A feature normalization enhancement-based risk control model construction method and system

Abstract

The application provides a risk control model construction method and system based on feature normalization enhancement, which comprises seven main steps of obtaining multi-source heterogeneous data, feature extraction and fusion, dynamic adaptation mechanism establishment, distribution deviation detection and parameter updating, multi-granularity feature normalization, feature feedback optimization and model training evaluation, and significantly improves the feature discrimination and stability of the risk control model.

Description

Technical Field

[0001] This invention relates to the field of risk control technology, and in particular to a method and system for constructing a risk control model based on feature normalization enhancement. This method improves the feature discriminativeness and stability of the risk control model through dynamic feature parameter updates and multi-granularity normalization processing, making it suitable for risk control applications in finance, e-commerce, and the internet. Background Technology

[0002] Risk control models play a crucial role in finance, e-commerce, and the internet, serving as key technological support for ensuring secure business operations. These models require processing large amounts of multi-source, heterogeneous data and generating features based on this data. They then use machine learning and other algorithms to identify abnormal behavior, fraud risks, and credit risks.

[0003] In actual online risk control system operation, risk control models typically need to continuously receive real-time data streams from multiple business platforms and complete risk identification and decision output based on these data streams. Due to factors such as changes in user behavior patterns, adjustments in business strategies, and changes in the external environment, the data distribution of model input features may change continuously or abruptly over time, leading to an imbalance in the range of model input values, a decrease in feature discrimination, and consequently, system operation problems such as unstable model output and increased false positive rate.

[0004] Traditional risk control model building typically employs simple feature engineering methods, such as performing basic statistical operations on the data, including summation, mean, maximum, and minimum values, and then directly inputting the generated features into the modeling process. Common data standardization methods mainly include Min-Max standardization and Z-score standardization. The former maps the data to the [0,1] interval, while the latter performs normalization based on the mean and standard deviation.

[0005] In more advanced risk control models, feature engineering typically employs a normalization strategy with fixed parameters to statically process features. These methods perform well in environments with relatively stable data distributions. However, in real-world production environments, data distribution can change significantly over time due to factors such as changes in user behavior patterns, seasonality, and policy adjustments—a phenomenon known as data drift. Furthermore, traditional methods have limited capabilities for processing non-numerical data and struggle to effectively integrate multi-source, heterogeneous data.

[0006] However, existing risk control model construction methods mostly employ fixed-parameter feature normalization strategies or offline statistical methods, making it difficult to respond promptly to changes in feature distribution during online model operation. When data distribution drifts, frequent full parameter recalculation or model retraining not only significantly increases computational resource consumption but also causes system response delays, failing to meet the engineering requirements of real-time performance and stability for online risk control systems. Summary of the Invention

[0007] The purpose of this invention is to address the shortcomings of existing technologies by providing a risk control model construction method and system based on feature normalization enhancement. This method significantly improves feature discrimination and model performance by establishing a dynamic adaptation mechanism, monitoring data distribution changes in real time, and employing a multi-granularity normalization strategy. This invention solves the problem of dynamically monitoring and adaptively adjusting the data distribution changes of model input features under limited computing resources and real-time response constraints during the online operation of a computer-based risk control system, thereby avoiding model input imbalance, system instability, or model performance degradation caused by feature distribution drift.

[0008] To address the aforementioned technical problems, this invention proposes a risk control model construction method based on feature normalization enhancement, executed in a computer-based risk control system. This method introduces a dynamic adaptation mechanism during model execution, maintains the statistical parameters of input features online, and adaptively adjusts the feature normalization strategy based on feature distribution offset detection results. This ensures that the model input remains within a stable numerical range, thereby improving the stability and reliability of the risk control model in an online operating environment.

[0009] The risk control model construction method based on feature normalization enhancement proposed in this invention includes: acquiring multi-source heterogeneous data from multiple platforms; cleaning, format conversion, and quality verification of the multi-source heterogeneous data to obtain structured data; performing feature extraction, feature selection, and high-order feature engineering on the structured data, wherein the high-order feature engineering includes constructing combined features, cross features, and domain features; fusing multiple extracted features to obtain transformed and fused feature data; configuring a normalization strategy and setting differentiated sliding window parameters for each type of feature data based on the transformed and fused feature data; calculating the mean, standard deviation, and quantile values ​​within the differentiated sliding window to establish a dynamic adaptation mechanism; and based on the dynamic adaptation mechanism, using the PSI index or KL divergence detection... The distribution offset of the feature data is measured, and the feature data is updated according to the preset hierarchical parameter update strategy based on the distribution offset to obtain updated dynamic feature parameters. The updated dynamic feature parameters are classified into enumerated features and numerical features. Label value mapping is performed on the enumerated features, and extreme value quantile mapping, interval mapping, and standard deviation correction are performed on the numerical features to obtain multi-granularity normalized features. Based on the multi-granularity normalized features, KS value or IV value is calculated in real time to monitor feature discrimination. When the discrimination decreases, sliding window parameter adjustment or normalization granularity switching is triggered to obtain high discrimination features with feedback. The high discrimination features are used to train and evaluate the performance of the risk control model to obtain an enhanced risk control model.

[0010] Furthermore, multi-source heterogeneous data from multiple platforms is acquired, and this data is cleaned, format-converted, and quality-verified to obtain structured data. This includes: formulating a data collection strategy based on the data dimensions required by the risk control model; generating a list of data sources for different platforms based on the data collection strategy; constructing API call interfaces, database connection interfaces, and log collection interfaces based on the data source list; collecting transaction behavior data, user attribute data, device behavior data, and external cooperation data from different platforms through these API call interfaces, database connection interfaces, and log collection interfaces; and sending the collected data to a data processing platform for further processing. The process involves storing the raw, multi-source heterogeneous data; performing structured analysis on the raw, multi-source heterogeneous data to identify data types, formats, and distribution characteristics; establishing a data dictionary based on the data types, formats, and distribution characteristics to obtain data structure information; performing missing value imputation, outlier correction, and duplicate data removal operations on the data structure information to obtain cleaned data; performing time format unification, enumeration value encoding, and text data structuring on the cleaned data to obtain format-standardized data; and calculating completeness, accuracy, and consistency indicators on the format-standardized data. If all indicators reach the expected threshold values, then the structured data is obtained.

[0011] Furthermore, feature extraction, feature selection, and high-order feature engineering are performed on the structured data. The high-order feature engineering includes constructing combined features, cross features, and domain features. Multiple extracted features are fused to obtain transformed and fused feature data, including: constructing statistical features such as mean, variance, maximum, minimum, and quantiles based on the structured data; constructing time-series features such as rate of change, year-on-year growth, and sliding window statistics; and constructing relational features such as network centrality and the number of associated accounts to obtain a basic feature set; performing arithmetic operations, ratio calculation, and cross feature construction based on the basic feature set to obtain a high-order feature set; performing cross-source association fusion on the high-order feature set to obtain a multi-dimensional feature matrix; calculating the correlation matrix index between features on the multi-dimensional feature matrix, and removing features with correlation matrix indices higher than a preset correlation threshold to obtain the transformed and fused feature data.

[0012] Furthermore, based on the transformed and fused feature data, a normalization strategy is configured for each type of feature data, and a differential sliding window parameter is set. The mean, standard deviation, and quantile values ​​within the differential sliding window are calculated, and a dynamic adaptation mechanism is established. This includes: classifying the transformed and fused feature data into continuous numerical features, sparse distribution features, skewed distribution features, and enumerated features; configuring a Z-score normalization strategy or a Min-Max normalization strategy for the continuous numerical features; configuring a robust quantile normalization strategy for the sparse distribution features; configuring a logarithmic transformation normalization strategy for the skewed distribution features; and configuring a label mapping strategy for the enumerated features, thus obtaining the normalization strategy configuration; based on the update frequency of the feature data, the feature data is reclassified into high-frequency changing feature data, periodic feature data, and stable feature data, with high-frequency features being classified as high-frequency changing feature data. A 7-day sliding window is set for variable feature data, a 30-day sliding window is set for periodic feature data, and a 90-day sliding window is set for stable feature data to obtain differential sliding window parameters. The mean, standard deviation, maximum, minimum, 25th percentile, 50th percentile, 75th percentile, and 95th percentile values ​​are calculated for the feature data. Based on these values ​​and the differential sliding window parameters, initial statistics are calculated. Based on these initial statistics, discrimination indices (KS, IV, and AUC values) are constructed for the feature data; stability indices (PSI and KL divergence) are constructed for the feature data; and outlier and missing rate indices are constructed for the feature data. A dynamic adaptation mechanism is established based on these discrimination, stability, and outlier detection indices.

[0013] Furthermore, based on the dynamic adaptation mechanism, the distribution shift of feature data is detected by the PSI index or KL divergence. A preset hierarchical parameter update strategy is then adopted to update the feature data according to the distribution shift, resulting in updated dynamic feature parameters. This includes: when new feature data flows in, updating the dataset within the sliding window according to the differentiated sliding window parameters while maintaining a constant sliding window size, resulting in updated window data; calculating the PSI index or KL divergence based on the updated window data to obtain a distribution shift detection index; when the PSI index is greater than 0.1, recalculating the mean, standard deviation, and quantile statistics based on the updated window data; when the PSI index is greater than 0.05 and less than or equal to 0.1, using an exponential moving average method to calculate the quantile statistics of the updated window data; when the PSI index is less than or equal to 0.05, keeping the existing window data unchanged, resulting in hierarchically updated parameters; verifying the hierarchically updated parameters by calculating the change in the discriminant properties of the feature data before and after the update; when the discriminant decreases beyond a preset discriminant threshold, using the feature data before the hierarchical update as the updated dynamic feature parameters.

[0014] Further, the updated dynamic feature parameters are classified into enumerated features and numerical features. Label value mapping is performed on the enumerated features, and extreme value quantile mapping, interval mapping, and standard deviation correction are performed on the numerical features to obtain multi-granularity normalized features. This includes: classifying the updated dynamic feature parameters into enumerated features and numerical features; establishing a business semantic mapping dictionary for the enumerated features, mapping enumerated values ​​to risk levels between 0 and 1; mapping high-cardinality enumerated features to numerical values ​​associated with the target variable using a target encoding method; and mapping time-based enumerated features using a periodic encoding method to obtain label-mapped enumerated features; based on the uniformity of the distribution of numerical features, the numerical features are classified into uniformly distributed numerical features and unevenly distributed numerical features; extreme values ​​and key quantiles are calculated for the uniformly distributed numerical features; and based on... The extreme values ​​and key quantiles are normalized and mapped using anchor quantiles to obtain standard normalized numerical features. For the non-uniform numerical features, their feature range is divided into multiple intervals. Each interval is mapped separately based on its boundary value and anchor quantile to obtain interval-mapped features. The standard deviation values ​​of enumerated features and numerical features are detected through a monitoring window. When the standard deviation of the features is detected to be greater than a preset standard deviation threshold, the feature mean of the monitoring window and the global feature mean are calculated. The enumerated features and numerical features are corrected based on the ratio of the feature mean within the monitoring window to the global feature mean to obtain standard deviation-corrected features. The enumerated features after label mapping, the standard normalized numerical features, the interval-mapped features, and the standard deviation-corrected features are fused to obtain the multi-granularity normalized features.

[0015] Further, based on the basic feature set, arithmetic operations, ratio calculations, and cross-feature construction are performed to obtain a higher-order feature set, including: constructing semantic representations for each feature data in the basic feature set to obtain semantic representation features; based on the semantic representation features, constructing the correlation relationships between features using graph structures, calculating the correlation matrix and conditional dependency probabilities between features to obtain feature association information; constructing high-dimensional dense context-enhanced features based on the feature association information and business scenario context information; compressing and selecting features from the context-enhanced features using LASSO and Elastic Net sparse coding methods to obtain sparse-coded features; performing nonlinear transformations on the sparse-coded features and introducing group sparsity and structural sparsity constraints to obtain a sparsified enhanced feature set; and performing arithmetic operations, ratio calculations, and cross-feature construction on the sparsified enhanced feature set to obtain the higher-order feature set.

[0016] Furthermore, the context enhancement features are compressed and selected using LASSO and Elastic Net sparse coding methods to obtain sparse coded features. This includes: applying LASSO regularization and Elastic Net regularization to the context enhancement features to convert the high-dimensional dense context enhancement features into a low-dimensional sparse representation, thus obtaining the sparsely coded transformed features; and performing feature dimensionality reduction and selection on the sparsely coded transformed features to retain the features with the most information, thus obtaining the sparse coded features.

[0017] Furthermore, the mean, standard deviation, maximum and minimum values, 25th quantile, 50th quantile, 75th quantile, and 95th quantile values ​​are calculated for the feature data. Based on these values ​​and the differential sliding window parameters, initial statistics are calculated, including: treating each feature data point as a dimension in a continuous space within the feature space, defining the feature value range as a continuous interval, thus obtaining a definition of the continuous feature space; and based on this definition, dividing the continuous space into multiple subdomains, each subdomain having boundaries defined by 5%, 25%, and 50%. The 75% and 95% quantiles are determined to obtain the domain discretization results. For each subdomain, a first-order linear basis function, a second-order polynomial basis function, or an exponential-logarithmic basis function is selected as the interpolation basis function to obtain a basis function selection scheme. Based on the basis function selection scheme, initialization parameters are calculated at the quantiles of each subdomain to obtain node parameter initialization results. Based on the node parameter initialization results, the local equations of each subdomain are assembled into a global equation, and the global equation is solved to obtain the coefficient values ​​at each quantile, thus obtaining the continuum model coefficients. Based on the continuum model coefficients and the differential sliding window parameters, initialization statistics are calculated using the continuum embedding mapping method.

[0018] Further, based on the node parameter initialization results, the local equations of each subdomain are assembled into a global equation, and the global equation is solved to obtain the coefficient values ​​at each quantile point, thus obtaining the continuum model coefficients. This includes: for each feature data to be processed, determining its subdomain in the continuous space, converting the feature data into the local coordinate system of the corresponding subdomain to obtain local coordinates; calculating the shape function value of the local coordinates using a preset interpolation basis function, calculating the mapping value of the feature data in the continuous space based on the shape function value and the quantile parameters to obtain the continuum embedding mapping value; standardizing the continuum embedding mapping value to the target interval between 0 and 1 to obtain the processed mapping value; calculating the mapping error in each subdomain, performing mesh refinement on subdomains with errors exceeding a threshold, recalculating the mapping value for the refined subdomains to obtain high-precision feature mapping values; and obtaining the continuum model coefficients based on the continuum embedding mapping value and the high-precision feature mapping value.

[0019] Further, when the PSI index is greater than 0.1, the mean, standard deviation, and quantile statistics are recalculated based on the updated window data, including: converting the updated window data into a sparse matrix representation and recording the positions and values ​​of non-zero elements to obtain a sparse data representation; calculating the covariance values ​​between feature data based on the sparse data representation, setting covariance values ​​less than a preset threshold to zero to obtain a sparse matrix representation of feature covariance; processing the sparse matrix representation of feature covariance, including using nested dissipative sorting or approximate minimum degree sorting to reduce padding, merging columns with the same non-zero pattern to form a supernode structure, and performing symbolic analysis to predict padding positions to obtain a processed sparse matrix; performing hybrid parallel sparse LU decomposition on the processed sparse matrix, decomposing the LU decomposition task into mutually independent subtasks, with multiple subtasks executed in parallel on multiple processors to obtain the decomposed L matrix and U matrix; and recalculating the mean, standard deviation, and quantile statistics based on the L matrix and the U matrix.

[0020] Further, a hybrid parallel sparse LU decomposition is performed on the processed sparse matrix, decomposing the LU decomposition task into independent subtasks. These subtasks are executed in parallel on multiple processors to obtain the decomposed L and U matrices. This includes: using a preset task-level parallel strategy to decompose the LU decomposition task into multiple independent subtasks, executing these subtasks in parallel on a multi-core CPU to obtain task-level parallel processing results; using a preset data-level parallel strategy to apply SIMD (Single Instruction Multiple Data) instruction set to the large, dense subtasks decomposed from the LU decomposition task to obtain data-level parallel processing results; using a preset heterogeneous computing acceleration strategy to transfer computationally intensive subtasks to the GPU for processing, while leaving communication-intensive subtasks on the CPU for processing, to obtain heterogeneous computing accelerated processing results; and based on the task-level parallel processing results, the data-level parallel processing results, and the heterogeneous computing accelerated processing results, a dedicated strip matrix solver is used for strip structures, a forward-backward replacement optimization algorithm is used for arrow-shaped structures, and a multiple frontier algorithm is used for general sparse structures to obtain the decomposed L and U matrices.

