Number monitoring method and device

Abstract

The invention provides a number monitoring method and device. The method comprises the following steps: generating a training sample set; based on each number contained in the training sample set, extracting historical call detailed list data of each number from a communication recording device storing historical call records of the numbers, and generating a number behavior data set; performing feature extraction and processing on the number behavior data set from three dimensions of number features, call behaviors and call behavior changes to generate a number feature variable set; based on the number characteristic variable set, through an XGBoost algorithm, constructing a number stability pre-judgment initial model, and generating a number stability prediction model; deploying the number stability prediction model to a service scene, obtaining yellow page number data in a number library, and performing stability pre-judgment on the yellow page number data by using the number stability prediction model to generate a number stability probability value; and updating a number library based on the number stability probability value. According to the embodiment of the invention, efficient maintenance of the number library is realized.

Description

Technical Field

[0001] The present application relates to the technical field of telecommunication data analysis, and in particular to a number monitoring method and device. Background Art

[0002] With the development of telecommunications technology and the growth of user demand, the construction and maintenance of various commercial number databases have become increasingly important. These number databases usually contain a large number of merchant contact numbers, which are used in a variety of scenarios such as customer service, marketing promotion, and industry information query.

[0003] At present, common number database maintenance technologies mainly rely on periodic manual verification or basic crawler methods for updating. For example, manually confirming the validity of the number by making a phone call, or using web crawlers to collect and update information one by one according to the data source. Although these methods can ensure a certain degree of data accuracy, they face the problems of long update cycles and low efficiency.

[0004] The current more advanced number database maintenance technology uses a simple regular batch verification strategy, that is, to periodically verify all numbers in the database without distinction. This technology verifies the validity of the number through manual or automatic outbound calling devices, and updates the database based on the verification results. This method can ensure a certain timeliness of the data, but it cannot accurately identify and predict the change characteristics of different numbers.

[0005] However, this indiscriminate verification method has obvious technical flaws: verifying all numbers with the same frequency consumes a lot of manpower and material resources, resulting in a waste of resources; due to the long update cycle, number changes cannot be discovered and processed in a timely manner, resulting in a large amount of outdated or erroneous information in the database, affecting the overall accuracy; it is impossible to perform intelligent management based on the actual usage and change patterns of the numbers, and lacks targeted maintenance strategies. Summary of the invention

[0006] In view of this, the present application provides a number monitoring method and device, which solves the problem in the prior art that verifying all numbers with the same frequency consumes a lot of manpower and material resources, has a long update cycle, and cannot timely discover and handle number changes.

[0007] The embodiment of the present application provides a number monitoring method, including: collecting outbound call verification data results, user error correction data, and positive samples and negative sample data in recovered number data to generate a training sample set; based on each number included in the training sample set, extracting historical call detail data of each number from a communication recording device that stores historical call records of the number, and preprocessing the historical call detail data to generate a number behavior data set; extracting and processing features of the number behavior data set from three dimensions of number features, call behavior, and call behavior changes to generate a number feature variable set; based on the number feature variable set, using the XGBoost algorithm, constructing an initial model for predicting number stability, and evaluating and tuning the prediction results of the initial model for predicting number stability to generate a number stability prediction model; deploying the number stability prediction model to a business scenario, obtaining yellow page number data in a number library, and using the number stability prediction model to predict the stability of the yellow page number data to generate a number stability probability value; based on the number stability probability value, focusing on verifying numbers above a preset probability threshold, and updating the number library according to the verification results.

[0008] Optionally, based on each number included in the training sample set, historical call details data of each number is extracted from a communication recording device that stores historical call records of the number, and the historical call details data are preprocessed to generate a number behavior data set, including: determining a preset time window range based on a verification time period of the number in the training sample set; for each number in the training sample set, extracting call details records within the preset time window from the communication recording device; performing format unification, missing value processing and outlier filtering on the call details records in turn to form preprocessed call details records; and performing structured processing on the preprocessed call details records to generate the number behavior data set.

[0009] Optionally, the feature extraction and processing of the number behavior data set from three dimensions of number characteristics, call behavior, and call behavior changes to generate a number feature variable set includes: using the number behavior data set to extract the number's city of origin, operator type, and number segment type information from the number feature dimension, and output a number attribute feature set; based on the number attribute feature set, combined with the number behavior data set, from the call behavior dimension, counting the number of calls and the number of call recipients to form a call behavior feature set; using the call behavior feature set to calculate the statistical characteristics of each behavior indicator in the call behavior feature set from the call behavior change dimension, and constructing a call behavior change feature set, the statistical characteristics including at least one of a mean or a variance; and performing feature fusion on the number attribute feature set, the call behavior feature set, and the call behavior change feature set to generate the number feature variable set.

[0010] Optionally, the prediction results of the initial model for predicting number stability are evaluated and parameters are optimized to generate a number stability prediction model, including: dividing the number feature variable set into a training set and a test set according to a preset ratio; using the test set to perform performance evaluation on the initial model for predicting number stability and calculate a performance evaluation index; according to the performance evaluation index, adjusting the parameters of the initial model for predicting number stability and generating the number stability prediction model.

[0011] Optionally, after generating the number stability prediction model, it also includes: inputting the prediction result of the number stability prediction model into a Bayesian confidence reconfigurable streaming accelerator, performing Monte Carlo sampling through the accelerator, and generating an uncertainty distribution of the prediction result; performing a probability calibration calculation according to the uncertainty distribution to form a 90% confidence interval of the prediction result; based on the 90% confidence interval value of the prediction result, using a dynamic weight adjustment mechanism to generate a calibrated and optimized prediction result, and updating the calibrated and optimized prediction result to the number stability prediction model for the number stability prediction process.

[0012] Optionally, deploying the number stability prediction model to a business scenario includes: using a serialization method to convert the number stability prediction model into a model file in a persistent format, and completing the environment configuration on the business server; using the configured business server environment to build a scheduled task scheduler and establish an automated prediction process; based on the automated prediction process, implementing a prediction process monitoring mechanism to continuously record and output prediction status and result distribution data.

[0013] Optionally, after using the number stability prediction model to predict the stability of the yellow pages number data, it also includes: using a large language model based on an active reasoning framework to perform in-depth analysis on the number stability probability value generated by the number stability prediction model to generate a semantic level analysis result; combining the semantic level analysis results, using a multi-objective optimization strategy to output optimized prediction parameters and an optimized verification strategy; based on the optimized prediction parameters and the optimized verification strategy, iteratively optimizing the number stability prediction model to improve the accuracy of number stability prediction.

[0014] Optionally, the large language model based on the active reasoning framework is used to perform in-depth analysis on the number stability probability value generated by the number stability prediction model to generate analysis results at the semantic level, including: obtaining numbers whose number stability probability values ​​are within a preset probability interval, and performing in-depth feature analysis on the numbers to form preliminary analysis results; using the preliminary analysis results, retrieving a historical case database, and analyzing the temporal changes in number usage patterns to generate number change risk features; based on the active reasoning framework, fusing the number change risk features, and simultaneously extracting and processing number-related outbound call record text and merchant status information, and outputting a comprehensive assessment result of number change risk as the analysis result at the semantic level.

[0015] Optionally, the multi-objective optimization strategy is adopted to output optimized prediction parameters and optimized verification strategies, including: collecting prediction accuracy, recognition coverage and resource consumption efficiency data, and establishing a multi-dimensional optimization indicator system; using the multi-dimensional optimization indicator system, calculating the gradient direction of each optimization target, constructing a Pareto optimal solution set, and determining the optimal balance point; based on the optimal balance point, generating updated prediction weights and threshold parameters, and applying the updated parameters to the number stability prediction process.

[0016] The embodiment of the present application also provides a number monitoring device, which includes: a sample collection module, which is used to collect outbound call verification data results, user error correction data, and positive samples and negative sample data in the recovered number data to generate a training sample set; a data processing module, which is used to extract historical call details data of each number from a communication recording device that stores historical call records of the number based on each number included in the training sample set, and pre-process the historical call details data to generate a number behavior data set; a feature extraction module, which is used to extract and process features of the number behavior data set from three dimensions: number features, call behavior, and call behavior changes, to generate a number feature change data set. A quantity set; a model training module, which is used to construct an initial model for predicting number stability based on the number feature variable set through the XGBoost algorithm, and evaluate the prediction results and optimize the parameters of the initial model for predicting number stability to generate a number stability prediction model; a prediction module, which is used to deploy the number stability prediction model to a business scenario, obtain the yellow pages number data in the number library, and use the number stability prediction model to predict the stability of the yellow pages number data to generate a number stability probability value; a verification and update module, which is used to focus on verifying numbers above a preset probability threshold based on the number stability probability value, and update the number library according to the verification results.

[0017] An embodiment of the present application also provides a computer device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the above-mentioned number monitoring method.

[0018] The embodiment of the present application also provides a computer-readable storage medium, which stores computer instructions, which are used to enable a computer to execute the method of the above number monitoring method. The embodiment of the present application also provides a computer program product, including computer instructions, which implement the steps of the above number monitoring method when executed by a processor.

[0019] This application has the following technical effects: by comprehensively capturing the number status characteristics from three dimensions: number features, call behavior, and call behavior changes, the accuracy of the prediction model is improved; the number stability prediction model is trained using the XGBoost algorithm, and the accurate prediction of number change risks is achieved through statistical analysis of number call behavior; a differentiated verification strategy based on probability thresholds is used to focus on monitoring high-risk numbers and significantly reduce the waste of resources caused by indiscriminate verification; a closed-loop device for number library maintenance is constructed, and the automation and intelligent management of number library maintenance is achieved through a cycle mechanism of model prediction, key verification, library update, and feedback optimization. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solutions of the embodiments of the present disclosure, the following is a brief introduction to the drawings required for use in the embodiments. The drawings herein are incorporated into the specification and constitute a part of the specification. These drawings illustrate embodiments consistent with the present disclosure and are used together with the specification to illustrate the technical solutions of the present disclosure. It should be understood that the following drawings only illustrate certain embodiments of the present disclosure and should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can also be obtained based on these drawings without creative work.

[0021] Figure 1 A flowchart of a number monitoring method provided in an embodiment of the present application; Figure 2 A schematic diagram of the process of step S1 provided in an embodiment of the present application; Figure 3 A schematic diagram of the process of step S2 provided in an embodiment of the present application; Figure 4 A schematic diagram of the process of step S3 provided in an embodiment of the present application; Figure 5 A schematic diagram of the process of step S4 provided in an embodiment of the present application. DETAILED DESCRIPTION

[0022] In order to make the purpose, technical scheme and advantages of the embodiments of the present disclosure clearer, the technical scheme in the embodiments of the present disclosure will be clearly and completely described below in conjunction with the drawings in the embodiments of the present disclosure. Obviously, the described embodiments are only part of the embodiments of the present disclosure, rather than all of the embodiments. The components of the embodiments of the present disclosure generally described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present disclosure provided in the drawings is not intended to limit the scope of the present disclosure for protection, but merely represents the selected embodiments of the present disclosure. Based on the embodiments of the present disclosure, all other embodiments obtained by those skilled in the art without making creative work belong to the scope of protection of the present disclosure.

