Telecommunication service quality evaluation method and system based on user emotion analysis

Abstract

The invention relates to the technical field of telecommunication services, in particular to a telecommunication service quality evaluation method and system based on user emotion analysis, which comprises the following steps of: deploying a digital tip receptor, monitoring and coding in real time to generate a first data stream and a second data stream which carry timestamps; processing the first data stream by using a hierarchical cross-modal attention model to generate an emotional trajectory, and determining an emotional abrupt change event and a time point thereof by analyzing the geometric morphology of the emotional trajectory; in response to the determined emotion mutation event, combining the second data streams in the backtracking time window before the event occurs into an observation symptom vector; matching the observation symptom vector with a causal attribution matrix, determining a technical root corresponding to a matrix row closest to the observation symptom vector in learning distance based on a matching result, taking the technical root as an evaluation result of a telecommunication service quality problem, and generating a structured diagnosis portrait, according to the invention, the telecommunication service quality problem which affects the user experience can be found and positioned more accurately and timely.

Description

Technical Field

[0001] This invention relates to the technical field of telecommunications services, and more specifically, to a method and system for evaluating the quality of telecommunications services based on user sentiment analysis. Background Technology

[0002] In the era of digital services, ensuring and improving user experience is a core competitive advantage for telecom operators and service providers. Traditional service quality assurance methods mainly rely on application performance monitoring (APM) and network monitoring systems, assessing system health by monitoring objective Quality of Service (QoS) indicators such as latency, packet loss rate, and server CPU utilization. However, there is a significant "semantic gap" between these server-side indicators and the user's actual subjective experience; that is, normal QoS indicators do not fully represent a satisfactory QoE for the user.

[0003] To bridge this gap, the video quality assessment method and device based on emotion recognition and machine learning disclosed in announcement number CN115690652A collects users' emotion values ​​while watching videos and combines them with post-event questionnaire scores to train a model for predicting video quality. This technology relies on post-event, holistic user evaluations, suffers from severe feedback lag, and cannot effectively assess specific problematic time periods during the experience. It averages or generalizes the emotions throughout the process, completely losing the correlation with specific instantaneous events. Essentially, it is a macroscopic model for predicting content popularity, rather than a real-time system for diagnosing service failures.

[0004] Furthermore, the customer service quality assessment method disclosed in announcement number CN119130219A acquires call text, voice, and feature information during a call. These are then processed separately using a first model within the service quality assessment model to obtain call content vectors, voice emotion vectors, and call feature vectors. A second model then fuses these vectors to generate the final user evaluation information. While this multimodal fusion method can integrate more information and arrive at a more accurate overall evaluation than a single dimension, its core idea remains information aggregation and weighting, aiming to answer the question of the overall quality of the service. Current technology suffers from the following problems:

[0005] Existing fusion models cannot immediately establish a direct and reliable association between the result of a user's negative emotion (such as frustration caused by application lag) and the specific technical event that caused it (e.g., a timeout of a particular API call or a TCP retransmission) the moment it occurs. The system knows the user is dissatisfied, but it doesn't know which technical fault caused this dissatisfaction at what point in time.

[0006] Many minor user experiences that degrade the user experience (such as a 0.5-second UI rendering delay or an illogical interface transition) do not trigger serious alerts in traditional APM systems, nor do users complain about them. However, the cumulative effect of these minor user experiences is a major cause of silent user churn. Existing technologies are almost powerless to perceive these minor user events themselves and understand how they work together to ultimately lead to a deterioration in user experience.

[0007] Therefore, this field urgently needs a completely new technological paradigm that should not merely remain at the level of macroscopic, lagging "assessment" or "scoring" of service quality, but rather endow telecommunications service systems with inherent, real-time "perception and diagnosis" capabilities. At the moment a user experience is impaired, it should not only be able to perceive the occurrence of pain, but also instantly trace and locate the physical lesion causing the pain along neural pathways. Summary of the Invention

[0008] Therefore, the purpose of this invention is to provide a method and system for evaluating the quality of telecommunications services based on user sentiment analysis, which can improve the overall quality of telecommunications services and user satisfaction.