[0021] Furthermore, based on the multi-granularity normalized features, the KS value or IV value is calculated in real time to monitor feature discrimination. When the discrimination decreases, the sliding window parameter is adjusted or the normalization granularity is switched to obtain high-discrimination features with feedback.

[0022] Furthermore, the high-discrimination features are used to train and evaluate the risk control model, resulting in an enhanced risk control model.

[0023] Furthermore, the method also includes: deploying the enhanced risk control model in the risk control system so that risk identification, risk disclosure and risk control operations are performed on the object to be processed through the enhanced risk control model.

[0024] Furthermore, embodiments of this application also provide a method for constructing a risk control model based on feature normalization enhancement, including: Obtain multi-source heterogeneous data, and perform cleaning, format conversion, and quality verification on the multi-source heterogeneous data to obtain structured data; The structured data is subjected to feature extraction, feature selection, and high-order feature engineering. Multiple extracted features are then fused to obtain transformed and fused feature data. Based on the transformed and fused feature data, normalization strategies are configured for different types of feature data during model operation, and differential sliding window parameters are set. Statistics within the sliding window are calculated, and a dynamic adaptation mechanism for controlling the distribution of model input features is established. Based on the dynamic adaptation mechanism, the degree of shift in the distribution of feature data is detected during model operation, and the feature parameters are updated hierarchically according to the degree of distribution shift to obtain the updated dynamic feature parameters. Multi-granularity feature normalization is performed on the updated dynamic feature parameters to keep the model input features within a preset numerical range. Based on the multi-granularity normalized features, the risk control model is trained and its performance is evaluated to obtain an enhanced risk control model. The enhanced risk control model is deployed in the risk control system to perform risk identification, risk disclosure, and risk control operations on the objects to be processed.

[0025] Furthermore, high-order feature engineering is performed on the structured data, including: Statistical features, time-series features, and relational features are constructed for the structured data; Based on the statistical features, time-series features and relational features, four arithmetic operations are combined, ratios are calculated and cross features are constructed to generate high-order features for characterizing risk behavior patterns. The higher-order features are correlated and fused across data sources to form a multi-dimensional feature matrix; Correlation analysis is performed on the multi-dimensional feature matrix to remove redundant features with correlations higher than a preset threshold, resulting in transformed and fused feature data for training the risk control model.

[0026] Furthermore, feature compression and selection are performed on the higher-order features, including: A feature selection method with regularization constraints is applied to the high-order features to compress the high-dimensional features and obtain a low-dimensional sparse feature representation. Based on the low-dimensional sparse feature representation, features that contribute more than a preset threshold to risk identification are retained as input features of the risk control model.

[0027] Furthermore, establishing a dynamic adaptation mechanism also includes: Within the differentiated sliding window, the mean, standard deviation, and multiple quantile statistics are calculated for the feature data. Based on the statistics, the feature range is segmented and modeled to generate mapping parameters for feature normalization. The original features are mapped to a preset standard numerical range using the mapping parameters to improve the stability of the model input features during system operation.

[0028] The technical solution of the present invention also includes a risk control model construction system based on feature normalization enhancement, comprising a data preprocessing module, a feature fusion module, a dynamic adaptation module, a parameter update module, a feature normalization module, a feedback optimization module, and a model construction module, each module implementing the corresponding steps of the above method.

[0029] Compared with the prior art, the present invention has the following technical effects: Through the above technical solution, the present invention achieves the following technical effects during the operation of a computer risk control system: (1) By using a differentiated sliding window and a hierarchical parameter update mechanism, the recalculation of all historical data is avoided every time the data distribution changes, which effectively reduces the system's computational load and memory usage; (2) By using multi-granularity feature normalization, the input features of the model are restricted to a preset numerical range, which reduces the problem of unstable model inference caused by abnormal amplification of input values. (3) By using a feature distribution monitoring mechanism based on PSI or KL divergence, real-time monitoring of the model input state is achieved, which improves the stability and robustness of the risk control system in long-term online operation scenarios; (4) While ensuring the predictive performance of the model, the frequency of model retraining was reduced, thereby improving the overall operating efficiency and availability of the risk control system. Attached Figure Description

[0030] Figure 1 This is a flowchart of the risk control model construction method based on feature normalization enhancement according to Embodiment 1 of the present invention; Figure 2 This is a structural diagram of the risk control model construction system based on feature normalization enhancement in Embodiment 2 of the present invention. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0032] Example 1 This embodiment presents a risk control model construction method based on feature normalization enhancement, which includes seven main steps: acquiring multi-source heterogeneous data, feature extraction and fusion, establishing a dynamic adaptation mechanism, distribution offset detection and parameter updating, multi-granularity feature normalization, feature feedback optimization, and model training evaluation. This method significantly improves the feature discrimination and stability of the risk control model.

[0033] like Figure 1 As shown in the figure, a risk control model construction method based on feature normalization enhancement in this embodiment specifically includes the following steps: S1: Acquire multi-source heterogeneous data from multiple platforms, clean, convert formats, and verify the quality of the multi-source heterogeneous data to obtain structured data; In the multi-source heterogeneous data acquisition phase, the system formulated a comprehensive data collection strategy based on risk control requirements. Multi-source heterogeneous data refers to data sets from different platforms with different structures and formats, including structured data (such as database tables), semi-structured data (such as JSON and XML), and unstructured data (such as text logs). Based on the dimensions required by the risk control model, such as user behavior, transaction characteristics, and relationship networks, the system formulated a data collection strategy, clarifying data sources, collection frequency, and priorities, and generating a list of data sources from different platforms. Based on this list, three types of core interfaces were constructed: an API call interface for accessing standardized data services provided by external systems; a database connection interface for directly connecting to database systems via protocols such as JDBC and ODBC; and a log collection interface for capturing system operation logs through a log collector. These interfaces worked together to collect four types of key data: transaction behavior data recording users' transaction activity trajectories; user attribute data containing users' basic information and historical characteristics; device behavior data capturing user operation characteristics; and external cooperation data providing industry risk indicators. The collected raw data was stored through a distributed storage system, forming a raw multi-source heterogeneous data pool.

[0034] After data acquisition, the system performs rigorous data preprocessing. First, structured analysis is conducted to identify data types (numerical, enumerable, time-based, etc.), format characteristics, and distribution characteristics (such as normal distribution and long-tailed distribution), establishing a data dictionary to clarify field meanings and value ranges. Then, based on the identified data structure information, three cleaning operations are performed: missing values ​​are filled using various strategies, such as mean / median filling, leading / tailed value filling, and model prediction filling; outliers are identified and corrected using box plots or Z-score methods; and duplicate data is deduplicated based on unique identifiers or content features. After cleaning, the system performs format standardization, including unifying time fields of different formats to a standard timestamp format, converting categorical variables to numerical codes using a mapping dictionary, and converting unstructured text into structured features through word segmentation and vectorization. Finally, the system calculates data quality indicators, including completeness (proportion of non-null values), accuracy (proportion of data conforming to business rules), and consistency (degree of consistency across source data). Only when these indicators reach preset thresholds are the data considered qualified structured data.

[0035] S2: Perform feature extraction, feature selection, and higher-order feature engineering on the structured data. The higher-order feature engineering includes the construction of combined features, cross features, and domain features. Multiple extracted features are fused to obtain transformed and fused feature data. Based on the preprocessed structured data, the system performed multi-level feature engineering. Feature engineering refers to the process of extracting, transforming, and constructing features from raw data, with the aim of creating a feature set that can effectively represent risk patterns. First, three types of basic features were constructed: statistical features, which capture the distribution characteristics of the data by calculating statistics such as mean, variance, maximum / minimum, and quantiles; time-series features, which capture the temporal evolution characteristics of the data by calculating statistics such as rate of change, chain growth, and sliding window; and relational features, which capture the network relationship characteristics between entities by calculating network centrality, number of associated accounts, etc. Building upon these basic features, the system constructed higher-order features: combining multiple features through arithmetic operations (addition, subtraction, multiplication, and division), such as "transfer amount / historical average transaction amount"; constructing normalized features through ratio calculations, such as "number of transactions today / average number of transactions in the past 30 days"; and capturing the interaction relationships between different dimensions through cross-features, such as the combined feature of "device type × transaction time period". These higher-order features were then fused through cross-source association, using unique identifiers such as user ID and device ID to link features from different data sources, forming a multi-dimensional feature matrix. Finally, by calculating the correlation matrix between features, highly correlated redundant features (correlation coefficients exceeding a preset threshold, such as 0.85) are removed, resulting in information-rich and relatively independent transformed and fused feature data.

[0036] S3: Based on the transformed and fused feature data, configure a normalization strategy for each type of feature data and set differential sliding window parameters, calculate the mean, standard deviation and quantile values ​​within the differential sliding window, and establish a dynamic adaptation mechanism. For the fused feature data, the system establishes a dynamic adaptation mechanism to enable feature normalization to adapt to changes in data distribution. The dynamic adaptation mechanism is a system that automatically adjusts normalization parameters based on data distribution characteristics. First, the system categorizes features into four types based on data characteristics: continuous numerical features (such as age and amount) are configured with Z-score normalization (calculated using (x-μ) / σ) or Min-Max normalization (calculated using (x-min) / (max-min)); sparsely distributed features (such as count features with most values ​​being 0) are configured with robust quantile normalization (mapping based on quantiles rather than extreme values); skewed distributed features (such as long-tailed monetary features) are configured with logarithmic transformation normalization (first taking the logarithm and then normalizing); and enumerated features (such as gender and city) are configured with a label mapping strategy (mapping categories to numerical values). Simultaneously, the system sets differentiated sliding windows based on feature update frequency: high-frequency changing features (such as recent login counts) are set with a 7-day window, periodic features (such as weekly transaction patterns) are set with a 30-day window, and stable features (such as user registration information) are set with a 90-day window. Based on these windows, the system calculates key statistics: mean, standard deviation, and various quantile values, as the basic parameters for normalization. Finally, the system constructs a monitoring indicator system, including discrimination indicators such as KS value (Kolmogorov-Smirnov statistic, measuring the ability of a feature to distinguish positive and negative samples) and IV value (Information Value, measuring the predictive ability of a feature); stability indicators such as PSI (Population Stability Index, measuring the stability of feature distribution) and KL divergence (measuring the difference between two probability distributions); and anomaly detection indicators such as outlier rate and missing rate, forming a complete dynamic adaptation mechanism.

[0037] S4: Based on the dynamic adaptation mechanism, the feature data distribution offset is detected based on the PSI index or KL divergence, and the feature data is updated according to the preset hierarchical parameter update strategy based on the distribution offset to obtain the updated dynamic feature parameters. When new data flows in, the system implements distribution shift detection and parameter updates based on a dynamic adaptation mechanism. Distribution shift refers to the phenomenon that feature distribution changes over time, commonly seen in the field of financial risk control. First, the system updates the dataset within the window according to the differentiated sliding window parameters, keeping the window size constant (new data enters, old data is removed). Then, it calculates the PSI indicator or KL divergence to assess the degree of change in feature distribution. The PSI indicator measures the difference between two distributions, and its calculation formula is PSI = Σ(actual percentage - expected percentage) × ln(actual percentage / expected percentage), where the actual percentage refers to the percentage of each interval in the new data, and the expected percentage refers to the percentage of each interval in the historical data.

[0038] When the PSI (Power Score) is greater than 0.1, it indicates a significant change in the feature distribution, possibly caused by factors such as changes in the market environment, shifts in user behavior, or adjustments to business rules. In this case, the system needs to recalculate the mean, standard deviation, and quantile statistics based on the updated window data to ensure that feature normalization and model input accurately reflect the current data distribution. The mean is a fundamental measure of central tendency in a dataset, calculated as the arithmetic mean of all sample values; the standard deviation measures the dispersion of the data, and is the square root of the arithmetic mean of the squared deviations of the samples from the mean; quantile statistics describe the shape of the data distribution and typically include multiple key percentiles (such as 10%, 25%, 50%, 75%, 90%, etc.). Recalculating these statistics is a computationally intensive task, especially for large-scale datasets. The system employs multiple optimization techniques to improve computational efficiency: online algorithms (such as the Welford algorithm) are used to simultaneously calculate the mean and standard deviation in a single scan of the data; selection algorithms (such as QuickSelect) are used to efficiently calculate quantiles, avoiding complete sorting; partition-based parallel computation is implemented, processing data in blocks and then merging the results; approximation algorithms (such as t-digest or KLL sketches) are used to process ultra-large-scale data, significantly reducing memory requirements and computation time while maintaining accuracy. For scenarios with high update frequencies, the system also implements a distributed computing framework, utilizing multiple machines to process data shards in parallel, and then efficiently merging the results through mergeable data structures (such as addition sketches). Through these optimizations, the system can complete the recalculation of statistics for large-scale window data within a reasonable time, ensuring that the model can adapt to changes in distribution in a timely manner.

[0039] When the PSI index is greater than 0.05 and less than or equal to 0.1, the data distribution has changed to some extent, but not to the point of requiring a complete recalculation. In this case, the system uses the exponential moving average method to calculate the quantile statistics of the updated window data, achieving a gradual update of the statistical parameters. Exponential Moving Average (EMA) is a time series smoothing technique that gives higher weight to recent data, while the weight of older data decays exponentially over time. The core formula of EMA is: EMA_t = α × Current_t + (1-α) × EMA_{t-1}, where α is the smoothing factor, typically set between 0.05 and 0.3; a larger α value makes the impact of new data more significant. In quantile updates, the system extends the traditional EMA method to implement exponentially weighted updates of quantiles. Specifically, the system maintains a special data structure (such as a t-digest or quantile sketch) that supports incremental updates and merging operations. When new window data arrives, the system first calculates a provisional quantile estimate based on the new data, and then merges it with the historical quantile structure according to exponential weights: NewQuantile = MergeWithWeight(CurrentQuantile, HistoricalQuantile, α). This method avoids storing and repeatedly processing all historical data, maintaining computational efficiency while ensuring that the quantile estimate can smoothly adapt to changes in distribution. To further improve accuracy, the system also implements an adaptive smoothing factor, dynamically adjusting the value of α according to the PSI index magnitude. This means that the more significant the change in distribution, the greater the impact of new data. This gradual update strategy is particularly suitable for features that exhibit slow-trending changes, as it can track the evolution of the distribution while avoiding overreaction to short-term fluctuations.

[0040] When the PSI index is less than or equal to 0.05, it indicates that the feature distribution is very stable, and the difference between the current window and the baseline window is minimal. In this case, the system chooses to keep the existing window data unchanged and does not update any parameters. This strategy is based on the principle of "don't fix it if it's not broken," avoiding unnecessary consumption of computational resources and preventing the model from frequently adjusting due to small random fluctuations, thus improving the overall system stability. Although no parameter updates are performed, the system continues to monitor the trend of PSI value changes to detect signs of accelerated distribution changes in a timely manner. The system implements several optimizations to improve monitoring efficiency: incremental PSI calculation is used, processing only newly arrived data; a stratified sampling strategy is implemented, where PSI is quickly estimated based on samples for large-scale datasets, and full calculation is only performed when the estimated value is close to the threshold; a dynamic monitoring frequency is set, automatically adjusting the inspection interval according to historical PSI volatility, with more frequent inspections for features with large fluctuations; and an asynchronous computing framework is used, where PSI monitoring is executed as a background task without affecting the performance of the main system. Furthermore, the system implements an early warning mechanism. When the PSI values ​​of multiple consecutive windows are all less than 0.05 but show a stable upward trend, an early warning of potential distribution changes is issued, reminding risk control analysts to pay attention to relevant characteristics. This strategy not only saves computational resources but also maintains the stability of model parameters, avoiding the risk of overfitting the current window data. Through this hierarchical update strategy, the system obtains the parameters after hierarchical updates.