[0023] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, further definition and explanation thereof is not required in subsequent drawings.

[0024] The term "and / or" herein only describes an association relationship, indicating that three relationships may exist. For example, A and / or B may represent the following three situations: A exists alone, A and B exist at the same time, and B exists alone. In addition, the term "at least one" herein represents any combination of at least two of any one or more of a plurality of. For example, including at least one of A, B, and C may represent including any one or more elements selected from the set consisting of A, B, and C.

[0025] like Figure 1 As shown, the embodiment of the present application provides a number monitoring method, including: S1: Collect outbound call verification data results, user error correction data, and positive and negative sample data from recovered number data to generate a training sample set.

[0026] In this step, the number monitoring device comprehensively collects samples from multiple data sources, including outbound call verification data results, user error correction data, and recovered number data, and extracts positive and negative samples from them.

[0027] Positive samples represent stable numbers, and negative samples represent unstable numbers. After cleaning and integration, these multi-source sample data form a reliable and well-labeled training sample set, which provides the necessary data foundation for subsequent machine learning models. This multi-source data fusion collection method ensures the diversity and representativeness of the samples, which helps to improve the generalization ability and prediction accuracy of the final model.

[0028] S2: Based on each number included in the training sample set, extract the historical call detail data of each number from the communication record number monitoring device that stores the historical call records of the number, and pre-process the historical call detail data to generate a number behavior data set.

[0029] The appropriate time window range is determined based on the time period of number verification, and then the call details records within the window are extracted from the communication record number monitoring device. These records contain rich communication behavior information, such as call time, call duration, call object, etc. After pre-processing these original call records such as format unification, missing value processing and outlier filtering, the number monitoring device converts them into a structured number behavior data set, laying a data foundation for subsequent feature extraction. This step focuses on solving the problems of data acquisition and quality assurance, ensuring that subsequent analysis is based on complete and consistent data.

[0030] S3: Extract and process the number behavior data set from three dimensions: number characteristics, call behavior, and call behavior changes to generate a number feature variable set.

[0031] S3 is the feature variable processing step, which is the process of extracting valuable features from the number behavior data. The number monitoring device performs feature engineering from three dimensions: first, the number feature dimension, extracting the basic attributes of the number such as the city of origin, operator type, number segment type, etc.; second, the call behavior dimension, calculating the number of calls and the number of call objects under different time periods and types; finally, the call behavior change dimension, calculating statistical features such as mean, variance, etc. for the behavior indicators to capture the stability and change pattern of the behavior. After the features of these three dimensions are fused, a set of feature variables that fully reflect the characteristics of the number is formed. This multi-dimensional feature extraction method is one of the core innovations of the embodiments of the present application, which enables the model to understand and predict the stability state of the number from multiple angles.

[0032] S4: Based on the number feature variable set, an initial model for predicting number stability is constructed through the XGBoost algorithm, and the prediction results of the initial model for predicting number stability are evaluated and parameters are optimized to generate a number stability prediction model.

[0033] First, the feature variable set is divided into a training set and a test set according to a preset ratio to ensure the balance and representativeness of the data. Then, the initial prediction model is built using the XGBoost algorithm, which has excellent performance in processing classification problems of structured data. By using the test set to evaluate the model performance and calculate key indicators such as AUC and accuracy, the number monitoring device can fully understand the prediction ability of the model. According to the evaluation results, the number monitoring device performs number monitoring device optimization on model parameters such as tree depth and learning rate, and finally generates an optimized number stability prediction model. This step ensures that the model can accurately capture the characteristics and patterns of number stability and provide a reliable prediction basis for practical applications.

[0034] In this embodiment, the construction of the XGBoost model adopts specific parameter settings and architecture design. The model base learner adopts a decision tree (CART tree), the maximum tree depth is tuned to 6, the number of subtrees is 150, and 8-thread parallel computing is enabled. The feature column sampling is set to 0.8, that is, 80% of the features are randomly used each time the tree is built. In terms of training parameters, the learning rate is set to 0.05. This lower learning rate makes the model training more stable; the minimum number of leaf node samples is set to 3, which effectively prevents overfitting; the regularization parameter lambda is set to 1.2 and alpha is set to 0.5, which control the L2 and L1 regularization strengths respectively; the sample sampling ratio is 0.8; the sample weight balance parameter is about 5.7, that is, the ratio of the number of negative samples to the number of positive samples. The training data scale is about 10,000 number samples, of which 8,500 are stable numbers and 1,500 are unstable numbers, divided into training sets and test sets at a ratio of 8:2. After feature screening, about 60 key features are retained. The training process uses 5-fold cross validation to ensure model stability, and early_stopping_rounds is set to 50 to avoid overfitting. It actually converges to about 120 iterations. The evaluation index mainly uses AUC, while monitoring precision and recall. The loss function uses logistic regression for binary classification problems. The model's AUC on the test set reaches 0.92, with an accuracy of 87% and a recall of 83%. The screening accuracy under the high threshold setting (0.8) reaches 95%. The training data sources include call result data recorded by the outbound call verification device (about 60%), user active error correction feedback data (about 25%), and number recovery data provided by operators (about 15%).

[0035] S5: deploy the number stability prediction model to a business scenario, obtain the yellow pages number data in the number database, and use the number stability prediction model to predict the stability of the yellow pages number data to generate a number stability probability value.

[0036] The model is converted into a persistent format through serialization methods, and the environment configuration is completed on the business server. Then a scheduled task scheduler is built, and an automated prediction process is established to achieve continuous monitoring of the prediction process. The number monitoring device obtains the yellow page number data that needs to be monitored from the number library, uses the deployed model to predict stability, and generates a stability probability value for each number. This step transforms the theoretical model into a practical application, establishes an efficient and automated prediction mechanism, and enables the model to continuously provide support for business decisions. The successful implementation of the model deployment ensures that the entire number monitoring device can operate stably in the actual environment and continuously output valuable prediction results.

[0037] S6: Based on the number stability probability value, the numbers with probability values ​​higher than a preset probability threshold are checked in detail, and the number database is updated according to the checking results.

[0038] According to the number stability probability value predicted by the model, the number monitoring device sets an appropriate threshold and screens out high-risk changed numbers for key verification. The verification can be performed by automatically calling the number monitoring device or manually dialing, with the purpose of confirming the actual status of the number. Based on the verification results, the number monitoring device updates the number library information, such as removing or marking invalid numbers, and updating the verification time of stable numbers. Of particular importance is that the verification results will also be fed back to the model training link to form a mechanism for continuous optimization. By regularly updating the model with new verification data, the number monitoring device can adapt to changes in number usage patterns and maintain long-term effectiveness. This closed-loop optimization mechanism is a significant advantage of the embodiments of the present application, which realizes the intelligent management and continuous improvement of the number library.

[0039] Among them, Figure 2 As shown, step S1 specifically includes: S1.1: Extract positive samples and negative samples from the outbound call verification data results. Positive samples are numbers that have been successfully verified and the number body has not changed, and negative samples include numbers that have failed verification.

[0040] Specifically, by analyzing the verification results recorded by the outbound call verification number monitoring device, the numbers are classified according to the verification status. Numbers that have been successfully verified and whose main body has not changed are marked as positive samples, representing stable numbers; while numbers that have failed verification (such as no one answered, the number has been deactivated, the number has been changed, etc.) are marked as negative samples, representing unstable numbers. These data come from the regular outbound call verification operations of the number monitoring device, with a high number and basic coverage.

[0041] S1.2: Collect user error correction data as a supplementary data source for negative samples. User error correction data refers to number error or change information reported by users. This type of data has high timeliness and accuracy.

[0042] In step S1.2, the number monitoring device collects error correction feedback data from users. This type of data is usually information that users actively feedback after discovering errors or changes in their numbers when using them, and has high timeliness and accuracy. As a supplement to negative samples, user error correction data can timely capture numbers that have not been routinely verified by the number monitoring device but have been changed, thereby improving the timeliness of the samples.

[0043] S1.3: Integrate recycled number data as another supplementary data source for negative samples. Recycled numbers refer to numbers that have been recycled by telecom operators and are no longer valid or have been used by someone else.

[0044] Integrate the recycled number data provided by telecom operators. This type of data refers to numbers that have been recycled by operators and are either no longer valid or have been reallocated to new users. As another important supplement to negative samples, recycled number data can effectively identify number failures caused by number recycling, further enriching the diversity of negative samples.

[0045] S1.4: Clean and integrate the collected sample data, remove duplicate data and abnormal data, and form a well-labeled training sample set to provide a data basis for subsequent model training.

[0046] In step S1.4, the sample data collected in the first three steps are cleaned and integrated. The number monitoring device removes duplicate data to ensure that the same number does not appear multiple times in the training set; filters abnormal data, such as samples with irregular formats and incomplete information; and annotates all samples to ensure that each sample has a clear label (stable or unstable) and necessary attribute information. Through this step, a training sample set with reliable quality and unified structure is formed, laying a data foundation for subsequent model training.

[0047] like Figure 3 As shown, step S2 specifically includes: S2.1: Determine a preset time window range based on the verification time period of the numbers in the training sample set.

[0048] Specifically, in step S2.1, an appropriate time window range is determined based on the verification time period of the numbers in the training sample set. The selection of the time window directly affects the quality of subsequent feature extraction and the accuracy of model prediction.

[0049] Usually, the number monitoring device will select a key time period based on the number verification period (such as monthly verification, quarterly verification), such as the data of the past three months. The setting of the time window needs to take into account the timeliness and sufficiency of the data. If the window is too short, it may lead to insufficient data and affect the stability of the feature; if the window is too long, it may contain outdated information and reduce the accuracy of the prediction.

[0050] S2.2: For each number in the training sample set, extract the call detail records within the preset time window from the communication record number monitoring device.

[0051] In step S2.2, the number monitoring device extracts the call details records within the specified time window from the communication record number monitoring device for each number in the training sample set. These detailed records contain the basic communication behavior information of the number, such as call time, call duration, call object (i.e. other numbers that call with the number), etc. This step requires efficient data interaction with the communication record number monitoring device, and the processing may involve a large number of numbers and huge historical call records.

[0052] S2.3: The call detail records are formatted, missing values ​​are processed, and outliers are filtered in sequence to form pre-processed call detail records.

