Cross-platform advertisement intelligent putting system based on user portraits

Abstract

The invention relates to the technical field of intelligent advertisement putting, in particular to an intelligent cross-platform advertisement putting system based on user portraits. The system comprises a portrait construction module, an RTB bidding module and a monitoring feedback module. According to the invention, through constructing the portrait module, generating the portrait label of the user, taking the feature data of the user as the input, establishing a multilayer perceptron model, outputting the label weight of the predicted user portrait, and adjusting the label weight of the corresponding user portrait in real time, the RTB bidding module establishes RTB real-time bidding interfaces with a plurality of advertisement transaction platforms; a bidding strategy of an RTB real-time bidding interface is dynamically adjusted, a monitoring feedback module monitors interaction information and feedback data of a user to an advertisement in real time, the bidding strategy of the RTB real-time bidding interface is further adjusted, the efficiency and competitiveness of advertisement putting are improved, and maximum benefits of advertisement putting are achieved.

Description

Technical Field

[0001] This invention relates to the field of intelligent advertising delivery technology, and more specifically, to a cross-platform intelligent advertising delivery system based on user profiles. Background Technology

[0002] With the continuous advancement of internet technology and the booming development of the e-commerce industry, social media platforms have experienced massive growth in information volume and user base, with users spending extended periods on these platforms. Therefore, more precise advertising targeting for products holds significant commercial value. Current advertising systems typically rely on data collection and analysis from a single platform, evaluating ad performance by monitoring user behavior data (such as click-through rates and dwell time). These systems usually utilize basic data analysis methods to generate user profiles and adjust ad targeting based on preset rules or simple algorithms. However, the control capabilities and accuracy of these existing systems are often limited by the data sources and analytical methods, failing to comprehensively reflect changes in user behavior and needs across platforms.

[0003] To enable precise ad delivery and improve click-through rates and conversion rates by collecting data from multiple platforms, generating user profiles, and adjusting tag weights based on users' latest behavior and feedback, we propose a cross-platform intelligent ad delivery system based on user profiles. This system combines a real-time bidding interface to formulate bidding strategies, participate in bidding in real time, and flexibly adjust bidding strategies based on user interaction information and feedback data, thereby improving the efficiency and competitiveness of ad delivery. Summary of the Invention

[0004] The purpose of this invention is to address the problem that advertising delivery systems typically rely on data collection and analysis from a single platform. Evaluating ad performance by monitoring user behavior data (such as click-through rate and dwell time) on that platform often limits the control capabilities and accuracy due to limitations in data sources and analysis methods. This makes it impossible to comprehensively reflect changes in user behavior and needs across platforms. To achieve precise ad delivery and improve click-through and conversion rates, this invention combines a real-time bidding (RTB) interface to formulate bidding strategies, participate in bidding in real-time, and flexibly adjust bidding strategies based on user interaction information and feedback data, thereby improving the efficiency and competitiveness of ad delivery.

[0005] To achieve the above objectives, the present invention provides a cross-platform intelligent advertising delivery system based on user profiles, including a profile building module, an RTB bidding module, and a monitoring and feedback module;

[0006] The user profile building module uses web crawler technology to collect basic user information, behavioral characteristics, and consumption characteristics from multiple platforms. It uses the K-means clustering algorithm to generate user profile labels. Using the user's feature data as input, it establishes a multilayer perceptron model, outputs the label weights of the predicted user profile, and adjusts the label weights of the corresponding user profiles in real time.

[0007] The RTB bidding module establishes RTB real-time bidding interfaces with multiple advertising trading platforms, quantifies the value of users based on user profile tags, calculates the user's comprehensive value score by combining the weight of the profile tags, and dynamically adjusts the bidding strategy of the RTB real-time bidding interface based on the user's comprehensive value score, with the advertiser's budget and goals as constraints.

[0008] The monitoring and feedback module monitors user interaction information and feedback data for advertisements in real time, establishes an evaluation system based on ad impressions, click-through rate, and conversion rate, and uses the indicator threshold method to further adjust the bidding strategy of the RTB real-time bidding interface.