[0041] S5: Classify the updated dynamic feature parameters into enumerated features and numerical features, perform label value mapping on the enumerated features, and perform extreme value quantile mapping, interval mapping and standard deviation correction on the numerical features to obtain multi-granularity normalized features. Based on the updated dynamic feature parameters, the system performs multi-granularity normalization. Multi-granularity normalization refers to using different normalization methods with different granularities and strategies for data of different types and distribution characteristics to maximize the preservation of discriminative information in the data. First, features are divided into two main categories: enumerable and numerical. For enumerable features, the system establishes a business semantic mapping dictionary, mapping enumerable values ​​to risk levels in the 0-1 range, such as "credit card" mapping to 0.3 and "cash" mapping to 0.7. For high-cardinality enumerable features (such as cities), a target encoding method is used, mapping each category to a statistic related to the target variable, such as "Beijing" mapping to the average default rate of Beijing users. For time-based enumerable features (such as days of the week), periodic encoding is used to preserve periodic characteristics, such as using sin and cos functions for encoding. For numerical features, the system further classifies them based on their distribution uniformity: for uniformly distributed features, extreme values ​​and key quantiles are calculated, and the new value is mapped using the formula "new value = (original value - minimum value) / (maximum value - minimum value) × (target interval size) + target interval minimum value"; for unevenly distributed features, the value range is divided into multiple intervals (e.g., based on quantiles), and the mapping formula is applied separately to each interval to ensure that even small differences within small intervals can be effectively distinguished. Furthermore, the system monitors the feature standard deviation through a monitoring window. When the standard deviation is abnormal, the feature is adjusted using a ratio correction method to ensure the rationality of the feature distribution. Finally, the system merges all normalized features to form a multi-granularity normalized feature set.

[0042] S6: Based on the multi-granularity normalized features, calculate the KS value or IV value in real time to monitor the feature discrimination. When the discrimination decreases, trigger the adjustment of sliding window parameters or the switching of normalization granularity to obtain high discrimination features after feedback. Based on the features normalized at multiple granularities, a feature feedback optimization mechanism is implemented. Feature feedback optimization is a mechanism that continuously monitors feature quality and dynamically adjusts processing strategies. The system calculates the KS value (measures the ability of a feature to distinguish between good and bad samples, with a value of 0-1) or IV value (measures the predictive ability of a feature for the target variable, generally considered to have significant predictive power >0.3) of the features in real time, monitoring changes in feature discriminative power. When a decrease in discriminative power is detected (e.g., a decrease in KS value of more than 10%), the system automatically triggers optimization strategies: adjusting sliding window parameters, such as reducing the window for highly variable features to improve sensitivity and expanding the window for unstable features to enhance stability; switching the normalization granularity, such as switching from global normalization to interval normalization, or from linear mapping to nonlinear mapping; in extreme cases, the system may even activate backup statistics to ensure the stability of feature normalization. Through this closed-loop feedback mechanism, the system can continuously optimize feature quality, cope with various data distribution changes, and ultimately output highly discriminative features.

[0043] S7: Use the high-discrimination features to train and evaluate the performance of the risk control model to obtain an enhanced risk control model.

[0044] Finally, the risk control model is trained using high-discrimination features optimized through feedback. The system first performs feature selection and combination, retaining features with high IV values. The optimal feature combination is determined through forward / backward feature selection, and domain knowledge is incorporated to ensure the integrity of the feature set. Then, a suitable algorithm (such as LR, GBDT, XGBoost, etc.) is selected for model training. The time series integrity is maintained when dividing the training / test sets, and the optimal hyperparameter combination is found through grid search. Cross-validation is used to verify the model's generalization ability. The system comprehensively evaluates model performance, including offline evaluation (AUC, KS, accuracy, recall, etc.), online A / B testing, stability testing, and comparison with benchmark models. Finally, the model is deployed to the production environment, enabling streaming processing, establishing monitoring and early warning systems, setting up automatic rollback mechanisms, and implementing regular retraining plans to ensure the long-term stable and efficient operation of the model.

[0045] Finally, the risk control model can be deployed in the risk control system and used to perform specific risk control operations, such as risk identification, risk disclosure, and risk control operations.

[0046] Through the detailed implementation steps described above, this risk control model construction method achieves high feature discriminativeness and stability, effectively addresses the data distribution drift problem, significantly improves the performance of the risk control model, increases the KS value by an average of more than 20%, and improves the model AUC by about 5-8 percentage points, providing strong technical support for risk control in the financial, e-commerce, and internet sectors.

[0047] In the process of building a risk control model, establishing a dynamic adaptation mechanism is a crucial step to ensure that feature processing can cope with changes in data distribution. Based on the transformed and fused feature data, the system first performs feature classification, configuring the most suitable normalization strategy for different types of features. Feature classification is the process of categorizing features according to their data characteristics and distribution patterns, with the aim of selecting the optimal processing method for each type of feature. The system divides features into four main categories: continuous numerical features, sparse distribution features, skewed distribution features, and enumerated features. Continuous numerical features refer to numerical features with continuous values ​​and a relatively uniform distribution, such as user age and account balance; sparse distribution features refer to features where most values ​​are zero or extremely small, with only a small portion taking non-zero values, such as the number of times a specific behavior is triggered; skewed distribution features refer to features with asymmetrical distributions and obvious skewness, such as transaction amounts which typically exhibit a right-skewed long-tail distribution; enumerated features refer to features with values ​​belonging to a finite number of discrete categories, such as gender, city, and device type.

[0048] For these four types of features, the system is configured with differentiated normalization strategies. For continuous numerical features, either Z-score or Min-Max normalization is configured. Z-score normalization (standardization) transforms the feature into a distribution with a mean of 0 and a standard deviation of 1, suitable for scenarios where data volatility needs to be considered; Min-Max normalization linearly maps the feature to the [0,1] interval, preserving the shape of the original distribution, suitable for scenarios where proportional relationships need to be maintained. For sparsely distributed features, a robust quantile normalization strategy is configured. This strategy maps based on the quantiles of the data rather than extreme values, effectively avoiding the distortion of normalization results by extreme values ​​and enhancing the expressive power of the feature in sparse regions. For skewedly distributed features, a logarithmic transformation normalization strategy is configured. The original data is first logarithmically transformed (e.g., log(1+x)) to compress the numerical range, and then the standard normalization method is applied, effectively handling the long-tail distribution problem. For enumerated features, a label mapping strategy is configured, mapping discrete categories to meaningful numerical representations, such as risk level coding and frequency coding. Through this differentiated configuration, the system selects the most suitable normalization method for each type of feature based on its data characteristics, forming a complete normalization strategy configuration.

[0049] Besides data characteristics, the temporal variation of features is also a crucial factor affecting normalization results. Based on the update frequency of feature data, the system reclassifies features into high-frequency changing feature data, periodic feature data, and stable feature data. High-frequency changing feature data refers to features that change significantly in the short term, such as recent login frequency or transaction amount in the last 7 days; periodic feature data refers to features with obvious periodic change patterns, such as weekday / weekend activity differences or beginning / end month consumption patterns; stable feature data refers to features that remain relatively stable over the long term, such as basic user attributes and long-term behavioral preferences. For these three types of features, the system sets differentiated sliding window parameters: a 7-day sliding window is set for high-frequency changing feature data, allowing the normalization parameters to quickly respond to recent changes; a 30-day sliding window is set for periodic feature data, covering the complete monthly cycle and capturing periodic patterns; and a 90-day sliding window is set for stable feature data, providing a more stable long-term statistical basis. The sliding window is a dynamic data processing mechanism. When new data enters, the window slides forward, maintaining a fixed size; as new data enters the window, the oldest data leaves. This differentiated sliding window setting allows the system to calculate normalization parameters using the most suitable data range based on the time-varying characteristics of the features, balancing sensitivity and stability.

[0050] Based on a defined sliding window, the system calculates key statistics to provide fundamental parameters for feature normalization. The system calculates the mean, standard deviation, maximum, minimum, and 25th, 50th, 75th, and 95th percentile values ​​for each feature within its corresponding sliding window. The mean reflects the central tendency of the feature; the standard deviation measures the dispersion of the feature; the maximum and minimum values ​​define the range of the feature's values; and the quantiles characterize the shape and central tendency of the feature distribution. These statistics together constitute a complete description of the feature distribution, providing a comprehensive reference for subsequent normalization processing. Especially for skewed distribution features, quantiles provide a more robust description of the distribution than the mean and standard deviation. The system combines these statistics with the sliding window parameters to calculate initialization statistics, i.e., the set of fundamental parameters for each feature within its respective sliding window, laying the data foundation for establishing a dynamic adaptation mechanism.

[0051] After initializing the statistics, the system constructs three key indicators to comprehensively monitor feature quality. First are the discrimination indicators, including the KS value, IV value, and AUC value. The KS value (Kolmogorov-Smirnov statistic) measures the feature's ability to distinguish between positive and negative samples. It is calculated by taking the maximum value of the cumulative distribution function difference between positive and negative samples, ranging from 0 to 1; a larger value indicates stronger discrimination ability. The IV value (Information Value) measures the feature's predictive ability for the target variable; generally, an IV > 0.3 indicates significant predictive power. The AUC value (Area Under Curve) reflects the feature's predictive performance as a univariate model, ranging from 0.5 to 1; a larger value indicates better predictive performance. Second are the stability indicators, including PSI and KL divergence. The Population Stability Index (PSI) measures the stability of a feature distribution over time, calculated by comparing the differences in feature distributions across different time windows. The Kullback-Leibler divergence (KL divergence) measures the difference between two probability distributions and is used to detect the degree of shift in the feature distribution. Finally, there are anomaly detection metrics, including outlier rate and missing rate. The outlier rate measures the proportion of outliers in a feature, typically defined based on statistical methods (such as the 3σ principle) or domain knowledge; the missing rate directly calculates the proportion of missing values ​​in a feature. Based on these three metrics, the system establishes a dynamic adaptation mechanism that can monitor feature quality in real time and trigger corresponding adjustment strategies based on metric changes, ensuring the stability and efficiency of the feature processing.

[0052] After the dynamic adaptation mechanism is established, the system can monitor changes in data distribution in real time and adopt corresponding parameter update strategies based on the degree of change. When new feature data flows in, the system first updates the dataset within the window according to the differentiated sliding window parameters. The sliding window update is a dynamic process: as new data enters the window, an equal amount of old data leaves the window, keeping the window size constant. For example, for high-frequency features with a 7-day sliding window, the system adds the latest day's data to the window each day while removing data older than 7 days; for periodic features with a 30-day window, the data from the most recent 30 days is retained for parameter calculation. This dynamic update mechanism ensures that the feature normalization parameters are always based on the most relevant time window data.

[0053] Based on the updated window data, the system calculates the PSI (Population Stability Index) or KL divergence to detect the degree of shift in the feature distribution. PSI is a widely used distribution stability indicator in risk control; it quantifies distribution change by comparing the differences in the proportions of two distributions across different bins. A higher PSI value indicates a more significant distribution shift. Generally, PSI < 0.1 is considered a stable distribution, 0.1 ≤ PSI < 0.25 indicates a slight change, and PSI ≥ 0.25 indicates a significant change. KL divergence (Kullback-Leibler divergence) is an indicator in information theory that measures the difference between two probability distributions and can also be used to detect distribution shifts. These two indicators quantify the degree of change in the feature distribution from different perspectives, providing a basis for the selection of subsequent parameter update strategies.

[0054] Based on the detected distribution shift, the system adopts a hierarchical parameter update strategy to ensure the effectiveness and stability of feature normalization. When the PSI index is greater than 0.1, it indicates a significant change in the feature distribution. The system recalculates the mean, standard deviation, and quantiles based on the updated window data, completely refreshing the normalization parameters to adapt to the new data distribution. When the PSI index is greater than 0.05 and less than or equal to 0.1, it indicates a slight change in the feature distribution. The system uses the exponential moving average method to calculate the quantile statistics of the updated window data. The exponential moving average is a smoothing method that gives higher weight to recent data. The calculation formula is: New parameter = α × Old parameter + (1-α) × New batch parameter, where α is the smoothing coefficient, usually set between 0.9 and 0.99. This method considers historical stability while responding appropriately to new changes, making it suitable for handling slight distribution shifts. When the PSI index is less than or equal to 0.05, it indicates that the feature distribution is basically stable. The system chooses to keep the existing window data and parameters unchanged to avoid unnecessary parameter fluctuations. Through this hierarchical update strategy, the system can take corresponding measures according to the severity of the distribution shift, which ensures both a sensitive response to significant changes and the stability of the overall parameters.

[0055] After parameter updates, the system undergoes rigorous verification to ensure that the updates do not reduce the discriminative power of the features. The system calculates the change in discriminative power of the feature data before and after the update, primarily focusing on changes in discriminative metrics such as KS and IV values. When the discriminative power decreases beyond a preset threshold (e.g., a decrease in KS value exceeding 10%), the system triggers a rollback mechanism, using the feature data before the hierarchical update as the updated dynamic feature parameters and abandoning the current update. This verification mechanism prevents parameter deterioration caused by data anomalies or temporary fluctuations, ensuring the quality and stability of the feature normalization process. Through a complete closed loop of distribution offset detection, hierarchical parameter updates, and verification rollback, the system achieves intelligent dynamic updates of feature parameters, effectively addressing the issue of data distribution changes.

[0056] Based on the updated dynamic feature parameters, the system performs multi-granularity normalization, transforming the original features into a more discriminative and stable form. First, the system classifies the updated dynamic feature parameters into enumerable features and numerical features, applying the most suitable normalization method to each type. Enumerable features are those with a finite number of discrete categories, such as gender, region, and product type; numerical features are those with continuous numerical values, such as age, amount, and frequency. This basic classification lays the foundation for subsequent differential processing.

[0057] For enumerated features, the system employs multiple mapping strategies to convert categories into meaningful numerical representations. First, the system establishes a business semantic mapping dictionary, mapping enumerated values ​​to risk levels between 0 and 1 based on business knowledge and historical data. For example, for payment method features, "credit card" might be mapped to 0.7 (higher risk), "debit card" to 0.3 (medium risk), and "Alipay" to 0.1 (low risk). For high-cardinality enumerated features (features with a wide variety of values, such as city or occupation), the system uses target encoding to map each category to statistics related to the target variable, such as the default rate or average risk score of that category. This method effectively addresses the sparsity problem of high-cardinality features while preserving the association between categories and the target variable. For time-based enumerated features (such as days of the week or months), the system uses periodic encoding for mapping, typically employing trigonometric function transformations (such as sin / cos transformations) to preserve the periodicity and continuity of time categories. Through these differentiated mapping strategies, the system converts enumerated features into numerical representations while preserving the business meaning and predictive power of the features to the greatest extent possible.

[0058] For numerical features, the system further classifies them into uniformly distributed and unevenly distributed numerical features based on their distribution uniformity, and applies different normalization methods. Distribution uniformity refers to whether the feature values ​​are distributed evenly across the entire range, usually determined through histogram analysis or statistical indicators (such as skewness and kurtosis). For relatively uniformly distributed numerical features, the system calculates extreme values ​​(maximum and minimum values) and key quantiles (such as the 25th, 50th, and 75th percentiles), and then performs normalization mapping based on these statistics using anchor quantiles. Anchor quantiles are reference points in the normalization process, used to control the shape of the distribution after normalization. For example, mapping the 75th percentile to 0.8 ensures that most data are distributed between 0 and 0.8, leaving a space of 0.8-1 for outliers, thus enhancing the ability to distinguish outliers. This quantile-based mapping method is more robust than traditional Min-Max normalization, less affected by extreme values, and maintains the relative ordering of the data.

[0059] For unevenly distributed numerical features, simple linear mapping often fails to preserve the discriminative information of the data. This system employs a more flexible interval mapping method. First, the system divides the feature value range into multiple intervals, typically based on quantiles (such as equal-frequency binning) or equal-interval binning. Then, a mapping formula is applied individually to each interval to achieve differentiated mapping between different intervals. For example, for long-tailed features, the system might apply a larger mapping ratio to low-value intervals and a smaller mapping ratio to high-value intervals. This maintains the overall ranking while enhancing the discriminative power of data in key intervals. Through this interval mapping, the system effectively addresses the normalization problem of unevenly distributed features, improving the expressive power and discriminative power of the features.

[0060] In addition to basic normalization, the system also implements standard deviation correction to further optimize feature distribution. The system continuously monitors the standard deviation of enumerated and numerical features through a monitoring window. When a feature's standard deviation exceeds a preset threshold (indicating an overly dispersed feature distribution), a correction mechanism is triggered. During correction, the system calculates the feature mean within the monitoring window and the global feature mean (historical long-term mean), then corrects the feature based on the ratio of these two means. The correction formula is: Corrected feature value = Original feature value × (Window mean / Global mean). This correction method effectively suppresses distribution anomalies caused by short-term fluctuations, making the feature distribution more stable and reasonable, while preserving the relative ranking relationship of features—a stabilization process that maintains discriminative power.