[0053] In step S2.3, the number monitoring device preprocesses the extracted call detail records. The first step is to unify the format to ensure that the formats of all data fields are consistent, such as converting timestamps in different formats to a standard format; the second step is to handle missing values, using appropriate filling strategies for missing data fields, such as mean filling, zero filling, or previous value filling; the last step is outlier filtering, identifying and processing abnormal data, such as records with abnormally long or short call durations. This step ensures the quality and consistency of the data and provides a reliable data foundation for subsequent feature extraction.

[0054] S2.4: Performing structured processing on the pre-processed call detail records to generate the number behavior data set.

[0055] The pre-processed call detail records are structured to form a number behavior data set that is convenient for feature extraction. Structural processing includes organizing and classifying call detail records by number and building a data structure suitable for subsequent analysis, such as a time series format or an aggregated statistical format. Through this step, the original call detail records are converted into a structured number behavior data set, which is convenient for subsequent feature extraction and model training.

[0056] like Figure 4 As shown, step S3 specifically includes: S3.1: Using the number behavior data set, extract the number's home city, operator type, and number segment type information from the number feature dimension, and output a number attribute feature set; Specifically, in step S3.1, the number monitoring device extracts features from the number feature dimension. The number's city of origin information is extracted, which reflects the number's geographical distribution characteristics; the operator type information is extracted, such as China Mobile, China Unicom or China Telecom, which reflects the number's basic network provider; the number segment type information is extracted, such as whether it is a virtual operator number, whether it is a special segment, etc. These basic attribute features together constitute the basic classification characteristics of the number, forming a number attribute feature set.

[0057] S3.2: Based on the number attribute feature set and in combination with the number behavior data set, the number of calls and the number of call recipients are counted from the call behavior dimension to form a call behavior feature set.

[0058] In the embodiment of the present application, in the process of S3.2, feature extraction is performed from the call behavior dimension. Based on the number behavior data set, the number monitoring device calculates the call behavior indicators of each number in different time periods (such as weekly dimension, monthly dimension) and different time types (such as working days and weekends). The main indicators include the number of calls, that is, the total number of calls initiated or received by the number in a specific time period; the number of call objects (uid number), that is, the number of different numbers that call with the number; and other possible behavioral indicators, such as average call duration, call frequency, etc. These behavioral indicators together constitute the usage pattern characteristics of the number, forming a call behavior feature set.

[0059] S3.3: Using the call behavior feature set, from the call behavior change dimension, calculate the statistical characteristics of each behavior indicator in the call behavior feature set, and construct a call behavior change feature set, wherein the statistical characteristics include at least one of a mean and a variance.

[0060] In step S3.3, the number monitoring device constructs characteristic variables from the dimension of call behavior changes.

[0061] The number monitoring device calculates the statistical characteristics of the behavior indicators extracted in S3.2 to capture the stability and change pattern of the number usage behavior. The key statistical characteristics include the mean, which reflects the average level of the indicator; the variance, which reflects the degree of fluctuation of the indicator; and other possible statistics, such as trend indicators and seasonal indicators. These statistical characteristics together constitute the change characteristics of the number usage behavior and form a call behavior change feature set.

[0062] S3.4: Fusing the number attribute feature set, the call behavior feature set, and the call behavior change feature set to generate the number feature variable set.

[0063] In step S3.4, the feature sets generated in the first three steps are fused to generate a complete number feature variable set. During the feature fusion process, the number monitoring device ensures the consistency of the scale of each feature, such as through standardization or normalization; at the same time, feature selection or dimensionality reduction may be performed to reduce redundant features and reduce model complexity. The final number feature variable set contains comprehensive features extracted from three dimensions: number features, call behavior, and call behavior changes, providing rich feature input for subsequent model training.

[0064] like Figure 5 As shown, step S4 specifically includes: S4.1: Divide the number feature variable set into a training set and a test set according to a preset ratio.

[0065] Specifically, in step S4.1, the number monitoring device divides the number feature variable set generated by S3 into a training set and a test set according to a preset ratio. When dividing the data, two key technical conditions need to be ensured: data balance and data representativeness. Data balance means that the ratio of positive and negative samples (stable numbers and unstable numbers) in the training set and the test set should be consistent to avoid the model being biased towards the majority category. Considering that the number of stable numbers is usually much larger than that of unstable numbers, the number monitoring device adopts stratified sampling technology and divides them according to the label ratio to ensure that the category distribution of the original data is maintained in both the training set and the test set.

[0066] S4.2: Using the test set, the performance of the initial model for predicting number stability is evaluated, and a performance evaluation index is calculated.

[0067] Data representativeness means that both the training set and the test set after division can fully represent the distribution characteristics of the overall data. To this end, in the division process, not only the stable / unstable label distribution should be considered, but also the distribution of other key features, such as the operator distribution and regional distribution of numbers. This comprehensive data division method ensures that the data set used for model training has good statistical characteristics.

[0068] In step S4.2, the number monitoring device evaluates the initialized XGBoost model.

[0069] XGBoost (e Xtreme Gradient Boosting) is an integrated learning algorithm that is particularly suitable for processing classification problems of structured data. It is used to construct a number stability prediction model in the embodiment of the present application. In the model evaluation stage, the number monitoring device uses the test set data to predict the trained model and calculate key performance indicators. In this embodiment, AUC (Area Under Curve) is mainly used as an indicator to evaluate the distinguishing ability of the model. The AUC value is between 0.5 and 1, and the larger the value, the stronger the distinguishing ability of the model. In addition, indicators such as accuracy, precision, and recall are also calculated to comprehensively evaluate the model performance.

[0070] S4.3: According to the performance evaluation index, adjust the parameters of the initial model for predicting number stability to generate the number stability prediction model.

[0071] In step S4.3, the number monitoring device optimizes the model parameters according to the evaluation results.

[0072] The number monitoring device uses parameter tuning technologies such as grid search or Bayesian optimization to adjust the key parameters of the model.

[0073] The tuning parameters include but are not limited to: maximum tree depth (max_depth), learning rate (eta), regularization parameters (lambda, alpha), minimum number of leaf node samples (min_child_weight), etc.

[0074] Through cross validation, the model is trained and the performance is evaluated multiple times using different parameter combinations to find the optimal parameter combination that can maximize the evaluation index. Parameter tuning is a computationally intensive process that requires balancing model performance and computing resource consumption. In an embodiment of the present application, an early-stopping mechanism is adopted, that is, when the performance indicators on the validation set do not improve significantly in multiple consecutive iterations, the training process is terminated in advance to avoid overfitting and unnecessary waste of computing resources.

[0075] In this embodiment, the construction of the XGBoost model adopts the following specific parameter settings and architecture design: Model Architecture: Basic learner: decision tree (CART tree); Maximum tree depth: 6 (optimal value after tuning); Number of subtrees: 150 trees; Parallel computing: Enabled, using 8 threads; Feature column sampling: 0.8 (80% of the features are randomly used each time a tree is built); Training parameters: Learning rate (eta): 0.05 (optimal value after tuning, a lower learning rate makes model training more stable); Minimum number of leaf node samples (min_child_weight): 3 (to prevent overfitting); Regularization parameter lambda: 1.2 (L2 regularization parameter, controlling model complexity); Regularization parameter alpha: 0.5 (L1 regularization parameter, promoting feature sparsity); Sample sampling ratio (subsample): 0.8 (80% of the samples are randomly selected each time the tree is built); Sample weight balance parameter (scale_pos_weight): about 5.7 (the ratio of the number of negative samples to the number of positive samples); Training data size: Total sample size: about 10,000 numbers, of which 8,500 are stable numbers and 1,500 are unstable numbers; Training set / test set division ratio: 8:2; Feature dimension: about 60 key features retained after feature screening; Training process: K-fold cross validation (k=5) was used to ensure the stability of the model; Set early_stopping_rounds to 50 to avoid overfitting; Number of training iterations: The actual convergence is about 120 iterations; Evaluation indicators: AUC is mainly used, while precision and recall are monitored; Loss function: logistic regression is used for binary classification problems; Feature Importance: During the training process, the model automatically identified key features and their importance scores; The top five features include: daily average of query times in the first week, weekly query times variance in the past four weeks, monthly query times variance in the past three months, longest call duration, and change rate of the number of call recipients; Model performance indicators: AUC on the test set: 0.92; Accuracy: 87% Recall rate: 83% Screening accuracy at high threshold setting (0.8): 95%; Training data source: Call result data recorded by the outbound call verification device (about 60%); User proactive error correction feedback data (about 25%); Number recycling data provided by operators (about 15%); Exemplarily, the number monitoring device divides the number feature variable set into a training set and a test set according to a preset ratio of 8:2.

[0076] Taking into account the imbalance in the number of stable and unstable numbers in the number sample (stable numbers usually account for more than 85%), the number monitoring device uses stratified sampling technology to ensure data balance.

[0077] For example, in a feature variable set containing 10,000 numbers, 8,500 of which are stable numbers and 1,500 are unstable numbers, the number monitoring device will divide them into training sets and test sets in the same proportion. That is, the training set contains 6,800 stable numbers and 1,200 unstable numbers, and the test set contains 1,700 stable numbers and 300 unstable numbers. This stratified sampling method ensures that both the training set and the test set can maintain the category distribution in the original data. At the same time, the number monitoring device will also ensure that the distribution of other key features of the training set and the test set is consistent, such as operator distribution (the proportion of China Mobile, China Unicom, and China Telecom), regional distribution (the proportion of different cities), etc., to ensure the representativeness of the data.

[0078] The model is initialized using the XGBoost algorithm and key parameters are configured. The number monitoring device sets the evaluation target (eval_metric) to AUC, which is an indicator for measuring the discrimination ability of the binary classification model and is particularly suitable for evaluating the performance of the number stability prediction model. The maximum tree depth (max_depth) of the model is initially set to 5 to control the complexity of the decision tree and avoid overfitting or underfitting. The learning rate (learning_rate) is set to a more conservative 0.1, so that the model can learn gradually and avoid converging to the local optimal solution too quickly. The sample sampling ratio (subsample) is set to 0.8, which means that 80% of the samples are randomly selected each time the tree is built, increasing the randomness of the model to reduce the risk of overfitting. In response to the problem of sample imbalance, the number monitoring device also configures the sample weight balance parameter (scale_pos_weight), whose value is set to the ratio of the number of negative samples to the number of positive samples, which is about 5.7 (8500 / 1500). This parameter configuration enables the model to better learn the characteristics of the minority category (unstable numbers) and reduce the negative impact of sample imbalance.

[0079] Use the configured XGBoost model to fit the training set data, and perform feature screening through feature importance evaluation.

[0080] From the 143 candidate features initially extracted, the feature variables with the greatest impact on the prediction results were screened out using the "feature_importances_" attribute of XGBoost.

[0081] For example, time series features such as "daily average of query times in the first week", "variance of weekly query times in the past four weeks", and "variance of monthly query times in the past three months" are identified as key features.

