Method for judging cell delay problem based on XGBoost decision tree
By using an XGBoost decision tree-based approach, the inefficiency of traditional cell latency analysis methods in high-dimensional data processing is addressed. This approach enables accurate root cause delineation and efficient model generalization for cell latency issues, adapting to dynamic network environments.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- 刘科汛
- Filing Date
- 2025-11-04
- Publication Date
- 2026-04-17
AI Technical Summary
Traditional cell latency analysis methods are inefficient when processing large-scale and multi-dimensional network data. They struggle to extract key features from massive amounts of KPI and XDR data, resulting in insufficient root cause localization accuracy and an inability to adapt to dynamically changing network environments.
An XGBoost decision tree-based approach was adopted, which involved data cleaning, feature selection, clustering, and model training to construct an XGBoost regression model. Pearson correlation coefficient analysis and affinity propagation clustering were combined to identify the key root causes affecting latency. The model complexity was controlled by hyperparameter tuning and early stopping.
It improves the accuracy and interpretability of root cause identification of cell latency problems, overcomes the efficiency bottleneck of traditional methods in high-dimensional data processing, enhances the generalization ability and robustness of the model, and can automatically learn the nonlinear relationship between network performance indicators and user-perceived latency.
Smart Images
Figure CN121887679A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of wireless communication technology, and in particular to a method for determining cell latency problems based on XGBoost decision trees. Background Technology
[0002] Identifying and analyzing the root causes of cell latency issues is one of the key challenges in the operation and maintenance of wireless communication networks. With the large-scale deployment of 5G networks and the diversification of services, users' demands for latency-sensitive applications (such as video calls, online games, and real-time interactive services) are increasing. Cells with abnormal latency in the network often directly affect user experience and may even lead to user complaints.
[0003] Traditional cell latency analysis methods largely rely on expert-based rules, threshold alarms, or basic statistical analysis. These methods have significant limitations when dealing with large-scale and multi-dimensional network data. On the one hand, manual rules struggle to cover complex and ever-changing network scenarios, and static, rigid threshold settings are prone to missed or false positives. On the other hand, traditional statistical methods are inefficient at handling high-dimensional features and non-linear relationships, making it difficult to effectively extract key features from massive KPI and XDR data, resulting in insufficient root cause localization accuracy and delayed response. Furthermore, existing methods often lack the ability to model the deep correlation between user perception and network metrics, making them unable to adapt to dynamically changing network environments.
[0004] Therefore, in response to the problems mentioned above, this invention proposes a method for determining cell latency issues based on XGBoost decision trees. Summary of the Invention
[0005] To overcome the inefficiency of traditional statistical methods in handling large datasets, this invention proposes a method for determining cell latency problems based on XGBoost decision trees.
[0006] The technical solution of this invention is: a method for determining cell latency issues based on XGBoost decision trees, comprising the following steps: S1 collects detailed user plane XDR call records and network management KPI data, and performs data cleaning and discretization processing. S2, using Pearson correlation coefficient analysis to screen feature indicators that are strongly correlated with "base station downlink latency"; S3 uses the affinity propagation clustering method to dynamically cluster key feature indicators, and prioritizes and assigns weights according to alarm class, parameter class, load class, interference class, and coverage class. S4. Construct an XGBoost regression model with user-perceived latency as the dependent variable and network KPI as the independent variable, and train the model using the gradient boosting algorithm. S5. Use grid search, random search or Bayesian optimization methods to tune the hyperparameters of the XGBoost model, including learning rate, tree depth, subsampling rate and regularization parameter; S6. Based on the importance ranking of the output features of the trained model, identify the key root causes affecting latency; S7 combines clustering results and feature importance to output bounded conclusions and optimization suggestions for the cell delay problem.
[0007] Preferably, the data cleaning and discretization process includes filling missing values with the mean or zero, removing invalid fields from samples with many characteristics, and discretizing continuous features into categorical variables.
