self, capsules):
[0113] # Apply the transformation matrix to the initial capsule to generate the transformed capsule vector
[0114] transformed_capsules = [self.transform(capsule) for capsulein capsules]
[0115] return transformed_capsules
[0116] def transform(self, capsule):
[0117] # Apply transformation matrix to transform
[0118] return np.dot(self.transform_matrix, capsule)
[0119] # Define dynamic routing layer
[0120] class DynamicRoutingLayer:
[0121] def __init__(self, num_iterations):
[0122] # Initialize the parameters of the layer, including the number of iterations
[0123] self.num_iterations = num_iterations
[0124] def forward(self, transformed_capsules):
[0125] # Use dynamic routing algorithm to pass information between capsules
[0126] routing_weights = self.initialize_routing_weights(transformed_capsules)
[0127] for _ in range(self.num_iterations):
[0128] routing_weights = self.update_routing_weights(transformed_capsules, routing_weights)
[0129] routed_capsules = self.calculate_routed_capsules(transformed_capsules, routing_weights)
[0130] return routed_capsules
[0131] def initialize_routing_weights(self, capsules):
[0132] # Initialize routing weights
[0133] num_capsules = len(capsules)
[0134] routing_weights = np.zeros((num_capsules, num_capsules))
[0135] return routing_weights
[0136] def update_routing_weights(self, capsules, routing_weights):
[0137] # Update routing weights
[0138] num_capsules = len(capsules)
[0139] for i in range(num_capsules):
[0140] for j in range(num_capsules):
[0141] routing_weights[i, j] = self.calculate_similarity(capsules[i], capsules[j])
[0142] return routing_weights
[0143] def calculate_similarity(self, capsule_i, capsule_j):
[0144] # Calculate the similarity between capsules
[0145] return np.dot(capsule_i, capsule_j)
[0146] def calculate_routed_capsules(self, capsules, routing_weights):
[0147] # Calculate the capsule vector after route adjustment
[0148] num_capsules = len(capsules)
[0149] routed_capsules = [self.calculate_routed_capsule(capsules,routing_weights, i) for i in range(num_capsules)]
[0150] return routed_capsules
[0151] def calculate_routed_capsule(self, capsules, routing_weights,index):
[0152] # Calculate the routing results for each capsule
[0153] weighted_sum = np.zeros_like(capsules[0])
[0154] for j in range(len(capsules)):
[0155] weighted_sum += routing_weights[index, j] * capsules[j]
[0156] return self.squash(weighted_sum)
[0157] def squash(self, vector):
[0158] # Apply the squash function to ensure that the length of the output vector is between 0 and 1
[0159] norm = np.linalg.norm(vector)
[0160] norm_squared = norm ** 2
[0161] return (norm_squared / (1 + norm_squared)) * (vector / norm)
[0162] # Main function, demonstrating the use of the dynamic routing layer
[0163] def main():
[0164] # Assume that the input data is the potential feature representation generated by the generative adversarial network layer
[0165] input_data = np.random.rand(10, 16) # 10 samples, 16 features per sample
[0166] # Initialize each layer
[0167] initial_capsule_layer = InitialCapsuleLayer()
[0168] transform_matrix = np.random.rand(16, 16) # Assume a 16-dimensional transformation matrix
[0169] capsule_transform_layer = CapsuleTransformLayer(transform_matrix)
[0170] dynamic_routing_layer = DynamicRoutingLayer(num_iterations=3)
[0171] # Forward propagation process
[0172] initial_capsules = initial_capsule_layer.forward(input_data)
[0173] transformed_capsules = capsule_transform_layer.forward(initial_capsules)
[0174] routed_capsules = dynamic_routing_layer.forward(transformed_capsules)
[0175] # Output the capsule vector after routing adjustment
[0176] print("Routed Capsules:", routed_capsules)
[0177] if __name__ == "__main__":
[0178] main()
[0179] Furthermore, the attention mechanism layer is implemented by the following steps:
[0180] Extracting the attributes of each feature based on the received route-adjusted feature vector generated by the dynamic routing layer;
[0181] Use a weighting mechanism to weight each feature and calculate the importance weight of each feature;
[0182] The calculated importance weights are applied to the corresponding feature vectors, and a weighted feature vector is generated by weighted averaging.