[0082] After the number monitoring device conducted a preliminary evaluation of the model, it was found that the model's AUC value on the test set reached 0.85, the accuracy was 83%, but the recall rate was only 76%. To further optimize the model performance, the number monitoring device used grid search technology to tune key parameters. After multiple rounds of tuning, the maximum tree depth was adjusted to 6, the learning rate was reduced to 0.05, the minimum number of leaf node samples (min_child_weight) was set to 3, and the regularization parameter lambda was set to 1.2.

[0083] These adjustments increased the model's AUC value to 0.92, the accuracy to 87%, and the recall to 83%. In addition, the number monitoring device also applied an early stopping mechanism, setting early_stopping_rounds to 50, that is, when the performance indicators on the validation set did not improve significantly in 50 consecutive iterations, the training process was terminated early to avoid overfitting and waste of computing resources. The number stability prediction model finally generated was able to achieve a 95% screening accuracy rate under a high threshold setting (0.8), successfully balancing accuracy and computing efficiency.

[0084] In addition, the following operations may be further performed after step S4: S4.4: Input the prediction result of the number stability prediction model into the Bayesian confidence reconfigurable streaming accelerator, perform Monte Carlo sampling through the accelerator, and generate the uncertainty distribution of the prediction result.

[0085] In step S4.4, the number monitoring device inputs the prediction results of the number stability prediction model trained by XGBoost into the Bayesian confidence reconfigurable streaming accelerator for further processing. The accelerator is an innovative computing architecture designed specifically for processing uncertainty estimation in real-time data streams.

[0086] When a number monitoring device generates a stability prediction probability value (such as 0.75) for a number, this single value cannot reflect the reliability of the prediction. The accelerator performs multiple random perturbations and resamples on the prediction results by executing the Monte Carlo sampling method, usually performing thousands of sampling operations. For example, for a number with a predicted probability of 0.75, the number monitoring device may generate 10,000 slightly different sample points to form a probability distribution. These sampling results concentrate on the "degree of confidence" of the model in the predicted value. The more concentrated the distribution, the more confident the model is in the prediction, and the more dispersed the distribution, the higher the uncertainty. In this way, the number monitoring device no longer outputs only a point estimate, but generates a complete probability distribution, which fully quantifies the uncertainty of the prediction results.

[0087] In one embodiment, The Bayesian-confident reconfigurable streaming accelerator is an innovative computing architecture whose detailed technical implementation includes multiple core components and workflows. The core architecture of the accelerator consists of three main functional units: Bayesian inference unit, streaming data processing unit, and dynamic resource allocation unit. The hardware implementation adopts a heterogeneous architecture of FPGA and GPU collaborative computing, equipped with 32GB high-bandwidth memory and 512 reconfigurable computing cores, supporting single-precision floating-point operations. In terms of the Bayesian inference unit design, the implementation algorithms include Markov chain Monte Carlo (MCMC) sampling and Hamiltonian Monte Carlo (HMC) sampling. The default configuration is 64 parallel sampling chains, each with 1,000 sample points. The Bayesian logistic regression model is used to capture the uncertainty of XGBoost predictions, and weak information priors are used for model parameters, specifically using the normal distribution N(0,5). In terms of the design of the streaming data processing unit, the data throughput can reach 10,000 numbers per second. It adopts an 8-stage pipeline design, equipped with a double buffer mechanism to ensure uninterrupted data flow, and realizes adaptive batch size adjustment (ranging between 32-256). The design of the dynamic resource allocation unit includes dynamic priority scheduling based on workload and accuracy requirements, real-time monitoring of computing unit utilization and memory usage, dynamic adjustment of the number of sampling chains and the number of samples per chain according to the complexity of the prediction results, and intelligent switching between high-performance mode and low-power mode for different types of prediction tasks. In terms of the specific implementation of Monte Carlo sampling, the sample perturbation method adds random noise that conforms to the normal distribution N(0,0.01) to the prediction results of the XGBoost model. The resampling strategy slightly perturbs the feature space to generate multiple similar but slightly different sample points. Each number prediction performs 10,000 independent samplings, and the Gelman-Rubin statistic is used to monitor the convergence of the sampling chain, which usually converges after 500-700 samplings. In terms of probability calibration calculation implementation, the calibration method combines quantile calibration with temperature scaling, determines the 90% confidence interval based on the 5% and 95% quantiles of the sampling distribution, and outputs the calibrated prediction value p' and its confidence interval [p'-δ, p'+δ] for the predicted value p. The calibration parameters are optimized on the validation set using the holdout method. The dynamic weight adjustment mechanism is based on the nonlinear function f(w)=a·exp(-b·w)+c of the confidence interval width w. The parameters are determined by Bayesian optimization as a=0.85, b=5.2, c=0.1. For predictions with confidence intervals that are too wide, the decision threshold is dynamically adjusted from 0.5 to 0.6-0.7, and the effect of each adjustment is recorded for continuous optimization of the adjustment function parameters. Through this complex and precise computing architecture, the Bayesian confidence reconfigurable streaming accelerator can efficiently process a large number of number prediction results, provide reliable uncertainty quantification, and provide more solid technical support for the decision-making of number monitoring devices. For example, the Bayesian-belief reconfigurable streaming accelerator is an innovative computing architecture whose detailed technical implementation includes the following core components and workflows:. Core architecture: The accelerator consists of three main functional units: Bayesian inference unit, streaming data processing unit, and dynamic resource allocation unit; The hardware implementation uses a heterogeneous architecture that uses FPGA (field programmable gate array) and GPU for collaborative computing; Total memory capacity: 32GB high bandwidth memory; Computing unit: 512 reconfigurable computing cores, supporting single-precision floating-point operations; Bayesian Inference Unit Design: Implementation algorithms: Markov Chain Monte Carlo (MCMC) sampling and Hamiltonian Monte Carlo (HMC) sampling; Number of sampling chains: The default configuration is 64 parallel sampling chains; Number of samples per chain: The default setting is 1000 sample points; Probabilistic model: A Bayesian logistic regression model is used to capture the uncertainty of XGBoost predictions; Prior distribution: Weakly Informative Prior is used for model parameters, specifically using normal distribution N(0,5); Streaming data processing unit design: Data throughput: can process prediction results for up to 10,000 numbers per second; Pipeline depth: 8-stage pipeline design; Cache mechanism: Double buffer design is adopted to ensure uninterrupted data flow; Batch optimization: adaptive batch size adjustment, ranging from 32-256; Dynamic resource allocation unit design: Task scheduling strategy: dynamic priority scheduling based on workload and accuracy requirements; Resource monitoring: real-time monitoring of computing unit utilization and memory usage; Adaptive configuration: dynamically adjust the number of sampling chains and the number of samples per chain according to the complexity of the prediction results; Energy efficiency optimization: intelligently switch between high-performance mode and low-power mode for different types of prediction tasks; Monte Carlo sampling is specifically implemented as follows: Sample perturbation method: add random noise conforming to the normal distribution N(0,0.01) to the prediction results of the XGBoost model; Resampling strategy: slightly perturb the feature space to generate multiple similar but slightly different sample points; Sampling depth: 10,000 independent samples are performed for each number prediction; Convergence determination: The Gelman-Rubin statistic is used to monitor the convergence of the sampling chain, which usually reaches convergence after 500-700 samplings; Probability calibration calculation implementation: Calibration method: IsotonicCalibration combined with TemperatureScaling; Confidence interval calculation: Determine the 90% confidence interval based on the 5% and 95% quantiles of the sampling distribution; Mathematical representation: For the predicted value p, output the calibrated predicted value p' and its confidence interval [p'-δ, p'+δ]; Calibration parameter optimization: Use the Hold-out Method to optimize the calibration parameters on the validation set; Dynamic weight adjustment mechanism: Adjustment function: nonlinear function f(w)=a·exp(-b·w)+c based on the confidence interval width w; Parameter values: a=0.85, b=5.2, c=0.1 were determined by Bayesian optimization; Threshold adjustment: For predictions with too wide confidence intervals, the decision threshold is dynamically adjusted from 0.5 to 0.6-0.7; Feedback mechanism: record the effect of each adjustment and use it to continuously optimize and adjust function parameters; Through this complex and precise computing architecture, the Bayesian confidence reconfigurable streaming accelerator can efficiently process a large number of number prediction results, provide reliable uncertainty quantification, and provide more solid technical support for the decision-making of number monitoring devices. In practical applications, the accelerator has increased the screening accuracy of unstable numbers from 87.4% to 92.1%, and the accuracy under high threshold settings from 95% to 97.3%, and significantly improved the robustness and reliability of the model.

[0088] S4.5: Based on the uncertainty distribution, perform a probability calibration calculation to form a 90% confidence interval for the prediction result.

[0089] In step S4.5, the number monitoring device performs a probability calibration calculation based on the uncertainty distribution generated in S4.4 to form a 90% confidence interval for the prediction result.

[0090] Specifically, the number monitoring device first performs a statistical analysis on the distribution generated by Monte Carlo sampling, calculates its 5% and 95% quantiles, and determines an interval range that includes 90% of the sample points in the distribution.

[0091] For example, for a number with a predicted probability of 0.75, the probability calibration calculation may yield a result of "0.75±0.15 (90% confidence interval)", indicating that the number monitoring device is 90% confident that the true stability probability value of the number falls between 0.6 and 0.9, which reflects the high uncertainty of the prediction.

[0092] For another number with a predicted probability of 0.75, the result may be "0.75±0.05 (90% confidence interval)", indicating that the number monitoring device is more confident in this prediction. The width of the confidence interval directly reflects the model's degree of confidence in the prediction result, providing more information basis for subsequent decision-making. This uncertainty quantification method based on probability theory enables the number monitoring device to distinguish between "knowing" and "not knowing", effectively improving the reliability of the prediction.

[0093] S4.6: Based on the 90% confidence interval value of the prediction result, a dynamic weight adjustment mechanism is used to generate a calibrated and optimized prediction result, and the calibrated and optimized prediction result is updated to the number stability prediction model for the number stability prediction process.

[0094] The number monitoring device generates calibrated and optimized prediction results based on the 90% confidence interval value calculated by S4.5 using a dynamic weight adjustment mechanism. The mechanism dynamically adjusts the weight of the prediction result according to the width of the confidence interval.

[0095] For prediction results with narrower confidence intervals (such as ±0.05), the number monitoring device gives higher weights because the model is more confident in these predictions; for prediction results with wider confidence intervals (such as ±0.15), the number monitoring device gives lower weights and may conservatively adjust the predicted values ​​to be closer to the middle value.