[0008] Preferably, the affinity propagation clustering method includes dynamically clustering the output wireless network indicators based on features using unsupervised learning techniques, clustering key feature indicators through the affinity propagation model, verifying and adjusting the attribution of unreasonable feature indicators based on network optimization experience, and allocating weight coefficients according to the priority order of alarm class, parameter class, load class, interference class and coverage class.
[0009] Preferably, the Pearson correlation coefficient analysis is used to screen feature indicators that are strongly correlated with "base station downlink delay". By calculating the Pearson correlation coefficient, indicators with a correlation greater than 0.25 with TCP connection establishment confirmation delay are used as model input features.
[0010] Preferably, the objective function of the XGBoost model comprises two parts: training loss and model complexity regularization. The objective function can be expressed as: ; in, It is a predicted value. and the true value The loss function between It is the complexity of the k-th tree, which is usually composed of the number of leaves in the tree and the L2 norm of the leaf weights.
[0011] As a preferred approach, to optimize the objective function, XGBoost uses a second-order Taylor expansion to approximate the loss function. For a given loss function l, its approximation with respect to the predicted value is... The first and second derivatives are respectively and Therefore, the loss function can be approximated as: ; The complexity Ω(f) of a tree takes into account the depth of the tree and the weights of the leaf nodes, and can be expressed as: ; Where T is the number of leaf nodes in the tree. It is the weight of the j-th leaf node. and It is the regularization parameter.
[0012] Preferably, the hyperparameter tuning process uses an early stopping method to control the number of training rounds, terminating training early when the performance on the validation set no longer improves.
[0013] Preferably, the feature importance is analyzed using the feature importance evaluation results output after training the XGBoost decision tree model.
[0014] Preferably, the method constructs a problem root cause determination device module, which includes a raw data acquisition module, a feature category acquisition module, a structural problem identification module, a structural problem location module, and an optimization scheme output module.
[0015] Preferably, the device combines in-memory computing, partitioning and load reduction, and data indexing technologies to support real-time data processing and root cause identification of user complaints.
[0016] The beneficial effects of this invention are: 1. This invention can efficiently process large-scale datasets consisting of massive user plane XDR call detail records and network-side KPI indicators. Through machine learning feature engineering and the powerful computing capabilities of the XGBoost model, it overcomes the efficiency bottleneck of traditional statistical methods when processing high-dimensional, multi-source wireless data.
[0017] 2. By combining Pearson correlation coefficient analysis with affinity propagation clustering, this invention achieves accurate screening and dynamic classification of key influencing features, significantly improving the accuracy and interpretability of root cause delineation for cell delay problems.
[0018] 3. The XGBoost regression model used in this invention can automatically learn the complex nonlinear relationship between network performance indicators and user-perceived latency, and effectively capture hidden fault modes that are difficult to identify by traditional methods by iteratively optimizing the model through gradient boosting algorithm.
[0019] 4. This invention effectively controls model complexity, prevents overfitting, and improves the model's generalization ability and robustness in real network environments by introducing hyperparameter tuning strategies such as grid search, random search, or Bayesian optimization, combined with an early stopping mechanism. Attached Figure Description
[0020] Figure 1 The diagram shown illustrates the data quality cleaning and discretization process of this invention. Figure 2 The diagram shown is a schematic of the module architecture of the device for determining the root cause of cell problems according to the present invention. Figure 3 The diagram shown is a schematic representation of the construction feature data of this invention; Figure 4 The diagram shown illustrates the key feature values of this invention. Detailed Implementation
[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0022] This invention provides an embodiment: a method for determining cell latency issues based on XGBoost decision trees. In this embodiment, the method of this study includes feature engineering, importance determination, and iterative mining of the XGBoost decision tree model. First, raw data is collected, then data scheduling is performed to identify structural problems. Next, the XGBoost decision tree model is used for root cause analysis of the problems, including machine learning modeling, clustering of key features of the community, correlation analysis, and simulation tuning. The XGBoost model is an ensemble learning framework that uses a gradient boosting algorithm to optimize the model. Its main function is to iteratively add trees, each adding a tree that fits the residuals predicted by all previous trees. The data input consists of a classification dataset and class labels. The model training process includes initializing the model to a constant value, calculating the residuals for each iteration and building a new tree to predict these residuals, adding the predictions of the new tree to the model, updating the residuals, and continuing to the next iteration. This process is repeated until a predetermined number of iterations is reached or the model performance no longer significantly improves. In XGBoost decision trees, information gain is used to select the best split feature, while XGBoost uses gradient boosting to optimize the loss function. The loss function is approximated by a second-order Taylor expansion, and this approximation is used to select the best split. Therefore, although information gain is not directly used in XGBoost, the process of gradient boosting is consistent with the goal of information gain (i.e., selecting the best split to reduce uncertainty). Finally, machine learning feature engineering, feature selection, importance determination, and iterative mining of network problems are introduced to fully promote the initiative in discovering network problems.
