A method for screening advertising resources
By crawling advertising inventory data in real time and dynamically updating user portraits, and adjusting the screening standards in combination with the DNN model and causal inference model, the problem of screening system lag caused by changes in the advertising environment is solved, and the flexibility and accuracy of advertising effects are achieved.
Patent Information
- Application Number
- CN202411794747.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-09
- Publication Date
- 2025-04-04
AI Technical Summary
Dynamic changes in the advertising environment and the lag of traditional screening systems lead to lower advertising effectiveness, algorithm bias, data loss and over-reliance on technology, etc., which affect the accuracy and fairness of screening results.
By establishing a data interface to capture key data of advertising resources in real time, dynamically update user portraits, calculate adaptive indicators of advertising resources based on the latest environmental changes, use the DNN model and causal inference model to dynamically adjust the screening criteria and priorities, conduct A/B testing and regularly analyze historical delivery data to tune the screening algorithm.
It realizes real-time and dynamic nature of advertising inventory screening, improves the flexibility and accuracy of advertising effects, reduces algorithm bias and data loss, and improves the automation level of the system.
Smart Images
Figure CN119762160A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of advertisement, and more specifically, to a method for screening advertisement resources. Background Art
[0002] The method of screening advertising resources has been evolving with the development of the advertising industry. At first, it relied on manual screening, and advertisers or advertising agencies selected suitable advertising channels based on experience. With the development of data analysis and Internet technology, the advertising screening method has gradually shifted to a data-driven model. The screening of advertising resources has transitioned from traditional manual screening to intelligent screening using algorithms, artificial intelligence and big data, which has promoted the improvement of the matching efficiency and accuracy of advertising resources. The screening method of advertising resources is mainly used to select advertising channels or media that meet specific conditions according to the goals and needs of advertisers to maximize advertising effects and input-output ratio. This method matches the advertiser's delivery goals with resource characteristics by evaluating indicators such as traffic, user characteristics, content quality, and coverage of potential advertising resources, and achieves accurate advertising delivery. The screening method is widely used in fields such as programmatic advertising, social media advertising, and search advertising, providing advertisers with effective resource selection tools to help them find the most suitable delivery channels in the complex advertising market.
[0003] However, there are also shortcomings in the screening methods of advertising resources. The main problem lies in the dynamic changes in the advertising environment and the lag of the screening system. Due to the constant changes in user needs and behaviors, and the advertising environment being affected by factors such as algorithms, data privacy and policy changes, traditional screening methods may not be able to adapt to new changes in a timely manner, resulting in reduced advertising effectiveness. Problems such as algorithm bias, data missing and over-reliance on technology may affect the accuracy and fairness of the screening results. Summary of the invention
[0004] The purpose of the present invention is to provide a method for screening advertising resources to solve the problems raised in the above background technology: the dynamic changes in the advertising environment and the lag of the screening system. Due to the constant changes in user needs and behaviors, and the advertising environment being affected by factors such as algorithms, data privacy and policy changes, traditional screening methods may not be able to adapt to new changes in a timely manner, resulting in reduced advertising effectiveness, algorithm bias, data missing and over-reliance on technology, which may affect the accuracy and fairness of the screening results.
[0005] Technical solution: The method for screening advertising resources comprises the following steps: S1. Capture the key data of advertising resources’ traffic, user characteristics, conversion rate, and click-through rate in real time by establishing a data interface, and dynamically update user portraits through user behavior data, including browsing history, clicks, and dwell time; collect the latest policies, platform rules, and data privacy to update external information, and adjust advertising screening strategies in a timely manner; S2. Extract the traffic, coverage population characteristics, and key content type characteristics of advertising resources based on the advertising needs of advertisers; calculate the adaptability indicators of advertising resources, including compliance and user interest matching, based on the latest environmental changes, and assign weights to adjust the screening conditions; calculate the user preference score of each advertising resource based on user behavior data and portraits to form a dynamically updated user interest model; S3. Preliminary screening of advertising resources based on preset criteria, including minimum click-through rate, user matching, and budget control, to eliminate resources that do not meet basic conditions; Use DNN models and causal inference models to dynamically adjust the screening criteria and priorities of advertising resources based on user behavior and environmental changes. If the advertising effect is not ideal, the screening strategy will be automatically adjusted to eliminate low-effect resources and introduce new alternative resources. S4. Conduct A / B testing, divide the screened advertising resources into different combinations for small-scale testing, and confirm the optimal advertising delivery combination; based on the test results and feedback data, update the weights and conditions in the screening model; S5. Regularly analyze historical delivery data, check the applicability and effectiveness of the algorithm, identify and improve deficiencies; regularly tune and iterate the screening algorithm based on the latest data and market feedback.