[0061] Finally, the system fuses all processed features to form the final multi-granularity normalized feature set. The system integrates enumerated features after label mapping, standard-normalized numerical features, features after interval mapping, and features corrected for standard deviation, ensuring that all features are mapped to an appropriate numerical range (usually 0-1) and maintain good discriminative power and stability. This multi-granularity normalization process employs the most suitable normalization strategy for different types and distributions of features, significantly improving the expressive power and predictive performance of the features, and providing high-quality feature input for subsequent risk control models.

[0062] Through a series of meticulous processes, including feature classification and normalization strategy configuration, differentiated sliding window settings, statistical calculation, monitoring indicator construction, distribution offset detection, hierarchical parameter updates, and multi-granularity feature normalization, the system establishes a complete dynamic feature adaptation and normalization enhancement mechanism. This effectively solves problems such as poor data drift adaptability and insufficient feature discriminative power in traditional feature engineering, significantly improving the performance and stability of the risk control model. Practice shows that this dynamic adaptation multi-granularity normalization method, compared to traditional static normalization, can improve the model's KS value by an average of 15-25%, significantly enhancing the risk control model's ability to identify fraud and credit risk.

[0063] Advanced feature engineering is a crucial step in risk control model building. By deeply processing and combining basic features, it can significantly improve the model's expressive power and predictive performance. The process of combining basic features through arithmetic operations, calculating ratios, and constructing cross-features involves multiple refined steps, from semantic representation construction to sparsity enhancement and finally to the formation of advanced feature sets, forming a complete advanced feature engineering system.

[0064] First, the system constructs semantic representations for each feature data in the basic feature set, resulting in semantic representation features. Semantic representation refers to converting features from their original numerical or categorical forms into vector representations that express their business meaning and inherent relationships. The key to this step is encoding the business semantic information of the features into the representation space, ensuring that features with similar semantics are close together, while features with significant semantic differences are far apart. For numerical features, the system uses methods such as binning, normalization, and semantic embedding for conversion; for categorical features, it uses methods such as label encoding, frequency encoding, or distribution encoding to convert them into numerical representations. In this process, the system also introduces domain knowledge enhancement, incorporating business experts' understanding of the features (such as their risk indication role or business importance) into the feature representations. For example, for the feature "user registration duration," in addition to its numerical value, the system considers its representation of user stability, mapping it to a specific semantic space. Through this semantic representation construction, the system ensures that each feature not only carries the original information but also contains rich business semantics, laying the foundation for subsequent feature association analysis.

[0065] Based on semantic representation features, the system uses a graph structure to construct the relationships between features, calculating the correlation matrix and conditional dependency probabilities to obtain feature association information. Graph structures are a powerful data representation method, composed of nodes and edges, naturally representing relationships between entities. Here, the system treats each feature as a node in the graph, and the relationships between features are represented by edges, with edge weights reflecting the strength of the relationship. Specifically, the system first calculates the correlation matrix between features, including multiple indicators such as Pearson correlation coefficient (suitable for linear relationships), Spearman correlation coefficient (suitable for monotonic relationships), and mutual information (suitable for nonlinear relationships) to comprehensively measure the statistical correlation between features. Then, the system further calculates the conditional dependency probability, i.e., the distribution of one feature given a value of another, which can capture more complex nonlinear dependencies. Based on these correlation and dependency indicators, the system constructs a weighted directed graph, where nodes represent features, edge directions represent dependencies, and edge weights represent the strength of the relationship. Through this graph structure representation, the system can intuitively display the complex network of relationships between features, providing guidance for subsequent feature combination and selection.

[0066] Based on feature association information and business scenario context information, the system constructs high-dimensional, dense context-enhanced features. Context-enhanced features refer to the result of expanding and enriching the original features after considering the business environment and related feature information. First, based on the previously constructed feature association graph, the system extracts the local subgraph structure of each feature, including directly connected neighbor features and second-order association features. Then, the system integrates business scenario context information, such as time period (e.g., weekday / weekend, daytime / nighttime), transaction type (e.g., large-amount consumption, cross-border transactions), and user group (e.g., high-risk users, new users), combining these contextual factors with the features and their neighboring information. During the integration process, the system employs an attention mechanism, dynamically adjusting the weights of different contextual factors based on the relevance of the feature to the context. For example, for the "login frequency" feature, a higher contextual weight may be assigned during nighttime hours, as abnormal logins at night usually have a higher risk indication effect. Finally, the system generates a high-dimensional, dense representation containing the original features, neighboring features, and contextual information. This representation can comprehensively capture the behavioral patterns of features in specific business scenarios, greatly enhancing the expressive power of the features. However, this high-dimensional dense representation also brings the curse of dimensionality and the risk of overfitting, requiring further compression and optimization.

[0067] To address the challenges posed by high-dimensional dense representations, the system employs LASSO and Elastic Net sparse coding methods to compress and select features from context-enhanced features, resulting in sparsely encoded features. LASSO (Least Absolute Shrinkage and Selection Operator) is a linear regression method with L1 regularization, capable of feature selection and coefficient compression. Its characteristic is a tendency to produce sparse solutions, where most feature coefficients are compressed to zero, retaining only a few important features. Elastic Net combines L1 and L2 regularization, possessing both the sparsity and feature selection capabilities of LASSO, and the ability to handle related feature groups like Ridge regression, making it particularly suitable for scenarios where the number of features far exceeds the number of samples. The system first applies LASSO regularization to the context-enhanced features, defining it as an optimization problem where the loss function is the sum of squared prediction errors, and the regularization term is the L1 norm of the feature coefficients (the sum of the absolute values ​​of the coefficients). By controlling the regularization strength, the system achieves a balance between model complexity and fitting ability, compressing most unimportant feature coefficients to zero. Next, the system applies Elastic Net regularization, introducing an L2 regularization term (sum of squared coefficients), which helps retain more information in relevant feature groups and avoids the excessive sparsity problem that LASSO may have. Through the combined application of LASSO and Elastic Net, the system transforms high-dimensional dense context-enhanced features into low-dimensional sparse representations, obtaining features after sparse coding transformation.

[0068] After sparse coding transformation, the system further performs feature dimensionality reduction and selection, retaining the features with the highest information content to obtain the final sparse coded features. In this step, the system employs multiple dimensionality reduction techniques, including Principal Component Analysis (PCA), Independent Component Analysis (ICA), and t-SNE, to map features to a lower-dimensional space while preserving the maximum data variance and structural information. PCA is a commonly used linear dimensionality reduction method that reduces dimensionality by finding the directions with the highest data variance (principal components) and projecting the original features onto these principal components. Independent Component Analysis focuses on finding mutually independent features, suitable for processing mutually independent signal sources. t-SNE (t-distributed stochastic neighbor embedding) is a non-linear dimensionality reduction method, particularly adept at preserving the local structure of data, suitable for visualizing high-dimensional data. The system adaptively selects the most suitable dimensionality reduction method based on data characteristics, or combines the results of multiple methods. After dimensionality reduction, the system performs feature selection based on information content indicators (such as variance contribution rate and mutual information), retaining those features carrying the most information, and sets a minimum information content threshold to ensure that the selected feature set has sufficient expressive power. Through this meticulous dimensionality reduction and selection process, the system ultimately obtains sparse coding features with moderate dimensionality and rich information content.

[0069] For sparse encoded features, the system further performs nonlinear transformations and introduces group sparsity and structural sparsity constraints to obtain a sparsity-enhanced feature set. Nonlinear transformations refer to using nonlinear functions to transform features to capture nonlinear relationships between them. Commonly used nonlinear transformations include logarithmic transformations, power transformations, and exponential transformations. The system selects the most suitable nonlinear transformation method for each feature based on its distribution characteristics and business meaning. For example, for long-tailed monetary features, a logarithmic transformation might be used; for ratio features with clear upper and lower bounds, a sigmoid transformation might be used. While performing nonlinear transformations, the system also introduces two important sparsity constraints: group sparsity and structural sparsity. Group sparsity refers to grouping features so that feature selection is done on a group-by-group basis. This means either selecting all features within a group or not selecting any features from that group at all. This is suitable for handling naturally grouped features, such as feature groups from the same data source or expressing similar information. Structured sparsity further considers the structural relationships between features, such as hierarchical structures and network structures, ensuring that the selected features meet predetermined structural constraints. This helps maintain the logical relationships and integrity between features. By introducing these sparsity constraints, the system not only controls the number of features but also ensures the structural rationality and business interpretability of the selected feature combinations, ultimately resulting in a feature set that is both sparse and richly expressive.

[0070] Finally, the system performs arithmetic operations, ratio calculations, and cross-feature construction on the sparsity-enhanced feature set to obtain the final high-order feature set. Arithmetic operations refer to combining different features into new features using basic mathematical operations such as addition, subtraction, multiplication, and division, such as the sum, difference, product, and quotient of two features. Ratio calculation is a special type of arithmetic operation that generates new features with normalization properties and business interpretability by calculating the ratio of two features; for example, "current transaction amount / historical average transaction amount" can reflect the degree of transaction anomalies. Cross-feature construction generates new features that can capture complex interaction relationships by combining different values ​​of multiple features. For example, for categorical features A and B, an A×B cross-feature can be constructed to represent the joint distribution of the two features. In practice, the system does not simply exhaustively enumerate all possible combinations, but rather, based on the previously constructed feature association graph and business knowledge, selectively chooses combinations that may have high predictive value, such as feature pairs with strong correlations or feature pairs that complement each other in terms of business. In addition, the system incorporates a complexity control mechanism to limit the highest order of combinations of high-order features (usually no more than 3 orders) to avoid combinatorial explosion and overfitting risks. Through this meticulous high-order feature construction process, the system ultimately generates a high-order feature set rich in information and with strong predictive power, providing robust support for risk control models.

[0071] The calculation of initialization statistics is fundamental to feature normalization and dynamic adaptation. The system employs an advanced method based on a continuum model to achieve fine-grained modeling and parameter initialization of feature distributions. First, the system treats each feature data point as a dimension within a continuous space in the feature space, defining the feature value range as a continuous interval, thus obtaining the definition of the continuous feature space. A continuous space refers to a space where feature values ​​can vary continuously within a certain range; for example, the age feature can take any real value between 0 and 120. By mapping features to a continuous space, the system can leverage the powerful tools of continuous mathematics to model the feature distribution, laying the foundation for subsequent fine-grained processing.

[0072] Based on the definition of a continuous feature space, the system divides the continuous space into multiple subdomains. The boundaries of each subdomain are determined by the 5%, 25%, 50%, 75%, and 95% quantiles, resulting in domain discretization. Domain discretization refers to dividing the continuous feature space into a finite number of discrete intervals (subdomains) to facilitate segmented processing of data with different distribution characteristics. The system selects key quantiles as subdomain boundaries instead of simple equidistant divisions, a method that better reflects the actual distribution characteristics of the data. Specifically, the system uses the 5%, 25%, 50% (median), 75%, and 95% quantiles to divide the feature space into six subdomains: [minimum, 5% quantile], [5% quantile, 25% quantile], [25% quantile, 50% quantile], [50% quantile, 75% quantile], [75% quantile, 95% quantile], and [95% quantile, maximum value]. This partitioning method has several advantages: First, it can adaptively adjust the size of subdomains according to data density, with smaller subdomains in data-dense areas and larger subdomains in data-sparse areas; second, by isolating extreme values ​​(such as values ​​below the 5% quantile or above the 95% quantile), it can mitigate the impact of extreme values ​​on the overall modeling; finally, this quantile partitioning method naturally corresponds to the risk stratification commonly used in the risk control field, which facilitates subsequent risk assessment and interpretation.

[0073] For each subdomain, the system selects a first-order linear basis function, a second-order polynomial basis function, or an exponential-logarithmic basis function as the interpolation basis function, thus obtaining the basis function selection scheme. Interpolation basis functions are mathematical tools used to construct continuous functions between discrete data points, approximating the relationship between eigenvalues ​​and the target variable within a subdomain. First-order linear basis functions are the simplest interpolation method, assuming a linear relationship between the eigenvalues ​​and the target variable within the subdomain, in the form f(x) = ax + b. Second-order polynomial basis functions introduce quadratic terms, enabling the description of more complex nonlinear relationships, in the form of... Exponential and logarithmic basis functions, through exponential or logarithmic transformations, are suitable for handling data with exponential growth or decay characteristics. The system automatically selects the most suitable basis function type based on the distribution characteristics of the data within each subdomain and its relationship with the target variable. For example, for subdomains with approximately linear relationships, first-order linear basis functions are selected to avoid overfitting; for subdomains with significant nonlinearity, second-order polynomial or exponential-logarithmic basis functions are selected to improve fitting accuracy. This adaptive basis function selection scheme enables the system to employ the most appropriate modeling method for data segments with different distribution characteristics, improving the overall model's accuracy and robustness.

[0074] Based on the basis function selection scheme, the system calculates initialization parameters at quantiles in each subdomain, obtaining the node parameter initialization results. Node parameters refer to the function values ​​and derivative values ​​at the subdomain boundary points (quantiles), used to determine the shape of the interpolation function. The system first calculates the characteristic statistics of each quantile point based on historical data, including the mean, standard deviation, and correlation coefficient with the target variable. Then, according to the selected basis function type, the system determines the number and type of parameters to be initialized. For example, for first-order linear basis functions, the function value and first derivative of each node need to be initialized; for second-order polynomial basis functions, the second derivative also needs to be initialized. The system calculates these derivative values ​​through local data fitting or numerical differentiation methods to ensure continuity and smoothness at the nodes. These node parameters collectively constitute the skeleton of the continuum model, determining the model's behavioral characteristics within each subdomain.

[0075] Based on the node parameter initialization results, the system assembles the local equations of each subdomain into a global equation, solves this global equation, and obtains the coefficient values ​​at each quantile, thus yielding the continuum model coefficients. The continuum model is an advanced modeling method that treats the discrete feature space as a continuous medium and uses partial differential equations to describe feature behavior. In this step, the system first establishes local equations for each subdomain, in the form Ax = b, where A is a coefficient matrix composed of basis functions and their derivatives, b is the node parameter vector, and x is the coefficient of the basis functions to be solved. Then, the system assembles these local equations into a global equation system, ensuring that continuity and smoothness conditions are satisfied at the boundaries of subdomains. The construction of the global equation follows the idea of ​​the finite element method, "gluing" the local equations into a whole by applying continuity constraints at the nodes. Finally, the system solves this global equation system to obtain the coefficient values ​​at each quantile, which completely determine the shape and behavior of the continuum model. Compared with traditional discrete feature processing methods, the continuum model can more accurately describe the behavior of features throughout the entire value domain, especially in sparse data regions, exhibiting better interpolation and extrapolation capabilities.

[0076] Based on continuum model coefficients and differentiated sliding window parameters, the system calculates initial statistics using a continuum embedding mapping method. Continuum embedding mapping is a technique that maps the original feature space to a new feature space, aiming to enhance the expressiveness and discriminative power of features. In its implementation, the system first locates each feature data to be processed within its subdomain and converts it into local coordinates within that subdomain. Then, the system uses pre-calculated basis function coefficients to calculate the mapping value of the feature in the continuum model. This mapping process considers the feature's position within the local subdomain, the influence of neighboring subdomains, and global distribution characteristics, generating a more expressive feature representation. For all feature data within the sliding window, the system calculates its mapping value using this method and then calculates initial statistics based on these values, including the mean, standard deviation, and various quantiles. These statistics not only reflect the distribution characteristics of the original features but also incorporate the complex nonlinear relationships captured by the continuum model, providing a more accurate and robust parameter basis for subsequent normalization processing.

[0077] Through the detailed high-order feature engineering and initial statistical calculation processes described above, the system constructs an information-rich and highly expressive feature set, providing a solid mathematical foundation for feature normalization and dynamic adaptation. The application of these methods enables the risk control model to capture more complex risk patterns and data relationships, significantly improving the model's predictive performance and stability, and providing strong support for intelligent decision-making in the financial risk control field. In practical applications, this methodology has increased the model's KS value by an average of over 20%, improved fraud detection accuracy by approximately 15%, and reduced the false positive rate by about 25%, fully demonstrating the crucial value of high-order feature engineering and refined statistical calculation in modern risk control systems.