[0183] First, the attention mechanism layer receives the route-adjusted feature vectors from the dynamic routing layer. These feature vectors contain rich information extracted from the input data and processed by multiple layers. The first step of the attention mechanism layer is to extract the specific attributes of each feature vector. These attributes reflect the performance and meaning of each feature in different dimensions, such as key indicators such as user click-through rate and conversion rate.
[0184] Next, the system uses a weighting mechanism to weight each feature. The core of the weighting mechanism is to calculate the importance weight of each feature, which represents the relative importance of the feature in the overall feature vector. To calculate these weights, the system usually uses a dot product operation combined with a Softmax function to ensure that the sum of all feature weights is 1, thereby achieving normalization. This weighting method is similar to evaluating the priority of each candidate among multiple candidates, ensuring that important features receive more attention, while the influence of secondary features is relatively weakened.
[0185] The calculated importance weights are then applied to the corresponding feature vectors. In this way, the system is able to perform a weighted average of the feature vectors to generate a weighted feature vector. Specifically, each feature vector is scaled up or down according to its weight, with important features contributing more to the final weighted feature vector and unimportant features contributing less. This process is similar to giving higher weights to key factors when scoring a comprehensive score, thereby obtaining a more representative comprehensive score.
[0186] Finally, the attention mechanism layer outputs weighted feature vectors that integrate the information of all important features and provide more accurate input data for subsequent prediction tasks. This weighted processing method ensures that the system can fully utilize the information of key features when predicting the effect of advertising, thereby improving the accuracy and reliability of the prediction.
[0187] Through the above steps, the attention mechanism layer not only improves the expressive power of the feature vector, but also enhances the system’s sensitivity to different features, enabling the advertising delivery adjustment system to more accurately predict and optimize advertising effects.
[0188] The output layer is used to receive the weighted feature vector and use four independent neurons to output the predicted values of the advertisement click-through rate, conversion rate, user retention rate and emotional response score.
[0189] The following is the implementation code of the output layer.
[0190] import numpy as np
[0191] # Define the output layer
[0192] class OutputLayer:
[0193] def __init__(self):
[0194] # Initialize the parameters of the output layer
[0195] self.weights = np.random.rand(4, 16) # Assume that the weight vector length of each neuron is 16, four independent neurons
[0196] self.biases = np.random.rand(4) # Each neuron corresponds to a bias value
[0197] def forward(self, weighted_features):
[0198] # Receive weighted feature vector
[0199] # Calculate the output of each neuron
[0200] outputs = []
[0201] for i in range(4): # Four independent neurons
[0202] output = self.calculate_neuron_output(weighted_features,self.weights[i], self.biases[i])
[0203] outputs.append(output)
[0204] # Convert the output into predicted values for ad click-through rate, conversion rate, user retention rate, and sentiment response score
[0205] predictions = {
[0206] 'CTR': outputs[0],
[0207] 'CVR': outputs[1],
[0208] 'URR': outputs[2],
[0209] 'ERS': outputs[3]
[0210] }
[0211] return predictions
[0212] def calculate_neuron_output(self, features, weights, bias):
[0213] # Calculate the output of a single neuron
[0214] # Linear transformation: weighting and biasing
[0215] linear_output = np.dot(features, weights) + bias
[0216] # Activation function: Use the Sigmoid function to limit the output between 0 and 1
[0217] activated_output = self.sigmoid(linear_output)
[0218] return activated_output
[0219] def sigmoid(self, x):
[0220] # Implement Sigmoid activation function
[0221] return 1 / (1 + np.exp(-x))
[0222] The dynamic adjustment module 104 is used to adjust the advertisement delivery using a reinforcement learning algorithm based on the evaluation result of the context-aware prediction evaluation, wherein the adjustment includes replacing the advertisement material, adjusting the delivery platform and optimizing the delivery time period, personalized content recommendation and user interaction mode adjustment.
