Advertisement risk pre-judgment and control method and device, storage medium and equipment

Abstract

The invention provides an advertisement risk pre-judgment and control method and device, a storage medium and equipment, and the method comprises the steps: collecting input data in real time, including advertisement platform data, user behavior data, scene feature data, user hierarchical data and external environment data, generating a core feature vector for pre-judging an advertisement putting risk according to the input data, and storing the core feature vector in a database; and inputting the core feature vector into a risk pre-judgment model trained in advance, outputting pre-judgment risk scores corresponding to at least two time windows with different time lengths in the future, calling a preset risk intervention strategy matrix, and matching and executing corresponding intervention actions according to the pre-judgment risk scores, the scene feature data and the user hierarchical data. According to the method, active intervention can be performed when the index does not exceed the standard, hysteresis consumption is reduced, the risk judgment accuracy is improved, the situation that short-time hesitation of a high-intention user is misjudged as a risk is avoided, the risk is moderately tolerated for the flow of a high-value user, the flow of a low-value user is strictly controlled, and the conversion efficiency is improved.

Description

Technical Field

[0001] This invention relates to the field of advertising placement technology, and in particular to a method, apparatus, storage medium and equipment for predicting and controlling advertising risks. Background Technology

[0002] In the digital advertising field, with intensifying competition for traffic and diversified user needs, advertisers are demanding increasingly stringent requirements for precise control over advertising risks. Effective advertising risk prediction and control can not only reduce ineffective cost consumption but also ensure the stable acquisition of high-quality traffic, directly impacting return on investment (ROI) and long-term business growth. However, existing advertising risk prediction and control methods still have many limitations and are insufficient to meet the refined operational needs in complex scenarios.

[0003] Existing methods largely rely on real-time metric monitoring, such as monitoring single data points like click-through rate (CTR) and cost per conversion (CPC), triggering price adjustments or pauses when these metrics exceed thresholds. While these methods achieve basic automation, they have significant limitations:

[0004] First, the prediction is delayed, only responding passively after a risk occurs (e.g., pausing only after costs have exceeded the limit), failing to avoid losses in advance. Second, the judgment dimensions are singular, ignoring the impact of user behavior (e.g., browsing depth, interaction intent) and scenario characteristics (e.g., time period, device), making it prone to misjudgment due to occasional fluctuations (e.g., classifying short-term hesitation of high-intent users as a risk). Third, the control strategy is rigid, applying the same standards to high-value users and ordinary users, and to promotional events and daily scenarios, leading to the loss of high-quality traffic or excessive intervention. Fourth, there is a lack of dynamic correction mechanisms, failing to track the effects after adjustments, making it difficult to adapt to complex and ever-changing advertising environments. Summary of the Invention

[0005] In view of this, the present invention provides a method, apparatus, storage medium and equipment for predicting and controlling advertising risks, which can realize early prediction of risks, reduce the misjudgment rate and reduce ineffective cost consumption.

[0006] In a first aspect, embodiments of the present invention provide a method for predicting and controlling advertising risks, the method comprising:

[0007] Real-time collection of input data, including advertising platform data, user behavior data, scene feature data, user segmentation data, and external environment data;

[0008] Generate a core feature vector for predicting advertising placement risks based on the input data;

[0009] The core feature vector is input into a pre-trained risk prediction model, which outputs the predicted risk score for at least two time windows of different durations, the confidence interval of the predicted risk score for each time window, and the risk trend.

[0010] The system invokes a preset risk intervention strategy matrix and matches and executes corresponding intervention actions based on the predicted risk score, scenario feature data, and user segmentation data.

[0011] Furthermore, the step of generating a core feature vector for predicting advertising placement risks based on the input data includes:

[0012] Extract user behavior features from the user behavior data;

[0013] Based on the advertising platform data, scene feature data, and external environment data, scene risk characteristics are generated.

[0014] Extract trend features from the advertising platform data;

[0015] Extract user segmentation features from the user segmentation data;

[0016] The user behavior features, scenario risk features, trend features, and user segmentation features are standardized and arranged in a fixed order to obtain the core feature vector.

[0017] Furthermore, the user behavior features include interaction depth score, dwell time deviation, and conversion intention probability. Extracting user behavior features from the user behavior data includes:

[0018] Interaction action sequences are extracted from the user behavior data, and scores matching each interaction action in the interaction action sequence are matched according to preset rules and summed to obtain an interaction depth score.

[0019] Extract the actual dwell time of the current user from the user behavior data, obtain the historical average dwell time of the same scene from the scene feature data, and obtain the dwell time deviation based on the actual dwell time and the historical average dwell time;

[0020] The user behavior data and historical conversion records of similar users are input into a pre-trained logistic regression model to obtain the conversion intention probability;

[0021] The scenario risk characteristics include a scenario risk coefficient and a time-period fluctuation index. These characteristics are generated based on advertising platform data, scenario feature data, and external environment data, and include:

[0022] Extract scene tags from the scene feature data, and extract competitor promotion status from the external environment data;

[0023] Call the preset scenario combination weight table to obtain the basic coefficients corresponding to the scenario label and competitor promotion status;

[0024] The scenario risk characteristics are obtained by multiplying the scenario tags and the base coefficients corresponding to the competitor's promotion status.

[0025] Extract the real-time consumption amount within a preset time period from the data of the advertising platform, and retrieve the historical consumption amount for the same period;

[0026] The time-period fluctuation index is obtained based on the real-time consumption amount and the historical consumption amount for the same period.