[0009] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0010] 1. This cross-platform intelligent advertising delivery system based on user profiles collects basic user information, behavioral characteristics, and consumption characteristics from multiple platforms by constructing a profile module. It uses the K-means clustering algorithm to generate user profile tags. Using user feature data as input, a multilayer perceptron model is established to output the predicted tag weights of user profiles. The tag weights of corresponding user profiles are adjusted in real time, realizing the collection of user data from multiple platforms. Based on the real-time feature data of users, the tag weights of user profiles are adjusted to ensure the real-time and accuracy of user profiles, providing a basis for subsequent participation in RTB bidding.

[0011] 2. The RTB bidding module establishes RTB real-time bidding interfaces with multiple advertising exchanges. It quantifies the value of users based on user profile tags, calculates the user's comprehensive value score by combining the weight of the profile tags, and dynamically adjusts the bidding strategy of the RTB real-time bidding interface based on the user's comprehensive value score, with the advertiser's budget and goals as constraints. It participates in bidding in real time and adjusts the bidding strategy according to user tags to improve the efficiency and competitiveness of advertising.

[0012] 3. The monitoring and feedback module monitors user interaction information and feedback data for advertisements in real time. It establishes an evaluation system based on ad impressions, click-through rate, and conversion rate. Using the indicator threshold method, it further adjusts the bidding strategy of the RTB real-time bidding interface to continuously optimize the advertising effect. Based on user interaction information and feedback data, it adjusts the strategy in a timely manner to maximize the benefits of advertising.

[0013] As a further improvement to this technical solution, the profile building module includes a user profile unit and a weight adjustment unit;

[0014] The user profiling unit uses the user's basic information, behavioral characteristics, and consumption characteristics as samples. It randomly selects K sample points as initial cluster centers, calculates the distance from each sample point to each cluster center, and assigns the sample points to the nearest cluster. For each cluster, it calculates the mean of all sample points in the cluster and uses it as the new cluster center. The process of assigning samples and updating cluster centers is repeated until the cluster centers no longer change significantly. Based on the characteristics of each cluster, a corresponding profile label is generated.

[0015] The weight adjustment unit trains the multilayer perceptron model using the training set, calculates the gradient of the loss function with respect to the model parameters using the backpropagation algorithm, and updates the parameters using the Adam optimizer to train the multilayer perceptron model for predicting user profile label weights.

[0016] The beneficial effect of adopting the above-mentioned further improvements is that, when generating user profiles, the K-means clustering algorithm clusters based on the characteristics of the data itself. It can automatically discover the natural grouping structure in the data and, by analyzing various behavioral data, attribute data and other multi-dimensional information of users, group users with similar characteristics into the same category, thereby achieving accurate classification of users.

[0017] When training a multilayer perceptron model, the multilayer perceptron model consists of an input layer, multiple hidden layers, and an output layer. The neurons in the hidden layers can perform nonlinear transformations on the input data with the help of activation functions. This allows the model to learn the complex nonlinear relationships in the data, thereby accurately predicting the weights of user profile labels.

[0018] As a further improvement to this technical solution, the user profiling unit uses the elbow method to determine the K value. By calculating the sum of squared errors within the cluster under different K values, a curve showing the relationship between the K value and the sum of squared errors within the cluster is plotted, and the K value corresponding to the inflection point of the curve is selected as the number of clusters.

[0019] The beneficial effect of the above-mentioned further improvement is that, when determining the number of clusters, the elbow method determines the optimal K value by plotting the relationship curve between the clustering error (such as the silhouette coefficient, distortion degree, etc.) and the K value under different K values. From the shape of the curve, as the K value increases, the clustering error usually gradually decreases. At first, the rate of decrease is relatively fast. When the K value reaches a certain point, the rate of decrease will slow down significantly. The curve shows a clear inflection point, just like a person's elbow. The K value corresponding to this inflection point is the better choice. This method does not require complex calculations or professional knowledge. You can roughly determine the optimal K value by simply observing the shape of the curve. It is very intuitive and simple.

[0020] As a further improvement to this technical solution, the weight adjustment unit establishes a real-time data access mechanism and adopts a message queue data transmission method to input the user's latest behavior and feature data into the trained multilayer perceptron model in a timely manner.

[0021] As a further improvement to this technical solution, the weight adjustment unit adopts a multi-class cross-entropy loss function to measure the difference between the predicted weights and the true weights of different image labels. The formula is as follows:

[0022] ;

[0023] in, The value of the cross-entropy loss function. This represents the probability distribution of the true label weights. To predict the probability distribution of label weights, This represents the number of image tags.