[0023] In this embodiment, the root cause analysis method is explained: First, machine learning modeling is used. This step uses existing user perception indicators as the data source and machine learning methods to model the data features. This helps to identify the wireless side reasons affecting various user complaints and ultimately deduce the root cause of the complaints.
[0024] The key feature clustering of the cell is performed. This step uses unsupervised learning technology based on the wireless network indicators output in the feature selection stage. The key feature indicators are dynamically clustered through the affinity propagation model. Then, the clustering results are validated based on network optimization experience. Unreasonable feature indicators are assigned to solidify the feature indicator classification table. Weight coefficients are allocated according to the priority of classification optimization processing. For example, the processing priority is sorted in descending order by alarm category, parameter category, load category, interference category, coverage category, etc., as shown in the table below.
[0025] Table 1 Classification of Curing Characteristic Indicators Correlation analysis is performed on different features or data to identify key influencing and driving factors by analyzing the relationships between them. This allows for predictions of business development. Many correlation analysis methods exist. Basic methods quickly identify relationships between data, such as positive, negative, or no correlation. Intermediate methods measure the strength of these relationships, such as perfect or imperfect correlation. Advanced methods transform these relationships into models and use these models to predict future business development. Machine learning is used to scientifically analyze latency issues, identifying wireless metrics affecting latency performance. Based on the specific needs of this latency problem and the analysis of data characteristics, correlation coefficients (Pearson correlation coefficient) and the XGBoost decision tree algorithm are employed as the primary research tools in this project.
[0026] Through simulation optimization of the XGBoost decision tree model, the XGBoost algorithm is particularly suitable for root cause identification of cell delay problems due to its excellent ability to handle a large number of features and provide accurate predictions. Learning rate adjustment: The learning rate is a parameter that controls the step size of each model update. In cell latency problems, appropriately reducing the learning rate can make the model training process more stable and avoid overfitting, especially when the dataset is small or noisy.
[0027] Tree depth and number: The maximum depth and number of trees directly affect the complexity of the model. In cell latency problems, adjusting these parameters can help the model capture the complex relationships in the data while avoiding overfitting.
[0028] Subsampling and column sampling: Subsampling and column sampling use random subsets of data and features in each iteration, which helps improve the model's generalization ability and prevent overfitting.
[0029] Regularization terms: L1 and L2 regularization terms can reduce the complexity of the model and prevent overfitting by penalizing the magnitude of the model weights.
[0030] Early stopping: Setting early stopping stops training when performance on the validation set no longer improves. This helps to find the optimal number of iterations and prevents overfitting.
[0031] In this embodiment, the experimental design and results are explained: Please see Figure 1 To evaluate the effectiveness of the proposed method, this study designed and implemented a series of experiments. Experimental results show that the method can effectively handle large-scale datasets and accurately identify the root causes of cell latency problems. The specific goal of this machine learning research is to determine the key factors affecting user-perceived latency. User-perceived latency is a continuous dependent variable, with data sourced from user call detail records (CDRs). The corresponding independent variable is the cell-level performance index in the network management system with a 15-minute granularity. The experiment collected over 120 feature values, and correlated and matched user CDR data from one week with the minute-granular wireless indicators of the network management system to form a... The foundation of machine learning data analysis is laid by preprocessing the data according to a predetermined data quality cleaning and discretization process. First, the accuracy, completeness, and timeliness of the data are checked, specifically verifying whether the data is the required data, whether the data was collected completely, and confirming the correct collection time. Next, the data is cleaned, including deleting null values, transforming or deleting outliers, and changing the attributes of inappropriate data types. Finally, data discretization is performed, specifically defining the inability of users to use 5G as core network, terminal, and wireless network categories, and analyzing wireless issues to prepare for subsequent modeling.
