A complaint preprocessing analysis method and device and a readable storage medium

Abstract

This application provides a complaint preprocessing analysis method, apparatus, and readable storage medium. The method includes: acquiring customer complaint data, network performance data, and network alarm data from network communication; performing data alignment processing on the customer complaint data, network performance data, and network alarm data to obtain multi-source aligned data; training a model based on the multi-source aligned data to obtain an intelligent complaint analysis model; and performing complaint analysis processing through the intelligent complaint analysis model. This application analyzes the perception differences in user service usage through data mining, trains samples based on implemented user data, explores input data, analyzes complaining users, discovers network problems through user perception dimensions, improves network coverage, and enhances the efficiency and accuracy of complaint processing.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method, apparatus and readable storage medium for complaint preprocessing analysis. Background Technology

[0002] As wireless networks expand and users demand higher network quality, the workload and difficulty of network optimization also increase. Handling user complaints is an indispensable part of routine network maintenance and a crucial way to identify problems and improve mobile network quality. Therefore, for mobile communication operators, improving service quality, enhancing user satisfaction, and reducing customer complaints are among the key priorities for current network optimization efforts.

[0003] Currently, mobile communication operators have accumulated a large amount of complaint data, but there is still a lack of effective utilization of this data.

[0004] Therefore, how to identify problems from the complex and varied complaint data, predict potential business issues in a timely manner, and prevent problems before they occur is a problem that needs to be solved. Summary of the Invention

[0005] The technical problem to be solved by this application is to provide a complaint preprocessing analysis method, apparatus and readable storage medium to address the above-mentioned shortcomings of the prior art.

[0006] Firstly, this application provides a complaint pre-processing analysis method, the method comprising: S1. Obtain customer complaint data, network performance data, and network alarm data from network communications; S2. Perform data alignment processing on customer complaint data, network performance data, and network alarm data to obtain multi-source aligned data; S3. Based on the multi-source aligned data, perform model training to obtain an intelligent complaint analysis model; S4. Complaint analysis and processing are performed using the intelligent complaint analysis model.

[0007] In some embodiments, S2 includes: A three-dimensional association key mechanism of user identifier-time window-base station cell is adopted to perform data alignment processing on customer complaint data, network performance data and network alarm data to obtain multi-source aligned data; Among them, the user identifier adopts a dual identifier system of communication number desensitized identity ID + SEQ user unique identifier, the time window is dynamically divided into time granularity according to data characteristics, and the base station cell adopts a combination of base station unique code + cell physical cell identifier (PCI) as the location identifier.

[0008] In some embodiments, a three-dimensional association key mechanism of user identifier-time window-base station cell is used to perform data alignment processing on customer complaint data, network performance data, and network alarm data, including: Customer complaint data, network performance data, and network alarm data are each labeled with a three-dimensional association key to form a standardized index; Establish a distributed association engine, perform hash mapping and range matching based on the association key, and automatically aggregate complaint texts, performance indicators and alarm events of the same user in the same time window and the same service cell; For cross-cell handover scenarios, the overlap of time windows and signal handover logs are used to correlate data from multiple cells before and after the handover.

[0009] In some embodiments, a user identifier-time window-base station cell three-dimensional association key mechanism is used to perform data alignment processing on customer complaint data, network performance data, and network alarm data, and the method further includes: Data normalization processing is performed on customer complaint data, network performance data, and network alarm data, specifically including: Customer complaint data normalization: A three-level processing approach is adopted, including text cleaning, keyword standardization, and semantic mapping. This includes filtering special characters and redundant expressions; standardizing synonyms and colloquial expressions through a custom dictionary; and associating standardized keywords with a network indicator dictionary. Network performance data normalization: To address the differences in the dimensions of different base station equipment and different indicator dimensions, a segmented normalization + dynamic threshold adaptation method is adopted, including: dividing the indicator value into sub-intervals according to the business scenario, and using Z-score standardization in each interval; and dynamically adjusting the normalization parameters in combination with the regional network quality benchmark value. Network alarm data normalization: Construct a standardized mapping table between alarm level and impact range: unify alarm codes from different equipment manufacturers into a three-dimensional label of alarm type-severity-affected user range; quantify alarm duration into three categories: instantaneous alarm, short-term alarm, and long-term alarm.

[0010] In some embodiments, a user identifier-time window-base station cell three-dimensional association key mechanism is used to perform data alignment processing on customer complaint data, network performance data, and network alarm data, and the method further includes: Missing value imputation is performed on customer complaint data, network performance data, and network alarm data, specifically including: Based on the classification and imputation strategies corresponding to different types of data, missing values ​​are imputed for customer complaint data, network performance data, and network alarm data. After the data is completed, cross-validation and anomaly detection are performed. The classification completion strategy includes: For missing customer complaint data: If the core issue description is missing in the complaint ticket, the Naive Bayes algorithm is used to infer possible complaint issues and generate standardized issue labels by associating the same user's historical complaint types, network alarm types and performance anomaly indicators during the same period. For missing network performance data: a time-series interpolation + similar cell reference fusion method is adopted, including: for short-term missing data, linear interpolation is used to fill in the gaps; for long-term missing data, similar cells with the same region, user density, and configuration are found based on the three-dimensional association key, and the gaps are filled in based on the average index of similar cells during the same period. For missing network alarm data: perform reverse completion based on the performance anomaly-alarm association rule base.