[0006] Preferably, the S1 collecting user behavior data includes the following steps: S1-1-1. Embed JavaScript code in the web page and use Google Analytics, Mixpanel, and Segment tools to capture users' browsing, click, and dwell time data; obtain page visits, click events, and dwell time information by analyzing the document and window objects provided by the browser; S1-1-2. Collect user-related behavior data in the back-end application, including through API requests and user login information, and integrate the data collected through API requests and the user login information with the front-end to form a complete user behavior log; S1-1-3. For mobile applications, collect user clicks, behavior data, and in-app dwell time by integrating SDKs, including Firebase Analytics and Adjust.
[0007] Extract features based on the user's browsing history, click data, and dwell time behavior data. The features include: Interest tags: Assign interest tags to users based on the content categories of the browsed pages (including technology, entertainment, fashion); Interaction strength: The weight and strength of each interest tag are calculated by the number of clicks and dwell time; Purchase intention prediction: Combine the user's click behavior on the advertisement to infer their purchase intention; Preferably, the step S1 of constructing and updating the user portrait includes the following steps: S1-2-1. Convert the collected user behavior data and features into structured portraits; use a rule-based approach to assign certain interest tags to each user, and adjust the weight of the tags based on the behavior for modeling; Weighting interest tags based on user behavior, including adding weight to a user's "Technology Interest" tag if the user browses technology content multiple times; Using a decision tree to train a user's interest prediction model, the interest prediction model automatically generates a user profile based on historical behavior data; S1-2-2. When a user visits a new page or clicks a new ad, only the relevant features are updated without rebuilding the entire user profile; set the time window for the behavior within the past 7 days, past 15 days, and past 30 days; S1-2-3. Based on the user's latest behavior data, the interest tags and behavior patterns in the portrait are weighted. For example, if the user has recently clicked on a certain type of advertisement frequently, the system will increase the weight of the advertisement category and adjust the user portrait.
[0008] Preferably, S2 obtains the user's preference score for different advertising resources by calculating the similarity between the user portrait vector and the advertising feature vector. The cosine similarity calculates the cosine value of the angle between the user portrait vector and the advertising feature vector, ranging from 0 to 1. The closer to 1, the higher the similarity. The formula is as follows: CosineSimilarity= Different user behaviors are assigned different weights to calculate preference scores; clicks indicate higher interest, which is set to 0.7; longer stays also indicate a certain interest, which is set to 0.5; the weight of browsing behavior alone is set to 0.3; the final user preference score is calculated using a weighted average, and the formula is as follows: User Preference Score=(Click×0.7+View×0.3+Time×0.5)×Similarity Preferably, based on each preference score calculation, a dynamically updated user interest model is constructed to continuously learn the user's latest preferences; collaborative filtering recommends advertising resources to similar users based on user behavior and preference scores; including user-based collaborative filtering and advertising-based collaborative filtering; Matrix decomposition decomposes the rating matrix of users and advertising resources into low-dimensional representations, and uses SVD to map the feature vectors of each user and advertising resource to the same vector space; The DNN model uses a neural network model to generate a user's interest model and predicts changes in interest based on the sequence of user behavior. The stream processing framework Kafkas outputs real-time calculation and update of user preference scores.
[0009] Preferably, the input layer of the DNN model takes the user portrait age, gender, advertising features and behavior sequence as input; the embedding layer embeds and encodes the discrete features advertising category and page category, represents them in vector space, and captures the association between features; the behavior sequence modeling layer uses the Attention mechanism to model the user behavior sequence, captures the changes in behavior over time, and generates an understanding of user interests; the fusion layer and the output layer fuse the outputs of the user portrait, the advertising features and the behavior sequence to generate a preference score.