[0009] To achieve the above objectives, the present invention provides the following technical solution:

[0010] The telecommunications service quality assessment method based on user sentiment analysis includes the following steps:

[0011] S1. Deploy multiple digital peripheral sensors in the client application. The sensors listen for and encode harmful technical events related to negative user experiences in real time to generate a first data stream and a second data stream carrying timestamps with nanosecond precision.

[0012] S2. Using a hierarchical cross-modal attention model, process the first data stream to generate an emotional trajectory that represents the changes in the user's emotional state in a high-dimensional vector space.

[0013] S3. By analyzing the geometric shape of the emotional trajectory, identify the state space transitions that occur within a preset time window to determine the emotional mutation events and their timing.

[0014] S4. In response to identifying a mood mutation event, combine the second data stream within the retrospective time window prior to the event into an observed symptom vector;

[0015] S5. Match the observed symptom vectors with the preset causal attribution matrix;

[0016] S6. Based on the matching results, determine the technical root cause corresponding to the matrix row that is closest to the observed symptom vector in terms of learning distance, use it as the assessment result of the telecommunications service quality problem, and generate a packet structured diagnostic profile.

[0017] The present invention is further configured such that: the digital peripheral receptor includes:

[0018] The smoothness damage sensor checks the state of the UI rendering thread at a frequency higher than the visual refresh rate of the user interface through the microtask scheduler. The sensor is activated when the rendering thread is continuously blocked or the event queue length exceeds a preset threshold.

[0019] The interactive entropy increase sensor monitors the sequence of user actions in a specific business process and calculates the rate of increase of information entropy in the sequence in real time. When the rate of increase of entropy exceeds a preset threshold, the sensor is activated to quantify the user's level of confusion and frustration.

[0020] The present invention is further configured such that the encoding step of a harmful technical event includes:

[0021] When any digital peripheral receptor is activated, a structured noxious stimulus signal is generated;

[0022] The data structure of the stimulus signal is a quadruple, which includes a nanosecond-level timestamp, a unique receptor ID, a pulse intensity vector, and a local context summary ID.

[0023] The pulse intensity vector is a vector of at least two values ​​used to quantify multiple dimensions of severity in this harmful technical incident.

[0024] The present invention is further configured such that the structure of the hierarchical cross-modal attention model includes:

[0025] The intramodal temporal layer contains at least two parallel gated recurrent unit networks, which perform temporal modeling on the speech acoustic feature sequence and the user operation behavior sequence, respectively, to extract the temporal dynamic features of each modality.

[0026] The cross-modal attention layer receives the output of the intramodal temporal layer and learns and fuses cross-modal cooperative change patterns by calculating the attention weights of the hidden state sequence of one modality to the hidden state sequence of another modality.

[0027] The present invention is further configured such that the step of analyzing the geometric shape of the emotional trajectory includes:

[0028] The emotional trajectory is viewed as a parametric curve in a high-dimensional space;

[0029] Calculate the average curvature of the curve within a preset time window;

[0030] A state space transition is determined when the average curvature exceeds a curvature threshold calibrated based on historical data.

[0031] The present invention is further configured such that: each row of the causal attribution matrix corresponds to a specific, locatable technical root cause identifier.

[0032] The technical root cause identifier points to fault types including slow database queries, content delivery network node failures, or timeouts of specific third-party interfaces;

[0033] Each column of the matrix corresponds to one dimension of the observed symptom vector;

[0034] The values ​​in the matrix represent the strength of the association between a particular technological root cause and a particular combination of symptoms.

[0035] The present invention is further configured to: match the observed symptom vector with the causal attribution matrix using a pre-trained learning distance metric model, and calculate the semantic distance between the observed symptom vector and each row vector in the causal attribution matrix;

[0036] The present invention is further configured such that the generated structured diagnostic profile includes a diagnostic session ID, an emotional mutation timestamp, a determined technical root cause identifier, and an evidence vector.