[0096] In specific implementation, the number monitoring device may use a nonlinear function based on the interval width to calculate the adjustment coefficient. For example, the prediction with an interval width of 0.1 may remain the original value, while the prediction with an interval width of 0.3 may be adjusted 15% toward the median. This calibrated prediction result is updated to the number stability prediction model, replacing or supplementing the original prediction value for the subsequent number stability prediction process. Through this calibration mechanism, the screening accuracy of unstable numbers has increased from the original 87.4% to 92.1%, and the accuracy under high threshold settings has increased from 95% to 97.3%, and the number of identified unstable numbers has increased by about 15%, significantly improving the prediction performance of the number monitoring device.

[0097] The Bayesian confidence reconfigurable streaming accelerator is an innovative computing architecture designed for handling uncertainty estimation in real-time data streams. In the scenario of number stability prediction, it is not enough to just get a probability value, but also to know the credibility of this probability value, that is, how much "confidence" the model has in its own prediction results. The accelerator uses the Monte Carlo sampling method to perform multiple random sampling of the prediction results to generate a probability distribution of the prediction results, thereby quantifying the uncertainty of the prediction.

[0098] Based on this uncertainty quantification, the number monitoring device performs a probability calibration calculation to form a 90% confidence interval for the prediction result, which means that there is a 90% confidence that the true result will fall within this interval. The number monitoring device dynamically adjusts the weight of the prediction result according to the width of the confidence interval, giving higher weights to prediction results with narrower confidence intervals (lower uncertainty) and lower weights to prediction results with wider confidence intervals (higher uncertainty), thereby generating a calibrated and optimized prediction result.

[0099] Step S5 specifically includes: S5.1: The number stability prediction model is converted into a model file in a persistent format using a serialization method, and the environment configuration is completed on the service server.

[0100] A serialization method is used to convert the trained and calibrated number stability prediction model into a model file in a persistent format, and the environment configuration is completed on the business server.

[0101] In terms of specific implementation, the number monitoring device uses Python's pickle module or XGBoost's built-in save_model method to convert the model object into a binary file or a JSON format file for easy storage and subsequent loading. At the same time, the number monitoring device will also save the model-related metadata information, such as feature lists, standardized parameters, probability calibration parameters, etc., to ensure the integrity of the model deployment. In terms of business server environment configuration, the number monitoring device will install necessary dependent libraries (such as XGBoost, NumPy, Pandas, etc.), configure appropriate computing resources (such as the number of CPU cores, memory size), set network access rights, test model loading and reasoning functions, and ensure that the model can run normally in the production environment. This step is the key link in converting the theoretical model into an actually usable number monitoring device, which determines whether the model can play a stable role in actual business.

[0102] S5.2: Using the configured business server environment, build a scheduled task scheduler and establish an automated prediction process.

[0103] Number monitoring devices usually use tools such as Cronjobs or Airflow to set up regularly executed prediction tasks, such as automatically triggering the prediction process at 2 a.m. every day or 8 a.m. every Monday.

[0104] The automated process includes multiple links: first, data extraction, obtaining the required data from the number database and the communication record number monitoring device; then feature calculation, generating the input features required for the model according to the predefined feature engineering process; followed by model prediction, loading the persistent model file to perform the prediction calculation; and finally, result storage, writing the predicted number stability probability value to the database or result file. The number monitoring device will also configure the dependencies between tasks, such as ensuring that feature calculation is performed only after data extraction is completed, as well as the task retry mechanism to handle possible execution failures. This automated scheduling mechanism reduces manual intervention, ensures the regular execution of prediction tasks, and provides technical support for the continuous monitoring of the number database.

[0105] S5.3: According to the automated prediction process, a prediction process monitoring mechanism is implemented to continuously record and output prediction status and result distribution data.

[0106] Specifically, the number monitoring device sets up multi-level monitoring points to track the execution status of the prediction task in real time, including: task start time, data loading progress, feature calculation time, model prediction speed, result storage status, etc.

[0107] At the same time, the number monitoring device will also record the statistical distribution of the prediction results, such as the number of numbers in each probability interval, the proportion of high-risk numbers, the risk distribution of numbers of different operators, etc. These monitoring data are displayed to management personnel through the log number monitoring device, data panel or alarm mechanism. In particular, the number monitoring device will set up an anomaly detection mechanism, which will automatically trigger an alarm notification when there is a significant anomaly in the distribution of the prediction results (such as a sudden and substantial increase in high-risk numbers).

[0108] In addition, the number monitoring device will also record resource consumption, such as CPU usage, memory usage, and predicted latency, in order to optimize the performance of the number monitoring device. This comprehensive monitoring mechanism ensures the transparency and visibility of the prediction process, helps to promptly discover and resolve potential problems, and ensures the stable operation of the number monitoring device.

[0109] S5.4: Obtain yellow page number data in the number database, and use the number stability prediction model to predict the stability of the yellow page number data to generate a number stability probability value.

[0110] The number monitoring device processes the number database in batch polling, divides the numbers in the database into multiple batches, and processes one batch every day to ensure that all numbers are covered within a complete cycle (such as one month).

[0111] For the extracted numbers, the number monitoring device first extracts relevant feature variables, which are consistent with the feature definition used in model training, including obtaining recent call details from the communication record number monitoring device, and calculating feature variables in three dimensions: number features, call behavior, and call behavior changes. The feature calculation process is highly automated to support large-scale data processing. The number monitoring device pays special attention to the consistency of feature processing and missing value processing to ensure consistency with the processing methods in the training phase. Model prediction uses vectorized operations to process multiple numbers at one time to improve computing efficiency. Finally, the number monitoring device generates a stability probability value between 0 and 1 for each number, where the closer the value is to 1, the more likely the number is to be in an unstable state and the higher the risk of change; the closer the value is to 0, the more likely the number is to be in a stable state and the lower the risk of change. These probability values ​​provide a scientific basis for subsequent differentiated verification.

[0112] The number monitoring device needs to serialize the trained XGBoost model into a persistent format, such as a pickle file or JSON format, for easy storage and loading. Then, configure the necessary dependency environment on the production server to ensure that the model can run normally. Next, the number monitoring device establishes an automated mechanism for the prediction process, sets up a scheduled task scheduler (such as Cronjobs or Airflow), and automatically triggers the prediction task at a predetermined time interval (such as daily or weekly). The number monitoring device also establishes a monitoring mechanism to monitor the model operation status and the distribution of prediction results in real time, and detect abnormal situations in a timely manner. The number monitoring device obtains the yellow pages number data that needs to be monitored from the number library, and uses the deployed model to predict stability and generate a stability probability value for each number.

[0113] In addition, after step S5, the following operations may be further performed: S5.5: Using a large language model based on an active reasoning framework, an in-depth analysis is performed on the number stability probability value generated by the number stability prediction model to generate an analysis result at the semantic level.

[0114] The number monitoring device uses a large language model (LLM) based on an active inference framework to perform in-depth analysis on the probability values ​​generated by the number stability prediction model. This step establishes a collaborative decision-making framework between the XGBoost model and the large language model.

[0115] When the XGBoost model predicts certain numbers in an uncertain range (such as a probability value between 0.4 and 0.6) or shows abnormal patterns, the active reasoning framework triggers the large language model to conduct a deeper analysis. The large language model generates a multi-angle reasoning chain by retrieving similar case histories, analyzing the temporal changes in number usage patterns, and integrating specific rules in the industry knowledge base.

[0116] For example, for a merchant number in the catering industry, LLM may find that the number's recent call pattern is abnormal and there are discussions on social media about the restaurant's possible relocation, so it judges that its risk of change is higher than the prediction value of the XGBoost model. In addition, LLM can understand and utilize richer semantic information, such as analyzing the voice text content in outbound call records and identifying semantic clues of number changes (such as "this number is no longer from XX company"). Through this deep semantic analysis, the number monitoring device generates semantic-level analysis results containing rich explanatory content, providing supplementary basis or correction suggestions for predictive decisions.

[0117] The large language model of the active reasoning framework is a deep neural network based on the Transformer architecture, which is specially trained for the field of number monitoring. The model infrastructure uses a pre-trained general large language model as the basis, specifically an autoregressive Transformer model with a parameter level of 175B, including 48 layers of Transformer encoders, 96 attention heads, each attention head dimension is 128, word embedding dimension is 12,288, and context window length is 4,096 tokens. The activation function uses GELU (Gaussian Error Linear Unit), and the layer normalization uses the Pre-LayerNorm structure to improve training stability. In terms of domain adaptability fine-tuning, the parameter efficient fine-tuning technology LoRA (Low-Rank Adaptation) is used, and the LoRA rank is set to 16, and only about 0.1% of the model parameters are updated. The fine-tuning data scale is about 50,000 labeled number analysis cases, and the data sources include historical number verification records, expert analysis reports, and telecommunications field knowledge bases. The learning rate is set to 2e-5, the cosine decay strategy is adopted, the training batch size is 32, and 3 rounds of full data set iterations are performed. The training hardware is 8×A100 80GB GPU, and mixed precision training is adopted. The active reasoning framework design adopts the Chain-of-Thought reasoning mechanism, the number of reasoning steps is 5-8 steps, which is dynamically adjusted according to the complexity of the problem. The reasoning control is based on the reasoning path selection of self-consistency verification. The knowledge integration combines the external structured knowledge base and the internal parameter knowledge. The output formatting adopts JSON structured output to ensure compatibility with subsequent devices. The integration mechanism with XGBoost adopts cascade integration, XGBoost as the first-level predictor, and the large language model as the second-level analyzer. The trigger condition is that when the XGBoost prediction probability is in the [0.4, 0.6] interval or the prediction result deviates significantly from the historical pattern, the large language model analysis is triggered. The XGBoost prediction result, key features and their importance scores are transmitted through structured API calls. The decision rule integrates the judgment of the two models based on the weighted voting mechanism, and the weight is dynamically adjusted according to the specific scenario. In terms of solving the "black box" problem, the model not only provides the final judgment, but also outputs a complete reasoning chain, recording in detail every step of the reasoning process from raw data to conclusion. For example, for a number judged to be high-risk, the model may provide the following reasoning chain: "The number's outgoing call volume has dropped by 50% in the past three weeks (abnormal behavior) → the merchant type is a small restaurant (a high-change rate industry) → the keyword 'change of boss' appears in the outbound call record (a direct change clue) → the business district has a demolition record (environmental risk factor) → the overall judgment of the change risk is high". This transparent reasoning process makes the model judgment auditable and verifiable, effectively reducing the opacity of AI decision-making.

[0118] The collaborative decision-making framework of the XGBoost model and the large language model is an innovative technical architecture that fully integrates the advantages of two different types of models to form a complementary and collaborative intelligent decision-making device. In terms of architectural design, the framework adopts a "dual-core + multi-level interaction" structure. As the first core, the XGBoost model is responsible for efficient quantitative analysis based on structured feature data, and is good at extracting stable statistical patterns from historical call records; as the second core, the large language model is responsible for semantic understanding, knowledge reasoning, and complex scenario judgment, and is good at processing unstructured data and identifying rare patterns. The two core models are connected through a multi-level interactive interface, including a feature sharing layer, a decision fusion layer, and a feedback optimization layer, to ensure that information can flow in both directions and be effectively integrated.