[0011] In some embodiments, S3 includes: S31. Construction of multi-label classification model, including: Feature extraction layer: XGBoost and LightGBM are trained in parallel to extract high-dimensional features from multi-source aligned data. XGBoost is used to capture linear and low-order nonlinear features, while LightGBM is used to mine high-order nonlinear features. The output features of the two models are concatenated to form a fused feature vector. Classification and prediction layer: Construct a lightweight neural network, input fused feature vectors, and output multi-dimensional probability of complaint problem types; S32. Model parameter tuning and optimization, including: Coarse tuning: A Bayesian optimization algorithm is used to perform a global search on the core parameters of XGBoost / LightGBM and the network parameters of the neural network to obtain the optimal parameter range; Fine-tuning: Within the optimal parameter range, a grid search combined with 5-fold cross-validation is used to determine the final parameters; S33. User-network relationship graph construction, including: Graph structure definition: Node types include user nodes and base station cell nodes. User node attributes include historical complaint records, package type, resident cell, and usage behavior; base station cell node attributes include network performance indicators, alarm records, coverage area, and user density. Edge types and weights: Spatial association edges connect user nodes in the same cell, with weights based on the percentage of time a user spends in the cell; Social association edges are built based on user relationships through calls, text messages, and shared data plans, with weights based on interaction frequency; Service association edges connect user nodes with nodes in their resident cell, with weights based on average signal strength. Graph dynamic update mechanism: Real-time synchronization of user location changes and cell network status, dynamically adjusting the existence and weight of edges; S34. Graph neural network prediction models, including: Neighbor sampling optimization: A hierarchical sampling + weight-first strategy is adopted, including: when sampling the neighbors of user nodes, neighbors with higher weights are selected first, and the number of samples per layer is limited; Feature fusion improvement: Introduce a node type attention mechanism into the aggregation function, including: when aggregating neighbor features, assign different attention to user node features and cell node features; Mass Complaint Risk Assessment: The model outputs a mass complaint risk score for user nodes, identifies high-risk users based on the risk score threshold, identifies the core propagation path of mass complaints by analyzing the connected subgraphs of high-risk users, and outputs the risk propagation probability to predict the scope of complaint spread.

[0012] In some embodiments, S31 further includes: a sample imbalance solution, specifically including at least one of the following: Sample expansion: A combined SMOTE+ synthetic sample generation method is adopted, including: generating similar samples for minority complaint samples using the SMOTE algorithm; and synthesizing virtual samples that conform to business logic based on real complaint texts and online data using generative adversarial networks. Loss function improvement: A weighted FocalLoss + cross-entropy fusion loss function is adopted, including: assigning higher weights to minority class samples; introducing a FocalLoss modulation factor to reduce the loss contribution of easily classified samples and focus on difficult-to-classify samples; Model training strategy: adopts a stratified sampling + incremental training mode, including: stratified sampling of the training set according to the type of complaint; establishing an incremental training mechanism to update the top-level neural network parameters of the model after adding new complaint data and network data.

[0013] Secondly, this application provides a complaint pre-processing analysis device, the device comprising: The data acquisition module is configured to acquire customer complaint data, network performance data, and network alarm data from network communications. The data alignment module is configured to perform data alignment processing on customer complaint data, network performance data, and network alarm data to obtain multi-source aligned data; The model training module is configured to train the model based on the multi-source aligned data to obtain an intelligent complaint analysis model. The analysis and processing module is configured to perform complaint analysis and processing through the intelligent complaint analysis model.

[0014] Thirdly, this application provides a complaint preprocessing analysis apparatus, including a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to implement the complaint preprocessing analysis method described in the first aspect above.

[0015] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the complaint preprocessing analysis method described in the first aspect.

[0016] This application provides a complaint preprocessing analysis method, apparatus, and readable storage medium. The method includes: acquiring customer complaint data, network performance data, and network alarm data from network communication; performing data alignment processing on the customer complaint data, network performance data, and network alarm data to obtain multi-source aligned data; training a model based on the multi-source aligned data to obtain an intelligent complaint analysis model; and performing complaint analysis processing through the intelligent complaint analysis model. This application provides a complaint preprocessing analysis method that analyzes the perception differences in user business usage through data mining, trains samples based on implemented user data, explores input data, analyzes complaining users, discovers network problems through user perception dimensions, shifts from KPI-oriented to KQI-oriented, and improves network coverage. This lowers the threshold for complaint processing, shortens the average processing time of a single work order, reduces operation and maintenance costs, and improves the efficiency and accuracy of complaint processing. Attached Figure Description

[0017] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0018] Figure 1 A flowchart of a complaint preprocessing analysis method provided in this application embodiment; Figure 2 A schematic diagram illustrating model training and application provided in the embodiments of this application; Figure 3 A schematic diagram illustrating intelligent complaint analysis provided in an embodiment of this application; Figure 4 A schematic diagram illustrating the acquisition of user complaint cells provided in this application embodiment; Figure 5 A schematic diagram illustrating complaint location provided in an embodiment of this application; Figure 6 A schematic diagram of the structure of a complaint preprocessing analysis device provided in this application embodiment; Figure 7 This is a schematic diagram of another complaint preprocessing analysis device provided in an embodiment of this application.

[0019] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0020] To enable those skilled in the art to better understand the technical solution of this application, the embodiments of this application will be further described in detail below with reference to the accompanying drawings.

[0021] It is understood that the specific embodiments and accompanying drawings described herein are merely for explaining this application and are not intended to limit this application.

[0022] It is understood that, without conflict, the various embodiments and features in the embodiments of this application can be combined with each other.

[0023] It is understood that, for ease of description, only the parts relevant to this application are shown in the accompanying drawings, while parts unrelated to this application are not shown in the drawings.

[0024] It is understood that each unit or module involved in the embodiments of this application may correspond to only one entity structure, or may be composed of multiple entity structures, or multiple units or modules may be integrated into one entity structure.

[0025] It is understood that the terms "first," "second," etc., used in the embodiments of this application are used to distinguish different objects or to distinguish different treatments of the same object, rather than to describe a specific order of objects.

[0026] It is understood that, without conflict, the functions and steps marked in the flowcharts and block diagrams of this application may occur in a different order than those marked in the accompanying drawings.