[0078] In the process of constructing a continuum model, assembling the nodal parameter initialization results into global equations and solving them is a crucial step in obtaining the continuum model coefficients. This process involves multiple technical aspects, including subdomain localization, local coordinate transformation, shape function calculation, error control, and mesh refinement. These steps together constitute a complete feature space continuous modeling workflow, providing an accurate and robust feature representation method for risk control models.

[0079] For each feature data to be processed, the system first needs to determine its subdomain in continuous space and convert the feature data into the local coordinate system of the corresponding subdomain to obtain local coordinates. A subdomain refers to a feature value interval divided by adjacent quantiles. For example, [25% quantile, 50% quantile] constitutes one subdomain, and [50% quantile, 75% quantile] constitutes another. Subdomain location is determined by comparing the size relationship between the feature value and each quantile value. For example, if a feature value is greater than the 25% quantile but less than the 50% quantile, it is located in the second subdomain. The local coordinate system is a standardized reference system established within each subdomain. Typically, the left boundary of the subdomain (e.g., the 25% quantile) is mapped to 0, and the right boundary (e.g., the 50% quantile) is mapped to 1. Any point within the subdomain can be represented by a relative position between 0 and 1. For the case where the feature value x is located within the subdomain [a, b], its local coordinate ξ is calculated as follows: This local coordinate transformation enables unified processing of subdomains of different sizes, simplifying subsequent numerical calculations. For example, for an eigenvalue of 200, if it lies within the subdomain [100, 300], its local coordinates are... This indicates that it is located at the midpoint of the subdomain. Through this subdomain localization and local coordinate transformation, the system maps discrete feature data into a unified computational framework, laying the foundation for subsequent continuum modeling.

[0080] Calculating shape function values ​​for local coordinates using predefined interpolation basis functions is a core step in constructing a continuum model. Shape functions are an important concept in finite element analysis; they are a set of interpolation functions defined on local elements, used to construct continuous approximate solutions between discrete nodes. In feature modeling, shape functions are used to achieve smooth interpolation between quantiles, ensuring the continuity of the feature space. Commonly used shape functions include Lagrange shape functions, Hermite shape functions, and spline shape functions. Lagrange shape functions are the simplest, ensuring the interpolation function passes through the given values ​​at the nodes; Hermite shape functions not only guarantee matching function values ​​at the nodes but also matching derivative values, providing smoother interpolation; spline shape functions provide guarantees on the continuity of higher-order derivatives, suitable for scenarios requiring high smoothness. In practical applications, the system typically selects the appropriate shape function type based on the nature of the feature and accuracy requirements. For example, for features with relatively gentle changes, a simple linear Lagrange shape function might be chosen; for features with drastic changes or requiring high-precision interpolation, Hermite shape functions or cubic spline shape functions might be selected. The specific calculation of the shape function depends on the type of shape function chosen, but the basic principle is to use local coordinate values ​​to calculate the values ​​of a set of basis functions at that point. These values ​​will be used for subsequent mapping calculations. For example, for a linear Lagrange shape function, the values ​​of the two basis functions are... ,in These are local coordinates. For a point with local coordinates of 0.3, its shape function value is... .

[0081] Based on shape function values ​​and quantile parameters, the system calculates the mapping values ​​of feature data in continuous space, obtaining continuum embedding mapping values. Continuum embedding mapping is the process of mapping discrete feature data to a continuous feature space, representing a high-level form of feature representation. In this step, the system uses the previously calculated shape function values, combined with parameters at quantile points (such as function values, first derivatives, second derivatives, etc.), to construct the mapping values ​​for feature points. For Lagrange shape functions, the calculation of the mapping values ​​is relatively simple, being a weighted sum of each node value and its corresponding shape function value. For example, for a point with local coordinates ξ, its mapping value can be expressed as... ,in Is the i-th shape function in The value at f i This is the function value at the i-th node. For Hermite-shaped functions, the calculation also needs to consider the contribution of the derivative, expressed as: ,in It is a shape function related to the derivative. It is the derivative value of the node. In this way, the system achieves a smooth mapping from discrete quantiles to a continuous feature space, so that any point in the feature space has a well-defined mapping value, forming a complete continuum representation. This continuum embedding mapping not only fills the "gaps" between quantiles, but also provides more detailed feature discrimination capabilities, which is of great significance for capturing subtle risk differences in risk control scenarios.

[0082] The continuum embedding mapping values ​​are standardized to the target interval of 0 to 1, resulting in the processed mapping values. Standardization is an important step in data preprocessing, aiming to unify data of different scales and ranges into a standard interval for easier subsequent processing and comparison. The system uses the Min-Max standardization method to linearly map the original mapping values ​​to the [0,1] interval. Specifically, for the original mapping value x, the standardized value is... Calculated as ,in and These are the minimum and maximum values ​​of all mapped values, respectively. In practical applications, to enhance robustness to outliers, the system may use quantiles instead of extreme values, such as using the 5th quantile to replace the minimum and the 95th quantile to replace the maximum. Standardization not only unifies the numerical range but also preserves the relative order and distribution shape of the data, representing a transformation that retains information. For risk control models, standardized features are easier to compare and fuse across features and better meet the input requirements of most machine learning algorithms. By standardizing the continuum embedding mapping values ​​to the 0-1 range, the system provides a unified and standardized input for subsequent feature usage and model training.

[0083] The system calculates the mapping error within each subdomain, refines the mesh for subdomains with errors exceeding a threshold, and recalculates the mapping values ​​for the refined subdomains to obtain high-precision feature mapping values. Mapping error refers to the difference between the predicted values ​​and actual feature values ​​of the continuum model, and is a key indicator for evaluating model accuracy. The system selects a set of validation points within each subdomain, compares the true feature values ​​of these points with the model's predicted values, and calculates various error indicators, such as mean squared error (MSE), mean absolute error (MAE), or maximum absolute error. When the error in a subdomain exceeds a preset threshold, the system performs mesh refinement. Mesh refinement is an adaptive mesh optimization technique that improves local modeling accuracy by adding nodes in high-error regions. Specifically, the system further divides subdomains with excessive errors into smaller sub-intervals; for example, [25% quantile, 50% quantile] is divided into [25% quantile, 37.5% quantile] and [37.5% quantile, 50% quantile]. For the refined subdomains, the system needs to recalculate the parameter values ​​of the newly added nodes, typically through interpolation of the original data or more refined local fitting. Then, the system reapplies shape function calculations and mapping values ​​to the refined mesh until the accuracy requirements are met or the maximum number of refinements is reached. This adaptive mesh refinement strategy can allocate more modeling resources to regions with complex variations in the feature space, while maintaining a simple model in regions with gradual changes, achieving efficient allocation of computational resources and a regional balance in modeling accuracy. In the field of risk control, this refined modeling is particularly important for accurately capturing feature changes in critical risk regions, helping to improve the accuracy of model judgments in boundary conditions.

[0084] Based on continuum embedding mapping values ​​and high-precision feature mapping values, the system ultimately obtains continuum model coefficients. These coefficients are the core of the continuum model, a set of parameters describing the behavior of features throughout the continuous space. These coefficients include parameters such as function values ​​and derivatives at the original quantiles and newly added nodes after refinement, fully defining the mapping behavior of features across the entire value domain. The system obtains optimal model coefficients by solving a global equation system, comprehensively considering the constraints and continuity requirements of all subdomains. The construction of the global equations considers interpolation conditions (function value matching), continuity conditions (continuity of function values ​​and derivatives at the boundaries of adjacent subdomains), and smoothness conditions (minimizing a certain energy functional). The solved model coefficients not only ensure accurate reproduction of given values ​​at quantiles but also guarantee the smoothness and continuity of the entire feature space. Compared with traditional discrete feature processing methods, the continuum model provides more accurate and smoother feature representations, especially exhibiting better interpolation and extrapolation capabilities in sparse data regions. This advanced feature modeling method provides risk control models with more detailed and stable feature inputs, effectively improving the model's predictive performance and generalization ability, and has significant advantages in capturing complex risk patterns and boundary conditions.

[0085] When the PSI index is greater than 0.1, it indicates a significant change in the feature distribution, and the system needs to recalculate the statistics based on the updated window data. PSI (Population Stability Index) measures the difference between two distributions. It is calculated by dividing the feature range into multiple intervals and comparing the proportion of each interval in the two distributions. A higher PSI value indicates a more significant change in distribution; generally, a PSI > 0.1 indicates a significant change in distribution, requiring model or parameter updates. To efficiently handle the recalculation of large-scale feature data, the system employs a series of optimization techniques based on sparse matrices.

[0086] First, the system converts the updated window data into a sparse matrix representation and records the positions and values ​​of non-zero elements, resulting in a sparse data representation. A sparse matrix is ​​a matrix where most elements are zero, and it is very common in risk control features, such as behavioral features and transaction records, which are typically highly sparsity (most users do not trigger specific behaviors, and no transactions occur at most points in time). Sparse matrix representation is an efficient data structure that stores only non-zero elements and their position information, significantly reducing storage space and computation compared to dense representation. Common sparse matrix storage formats include coordinate (COO), compressed row (CSR), and compressed column (CSC) formats. The coordinate format is the most intuitive, storing the row index, column index, and value of each non-zero element; the compressed row format optimizes row access efficiency, making it particularly suitable for scenarios with frequent row operations; the compressed column format optimizes column access, making it suitable for scenarios with frequent column operations. The system typically selects the most suitable storage format based on subsequent computational needs. For example, if frequent calculations of correlation between features (calculated by column) are required, the CSC format may be more suitable; if frequent retrieval of the complete feature vector of a sample (calculated by row) is required, the CSR format may be more efficient. By converting to a sparse representation, the system can efficiently process large-scale feature data, creating conditions for subsequent high-performance statistical computing.

[0087] Based on sparse data representation, the system calculates the covariance values ​​between feature data, sets covariance values ​​less than a preset threshold to zero, and obtains a sparse matrix representation of the feature covariance. The covariance matrix is ​​an important statistical tool for describing the linear relationship between features; its element (i,j) represents the covariance between feature i and feature j, and the calculation formula is as follows: ,in and It is an eigenvalue. and Here, E represents the feature mean, and E represents the expected value. In high-dimensional feature spaces, most feature pairs are weakly correlated, with corresponding covariances close to zero. The system sets a small threshold... (like ), with absolute values ​​less than The covariance value is treated as zero, further increasing the sparsity of the matrix. This sparsity treatment not only improves computational efficiency but also helps reduce the influence of noise and highlight significant feature relationships. The calculation of the sparse covariance matrix can be optimized by leveraging the characteristics of sparse matrices; for example, only the contributions of non-zero elements can be calculated, avoiding the participation of a large number of zero elements in the calculation. In addition, the system can also adopt a block-based computation strategy, decomposing the large matrix into multiple smaller blocks for parallel computation and then merging the results to improve computational efficiency. Through these optimizations, the system can efficiently calculate and store the covariance matrix of large-scale features, providing a basis for the recalculation of subsequent statistics.

[0088] To process the sparse matrix representation of feature covariance, the system employs several specialized optimization techniques. Nested Dispersion, a graph-partition-based matrix reordering algorithm, recursively divides the computation graph into relatively independent sub-regions, reducing connections between them. In matrix computation, this reordering significantly reduces "filling"—that is, positions that were originally zero become non-zero during computation—alleviating computational burden and storage pressure. Approximate Minimum Degree (AMD) is another important sparse matrix reordering algorithm. It uses a greedy strategy to select the column / row that results in the least filling at each step for elimination. AMD algorithms do not require complete symbolic decomposition but use approximate degree update rules, greatly reducing computational overhead and making them particularly suitable for large-scale sparse matrix processing. Supernode technology merges columns with the same non-zero pattern into a single processing unit, reducing indirect addressing overhead and improving cache utilization. For example, if several consecutive columns in a matrix have the exact same distribution pattern of non-zero elements, they can be treated as a single supernode, reducing indexing operations and improving data locality. Symbolic analysis, performed before numerical computation, analyzes the non-zero structure of the matrix to predict potential filling positions during decomposition, pre-allocates memory, and optimizes access patterns. The combined application of these optimization techniques enables the system to significantly improve the efficiency of sparse matrix processing while maintaining numerical accuracy, creating favorable conditions for subsequent parallel LU decomposition.

[0089] For the processed sparse matrix, the system performs hybrid parallel sparse LU decomposition, breaking down the LU decomposition task into independent subtasks. These subtasks are executed in parallel on multiple processors, yielding the decomposed L and U matrices. LU decomposition is a fundamental operation in linear algebra, decomposing matrix A into a lower triangular matrix L and an upper triangular matrix U such that A = LU. This decomposition simplifies the solution process of linear equation systems, transforming Ax = b into solving two simpler triangular systems: Ly = b (forward substitution) and Ux = y (backward substitution). Traditional LU decomposition algorithms are inefficient on large-scale sparse matrices, while the hybrid parallel strategy significantly improves computational performance by combining multiple parallel computing techniques. The system employs a hybrid application of task-level parallelism, data-level parallelism, and heterogeneous computing strategies, forming a highly efficient parallel computing framework that fully utilizes the computing power of modern multi-core processors and GPU accelerators.

[0090] Task-level parallelism is a parallelization approach that decomposes the LU decomposition task into multiple relatively independent subtasks, which are then executed in parallel by multiple processing units. In sparse LU decomposition, the system uses dependency graph analysis to identify dependencies in the computation process, marking subtasks without mutual dependencies as parallelizable. For example, for matrices with block diagonal structures, the decomposition of each diagonal block can be performed in parallel; for general sparse matrices, the system can construct an elimination tree, where computations on different branches can be performed in parallel. The system employs a work-stealing scheduling algorithm to dynamically allocate tasks and balance the load on each processor. In work-stealing mode, each processor maintains its own task queue. When a processor completes its own task queue, it can "steal" tasks from other processor queues, improving overall resource utilization. Furthermore, the system considers task granularity optimization, merging excessively small tasks to reduce scheduling overhead and splitting excessively large tasks to increase parallelism. Through this fine-grained task parallelism strategy, the system can fully utilize the computing power of multi-core CPUs to accelerate the decomposition process of large-scale sparse matrices.

[0091] Data-level parallelism is a parallelization method that performs the same operation on multiple data elements. Modern processors support this parallelism through the SIMD (Single Instruction Multiple Data) instruction set. SIMD instructions allow a single instruction to perform the same operation on multiple data elements simultaneously, significantly improving computational efficiency. Common SIMD instruction sets include SSE, AVX, and AVX-512, capable of processing 2 to 16 floating-point numbers simultaneously, depending on the data type and instruction set. In sparse LU decomposition, although the overall matrix is ​​sparse, local computations may involve dense submatrix operations, such as diagonal block inversion and matrix-matrix multiplication. The system identifies these computationally intensive dense subtasks and optimizes computation using the SIMD instruction set. For example, for multiplication operations on small dense matrices, the system can significantly accelerate computation using loop expansion and SIMD instructions. Furthermore, the system optimizes memory access patterns, reducing the impact of cache misses and memory latency through techniques such as data blocking, prefetching, and alignment. Data blockization divides large matrices into smaller blocks suitable for cache size, improving cache hit rate; prefetching loads data into the cache before it is actually needed, hiding memory latency; alignment ensures that the starting address of the data meets the alignment requirements of SIMD instructions, avoiding performance loss from misaligned accesses. Through these data-level parallel optimizations, the system significantly improves the efficiency of locally dense computations and accelerates the overall LU decomposition process.

[0092] Heterogeneous computing is a computing paradigm that utilizes different types of processors to work collaboratively, such as the collaboration between CPUs and GPUs. In large-scale sparse matrix computation, different types of subtasks are best suited for execution on different hardware. GPUs excel at handling regular, computationally intensive tasks, such as operations on large, dense submatrices; while CPUs are better suited for irregular, control-flow-complex, or communication-intensive tasks, such as task scheduling and sparse data rearrangement. The system dynamically determines the execution location based on the characteristics of the subtasks. For identified computationally intensive tasks, such as LU decomposition or matrix multiplication of large, dense submatrices, the system transfers them to the GPU, using parallel computing frameworks such as CUDA or OpenCL to accelerate computation; for tasks with complex dependencies or poor data locality, such as sparse matrix reordering or task scheduling, they remain on the CPU. The system also optimizes data transfer between the CPU and GPU, a common bottleneck in heterogeneous computing. Optimization strategies include pipelined processing (overlapping data transfer and computation), batch transfer (reducing startup overhead), data reuse (reducing redundant transfers), and memory pool management (reducing dynamic allocation overhead). Furthermore, the system employs an asynchronous execution model, allowing the CPU to perform other tasks immediately after submitting a GPU task, without waiting for the GPU to complete, thus improving overall resource utilization. Through this heterogeneous computing strategy, the system fully leverages the advantages of different hardware components, achieving optimal overall performance.