[0027] Furthermore, the trend feature includes the sliding window slope value. Extracting trend features from the advertising platform data includes:

[0028] Click-through rate and conversion cost are extracted from the data of the advertising platform according to a preset cycle;

[0029] Calculate the linear regression slope of the click-through rate within the first sliding window and the linear regression slope of the conversion cost within the second sliding window, respectively, for the first and second duration sliding windows, to obtain trend characteristics;

[0030] Extracting user segmentation features from the user segmentation data includes:

[0031] Extract user tags from the user stratification data;

[0032] The preset user stratification weight mapping table is called to obtain the weights corresponding to the extracted user tags. When a user has multiple user tags at the same time, the maximum weight is taken as the user stratification feature.

[0033] Furthermore, a preset risk intervention strategy matrix is ​​invoked, and corresponding intervention actions are matched and executed based on the predicted risk score, scenario feature data, and user segmentation data, including:

[0034] When the predicted risk score is greater than or equal to the first threshold, and the scenario feature data is an e-commerce promotion and the user segmentation data is a new user, the advertising platform will call the pause API interface to pause advertising and push new customer coupons to the user.

[0035] When the predicted risk score is greater than the second threshold and less than the first threshold, and the scene feature data is an educational trial scene and the user segmentation data shows that watching the trial video exceeds the preset time period, the advertising platform bidding API interface is called to maintain the bid and add an "Instant Reservation" floating window to the page.

[0036] When the predicted risk score is below the third threshold, and the scenario feature data is a weekday PC terminal and the user segmentation data is a high-value user, the advertising platform bidding API interface is called to increase the bid by a preset amount and the advertisement is displayed first.

[0037] Furthermore, the method also includes:

[0038] The actual risk score after the intervention action is performed is collected at preset time intervals, and the actual risk score is compared with the predicted risk score to obtain the prediction deviation rate.

[0039] When the predicted deviation rate exceeds the preset deviation rate, the size of the first threshold and / or the second threshold and / or the third threshold in the risk intervention strategy matrix is ​​automatically adjusted.

[0040] Furthermore, the method also includes generating an intervention report.

[0041] Secondly, embodiments of the present invention provide an advertising risk prediction and control device, the device comprising:

[0042] The data acquisition module is used to collect input data in real time, including advertising platform data, user behavior data, scene feature data, user segmentation data, and external environment data.

[0043] The core feature vector generation module is used to generate a core feature vector for predicting the risk of ad placement based on the input data.

[0044] The risk prediction module is used to input the core feature vector into a pre-trained risk prediction model and output the predicted risk score corresponding to at least two time windows of different durations in the future.

[0045] The intervention module is used to invoke a preset risk intervention strategy matrix, and match and execute corresponding intervention actions based on the predicted risk score, scenario feature data and user segmentation data.

[0046] Thirdly, embodiments of the present invention provide a computer-readable storage medium storing a computer program, wherein the computer program is configured to execute the method described in any one of the first aspects when running.

[0047] Fourthly, embodiments of the present invention provide an electronic device, including a memory and a processor, characterized in that the memory stores a computer program, and the processor is configured to run the computer program to perform the method described in any one of the first aspects.

[0048] The technical solution provided by this invention predicts risk trends in advance through a risk prediction model and proactively intervenes when indicators are not exceeded (e.g., reducing prices in advance when costs are predicted to surge), thereby reducing delayed costs. Simultaneously, it integrates multi-dimensional data such as user behavior (e.g., interaction depth, dwell time), scenario characteristics (e.g., peak promotional periods), and user segmentation (e.g., high-value users) to improve the accuracy of risk assessment and avoid misjudging short-term hesitation of high-intent users as risk. Through a risk intervention strategy matrix, it moderately tolerates risks in high-value user traffic (e.g., continuing to run ads even with slight cost overruns) while strictly controlling low-value traffic, thus improving conversion efficiency. Furthermore, it provides feedback on intervention effects at set time intervals and automatically adjusts strategy thresholds to ensure stable operation even in complex scenarios such as competitor promotions and traffic fluctuations. Attached Figure Description

[0049] Figure 1 This is a flowchart of an advertising placement control method provided in an embodiment of the present invention.

[0050] Figure 2 This is a flowchart illustrating the process of generating core feature vectors in an embodiment of the present invention.

[0051] Figure 3 This is a flowchart illustrating the implementation of corresponding intervention actions in an embodiment of the present invention.

[0052] Figure 4 This is a schematic diagram of the structure of an advertising risk prediction and control device provided in an embodiment of the present invention.

[0053] Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0055] See Figure 1 , Figure 1 This is a flowchart of an advertising risk prediction and control method provided by an embodiment of the present invention. The method includes the following steps:

[0056] Step 11: Collect input data in real time. Input data includes advertising platform data, user behavior data, scene feature data, user segmentation data, and external environment data.

[0057] In step 11, advertising platform data refers to real-time performance metrics returned by the advertising platform, including impressions, clicks, conversions, current bids, and spending; user behavior data refers to user interaction data with ads and landing pages, such as dwell time, clicks on the "add to cart" button, and browsing paths (e.g., browsing paths from the homepage to the details page); scenario feature data refers to quantitative information describing the advertising scenario, such as time period tags (e.g., weekends 19:00-22:00), device type (e.g., mobile / PC), network environment (e.g., 5G), and industry scenario (e.g., e-commerce 618 promotion); user segmentation data refers to classification information based on users' historical value, such as high-value users (e.g., historical LTV ≥ 500 yuan), new users (e.g., registered ≤ 7 days), and churn warning users (e.g., inactive for 30 days); external environment data refers to external factors affecting advertising, such as competitor promotion status (e.g., competitors are having discount promotions) and industry traffic index (e.g., 0-100, with higher values ​​indicating stronger traffic).