[0010] Preferably, Kafka's architecture is set up as follows: Data stream topics: Create separate Kafka topics for different types of user behavior data, including "page_views", "clicks", and "stays"; Producers: The front-end and back-end systems act as Kafka producers, continuously sending user behavior data to Kafka topics; Consumers: The processing system of the deep learning model acts as a consumer, subscribes to Kafka topics, receives user behavior data in real time, and triggers model updates and recalculations; Use the Flink stream processing framework to connect to the Kafka topic and organize the user behavior data into time windows; construct the behavior data of the last 10 minutes into a window and input it into the DNN model for preference score calculation; whenever Kafka receives new user behavior data, the DNN model extracts the latest behavior sequence and adds the latest user behavior data to the behavior sequence according to the latest behavior sequence to form a time series data input model; The updated behavior sequence is input into the DNN model together with the advertising features and the user portrait features to calculate the user's preference score for each advertisement; the DNN model sends the latest preference score to the backend system.
[0011] Preferably, the S3DNN model automatically optimizes advertisement screening and priority through environmental feedback; the DNN model state definition includes user portrait, advertisement state and environmental characteristics including external hot spots; the priority of advertisement resources is adjusted, low-effect resources are eliminated and new advertisements are introduced; click-through rate and conversion rate are considered in the short term, and the impact of advertisements on user activity is focused on in the long term; the DNN model dynamically optimizes the screening criteria based on feedback, so that advertisement resources match changes in user interests in real time; The causal inference model reveals the relationship between user behavior, advertising features and external environment through a causal graph; intervenes in low-effective advertisements, uses causal inference methods to calculate the real impact of advertising effects, and identifies key factors; the model eliminates confirmed low-effective advertisements and selects new resources that are more in line with user interests.
[0012] Preferably, the combination of the DNN model and the causal inference model comprises the following steps: S3-1. The system monitors the click-through rate and conversion rate of advertisements in real time, and records the browsing and clicking behavior data of users; and uses the real-time feedback as the reward input for reinforcement learning; S3-2. After each round of user interaction, the DNN model adjusts the advertising priority based on the reward function; When it is found that the click-through rate and conversion rate of a certain advertising resource are significantly lower than the expected value, the DNN model gradually reduces its exposure frequency and triggers the causal inference model for further evaluation; S3-3. When the advertising effect is lower than the threshold, the causal inference model analyzes the reasons for the low effect; based on the causal graph, the contribution of each factor to the effect is calculated to confirm whether the poor effect is caused by the advertising content, display time, or changes in user interests; If the causal inference model confirms that a certain advertisement does affect the user experience, it will automatically remove the advertisement from the priority resource pool, and introduce new advertising resources through user behavior and environmental changes; S3-4. When new advertising resources are added, the causal inference first determines the initial priority of the resource based on causal inference and historical behavior; as the user and environment change, the DNN model will continue to adjust the priority of the advertising resources.
[0013] Compared with the prior art, the advantages of the present invention are: (1) The artificial learning model dynamically updates the screening strategy through environmental feedback. It can quickly adjust the priority and screening criteria of advertisements when user preferences change or advertising effects fluctuate, thus achieving more flexible recommendations.
[0014] (2) Causal inference models can identify the key factors that affect advertising effectiveness through causal analysis and help eliminate advertisements that actually cause negative impacts, thereby reducing the misuse of inefficient resources.
[0015] (3) With the help of reinforcement learning and causal inference, when the advertising effect is not ideal, it can automatically eliminate low-performance advertising resources and introduce resources that better meet user needs in real time, making the advertising resource pool more dynamically adaptable.
[0016] (4) Through the delayed reward mechanism of reinforcement learning, the system can optimize the advertising screening strategy in a direction that is beneficial to the long-term value of users, balancing short-term effects and long-term user satisfaction.