[0037] A telecommunications service quality assessment system based on user sentiment analysis includes:

[0038] A client-side digital sensor deployment module is used to instantiate and manage multiple functionally specific digital peripheral sensors and to encode harmful technical events.

[0039] The hierarchical cross-modal attention analysis module is used to receive and process the first data stream encoded by the receptors to generate emotional state trajectories and detect their state space transitions.

[0040] The causal attribution matrix matching engine is used to respond to state space transitions, combine the second data stream into an observed symptom vector, and match it with the built-in causal attribution matrix to pinpoint the root cause of the technology.

[0041] The diagnostic profile generation module is used to generate a structured business quality diagnostic profile based on the technical root causes.

[0042] The present invention is further configured such that: the causal attribution matrix matching engine includes a learning distance metric neural network. When performing matching, the engine calculates the non-Euclidean distance between the observed symptom vector and each row vector in the causal attribution matrix through the neural network to determine the technical root that best matches the symptom.

[0043] Compared with the shortcomings of the prior art, the beneficial effects of the present invention are as follows:

[0044] This invention, by introducing the perception and analysis of user emotional states and combining it with underlying technical events, can more accurately and promptly identify and locate telecommunications service quality issues affecting user experience. Compared to traditional alarm methods based on technical indicator thresholds, this invention can detect situations where technical indicators have not yet reached preset thresholds. Through a causal attribution matrix, it can automatically associate user-perceived problems with specific technical root causes, greatly improving the efficiency of fault diagnosis and troubleshooting. Furthermore, the structured diagnostic profile provides maintenance personnel with clear problem information and repair suggestions, shortening fault recovery time and ultimately improving overall telecommunications service quality and user satisfaction. Attached Figure Description

[0045] Figure 1 This is a diagram of the overall system architecture of the present invention;

[0046] Figure 2 This is a diagram illustrating the operation of the digital peripheral receptor of the present invention.

[0047] Figure 3 This is a structural diagram of the hierarchical cross-modal attention model of the present invention;

[0048] Figure 4 This is the output image of the structured diagnostic profile of the present invention. Detailed Implementation

[0049] Reference Figures 1 to 4 The telecommunications service quality assessment method based on user sentiment analysis of this invention specifically includes the following steps:

[0050] Step 1: Deploy multiple functionally specialized digital peripheral sensors in the client application. The sensors listen for and encode harmful technical events related to negative user experiences in real time to generate a first data stream and a second data stream carrying timestamps with nanosecond precision.

[0051] Digital peripheral receptors specifically include fluency injury receptors, interaction entropy increase receptors, and other potential receptors.

[0052] Smoothness Noxious Sensor: This sensor focuses on detecting stuttering and unresponsiveness in the user interface (UI). It checks the state of the UI rendering thread at a dynamically adjusted frequency (e.g., 100Hz on high-end devices, 60Hz on low-end devices) via a microtask scheduler (e.g., an asynchronous task executor based on a priority queue). The rendering thread is responsible for drawing the application's UI onto the screen; its blocking causes the interface to become unresponsive. An event queue stores user input (such as touches and swipes) and system events. When the rendering thread is continuously detected to be blocked for multiple check cycles (e.g., three consecutive check cycles, or 30 milliseconds) or the event queue length exceeds a preset threshold (e.g., 50-200 events, adjusted according to device performance), it indicates insufficient system processing power, and the user will perceive stuttering. At this point, the smoothness noxious sensor is activated, generating a noxious stimulus signal.

[0053] The microtask scheduler uses low-latency timers provided by the platform (such as Android's Choreographer callback mechanism or iOS's CADisplayLink) to implement high-frequency state checks. The rendering thread state is determined by checking the thread's execution stack or waiting lock state. The event queue length is obtained by hooking the application framework's event dispatch mechanism. To reduce performance overhead, the event sampling mechanism only records critical events (such as events that block for more than 20 milliseconds). An interaction entropy increase sensor quantifies the degree of confusion and frustration experienced by users in specific business processes (e.g., online top-up, package change) by monitoring user action sequences (e.g., clicks, swipes, input, back, etc.). The information entropy growth rate is used as an indicator of increased uncertainty in user action sequences. When users encounter problems or confusion, their actions become random, repetitive, or deviate from the normal path, leading to a significant increase in information entropy. This sensor calculates the information entropy growth rate of the action sequence in real time. When the rate exceeds an adaptive threshold (e.g., 0.2-1.0 bits / second, dynamically adjusted based on user history), it indicates that the user has encountered difficulties or a process design flaw, generating a harmful stimulus signal.

