Advertisement consumption estimation method and device for AppStore application
By using intelligent association of multi-source data and dual-model fusion training, the problems of weak correlation of indicators, non-standard data processing and insufficient model accuracy in App Store advertising have been solved, achieving higher accuracy and stable advertising consumption prediction and adapting to market changes.
Patent Information
- Application Number
- CN202610062194.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-01-17
- Publication Date
- 2026-04-28
AI Technical Summary
Existing technologies for App Store advertising suffer from weak correlation of metrics, non-standard data processing, insufficient model accuracy, and lack of iteration mechanisms, leading to inaccurate keyword selection decisions and insufficient prediction accuracy.
We employ a method of intelligent association of multi-source data, multi-dimensional feature engineering, and dual-model fusion training. We generate a feature matrix through multi-dimensional feature engineering, combine gradient boosting regression and random forest regression models to predict consumption, and regularly update the model parameters to adapt to market changes.
It improved data quality and prediction accuracy, reduced prediction errors, achieved more accurate advertising expenditure prediction and adaptability, and enhanced the stability and practicality of the model.
Smart Images
Figure CN121935870A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of digital advertising technology, specifically to a method and apparatus for estimating ad spending on the App Store. Background Technology
[0002] In App Store advertising (ASA) scenarios, advertisers' core objective is to acquire users' downloads at a reasonable cost. To achieve this goal, advertisers need to select search terms, or keywords, for ad placement. Existing technologies typically rely on a single or a few independent metrics for keyword selection decisions, such as popularity or trending. The specific implementation logic involves: collecting basic metrics such as keyword popularity and search index, setting threshold ranges for these metrics, prioritizing keywords that meet the threshold requirements, and using a single model to make a simple prediction of ad spend after placement.
[0003] However, existing technologies have the following drawbacks: Weak correlation between metrics: Reliable metrics such as popularity and heat may contradict each other, for example, "low popularity but high heat", and the direct correlation between these metrics and ad spending has not been fully explored, leading to inaccurate keyword selection decisions.
[0004] Non-standard data processing: The data sources are diverse and scattered, with problems such as inconsistent formats, outliers (such as a search index of 0), and missing data. The lack of a unified and standardized preprocessing process affects the reliability of subsequent analysis.
[0005] Insufficient model accuracy: The prediction models used are often singular and cannot simultaneously capture the complex linear and nonlinear relationships between features and ad spending, resulting in insufficient prediction accuracy.
[0006] Lack of iteration mechanism: Once the model is trained, it remains unchanged and cannot adapt to dynamic changes such as market trends, user behavior and store algorithm rules, resulting in a significant decline in long-term prediction performance. Summary of the Invention
[0007] The technical problem to be solved by this invention is that existing methods for predicting advertising consumption suffer from weak correlation of indicators, non-standard data processing, insufficient model accuracy, and lack of iterative mechanisms.
[0008] This invention provides the following technical invention: The App Store application advertising consumption prediction method proposed in this invention specifically includes the following steps: S1: Intelligent correlation of multi-source data Extract raw data containing application identifiers, keywords, consumption values, and multiple keyword feature indicators from the database; establish data associations, clean and filter invalid data, unify data types, and obtain a cleaned benchmark dataset; S2: Multi-dimensional Feature Engineering The benchmark dataset is divided into a training subset and a prediction subset; statistical values are calculated based on the training subset; missing and outlier imputation, numerical transformation, statistical feature construction, interactive feature construction, and feature filtering are performed on the training subset and the prediction subset to generate the final feature matrix and target variable. S3: Dual-model fusion training Based on the feature matrix and target variables of the training subset, a hierarchical sampling method is used to divide the internal training set and validation set; a gradient boosting regression model is constructed and trained as the main model and a random forest regression model is trained as the auxiliary model; the prediction results of the two models are combined through a weighted fusion strategy, and the performance of the fused model is evaluated on the validation set. S4: Consumption Prediction Application and Iteration The fusion model trained is used to estimate the consumption of new keyword features; steps S1 to S3 are periodically re-executed based on new data to update the model parameters and maintain the estimation accuracy.
[0009] Preferred technical solution one: In step S2, the construction of statistical features includes: for each application identifier, calculating the mean and median of the consumption of all keywords under it; for each keyword, calculating the mean, median and quantile of its consumption under different applications.
[0010] Preferred technical solution 2: The interactive feature construction includes: multiplying the standardized keyword feature indicators with the logarithmically transformed statistical features to generate interactive features.
[0011] Preferred technical solution three: In step S3, the hierarchical sampling is based on the binning results of the target variable to ensure that the data distribution of different consumption levels in the training set and the validation set is consistent.
[0012] Preferred technical solution four: Step S4 further includes: persistently saving the converter, model parameters and feature list used for data preprocessing to support offline and online prediction calls.