[0032] A thorough understanding of the theoretical derivation and formula examples of the XGBoost algorithm is crucial for mastering its principles and applications. Therefore, this study elucidates its key concepts and derivation process. The XGBoost objective function consists of two parts: the training loss and a regularization term used to control model complexity. Its mathematical expression is the sum of the loss function term and the model complexity term. The loss function term measures the difference between the predicted and actual values, while the regularization term penalizes the complexity of each tree, typically calculated based on the number of leaf nodes and the L2 norm of the node weights. To efficiently optimize this objective function, XGBoost uses a second-order Taylor expansion to approximate the loss function. Utilizing the first and second derivatives of the loss function with respect to the predicted values, the loss function is approximated as a quadratic function with respect to newly added base learners, allowing for the analytical solution of the optimal leaf node weights and structural gain. The tree complexity regularization term explicitly considers the number of leaf nodes and the sum of squares of the weights of each leaf node, and introduces two hyperparameters, gamma and lambda, to control the number of leaf nodes and the weight values, respectively. By adjusting these hyperparameters, fine-grained control over model complexity can be achieved, preventing overfitting. Hyperparameter tuning itself is a key step in improving model performance. The following are some key concepts and formulas of XGBoost, along with their derivation: The objective function of the XGBoost model consists of two parts: training loss and model complexity regularization. The objective function can be expressed as: ; in, It is a predicted value. and the true value The loss function between It is the complexity of the k-th tree, which is usually composed of the number of leaves in the tree and the L2 norm of the leaf weights.
[0033] As a preferred approach, to optimize the objective function, XGBoost uses a second-order Taylor expansion to approximate the loss function. For a given loss function l, its approximation with respect to the predicted value is... The first and second derivatives are respectively and Therefore, the loss function can be approximated as: ; The complexity Ω(f) of a tree takes into account the depth of the tree and the weights of the leaf nodes, and can be expressed as: ; Where T is the number of leaf nodes in the tree. It is the weight of the j-th leaf node. and It is the regularization parameter.
[0034] This study explores in detail commonly used hyperparameter tuning methods to optimize XGBoost model performance. Grid search is a systematic parameter optimization method that predefines a grid containing different values for key hyperparameters (such as learning rate, maximum tree depth, subsampling ratio, etc.), exhaustively traverses all parameter combinations in the grid, and uses cross-validation to evaluate the model performance of each parameter set. Finally, it selects the parameter set with the best evaluation metric as the final configuration. Random search differs from grid search; instead of trying all possible parameter combinations, it randomly selects a certain number of parameter combinations from a specified parameter distribution for evaluation. This method is often more efficient than grid search in finding satisfactory parameter combinations when the parameter space has high dimensionality. Bayesian optimization is a more advanced optimization technique that constructs a surrogate model (usually a probabilistic model) to simulate the objective function (i.e., the relationship between model performance and hyperparameters) and intelligently selects the next most promising hyperparameter combination for evaluation using existing evaluation results. This sequential optimization strategy aims to find the global optimum or near-optimal solution with as few evaluations as possible.
[0035] Please see Figure 2 In this embodiment, the system device is described as follows: The following modules together constitute a machine learning-based optimization device for wireless network cell structure problems, aiming to improve the accuracy and efficiency of network operation and maintenance while adapting to the big data processing needs of the 5G network environment. By introducing technologies such as information gain decision trees and correlation analysis, this device can process data in real time, uncover the true causes of problems, and promote proactive network problem discovery. Furthermore, by combining in-memory computing, partitioned load balancing, and data indexing, this device can improve the accuracy of call detail record (CDR) delineation and localization, thereby providing effective problem identification and optimization solutions for determining the root cause of user complaints regarding the lack of 5G connectivity.