[0054] The sequence of actions is captured by intercepting UI events. To reduce the computational burden on the client, the information entropy calculation is moved to edge computing nodes or the cloud, and the client only collects the raw action sequence. Information entropy is calculated based on a sliding time window (10-60 seconds), using the following formula:

[0055] ,in For operation type The frequency of entropy growth rate is defined as the change in entropy value per unit time, and the threshold is determined by user segmentation (novice vs. experienced users).

[0056] Other potential sensors: Deploy other sensors to monitor network latency (request round-trip time), page load failure rate, application crashes, abnormal battery consumption (such as excessive background tasks), and audio / video playback stuttering, etc., to comprehensively understand the user experience. For example, a battery consumption sensor can detect abnormal power consumption behavior (such as battery drop exceeding 0.5% per minute) and record relevant context (such as the number of background processes).

[0057] Encoding of noxious technical events: When any digital peripheral receptor is activated, a structured noxious stimulus signal is generated. The data structure is a quadruple with the following format: (nanosecond-level timestamp, receptor unique identifier ID, pulse intensity vector, local context summary ID).

[0058] Nanosecond-level timestamps: Use system-provided high-precision nanosecond-level timestamps (such as Unix timestamps, accurate to nanoseconds).

[0059] Receptor unique identifier ID: A unique string or integer used to identify which receptor is activated (1 for fluency receptors, 2 for interaction entropy increase receptors).

[0060] Pulse intensity vector: A multi-dimensional vector that quantifies the severity of an event.

[0061] Smoothness impact: Includes blocking duration (milliseconds) and maximum event queue length. When a rendering thread is detected to be blocked for 50 milliseconds and the maximum event queue length reaches 200, the pulse intensity vector is (50, 200).

[0062] Interaction entropy increase: This includes the current information entropy growth rate (bits / second) and the confusion index, which can be calculated based on the number of invalid operations (such as repeated clicks or frequent returns) performed by the user in a specific business process. If the information entropy growth rate is detected to be 0.8 bits / second and the user has performed 5 invalid operations in the current process, the impulse intensity vector is (0.8, 5).

[0063] Local Context Summary ID: An identifier pointing to context information, including screen / page name, business process stage, network type, device model, operating system version, etc. To protect privacy, the context information is stored after being anonymized on the client side (e.g., hashed user ID), and queries are performed by association with the ID.

[0064] Data stream generation:

[0065] The first data stream contains receptor signals directly related to the user's emotional state, such as fluency impairment signals and interaction entropy increase signals, which are used for emotion trajectory analysis.

[0066] The second data stream contains signals of all types of harmful technical events, used to construct observational symptom vectors.

[0067] The data stream is temporarily stored locally and uploaded in batches when a Wi-Fi connection is established, reducing bandwidth usage in weak network environments.

[0068] Step 2: Use a hierarchical cross-modal attention model to process the first data stream to generate an emotion trajectory that represents the changes in the user's emotional state in a high-dimensional vector space.

[0069] Hierarchical cross-modal attention model: includes intra-modal temporal layers and cross-modal attention layers.

[0070] The intramodal temporal layer contains two parallel gated recurrent unit (GRU) networks that process the fluency and interactive entropy-increasing signal sequences respectively, capturing temporal dependencies.

[0071] GRU model:

[0072] Update Gate: ;

[0073] Reset Door: ;

[0074] Candidate hidden state: ;

[0075] Current hidden state: ;

[0076] in, It is the input vector at the current moment (the sensor signal from the first data stream). It was the hidden state from a previous moment. It is the Sigmoid activation function. It is the hyperbolic tangent activation function. , , , , , It is a weight matrix. , , It is a bias vector. This represents element-wise multiplication.