[0058] In implementation, the computer can collect data in parallel at a high frequency of 10-30 seconds per instance. It pulls data from advertising platforms in real time via open APIs (such as Baidu Promotion API and ByteDance Engine API), aggregating data every minute (e.g., merging 10-second data into a minute-level average). It connects to front-end tracking systems (e.g., via JavaScript tracking) to collect user behavior data, associating it with user IDs to form behavior sequences (e.g., User A: stays for 2 minutes → clicks to add to cart → browses the payment page) to obtain user behavior data. It matches time period tags with system time, identifies device types through device user agents, obtains network environment data through network status interfaces, and combines industry calendars to mark major promotional scenarios to obtain scene feature data. It calls internal user tagging system interfaces (such as CRM system interfaces) to obtain user segmentation results and historical LTV, associating them with current advertising traffic to obtain user segmentation data. Finally, it obtains external environment data through competitor monitoring tools (such as third-party data crawlers) and aligns it with scene feature data by timestamp. After collection, the computer automatically filters out outliers (such as invalid data where the number of clicks in a single instance exceeds three times the average for the same scene) and fills in missing values ​​(such as using the average dwell time of the same device type to fill in missing values) to ensure data integrity.

[0059] Step 11 provides raw data for subsequent feature generation and risk prediction, ensuring that the input data is comprehensive, real-time, and accurate, and avoiding biases in subsequent analysis due to missing or delayed data. For example, if competitor promotional data is omitted, it may lead to the misjudgment that high spending on one's own advertising is normal, missing the opportunity for intervention.

[0060] Step 12: Generate a core feature vector for predicting advertising placement risks based on the input data.

[0061] In step 12, the core feature vector is a one-dimensional array formed by combining multiple quantified features in a fixed order. It is the input of the risk prediction model and includes user behavior features, scenario risk features, trend features, user segmentation features, etc.

[0062] This step transforms unstructured raw data (such as user add-to-cart actions and weekend evening scenarios) into quantifiable features that the risk prediction model can recognize, solving the problem that raw data cannot be directly input into the model. For example, only after a user's add-to-cart action is quantified through interaction depth scoring can the risk prediction model identify its impact on risk prediction.

[0063] Step 13: Input the core feature vector into the pre-trained risk prediction model and output the predicted risk score corresponding to at least two time windows of different durations in the future.

[0064] In step 13, the risk prediction model is a pre-trained machine learning model (such as XGBoost or LSTM time series model), which takes a core feature vector as input and outputs a quantitative score of future risk. The time window refers to a preset future time period, such as the next 30 minutes or the next hour, used to define the time range for prediction. The predicted risk score is a quantitative value of 0-100 points output by the risk prediction model. The higher the score, the higher the risk of advertising within the future time window (e.g., 80 points indicates high risk, which may lead to a surge in costs).

[0065] The computer loads the pre-trained risk prediction model and performs the following operations:

[0066] The core feature vector generated in step 12 is input into the pre-trained risk prediction model. The risk prediction model calculates the risk score for at least two future time windows using an internal algorithm (such as a decision tree combination in XGBoost) (e.g., a risk score of 80 for the next 30 minutes and 85 for the next hour). In other embodiments of the present invention, the risk prediction model can also output the risk score and confidence interval for each time window (e.g., at a 90% confidence level, the risk score for the next 30 minutes is 75-85); it can also mark risk trends (e.g., if the risk score for the next 30 minutes is 80 and the risk score for the next hour is 85, then the risk trend is marked as increased risk).

[0067] In other embodiments of the present invention, the risk prediction model is incrementally trained every hour using newly generated actual data to update parameters and optimize prediction accuracy (e.g., using actual data from the morning to correct the prediction model for the afternoon).

[0068] This step enables early risk prediction, overcoming the limitation of responding only after risks occur in existing technologies. For example, if the risk prediction model predicts a risk score of 80 (high risk) in the next 30 minutes, intervention can be made before costs exceed the budget, reducing ineffective spending.

[0069] Step 14: Invoke the preset risk intervention strategy matrix, match and execute the corresponding intervention actions according to the predicted risk score, scenario feature data and user segmentation data.

[0070] In step 14, the risk intervention strategy matrix refers to the pre-defined correspondence table (which can be a two-dimensional or three-dimensional matrix) between the predicted risk score, scenario characteristics, user segmentation, and intervention actions; the intervention actions refer to the specific operations taken to control the risk, such as adjusting prices, suspending keywords, and pushing coupons.

[0071] Specifically, the computer invokes the risk intervention strategy matrix, matches and executes actions according to the following logic:

[0072] From the risk intervention strategy matrix, identify the predicted risk range (e.g., above 80 points) + scenario characteristics (e.g., e-commerce promotions) + user segmentation (e.g., new users), and the corresponding intervention actions are to suspend advertising and push coupons to new customers.

[0073] Then, the bidding API of the advertising platform is called to adjust the bid, the keyword status API is called to pause non-core keywords, and the internal coupon system API is called to push coupons.

[0074] Finally, record the action execution time and parameters (e.g., price reduction of 10%) for subsequent effect tracking.

[0075] This step controls risk based on risk level, scenario characteristics, and user value differences, avoiding a one-size-fits-all approach. Through precise matching of actions, it balances risk control and traffic value, improving ad delivery efficiency. For example:

[0076] For users with a risk score of 80 and who are classified as high-value users, a strategy of slightly reducing the price while prioritizing the display of the advertisement will be implemented (to avoid losing core users).

[0077] For users with a predicted risk score of 80 and who are classified as low-value users, the strategy of pausing ad spending will be implemented (to control ineffective spending).

[0078] The implementation methods of each step in the embodiments of the present invention will be described in detail below.

[0079] Please see Figure 2 , Figure 2 This is a flowchart illustrating the process of generating the core feature vector in this embodiment of the invention, specifically step 12, which may include the following steps:

[0080] Step 121: Extract user behavior features from user behavior data.