[0017] (5) It has achieved a high degree of automation, using reinforcement learning and causal inference to continuously self-optimize the screening criteria, allowing the system to operate efficiently with less human intervention. BRIEF DESCRIPTION OF THE DRAWINGS
[0018] Figure 1 It is a schematic diagram of the overall system of a method for screening advertising resources of the present invention; DETAILED DESCRIPTION
[0019] For examples, see Figure 1 A method for screening advertising resources includes the following steps: S1. Capture the key data of advertising resources’ traffic, user characteristics, conversion rate, and click-through rate in real time by establishing a data interface, and dynamically update user portraits through user behavior data, including browsing history, clicks, and dwell time; collect the latest policies, platform rules, and data privacy to update external information, and adjust advertising screening strategies in a timely manner; Specifically, the implementation steps for establishing a data interface are as follows: Select interface type: Select RESTful, SOAP, GraphQL, gRPC and other interface types according to your needs.
[0020] Design the interface path and request method: define the interface URL path, request method (GET, POST, PUT, DELETE, etc.), and required path parameters and query parameters.
[0021] Design request and response formats: define the data format of the request body and response body (such as JSON, XML), and clarify the data fields and their meanings.
[0022] Design data model: Build the database model of the interface and determine the data table structure, fields, relationships, and storage methods.
[0023] Security design: Add identity authentication, permission control, data encryption and other mechanisms to ensure interface security.
[0024] Build a development environment: prepare the development framework, development tools, and database environment.
[0025] Implement interface logic: Implement the business logic of the interface according to the design, including data query, data writing, data conversion and data validation.
[0026] Write interface documentation: Gradually improve the interface documentation during the development process, record the interface URL, parameters, request samples, response samples and error code descriptions.
[0027] Data verification: Implement data verification on interface requests to prevent invalid or erroneous data from entering the system and ensure data accuracy.
[0028] Unit testing: Perform unit testing on each functional module of the interface to verify the correctness of each functional point.
[0029] Integration test: Test after integrating the interface with other modules to check whether the data flow is smooth and verify the compatibility of the interface with other systems.
[0030] Stress testing: Evaluate the performance of the interface under high concurrent requests and determine its response speed and stability.
[0031] Security testing: Check whether the interface has security vulnerabilities, such as SQL injection, XSS attacks, etc., and verify the reliability of authentication and authorization mechanisms.
[0032] Exception handling test: Test the interface's handling of exceptions (such as invalid parameters, request timeouts, etc.) to ensure that error responses are standardized and recorded in logs.
[0033] Configure the production environment: Prepare the servers, databases, and network resources for the production environment, and ensure that the configuration meets security standards.
[0034] API release: deploy the API code to the production environment and set up load balancing, monitoring, and logging.
[0035] API document publishing: Publish the interface document to the API management platform or shared space for use by relevant personnel.
[0036] S2. Extract the traffic, coverage population characteristics, and key content type characteristics of advertising resources based on the advertising needs of advertisers; calculate the adaptability indicators of advertising resources, including compliance and user interest matching, based on the latest environmental changes, and assign weights to adjust the screening conditions; calculate the user preference score of each advertising resource based on user behavior data and portraits to form a dynamically updated user interest model; S3. Preliminary screening of advertising resources based on preset criteria, including minimum click-through rate, user matching, and budget control, to eliminate resources that do not meet basic conditions; Use DNN models and causal inference models to dynamically adjust the screening criteria and priorities of advertising resources based on user behavior and environmental changes. If the advertising effect is not ideal, the screening strategy will be automatically adjusted to eliminate low-effect resources and introduce new alternative resources. S4. Conduct A / B testing, divide the screened advertising resources into different combinations for small-scale testing, and confirm the optimal advertising delivery combination; based on the test results and feedback data, update the weights and conditions in the screening model; Specifically, conducting an A / B test includes the following steps: Grouping design: Grouping based on the attributes of different advertising resources (such as advertising content, advertising format, target keywords, etc.). Divide advertising resources into different combinations (such as A, B, C groups), each containing different advertising content or combinations.
[0037] Allocation ratio: The ratio of test users to each group. Usually, A / B testing uses equal allocation (e.g. 50% of users are allocated to group A and 50% to group B). However, in multi-combination testing, the allocation can be even or based on a specific strategy.
[0038] Generate ad combinations: Allocate ad resources to different test combinations to form multiple combinations. For example, Group A contains ad resources 1 and 2, Group B contains ad resources 3 and 4, and so on.