[0119] Specifically, the collaborative decision-making framework of XGBoost and large language models is an innovative technical architecture, and its detailed design adopts a "dual-core + multi-level interaction" structure. As the first core, the XGBoost model is responsible for efficient quantitative analysis based on structured feature data, and is good at extracting stable statistical patterns from historical call records; as the second core, the large language model is responsible for semantic understanding, knowledge reasoning and complex scenario judgment, and is good at processing unstructured data and identifying rare patterns. The two core models are connected through a multi-level interactive interface, including a feature sharing layer, a decision fusion layer, and a feedback optimization layer, to ensure that information can flow in both directions and be effectively integrated. The working principle of this framework is based on the mechanism of "hierarchical processing + collaborative decision-making". First, all number data is preliminarily predicted by the XGBoost model to obtain the basic stability probability value. Then, the device screens out complex cases that need further analysis according to preset rules (such as the prediction result is between the uncertainty interval of 0.4-0.6, or the prediction result deviates significantly from the historical pattern). For these cases, the device triggers the large language model for in-depth analysis. The large language model receives the structured feature data of the number, the prediction results of XGBoost, and additional unstructured information (such as outbound call record text, merchant information, etc.), and generates supplementary judgments and explanations through multi-step reasoning. Finally, the device integrates the judgment results of the two models through weighted fusion or rule switching to form a final decision. To ensure the controllability and consistency of the output of the large language model, the device interacts with the model through a defined function calling interface (Function Calling), defines multiple dedicated functions such as analyzeNumberPattern, retrieveSimilarCases, assessIndustryRisk and other core functions, and ensures the structuring and consistency of input and output through parameter passing in JSON format. This multi-objective optimization strategy of the Framework model collects multi-dimensional data including prediction accuracy, recognition coverage and resource consumption efficiency, and establishes a comprehensive optimization indicator system. The device uses this indicator system to calculate the gradient direction of each optimization target, constructs the Pareto optimal solution set through mathematical optimization methods, and determines the optimal balance point based on business needs and resource constraints. Based on this balance point, the device generates updated prediction weight parameters (adjusting the weights of XGBoost and the large language model in decision-making) and threshold parameters for different business scenarios to achieve precise allocation of resources and improve the accuracy of number stability prediction.

[0120] The working principle of this framework is based on the mechanism of "tiered processing + collaborative decision-making".

[0121] First, all number data is preliminarily predicted by the XGBoost model to obtain a basic stability probability value. Then, the device screens out complex cases that require further analysis based on preset rules (such as prediction results between the uncertainty interval of 0.4-0.6, or prediction results that deviate significantly from historical patterns). For these cases, the device triggers the large language model for in-depth analysis. The large language model receives the structured feature data of the number, the prediction results of XGBoost, and additional unstructured information (such as outbound call record text, merchant information, etc.), and generates supplementary judgments and explanations through multi-step reasoning. Finally, the device integrates the judgment results of the two models through weighted fusion or rule switching to form a final decision.

[0122] The synergy mechanism is the core of the framework, which is specifically reflected in three levels: information complementarity, decision fusion and learning optimization.

[0123] At the information complementarity level, XGBoost provides precise quantitative analysis results, while the large language model provides semantic understanding and background knowledge; at the decision fusion level, the device dynamically adjusts the weights of the two models according to different scenarios. For example, for regular numbers with rich historical data, it is more inclined to trust the judgment of XGBoost, while for special numbers with sparse data or abnormal behavior, it pays more attention to the analysis of the large language model; at the learning optimization level, the two models provide learning feedback to each other. For example, the quantitative features of XGBoost can guide the large language model to focus on key factors, and the new patterns recognized by the large language model can be converted into new features of XGBoost, forming a mutually reinforcing evolutionary mechanism.

[0124] In order to ensure the technical feasibility and transparency of the collaborative framework, the application of large language models adopts a specific technical implementation path. First, the basic large language model (such as deepseek or GPT-4o) is fine-tuned for domain adaptability and further trained using data sets related to number monitoring. The fine-tuning process uses parameter-efficient fine-tuning techniques (such as LoRA and P-Tuning) to update only a small number of key parameters in the model, maintaining basic capabilities while strengthening specific domain knowledge. The fine-tuning data includes annotated number cases, expert analysis reports, and related industry knowledge, and guides the model to learn how to analyze number stability issues through instruction fine-tuning.

[0125] In actual applications, the large language model interacts with the device through a defined function calling interface (FunctionCalling).

[0126] For example, the device defines several dedicated functions such as analyzeNumberPattern(numberFeatures,historyData) for analyzing number patterns, retrieveSimilarCases(numberInfo,topK) for retrieving similar historical cases, and assessIndustryRisk(merchantInfo,industryType) for assessing industry risks. When a large language model is needed for analysis, the device passes in relevant data through these structured interfaces, and the model generates a JSON response in a standard format, including analysis conclusions, confidence scores, and reasoning steps. This FunctionCalling mechanism ensures the controllability and consistency of the output of the large language model and avoids the uncertainty that may be caused by free text generation.

[0127] In addition, to reduce the risk of "black box", the device implements an explainability mechanism. When generating analysis results, the large language model not only provides the final judgment, but also outputs a complete chain of thought, recording in detail every step of the reasoning process from raw data to conclusion.

[0128] For example, for numbers judged to be high-risk, the model may provide the following reasoning chain: "The number's outgoing call volume has dropped by 50% in the past three weeks (abnormal behavior) → the merchant type is a small restaurant (a high-change rate industry) → the keyword 'change of owner' appears in the outbound call record (a direct change clue) → the business district has a demolition record (an environmental risk factor) → the overall judgment of the change risk is high". This transparent reasoning process makes the model judgment auditable and verifiable, significantly reducing the opacity of AI decision-making, while also providing professional knowledge support for human operators, enabling them to understand and verify the rationality of the device's judgment.

[0129] In a specific example, the active inference framework large language model used in this embodiment is a deep neural network based on the Transformer architecture, which is specially trained for the field of number monitoring. It includes: Model infrastructure: Basic model: The pre-trained general large language model is used as the basis, specifically the autoregressive Transformer model with 175B parameters; Number of hidden layers: 48-layer Transformer encoder; Number of attention heads: 96; Each attention head dimension: 128; Word embedding dimension: 12,288; Context window length: 4,096 tokens; Activation function: GELU (GaussianErrorLinearUnit); Layer normalization: Use Pre-LayerNorm structure to improve training stability; Domain Adaptation Fine-tuning: Fine-tuning method: parameter efficient fine-tuning technology LoRA (Low-Rank Adaptation); LoRA rank: 16; Fine-tuning parameter ratio: only update about 0.1% of the model parameters; Fine-tuning data size: about 50,000 annotated number analysis cases; Data sources: historical number verification records, expert analysis reports, and telecommunications knowledge base; Learning rate: 2e-5, using cosine decay strategy; Training batch size: 32; Number of training rounds: 3 rounds of full dataset iterations; Training hardware: 8×A10080GB GPU, mixed precision training; Active reasoning framework design: Reasoning chain structure: adopts the Chain-of-Thought reasoning mechanism; Number of reasoning steps: 5-8 steps, dynamically adjusted according to the complexity of the problem; Reasoning control: Reasoning path selection based on self-consistency verification; Knowledge integration: combining external structured knowledge base with internal parameter knowledge; Output formatting: JSON structured output to ensure compatibility with subsequent devices; Model and XGBoost integration mechanism: Integration architecture: Cascade integration is adopted, with XGBoost as the first-level predictor and the large language model as the second-level analyzer; Trigger condition: When the XGBoost prediction probability is in the range of [0.4, 0.6] or the prediction result deviates significantly from the historical pattern, the large language model analysis is triggered; Information transmission: XGBoost prediction results, key features and their importance scores are transmitted through structured API calls; Decision rules: Integrate the judgments of the two models based on the weighted voting mechanism, and the weights are dynamically adjusted according to the specific scenario.

[0130] FunctionCalling implementation: Define function set: 8 core functions including analyzeNumberPattern, retrieveSimilarCases, and assessIndustryRisk; Calling method: Parameters are passed in JSON format to ensure the structure and consistency of input and output; Execution control: supports function chain calling to implement complex analysis processes; Error handling: set up timeout mechanism and exception capture to ensure device stability; Model evaluation metrics: Explanation accuracy: 94.7% (the proportion of explanations provided by the model that agree with human experts); Decision-making support capability: In borderline cases that XGBoost has difficulty in judging, the decision accuracy rate is improved by 17.3 percentage points; Reasoning latency: On average, each number analysis takes 0.8 seconds; Resource consumption: Peak memory usage is 16GB, supporting batch processing mode; S5.6: In combination with the analysis results at the semantic level, a multi-objective optimization strategy is adopted to output optimized prediction parameters and optimized verification strategies.

[0131] The number monitoring device first constructed a multi-dimensional optimization indicator system including prediction accuracy, recognition coverage, and resource consumption efficiency. This multi-objective perspective breaks the traditional single accuracy optimization idea and enables the number monitoring device to make more balanced decisions in a complex business environment. Due to conflicts between different optimization objectives (for example, improving accuracy may reduce coverage or increase resource consumption), the number monitoring device calculates the gradient direction of each optimization objective and constructs a Pareto optimal solution set. A Pareto optimal solution refers to a set of solutions that cannot improve all other objectives at the same time without compromising at least one objective.

[0132] Through this multi-objective optimization, the number monitoring device can improve the overall work efficiency by about 35% with the same resource input. The number monitoring device combines business needs and resource constraints, selects the best balance point from the Pareto frontier, generates optimized prediction weight parameters (adjusts the weights of XGBoost and LLM in decision-making) and threshold parameters in different scenarios, formulates differentiated verification strategies for different types of numbers, and realizes accurate allocation of resources.

[0133] S5.7: Based on the optimized prediction parameters and the optimized verification strategy, the number stability prediction model is iteratively optimized to improve the accuracy of number stability prediction.

[0134] This step realizes a closed loop from analysis and decision-making to model optimization. The number monitoring device applies the parameters generated by multi-objective optimization to the prediction process and adjusts the decision-making mechanism of the model.

[0135] For example, for numbers in high-risk industries identified by LLM (such as catering, retail and other industries with frequent business turnover), the number monitoring device will automatically lower the instability judgment threshold and increase the monitoring sensitivity; while for numbers in stable industries (such as large enterprises), the threshold will be increased to reduce unnecessary verification.

[0136] The number monitoring device has also established a feedback iteration mechanism to feed back the verification results to the model training phase, continuously update the training samples, optimize the feature weights of the XGBoost model, and enrich the case knowledge base of the LLM. Through continuous learning and adaptation, the number monitoring device can continuously improve itself as the business environment changes.