[0081] Step 121 involves extracting user behavior features from user behavior data. These features are used to quantify users' intention to engage with and the depth of their interaction with the advertisement.

[0082] In other embodiments of the present invention, user behavior characteristics include interaction depth score, dwell time deviation, and conversion intention probability. Step 121 can be implemented through the following steps:

[0083] Step 1211: Extract the interaction action sequence from the user behavior data, match the scores of each interaction action in the interaction action sequence according to preset rules, and sum them to obtain the interaction depth score.

[0084] Step 1212: Extract the current user's actual dwell time from user behavior data, obtain the historical average dwell time for the same scene from scene feature data, and obtain the dwell time deviation based on the actual dwell time and the historical average dwell time.

[0085] Step 1213: Input user behavior data and historical conversion records of similar users into the pre-trained logistic regression model to obtain the conversion intention probability.

[0086] Specifically, the computer generates quantifiable features from the user behavior data (such as dwell time, interaction actions, and browsing paths) collected in step 11 using preset rules. For example, when calculating the interaction depth score, the user's browsing, adding to cart, and inquiry actions are weighted and accumulated (1 point for browsing 3 pages, 3 points for adding to cart, and 5 points for inquiring; if the user completes browsing 3 pages and adding to cart, the score is 1+3=4 points); when calculating the dwell time deviation, the current user's dwell time (e.g., 2 minutes) is compared with the average dwell time in the same scenario (e.g., 1.5 minutes), resulting in a deviation of 33%; a logistic regression model can also output the conversion intention probability (e.g., when the user adds to cart and browses the payment page, the logistic regression model outputs 75%). These features directly reflect the user's potential conversion intention, providing a user-dimensional basis for risk prediction.

[0087] Step 122: Generate scenario risk characteristics based on advertising platform data, scenario feature data, and external environment data.

[0088] Step 122 involves combining advertising platform data, scenario feature data, and external environment data to generate scenario risk characteristics, which are used to assess the potential risk level under different scenarios. Scenario risk characteristics include scenario risk coefficients and time-period fluctuation indices. Step 122 can be achieved through the following steps:

[0089] Step 1221: Extract scene tags from the scene feature data and extract competitor promotion status from the external environment data.

[0090] Step 1222: Call the preset scene combination weight table to obtain the basic coefficients corresponding to the scene tags and competitor promotion status.

[0091] Step 1223: Multiply the scene tag and the base coefficient corresponding to the competitor's promotion status to obtain the scene risk coefficient.

[0092] Step 1224: Extract the real-time consumption amount within the preset time period from the advertising platform data, and retrieve the historical consumption amount for the same period.

[0093] Step 1225: Obtain the time period fluctuation index based on the real-time consumption amount and the historical consumption amount for the same period.

[0094] Computers integrate various types of data to construct quantitative indicators strongly correlated with specific scenarios. For example, the scenario risk coefficient is generated by multiplying the base coefficients (e.g., 1.1, 1.2, and 1.3 respectively) corresponding to scenario tags such as weekend evenings, mobile devices, and competitor promotions, resulting in 1.716 (simplified to 1.7). A higher base coefficient indicates a higher scenario risk. The time-period fluctuation index is calculated by comparing the consumption in the past hour (e.g., 1000 yuan) with the historical consumption in the same period (e.g., 800 yuan), resulting in 1.25, reflecting the degree of deviation between the current time-period consumption and historical norms. These features address the problem of existing technologies neglecting scenario differences, making risk assessment more closely aligned with the actual deployment environment.

[0095] Step 123: Extract trend features from advertising platform data.

[0096] Step 123 involves extracting trend features from the advertising platform data to capture the dynamic trends of campaign metrics. Trend features include the sliding window slope value. Step 123 can be achieved through the following steps:

[0097] Step 1231: Extract click-through rate and conversion cost from the advertising platform data according to the preset cycle.

[0098] Step 1232: Set up a first duration sliding window and a second duration sliding window, calculate the linear regression slope of the click-through rate in the first sliding window, calculate the linear regression slope of the conversion cost in the second sliding window, and obtain trend characteristics.

[0099] Computers use sliding window technology to perform short- and medium-term trend analysis on real-time metrics of advertising platforms (such as click-through rate (CTR) and cost-to-conversion (CPC)). For example, by setting up two sliding windows—one for the last 5 minutes and one for the last 30 minutes—linear regression is performed on the CTR and CPC data within each window to calculate the slope: if the CTR rises from 2% to 3% in the last 5 minutes with a slope of 0.2 (an increase of 0.2 percentage points per minute), it indicates a short-term upward trend in CTR; if the CPC rises from 100 yuan to 120 yuan in the last 30 minutes with a slope of 0.67, it indicates a continued increase in medium-term costs. These two slope values ​​together constitute the trend characteristics, providing a dynamic basis for risk prediction and avoiding the need to judge risk solely based on static indicators.

[0100] Step 124: Extract user segmentation features from user segmentation data.

[0101] Step 124 involves extracting user segmentation features from user segmentation data to achieve differentiated treatment of user traffic of different value. Step 124 can be achieved through the following steps:

[0102] Step 1241: Extract user tags from user segmentation data.

[0103] Step 1242: Call the preset user stratification weight mapping table to obtain the weights corresponding to the extracted user tags. When a user has multiple user tags at the same time, the maximum weight is taken as the user stratification feature.

[0104] The computer extracts user tags (such as high-value users, new users, and churn warning users) from user segmentation data and matches them with corresponding value weights (e.g., 1.2 for high-value users, 0.8 for new users, and 0.5 for churn warning users) using a pre-defined weight mapping table. If a user has multiple tags (e.g., new user and high-potential user, with corresponding weights of 0.8 and 1.0 respectively), the maximum value of 1.0 is taken as its segmentation feature. This feature ensures that the system prioritizes retaining high-value user traffic during risk control, avoiding the loss of high-quality traffic due to a one-size-fits-all approach.