[0223] The dynamic adjustment module 104 plays a key role in the advertising delivery adjustment system, and dynamically adjusts the advertising delivery using a reinforcement learning algorithm based on the context-aware prediction evaluation results provided by the advertising effect evaluation module 103. The main goal of this module is to optimize the advertising delivery strategy to improve the accuracy of advertising and user satisfaction.
[0224] First, the dynamic adjustment module 104 receives the evaluation results transmitted by the advertising effect evaluation module 103. These evaluation results include key indicators such as the advertising click-through rate, conversion rate, user retention rate, and emotional response score. Based on these data, the dynamic adjustment module 104 can fully understand the effect of the current advertising delivery and user response.
[0225] Next, the dynamic adjustment module 104 optimizes the advertising delivery strategy using a reinforcement learning algorithm. Reinforcement learning is a machine learning method that obtains feedback from the environment by continuously testing and adjusting strategies, and optimizes the decision-making process based on the feedback information. In this system, the reinforcement learning algorithm adjusts the advertising delivery parameters based on the evaluation results, such as changing the advertising material, adjusting the delivery platform, optimizing the delivery time period, etc. Specifically, changing the advertising material means selecting more attractive and relevant advertising content based on user interests and feedback; adjusting the delivery platform means delivering advertisements on appropriate platforms based on the user characteristics and advertising effects of different platforms; optimizing the delivery time period means selecting the best time to deliver advertisements based on the user's active time to maximize the exposure and click-through rate of the advertisement.
[0226] In addition, the dynamic adjustment module 104 can also perform personalized content recommendation and user interaction mode adjustment. Personalized content recommendation is to provide users with customized advertising content by analyzing user portraits and behavior data, so that the advertisements are more in line with the interests and needs of users. User interaction mode adjustment is to optimize the interaction between users and advertisements, such as guiding users to click, share or comment more, thereby improving the engagement and conversion rate of advertisements.
[0227] In order to ensure the timeliness and accuracy of the adjustment, the dynamic adjustment module 104 uses real-time data processing technology. The system will continuously monitor and analyze the latest behavior data of users, and adjust the advertising delivery strategy in real time based on these data. In this way, the system can quickly respond to changes in user behavior and optimize the advertising delivery effect in real time.
[0228] Finally, all adjustment operations and results of the dynamic adjustment module 104 are recorded in the system log. These records not only help further optimize and improve the system, but also provide advertisers with detailed delivery reports and effect analysis to help them formulate more effective advertising strategies.
[0229] In the ad delivery adjustment system, the training of the neural network model is a multi-stage, iterative process involving a series of steps from data collection to model deployment. This process ensures that the model can effectively learn and adapt to the dynamic changes in user behavior and ad effects to achieve an optimized ad delivery strategy. The following are the specific steps of model training:
[0230] First, the system needs to collect user portrait and behavior data, including ad click-through rate (CTR), conversion rate (CVR), user retention rate (URR), and emotional response score (ERS). These data are pre-processed by standardization and normalization to ensure the consistency and quality of the input data. This step is completed by the input layer, which accepts user portrait and behavior data and pre-processes them to obtain data suitable for model training.
[0231] Next, the preprocessed data is passed to the conditional generative adversarial network layer, which includes a generator and a discriminator. The generator generates a latent feature representation based on the preprocessed data, while the discriminator distinguishes true from false on the output of the generator and feeds back the error to the generator for improvement. During the training process, the generator and the discriminator are continuously optimized through adversarial training to generate more realistic and useful latent feature representations.
[0232] The latent feature representation generated by the generator is then passed to the dynamic routing layer. The dynamic routing layer uses a capsule network for processing. The purpose of this layer is to dynamically adjust the routing coefficients so that the feature representation can more accurately reflect the nature of the input data. After being processed by the dynamic routing layer, a feature vector with adjusted routing is generated.
[0233] These route-adjusted feature vectors are then passed to the attention layer. The attention layer generates weighted feature vectors by calculating the weighted average of the feature vectors and dynamically assigning importance weights to different features. Specifically, the attention layer calculates the importance weight of each feature based on its attributes and applies these weights to the feature vector to generate a comprehensive weighted feature representation.