[0024] The beneficial effects of adopting the above-mentioned further improvements are that when using the multilayer perceptron model to predict the weight of user profile tags, user behavior and characteristics are constantly changing over time. The real-time data access mechanism can capture the latest dynamics of users and input new data into the multilayer perceptron model in a timely manner. The model can be adjusted and optimized according to the latest data. The latest data can help the model capture short-term trends and sudden changes in user behavior, thereby improving the accuracy of prediction, making advertising more precise, and improving advertising effectiveness and conversion rate.

[0025] The multi-class cross-entropy loss function accurately measures the difference between the model's predictions and the true labels. In user profile label weight prediction problems, there are usually multiple different label categories, each corresponding to different user features or attributes. This loss function considers the deviation between the predicted probability and the true probability for each category, thereby prompting the model to minimize this deviation during training and make the prediction results closer to reality.

[0026] As a further improvement to this technical solution, the RTB bidding module includes an evaluation value unit and a dynamic bidding unit;

[0027] The evaluation value unit quantifies the user's value using the dimensions of spending power, loyalty, and activity, and calculates the user's comprehensive value score by combining the weights of profile tags.

[0028] The dynamic bidding unit categorizes users into high-value, medium-value, and low-value users based on their overall value score, and adjusts the bidding strategy according to the advertiser's budget and objectives.

[0029] The beneficial effect of adopting the above-mentioned further improvements is that when calculating the user's comprehensive value score, evaluation is carried out through three dimensions, which can comprehensively depict the user's behavioral characteristics and value contribution from different perspectives. Spending power reflects the user's purchasing strength and consumption potential, loyalty reflects the user's dependence on the brand or product and long-term purchase intention, and activity shows the user's participation and interaction frequency on the platform or product. By comprehensively considering these three dimensions, the one-sidedness of single-dimensional evaluation can be avoided, and the overall value of the user can be grasped more accurately.

[0030] When adjusting bidding strategies, high-value users typically have higher spending potential and stronger purchasing intentions, making them more valuable to advertisers. By differentiating them from medium- and low-value users, advertisers can target high-value user groups with increased advertising budgets and higher bids, ensuring that ads are shown to these users more frequently and accurately. This increases ad exposure and click-through rates, and raises the likelihood of high-value users paying attention to and purchasing the advertiser's products or services.

[0031] As a further improvement to this technical solution, the formula for calculating the user's comprehensive value score by the evaluation value unit is as follows:

[0032] ;

[0033] in, To calculate the overall value score, For tags The weight, The quantized label value, This represents the number of tags.

[0034] As a further improvement to this technical solution, the dynamic bidding unit divides the advertiser's objectives into brand exposure, sales conversion, and potential customer acquisition, and restricts the bidding strategy with budget allocation and bid cap.

[0035] The beneficial effect of adopting the above-mentioned further improvements is that different advertising objectives correspond to different audience groups. By clearly defining objectives, advertisers can more accurately target their audiences, and budget allocation is a key factor in achieving advertising objectives. Rationally allocating the budget according to the three objectives of brand exposure, sales conversion, and lead acquisition allows resources to be allocated based on the importance and expected returns of each objective.

[0036] As a further improvement to this technical solution, the monitoring and feedback module sets up a log recording function on the server side of the advertising delivery system to monitor user interaction information and feedback data of the advertisement in real time.

[0037] The beneficial effects of adopting the above-mentioned further improvements are that, based on the user interaction information recorded in the logs, more personalized advertising recommendations can be provided to users. After understanding the user's preferences and behavioral habits, the system can accurately push advertisements that match the user's interests, reduce the frequency of users seeing irrelevant advertisements, make the advertising content more valuable to users, and thus improve users' acceptance and satisfaction with the advertisements.

[0038] As a further improvement to this technical solution, the monitoring and feedback module uses the upper and lower limits of the thresholds for ad impressions, click-through rate, and conversion rate as boundaries to further adjust the bidding strategy.

[0039] The beneficial effects of adopting the above-mentioned further improvements are that setting upper and lower limits for the number of impressions can ensure that the advertisement has a certain exposure base, while avoiding excessive exposure that would waste resources. When the actual number of impressions is lower than the lower limit, the bid can be appropriately increased to increase the opportunity for the advertisement to be displayed, ensuring that the brand or product has sufficient market exposure. If it is higher than the upper limit, it means that the advertisement may be displayed too frequently and exceeds the acceptance range of the target audience. At this time, the bid can be reduced, which can save costs and avoid users getting annoyed.