[0039] Data tracking and logging: Configure a data tracking system to record user interactions with each ad combination, including impressions, clicks, conversions, etc. You can use data tracking tools (such as Google Analytics, Mixpanel) or a custom logging system.
[0040] Control variables: Keep other variables (such as display location and display frequency) consistent across groups to ensure the fairness of the test and eliminate the impact of interfering factors on the test results.
[0041] Time window selection: Make sure the test time is long enough to cover user behaviors at different times to obtain reliable statistical data. Usually a time window of one week or one month is selected to ensure that the changes in user behavior on weekends and weekdays are included.
[0042] Data cleaning: Before analyzing the data, the log data is cleaned and abnormal data (such as accidental clicks, robot traffic, etc.) is removed to improve the reliability of the data.
[0043] Statistical analysis: Conduct statistical analysis on the key indicators of each test group and calculate the mean and standard error of indicators such as click-through rate and conversion rate.
[0044] Significance test: Use statistical tests (such as t-tests or chi-square tests) to verify whether the differences between different advertising combinations are statistically significant. For example, calculate the p-value to determine whether combination A is significantly better than combination B.
[0045] Determine the best combination: Based on the test results, select the ad combination that achieves the target indicator or has the highest conversion rate as the preferred combination. If the test does not show a clear winner, you can adjust the combination content or extend the test time.
[0046] Effect evaluation and recording: Record the effect of each advertising combination and compare it with the benchmark value (the current optimal solution) to find out the actual effect differences of different combinations.
[0047] Feedback iteration: If the A / B test fails to find a significant optimal combination, you can optimize the combination based on the analysis feedback, such as adjusting the advertising content and resource matching, generate a new combination and conduct the next round of testing S5. Regularly analyze historical delivery data, check the applicability and effectiveness of the algorithm, identify and improve deficiencies; regularly tune and iterate the screening algorithm based on the latest data and market feedback.
[0048] S1 collects user behavior data including the following steps: S1-1-1. Embed JavaScript code in the web page and use Google Analytics, Mixpanel, and Segment tools to capture users' browsing, click, and dwell time data; obtain page visits, click events, and dwell time information by analyzing the document and window objects provided by the browser; S1-1-2. Collect user-related behavior data in the back-end application, including through API requests and user login information, and integrate the data collected through API requests and user login information with the front-end to form a complete user behavior log; S1-1-3. For mobile applications, collect user clicks, behavior data, and in-app dwell time by integrating SDKs, including Firebase Analytics and Adjust.
[0049] S1 builds and updates user profiles including the following steps: S1-2-1. Convert the collected user behavior data and features into structured portraits; use a rule-based approach to assign certain interest tags to each user, and adjust the weight of the tags based on the behavior for modeling; Weighting interest tags based on user behavior, including adding weight to a user's "Technology Interest" tag if the user browses technology content multiple times; Use decision trees to train a user's interest prediction model, which automatically generates user portraits based on historical behavior data; S1-2-2. When a user visits a new page or clicks a new ad, only the relevant features are updated without rebuilding the entire user profile; set the time window for the behavior within the past 7 days, past 15 days, and past 30 days; S1-2-3. Based on the user's latest behavior data, the interest tags and behavior patterns in the portrait are weighted. For example, if the user has recently clicked on a certain type of advertisement frequently, the system will increase the weight of the advertisement category and adjust the user portrait.
[0050] S2 obtains the user's preference score for different advertising resources by calculating the similarity between the user portrait vector and the advertising feature vector. The cosine similarity calculates the cosine value of the angle between the user portrait vector and the advertising feature vector, ranging from 0 to 1. The closer to 1, the higher the similarity. The formula is as follows: CosineSimilarity= Different user behaviors are assigned different weights to calculate preference scores; clicks indicate higher interest, which is set to 0.7; longer stays also indicate a certain interest, which is set to 0.5; the weight of browsing behavior alone is set to 0.3; the final user preference score is calculated using a weighted average, and the formula is as follows: User Preference Score=(Click×0.7+View×0.3+Time×0.5)×Similarity Based on each preference score calculation, a dynamically updated user interest model is constructed to continuously learn the user's latest preferences; collaborative filtering recommends advertising resources to similar users based on user behavior and preference scores; including user-based collaborative filtering and advertising-based collaborative filtering; Specifically, collaborative filtering based on user behavior and preference scoring includes the following steps: User-based collaborative filtering: recommend advertising resources through similar users. Calculate the similarity between users. Common similarity indicators include Pearson correlation coefficient, cosine similarity, etc.