[0234] Finally, the weighted feature vector is passed to the output layer. The output layer contains four independent neurons, each corresponding to a prediction value: ad click-through rate, conversion rate, user retention rate, and emotional response score. The output layer processes the weighted feature vector to generate the final prediction result.
[0235] During the training process, the entire model needs to undergo rigorous evaluation and testing, using cross-validation and various performance indicators (such as accuracy, response time, etc.) to verify the effectiveness and reliability of the model. After the training is completed, the evaluated and tested model will be deployed to the advertising delivery adjustment system to process user behavior and advertising effect data in real time, dynamically adjust the advertising display strategy based on the results predicted by the model, optimize the advertising delivery effect and meet the dynamic needs of users.
[0236] After deployment, the system will continue to monitor the performance and actual application effects of the model, collect new data for iterative updates and optimization of the model, ensure that the model can adapt to environmental changes and new needs, and continue to provide efficient and accurate advertising solutions. Through this rigorous and detailed training and implementation process, the advertising adjustment system can fully utilize the powerful learning and prediction capabilities of neural networks, continuously optimize and improve the effects of advertising, and achieve accurate advertising and improved user experience.
[0237] Furthermore, the reinforcement learning algorithm used in the dynamic adjustment module is implemented using a deep Q network algorithm; the state space includes ad click-through rate, conversion rate, user retention rate, and emotional response score; the action space includes changing ad materials, adjusting delivery platforms and optimizing delivery time periods, personalized content recommendations, and user interaction mode adjustments; and the reward function is implemented using the following formula 1:
[0238]
[0239] in, Indicates status; Indicates action; , , and is the adjustment factor; Indicates in status and actions The click-through rate of the advertisement is calculated as described in Formula 2 below:
[0240]
[0241] in, Indicates in status and actions The number of ad clicks under Indicates in status and actions The number of ad impressions under Indicates status and actions The conversion rate is calculated as shown in the following formula 3:
[0242]
[0243] in, In Status and actions The actual number of purchases made under In Status and actions The number of ad clicks under represents the user retention rate, which is calculated as shown in Formula 4 below:
[0244]
[0245] in, In Status and actions The number of users who continue to interact with the platform after a specified period of time; Indicates status The total number of users under represents the emotional response score, which is calculated as shown in Formula 5 below:
[0246]
[0247] in, Indicates Users in the state and actions The following are the emotional responses to the advertisements: Indicates the total number of users.
[0248] The reinforcement learning algorithm used in the dynamic adjustment module is implemented using the Deep Q Network (DQN) algorithm, which optimizes the advertising delivery strategy through the definition of state space and action space, as well as a specific reward function.
[0249] In this system, the state space includes ad click-through rate (CTR), conversion rate (CVR), user retention rate (URR) and emotional response score (ERS). These indicators are key parameters for evaluating advertising effectiveness and can fully reflect the performance of advertising and user reactions. The action space includes changing advertising materials, adjusting the delivery platform and optimizing the delivery time period, personalized content recommendation and user interaction mode adjustment. These operations can directly affect the display method of advertising and the user's interactive experience.
[0250] To evaluate the effect of each state-action pair, the system uses a reward function as follows:
[0251]
[0252] in, Indicates the status, Indicates action, , , and is the adjustment coefficient used to balance the weights of different indicators and can be directly obtained through expert knowledge. The specific calculation method is as follows:
[0253] 1. Ad Click-Through Rate (CTR):
[0254]
[0255] Among them, among them, Indicates in status and actions The number of ad clicks under Indicates in status and actions The number of ad impressions for .
[0256] In the advertising delivery system, the number of ad clicks and ad impressions These are two key indicators used to evaluate advertising effectiveness. The following is a detailed explanation of how to obtain these two indicators.
[0257] Ad impressions represent the total number of times an ad is displayed in a specific state s and taking a specific action a. To obtain ad impressions, you can follow these steps:
[0258] 1. Status: The specific status of the ad when it is displayed, including the ad click-through rate (CTR), conversion rate (CVR), user retention rate (URR) and emotional response score (ERS).