[0105] Step 125: Standardize user behavior features, scenario risk features, trend features, and user segmentation features, and arrange them in a fixed order to obtain the core feature vector.

[0106] Step 125 involves standardizing the four types of features and combining them into a core feature vector, providing a unified input format for the risk prediction model. The computer first performs Z-score standardization on all features (e.g., an interaction depth score of 4 becomes 1.33 after processing), eliminating the magnitude differences between different features. Then, the standardized features are arranged in a fixed order (e.g., interaction depth score, dwell time deviation, scene risk coefficient, 30-minute cost slope, value weight, etc.) to form a one-dimensional vector (e.g., [1.33, 0.25, 1.2, 0.05, 1.2]). Standardization ensures that the risk prediction model treats each feature fairly, while the fixed order guarantees the consistency of the feature vector, enabling the risk prediction model to stably receive input and output reliable prediction results. Ultimately, this achieves a standardized transformation from raw data to the risk prediction model input, laying the foundation for subsequent risk prediction.

[0107] In some embodiments of the present invention, please refer to Figure 3 , Figure 3 This is a flowchart illustrating the implementation of the corresponding intervention action in this embodiment of the invention. Specifically, step 14 can be achieved through the following steps:

[0108] Step 141: When the predicted risk score is greater than or equal to the first threshold, and the scenario feature data is an e-commerce promotion and the user segmentation data is a new user, then call the advertising platform's pause API interface to pause advertising and push new customer coupons to the user.

[0109] Step 141 involves precise intervention in high-risk scenarios. When the predicted risk score reaches or exceeds the first threshold (e.g., 80 points represents high risk), and the scenario is an e-commerce promotion with a new user, the system will perform two actions: First, it will pause the current ad campaign via the advertising platform's API interface, quickly cutting off the consumption of high-risk traffic and avoiding waste due to soaring costs in highly competitive promotional scenarios. Second, it will push exclusive new customer coupons to the new user through the internal coupon system. While controlling risk, it will use discounts to stimulate new users to complete conversions, balancing loss mitigation and retention, thus solving the problem of potential user churn caused by traditional direct suspension of high-risk campaigns. For example, during the 618 e-commerce promotion, if a new user's traffic risk score is predicted to be 85 points, the system will pause ad campaigns and push a "30 RMB off for every 100 RMB spent" new customer coupon, reducing ineffective spending and increasing the probability of the new user placing an order for the first time.

[0110] Step 142: When the predicted risk score is greater than the second threshold but less than the first threshold, and the scene feature data is an educational trial scenario and the user segmentation data shows that the viewing of the trial video exceeds the preset time period, the advertising platform bidding API interface is called to maintain the bid and an "Instant Reservation" floating window is added to the page.

[0111] Step 142 addresses flexible responses to medium-risk scenarios. When the predicted risk score falls between the second threshold (e.g., 60 points) and the first threshold (80 points) (medium risk), and the scenario is an educational trial lesson where the user watches the trial video for more than a preset time period (e.g., 5 minutes, indicating high user intent), the system will adopt a strategy combining maintaining the bid and promoting conversion: maintaining the current bid through the advertising platform's bidding API to avoid reduced exposure due to price reductions; and simultaneously adding a "Book Now" floating window on the landing page to lower the user's decision-making threshold and encourage high-intent users to act quickly. The core of this design is to distinguish between effective intent fluctuations and real risk. For example, during an educational institution's enrollment period, if a user watches a math trial lesson for 7 minutes (exceeding the preset 5-minute value), even with a risk score of 70 (medium risk), the system will maintain the bid and display the booking entry to avoid missing potential students due to excessive intervention.

[0112] Step 143: When the predicted risk score is below the third threshold, and the scenario feature data is a weekday PC terminal and the user segmentation data is a high-value user, the advertising platform bidding API interface is called to increase the bid by a preset amount and the advertisement is displayed first.

[0113] Step 143 focuses on capturing traffic in low-risk scenarios. When the predicted risk score is below the third threshold (e.g., 50 points, low risk), and the scenario is a weekday PC platform with high-value users, the system will proactively increase the ad placement: by increasing the bid by a preset margin (e.g., 10%) through the ad platform's bidding API interface, and triggering a priority display mechanism (e.g., the ad ranking is moved up 3 positions). The logic behind this action is that high-value user traffic in low-risk environments has high conversion potential and needs to be proactively captured to improve overall ad placement efficiency. For example, at 9 AM on a weekday, if a high-value user (e.g., historical LTV exceeding 2000 yuan) of a certain enterprise service ad has a risk score of 40, the system will increase the bid by 10% and prioritize display, ensuring that these high-quality users see the ad first, thereby increasing the reach and conversion rate of high-value users.

[0114] In other embodiments of the present invention, after step 143, the method may further include:

[0115] Step 144: Collect the actual risk score after the intervention action is performed at preset time intervals, compare the actual risk score with the predicted risk score, and obtain the prediction deviation rate.

[0116] Step 144 involves dynamic tracking and deviation assessment of the intervention's effectiveness, aiming to quantify the accuracy of the risk prediction model. After the intervention action (such as increasing bids or prioritizing display) is executed in step 143, the computer collects actual campaign data (such as click-through rate and conversion cost) at preset time intervals (such as every 10 minutes) and recalculates the actual risk score based on this data (the calculation logic is consistent with steps 12 and 13). Subsequently, the actual risk score is compared with the predicted risk score output in step 13, and the result is calculated according to the formula. Calculate the deviation rate. For example, if the predicted risk score is 40 (low risk) and the actual risk score is calculated to be 60, the deviation rate is (60-40) / 40×100%=50%. This value directly reflects the degree of deviation between the prediction and the actual risk, providing a basis for subsequent strategy adjustments.