[0040] In addition to the objectives, features, and advantages described above, the present invention has other objectives, features, and advantages. The invention will now be described in further detail with reference to the figures. Attached Figure Description

[0041] Figure 1 This is a schematic diagram of the overall process of the present invention;

[0042] Figure 2 This is a schematic diagram illustrating the overall detailed process of the present invention.

[0043] The meanings of the labels in the diagram are as follows:

[0044] 100. User Profile Building Module; 110. User Profile Unit; 120. Weight Adjustment Unit; 200. RTB Bidding Module; 210. Value Evaluation Unit; 220. Dynamic Bidding Unit; 300. Monitoring and Feedback Module. Detailed Implementation

[0045] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0046] Currently, advertising delivery systems typically rely on data collection and analysis from a single platform. They evaluate ad performance by monitoring user behavior data (such as click-through rate and dwell time) on that platform. However, the control capabilities and accuracy are often limited by the data source and analysis methods, failing to fully reflect changes in user behavior and needs across platforms. To achieve precise ad delivery and improve click-through and conversion rates by collecting data from multiple platforms, generating user profiles, and adjusting tag weights based on users' latest behavior and feedback, a real-time bidding (RTB) interface can be used to formulate bidding strategies, participate in bidding in real-time, and flexibly adjust bidding strategies based on user interaction information and feedback data, thereby improving the efficiency and competitiveness of ad delivery.

[0047] Therefore, this invention proposes a method to collect basic user information, behavioral characteristics, and consumption characteristics from multiple platforms by constructing a user profiling module. Using the K-means clustering algorithm, user profile tags are generated. A multilayer perceptron model is established using user characteristic data as input, outputting the tag weights for predicting user profiles. The tag weights for corresponding user profiles are adjusted in real time. An RTB bidding module establishes RTB real-time bidding interfaces with multiple advertising trading platforms. Based on the user's comprehensive value score and constrained by the advertiser's budget and goals, the bidding strategy of the RTB real-time bidding interface is dynamically adjusted. A monitoring and feedback module monitors user interaction information and feedback data for advertisements in real time. An evaluation system is established using ad impressions, click-through rate, and conversion rate as indicators. An indicator threshold method is used to further adjust the bidding strategy of the RTB real-time bidding interface.

[0048] Specifically as follows:

[0049] Please see Figure 1 As shown, the present invention provides a cross-platform intelligent advertising delivery system based on user profiles, including a profile building module 100, an RTB bidding module 200, and a monitoring and feedback module 300.

[0050] The user profile module 100 uses web crawler technology to collect basic user information, behavioral characteristics, and consumption characteristics from multiple platforms. It uses the K-means clustering algorithm to generate user profile labels. Using user feature data as input, it establishes a multilayer perceptron model, outputs the label weights of the predicted user profiles, and adjusts the label weights of the corresponding user profiles in real time.

[0051] Using web crawler technology, we collect basic user information such as age, gender, region, education level, and occupation; behavioral characteristics such as page browsing frequency, dwell time, click behavior, and search keywords; and consumption characteristics such as spending amount, purchase frequency, type of products or services purchased, and coupon usage. The collected data is then organized into a structured table format, with each row representing a user and each column representing a characteristic.

[0052] like Figure 2 As shown, the profile building module 100 includes a user profile unit 110 and a weight adjustment unit 120;

[0053] User profiling unit 110 uses user basic information, behavioral characteristics, and consumption characteristics as samples, randomly selects K sample points as initial cluster centers, calculates the distance from each sample point to each cluster center, assigns the sample points to the nearest cluster, calculates the mean of all sample points in each cluster, uses it as the new cluster center, and repeats the above process of assigning samples and updating cluster centers until the cluster centers no longer change significantly, and generates corresponding profile labels based on the characteristics of each cluster.

[0054] The weight adjustment unit 120 uses the training set to train the multilayer perceptron model, calculates the gradient of the loss function with respect to the model parameters through the backpropagation algorithm, and updates the parameters using the Adam optimizer to train the multilayer perceptron model for predicting user profile label weights.