[0027] It is understood that the flowcharts and block diagrams of this application illustrate the possible architecture, functions, and operations of systems, apparatuses, devices, and methods according to various embodiments of this application. Each block in a flowchart or block diagram may represent a unit, module, program segment, or code, containing executable instructions for implementing the specified function. Furthermore, each block or combination of blocks in the block diagrams and flowcharts may be implemented using a hardware-based system to implement the specified function, or using a combination of hardware and computer instructions.

[0028] It is understood that the units and modules involved in the embodiments of this application can be implemented by software or by hardware. For example, the units and modules can be located in the processor.

[0029] It is understood that the specific values ​​of each parameter in this application are merely illustrative examples, and in practical applications, the parameters can be optimized and adjusted based on specific requirements.

[0030] This application employs data mining technology to build a predictive model using data information, effectively predicting customer complaints so as to better utilize this data information to provide decision support for network optimization.

[0031] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0032] This application provides a complaint preprocessing analysis method. The workflow of this method can be implemented by electronic devices, such as computers, handheld smart terminals, etc. For ease of explanation, the embodiments of this application are described with the computer as the subject of the method execution.

[0033] Figure 1 This is a schematic diagram of the complaint preprocessing analysis method provided in the embodiments of this application, as shown below. Figure 1 As shown, this application provides a complaint pre-processing analysis method, which includes steps S1-S4, as follows: S1. Obtain customer complaint data, network performance data, and network alarm data from network communications; Specifically, it acquires customer complaint data (unstructured text), network performance data (time-series), and network alarm data (event-based).

[0034] S2. Perform data alignment processing on customer complaint data, network performance data, and network alarm data to obtain multi-source aligned data; In some embodiments, S2 includes: A three-dimensional association key mechanism of user identifier-time window-base station cell is adopted to perform data alignment processing on customer complaint data, network performance data and network alarm data to obtain multi-source aligned data; Among them, the user identifier adopts a dual identifier system of communication number desensitized identity ID + SEQ user unique identifier, the time window is dynamically divided into time granularity according to data characteristics, and the base station cell adopts a combination of base station unique code + cell physical cell identifier (PCI) as the location identifier.

[0035] This application addresses the heterogeneous characteristics of customer complaint data (unstructured text), network performance data (time-series), and network alarm data (event-based), and innovatively proposes a three-dimensional association key mechanism of "user identifier - time window - base station cell" to achieve accurate alignment of multi-source data.

[0036] Specifically, the data alignment mechanism employs a unified mapping driven by three-dimensional association keys. The association key definition and standardization process includes: 1. User Identification: A unified dual identification system of "de-identified mobile phone number ID + SEQ user unique identifier" is adopted to solve the problem of inconsistency between user account in complaint data and user number in network data, and cross-system matching is completed through the internal user identity mapping table of China Unicom.

[0037] 2. Time Window: Dynamically divide the time granularity according to data characteristics: network performance data (15 minutes / window), network alarm data (1 minute / window, supporting alarm duration interval matching), and complaint data (time window of 1 hour before and after the complaint submission time), ensuring that the time dimension of time series data and event data overlaps.

[0038] 3. Base station cell: The combination of "unique base station code + cell PCI (Physical Cell Identifier)" is used as the location identifier. The GIS geocoding system spatially matches the fuzzy address (such as "near XX cell") and latitude and longitude coordinates in the complaint data with the coverage area of ​​the base station cell in the network data, and accurately associates it with the specific service cell.

[0039] In some embodiments, a three-dimensional association key mechanism of user identifier-time window-base station cell is used to perform data alignment processing on customer complaint data, network performance data, and network alarm data, including: Customer complaint data, network performance data, and network alarm data are each labeled with a three-dimensional association key to form a standardized index; Establish a distributed association engine, perform hash mapping and range matching based on the association key, and automatically aggregate complaint texts, performance indicators and alarm events of the same user in the same time window and the same service cell; For cross-cell handover scenarios, the overlap of time windows and signal handover logs are used to correlate data from multiple cells before and after the handover.

[0040] Specifically, the alignment execution process of this application includes: 1. In the data preprocessing stage, three-dimensional association key labels are assigned to the three types of data to form a standardized index; 2. Establish a distributed association engine, perform hash mapping and range matching based on the association key, and automatically aggregate complaint texts, performance indicators and alarm events of the same user in the same time window and the same service cell; 3. For cross-cell handover scenarios (such as when a user complains during a cell handover process), the overlap of time windows (with a threshold set to 70%) and signal handover logs are used to associate data from multiple cells before and after the handover to ensure data integrity.

[0041] In some embodiments, a user identifier-time window-base station cell three-dimensional association key mechanism is used to perform data alignment processing on customer complaint data, network performance data, and network alarm data, and the method further includes: Data normalization processing is performed on customer complaint data, network performance data, and network alarm data, specifically including: Customer complaint data normalization: A three-level processing approach is adopted, including text cleaning, keyword standardization, and semantic mapping. This includes filtering special characters and redundant expressions; standardizing synonyms and colloquial expressions through a custom dictionary; and associating standardized keywords with a network indicator dictionary. Network performance data normalization: To address the differences in the dimensions of different base station equipment and different indicator dimensions, a segmented normalization + dynamic threshold adaptation method is adopted, including: dividing the indicator value into sub-intervals according to the business scenario, and using Z-score standardization in each interval; and dynamically adjusting the normalization parameters in combination with the regional network quality benchmark value. Network alarm data normalization: Construct a standardized mapping table between alarm level and impact range: unify alarm codes from different equipment manufacturers into a three-dimensional label of alarm type-severity-affected user range; quantify alarm duration into three categories: instantaneous alarm, short-term alarm, and long-term alarm.