[0259] 2. Action a: Actions taken under specific conditions, such as changing advertising creatives, adjusting delivery platforms, or optimizing delivery time periods.
[0260] 3. Record the number of impressions: Every time an ad is displayed, the system will record an impression event. You can use the statistical tools or logging system provided by the ad management platform to track each ad impression.
[0261] Ad clicks Indicates the total number of times users click on ads in a specific state s and taking a specific action a. To get the number of ad clicks, you can follow these steps:
[0262] 1. State s and action a: Same as s and a in the ad impressions above.
[0263] 2. Record clicks: Every time a user clicks on an ad, the system records a click event. You can use the click monitoring tools or logging systems provided by the ad management platform to track each ad click.
[0264] 3. Conversion Rate (CVR):
[0265]
[0266] in, In Status and actions The actual number of purchases made under In Status and actions The number of ad clicks under .
[0267] In the advertising delivery system, the actual number of purchases The system tracks these post-click purchase conversions and records the number of times users complete purchases, thereby obtaining These data will be automatically collected and counted through the advertising management platform or system logs.
[0268] 4. User Retention Rate (URR):
[0269]
[0270] in, In Status and actions The number of users who continue to interact with the platform after a specified period of time. Indicates status The total number of users under .
[0271] In the advertising delivery system, the calculation of user retention rate depends on two key data. First, It represents the number of users who continue to interact with the platform within a specified time after being in a specific state s and taking a specific action a. This data is obtained by tracking the user's behavior after interacting with the ad and counting the number of users who are still active within a given time. Represents the total number of users in state s, which refers to the total number of users who interacted with the ad in that state. These data can be automatically collected and counted through ad management platforms and user behavior analysis tools.
[0272] 5. Emotional Response Rating (ERS):
[0273]
[0274] in, Indicates Users in the state and actions The following are the emotional responses to the advertisements: Indicates the total number of users.
[0275] The following example explains how to obtain the state space composed of the click-through rate (CTR), conversion rate (CVR), user retention rate (URR), and emotional response score (ERS). .
[0276] The example scenario is as follows:
[0277] Assume there is an online advertising platform, the current status It consists of the following indicators:
[0278] Ad click-through rate (CTR): 0.05 (i.e. 5% of displayed ads are clicked)
[0279] Conversion rate (CVR): 0.02 (i.e. 2% of users who clicked on the ad completed the purchase)
[0280] User Retention Rate (URR): 0.1 (i.e. 10% of users continue to interact with the platform after a specified period of time)
[0281] Emotional Response Score (ERS): Initially 0, the action taken by the system needs to be obtained through sentiment analysis It is to replace the creative.
[0282] Get emotional response score :
[0283] 1. Advertising display and user feedback:
[0284] In state s, the ad creative is replaced and displayed to the user.
[0285] After seeing the ad, user i may post a comment on social media or express his emotional response to the ad through the feedback mechanism provided by the platform.
[0286] 2. Collect user feedback:
[0287] Suppose user i makes a comment: "This new ad is great, I like it."
[0288] 3. Sentiment Analysis:
[0289] The comments are analyzed using a pre-trained sentiment analysis model to obtain sentiment scores.
[0290] In the above-mentioned embodiment, an advertisement delivery adjustment system is provided. Correspondingly, the present application also provides an advertisement delivery adjustment method. Figure 2 , which is a flow chart of an embodiment of an advertisement delivery adjustment method of the present application. Since this embodiment, namely the second embodiment, is basically similar to the first embodiment, the description is relatively simple, and the relevant parts can refer to the partial description of the first embodiment. The method embodiment described below is only illustrative.
[0291] A second embodiment of the present application provides an advertisement delivery adjustment method, including:
[0292] Step S201: real-time monitoring and collection of user behavior data on multiple media platforms, wherein the behavior data includes the user's click rate, dwell time, interaction frequency, browsing path, and purchase history;
[0293] Step S202: integrating the collected behavior data with a third-party data source to generate a user profile; wherein the third-party data source includes social media data and geographic location data;
[0294] Step S203: Performing a context-aware prediction evaluation on the advertising effect based on the user portrait and behavior data, wherein the evaluation results of the context-aware prediction evaluation include the advertising click-through rate, conversion rate, user retention rate and emotional response score;
[0295] Step S204: Based on the evaluation result of the context-aware prediction evaluation, the advertisement delivery is adjusted using a reinforcement learning algorithm, wherein the adjustment includes replacing the advertising material, adjusting the delivery platform and optimizing the delivery time period, personalized content recommendation and user interaction mode adjustment.