[0055] The data within each cluster is analyzed to calculate statistical measures of basic information, behavioral characteristics, and consumption characteristics, such as the mean, median, and mode. Based on these statistical measures, combined with business knowledge and market research, representative user profile tags are generated for each cluster. For example, if the users in a cluster are mainly aged 20-30, predominantly female, frequently browse fashion websites, have a high purchase frequency, and mainly buy clothing and cosmetics, then the user profile tag for this cluster can be defined as "young, fashionable female consumer."

[0056] For the multilayer perceptron model, the collected data is divided into a training set, a validation set, and a test set. The training set is used for learning the model's parameters, the validation set is used to adjust the model's hyperparameters (such as the learning rate, the number of hidden layer neurons, etc.), and the test set is used to evaluate the model's final performance.

[0057] In order to better determine the K value, the user profiling unit 110 uses the elbow method to determine the K value. By calculating the sum of squared errors within the cluster under different K values, a curve showing the relationship between the K value and the sum of squared errors within the cluster is plotted, and the K value corresponding to the inflection point of the curve is selected as the number of clusters.

[0058] Gradually increase the value of K from a small value (usually starting from 1), for example, K=1,2,3,⋯,p, where p is a pre-set maximum value. For each K value, use the K-means clustering algorithm to cluster the data and calculate the sum of squared errors within each cluster (SSE). Plot a relationship curve with K value on the x-axis and SSE on the y-axis and observe the shape of the curve. When the slope of the curve suddenly becomes gentler, the corresponding K value is the appropriate number of clusters. This point is like a person's elbow. Before this point, as K increases, SSE decreases rapidly. After this point, as K increases, the rate of decrease of SSE slows down significantly.

[0059] In order to input real-time data into the multilayer perceptron model, the weight adjustment unit 120 establishes a real-time data access mechanism and adopts a message queue data transmission method to input the user's latest behavior and feature data into the trained multilayer perceptron model in a timely manner.

[0060] Collect the latest user behavior and characteristic data from various data sources (such as websites, mobile applications, server logs, etc.). Log collection tools (such as Fluentd, Logstash) or custom scripts can be used to collect user behavior data (such as click events, browsing history, dwell time, etc.) and characteristic data (such as user age, gender, region, etc.). The collected data is then converted into a unified format, such as JSON or Protobuf, for easy subsequent processing and transmission. A message queue client library is used to send the processed data to a specified topic or queue in the message queue.

[0061] To better measure the difference between the predicted and true weights of different image labels, the weight adjustment unit 120 uses a multi-class cross-entropy loss function to measure the difference between the predicted and true weights of different image labels. The formula is as follows:

[0062] ;

[0063] in, The value of the cross-entropy loss function. This represents the probability distribution of the true label weights. To predict the probability distribution of label weights, This represents the number of image tags.

[0064] The multi-class cross-entropy loss function is often used to measure the difference between the model's prediction and the true label in multi-class problems. When processing portrait labels, each portrait label can be regarded as a category, and this loss function can be used to measure the difference between the predicted weight and the true weight.

[0065] The prepared true weight vector and predicted weight vector are substituted into the multi-class cross-entropy loss function formula for calculation. For each sample, the cross-entropy loss between its true weight and predicted weight is calculated separately. Then, as needed, the loss values ​​of multiple samples can be averaged or summed to obtain an overall loss value, which is used to measure the model's predictive performance on this batch of data.

[0066] The smaller the loss value, the smaller the difference between the predicted weights and the true weights, and the better the model predicts the image labels. Conversely, the larger the loss value, the greater the deviation between the model's prediction results and the actual situation. Further adjustments to the model's parameters or improvements to the model structure are needed to reduce the loss value and improve the model's accuracy.

[0067] In addition, the RTB bidding module 200 establishes RTB real-time bidding interfaces with multiple advertising exchange platforms, quantifies the value of users based on user profile tags, calculates the comprehensive value score of users by combining the weight of profile tags, and dynamically adjusts the bidding strategy of the RTB real-time bidding interface based on the comprehensive value score of users, with the advertiser's budget and goals as the limit.

[0068] The RTB bidding module 200 includes an evaluation value unit 210 and a dynamic bidding unit 220.

[0069] The evaluation value unit 210 quantifies the value of users based on the dimensions of spending power, loyalty, and activity, and calculates the user's comprehensive value score by combining the weights of profile tags.