[0051] Similarity matrix generation: Use user behavior data to calculate the similarity matrix and identify user groups with a high degree of similarity to the current user.
[0052] Recommended advertising resources: Recommend advertising resources that are preferred by similar users to the current user, while filtering out ads that the current user has viewed or clicked to ensure the novelty of the recommendations.
[0053] Content-based collaborative filtering: If the ad resources contain information such as tags, descriptions, or categories, these content features can be used to recommend similar types of ad resources to users. Content-based recommendations are usually combined with user interest tags to improve the relevance of recommendations.
[0054] Matrix decomposition model: When the amount of data is large, matrix decomposition techniques such as SVD and ALS can be used to compress the user interest model and further extract the implicit features of users and advertising resources. Matrix decomposition helps capture the relationship between users' implicit interests and advertising resources and improves the recommendation effect.
[0055] The input layer of the DNN model takes the user portrait age, gender, advertising features and behavior sequence as input; the embedding layer embeds and encodes the discrete features advertising category and page category, represents them in vector space, and captures the association between features; the behavior sequence modeling layer uses the Attention mechanism to model the user behavior sequence, captures the changes in behavior over time, and generates an understanding of user interests; the fusion layer and the output layer fuse the outputs of the user portrait, advertising features and behavior sequence to generate a preference score.
[0056] Specifically, the model training process is as follows: Loss function: Binary cross entropy loss function and mean square error (MSE) are selected as loss functions to optimize the accuracy of preference scoring; Optimizer: Use Adam and SGD optimization algorithms for gradient descent to speed up the convergence of the model; Training process: historical behavior data is divided into training set, validation set and test set; the model goes through multiple rounds of iterations to optimize the weights of each layer and improve accuracy; Use the Flink stream processing framework to connect to Kafka topics and organize user behavior data into time windows. Build the behavior data of the last 10 minutes into windows and input them into the DNN model for preference score calculation. Whenever Kafka receives new user behavior data, the DNN model extracts the latest behavior sequence and adds the latest user behavior data to the behavior sequence based on the latest behavior sequence to form a time series data input model. The updated behavior sequence is input into the DNN model together with the advertising features and user portrait features to calculate the user's preference score for each advertisement; the DNN model sends the latest preference score to the backend system.
[0057] The S3DNN model automatically optimizes ad screening and priority through environmental feedback; the DNN model state definition includes user portraits, ad status, and environmental features including external hot spots; it adjusts the priority of ad resources, removes low-performance resources, and introduces new ads; it considers click-through rate and conversion rate in the short term, and focuses on the impact of ads on user activity in the long term; the DNN model dynamically optimizes screening criteria based on feedback, so that ad resources match changes in user interests in real time; The causal inference model reveals the relationship between user behavior, advertising characteristics and the external environment through causal graphs; intervenes in low-effective advertising, uses causal inference methods to calculate the real impact of advertising effects, and identifies key factors; the model eliminates confirmed low-effective advertisements and selects new resources that are more in line with user interests.
[0058] The combination of DNN model and causal inference model includes the following steps: S3-1. The system monitors the click-through rate and conversion rate of advertisements in real time, and records the browsing and clicking behavior data of users; and uses the real-time feedback as the reward input for reinforcement learning; S3-2. After each round of user interaction, the DNN model adjusts the advertising priority based on the reward function; When it is found that the click-through rate and conversion rate of a certain advertising resource are significantly lower than expected, the DNN model gradually reduces its exposure frequency and triggers the causal inference model for further evaluation; S3-3. When the advertising effect is lower than the threshold, the causal inference model analyzes the reasons for the low effect; based on the causal graph, the contribution of each factor to the effect is calculated to confirm whether the poor effect is caused by the advertising content, display time, or changes in user interests; If the causal inference model confirms that an ad does affect the user experience, it will automatically remove the ad from the priority resource pool, and introduce new ad resources based on user behavior and environmental changes; S3-4. When new advertising resources are added, causal inference first determines the initial priority of the resource based on causal inference and historical behavior; as users and the environment change, the DNN model will continue to adjust the priority of advertising resources.