[0036] The device includes a raw data acquisition module, a feature category acquisition module, a structural problem identification module, a structural problem localization module, and an optimization scheme output module. The raw data acquisition module is responsible for real-time acquisition of raw data such as XDR, 4 / 5G KPI of the wireless network cell gateway, full 4 / 5G MR / MDT, and alarm faults. The feature category acquisition module classifies the characteristics of wireless network cell problems using clustering algorithms and decision trees. It processes the first batch of raw data using clustering algorithms (such as K-means or DBSCAN) to identify and classify different scenarios of the wireless network cell. The structural problem identification module uses classification algorithms to identify network structural problems. The module applies algorithms such as decision trees and random forests to identify and classify potential structural problems based on scenario categories. The structural problem localization module constructs a tree model to locate the root cause of structural problems. It uses algorithms such as Light GBM or XGBoost to build a model based on problem categories and uses the information gain of decision trees to accurately locate the root cause of the problem. The optimization scheme output module uses classification algorithms to identify wireless network cell problems, obtain the root cause of the problem, output the most suitable optimization scheme based on the root cause, and send the scheme back to the data storage server.
[0037] Furthermore, the present invention provides an embodiment: Please see Figure 3 The first step is to construct feature data and perform Pearson correlation coefficient analysis on KQI and KPI. It was found that, based on the previous analysis, the wireless "base station downlink latency" indicator is positively correlated with high platform latency and TCP connection establishment confirmation latency. It can be used for wireless problem delimitation and latency estimation. The correlation coefficient with other KPI indicators is below 0.25, and there is no strong linear correlation between them.
[0038] The second step involves clustering key cell features to establish the required characteristic relationship between the platform indicator "TCP connection confirmation delay" and the wireless network management indicator. Statistical analysis is performed on the raw data, totaling approximately 1.89 million records, with approximately 100,000 users. Each user reports an average of about 18 call detail records (CDRs), with an average delay of 2.3 seconds. The number of samples with delays greater than 4 seconds is approximately 80,000. The statistical counters and formulas for various delay types on the base station side are used as features. The statistical counters for various delay types on the base station side are shown in the table below.
[0039] Table 2. Statistical counters for various types of delays on the base station side. The third step involves analyzing the relationships between different features or data, including the removal of factors contributing to wireless complaint hotspots, and correlation analysis between the user sample (resident cells) for perception scoring and wireless complaint hotspot areas / cells. Data sources include the <Complaint List>, geographic latitude and longitude crawling, and the association between the user sample for perception scoring and resident cells. Wireless-side KPI factors are also removed, and the user sample for perception scoring (resident cells) is correlated with wireless cell download speed, large packet rate, and latency.
[0040] Step 4: XGBoost Decision Tree Process Input files: XGBoost can load various data formats, including text, arrays, and binary cache files. The loaded data is stored in the object DMatrix. Data processing: 0 can be filled, missing mean values can be filled, and invalid fields can be dropped for samples with many characteristics. To avoid errors, the length of the processed data must be the same. BOOSTER parameters: Select the booster model, learning objective, tree depth, learning rate, penalty coefficient, etc. Here, we mainly adopt some empirical parameters from case studies. Training process: Select training set, sample set, and sample ratio; Model training and output: Start model training, output accuracy results, test sample result predictions, model output, and important features.
[0041] The XGBoost decision tree diagram can be used to roughly select the threshold values of the identification rules with high accuracy. Pearson correlation coefficient analysis of each feature field (KPI) and "base station downlink latency" reveals a strong positive correlation with many load-related indicators and a strong negative correlation with uplink and downlink quality indicators.