[0070] Dynamic bidding unit 220 categorizes users into high-value, medium-value, and low-value users based on their overall value score, and adjusts the bidding strategy according to the advertiser's budget and objectives.

[0071] Spending power dimension:

[0072] Related tags: income level, asset status, consumption frequency, price range of purchased goods or services;

[0073] Assessment Method: Users with high income, substantial assets, frequent consumption, and a preference for high-priced goods typically possess greater purchasing power and are likely to be more valuable to businesses. For example, a user who frequently purchases high-end electronic products and has a monthly income exceeding 20,000 yuan has higher value in terms of purchasing power.

[0074] Loyalty dimension:

[0075] Related tags: purchase frequency, repeat purchase rate, brand preference, membership level;

[0076] Evaluation Method: Users with high purchase frequency, high repurchase rate, strong preference for specific brands, and high membership levels have high loyalty and are important valuable users for the company. For example, users who have purchased skincare products from a certain brand for many consecutive years and only choose that brand have outstanding value in terms of loyalty.

[0077] Activity level:

[0078] Related tags: login frequency, interaction behavior (such as comments, likes, shares, etc.), number of times participating in activities;

[0079] Evaluation Method: Users who frequently log in to the platform and actively participate in interactions and activities bring more traffic and vitality to the platform, and have high activity value. For example, users who log in to social media platforms multiple times a day and frequently post comments and share content have high activity levels.

[0080] Develop a bidding strategy:

[0081] High-value users: For high-value users, given their higher potential value to advertisers, bids can be appropriately increased, within budget limits, to increase the chances of gaining exposure and clicks from these users. For example, the bid could be 1.5 times the average bid.

[0082] Mid-value users: For mid-value users, bids can be kept close to the average bid level to balance the effectiveness and cost of ad campaigns;

[0083] Low-value users: For low-value users, bids can be appropriately reduced, or when budgets are tight, advertising to this group of users can be decreased. For example, the bid can be 0.8 times the average bid.

[0084] To better calculate the overall value score, the formula for calculating the user's overall value score in the evaluation value unit 210 is as follows:

[0085] ;

[0086] in, To calculate the overall value score, For tags The weight, The quantized label value, This represents the number of tags.

[0087] For example, if a user's spending amount tag has a quantified value of 0.8 and a weight of 0.4, browsing frequency tag has a quantified value of 0.6 and a weight of 0.3, and interaction count tag has a quantified value of 0.7 and a weight of 0.3, then the user's overall value score S = 0.4 × 0.8 + 0.3 × 0.6 + 0.3 × 0.7 = 0.71.

[0088] In order to better restrict bidding strategies, the dynamic bidding unit 220 divides advertisers’ objectives into brand exposure, sales conversion and lead generation, and restricts bidding strategies with budget allocation and bid cap;

[0089] Brand exposure aims to increase brand awareness and influence, allowing more potential customers to learn about the brand and its products or services. Metrics typically focused on under this objective include impressions, reach, and brand recognition. For example, this involves displaying ads on social media platforms, search engine results pages, or video websites to ensure the brand information is seen by as many users as possible.

[0090] Sales conversion: This focuses on directly driving the sale of products or services, converting potential customers into actual buyers. Key metrics for measuring this goal include conversion rate, sales revenue, and order volume. For example, this can be achieved through advertising on e-commerce platforms to guide users to purchase goods, or through offline advertising events to attract customers to in-store.

[0091] Acquiring potential customers: The focus is on collecting information about potential customers, such as contact information and interests, to facilitate subsequent marketing follow-up and nurturing, ultimately leading to conversion. Relevant metrics include the number of leads, form submissions, and registrations. Common methods include setting up form ads and offline event registration ads to attract users to leave their information;

[0092] Determine the acceptable cost range for each objective and set a bid cap based on this. For brand exposure objectives, since the focus is primarily on impressions, the cost per impression is relatively low. A reasonable bid cap can be set based on market averages and expected impression results, such as a bid cap of 10 yuan per thousand impressions. For sales conversion objectives, considering the need to achieve actual purchases, the cost is relatively high. A bid cap can be set based on factors such as product profit and expected conversion rate, such as 20 yuan per click. The bid cap for lead acquisition objectives falls between the two, and can be set based on the expected cost of acquiring a lead, such as a bid cap of 15 yuan per action (e.g., form submission).