[0077] Input sequence: The pulse intensity vector in the quadruple is extracted and standardized, and then used as the input to the two GRU networks.

[0078] Output: Each GRU network outputs a sequence of hidden states, representing the temporal dynamics of the fluency mode and the interaction entropy increase mode, respectively.

[0079] Cross-modal attention layer: Receives hidden state sequences from the outputs of two GRU networks. By calculating the attention weights of the hidden state sequence of one modality to the hidden state sequence of the other modality, it learns and fuses cross-modal cooperative variation patterns. It can learn that a sharp drop in fluency is often accompanied by frequent invalid clicks from the user.

[0080] Calculate attention score: ;

[0081] Calculate attention weights: ;

[0082] Calculate the weighted context vector: ;

[0083] Computational fusion representation: ;

[0084] in, Is the first modality (e.g., fluency) at time step The hidden state, Is the second mode (e.g., interaction entropy increase) at time step The hidden state, , , , , , These are learnable parameters. This represents the length of the second modality sequence, and [;] indicates vector concatenation. This demonstrates the attention calculation of the first modality on the second modality, and vice versa.

[0085] Output: The cross-modal attention layer outputs a hidden state sequence that incorporates cross-modal information.

[0086] Emotional trajectory generation: The output sequence of the cross-modal attention layer is further processed (through a fully connected layer) and mapped to a high-dimensional vector space (128-512 dimensional), representing the emotional trajectory of the user's emotional state changing in the high-dimensional space.

[0087] Step 3: By analyzing the geometric shape of the emotional trajectory, identify the state space transitions that occur within a preset time window to determine the emotional mutation events and their timing.

[0088] Analyzing the geometric shape of emotional trajectories:

[0089] The generated high-dimensional vector sequence is viewed as a parametric curve that changes with time in a high-dimensional vector space. , It is the dimension of the emotion vector.

[0090] The average curvature within a preset time window (1-10 seconds) is calculated using the following approximate method:

[0091] Discrete point sequence: Calculate three consecutive points The cosine of the angle between the vectors measures the degree of curvature.

[0092] Identifying state space transitions:

[0093] A curvature threshold is determined based on historical data. This threshold represents the range of curvature changes in a user's emotions under normal, stable conditions.

[0094] Within a preset time window, when the calculated average curvature exceeds this calibrated curvature threshold... At that point, a state space transition was confirmed. This means that the user's emotional state underwent a significant, non-linear change in the high-dimensional space, corresponding to the user rapidly shifting from a normal or slightly uncomfortable state to a negative emotional state.

[0095] When the curvature exceeds the threshold, the starting point of the time window or the point when the curvature reaches its peak within the window is determined as the time point of the emotional mutation event.

[0096] Step 4: In response to identifying the mood mutation event, combine the second data stream within the retrospective time window prior to the event into an observation symptom vector.

[0097] Once a sudden emotional event and its timing are detected, a preset time window (30-120 seconds) prior to that event is immediately traced back. Within this traceback window, harmful technical event signals from the second data stream are collected and combined into an observation symptom vector.

[0098] Retrospective time window: The size of the window needs to be carefully considered. Too short a window will fail to capture the complete technical problem information that caused the emotional shift; too long a window will introduce irrelevant noise. The appropriate window size can be determined through experiments and analysis of historical data.

[0099] Combined Observational Symptom Vector: The second data stream within the retrospective time window contains various harmful technical event signals, each signal being a four-tuple. Information from these signals is extracted to construct a fixed-dimensional observational symptom vector. Each dimension of the vector corresponds to a symptom type, and its value represents the intensity or frequency of that symptom within the retrospective time window. Symptom types include: number of smoothness stutters, average stutter duration, number of event queue overflows, peak rate of interaction entropy growth, number of invalid operations, number of network request timeouts, number of page load failures, and number of application crashes.

[0100] If the signal comes from a smoothness sensor and the pulse intensity vector is (50, 200), then add statistical values ​​for the "average stutter duration" and "maximum event queue length" dimensions.