[0117] Step 145: When the predicted deviation rate exceeds the preset deviation rate, automatically adjust the size of the first threshold and / or the second threshold and / or the third threshold in the risk intervention strategy matrix.

[0118] Step 145 is a threshold adaptive optimization based on the deviation rate, ensuring that the risk intervention strategy matrix can dynamically adapt to the actual deployment environment. When the predicted deviation rate calculated in step 144 exceeds the preset deviation rate (e.g., 20%), the system will automatically adjust the first, second, or third threshold in the risk intervention strategy matrix. For example, if there are multiple instances where the predicted risk score is lower than the third threshold (50 points) but the actual risk score reaches 60 points (deviation rate above 20%), it indicates that the original third threshold was set too low, which may lead to misjudgment of low risk. In this case, the system will raise the third threshold from 50 points to 55 points, making the low-risk judgment standard more stringent. Conversely, if the actual risk score is consistently lower than the predicted value, the threshold may be lowered. This step solves the problem that fixed thresholds are difficult to adapt to complex deployment scenarios, making risk classification more accurate through dynamic correction, and ensuring that intervention actions always match the actual risk level.

[0119] In other embodiments of the present invention, the advertising risk prediction and control method may further include:

[0120] Step 15: Generate an intervention report.

[0121] Step 15 generates an intervention report, which is a structured record and review of the entire risk prediction and intervention process, aiming to achieve process traceability, effect analysis, and strategy optimization.

[0122] For example, after each intervention action is completed (or at a fixed period, such as daily / per batch), the computer automatically summarizes the following core information to generate a report:

[0123] Basic information: intervention time, advertising plans / keywords involved, corresponding predicted risk score and actual risk score;

[0124] Scenarios and user characteristics: Scenarios at the time of intervention (e.g., "e-commerce promotion", "weekday PC terminal"), user segmentation results (e.g., new users, high-value users);

[0125] Intervention details: the specific operations performed (such as pausing ad delivery, increasing bids by 10%, or sending coupons), the API interfaces called, and the parameters;

[0126] Results data: Comparison of key metrics before and after intervention (e.g., click-through rate increased from 2% to 3%, conversion cost decreased from 100 yuan to 80 yuan), prediction deviation rate, and user conversion results (e.g., order rate of new users using coupons);

[0127] Note on abnormalities: If the prediction deviation rate exceeds the threshold or the intervention effect does not meet expectations, please indicate the possible reasons (such as competitors temporarily increasing promotional efforts, resulting in actual risks being higher than predicted).

[0128] Once the report is generated, it will be automatically stored in the system log library and can be displayed through a visual interface (such as trend charts and scenario-effect correlation analysis diagrams) for operations personnel to review. For example, by analyzing the report, it can be found that in the educational trial scenario, the intervention of maintaining the bid and adding an appointment entry increased the conversion rate of high-intent users by 20%, and the strategy matrix for this scenario can be further optimized. If a certain type of scenario frequently shows prediction deviations, the feature weights of the risk prediction model can be adjusted accordingly.

[0129] This step solves the problems of unrecorded intervention processes and difficulty in evaluating effects in traditional methods, providing data support for manual optimization and model iteration, and continuously improving the accuracy and adaptability of the system.

[0130] See Figure 4 , Figure 4 This is a schematic diagram of an advertising risk prediction and control device provided in an embodiment of the present invention. The device includes:

[0131] The data acquisition module 21 is used to collect input data in real time, including advertising platform data, user behavior data, scene feature data, user segmentation data, and external environment data.

[0132] The core feature vector generation module 22 is used to generate a core feature vector for predicting the risk of ad placement based on the input data.

[0133] Risk prediction module 23 is used to input the core feature vector into a pre-trained risk prediction model and output the predicted risk score corresponding to at least two time windows of different durations in the future.

[0134] Intervention module 24 is used to call the preset risk intervention strategy matrix, match and execute corresponding intervention actions based on the predicted risk score, scenario feature data and user segmentation data.

[0135] In some embodiments of the present invention, the core feature vector generation module 22 may include:

[0136] User behavior feature unit 221 is used to extract user behavior features from the user behavior data;

[0137] The scenario risk feature unit 222 is used to generate scenario risk features based on the advertising platform data, scenario feature data and external environment data;

[0138] Trend feature unit 223 is used to extract trend features from the advertising platform data;

[0139] User stratification feature unit 224 is used to extract user stratification features from the user stratification data;

[0140] The standardization unit 225 is used to standardize the user behavior features, scenario risk features, trend features and user segmentation features, and arrange them in a fixed order to obtain the core feature vector.

[0141] In other embodiments of the present invention, the user behavior features include interaction depth score, dwell time deviation, and conversion intention probability, and the user stratification feature unit 221 may include:

[0142] The interaction depth unit 2211 is used to extract the interaction action sequence from the user behavior data, match the scores of each interaction action in the interaction action sequence according to preset rules and sum them to obtain the interaction depth score.

[0143] The dwell time deviation subunit 2212 is used to extract the actual dwell time of the current user from the user behavior data, obtain the historical average dwell time of the same scene from the scene feature data, and obtain the dwell time deviation based on the actual dwell time and the historical average dwell time.

[0144] The conversion intention probability subunit 2213 is used to input the user behavior data and historical conversion records of similar users into a pre-trained logistic regression model to obtain the conversion intention probability.

[0145] In other embodiments of the present invention, the scenario risk characteristics include a scenario risk coefficient and a time-period fluctuation index, and the scenario risk characteristic unit 222 may include:

[0146] The competitor promotion status subunit 2221 is used to extract scene tags from the scene feature data and extract competitor promotion status from the external environment data.