[0296] Although the present application is disclosed as above in the form of a preferred embodiment, it is not intended to limit the present application. Any technical personnel in this field may make possible changes and modifications without departing from the spirit and scope of the present application. Therefore, the scope of protection of the present application shall be based on the scope defined by the claims of the present application.
Claims
1. An advertisement delivery adjustment system, characterized in that: include: A user behavior monitoring module, which is used to monitor and collect user behavior data in real time on multiple media platforms, wherein the behavior data includes the user's click rate, dwell time, interaction frequency, browsing path and purchase history; A data fusion module, used to fuse the collected behavioral data with third-party data sources to generate user portraits; wherein the third-party data sources include social media data and geographic location data; An advertising effect evaluation module, which is used to perform a context-aware prediction evaluation on the advertising effect based on the user portrait and behavior data, wherein the evaluation results of the context-aware prediction evaluation include the advertising click-through rate, conversion rate, user retention rate and emotional response score; A dynamic adjustment module, configured to adjust the advertisement delivery using a reinforcement learning algorithm based on the evaluation result of the context-aware prediction evaluation, wherein the adjustment includes replacing the advertisement material, adjusting the delivery platform and optimizing the delivery time period, personalized content recommendation and user interaction mode adjustment; The reinforcement learning algorithm used in the dynamic adjustment module is implemented by a deep Q network algorithm; the state space includes ad click-through rate, conversion rate, user retention rate and emotional response score; the action space includes changing ad materials, adjusting delivery platforms and optimizing delivery time periods, personalized content recommendations and user interaction mode adjustments; the reward function is implemented by the following formula 1: in, Indicates status; Indicates action; , , and is the adjustment factor; Indicates in status and actions The click-through rate of the advertisement is calculated as described in Formula 2 below: in, Indicates in status and actions The number of ad clicks under Indicates in status and actions The number of ad impressions under Indicates status and actions The conversion rate is calculated as shown in the following formula 3: in, In Status and actions The actual number of purchases made under In Status and actions The number of ad clicks under represents the user retention rate, which is calculated as shown in Formula 4 below: in, In Status and actions The number of users who continue to interact with the platform after a specified period of time; Indicates status The total number of users under represents the emotional response score, which is calculated as shown in Formula 5 below: in, Indicates Users in the state and actions The following are the ratings of emotional responses to the ads: Indicates the total number of users.
2. The advertisement delivery adjustment system according to claim 1, characterized in that: The user behavior monitoring module includes a data storage unit for storing user behavior data monitored and collected in real time.
3. The advertisement delivery adjustment system according to claim 1, characterized in that: The data fusion module includes a data cleaning unit for cleaning the collected behavior data and data from a third-party data source to remove noise and redundant data.
4. The advertisement delivery adjustment system according to claim 1, characterized in that: The advertising effect evaluation module uses a trained hybrid neural network model to perform context-aware prediction and evaluation on the advertising effect based on the user portrait and behavior data; the hybrid neural network model includes an input layer, a conditional generative adversarial network layer, a dynamic routing layer, an attention mechanism layer and an output layer; wherein the input layer accepts user portraits and behavior data as input, and performs standardization and normalization preprocessing on the input data to obtain preprocessed data; the generative adversarial network layer includes a generator and a discriminator, and the generator generates a potential feature representation based on the preprocessed data; the discriminator distinguishes the true and false output of the generator, and feeds back the error to the generator for improvement; the dynamic routing layer is implemented using a capsule network, processes the potential feature representation, and obtains a feature vector after routing adjustment; the attention mechanism layer dynamically assigns importance weights of different features by calculating the feature weighted average of the feature vector after routing adjustment, and generates a weighted feature vector; the output layer is used to receive the weighted feature vector, and uses four independent neurons to output the predicted values of the advertising click-through rate, conversion rate, user retention rate and emotional response score.