[0042] Specifically, this application adopts a data normalization strategy, including: 1. Normalization of unstructured complaint texts The process employs a three-tiered approach: text cleaning, keyword standardization, and semantic mapping. First, special characters and redundant expressions are filtered out (e.g., "Please resolve the slow internet issue" is cleaned to "slow internet"). Then, synonyms and colloquial expressions are standardized using a custom dictionary (e.g., "lag," "stuck," and "slow loading" are uniformly mapped to "high network latency"). Finally, standardized keywords are associated with a network metrics dictionary (e.g., "unable to access the internet" is mapped to "downlink speed = 0" and "access success rate < 10%)," thus achieving a structured transformation of the text data.

[0043] 2. Normalization of Time-Series Network Performance Data

[0044] To address the differences in metrics across different base station devices and metrics dimensions (such as rate, latency, and call drop rate), a "segmented normalization + dynamic threshold adaptation" approach is adopted: metric values ​​are divided into sub-intervals based on service scenarios (voice, internet access, video), and Z-score standardization is applied within each interval; simultaneously, the normalization parameters are dynamically adjusted in conjunction with regional network quality benchmarks (such as the difference in rate thresholds between urban and suburban areas) to avoid metric distortion caused by regional differences.

[0045] 3. Normalization of event-based network alarm data

[0046] Construct a standardized mapping table between alarm levels and impact scope: unify the alarm codes of different equipment manufacturers (such as base station fault codes of Huawei and Ericsson) into a three-dimensional label of "alarm type - severity - scope of affected users" (such as "transmission interruption - emergency - all users in the cell"); at the same time, quantify the alarm duration into three categories: "instantaneous alarm (<5 minutes)", "short-term alarm (5-30 minutes)" and "long-term alarm (>30 minutes)" to facilitate matching with time windows.

[0047] In some embodiments, a user identifier-time window-base station cell three-dimensional association key mechanism is used to perform data alignment processing on customer complaint data, network performance data, and network alarm data, and the method further includes: Missing value imputation is performed on customer complaint data, network performance data, and network alarm data, specifically including: Based on the classification and imputation strategies corresponding to different types of data, missing values ​​are imputed for customer complaint data, network performance data, and network alarm data. After the data is completed, cross-validation and anomaly detection are performed. The classification completion strategy includes: For missing customer complaint data: If the core issue description is missing in the complaint ticket, the Naive Bayes algorithm is used to infer possible complaint issues and generate standardized issue labels by associating the same user's historical complaint types, network alarm types and performance anomaly indicators during the same period. For missing network performance data: a time-series interpolation + similar cell reference fusion method is adopted, including: for short-term missing data, linear interpolation is used to fill in the gaps; for long-term missing data, similar cells with the same region, user density, and configuration are found based on the three-dimensional association key, and the gaps are filled in based on the average index of similar cells during the same period. For missing network alarm data: perform reverse completion based on the performance anomaly-alarm association rule base.

[0048] Specifically, this application employs a missing value imputation scheme, and the classification imputation strategy includes: 1. Missing network performance data: A fusion method of "time-series interpolation + similar cell reference" is adopted. For short-term missing data (≤1 time window), linear interpolation is used to fill in the gaps; for long-term missing data (>1 time window), similar cells with the same region, user density, and configuration are found based on the three-dimensional correlation key, and the data is filled in using the average of their corresponding indicators, with the filling error controlled within 10%.

[0049] 2. Missing complaint text data: If the core problem description is missing in the complaint work order, the Naive Bayes algorithm is used to infer the possible complaint problem by associating the same user's historical complaint types, the same network alarm types and performance anomaly indicators, and generating standardized problem labels (e.g., if the user has complained multiple times about "voice lag" in the past, and there is a "voice channel interference" alarm in the same cell at the same time, the missing problem is inferred to be "voice quality problem".

[0050] 3. Missing network alarm data: Reverse completion using the "Performance Anomaly - Alarm Association Rule Base". If a cell experiences significant performance degradation within a specific time window (e.g., a sudden 30% increase in call drop rate) but there are no corresponding alarm records, the system matches possible alarm types (e.g., "wireless link failure") based on the rule base, marks them as "latent alarms", and adds them to the data.

[0051] In addition, this application adopts a missing value verification mechanism, including: after completion, the reliability of the data is ensured by "cross-validation + anomaly detection": the completed data is checked for consistency with adjacent time windows and similar cell data. If the deviation exceeds a preset threshold (such as 20%), manual review is triggered; at the same time, the isolated forest algorithm is used to identify abnormal completed data to avoid incorrect filling affecting the model performance.

[0052] S3. Based on the multi-source aligned data, perform model training to obtain an intelligent complaint analysis model; Figure 2 This is a schematic diagram illustrating model training and application provided in the embodiments of this application, such as... Figure 2 As shown, in this application, the intelligent complaint analysis model adopts the random forest algorithm combined with machine learning to continuously optimize the accuracy of complaint analysis. Different judgment algorithms are used in each step of the decision-making process to achieve the goal of accurately judging customer problems. Taking user complaints about network problems as an example: the initial data of this model comes from user complaint work orders. The data content of the work orders includes the direction of the complaint, 3G / 4G signal level and quality, base station location, and problem results.

[0053] After the program starts running, the crawler subroutine and the user-associated residential cell subroutine run synchronously via coroutines. The crawler subroutine analyzes and reverse-engineers JavaScript to simulate logging into the indicator monitoring platform. After analyzing network requests, it constructs parameters to simulate GET and POST requests, sending requests to the backend to obtain data on cell performance indicators and monitoring alarms. The user-associated residential cell subroutine matches daily complaint details with the residential cells, accurately locating the 3G and 4G cells frequently used by the complaining user. The crawler and user-associated residential cell subroutine run concurrently via coroutines, then associate indicator and alarm data with the complaining user's residential cell, providing a clear view for complaint handling personnel to analyze and locate problems. Finally, it outputs solutions and terminates the program.