[0093] In order to better monitor user interaction information and feedback data with the advertisement, the monitoring and feedback module 300 sets up a log recording function on the server side of the advertising delivery system to monitor user interaction information and feedback data with the advertisement in real time.

[0094] Record detailed user interactions with ads, such as ad display time, number of displays, click time, number of clicks, duration of video ad viewing (if it is a video ad), whether swipe or scroll actions were performed, and if a user submits feedback through the feedback channels provided by the ad delivery system (such as comments, reports, etc.), record the content, submission time, and feedback type of the feedback.

[0095] For example, in Python, the logging library can be used, which provides flexible logging configuration options, allowing you to set the log level (such as DEBUG, INFO, WARNING, ERROR, etc.), output format, and output destination (file, console, etc.). In Java, log4j, logback, and other commonly used logging frameworks are powerful and widely applied.

[0096] In order to further adjust the bidding strategy, the monitoring and feedback module 300 uses the upper and lower limits of the thresholds for ad impressions, click-through rate and conversion rate as boundaries to further adjust the bidding strategy;

[0097] Ad impressions:

[0098] Threshold setting: Lower limit: Refer to the minimum number of impressions in historical data that show the ad to be displayed normally and bring a certain effect, while also taking into account the industry average. For example, if the average number of impressions for similar ads in the same industry is 5,000 times / day, and your own historical data shows that the ad effect is not obvious if it is below 3,000 times / day, then the lower limit threshold can be set to 3,000 times / day;

[0099] Upper limit: Consider the advertising budget, platform capacity, and actual business needs. If the advertising budget is limited, increasing it after the number of impressions reaches a certain level may lead to excessive costs and diminishing marginal returns. For example, if the cost increases too rapidly after exceeding 10,000 impressions / day, the upper limit threshold can be set to 10,000 impressions / day.

[0100] Adjustment measures:

[0101] Below the minimum: Increase bids to enhance ad competitiveness; check ad settings, such as whether targeting criteria are too narrow, and broaden the audience reach; optimize ad creatives to increase their appeal and thus increase impressions.

[0102] Exceeding the upper limit: Lower bids to control costs; analyze the reasons for high impressions but poor performance in other metrics, such as whether the ad creative is misleading, leading to invalid impressions; optimize ad targeting to make ad display more accurate and improve display effect.

[0103] In summary, the working principle of this solution is as follows:

[0104] This cross-platform intelligent advertising delivery system based on user profiles collects basic user information, behavioral characteristics, and consumption characteristics from multiple platforms by constructing a profile module 100. It uses the K-means clustering algorithm to generate user profile tags. Using user feature data as input, a multilayer perceptron model is established to output the predicted tag weights of user profiles. The tag weights of corresponding user profiles are adjusted in real time, enabling the collection of user data from multiple platforms. Based on the real-time feature data of users, the tag weights of user profiles are adjusted to ensure the real-time and accuracy of user profiles, providing a basis for subsequent participation in RTB bidding.

[0105] The RTB bidding module 200 establishes RTB real-time bidding interfaces with multiple advertising exchanges. It quantifies and evaluates user value based on user profile tags, calculates the user's comprehensive value score by combining the weight of the profile tags, and dynamically adjusts the bidding strategy of the RTB real-time bidding interface based on the user's comprehensive value score, with the advertiser's budget and goals as constraints. It participates in bidding in real time and adjusts the bidding strategy according to user tags to improve the efficiency and competitiveness of advertising.

[0106] The monitoring and feedback module 300 monitors user interaction information and feedback data for advertisements in real time. It establishes an evaluation system based on ad impressions, click-through rate, and conversion rate. Using the indicator threshold method, it further adjusts the bidding strategy of the RTB real-time bidding interface, continuously optimizes the advertising effect, and adjusts the strategy in a timely manner based on user interaction information and feedback data to maximize the benefits of advertising.