[0093] Based on the processing results of task-level parallelism, data-level parallelism, and heterogeneous computing, the system employs specialized optimization algorithms for different matrix structures. A banded matrix is ​​a matrix where non-zero elements are concentrated near the main diagonal, and the bandwidth *b* represents the distribution of non-zero elements within *b* diagonals above and below the main diagonal. For banded matrices, a dedicated banded matrix solver significantly reduces computation and storage requirements by utilizing the limited bandwidth, employing an efficient storage format (storing only elements within the band), and specialized algorithms. An arrow matrix is ​​a matrix where non-zero elements are concentrated in certain rows and columns, forming an "arrow" shape. This structure is common in hierarchical problems, such as domain decomposition methods or tree structure problems. The forward-backward replacement optimization algorithm is specifically optimized for arrow-shaped structures; for example, it can solve the "shaft" part first, and then use these results to efficiently solve the "arrow" part, avoiding redundant computations in general methods. The multifrontal method is a general direct solution method for sparse matrices, applicable to general sparse structures. It utilizes the concept of a frontal matrix to decompose and organize a sparse matrix into a tree structure, with each front corresponding to a node in the elimination tree. The algorithm processes the frontal matrix from the bottom up, passing some results to the parent nodes, thus achieving efficient parallel computation. These structure-aware optimization algorithms enable the system to automatically select the most suitable solution strategy based on the actual structural characteristics of the matrix, significantly improving computational efficiency. Through the application of these specialized algorithms, the system ultimately obtains efficiently computed L and U matrices, providing a solid foundation for subsequent statistical calculations.

[0094] Based on the L and U matrices, the system recalculates the mean, standard deviation, and quantile statistics. The L and U matrices provide an efficient method for solving linear equations, and statistical calculations can often be transformed into solving linear equations. For example, calculating the weighted mean might require solving an equation of the form Ax = b, where A is the covariance matrix, b is a known vector, and x is the weight vector to be calculated. With LU decomposition, the system can efficiently solve this in two steps: forward and backward substitution, without directly inverting the original matrix (which has high computational complexity and poor numerical stability). For quantile calculations, the system might use an iterative method, where each iteration involves solving a system of linear equations; LU decomposition makes this process more efficient. Mean calculation is relatively straightforward and can be represented as a weighted average of eigenvalues; standard deviation calculation involves the square root of the sum of squares of the mean deviations; quantile calculation requires interpolation of the sorted data or the use of approximate algorithms such as quantile regression. The system also fully utilizes the correlation structure between features, obtaining statistics for multiple features at once through matrix calculations, rather than calculating them one by one, further improving efficiency. By using this statistical calculation method based on sparse matrix optimization, the system can quickly recalculate key statistical parameters when the feature distribution changes significantly, ensuring timely adaptation and high-quality updates of feature normalization.

[0095] Hybrid parallel sparse LU decomposition is a highly efficient method for large-scale sparse matrix factorization. By combining multiple parallel strategies and structural optimization techniques, it significantly improves computational performance. This method first decomposes the LU decomposition task into multiple independent subtasks, employing a task-level parallel strategy for parallel execution on multi-core CPUs. Task decomposition is based on matrix structure analysis and dependency identification, typically using elimination trees or task dependency graphs to represent dependencies between tasks. The system uses a work-stealing scheduling algorithm to dynamically allocate tasks, ensuring balanced load across processors and improving resource utilization. For large, dense subtasks derived from the LU decomposition task, the system employs a data-level parallel strategy, applying the SIMD instruction set to accelerate computation. SIMD instructions can process multiple data elements simultaneously, significantly improving the efficiency of computationally intensive operations. The system also optimizes memory access patterns, reducing the impact of cache misses and memory latency through data blockization, prefetching, and alignment techniques. For computationally intensive subtasks, the system employs a heterogeneous computing acceleration strategy, transferring them to GPUs for processing to fully utilize their parallel computing capabilities; simultaneously, communication-intensive subtasks are handled on the CPU to avoid frequent data transfer overhead. The system optimizes data transfer between the CPU and GPU, using techniques such as pipelining, batch transfer, and memory pooling to reduce communication overhead. For different matrix structures, the system employs specially optimized algorithms: a dedicated striped matrix solver is used for striped matrices, a forward-backward replacement optimization algorithm is used for arrow-shaped structures, and a multi-front algorithm is used for general sparse structures. Through this multi-level, multi-dimensional optimization strategy, the system significantly improves the performance of sparse LU decomposition, providing solid support for efficient recalculation of statistics.

[0096] By comprehensively applying continuum model construction and sparse matrix optimization techniques, the system achieves efficient processing and refined modeling of large-scale feature data. These advanced mathematical and computational methods enable the system to significantly improve computational efficiency while maintaining high accuracy, providing high-quality feature support for risk control models. In practical applications, this approach has increased feature processing speed by 5-10 times and reduced storage requirements by more than 70%, while maintaining or improving the expressiveness and discriminative power of features, providing strong support for the accuracy and real-time performance of risk control decisions. More importantly, when feature distribution changes, the system can respond quickly, recalculating the optimal statistical parameters to ensure the continued effectiveness of feature normalization and model prediction, greatly enhancing the adaptability and stability of the risk control system in complex and ever-changing environments.

[0097] Hybrid parallel sparse LU decomposition is an advanced algorithm for efficiently processing large-scale sparse matrices. By comprehensively utilizing multiple parallel computing techniques and structural optimization strategies, it significantly improves decomposition speed and computational efficiency. The core idea of ​​this method is to break down the LU decomposition task into parallelizable subtasks and select the most suitable parallel strategy and hardware platform based on the characteristics of each subtask, while also employing specially optimized algorithms considering the matrix structure features. This multi-dimensional optimization approach fully leverages the advantages of modern computing architectures, providing robust support for statistical calculations in risk control systems.

[0098] First, the system employs a pre-defined task-level parallelism strategy to decompose the LU decomposition task into multiple independent subtasks. These subtasks are then executed in parallel on a multi-core CPU, yielding the task-level parallel processing result. Task-level parallelism is a high-level parallelization method that improves overall computational speed by identifying independent parts of the computational task and having multiple processing units execute different tasks simultaneously. In sparse LU decomposition, the foundation of task decomposition is matrix dependency analysis. The system first constructs an elimination tree, a tree structure representing the dependency relationships of variables during matrix decomposition. Each node in the elimination tree corresponds to a column of the matrix, and the parent-child relationship between nodes indicates the dependency order of elimination operations. Operations corresponding to nodes without direct dependencies in the elimination tree can be executed in parallel, providing a theoretical basis for task decomposition. Based on the elimination tree, the system uses a task partitioning algorithm to divide the LU decomposition process into multiple subtasks, each responsible for processing a portion of the matrix. The partitioning strategy considers load balancing and communication costs, striving to ensure that the computational load of each subtask is similar and that inter-task communication is minimized.

[0099] After task decomposition, the system employs various scheduling techniques to efficiently execute these subtasks on multi-core CPUs. Static scheduling pre-allocates tasks before execution, suitable for scenarios with predictable computational loads; dynamic scheduling allocates tasks at runtime based on actual conditions, making it more suitable for irregular sparse matrices. Work stealing is a commonly used dynamic scheduling strategy where each processor maintains its own task queue. After completing its own tasks, a processor can "steal" tasks from other processor queues, effectively balancing the load. The system also implements a task priority mechanism, assigning priorities based on the task's position on the critical path, prioritizing critical tasks that affect overall execution time. Furthermore, the system considers the characteristics of NUMA (Non-Uniform Memory Access) architecture, allocating tasks to processors closest to their data to reduce memory access latency. Through this refined task-level parallelism strategy, the system fully utilizes the computing power of multi-core CPUs, significantly accelerating the sparse LU decomposition process, especially for matrix structures with good parallelism, where the speedup is almost linearly related to the number of processors.

[0100] Next, the system employs a pre-defined data-level parallelism strategy to apply the SIMD (Single Instruction Multiple Data) instruction set to the large, dense subtasks decomposed from the LU decomposition task, obtaining data-level parallel processing results. Data-level parallelism refers to a parallel approach that performs the same operation on multiple data elements simultaneously, implemented through SIMD instructions. SIMD instructions are a high-efficiency computing technique supported by modern processors, allowing a single instruction to process multiple data elements concurrently, significantly improving the throughput of computationally intensive operations. In sparse LU decomposition, although the overall matrix is ​​sparse, the decomposition process generates some locally dense submatrices, such as diagonal block matrices or front-edge matrices; these parts are suitable for acceleration using SIMD instructions.

[0101] The system first identifies the dense computational cores in the LU decomposition process, mainly including submatrix update operations (also known as Schur complement operations), triangulation, and matrix-matrix multiplication. For these operations, the system uses algorithm variants optimized for SIMD. For example, in matrix-matrix multiplication, the system employs a block-based algorithm to decompose the large matrix into smaller blocks, and then uses SIMD instructions to process multiple elements within each block in parallel. The SIMD instruction set used by the system varies depending on the hardware platform, including SSE, AVX, and AVX-512 for x86 architecture, or NEON for ARM architecture. These instruction sets can process 2 to 16 floating-point numbers simultaneously, greatly improving computational density. To fully leverage the performance of SIMD instructions, the system also implements several optimization techniques: data alignment ensures that memory accesses meet the alignment requirements of SIMD instructions; data packing reassembles discontinuous data into contiguous blocks suitable for SIMD operations; loop unrolling reduces loop control overhead; and instruction pipeline optimization ensures efficient execution of SIMD instructions. Furthermore, the system offers multiple implementations to meet different precision requirements, such as single-precision and double-precision versions, allowing for the selection of the more efficient single-precision computation when feasible. Through this fine-grained data-level parallel optimization, the system significantly improves the efficiency of locally dense computations, providing crucial support for overall LU decomposition performance.

[0102] The system employs a pre-defined heterogeneous computing acceleration strategy to offload computationally intensive subtasks to the GPU for processing, while leaving communication-intensive subtasks on the CPU, resulting in heterogeneous computing accelerated processing results. Heterogeneous computing refers to a computing model that utilizes different types of processors to work collaboratively, such as a combination of CPU and GPU. GPUs (Graphics Processing Units), with their numerous parallel processing cores and high memory bandwidth, are particularly suitable for handling computationally intensive tasks with high rules and data parallelism; while CPUs are better suited for tasks with complex control flows or high communication intensity. This difference in hardware characteristics provides a natural guide for task allocation.

[0103] The system first analyzes the computational characteristics of each subtask in the LU decomposition process to identify tasks suitable for GPU acceleration. Typically, computationally intensive tasks, such as updating large-scale dense submatrices and decomposing large frontier matrices, involve high computational demands and data parallelism, making them suitable for GPU execution. Conversely, control-intensive tasks such as task scheduling, sparse data structure management, and dependency analysis, as well as communication-intensive tasks requiring frequent small-scale data exchanges, are better suited for CPU execution. The system implements a task characteristic analysis framework that considers factors such as computational density (operations per byte of data), data locality, computational regularity, and data transfer volume. It evaluates the potential benefits of GPU acceleration for each subtask and only assigns the task to the GPU when the benefits exceed the data transfer overhead.

[0104] For tasks executed by the GPU, the system uses parallel computing frameworks such as CUDA or OpenCL for efficient processing. In the GPU parallel computing model, computational tasks are organized into a multi-layered parallel structure: a grid contains multiple blocks, and each block contains multiple threads. The system designs the optimal thread organization based on task characteristics and data scale, balancing parallelism and resource utilization. For dense matrix operations, the system implements highly optimized GPU kernel functions, utilizes shared memory to reduce global memory access, uses cooperative groups for thread synchronization, and optimizes thread communication through warp shuffle instructions. For large-scale tasks, the system adopts a pipelined execution model, decomposing the task into multiple stages and establishing multiple streams on the GPU for parallel execution, while overlapping with CPU computation to maximize resource utilization.

[0105] The system also optimizes data transfer between the CPU and GPU, a common bottleneck in heterogeneous computing. Optimization strategies include: batch data transfer to reduce startup overhead; zero-copy memory, allowing the GPU to directly access host memory, suitable for small datasets; unified memory, simplifying memory management with automatic data migration handled by the system; asynchronous data transfer overlapping with computation to hide latency; and data compression to reduce transfer volume, particularly suitable for sparse data. Through this sophisticated heterogeneous computing strategy, the system fully leverages the strengths of both the CPU and GPU, achieving optimal overall performance. Especially for computationally intensive tasks, GPU acceleration can provide 5-10 times or even higher performance improvements.

[0106] Based on the results of task-level parallel processing, data-level parallel processing, and heterogeneous computing acceleration, the system employs specially optimized algorithms for different matrix structures. These include a dedicated strip matrix solver for striped structures, a forward-backward replacement optimization algorithm for arrow-shaped structures, and a multi-front algorithm for general sparse structures, ultimately yielding the decomposed L and U matrices. This structure-aware algorithm selection strategy recognizes that different sparsity patterns require different processing methods, improving computational efficiency through specialization.

[0107] A banded matrix is ​​a matrix whose non-zero elements are concentrated within a finite width around the main diagonal. Its bandwidth is defined as the sum of the upper and lower bandwidths plus one (the main diagonal). Banded matrices are common in various applications, such as finite difference methods and time series analysis. Dedicated banded matrix solvers leverage the special characteristics of the banded structure, employing efficient storage formats and algorithms. In terms of storage, the system only stores elements within the band, typically using a banded storage format that compactly arranges these elements, significantly reducing memory requirements. Algorithm-wise, the system implements LU decomposition optimized for banded structures, avoiding unnecessary operations on known zero-valued out-of-band elements, reducing computational complexity from that of general sparse matrices. Reduce to Where n is the matrix dimension and b is the bandwidth. For solving the decomposed triangular matrix, the system uses a specialized striped forward / backward replacement algorithm, which also avoids operations on out-of-band zero elements. Furthermore, the system implements parallel solution strategies for striped matrices, such as the wavefront method and domain decomposition, further improving computational efficiency. These optimizations make the LU decomposition and solution speed of striped matrices several to tens of times faster than general sparse algorithms, especially for cases with small bandwidth.

[0108] An arrow matrix is ​​a matrix whose non-zero elements are concentrated in certain rows and columns, forming an "arrow" shape. It is typically represented by the main diagonal plus a dense array of the last few rows and columns. This structure is common in hierarchical problems, boundary constraint problems, or reduced systems. The forward-backward replacement optimization algorithm is specifically designed for arrow-shaped structures, leveraging its unique pattern to achieve efficient solutions. The key idea of ​​the algorithm is to separate the solution of the main diagonal block (the body of the arrow) and the boundary block (the arrowhead). The specific steps include: first, representing the arrow matrix A as a 2×2 block matrix; then, using the block LU decomposition formula, calculating the LU decomposition and Schur complement of the diagonal block separately; finally, solving the entire system using block forward / backward replacement. Since the main diagonal block is usually sparse and may even be diagonal, its decomposition can be performed efficiently; while the Schur complement is usually a small-scale dense matrix, which can be handled using optimized dense matrix routines. This separation strategy avoids the overhead of performing general sparse operations on the entire matrix, making it particularly suitable for cases where the dimension of the "arrow" part (boundary block) is much smaller than the entire matrix. The system also implements a parallel version, where the processing of the main diagonal blocks can be fully parallelized, with only the final Schur complement solution requiring centralized processing. Through this structure-aware optimization, the solution speed for arrow-shaped matrices is much faster than the general method, especially when the "arrow" portion is small.