[0042] Please see Figure 4 Key features show that wireless attach latency is the most important factor in user ratings, followed by speed and HTTP / DNS latency. Complaint hotspots have the least impact on user perception scores, but this may be related to the classification of the complained cell.
[0043] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.
Claims
1. A method for judging a cell delay type problem based on an XGBoost decision tree, characterized in that, It includes the following steps: S1 collects detailed user plane XDR call records and network management KPI data, and performs data cleaning and discretization processing. S2, using Pearson correlation coefficient analysis to screen feature indicators that are strongly correlated with "base station downlink delay"; S3 uses the affinity propagation clustering method to dynamically cluster key feature indicators, and prioritizes and assigns weights according to alarm class, parameter class, load class, interference class, and coverage class. S4. Construct an XGBoost regression model with user-perceived latency as the dependent variable and network KPI as the independent variable, and train the model using the gradient boosting algorithm. S5. Use grid search, random search or Bayesian optimization methods to tune the hyperparameters of the XGBoost model, including learning rate, tree depth, subsampling rate and regularization parameter; S6. Based on the importance ranking of the output features of the trained model, identify the key root causes affecting latency; S7 combines clustering results and feature importance to output bounded conclusions and optimization suggestions for the cell delay problem.
2. The method for judging cell delay type problems based on XGBoost decision tree according to claim 1, characterized in that: The data cleaning and discretization process includes filling missing values with the mean or zero, removing invalid fields from samples with many characteristics, and discretizing continuous features into categorical variables.
3. The method for determining cell latency issues based on XGBoost decision trees according to claim 1, characterized in that: The affinity propagation clustering method includes dynamic clustering of the output wireless network indicators based on feature selection using unsupervised learning techniques. After clustering key feature indicators through the affinity propagation model, data verification and adjustment of unreasonable feature indicator attribution are performed based on network optimization experience. Weight coefficients are then allocated according to the priority order of alarm class, parameter class, load class, interference class and coverage class.
4. The method for determining cell latency issues based on XGBoost decision trees according to claim 1, characterized in that: The Pearson correlation coefficient analysis was used to screen feature indicators that are strongly correlated with "base station downlink delay". By calculating the Pearson correlation coefficient, indicators with a correlation greater than 0.25 with TCP connection establishment confirmation delay were used as input features for the model.
5. The method for determining cell latency issues based on XGBoost decision trees according to claim 1, characterized in that, The objective function of the XGBoost model consists of two parts: training loss and model complexity regularization. The objective function can be expressed as: ； in, It is a predicted value. and the true value The loss function between It is the complexity of the k-th tree, which is usually composed of the number of leaves in the tree and the L2 norm of the leaf weights.
6. The method for determining cell latency issues based on XGBoost decision trees according to claim 5, characterized in that: To optimize the objective function, XGBoost uses a second-order Taylor expansion to approximate the loss function. For a given loss function l, its approximation with respect to the predicted value is... The first and second derivatives are respectively and Therefore, the loss function can be approximated as: ； The complexity Ω(f) of a tree takes into account the depth of the tree and the weights of the leaf nodes, and can be expressed as: ； Where T is the number of leaf nodes in the tree. It is the weight of the j-th leaf node. and It is the regularization parameter.
7. The method for determining cell latency issues based on XGBoost decision trees according to claim 1, characterized in that: The hyperparameter tuning process employs an early stopping method to control the number of training rounds, terminating training prematurely when the performance on the validation set no longer improves.
8. The method for determining cell latency issues based on XGBoost decision trees according to claim 1, characterized in that: The feature importance is analyzed using the feature importance evaluation results output after training the XGBoost decision tree model.
9. The method for determining cell latency issues based on XGBoost decision trees according to claim 1, characterized in that: This method constructs a problem root cause determination device module, which includes a raw data acquisition module, a feature category acquisition module, a structural problem identification module, a structural problem localization module, and an optimization scheme output module.
10. The method for determining cell latency problems based on XGBoost decision trees according to claim 9, characterized in that: The device combines in-memory computing, partitioning and load reduction, and data indexing technologies to support real-time data processing and root cause identification of user complaints.