[0137] This iterative optimization mechanism has increased the prediction accuracy from 95% of the original solution to 98.2%, especially when dealing with borderline cases and atypical patterns. It has also increased the average recognition lead time from 7 days to 15 days, leaving more time for business adjustments. This closed-loop optimization reflects the adaptive learning ability of the number monitoring device, allowing the prediction of number stability to continue to improve and adapt to changing business needs.

[0138] The big language model of the active reasoning framework is an artificial intelligence number monitoring device capable of deep semantic analysis and reasoning. In the number monitoring scenario, when the XGBoost model's prediction results for certain numbers are in an uncertain range (such as a probability value between 0.4-0.6) or an abnormal pattern appears, the active reasoning framework triggers the big language model to perform a deeper analysis. The big language model generates a multi-angle reasoning chain by retrieving similar case histories, analyzing the temporal changes in number usage patterns, and integrating specific rules in the industry knowledge base (such as the number update cycle characteristics of a specific industry), providing supplementary basis or correction suggestions for XGBoost's predictions.

[0139] Specifically, the number monitoring device first obtains numbers with probability values ​​within a specific range (such as 0.4-0.6), performs in-depth feature analysis on these numbers, and forms preliminary analysis results. Then, the number monitoring device searches the historical case database to find historical cases with similar features to the current number, and analyzes the temporal changes in the number usage pattern to generate number change risk features. Based on the active reasoning framework, the number monitoring device integrates all analysis results, takes into account the outbound call record text and merchant status information, and comprehensively evaluates the risk of number change.

[0140] The number monitoring device also adopts a multi-objective optimization strategy, taking into account multiple objectives such as prediction accuracy, recognition coverage, and resource consumption efficiency. Through the Pareto optimal principle, the number monitoring device finds the best balance between these potentially conflicting objectives and generates optimized prediction parameters and verification strategies. These optimization results are applied to the number stability prediction process to continuously improve model performance.

[0141] Among them, in S5.5, a large language model based on an active reasoning framework is used to perform in-depth analysis on the number stability probability value generated by the number stability prediction model to generate analysis results at the semantic level, including: A1: Obtain a number whose number stability probability value is within a preset probability interval, and perform a deep feature analysis on the number to form a preliminary analysis result.

[0142] First, we filter out numbers whose probability of number stability is in a specific uncertainty interval (e.g., between 0.4 and 0.6). These numbers are boundary cases that are difficult for the XGBoost model to determine and require further in-depth analysis.

[0143] For these numbers, the number monitoring device performs in-depth feature analysis, including detailed analysis of number attributes, historical call behavior, and change patterns. The analysis process not only considers existing feature variables, but also further explores potential related features, such as the time distribution characteristics of number usage, group characteristics of interactive objects, etc. Through this in-depth feature exploration, the number monitoring device forms preliminary analysis results for each boundary number, laying the foundation for subsequent historical case matching and semantic analysis. This step reflects the number monitoring device's special attention to boundary situations that are difficult for the model to judge, and improves the accuracy of the overall prediction.

[0144] A2: Using the preliminary analysis results, search the historical case database, analyze the temporal changes in number usage patterns, and generate number change risk characteristics.

[0145] The number monitoring device uses the preliminary analysis results to search for similar historical cases in the historical case database. These historical cases contain number samples with known results in the past and their complete feature change trajectories. Through the case matching algorithm, the number monitoring device finds the historical cases most similar to the currently analyzed number and analyzes the results and reasons for the changes in number stability in these cases.

[0146] At the same time, the number monitoring device conducts a time-series analysis of the usage pattern of the current number to detect the changing trend of behavioral characteristics over time, such as sudden changes in call frequency and changes in the composition of call objects. This analysis method based on historical experience and time-series changes can reveal dynamic risk signals that may be ignored by static features and generate more comprehensive number change risk features. This step makes full use of historical experience and time dimension information to enhance the ability to perceive number change risks.

[0147] A3: Based on the active reasoning framework, the number change risk features are integrated, and the number-related outbound call record text and merchant status information are extracted and processed, and a comprehensive assessment result of the number change risk is output as the analysis result at the semantic level.

[0148] Based on the active inference framework, the number change risk features generated in the first two steps are intelligently fused.

[0149] The active reasoning framework simulates the thinking process of human experts to evaluate the risk status of the number from multiple angles and dimensions. Of particular importance is that in this step, the number monitoring device also extracts and processes the outbound call record text related to the number, such as the voice-to-text content in the call record, from which semantic clues that may indicate a number change (such as "this number is no longer owned by XX company").

[0150] At the same time, the number monitoring device analyzes the merchant status information associated with the number, such as the merchant's business conditions, online comments, news reports, etc., to find external factors that may affect the stability of the number. By comprehensively considering these multi-dimensional information, the number monitoring device outputs a comprehensive risk assessment result with rich semantic interpretation. This step introduces semantic understanding and knowledge reasoning capabilities, greatly improving the analysis depth and prediction capabilities of the number monitoring device.

[0151] It should be noted that in S5.6, a multi-objective optimization strategy is adopted to output the optimized prediction parameters and the optimized verification strategy, including: B1. Collect data on prediction accuracy, recognition coverage, and resource consumption efficiency, and establish a multi-dimensional optimization indicator system.

[0152] A comprehensive multi-dimensional optimization indicator system has been constructed, no longer focusing on a single prediction accuracy indicator.

[0153] Specifically, the number monitoring device collects three types of key data at the same time: prediction accuracy data, which reflects the correctness of the model's judgment; identification coverage data, which indicates the proportion of actual unstable numbers that can be identified; and resource consumption efficiency data, which measures the verification resource consumption required to achieve a specific accuracy and coverage.

[0154] In addition, the number monitoring device may also consider other auxiliary indicators, such as forecast interpretability, forecast lead time, etc. Through the unified collection and standardized processing of these multi-dimensional indicators, the number monitoring device has established an indicator system that can comprehensively evaluate the performance of models and strategies. This multi-dimensional perspective enables the number monitoring device to make more balanced and comprehensive optimization decisions in a complex business environment, rather than simply pursuing the maximization of a certain indicator.

[0155] B2. Using the multi-dimensional optimization index system, calculate the gradient direction of each optimization objective, construct a Pareto optimal solution set, and determine the optimal balance point.

[0156] Based on the established multi-dimensional indicator system, a Pareto optimal analysis is performed. The number monitoring device calculates the gradient direction of each optimization target, that is, determines the direction and degree of influence of changing a certain parameter on each indicator. Since there may be conflicts between different optimization targets (for example, improving accuracy may reduce coverage), the number monitoring device constructs a Pareto optimal solution set through mathematical optimization methods. The Pareto optimal solution refers to a set of solutions that cannot improve all other targets at the same time without compromising at least one target. On this basis, the number monitoring device combines business needs and resource constraints to determine the optimal balance point, that is, selects the solution that best suits the current business scenario on the Pareto frontier. This decision-making method based on multi-objective optimization theory enables the number monitoring device to find a scientific and reasonable balance between multiple conflicting targets, avoiding the one-sidedness that may be caused by the optimization of a single indicator.

[0157] B3. Generate updated prediction weights and threshold parameters based on the optimal balance point, and apply the updated parameters to the number stability prediction process.

[0158] Based on the determined optimal balance point, specific operating parameters and strategies are generated.

[0159] It mainly includes two types of parameters: one is the prediction weight parameter, which adjusts the weight distribution of the XGBoost model and the large language model in the final decision. Different weight combinations may be used for different types of numbers; the other is the threshold parameter, which sets differentiated judgment thresholds for different business scenarios and number types. For example, a lower threshold may be used for high-risk industries to increase sensitivity. These updated parameters are directly applied to the number stability prediction process, affecting subsequent judgment results and verification strategies. The number monitoring device also establishes a real-time monitoring mechanism for parameter effects, continuously evaluates the effects of parameter adjustments, and provides feedback for the next round of optimization. This step transforms abstract optimization theory into specific execution parameters, realizes the transition from theoretical analysis to practical application, and enables the results of multi-objective optimization to effectively improve the actual performance of the number monitoring device.

[0160] Step S6 specifically includes: S6.1: Based on the number stability probability value, set an appropriate instability probability threshold to screen out high-risk change numbers.

[0161] In step S6.1, the number monitoring device sets an appropriate instability probability threshold according to business requirements and resource conditions. By analyzing the ROC curve (Receiver Operating Characteristic Curve) and the PR curve (Precision-Recall Curve), the number monitoring device determines the model performance under different thresholds.

[0162] For example, when the probability threshold is set to 0.5, the screening accuracy of unstable numbers can reach 87.4%; when the threshold is raised to 0.8, the screening accuracy can be increased to 95%, but the number of unstable numbers identified will be reduced accordingly. The selection of the threshold needs to balance accuracy and coverage, and take into account the available verification resources.

[0163] S6.2: Focus on verifying the high-risk numbers screened out to confirm their actual stable status.

[0164] In step S6.2, the number monitoring device conducts key verification on the selected high-risk numbers. The verification can be performed by automatic outbound call number monitoring device for basic validity verification, or by manual dialing for in-depth information confirmation. The purpose of the verification is to confirm the actual status of the number, including whether it is in use, whether the owner has changed, whether it is still associated with the original recorded merchant, etc.

[0165] S6.3: Update the number database information based on the verification results, and feed the verification results back to the model training phase to achieve continuous optimization of the model.

[0166] In step S6.3, the number monitoring device updates the number database information according to the verification results. For numbers that are confirmed to have changed, the number monitoring device removes them from the database or marks them as invalid; for numbers that are confirmed to be stable, the number monitoring device updates their last verification time and adjusts their priority in subsequent monitoring.

[0167] In addition, the verification results are fed back to the model training process to form a closed-loop optimization mechanism. The number monitoring device compares the verification results with the model prediction results to generate new labeled data, and uses this new data to update the model regularly (such as monthly or quarterly) to improve the model's prediction accuracy. This continuous learning mechanism enables the model to adapt to changes in number usage patterns and emerging features, maintaining long-term effectiveness.

[0168] The present application also provides a number monitoring device, the device comprising: The sample collection module is used to collect positive and negative sample data from outbound call verification data results, user error correction data, and recovered number data to generate a training sample set; A data processing module, configured to extract historical call detail data of each number from a communication recording device storing historical call records of the number based on each number included in the training sample set, and pre-process the historical call detail data to generate a number behavior data set; A feature extraction module is used to extract and process the number behavior data set from three dimensions: number feature, call behavior, and call behavior change, and generate a number feature variable set; A model training module is used to construct an initial model for predicting number stability based on the number feature variable set through the XGBoost algorithm, and to evaluate and optimize the parameters of the prediction results of the initial model for predicting number stability to generate a number stability prediction model; A prediction module, used to deploy the number stability prediction model to a business scenario, obtain the yellow page number data in the number library, and use the number stability prediction model to predict the stability of the yellow page number data, and generate a number stability probability value; The verification and updating module is used to focus on verifying the numbers that are higher than a preset probability threshold based on the number stability probability value, and update the number database according to the verification results.