[0109] The Multifrontal Method (MFM) is an efficient and general-purpose direct solution method for sparse matrices, applicable to various sparse patterns, especially those lacking regular structure. This algorithm organizes computation based on an elimination tree, achieving efficient decomposition through the concept of a frontal matrix. The frontal matrix is ​​a small, dense matrix representing the set of active variables in the current elimination step. The MFM workflow comprises several main stages: first, an elimination tree is constructed by analyzing the sparse patterns of the matrix; then, the elimination tree is traversed bottom-up, constructing and processing the frontal matrix for each node; finally, the elimination tree is traversed top-down to complete the solution process. In constructing and processing the frontal matrix, the system uses highly optimized dense matrix computation routines, such as BLAS (Basic Linear Algebra Subprograms) or their custom implementations, fully leveraging the capabilities of modern processors. A key advantage of the MFM is its natural parallelism: different branches of the elimination tree can be processed in parallel, making it particularly suitable for multi-core and distributed environments. The system implements several optimization techniques to improve the performance of the multi-front algorithm: supernode merging combines columns with similar sparse patterns into supernodes, reducing the number of front matrices; front matrix compression removes fully updated rows / columns, reducing memory requirements; dynamic scheduling dynamically allocates tasks based on the size of the front matrix and computational cost; and mixed-precision computation uses reduced precision to accelerate some computations where permissible. These optimizations make the multi-front algorithm an efficient choice for handling general sparse matrices, capable of handling various complex sparse patterns while maintaining good numerical stability and parallel scalability.

[0110] In the practical implementation of hybrid parallel sparse LU decomposition, the system integrates these parallel strategies and structure optimization algorithms to form a complete processing flow. First, the system analyzes the structural characteristics of the input matrix to determine the most suitable solution strategy: a dedicated strip solver is used for striped matrices, a forward-backward replacement optimization algorithm is used for arrow-shaped matrices, and a multiple front algorithm is used for other general sparse matrices. Then, the system applies appropriate parallel strategies: task-level parallel decomposition computation, data-level parallel optimization of local computation, and heterogeneous computing to accelerate suitable subtasks. The combination of these strategies and algorithms forms a highly optimized hybrid parallel framework capable of adaptively handling sparse matrices of various sizes and structures. Finally, the system obtains the decomposed L matrix (lower triangular matrix) and U matrix (upper triangular matrix), both of which satisfy A=LU, providing a foundation for subsequent linear system solutions.

[0111] Practice has shown that this hybrid parallel sparse LU decomposition method performs exceptionally well when processing large-scale risk control feature data. For a typical-sized risk control feature covariance matrix (dimension...),... With a sparsity exceeding 99%, the computation speed is 10-100 times faster than traditional methods, while memory requirements are reduced by 60-80%. The system can decompose a million-dimensional sparse matrix within seconds to minutes, meeting the needs of risk control systems for rapid updates of statistics. This high-performance computing capability is crucial for real-time risk control decision-making and large-scale feature engineering, enabling the system to quickly adjust when feature distributions change, maintaining the effectiveness and accuracy of the model.

[0112] Beyond its performance advantages, hybrid parallel sparse LU decomposition also offers excellent scalability and adaptability. As the number of processor cores increases, the system's processing power increases almost linearly; with the emergence of new hardware (such as more advanced GPUs or dedicated accelerators), the system can easily adapt by updating the corresponding computing cores. This scalability and adaptability ensures that the system can keep pace with the growth of data volume and the development of hardware technology, providing a solid foundation for the long-term evolution of risk control systems.

[0113] Example 2 like Figure 2 As shown, this embodiment provides a risk control model construction system based on feature normalization enhancement, including: a data preprocessing module 10, used to acquire multi-source heterogeneous data from multiple platforms, clean, convert formats, and perform quality verification on the multi-source heterogeneous data to obtain structured data; a feature fusion module 20, used to perform feature extraction, feature selection, and high-order feature engineering on the structured data, wherein the high-order feature engineering includes the construction of combined features, cross features, and domain features, and the fusion of multiple extracted features to obtain transformed and fused feature data; a dynamic adaptation module 30, used to configure a normalization strategy for each type of feature data and set differentiated sliding window parameters based on the transformed and fused feature data, calculate the mean, standard deviation, and quantile values ​​within the differentiated sliding window, and establish a dynamic adaptation mechanism; and a parameter update module 40, used to update the parameter based on the dynamic adaptation mechanism and PS. The I-index or KL divergence detects the degree of distribution shift in feature data. Based on the degree of distribution shift, a preset hierarchical parameter update strategy is adopted to update the feature data, resulting in updated dynamic feature parameters. The feature normalization module 50 is used to classify the updated dynamic feature parameters into enumerated features and numerical features. Label value mapping is performed on the enumerated features, and extreme value quantile mapping, interval mapping, and standard deviation correction are performed on the numerical features to obtain multi-granularity normalized features. The feedback optimization module 60 calculates KS value or IV value in real time based on the multi-granularity normalized features to monitor feature discrimination. When the discrimination decreases, the sliding window parameter adjustment or normalization granularity switching is triggered to obtain high discrimination features after feedback. The model building module 70 is used to train and evaluate the performance of the risk control model using the high discrimination features to obtain an enhanced risk control model.

[0114] Example 3 In this embodiment, the risk control model construction method based on feature normalization enhancement is executed online in a computer risk control system. The risk control system includes a data acquisition module, a feature processing module, a model inference module, and a risk control execution module, all of which work collaboratively via a bus or network connection.

[0115] During system operation, the data acquisition module continuously receives real-time data streams from multiple business platforms, including transaction behavior data, user attribute data, and device behavior data. This data is then cleaned, format-converted, and quality-verified before being input into the feature processing module.

[0116] The feature processing module performs feature extraction and high-order feature engineering on the structured data to generate a feature set for characterizing risk behavior patterns. Through correlation analysis and feature compression, it selects input features suitable for the operation of the online risk control system.

[0117] To ensure the stability of the risk control system during long-term operation, the feature processing module maintains feature statistical parameters based on a differentiated sliding window. When a shift in feature distribution is detected, the statistical parameters are updated hierarchically to dynamically adjust the feature normalization strategy, ensuring that the model input features always remain within a preset numerical range.

[0118] Based on features processed through multi-granularity normalization, the model inference module trains and evaluates the performance of the risk control model, resulting in an enhanced risk control model. This enhanced risk control model is deployed into the risk control system and, during system operation, is used to perform risk identification and control operations on the objects to be processed, including intercepting high-risk behaviors, alerting suspicious behaviors, or adjusting risk levels.

[0119] Through the above methods, this embodiment improves the stability and reliability of the risk control model in the online operating environment without increasing the system's computational load, avoids the model performance degradation problem caused by feature distribution drift, and thus improves the overall operating effect of the risk control system.

[0120] Example 4 This application also provides a method for constructing a risk control model based on feature normalization enhancement, including: Obtain multi-source heterogeneous data, and perform cleaning, format conversion, and quality verification on the multi-source heterogeneous data to obtain structured data; The structured data is subjected to feature extraction, feature selection, and high-order feature engineering. Multiple extracted features are then fused to obtain transformed and fused feature data. Based on the transformed and fused feature data, normalization strategies are configured for different types of feature data during model operation, and differential sliding window parameters are set. Statistics within the sliding window are calculated, and a dynamic adaptation mechanism for controlling the distribution of model input features is established. Based on the dynamic adaptation mechanism, the degree of shift in the distribution of feature data is detected during model operation, and the feature parameters are updated hierarchically according to the degree of distribution shift to obtain the updated dynamic feature parameters. Multi-granularity feature normalization is performed on the updated dynamic feature parameters to keep the model input features within a preset numerical range. Based on the multi-granularity normalized features, the risk control model is trained and its performance is evaluated to obtain an enhanced risk control model. The enhanced risk control model is deployed in the risk control system to perform risk identification, risk disclosure, and risk control operations on the objects to be processed.

[0121] Furthermore, high-order feature engineering is performed on the structured data, including: Statistical features, time-series features, and relational features are constructed for the structured data; Based on the statistical features, time-series features and relational features, four arithmetic operations are combined, ratios are calculated and cross features are constructed to generate high-order features for characterizing risk behavior patterns. The higher-order features are correlated and fused across data sources to form a multi-dimensional feature matrix; Correlation analysis is performed on the multi-dimensional feature matrix to remove redundant features with correlations higher than a preset threshold, resulting in transformed and fused feature data for training the risk control model.

[0122] Furthermore, feature compression and selection are performed on the higher-order features, including: A feature selection method with regularization constraints is applied to the high-order features to compress the high-dimensional features and obtain a low-dimensional sparse feature representation. Based on the low-dimensional sparse feature representation, features that contribute more than a preset threshold to risk identification are retained as input features of the risk control model.

[0123] Furthermore, establishing a dynamic adaptation mechanism also includes: Within the differentiated sliding window, the mean, standard deviation, and multiple quantile statistics are calculated for the feature data. Based on the statistics, the feature range is segmented and modeled to generate mapping parameters for feature normalization. The original features are mapped to a preset standard numerical range using the mapping parameters to improve the stability of the model input features during system operation.

[0124] Furthermore, in this embodiment, the risk control model construction method based on feature normalization enhancement is executed online in a computer risk control system. The risk control system is deployed in a server or cloud computing environment and includes a data acquisition module, a feature processing module, a model training and inference module, and a risk control execution module. Each module operates collaboratively through a processor and memory to perform risk identification and risk control operations on the objects to be processed during system operation.

[0125] During the operation of the risk control system, the data acquisition module continuously acquires multi-source heterogeneous data from multiple business platforms. This multi-source heterogeneous data includes transaction behavior data, user attribute data, device behavior data, and external risk data. After the data is accessed via an interface, it first undergoes data cleaning, format conversion, and quality verification to eliminate outliers, duplicate data, and format inconsistencies, thereby obtaining structured data, which serves as input data for subsequent feature processing.

[0126] The feature processing module performs feature extraction and high-order feature engineering on the structured data. Specifically, it constructs statistical features, time-series features, and relational features based on the structured data. Statistical features reflect the central tendency and dispersion of the data, time-series features depict patterns of behavior over time, and relational features describe the relationships between accounts or devices. Based on this, high-order features are generated to characterize risk behavior patterns by performing arithmetic operations, ratio calculations, and cross-feature construction on these features. Subsequently, these high-order features are fused across data sources to form a multi-dimensional feature matrix. Redundant features with correlations exceeding a preset threshold are removed through correlation analysis, resulting in transformed and fused feature data for training the risk control model.

[0127] To meet the real-time and computational resource requirements of online risk control systems, this embodiment further performs feature compression and selection processing on the high-order features. Specifically, a feature selection method with regularization constraints is used to compress the high-dimensional features, mapping them to low-dimensional sparse feature representations. Based on the evaluation results of the features' contribution to risk identification, features with a contribution higher than a preset threshold are retained as input features of the risk control model, thereby reducing the system's computational load and improving the model's stability in the online operating environment.

[0128] During model operation, the feature processing module establishes a dynamic adaptation mechanism based on the input features. Specifically, it sets differentiated sliding window parameters according to the data type and update frequency of different features, and calculates the mean, standard deviation, and multiple quantile statistics for the feature data within the sliding window. Based on these statistics, it performs segmented modeling of the feature value range, generates mapping parameters for feature normalization, and uses these mapping parameters to map the original features to a preset standard numerical range, thereby controlling the numerical range of the model's input features.

[0129] When the risk control system detects continuous input of new data during operation, the feature processing module monitors the feature distribution within the sliding window and determines whether the feature distribution has changed based on a preset distribution offset detection index. When the detected feature distribution offset reaches a preset threshold, the system performs hierarchical updates to the statistical parameters, thereby dynamically adjusting the feature normalization strategy. This allows the model input features to adaptively adjust with changes in data distribution, avoiding model input imbalance or unstable inference results due to feature distribution drift.

[0130] Based on the input features after the above multi-granularity feature normalization processing, the model training and inference module trains and evaluates the performance of the risk control model to obtain an enhanced risk control model. This enhanced risk control model is deployed in the risk control system and, during system operation, is used to perform risk identification, risk disclosure, and risk control operations on the objects to be processed, including but not limited to intercepting high-risk behaviors, providing early warnings for suspicious behaviors, and dynamically adjusting risk levels.

[0131] Through the above implementation methods, this embodiment achieves dynamic control and adaptive adjustment of the model input feature distribution during the online operation of the computer risk control system. Without relying on frequent model retraining, it improves the stability and reliability of the risk control model in the long-term operating environment, thereby effectively enhancing the overall effect of the risk control system in performing risk control operations.

Claims

1. A risk control model construction method based on feature normalization enhancement, characterized in that, include: Acquire multi-source heterogeneous data from multiple platforms, including transaction behavior data, user attribute data, device behavior data, and external cooperation data from multiple business platforms. Clean, convert, and verify the quality of the multi-source heterogeneous data to obtain structured data. The structured data is subjected to feature extraction, feature selection, and higher-order feature engineering. The higher-order feature engineering includes the construction of combined features, cross features, and domain features. Multiple extracted features are fused to obtain transformed and fused feature data. Based on the transformed and fused feature data, a normalization strategy is configured for each type of feature data, and a differential sliding window parameter is set. The mean, standard deviation, and quantile values ​​within the differential sliding window are calculated, and a dynamic adaptation mechanism is established. This includes: classifying the transformed and fused feature data into continuous numerical features, sparse distribution features, skewed distribution features, and enumerated features; configuring a Z-score normalization strategy or a Min-Max normalization strategy for the continuous numerical features; configuring a robust quantile normalization strategy for the sparse distribution features; configuring a logarithmic transformation normalization strategy for the skewed distribution features; and configuring a label mapping strategy for the enumerated features, thus obtaining the normalization strategy configuration; based on the update frequency of the feature data, the feature data is reclassified into high-frequency changing feature data, periodic feature data, and stable feature data; setting a 7-day sliding window for high-frequency changing feature data, a 30-day sliding window for periodic feature data, and a 9-day sliding window for stable feature data. A 0-day sliding window is used to obtain differentiated sliding window parameters. High-frequency change features include recent login frequency and transaction amount over the last 7 days; periodic features include weekday / weekend activity differences and beginning / end month consumption patterns; and stable features include user basic attributes and long-term behavioral preferences. The mean, standard deviation, maximum, minimum, 25th percentile, 50th percentile, 75th percentile, and 95th percentile values ​​are calculated for the feature data. Based on these values ​​and the differentiated sliding window parameters, initial statistics are calculated. Based on these initial statistics, KS, IV, and AUC values ​​are constructed as discrimination indices for the feature data. PSI and KL divergence stability indices are also constructed. Outlier and missing rate anomaly detection indices are constructed. A dynamic adaptation mechanism is established based on these discrimination, stability, and anomaly detection indices. During the online operation of the computer risk control system, based on the dynamic adaptation mechanism, the distribution offset of feature data is detected based on the PSI index or KL divergence, and the feature data is updated according to the distribution offset by a preset hierarchical parameter update strategy to obtain the updated dynamic feature parameters. The updated dynamic feature parameters are classified into enumerated features and numerical features. Label value mapping is performed on the enumerated features, and extreme value quantile mapping, interval mapping and standard deviation correction are performed on the numerical features to obtain multi-granularity normalized features. Based on the multi-granularity normalized features, KS values ​​or IV values ​​are calculated in real time to monitor feature discrimination. When the discrimination decreases, the sliding window parameters are adjusted or the normalization granularity is switched to obtain high discrimination features after feedback. The high-discrimination features are used to train and evaluate the risk control model, resulting in an enhanced risk control model. The enhanced risk control model is deployed in a computer risk control system so that risk identification, risk disclosure, and risk control operations are performed on the objects to be processed through the enhanced risk control model.

2. The method according to claim 1, characterized in that, The process involves acquiring heterogeneous data from multiple platforms, cleaning, format conversion, and quality verification to obtain structured data. This includes: developing a data collection strategy based on the data dimensions required by the risk control model; generating a list of data sources from different platforms based on the data collection strategy; constructing API call interfaces, database connection interfaces, and log collection interfaces based on the data source list; collecting transaction behavior data, user attribute data, device behavior data, and external cooperation data from different platforms through these interfaces; and sending the collected data to a data processing platform for storage to obtain raw heterogeneous data; performing structured analysis on the raw heterogeneous data to identify data types, formats, and distribution characteristics; establishing a data dictionary based on these data types, formats, and distribution characteristics to obtain data structure information; and performing missing value imputation, outlier correction, and duplicate data removal operations on the data structure information to obtain cleaned data. The cleaned data is subjected to time format unification, enumeration value encoding, and text data structuring to obtain format-standardized data; the completeness, accuracy, and consistency indicators of the format-standardized data are calculated, and if all indicators reach the expected threshold, the structured data is obtained.