[0054] Figure 3 This is a schematic diagram of the intelligent complaint analysis provided in the embodiments of this application, such as... Figure 3 As shown, other functions of complaint analysis in this application include: The map display function geographically links problematic base stations with complaint points, enabling faster indicator analysis based on complaint issues.

[0055] The complaint cleaning function automatically filters out non-moveable mesh quality issues and intelligently corrects the latitude and longitude of user complaint points. The indicator integration module performs data mining and cross-analysis on network data and presents the final learning selection results.

[0056] Figure 4 This is a schematic diagram illustrating the acquisition of user complaint cells provided in an embodiment of this application, such as... Figure 4 As shown, this application first extracts the user complaint area, crawls and converts the data using the Baidu API to obtain the latitude and longitude of the user complaint. The user's latitude and longitude are then matched with existing network parameters to identify five surrounding sites and 15 corresponding residential areas. These are then compared with the residential areas where the user is usually located, and the intersection is used to accurately pinpoint the residential area occupied by the user.

[0057] Figure 5 A schematic diagram illustrating complaint location provided in the embodiments of this application, such as... Figure 5 As shown, this application analyzes and judges various KPIs and KQI indicators of the target cell based on user complaints to determine whether the indicators of the target cell are normal. First, based on the complained cell, alarm data during the same period is matched to determine whether the problem is caused by an alarm. Then, the cell-level noise floor is judged to determine whether the problem is caused by interference. Finally, based on the network management and SEQ platform data, the KPI and KQI indicators are compared to locate and analyze the factors affecting the indicators.

[0058] In some embodiments, S3 includes S31-S34, specifically as follows: S31. Construction of multi-label classification model, including: Feature extraction layer: XGBoost and LightGBM are trained in parallel to extract high-dimensional features from multi-source aligned data. XGBoost is used to capture linear and low-order nonlinear features, while LightGBM is used to mine high-order nonlinear features. The output features of the two models are concatenated to form a fused feature vector. Classification and prediction layer: Construct a lightweight neural network, input fused feature vectors, and output multi-dimensional probability of complaint problem types; Specifically, in constructing the multi-label classification model, this application innovatively adopts a hybrid architecture of GBDT + Neural Network, including: The first stage (feature extraction layer) employs parallel training of XGBoost and LightGBM to extract high-dimensional features. XGBoost is responsible for capturing linear features and low-order nonlinear features (such as the relationship between indicator thresholds), while LightGBM is responsible for mining high-order nonlinear features (such as the interaction of multiple indicators). The output features of the two models are then concatenated to form a fused feature vector.

[0059] The second stage (classification prediction layer): Constructs a lightweight neural network, inputs a fused feature vector, and outputs multi-dimensional probability of complaint types (8 core issues including voice quality, internet speed, service coverage, and package pricing). The neural network adopts a 3-layer fully connected structure, with ReLU activation function in the hidden layers and a Sigmoid function in the output layer (supporting single-user prediction of multiple issue types, such as simultaneously complaining about "slow internet" and "poor signal").

[0060] In some embodiments, S31 further includes: a sample imbalance solution, specifically including at least one of the following: Sample expansion: A combined SMOTE+ synthetic sample generation method is adopted, including: generating similar samples for minority complaint samples using the SMOTE algorithm; and synthesizing virtual samples that conform to business logic based on real complaint texts and online data using generative adversarial networks. Loss function improvement: A weighted FocalLoss + cross-entropy fusion loss function is adopted, including: assigning higher weights to minority class samples; introducing a FocalLoss modulation factor to reduce the loss contribution of easily classified samples and focus on difficult-to-classify samples; Model training strategy: adopts a stratified sampling + incremental training mode, including: stratified sampling of the training set according to the type of complaint; establishing an incremental training mechanism to update the top-level neural network parameters of the model after adding new complaint data and network data.

[0061] Specifically, this application employs a solution to the imbalanced sample problem, including at least one of the following: 1. Sample expansion A combined approach of "SMOTE + synthetic sample generation" is adopted. For a few types of complaint samples (such as package pricing issues), similar samples are generated using the SMOTE algorithm; at the same time, based on real complaint texts and network data, virtual samples that conform to business logic are synthesized using generative adversarial networks (GANs) (such as complaint texts simulating "weak signal in remote areas" and corresponding network performance data), so that the ratio of various types of samples tends to be 1:1.

[0062] 2. Improved loss function

[0063] We employ a "weighted Focal Loss + cross-entropy fusion loss function" to address the issues of sample imbalance and difficulty in distinguishing between easy and difficult samples in multi-label scenarios. We assign higher weights to minority class samples (e.g., a weight of 3 for package pricing issues and a weight of 1 for internet speed issues); simultaneously, we introduce a modulation factor from Focal Loss to reduce the loss contribution of easily classified samples and focus on difficult-to-classify samples (e.g., the ambiguous boundaries between "voice stuttering" and "voice interruption").

[0064] 3. Model Training Strategy

[0065] A "stratified sampling + incremental training" model is adopted. The training set is stratified by complaint type to ensure that each batch of training data includes samples of all types. At the same time, an incremental training mechanism is established. After adding new complaint data and network data each day, only the top-level neural network parameters of the model are updated, avoiding the waste of resources caused by full retraining and improving the efficiency of model iteration.