5. The advertisement delivery adjustment system according to claim 4, characterized in that: The generator adopts a deep convolution generator network, including a first input layer, a first convolution layer, a first deconvolution layer and a first output layer; the first input layer is used to receive the preprocessed data; The first convolution layer is implemented by a plurality of stacked convolution layers, and is used to process the output of the first input layer to obtain a high-order feature map; the first deconvolution layer includes a plurality of deconvolution layers, and is used to upsample the high-order feature map extracted by the convolution layer to the size of the original input data to obtain a deconvolution feature representation; The first output layer is used to generate a latent feature representation based on the deconvolution feature representation.
6. The advertisement delivery adjustment system according to claim 4, characterized in that: The discriminator is implemented using a multilayer perceptron structure, which includes a second input layer, a second hidden layer, and a second output layer; wherein the second input layer is used to accept potential feature representations and real data generated by the generator; the second hidden layer includes multiple neurons for processing the data of the second input layer; and the output layer is used to process the output of the second hidden layer and output a true or false discrimination result.
7. The advertisement delivery adjustment system according to claim 4, characterized in that: The dynamic routing layer is implemented using a capsule network, which includes an initial capsule layer, a capsule transformation layer and a dynamic routing layer; the initial capsule layer is used to receive the potential feature representation generated by the conditional generative adversarial network layer and decompose it into multiple initial capsules, each capsule containing a vector; the capsule transformation layer is used to apply a transformation matrix to the initial capsule to obtain a transformed capsule vector; the dynamic routing layer uses a dynamic routing algorithm to transfer information between capsules to generate a capsule vector after routing adjustment.
8. The advertisement delivery adjustment system according to claim 4, characterized in that: The attention mechanism layer is implemented by the following steps: Extracting the attributes of each feature based on the received route-adjusted feature vector generated by the dynamic routing layer; Use a weighting mechanism to weight each feature and calculate the importance weight of each feature; The calculated importance weights are applied to the corresponding feature vectors, and a weighted feature vector is generated by weighted averaging.
9. A method for adjusting advertisement delivery, characterized in that: include: Real-time monitoring and collection of user behavior data on multiple media platforms, wherein the behavior data includes user click-through rate, dwell time, interaction frequency, browsing path and purchase history; Fusion of the collected behavioral data with third-party data sources to generate user profiles; wherein the third-party data sources include social media data and geographic location data; Based on the user portrait and behavior data, a situation-aware prediction evaluation is performed on the advertising effect, wherein the evaluation results of the situation-aware prediction evaluation include the advertising click-through rate, conversion rate, user retention rate and emotional response score; Based on the evaluation result of the scenario-aware prediction evaluation, the advertisement delivery is adjusted using a reinforcement learning algorithm, wherein the adjustment includes replacing the advertisement material, adjusting the delivery platform and optimizing the delivery time period, personalized content recommendation and user interaction mode adjustment; wherein the reinforcement learning algorithm is implemented using a deep Q network algorithm; the state space includes advertisement click-through rate, conversion rate, user retention rate and emotional response score; the action space includes replacing the advertisement material, adjusting the delivery platform and optimizing the delivery time period, personalized content recommendation and user interaction mode adjustment; the reward function is implemented using the following formula 1: in, Indicates status; Indicates action; , , and is the adjustment factor; Indicates in status and actions The click-through rate of the advertisement is calculated as described in Formula 2 below: in, Indicates in status and actions The number of ad clicks under Indicates in status and actions The number of ad impressions under Indicates status and actions The conversion rate is calculated as shown in the following formula 3: in, In Status and actions The actual number of purchases made under In Status and actions The number of ad clicks under represents the user retention rate, which is calculated as shown in Formula 4 below: in, In Status and actions The number of users who continue to interact with the platform after a specified period of time; Indicates status The total number of users under represents the emotional response score, which is calculated as shown in Formula 5 below: in, Indicates Users in the state and actions The following are the ratings of emotional responses to the ads: Indicates the total number of users.