[0059] The above shows and describes the basic principles, main features and advantages of the present invention; those skilled in the art should understand that the present invention is not limited to the above embodiments, and the above embodiments and descriptions are only preferred examples of the present invention and are not used to limit the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, and these changes and improvements all fall within the scope of the present invention to be protected; the scope of protection claimed by the present invention is defined by the attached claims and their equivalents.
Claims
1. A method for screening advertising resources, characterized in that: The method for screening advertising resources comprises the following steps: S1. Capture the key data of advertising resources’ traffic, user characteristics, conversion rate, and click-through rate in real time by establishing a data interface, and dynamically update user portraits through user behavior data, including browsing history, clicks, and dwell time; collect the latest policies, platform rules, and data privacy to update external information, and adjust advertising screening strategies in a timely manner; S2. Extract the traffic, coverage population characteristics, and key content type characteristics of advertising resources based on the advertising needs of advertisers; calculate the adaptability indicators of advertising resources, including compliance and user interest matching, based on the latest environmental changes, and assign weights to adjust the screening conditions; calculate the user preference score of each advertising resource based on user behavior data and portraits to form a dynamically updated user interest model; S3. Preliminary screening of advertising resources based on preset criteria, including minimum click-through rate, user matching, and budget control, to eliminate resources that do not meet basic conditions; Use DNN models and causal inference models to dynamically adjust the screening criteria and priorities of advertising resources based on user behavior and environmental changes. If the advertising effect is not ideal, the screening strategy will be automatically adjusted to eliminate low-effect resources and introduce new alternative resources. S4. Conduct A / B testing, divide the screened advertising resources into different combinations for small-scale testing, and confirm the optimal advertising delivery combination; based on the test results and feedback data, update the weights and conditions in the screening model; S5. Regularly analyze historical delivery data, check the applicability and effectiveness of the algorithm, identify and improve deficiencies; regularly tune and iterate the screening algorithm based on the latest data and market feedback.
2. The method for screening advertising resources according to claim 1, characterized in that: The S1 collecting user behavior data includes the following steps: S1-1-1. Embed JavaScript code in the web page and use Google Analytics, Mixpanel, and Segment tools to capture users' browsing, click, and dwell time data; obtain page visits, click events, and dwell time information by analyzing the document and window objects provided by the browser; S1-1-2. Collect user-related behavior data in the back-end application, including through API requests and user login information, and integrate the data collected through API requests and the user login information with the front-end to form a complete user behavior log; S1-1-3. For mobile applications, collect user clicks, behavior data, and in-app dwell time by integrating SDKs, including Firebase Analytics and Adjust.
3. The method for screening advertising resources according to claim 1, characterized in that: The S1 construction and update of user portraits includes the following steps: S1-2-1. Convert the collected user behavior data and features into structured portraits; use a rule-based approach to assign certain interest tags to each user, and adjust the weight of the tags based on the behavior for modeling; Weighting interest tags based on user behavior, including adding weight to a user's "Technology Interest" tag if the user browses technology content multiple times; Using a decision tree to train a user's interest prediction model, the interest prediction model automatically generates a user profile based on historical behavior data; S1-2-2. When a user visits a new page or clicks a new ad, only the relevant features are updated without rebuilding the entire user profile; Set the time window for the behavior within the past 7 days, past 15 days, and past 30 days; S1-2-3. Based on the user's latest behavior data, adjust the weights of interest tags and behavior patterns in the portrait; This includes when a user has recently clicked on a certain type of advertisement frequently. The system will increase the weight of that advertisement category and adjust the user portrait.