[0066] S32. Model parameter tuning and optimization, including: Coarse tuning: A Bayesian optimization algorithm is used to perform a global search on the core parameters of XGBoost / LightGBM and the network parameters of the neural network to obtain the optimal parameter range; Fine-tuning: Within the optimal parameter range, a grid search combined with 5-fold cross-validation is used to determine the final parameters; Specifically, this application employs a two-stage parameter tuning strategy for model optimization, including: Phase 1 (Coarse Adjustment): Using the Bayesian optimization algorithm, a global search is performed on the core parameters of XGBoost / LightGBM (learning rate, tree depth, number of leaf nodes, etc.) and the number of hidden layer neurons and learning rate of the neural network to quickly locate the optimal parameter range.

[0067] The second stage (fine-tuning): Within the optimal interval, a grid search combined with 5-fold cross-validation is used to determine the final parameters. The focus is on optimizing LightGBM's "max_bin" (which controls the granularity of feature discretization) and the neural network's "regularization coefficient" (to avoid overfitting).

[0068] Furthermore, the model performance evaluation metrics in this application adopt multi-label classification-specific metrics: Hamming Loss, Macro-F1, and Micro-F1. The objectives are Hamming Loss ≤ 0.05, Macro-F1 ≥ 0.85, and Micro-F1 ≥ 0.90, to ensure a balanced prediction accuracy for various types of complaints.

[0069] S33. User-network relationship graph construction, including: Graph structure definition: Node types include user nodes and base station cell nodes. User node attributes include historical complaint records, package type, resident cell, and usage behavior; base station cell node attributes include network performance indicators, alarm records, coverage area, and user density. Edge types and weights: Spatial association edges connect user nodes in the same cell, with weights based on the percentage of time a user spends in the cell; Social association edges are built based on user relationships through calls, text messages, and shared data plans, with weights based on interaction frequency; Service association edges connect user nodes with nodes in their resident cell, with weights based on average signal strength. Graph dynamic update mechanism: Real-time synchronization of user location changes and cell network status, dynamically adjusting the existence and weight of edges; Specifically, this application uses a user-network relationship graph for construction, including: 1. Definition of graph structure Node types include user nodes and base station cell nodes. User node attributes include historical complaint records, package type, resident cell, and usage behavior (such as call duration and data usage); base station cell node attributes include network performance indicators, alarm records, coverage area, and user density.

[0070] 2. Edge type and weight

[0071] Spatial association edge: connects user nodes in the same community, with weights based on the percentage of time a user spends in the community (e.g., if user A spends 80% of their time in community X, the weight is set to 0.8). Social connection edge: Based on user relationships such as calls, text messages, and shared packages, the weight is based on the frequency of interaction (e.g., if user A and user B make an average of 3 calls per day, the weight is set to 0.3). Service association edge: connects the user node with the node of the cell where it resides, and the weight is based on the average signal strength (e.g., if the average RSRP of user A in cell X is -85dBm, the weight is set to 0.9).

[0072] 3. Dynamic graph update mechanism

[0073] Real-time synchronization of user location changes (based on SEQ data) and cell network status dynamically adjusts the existence and weight of edges. For example, when a user switches to a new cell, spatial association edges of the original cell are deleted and spatial association edges of the new cell are added; when cell performance deteriorates, the weight of service association edges between the cell and the user node is reduced.

[0074] S34. Graph neural network prediction models, including: Neighbor sampling optimization: A hierarchical sampling + weight-first strategy is adopted, including: when sampling the neighbors of user nodes, neighbors with higher weights are selected first, and the number of samples per layer is limited; Feature fusion improvement: Introduce a node type attention mechanism into the aggregation function, including: when aggregating neighbor features, assign different attention to user node features and cell node features; Mass Complaint Risk Assessment: The model outputs a mass complaint risk score for user nodes, identifies high-risk users based on the risk score threshold, identifies the core propagation path of mass complaints by analyzing the connected subgraphs of high-risk users, and outputs the risk propagation probability to predict the scope of complaint spread.

[0075] Specifically, this application employs a graph neural network prediction model, and the model selection and improvement are based on an enhanced version of GraphSAGE, including: 1. Neighbor sampling optimization A "hierarchical sampling + weight-first" strategy is adopted. When sampling the neighbors of user nodes, neighbors with high weights (such as users who stay in the same community for a long time and social friends who interact frequently) are selected first to ensure that the sampled nodes have stronger relevance; at the same time, the number of samples in each layer is limited (20 in the first layer and 10 in the second layer) to improve the model training efficiency.

[0076] 2. Feature fusion improvement

[0077] A "node type attention mechanism" is introduced into the aggregation functions of GraphSAGE. When aggregating neighbor features, different attention weights are assigned to user node features and cell node features (e.g., when network alarms occur frequently, the weight of cell node features is increased to 0.6, and the weight of user node features is set to 0.4), thereby enhancing the influence of key information.

[0078] 3. Risk assessment of group complaints

[0079] The model outputs a "group complaint risk score" (0-100) for user nodes, with a score ≥60 indicating a high-risk user. By analyzing the connected subgraphs of high-risk users, the model identifies the core propagation path of group complaints (e.g., multiple high-risk users in a community are connected by spatial association edges, suggesting that a group complaint will occur in that community); it also outputs the risk propagation probability to predict the scope of complaint spread (e.g., "Within the next 24 hours, 30% of users in this community will complain about network coverage issues").

[0080] After modeling is completed, the data model is tested using actual data samples, the model is evaluated, and the model is optimized based on the evaluation results.

[0081] In this application, the programming runtime libraries NumPy and Pandas are used to process large amounts of data.

[0082] NumPy (Numerical Python) is an extension library for the Python language that supports large-dimensional arrays and matrix operations, and also provides a large library of mathematical functions for array operations. It boasts powerful scientific computing capabilities, including: 1. A powerful N-dimensional array object (Array); 2. A relatively mature (broadcast) function library; 3. A toolkit for integrating C / C++ and Fortran code; 4. Practical linear algebra, Fourier transform, and random number generation functions. Using NumPy in conjunction with the sparse matrix operation package SciPy makes the process even more convenient. NumPy (Numeric Python) provides many advanced numerical programming tools, such as matrix data types, vector processing, and sophisticated computation libraries. It was specifically designed for rigorous numerical processing.