[0147] The basic coefficient subunit 2222 is used to call a preset scene combination weight table to obtain the basic coefficients corresponding to the scene label and competitor promotion status;

[0148] The scenario risk feature subunit 2223 is used to multiply the scenario label and the basic coefficient corresponding to the competitor's promotion status to obtain the scenario risk feature;

[0149] The time-period fluctuation index subunit 2224 is used to extract the real-time consumption amount within a preset time period from the data of the advertising platform, retrieve the historical consumption amount for the same period, and obtain the time-period fluctuation index based on the real-time consumption amount and the historical consumption amount for the same period.

[0150] In other embodiments of the present invention, the trend feature includes a sliding window slope value, and the trend feature unit 223 may include:

[0151] Extraction subunit 2231 is used to extract click-through rate and conversion cost from the advertising platform data according to a preset period;

[0152] The calculation subunit 2232 is used to set a first duration sliding window and a second duration sliding window, calculate the linear regression slope of the click-through rate in the first sliding window, calculate the linear regression slope of the conversion cost in the second sliding window, and obtain trend characteristics.

[0153] In other embodiments of the present invention, the user hierarchical feature unit 224 may include:

[0154] User tag extraction subunit 2241 is used to extract user tags from the user hierarchical data;

[0155] Subunit 2242 is invoked to call the preset user stratification weight mapping table to obtain the weights corresponding to the extracted user tags. When a user has multiple user tags at the same time, the maximum weight is taken as the user stratification feature.

[0156] In some embodiments of the present invention, the intervention module 24 may include:

[0157] The first intervention subunit 241 is used to call the advertising platform's pause API interface to pause advertising and push new customer coupons to the user when the predicted risk score is greater than or equal to the first threshold and the scenario feature data is an e-commerce promotion and the user segmentation data is a new user.

[0158] The second intervention subunit 242 is used to call the advertising platform bidding API interface to maintain the bid while adding an immediate appointment floating window on the page when the predicted risk score is greater than the second threshold and less than the first threshold, and the scene feature data is an educational auditing scene and the user stratification data shows that the viewing of the auditing video exceeds a preset time period.

[0159] The third intervention subunit 243 is used to call the advertising platform bidding API interface to increase the bid by a preset amount and prioritize displaying the advertisement when the predicted risk score is below the third threshold, and the scene feature data is a weekday PC terminal and the user segmentation data is a high-value user.

[0160] In other embodiments of the present invention, the intervention module 24 may further include:

[0161] The comparison subunit 244 is used to collect the actual risk score after the intervention action is performed at preset time intervals, and compare the actual risk score with the predicted risk score to obtain the prediction deviation rate.

[0162] The adjustment subunit 245 is used to automatically adjust the size of the first threshold and / or the second threshold and / or the third threshold in the risk intervention strategy matrix when the predicted deviation rate exceeds the preset deviation rate.

[0163] In other embodiments of the invention, the apparatus may further include generating an intervention report 25.

[0164] It should be noted that the advertising risk prediction and control device in this embodiment of the invention and the advertising risk prediction and control method in the above embodiments belong to the same inventive concept. Technical details not described in detail in this device can be found in the previous description of the method, and will not be repeated here.

[0165] Furthermore, embodiments of the present invention also provide a computer-readable storage medium storing a computer program, wherein the computer program is configured to execute the methods described above when running.

[0166] Figure 5 This is a schematic diagram of the structure of an electronic device 10 provided in an embodiment of the present invention. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices (such as helmets, glasses, watches, etc.), and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0167] like Figure 5As shown, the electronic device 10 includes at least one processor 11 and a memory, such as a read-only memory (ROM) 12 or a random access memory (RAM) 13, communicatively connected to the at least one processor 11. The memory stores computer programs executable by the at least one processor. The processor 11 can perform various appropriate actions and processes based on the computer program stored in the ROM 12 or loaded from storage unit 18 into the RAM 13. The RAM 13 may also store various programs and data required for the operation of the electronic device 10. The processor 11, ROM 12, and RAM 13 are interconnected via a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14.

[0168] Multiple components in electronic device 10 are connected to I / O interface 15, including: input unit 16, such as keyboard, mouse, etc.; output unit 17, such as various types of displays, speakers, etc.; storage unit 18, such as disk, optical disk, etc.; and communication unit 19, such as network card, modem, wireless transceiver, etc. Communication unit 19 allows electronic device 10 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0169] Processor 11 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of processor 11 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. Processor 11 performs the various methods and processes described above, such as the idle detection method.

[0170] In some embodiments, the idle detection method may be implemented as a computer program tangibly contained in a computer-readable storage medium, such as storage unit 18. In some embodiments, part or all of the computer program may be loaded and / or mounted on electronic device 10 via ROM 12 and / or communication unit 19. When the computer program is loaded into RAM 13 and executed by processor 11, one or more steps of the idle detection method described above may be performed. Alternatively, in other embodiments, processor 11 may be configured to perform the idle detection method by any other suitable means (e.g., by means of firmware).

[0171] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0172] Computer programs used to implement the methods of the present invention can be written in any combination of one or more programming languages. These computer programs can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the processor, the computer programs cause the functions / operations specified in the flowcharts and / or block diagrams to be performed. The computer programs can be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0173] In the context of this invention, a computer-readable storage medium can be a tangible medium that may contain or store a computer program for use by or in conjunction with an instruction execution system, apparatus, or device. A computer-readable storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination thereof. Alternatively, a computer-readable storage medium may be a machine-readable signal medium. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0174] To provide interaction with a user, the systems and techniques described herein can be implemented on an electronic device having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the electronic device. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0175] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as data servers), or computing systems that include middleware components (e.g., application servers), or computing systems that include frontend components (e.g., user computers with graphical user interfaces or web browsers through which users can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., communication networks). Examples of communication networks include local area networks (LANs), wide area networks (WANs), blockchain networks, and the Internet.