4. The method for screening advertising resources according to claim 1, characterized in that: S2 obtains the user's preference score for different advertising resources by calculating the similarity between the user portrait vector and the advertising feature vector. The cosine similarity calculates the cosine value of the angle between the user portrait vector and the advertising feature vector, ranging from 0 to 1. The closer to 1, the higher the similarity. The formula is as follows: CosineSimilarity= ; Different user behaviors are assigned different weights to calculate preference scores; clicks indicate higher interest, which is set to 0.7; longer stays also indicate a certain interest, which is set to 0.5; the weight of browsing behavior alone is set to 0.3; the final user preference score is calculated using a weighted average, and the formula is as follows: User Preference Score=(Click×0.7+View×0.3+Time×0.5)×Similarity.
5. The method for screening advertising resources according to claim 4, characterized in that: Based on each preference score calculation, a dynamically updated user interest model is constructed to continuously learn the user's latest preferences; Collaborative filtering recommends advertising resources to similar users based on user behavior and preference scores; including user-based collaborative filtering and advertising-based collaborative filtering; Matrix decomposition decomposes the rating matrix of users and advertising resources into low-dimensional representations, and uses SVD to map the feature vectors of each user and advertising resource to the same vector space; The DNN model uses a neural network model to generate a user's interest model and predicts changes in interest based on the sequence of user behavior. The stream processing framework Kafkas outputs real-time calculation and update of user preference scores.
6. The method for screening advertising resources according to claim 5, characterized in that: The input layer of the DNN model takes the user portrait age, gender, advertising features and behavior sequence as input; the embedding layer embeds and encodes the discrete feature advertising category and page category, and expresses them in vector space to capture the association between features; The behavior sequence modeling layer uses the Attention mechanism to model user behavior sequences, capture changes in behavior over time, and generate an understanding of user interests; The fusion layer and the output layer fuse the user portrait, the advertisement features and the output of the behavior sequence to generate a preference score.
7. The method for screening advertising resources according to claim 6, characterized in that: Use the Flink stream processing framework to connect to the Kafka topic and organize the user behavior data into time windows; construct the behavior data of the last 10 minutes into a window and input it into the DNN model for preference score calculation; whenever Kafka receives new user behavior data, the DNN model extracts the latest behavior sequence and adds the latest user behavior data to the behavior sequence according to the latest behavior sequence to form a time series data input model; The updated behavior sequence is input into the DNN model together with the advertising features and the user portrait features to calculate the user's preference score for each advertisement; the DNN model sends the latest preference score to the backend system.
8. The method for screening advertising resources according to claim 1, characterized in that: The S3DNN model automatically optimizes ad screening and priority through environmental feedback; the DNN model state definition includes user portrait, ad state and environmental features including external hotspots; Adjust the priority of advertising resources, remove low-performance resources and introduce new ads; consider click-through rate and conversion rate in the short term, and focus on the impact of advertising on user activity in the long term; the DNN model dynamically optimizes the screening criteria based on feedback to match advertising resources with changes in user interests in real time; The causal inference model reveals the relationship between user behavior, advertising features and external environment through a causal graph; Intervene ineffective ads, use causal inference methods to calculate the true impact of ad effectiveness, and identify key factors; the model eliminates confirmed ineffective ads and selects new resources that are more in line with user interests.
9. The method for screening advertising resources according to claim 8, characterized in that: The combination of the DNN model and the causal inference model includes the following steps: S3-1. The system monitors the click-through rate and conversion rate of advertisements in real time, and records the browsing and clicking behavior data of users; Use real-time feedback as reward input for reinforcement learning; S3-2. After each round of user interaction, the DNN model adjusts the advertising priority based on the reward function; When it is found that the click-through rate and conversion rate of a certain advertising resource are significantly lower than the expected value, the DNN model gradually reduces its exposure frequency and triggers the causal inference model for further evaluation; S3-3. When the advertising effect is lower than the threshold, the causal inference model analyzes the reasons for the low effect; based on the causal graph, the contribution of each factor to the effect is calculated to confirm whether the poor effect is caused by the advertising content, display time, or changes in user interests; If the causal inference model confirms that a certain advertisement does affect the user experience, it will automatically remove the advertisement from the priority resource pool, and introduce new advertising resources through user behavior and environmental changes; S3-4. When new advertising resources are added, the causal inference first determines the initial priority of the resource based on causal inference and historical behavior; as the user and environment change, the DNN model will continue to adjust the priority of the advertising resources.