[0083] Pandas is a Python data analysis package, part of the PyData project. Originally developed as a financial data analysis tool, Pandas provides excellent support for time series analysis. The name Pandas comes from panel data and Python data analysis. Panel data is an economics term for multidimensional datasets, and Pandas also provides a panel data type. Furthermore, Pandas incorporates a large number of libraries and standard data models, providing the tools needed to efficiently manipulate large datasets. Pandas offers a wealth of functions and methods that enable us to process data quickly and easily.

[0084] S4. Complaint analysis and processing are performed using the intelligent complaint analysis model.

[0085] This application utilizes big data for complaint preprocessing analysis, predicting potential complaints from all users and identifying potential issues that might arise before users experience them. Network optimization efforts are initiated ahead of time, prioritizing optimization for users potentially switching networks while keeping their numbers. Based on historical complaint records, proactive customer care is provided to improve user experience and satisfaction.

[0086] This application provides a complaint pre-processing analysis method. It analyzes user perception discrepancies during business usage through data mining, trains samples based on implemented user data, explores input data, analyzes complaining users, and identifies network issues through user perception dimensions. This shifts the focus from KPI-driven to KQI-driven approaches, improving network coverage. This lowers the barrier to complaint handling, shortens the average processing time for a single work order, reduces operational costs, and improves the efficiency and accuracy of complaint handling.

[0087] It should be understood that although the steps in the flowcharts of the above embodiments are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some of the steps in the figures may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times, and their execution order is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0088] Figure 6 This is a schematic diagram of the complaint preprocessing analysis device provided in the embodiments of this application, as shown below. Figure 6 As shown, this application provides a complaint preprocessing analysis device, the device comprising: Data acquisition module 11 is configured to acquire customer complaint data, network performance data, and network alarm data in network communication; The data alignment module 12 is configured to perform data alignment processing on customer complaint data, network performance data, and network alarm data to obtain multi-source aligned data; Model training module 13 is configured to train the model based on the multi-source aligned data to obtain an intelligent complaint analysis model; The analysis and processing module 14 is configured to perform complaint analysis and processing through the intelligent complaint analysis model.

[0089] Regarding the limitations on the complaint pre-processing analysis device, please refer to the limitations on the complaint pre-processing analysis method in the above embodiments of this application, which will not be repeated here.

[0090] Figure 7 Another schematic diagram of the complaint preprocessing analysis device provided in the embodiments of this application is shown below. Figure 7 As shown, the device includes a memory 22 and a processor 21. The memory stores a computer program, and the processor is configured to run the computer program to perform the methods described in the above embodiments of this application.

[0091] The memory is connected to the processor. The memory can be flash memory, read-only memory or other types of memory. The processor can be a central processing unit or a microcontroller.

[0092] In some embodiments, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the methods described in the above embodiments of this application.

[0093] The computer-readable storage medium includes volatile or non-volatile, removable or non-removable media implemented in any method or technology for storing information, such as computer-readable instructions, data structures, computer program modules or other data. Computer-readable storage media include, but are not limited to, RAM (Random Access Memory), ROM (Read-Only Memory), EEPROM (Electrically Erasable Programmable Read-Only Memory), flash memory or other memory technologies, CD-ROM (Compact Disc Read-Only Memory), DVD or other optical disc storage, cartridges, magnetic tapes, disk storage or other magnetic storage devices, or any other medium that can be used to store desired information and is accessible to a computer.

[0094] It is understood that the above embodiments are merely exemplary implementations used to illustrate the principles of this application, and this application is not limited thereto. For those skilled in the art, various modifications and improvements can be made without departing from the spirit and substance of this application, and these modifications and improvements are also considered to be within the scope of protection of this application.

Claims

1. A method for pre-processing and analyzing complaints, characterized in that, The method includes: S1. Obtain customer complaint data, network performance data, and network alarm data from network communications; S2. Perform data alignment processing on customer complaint data, network performance data, and network alarm data to obtain multi-source aligned data; S3. Based on the multi-source aligned data, perform model training to obtain an intelligent complaint analysis model; S4. Complaint analysis and processing are performed using the intelligent complaint analysis model.

2. The complaint pre-processing analysis method according to claim 1, characterized in that, S2 includes: A three-dimensional association key mechanism of user identifier-time window-base station cell is adopted to perform data alignment processing on customer complaint data, network performance data and network alarm data to obtain multi-source aligned data; Among them, the user identifier adopts a dual identifier system of communication number desensitized identity ID + SEQ user unique identifier, the time window is dynamically divided into time granularity according to data characteristics, and the base station cell adopts a combination of base station unique code + cell physical cell identifier (PCI) as the location identifier.

3. The complaint pre-processing analysis method according to claim 2, characterized in that, A three-dimensional association key mechanism of user identifier-time window-base station cell is adopted to perform data alignment processing on customer complaint data, network performance data, and network alarm data, including: Customer complaint data, network performance data, and network alarm data are each labeled with a three-dimensional association key to form a standardized index; Establish a distributed association engine, perform hash mapping and range matching based on the association key, and automatically aggregate complaint texts, performance indicators and alarm events of the same user in the same time window and the same service cell; For cross-cell handover scenarios, the overlap of time windows and signal handover logs are used to correlate data from multiple cells before and after the handover.