3. The method according to claim 1, characterized in that, The structured data undergoes feature extraction, feature selection, and high-order feature engineering. The high-order feature engineering includes constructing combined features, cross-features, and domain features. Multiple extracted features are fused to obtain transformed and fused feature data. This process includes: constructing statistical features (mean, variance, maximum, minimum, and quantiles) based on the structured data; constructing time-series features (rate of change, year-on-year growth, and sliding window statistics); and constructing relational features (network centrality and number of associated accounts) to obtain a basic feature set. Based on the basic feature set, arithmetic operations, ratio calculations, and cross-feature construction are performed to obtain a high-order feature set. The high-order feature set is then cross-source correlated and fused to obtain a multi-dimensional feature matrix. Finally, a correlation matrix index is calculated between features in the multi-dimensional feature matrix, and features with correlation matrix indices exceeding a preset correlation threshold are removed to obtain the transformed and fused feature data.

4. The method according to claim 1, characterized in that, Based on the dynamic adaptation mechanism, the distribution shift of feature data is detected by the PSI index or KL divergence. A preset hierarchical parameter update strategy is then applied to update the feature data according to the distribution shift, resulting in updated dynamic feature parameters. This includes: When new feature data flows in, the dataset within the sliding window is updated according to the differentiated sliding window parameters, while maintaining a constant sliding window size, resulting in updated window data; Based on the updated window data, the PSI index or KL divergence is calculated to obtain a distribution shift detection index; When the PSI index is greater than 0.1, the mean, standard deviation, and quantile statistics are recalculated based on the updated window data; when the PSI index is greater than 0.05 and less than or equal to 0.1, the quantile statistics of the updated window data are calculated using an exponential moving average method; when the PSI index is less than or equal to 0.05, the existing window data remains unchanged, resulting in hierarchically updated parameters; The hierarchically updated parameters are verified by calculating the change in the discriminant strength of the feature data before and after the update. When the discriminant strength decreases beyond a preset discriminant strength threshold, the feature data before the hierarchical update is used as the updated dynamic feature parameters.

5. The method according to claim 1, characterized in that, The updated dynamic feature parameters are classified into enumerated features and numerical features. Label value mapping is performed on the enumerated features, and extreme value quantile mapping, interval mapping, and standard deviation correction are performed on the numerical features to obtain multi-granularity normalized features. This includes: classifying the updated dynamic feature parameters into enumerated features and numerical features; establishing a business semantic mapping dictionary for the enumerated features, mapping enumerated values ​​to risk levels between 0 and 1; mapping high-cardinality enumerated features to numerical values ​​associated with the target variable using a target encoding method; and mapping time-based enumerated features using a periodic encoding method to obtain the label-mapped enumerated features. Based on the uniformity of the distribution of numerical features, the numerical features are classified into uniform numerical features and non-uniformly distributed numerical features. For the uniform numerical features, extreme values ​​and key quantiles are calculated. Based on the extreme values ​​and key quantiles, anchor quantiles are used for normalization mapping to obtain standard normalized numerical features. For the non-uniform numerical features, their feature value range is divided into multiple intervals. Based on the boundary values ​​and anchor quantiles of each interval, each interval is mapped separately to obtain interval-mapped features. The standard deviation values ​​of enumerated features and numerical features are detected through a monitoring window. When the standard deviation of a feature is detected to be greater than a preset standard deviation threshold, the feature mean of the monitoring window and the global feature mean are calculated. The enumerated features and numerical features are corrected based on the ratio of the feature mean within the monitoring window to the global feature mean to obtain standard deviation-corrected features. The enumerated features after label mapping, the standard normalized numerical features, the interval-mapped features, and the standard deviation-corrected features are fused to obtain the multi-granularity normalized features.

6. The method according to claim 3, characterized in that, Based on the basic feature set, arithmetic operations, ratio calculations, and cross-feature construction are performed to obtain a higher-order feature set, including: constructing semantic representations for each feature data in the basic feature set to obtain semantic representation features; constructing the correlation relationships between features using graph structures based on the semantic representation features, calculating the correlation matrix and conditional dependency probabilities between features to obtain feature association information; constructing high-dimensional dense context-enhanced features based on the feature association information and business scenario context information; compressing and selecting features from the context-enhanced features using LASSO and Elastic Net sparse coding methods to obtain sparse-coded features; performing nonlinear transformations on the sparse-coded features and introducing group sparsity and structural sparsity constraints to obtain a sparsified enhanced feature set; and performing arithmetic operations, ratio calculations, and cross-feature construction on the sparsified enhanced feature set to obtain the higher-order feature set.

7. The method according to claim 6, characterized in that, The context enhancement features are compressed and selected using LASSO and Elastic Net sparse coding methods to obtain sparse coded features. This includes: applying LASSO regularization and Elastic Net regularization to the context enhancement features, converting the high-dimensional, dense context enhancement features into a low-dimensional, sparse representation, resulting in sparsely coded features; and performing feature dimensionality reduction and selection on the sparsely coded features, retaining the features with the most information, to obtain the sparse coded features.

8. The method according to claim 1, characterized in that, The mean, standard deviation, maximum and minimum values, 25th quantile, 50th quantile, 75th quantile, and 95th quantile values ​​are calculated for the feature data. Based on these values ​​and the differential sliding window parameters, initialization statistics are calculated, including: treating each feature data as a dimension in a continuous space within the feature space, and defining the feature value range as a continuous interval to obtain the definition of the continuous feature space; dividing the continuous space into multiple subdomains based on the definition of the continuous feature space, with the boundaries of each subdomain determined by the 5th, 25th, 50th, 75th, and 95th quantiles to obtain the domain discretization result; selecting a first-order linear basis function, a second-order polynomial basis function, or an exponential-logarithmic basis function as the interpolation basis function for each subdomain to obtain the basis function selection scheme; and calculating initialization parameters at the quantiles of each subdomain based on the basis function selection scheme to obtain the node parameter initialization result. Based on the node parameter initialization results, the local equations of each subdomain are assembled into a global equation, and the global equation is solved to obtain the coefficient values ​​at each quantile, thus obtaining the continuum model coefficients; based on the continuum model coefficients and the differential sliding window parameters, the initialization statistics are calculated using the continuum embedding mapping method.

9. The method according to claim 8, characterized in that, Based on the node parameter initialization results, the local equations of each subdomain are assembled into a global equation, and the global equation is solved to obtain the coefficient values ​​at each quantile point, thus obtaining the continuum model coefficients. This includes: for each feature data to be processed, determining its subdomain in the continuous space, converting the feature data into the local coordinate system of the corresponding subdomain to obtain local coordinates; calculating the shape function value of the local coordinates using a preset interpolation basis function, calculating the mapping value of the feature data in the continuous space based on the shape function value and the quantile parameters to obtain the continuum embedding mapping value; standardizing the continuum embedding mapping value to a target interval between 0 and 1 to obtain the processed mapping value; calculating the mapping error in each subdomain, performing mesh refinement on subdomains with errors exceeding a threshold, recalculating the mapping value for the refined subdomains to obtain high-precision feature mapping values; and obtaining the continuum model coefficients based on the continuum embedding mapping value and the high-precision feature mapping value.

10. The method according to claim 4, characterized in that, When the PSI index is greater than 0.1, the mean, standard deviation, and quantile statistics are recalculated based on the updated window data, including: converting the updated window data into a sparse matrix representation and recording the positions and values ​​of non-zero elements to obtain a sparse data representation; calculating the covariance values ​​between feature data based on the sparse data representation, setting covariance values ​​less than a preset threshold to zero to obtain a sparse matrix representation of the feature covariance; processing the sparse matrix representation of the feature covariance, including using nested dissipative sorting or approximate minimum degree sorting to reduce padding, merging columns with the same non-zero pattern to form a supernode structure, and performing symbolic analysis to predict padding positions to obtain a processed sparse matrix; performing hybrid parallel sparse LU decomposition on the processed sparse matrix, decomposing the LU decomposition task into mutually independent subtasks, with multiple subtasks executed in parallel on multiple processors to obtain the decomposed L matrix and U matrix; and recalculating the mean, standard deviation, and quantile statistics based on the L matrix and the U matrix.

11. The method according to claim 10, characterized in that, The processed sparse matrix is ​​subjected to hybrid parallel sparse LU decomposition, which decomposes the LU decomposition task into independent subtasks. These subtasks are executed in parallel on multiple processors to obtain the decomposed L and U matrices. This process includes: 1) Decomposing the LU decomposition task into multiple independent subtasks using a preset task-level parallel strategy, executing these subtasks in parallel on a multi-core CPU to obtain task-level parallel processing results; 2) Applying a preset data-level parallel strategy to the large, dense subtasks decomposed from the LU decomposition task using SIMD (Single Instruction Multiple Data) instruction set to obtain data-level parallel processing results; 3) Using a preset heterogeneous computing acceleration strategy, transferring computationally intensive subtasks to the GPU for processing, while leaving communication-intensive subtasks on the CPU for processing, to obtain heterogeneous computing accelerated processing results; 4) Based on the task-level parallel processing results, the data-level parallel processing results, and the heterogeneous computing accelerated processing results, a dedicated strip matrix solver is used for strip structures, a forward-backward replacement optimization algorithm is used for arrow-shaped structures, and a multiple frontier algorithm is used for general sparse structures to obtain the decomposed L and U matrices.

12. A risk control model construction method based on feature normalization enhancement, characterized in that, include: Acquire multi-source heterogeneous data, including transaction behavior data, user attribute data, device behavior data, and external cooperation data from multiple business platforms. Clean, convert, and verify the quality of the multi-source heterogeneous data to obtain structured data. The structured data is subjected to feature extraction, feature selection, and high-order feature engineering. Multiple extracted features are then fused to obtain transformed and fused feature data. Based on the transformed and fused feature data, normalization strategies are configured for different types of feature data during model operation, and differential sliding window parameters are set. Statistics within the sliding window are calculated, and a dynamic adaptation mechanism for controlling the distribution of model input features is established. This includes: classifying the transformed and fused feature data into continuous numerical features, sparse distribution features, skewed distribution features, and enumerated features; configuring Z-score normalization or Min-Max normalization strategies for continuous numerical features; configuring robust quantile normalization strategies for sparse distribution features; configuring logarithmic transformation normalization strategies for skewed distribution features; and configuring label mapping strategies for enumerated features, thus obtaining the normalization strategy configuration; based on the feature data update frequency, the feature data is reclassified into high-frequency changing feature data, periodic feature data, and stable feature data; setting a 7-day sliding window for high-frequency changing feature data, a 30-day sliding window for periodic feature data, and a stable feature data... The feature data is set with a 90-day sliding window to obtain differentiated sliding window parameters. High-frequency changing feature data includes recent login frequency and transaction amount in the last 7 days; periodic feature data includes weekday / weekend activity differences and beginning / end month consumption patterns; and stable feature data includes user basic attributes and long-term behavioral preferences. The mean, standard deviation, maximum, minimum, 25th percentile, 50th percentile, 75th percentile, and 95th percentile values ​​are calculated for the feature data. Based on these values ​​and the differentiated sliding window parameters, initial statistics are calculated. Based on these initial statistics, KS, IV, and AUC values ​​are constructed as discrimination indices for the feature data; PSI and KL divergence stability indices are constructed; and outlier and missing rate anomaly detection indices are constructed. A dynamic adaptation mechanism is established based on these discrimination, stability, and anomaly detection indices. Based on the dynamic adaptation mechanism, the degree of shift in the distribution of feature data is detected during model operation, and the feature parameters are updated hierarchically according to the degree of distribution shift to obtain the updated dynamic feature parameters. Multi-granularity feature normalization is performed on the updated dynamic feature parameters to keep the model input features within a preset numerical range. Based on the multi-granularity normalized features, the risk control model is trained and its performance is evaluated to obtain an enhanced risk control model. The enhanced risk control model is deployed in a computer risk control system to perform risk identification, risk disclosure, and risk control operations on the objects to be processed.

13. The method according to claim 12, characterized in that, The high-order feature engineering of the structured data includes: constructing statistical features, time-series features, and relational features for the structured data; performing arithmetic operations, ratio calculations, and cross-feature construction based on the statistical features, time-series features, and relational features to generate high-order features for characterizing risk behavior patterns; performing cross-data source correlation and fusion on the high-order features to form a multi-dimensional feature matrix; and performing correlation analysis on the multi-dimensional feature matrix to remove redundant features with correlations higher than a preset threshold, obtaining transformed and fused feature data for risk control model training.

14. The method according to claim 13, characterized in that, The high-order features are compressed and selected, including: using a feature selection method with regularization constraints to compress the high-dimensional features and obtain a low-dimensional sparse feature representation; based on the low-dimensional sparse feature representation, retaining features that contribute more than a preset threshold to risk identification as input features of the risk control model.

15. The method according to claim 12, characterized in that, The dynamic adaptation mechanism also includes: calculating the mean, standard deviation, and multiple quantile statistics for the feature data within the differentiated sliding window; performing segmented modeling of the feature value domain based on the statistics to generate mapping parameters for feature normalization; and using the mapping parameters to map the original features to a preset standard value range to improve the stability of the model input features during system operation.

16. A risk control model construction system based on feature normalization enhancement, characterized in that, For performing the method according to any one of claims 1 to 11, comprising: The data preprocessing module is used to acquire multi-source heterogeneous data from multiple platforms. The multi-source heterogeneous data includes transaction behavior data, user attribute data, device behavior data, and external cooperation data from multiple business platforms. The multi-source heterogeneous data is cleaned, format converted, and quality verified to obtain structured data. The feature fusion module is used to perform feature extraction, feature selection, and higher-order feature engineering on the structured data. The higher-order feature engineering includes the construction of combined features, cross features, and domain features, and the fusion of multiple extracted features to obtain transformed and fused feature data. The dynamic adaptation module is used to configure a normalization strategy and set differential sliding window parameters for each type of feature data based on the transformed and fused feature data, calculate the mean, standard deviation, and quantile values ​​within the differential sliding window, and establish a dynamic adaptation mechanism. This includes: classifying the transformed and fused feature data into continuous numerical features, sparse distribution features, skewed distribution features, and enumerated features; configuring a Z-score normalization strategy or a Min-Max normalization strategy for continuous numerical features; configuring a robust quantile normalization strategy for sparse distribution features; configuring a logarithmic transformation normalization strategy for skewed distribution features; and configuring a label mapping strategy for enumerated features, thus obtaining a normalization strategy configuration; and reclassifying the feature data into high-frequency changing feature data, periodic feature data, and stable feature data based on the feature data update frequency; setting a 7-day sliding window for high-frequency changing feature data, a 30-day sliding window for periodic feature data, and a 30-day sliding window for stable feature data. A 90-day sliding window is set to obtain differentiated sliding window parameters. High-frequency change features include recent login frequency and transaction amount in the last 7 days; periodic features include weekday / weekend activity differences and beginning / end month consumption patterns; and stable features include user basic attributes and long-term behavioral preferences. The mean, standard deviation, maximum, minimum, 25th percentile, 50th percentile, 75th percentile, and 95th percentile values ​​are calculated for the feature data. Based on these values ​​and the differentiated sliding window parameters, initial statistics are calculated. Based on these initial statistics, KS, IV, and AUC values ​​are constructed as discrimination indices for the feature data; PSI and KL divergence stability indices are constructed; and outlier and missing rate anomaly detection indices are constructed. A dynamic adaptation mechanism is established based on these discrimination, stability, and anomaly detection indices. The parameter update module is used to detect the degree of feature data distribution shift based on the dynamic adaptation mechanism, the PSI index or KL divergence, and update the feature data according to the degree of distribution shift by adopting a preset hierarchical parameter update strategy to obtain the updated dynamic feature parameters. The feature normalization module is used to classify the updated dynamic feature parameters into enumerated features and numerical features, perform label value mapping on the enumerated features, and perform extreme value quantile mapping, interval mapping and standard deviation correction on the numerical features to obtain multi-granularity normalized features. The feedback optimization module is used to calculate the KS value or IV value in real time based on the multi-granularity normalized features to monitor the feature discrimination. When the discrimination decreases, it triggers the adjustment of sliding window parameters or the switching of normalization granularity to obtain high discrimination features after feedback. The model building module is used to train and evaluate the performance of the risk control model using the high-discrimination features to obtain an enhanced risk control model, and to deploy the enhanced risk control model in the computer risk control system so that risk identification, risk disclosure and risk control operations are performed on the object to be processed through the enhanced risk control model.

Images