[0176] A computing system can include clients and servers. Clients and servers are generally located far apart and typically interact through communication networks. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. The server can be a cloud server, also known as a cloud computing server or cloud host, which is a hosting product within the cloud computing service system to address the shortcomings of traditional physical hosts and VPS services, such as high management difficulty and weak business scalability.

[0177] It should be understood that the sorting, addition, or deletion steps can be re-executed using the various forms of processes shown above. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0178] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for predicting and controlling advertising risks, characterized in that, The method includes: Real-time collection of input data, including advertising platform data, user behavior data, scene feature data, user segmentation data, and external environment data; Generate a core feature vector for predicting advertising placement risks based on the input data; The core feature vector is input into a pre-trained risk prediction model, which outputs the predicted risk scores for at least two time windows of different durations in the future. The system invokes a preset risk intervention strategy matrix and matches and executes corresponding intervention actions based on the predicted risk score, scenario feature data, and user segmentation data.

2. The method according to claim 1, characterized in that, The process of generating a core feature vector for predicting advertising placement risks based on the input data includes: Extract user behavior features from the user behavior data; Based on the advertising platform data, scene feature data, and external environment data, scene risk characteristics are generated. Extract trend features from the advertising platform data; Extract user segmentation features from the user segmentation data; The user behavior features, scenario risk features, trend features, and user segmentation features are standardized and arranged in a fixed order to obtain the core feature vector.

3. The method according to claim 2, characterized in that, The user behavior features include interaction depth score, dwell time deviation, and conversion intention probability. User behavior features are extracted from the user behavior data, including: Interaction action sequences are extracted from the user behavior data, and scores matching each interaction action in the interaction action sequence are matched according to preset rules and summed to obtain an interaction depth score. Extract the actual dwell time of the current user from the user behavior data, obtain the historical average dwell time of the same scene from the scene feature data, and obtain the dwell time deviation based on the actual dwell time and the historical average dwell time; The user behavior data and historical conversion records of similar users are input into a pre-trained logistic regression model to obtain the conversion intention probability; The scenario risk characteristics include a scenario risk coefficient and a time-period fluctuation index. These characteristics are generated based on advertising platform data, scenario feature data, and external environment data, and include: Extract scene tags from the scene feature data, and extract competitor promotion status from the external environment data; Call the preset scenario combination weight table to obtain the basic coefficients corresponding to the scenario label and competitor promotion status; The scenario risk coefficient is obtained by multiplying the scenario label and the base coefficient corresponding to the competitor's promotion status. Extract the real-time consumption amount within a preset time period from the data of the advertising platform, and retrieve the historical consumption amount for the same period; The time-period fluctuation index is obtained based on the real-time consumption amount and the historical consumption amount for the same period.

4. The method according to claim 2, characterized in that, The trend features include the sliding window slope value. Extracting trend features from the advertising platform data includes: Click-through rate and conversion cost are extracted from the data of the advertising platform according to a preset cycle; Set up a first-duration sliding window and a second-duration sliding window, calculate the linear regression slope of the click-through rate within the first sliding window, calculate the linear regression slope of the conversion cost within the second sliding window, and obtain trend characteristics; Extracting user segmentation features from the user segmentation data includes: Extract user tags from the user stratification data; The preset user stratification weight mapping table is called to obtain the weights corresponding to the extracted user tags. When a user has multiple user tags at the same time, the maximum weight is taken as the user stratification feature.

5. The method according to claim 1, characterized in that, Invoke the preset risk intervention strategy matrix, and match and execute corresponding intervention actions based on the predicted risk score, scenario feature data, and user segmentation data, including: When the predicted risk score is greater than or equal to the first threshold, and the scenario feature data is an e-commerce promotion and the user segmentation data is a new user, the advertising platform will call the pause API interface to pause advertising and push new customer coupons to the user. When the predicted risk score is greater than the second threshold and less than the first threshold, and the scene feature data is an educational trial scenario and the user segmentation data shows that watching the trial video exceeds the preset time period, the advertising platform bidding API interface is called to maintain the bid and an "Instant Reservation" floating window is added to the page. When the predicted risk score is below the third threshold, and the scenario feature data is a weekday PC terminal and the user segmentation data is a high-value user, the advertising platform bidding API interface is called to increase the bid by a preset amount and the advertisement is displayed first.

6. The method according to claim 5, characterized in that, The method further includes: The actual risk score after the intervention action is performed is collected at preset time intervals, and the actual risk score is compared with the predicted risk score to obtain the prediction deviation rate. When the predicted deviation rate exceeds the preset deviation rate, the size of the first threshold and / or the second threshold and / or the third threshold in the risk intervention strategy matrix is ​​automatically adjusted.

7. The method according to any one of claims 1-6, characterized in that, The method also includes generating intervention reports.

8. An advertising risk prediction and control device, characterized in that, The device includes: The data acquisition module is used to collect input data in real time, including advertising platform data, user behavior data, scene feature data, user segmentation data, and external environment data. The core feature vector generation module is used to generate a core feature vector for predicting the risk of ad placement based on the input data. The risk prediction module is used to input the core feature vector into a pre-trained risk prediction model and output the predicted risk score corresponding to at least two time windows of different durations in the future. The intervention module is used to invoke a preset risk intervention strategy matrix, and match and execute corresponding intervention actions based on the predicted risk score, scenario feature data and user segmentation data.

9. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, wherein the computer program is configured to execute the method of any one of claims 1 to 7 when it is run.

10. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to run the computer program to perform the method of any one of claims 1 to 7.

Images