[0169] An embodiment of the present application also provides a computer device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the above-mentioned number monitoring method.

[0170] The embodiment of the present application also provides a computer-readable storage medium, which stores computer instructions, which are used to enable a computer to execute the method of the above number monitoring method. The embodiment of the present application also provides a computer program product, including computer instructions, which implement the steps of the above number monitoring method when executed by a processor.

[0171] This application has the following technical effects: Efficient resource utilization: Through the differentiated verification strategy based on model prediction, the number monitoring device only focuses on verifying high-risk numbers, greatly reducing the waste of resources caused by indiscriminate verification and improving resource utilization efficiency. Experimental results show that compared with traditional full verification, this number monitoring device can save about 70% of verification resources while maintaining considerable data accuracy.

[0172] Improve prediction accuracy: By extracting comprehensive features from three dimensions: number features, call behavior, and call behavior changes, combined with the powerful learning ability of the XGBoost algorithm, the number monitoring device can accurately identify the risk of number changes. Under high threshold settings, the number monitoring device can achieve a 95% screening accuracy rate, which greatly exceeds the accuracy of traditional methods.

[0173] Intelligent management: The number monitoring device has built a complete closed loop for number database maintenance. Through the cycle mechanism of model prediction, key verification, database update, and feedback optimization, it has realized the automation and intelligent management of number database maintenance. In particular, by introducing the Bayesian confidence accelerator and the active reasoning framework of the large language model, the number monitoring device can continuously learn and optimize to adapt to the changing business environment.

[0174] Early warning capability: With in-depth analysis of number behavior patterns, the number monitoring device can identify numbers with high change risks before the numbers are actually changed, providing early warning for business adjustments. Compared with the traditional post-verification method, this number monitoring device increases the average identification lead time from 0 days to 7-15 days, leaving sufficient time for business adjustments.

[0175] To sum up, the number monitoring method and device provided in this application, by deeply mining the number behavior characteristics and combining advanced machine learning technology and intelligent optimization strategies, realize efficient and accurate maintenance of the commercial number database, solve the technical problems such as resource waste, update lag, and low accuracy existing in traditional methods, greatly improve the maintenance efficiency and quality of the number database, and provide reliable data guarantee for the application of the commercial number database.

[0176] The embodiment of the present disclosure also provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of a number monitoring method described in the above method embodiment are executed. The storage medium can be a volatile or non-volatile computer-readable storage medium.

[0177] In addition, an embodiment of the present disclosure also provides a computer program product, which stores a computer program. When the computer program is executed by a processor, the steps of a number monitoring method provided in any of the above embodiments of the present disclosure are executed. For details, please refer to the above method embodiments, which will not be repeated here.

[0178] The computer program product may be implemented in hardware, software or a combination thereof. In an optional embodiment, the computer program product is embodied as a computer storage medium, which may be a volatile or non-volatile computer-readable storage medium. In another optional embodiment, the computer program product is embodied as a software product, such as a software development kit (SDK) and the like.

[0179] Those skilled in the art can clearly understand that, for the convenience and simplicity of description, the specific working process of the above-described equipment and devices can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here. In the several embodiments provided in the present disclosure, it should be understood that the disclosed equipment, devices and methods can be implemented in other ways. The device embodiments described above are merely schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another device, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some communication interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0180] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0181] In addition, each functional unit in each embodiment of the present disclosure may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0182] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a non-volatile computer-readable storage medium that is executable by a processor. Based on this understanding, the technical solution of the present disclosure, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present disclosure. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0183] Finally, it should be noted that the above-described embodiments are only specific implementation methods of the present disclosure, which are used to illustrate the technical solutions of the present disclosure, rather than to limit them. The protection scope of the present disclosure is not limited thereto. Although the present disclosure is described in detail with reference to the aforementioned embodiments, ordinary technicians in the field should understand that any technician familiar with the technical field can still modify the technical solutions recorded in the aforementioned embodiments within the technical scope disclosed in the present disclosure, or can easily think of changes, or make equivalent replacements for some of the technical features therein; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present disclosure, and should be included in the protection scope of the present disclosure. Therefore, the protection scope of the present disclosure shall be based on the protection scope of the claims.

Claims

1. A number monitoring method, characterized in that: include: Collect outbound call verification data results, user error correction data, and positive and negative sample data from recovered number data to generate a training sample set; Based on each number included in the training sample set, extracting historical call detail data of each number from a communication recording device storing historical call records of the number, and preprocessing the historical call detail data to generate a number behavior data set; Extract and process the number behavior data set from three dimensions: number characteristics, call behavior, and call behavior changes, and generate a number characteristic variable set; Based on the number feature variable set, an initial model for predicting number stability is constructed by using the XGBoost algorithm, and the prediction results of the initial model for predicting number stability are evaluated and parameters are tuned to generate a number stability prediction model; Deploy the number stability prediction model to a business scenario, obtain the yellow page number data in the number library, and use the number stability prediction model to predict the stability of the yellow page number data to generate a number stability probability value; Based on the number stability probability value, numbers with probability values ​​higher than a preset threshold are checked in detail, and the number database is updated according to the checking results.

2. The method according to claim 1, characterized in that The method of extracting the historical call details data of each number from a communication recording device storing the historical call records of the number based on each number included in the training sample set, and preprocessing the historical call details data to generate a number behavior data set includes: Determining a preset time window range based on the verification time period of the numbers in the training sample set; For each number in the training sample set, extracting call detail records within the preset time window from the communication recording device; The call detail records are sequentially formatted, missing value processed, and outlier filtered to form pre-processed call detail records; The pre-processed call detail records are structured to generate the number behavior data set.

3. The method according to claim 1, characterized in that The feature extraction and processing of the number behavior data set from the three dimensions of number features, call behavior, and call behavior changes to generate a number feature variable set includes: Utilizing the number behavior data set, extracting the number's home city, operator type, and number segment type information from the number feature dimension, and outputting a number attribute feature set; According to the number attribute feature set and in combination with the number behavior data set, the number of calls and the number of call objects are counted from the call behavior dimension to form a call behavior feature set; Using the call behavior feature set, from the call behavior change dimension, calculating the statistical characteristics of each behavior indicator in the call behavior feature set, and constructing a call behavior change feature set, wherein the statistical characteristics include at least one of a mean value and a variance; The number attribute feature set, the call behavior feature set and the call behavior change feature set are feature-fused to generate the number feature variable set.

4. The method according to claim 1, characterized in that The step of evaluating the prediction result of the initial model for predicting number stability and optimizing parameters to generate a number stability prediction model includes: Dividing the number feature variable set into a training set and a test set according to a preset ratio; Using the test set, the performance of the initial model for predicting number stability is evaluated, and a performance evaluation index is calculated; According to the performance evaluation index, the parameters of the initial model for predicting number stability are adjusted to generate the number stability prediction model.

5. The method according to claim 4, characterized in that After generating the number stability prediction model, the method further includes: Inputting the prediction result of the number stability prediction model into a Bayesian confidence reconfigurable streaming accelerator, performing Monte Carlo sampling through the accelerator, and generating an uncertainty distribution of the prediction result; Based on the uncertainty distribution, a probability calibration calculation is performed to form a 90% confidence interval for the prediction result; Based on the 90% confidence interval value of the prediction result, a dynamic weight adjustment mechanism is used to generate a calibrated and optimized prediction result, and the calibrated and optimized prediction result is updated to the number stability prediction model for the number stability prediction process.

6. The method according to claim 1, characterized in that The deploying the number stability prediction model to a business scenario includes: Using a serialization method, the number stability prediction model is converted into a model file in a persistent format, and the environment configuration is completed on the business server; Using the configured business server environment, build a scheduled task scheduler and establish an automated prediction process; According to the automated prediction process, a prediction process monitoring mechanism is implemented to continuously record and output prediction status and result distribution data.

7. The method according to claim 1, characterized in that After using the number stability prediction model to predict the stability of the yellow page number data, the method further includes: Using a large language model based on an active reasoning framework, the number stability probability value generated by the number stability prediction model is deeply analyzed to generate an analysis result at the semantic level; Combining the analysis results at the semantic level, a multi-objective optimization strategy is adopted to output optimized prediction parameters and optimized verification strategies; Based on the optimized prediction parameters and the optimized verification strategy, the number stability prediction model is iteratively optimized to improve the accuracy of number stability prediction.

8. The method according to claim 7, characterized in that The large language model based on the active inference framework is used to perform in-depth analysis on the number stability probability value generated by the number stability prediction model to generate semantic level analysis results, including: Acquire numbers whose number stability probability values ​​are within a preset probability interval, and perform in-depth feature analysis on the numbers to form preliminary analysis results; Using the preliminary analysis results, searching the historical case database and analyzing the temporal changes in number usage patterns to generate number change risk characteristics; Based on the active reasoning framework, the number change risk features are integrated, and the number-related outbound call record text and merchant status information are extracted and processed, and a comprehensive assessment result of the number change risk is output as the analysis result at the semantic level.

9. The method according to claim 7, characterized in that: The multi-objective optimization strategy is adopted to output the optimization prediction parameters and the optimization verification strategy, including: Collect data on prediction accuracy, recognition coverage, and resource consumption efficiency, and establish a multi-dimensional optimization indicator system; Using the multi-dimensional optimization index system, the gradient direction of each optimization objective is calculated, a Pareto optimal solution set is constructed, and an optimal balance point is determined; According to the optimal balance point, updated prediction weights and threshold parameters are generated, and the updated parameters are applied to the number stability prediction process.

10. A number monitoring device, characterized in that: include: The sample collection module is used to collect positive and negative sample data from outbound call verification data results, user error correction data, and recovered number data to generate a training sample set; A data processing module, configured to extract historical call detail data of each number from a communication recording device storing historical call records of the number based on each number included in the training sample set, and pre-process the historical call detail data to generate a number behavior data set; A feature extraction module is used to extract and process the number behavior data set from three dimensions: number feature, call behavior, and call behavior change, and generate a number feature variable set; A model training module is used to construct an initial model for predicting number stability based on the number feature variable set through the XGBoost algorithm, and to evaluate and optimize the parameters of the prediction results of the initial model for predicting number stability to generate a number stability prediction model; A prediction module, used to deploy the number stability prediction model to a business scenario, obtain the yellow page number data in the number library, and use the number stability prediction model to predict the stability of the yellow page number data, and generate a number stability probability value; The verification and updating module is used to focus on verifying the numbers that are higher than a preset probability threshold based on the number stability probability value, and update the number database according to the verification results.

Images