[0013] This invention relates to an App Store application advertising consumption prediction device, comprising: The multi-source data intelligent association module is used to execute step S1, integrate scattered data sources, clean invalid data, unify data format, and provide high-quality and consistent raw data for subsequent feature engineering, thus solving the problems of scattered and low-quality data. The multi-dimensional feature engineering module is used to execute step S2, which processes the cleaned raw data by performing features, including missing value repair, feature encoding, feature enhancement, and feature selection, transforming the raw data into effective features that meet the requirements of model training and improving the quality of model input. The dual-model fusion training module is used to execute step S3, constructing a gradient boosting regression (GBR) main model and a random forest regression (RF) auxiliary model, taking into account both linear and nonlinear relationships between features and costs, and improving prediction accuracy through weighted fusion to solve the problem of poor adaptability of a single model; The consumption prediction iteration module is used to execute step S4, load the trained model for real-time prediction, and manage the regular updates and iterations of the model.
[0014] The App Store application advertising consumption prediction method and apparatus proposed in this invention achieve the following beneficial effects: (1) Data accuracy: Multi-source data association and integration: Unified data sources and data formats, and the elimination of invalid data through strict filtering rules, improve data integrity to over 95%; Refined data processing: The "reuse of training set statistical values" strategy is adopted to avoid data leakage, and outliers are filled with the median, significantly improving data quality; Precise feature mining: Consumption-related factors are captured in all aspects through statistical features, interaction features, and binning features, and the correlation between features and consumption is improved by 30%; (2) Prediction accuracy: Advantages of dual-model fusion: GBR model captures linear relationships, RF model captures nonlinear relationships, and after fusion, MAPE is reduced by 20%-40% compared with a single model; Parameter optimization guarantee: Through RandomizedSearchCV and hierarchical cross-validation, the model stability is improved, and the validation set R² can reach more than 0.75; Error controllable: Custom weighted MAPE index limits the impact of extreme errors, and the prediction results are more in line with the actual business scenario.
[0015] (3) Practicality and scalability: Real-time prediction capability: Preprocessing logic and model are stored separately, making it convenient to call. The prediction response time for a single data item is ≤0.1 seconds; Continuous iteration mechanism: Supports incremental data updates, automatically adapts to business changes, and the long-term prediction accuracy decay rate is ≤5% / month; Strong scalability: The feature engineering module supports adding new feature types, and the model module can be replaced with other regression models to adapt to more business scenarios.
[0016] (4) System stability: Resource management is standardized: Database connections are closed in a timely manner after use to avoid resource leakage and support large-scale data processing; Anomaly handling is well done: Built-in safe functions such as safe_log1p and safe_to_float are used to avoid program crashes caused by abnormal data types and extreme values; Results are traceable: Statistical values, bucket boundaries and model parameters are saved during the training process, and the predicted results can be backtracked for analysis. Attached Figure Description
[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart illustrating a high-safety sheet metal parts production equipment and its production method proposed in this invention. Detailed Implementation
[0018] The technologies described below will be clearly and completely described in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.
[0019] like Figure 1 As shown in the figure, the App Store application advertising consumption estimation method proposed in this invention specifically includes the following steps. S1: Intelligent correlation of multi-source data (a) Database connection and query: Connect to the target MySQL database using the pymysql library, execute SQL query statements, and extract core fields; (b) Data conversion: Call the safe_to_float function to convert numeric fields such as consume and search_index to float type. Fields that cannot be converted are marked as missing values. (c) Data filtering: Filter records where consume>0 and are not empty, remove invalid data, and output the cleaned original data set; S2: Multi-dimensional Feature Engineering (a) Data splitting and labeling: Label the raw data output from step S1 as either the training set (90%) or the prediction set (10%). The training set needs to record statistical characteristics (median, quantiles, etc.). (b) Missing / outlier repair: Training set: Fill the 0 and missing values of search_index with its own median, fill the missing values of popularity with its own median, and fill the missing values of highImpression_share with its own 25th quantile; Prediction set: Reuse the statistical values from the training set to fill in the corresponding missing and outlier values; (c) Feature generation: Statistical features: Calculate consumption-related statistics by grouping by app_id and keyword, and generate features such as app_consume_mean and kw_consume_q75; Enhanced features: Numerical features are processed using QuantileTransformer and StandardScaler, and statistical features are processed using safe_log1p transformation to construct four types of interactive features; Bucketing feature: The training set `consume` is divided into quartile buckets, and the prediction set reuses the bucket boundaries of the training set to generate `consume_bucket`. (d) Feature selection: Select from Model (based on RF model) to select core features and output feature matrix (X) and target variable (y); The statistical feature construction includes: for each application identifier, calculating the mean and median of the consumption of all keywords under it; for each keyword, calculating the mean, median and quantile of its consumption in different applications; S3: Dual-model fusion training (a) Internal splitting of the training set: The training set feature matrix (X_train) output in step S3 is split into an internal training set and a validation set in an 8:2 ratio, and stratified sampling is performed based on consume_bucket; (b) Model construction: Construct the GBR main model and the RF auxiliary model respectively, and configure the initial parameters; (c) Parameter optimization: Perform RandomizedSearchCV (5-fold stratified cross-validation) on the GBR model to search for the optimal parameter combination; (d) Model training: Train the optimized GBR model (with the transformed target variable y_transformed as input) and the RF model (with the log-transformed y as input) using the internal training set. (e) Model fusion: Define a weighted fusion function (GBR prediction result weight 0.7, RF prediction result weight 0.3), and output the fused prediction function; (f) Model evaluation: Use the validation set to evaluate the MSE, RMSE, MAPE and other metrics of the fusion model, and output the feature importance ranking; (g) Model saving: Save the GBR pipeline, RF pipeline, y_transformer, and feature column list to the specified directory; The stratified sampling is based on the binning results of the target variable to ensure that the data distribution of different consumption levels in the training set and the validation set is consistent. S4: Consumption Prediction Application and Iteration (a) Real-time prediction: Input the feature to be predicted, call the preprocessed instance saved in the training set to perform feature processing; call the fusion prediction function, and output the consumption prediction result and the interpretation of feature importance; (b) Result evaluation: Estimate the predicted set (10%), output the comparison results between the actual value and the predicted value, grouped statistical indicators, and error distribution; (c) Model iteration: Regularly (weekly) acquire incremental data, repeat steps S1 to S3, and update the model; monitor the prediction error, and if it exceeds the threshold, trigger feature re-selection or parameter tuning; (d) Resource release: After all operations are completed, close the database connection and release system resources; Step S4 also includes: persistently saving the converter, model parameters and feature list used for data preprocessing to support offline and online prediction calls.