4. The complaint pre-processing analysis method according to claim 2, characterized in that, A three-dimensional association key mechanism of user identifier-time window-base station cell is adopted to perform data alignment processing on customer complaint data, network performance data, and network alarm data, and also includes: Data normalization processing is performed on customer complaint data, network performance data, and network alarm data, specifically including: Customer complaint data normalization: A three-level processing approach is adopted, including text cleaning, keyword standardization, and semantic mapping. This includes filtering special characters and redundant expressions; standardizing synonyms and colloquial expressions through a custom dictionary; and associating standardized keywords with a network indicator dictionary. Network performance data normalization: To address the differences in the dimensions of different base station equipment and different indicator dimensions, a segmented normalization + dynamic threshold adaptation method is adopted, including: dividing the indicator value into sub-intervals according to the business scenario, and using Z-score standardization in each interval; and dynamically adjusting the normalization parameters in combination with the regional network quality benchmark value. Network alarm data normalization: Construct a standardized mapping table between alarm level and impact range: unify alarm codes from different equipment manufacturers into a three-dimensional label of alarm type-severity-affected user range; quantify alarm duration into three categories: instantaneous alarm, short-term alarm, and long-term alarm.

5. The complaint pre-processing analysis method according to claim 2, characterized in that, A three-dimensional association key mechanism of user identifier-time window-base station cell is adopted to perform data alignment processing on customer complaint data, network performance data, and network alarm data, and also includes: Missing value imputation is performed on customer complaint data, network performance data, and network alarm data, specifically including: Based on the classification and imputation strategies corresponding to different types of data, missing values ​​are imputed for customer complaint data, network performance data, and network alarm data. After the data is completed, cross-validation and anomaly detection are performed. The classification completion strategy includes: For missing customer complaint data: If the core issue description is missing in the complaint ticket, the Naive Bayes algorithm is used to infer possible complaint issues and generate standardized issue labels by associating the same user's historical complaint types, network alarm types and performance anomaly indicators during the same period. For missing network performance data: a time-series interpolation + similar cell reference fusion method is adopted, including: for short-term missing data, linear interpolation is used to fill in the gaps; for long-term missing data, similar cells with the same region, user density, and configuration are found based on the three-dimensional association key, and the gaps are filled in based on the average index of similar cells during the same period. For missing network alarm data: perform reverse completion based on the performance anomaly-alarm association rule base.

6. The complaint pre-processing analysis method according to claim 1, characterized in that, S3 includes: S31. Construction of multi-label classification model, including: Feature extraction layer: XGBoost and LightGBM are trained in parallel to extract high-dimensional features from multi-source aligned data. XGBoost is used to capture linear and low-order nonlinear features, while LightGBM is used to mine high-order nonlinear features. The output features of the two models are concatenated to form a fused feature vector. Classification and prediction layer: Construct a lightweight neural network, input fused feature vectors, and output multi-dimensional probability of complaint problem types; S32. Model parameter tuning and optimization, including: Coarse tuning: A Bayesian optimization algorithm is used to perform a global search on the core parameters of XGBoost / LightGBM and the network parameters of the neural network to obtain the optimal parameter range; Fine-tuning: Within the optimal parameter range, a grid search combined with 5-fold cross-validation is used to determine the final parameters; S33. User-network relationship graph construction, including: Graph structure definition: Node types include user nodes and base station cell nodes. User node attributes include historical complaint records, package type, resident cell, and usage behavior; base station cell node attributes include network performance indicators, alarm records, coverage area, and user density. Edge types and weights: Spatial association edges connect user nodes in the same cell, with weights based on the percentage of time a user spends in the cell; Social association edges are built based on user relationships through calls, text messages, and shared data plans, with weights based on interaction frequency; Service association edges connect user nodes with nodes in their resident cell, with weights based on average signal strength. Graph dynamic update mechanism: Real-time synchronization of user location changes and cell network status, dynamically adjusting the existence and weight of edges; S34. Graph neural network prediction models, including: Neighbor sampling optimization: A hierarchical sampling + weight-first strategy is adopted, including: when sampling the neighbors of user nodes, neighbors with higher weights are selected first, and the number of samples per layer is limited; Feature fusion improvement: Introduce a node type attention mechanism into the aggregation function, including: when aggregating neighbor features, assign different attention to user node features and cell node features; Mass Complaint Risk Assessment: The model outputs a mass complaint risk score for user nodes, identifies high-risk users based on the risk score threshold, identifies the core propagation path of mass complaints by analyzing the connected subgraphs of high-risk users, and outputs the risk propagation probability to predict the scope of complaint spread.

7. The complaint pre-processing analysis method according to claim 6, characterized in that, S31 also includes: solutions to sample imbalance, specifically including at least one of the following: Sample expansion: A combined SMOTE+ synthetic sample generation method is adopted, including: generating similar samples for minority complaint samples using the SMOTE algorithm; and synthesizing virtual samples that conform to business logic based on real complaint texts and online data using generative adversarial networks. Loss function improvement: A weighted FocalLoss + cross-entropy fusion loss function is adopted, including: assigning higher weights to minority class samples; introducing a FocalLoss modulation factor to reduce the loss contribution of easily classified samples and focus on difficult-to-classify samples; Model training strategy: adopts a stratified sampling + incremental training mode, including: stratified sampling of the training set according to the type of complaint; establishing an incremental training mechanism to update the top-level neural network parameters of the model after adding new complaint data and network data.

8. A complaint pre-processing analysis device, characterized in that, The device includes: The data acquisition module is configured to acquire customer complaint data, network performance data, and network alarm data from network communications. The data alignment module is configured to perform data alignment processing on customer complaint data, network performance data, and network alarm data to obtain multi-source aligned data; The model training module is configured to train the model based on the multi-source aligned data to obtain an intelligent complaint analysis model. The analysis and processing module is configured to perform complaint analysis and processing through the intelligent complaint analysis model.

9. A complaint pre-processing analysis device, characterized in that, It includes a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to implement the complaint preprocessing analysis method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the complaint preprocessing analysis method as described in any one of claims 1-7.

Images