[0107] The foregoing has shown and described 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. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A cross-platform intelligent advertising delivery system based on user profiles, characterized in that: It includes a profile building module (100), an RTB bidding module (200), and a monitoring and feedback module (300). The profile building module (100) uses web crawler technology to collect basic information, behavioral characteristics and consumption characteristics of users from multiple platforms, uses K-means clustering algorithm to generate profile labels for users, uses user feature data as input to build a multilayer perceptron model, outputs the label weights of predicted user profiles, and adjusts the label weights of corresponding user profiles in real time. The RTB bidding module (200) establishes RTB real-time bidding interfaces with multiple advertising trading platforms, quantifies the value of users based on user profile tags, calculates the comprehensive value score of users by combining the weight of profile tags, and dynamically adjusts the bidding strategy of the RTB real-time bidding interface based on the comprehensive value score of users and the budget and goals of advertisers. The monitoring and feedback module (300) monitors the user's interaction information and feedback data with the advertisement in real time, establishes an evaluation system with the number of ad impressions, click-through rate and conversion rate as indicators, and further adjusts the bidding strategy of the RTB real-time bidding interface by using the indicator threshold method.

2. The cross-platform intelligent advertising delivery system based on user profiles according to claim 1, characterized in that: The profile building module (100) includes a user profile unit (110) and a weight adjustment unit (120). The user profiling unit (110) uses the user's basic information, behavioral characteristics information and consumption characteristics data as samples, randomly selects K sample points as initial cluster centers, calculates the distance from each sample point to each cluster center, assigns the sample points to the nearest cluster, calculates the mean of all sample points in each cluster for each cluster, uses it as the new cluster center, and repeats the above process of assigning samples and updating cluster centers until the cluster centers no longer change significantly, and generates corresponding profile labels based on the characteristics of each cluster. The weight adjustment unit (120) trains the multilayer perceptron model using the training set, calculates the gradient of the loss function with respect to the model parameters through the backpropagation algorithm, and updates the parameters using the Adam optimizer to train the multilayer perceptron model for predicting user profile label weights.

3. The cross-platform intelligent advertising delivery system based on user profiles according to claim 2, characterized in that: The user profiling unit (110) uses the elbow method to determine the K value. By calculating the sum of squared errors within the cluster under different K values, it plots the relationship curve between the K value and the sum of squared errors within the cluster, and selects the K value corresponding to the inflection point of the curve as the number of clusters.

4. The cross-platform intelligent advertising delivery system based on user profiles according to claim 2, characterized in that: The weight adjustment unit (120) establishes a real-time data access mechanism and adopts a message queue data transmission method to input the user's latest behavior and feature data into the trained multilayer perceptron model in a timely manner.

5. The cross-platform intelligent advertising delivery system based on user profiles according to claim 2, characterized in that: The weight adjustment unit (120) uses a multi-class cross-entropy loss function to measure the difference between the predicted weights and the true weights of different image labels. The formula is as follows: ； in, The value of the cross-entropy loss function. This represents the probability distribution of the true label weights. To predict the probability distribution of label weights, This represents the number of image tags.

6. The cross-platform intelligent advertising delivery system based on user profiles according to claim 1, characterized in that: The RTB bidding module (200) includes an evaluation value unit (210) and a dynamic bidding unit (220). The evaluation value unit (210) quantifies the value of users based on the dimensions of spending power, loyalty, and activity, and calculates the user's comprehensive value score by combining the weights of profile tags. The dynamic bidding unit (220) classifies users into high-value users, medium-value users and low-value users based on the user's comprehensive value score, and adjusts the bidding strategy based on the advertiser's budget and goals.

7. The cross-platform intelligent advertising delivery system based on user profiles according to claim 6, characterized in that: The formula for calculating the user's overall value score by the evaluation value unit (210) is as follows: ； in, To calculate the overall value score, For tags The weight, The quantized label value, This represents the number of tags.

8. The cross-platform intelligent advertising delivery system based on user profiles according to claim 6, characterized in that: The dynamic bidding unit (220) divides the advertiser's objectives into brand exposure, sales conversion and lead generation, and limits the bidding strategy with budget allocation and bid cap.

9. The cross-platform intelligent advertising delivery system based on user profiles according to claim 1, characterized in that: The monitoring and feedback module (300) monitors user interaction information and feedback data with respect to advertisements in real time by setting a log recording function on the server side of the advertising delivery system.

10. The cross-platform intelligent advertising delivery system based on user profiles according to claim 1, characterized in that: The monitoring and feedback module (300) further adjusts the bidding strategy based on the upper and lower limits of the thresholds for ad impressions, click-through rate, and conversion rate.

Images