[0020] This invention relates to an App Store application advertising consumption prediction device, comprising: The multi-source data intelligent association module is used to execute step S1, integrate scattered data sources, clean invalid data, unify data format, and provide high-quality and consistent raw data for subsequent feature engineering, thus solving the problems of scattered and low-quality data. The multi-dimensional feature engineering module is used to execute step S2, which processes the cleaned raw data by performing features, including missing value repair, feature encoding, feature enhancement, and feature selection, transforming the raw data into effective features that meet the requirements of model training and improving the quality of model input. The dual-model fusion training module is used to execute step S3, constructing a gradient boosting regression (GBR) main model and a random forest regression (RF) auxiliary model, taking into account both linear and nonlinear relationships between features and costs, and improving prediction accuracy through weighted fusion to solve the problem of poor adaptability of a single model; The consumption prediction iteration module is used to execute step S4, load the trained model for real-time prediction, and manage the regular updates and iterations of the model.
[0021] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, material, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, material, or apparatus.
[0022] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A method for estimating ad spend in the App Store, characterized in that, Includes the following steps: S1: Intelligent correlation of multi-source data Extract raw data containing application identifiers, keywords, consumption values, and multiple keyword feature indicators from the database; establish data associations, clean and filter invalid data, unify data types, and obtain a cleaned benchmark dataset; S2: Multi-dimensional Feature Engineering The benchmark dataset is divided into a training subset and a prediction subset, and statistical features and interaction features are constructed based on the training subset to generate a feature matrix for model training and prediction. S3: Dual-model fusion training Based on the feature matrix and target variable of the training subset, a hierarchical sampling method is used to divide the internal training set and validation set; a gradient boosting regression model is constructed and trained as the main model and a random forest regression model is trained as the auxiliary model; the prediction results of the two models are combined to obtain the final prediction model. S4: Consumption Prediction Application and Iteration The fusion model, after training, is used to predict the consumption of new keyword features; and the final prediction model is updated periodically based on the new data.
2. The App Store application advertising consumption estimation method according to claim 1, characterized in that, In step S2, the construction of statistical features includes: for each application identifier, calculating the mean and median of the consumption of all keywords under it; for each keyword, calculating the mean, median and quantile of its consumption in different applications.
3. The App Store application advertising consumption prediction method according to claim 1, characterized in that, The interactive feature construction includes: multiplying the standardized keyword feature indicators with the logarithmically transformed statistical features to generate interactive features.
4. The App Store application advertising consumption prediction method according to claim 1, characterized in that, In step S3, the stratified sampling is performed based on the binning results of the target variable to ensure that the data distribution of different consumption levels in the training set and the validation set is consistent.
5. The App Store application advertising consumption prediction method according to claim 1, characterized in that, Step S4 also includes: persistently saving the converter, model parameters and feature list used for data preprocessing to support offline and online prediction calls.
6. The apparatus for the App Store application advertising consumption estimation method according to any one of claims 1-5, characterized in that, include: The multi-source data intelligent association module is used to perform step S1 in claim 1, integrate scattered data sources, clean invalid data, unify data format, and provide high-quality and consistent raw data for subsequent feature engineering, thereby solving the problems of scattered and low-quality data. The multi-dimensional feature engineering module is used to perform step S2 in claim 1, which processes the cleaned raw data by performing feature processing, including missing value repair, feature encoding, feature enhancement, and feature filtering, and transforms the raw data into effective features that meet the requirements of model training, thereby improving the quality of model input. The dual-model fusion training module is used to execute step S3 in claim 1, constructing a gradient boosting regression main model and a random forest regression auxiliary model, taking into account both linear and nonlinear relationships between features and consumption, and improving prediction accuracy through weighted fusion. The consumption prediction iteration module is used to execute step S4 in claim 1, load the trained model for real-time prediction, and manage the periodic updates and iterations of the model.