[0101] If the signal comes from the interactive entropy increase sensor and the pulse intensity vector is (0.8,5), then record the "peak value of the interactive entropy increase rate" and the "number of invalid operations".

[0102] The final observed symptom vector can include statistical information such as the total number of symptom occurrences, average intensity, maximum intensity, and duration.

[0103] Step 5: Match the observed symptom vectors with the preset causal attribution matrix. The row vectors of the matrix represent the technical root causes, and the column vector dimensions correspond to the dimensions of the symptom vectors.

[0104] Construction of the causal attribution matrix: This matrix was obtained through offline causal inference statistical analysis of massive historical fault log data. This included collecting a large number of user problem reports, client and server technical logs, monitoring data, etc. Using data mining and statistical methods (e.g., association rule mining, Granger causality testing, Bayesian network inference), the frequency and intensity of different combinations of technical root causes and symptoms were analyzed in various past technical fault scenarios, thereby establishing a model of the correlation strength between technical root causes and symptoms.

[0105] Matrix content:

[0106] Row vectors: Each row of the matrix corresponds to a specific, locatable technical root cause identifier. These identifiers point to specific types of failures, such as slow database queries, content delivery network node failures, timeouts of specific third-party interfaces, application client memory leaks, excessive server load, and network instability.

[0107] Column vectors: Each column of the matrix corresponds to a dimension of the observed symptom vector, including dimensions such as the number of smoothness stutters, average stutter duration, peak rate of interaction entropy growth, and number of network request timeouts.

[0108] Matrix values: The values ​​in the matrix are obtained through offline causal inference statistical analysis and represent the strength of the association between a specific technical root cause and a specific combination of symptoms. These values ​​can be probability values, correlation coefficients, or normalized statistical indicators.

[0109] Matching observed symptom vectors with causal attribution matrices:

[0110] A pre-trained learning distance metric model is used to calculate the semantic distance between the observed symptom vector and the row vector corresponding to each row (representing a technology root cause) in the causal attribution matrix.

[0111] The distance metric model is a feedforward neural network whose structure may include multiple fully connected layers and non-linear activation functions. It is trained to minimize the distance between the row vector of the correct root cause and its corresponding symptom vector in a known fault scenario. The training data includes a large number of labeled fault cases, each containing: the observed symptom vector, the actual technical root cause, and the corresponding row vector of the technical root cause in the causal attribution matrix. Through training, the neural network learns a distance metric that ensures, within the same fault scenario, the distance between the observed symptom vector and the row vector of the correct technical root cause is less than the distance between the row vectors of other incorrect technical root causes.

[0112] Matching process: The observed symptom vectors are input into the learning distance metric model, along with each row vector of the causal attribution matrix (or the model calculates the embedding representation of each row vector). The model outputs the distance (or similarity score) between the observed symptom vectors and the row vectors of each technology root cause. The smaller the distance (higher the similarity), the better the technology root cause matches the observed symptoms.

[0113] Causal Attribution Matrix Matching Engine: The causal attribution matrix matching engine in the system is responsible for performing this matching process. It has a built-in causal attribution matrix and a neural network for learning distance metrics. When an observed symptom vector is received, the engine calls the neural network to calculate the distance and find the nearest matrix row.

[0114] Step 6: Based on the matching results, determine the technical root cause corresponding to the matrix row that is closest to the observed symptom vector in terms of learning distance. This will serve as the assessment result for the telecommunications service quality problem, and generate a structured diagnostic profile that includes the technical root cause, evidence path, and confidence score.

[0115] Identify the technical root cause: Select the technical root cause identifier corresponding to the row of the causal attribution matrix that has the smallest learning distance (or the highest similarity score) to the observed symptom vector.

[0116] Generate a structured diagnostic profile: Based on the identified technical root cause and relevant information in the matching process, generate a structured diagnostic profile. This profile aims to clearly present the problem, cause, and relevant evidence to facilitate subsequent troubleshooting and repair.

[0117] The generated structured diagnostic profile includes: diagnostic session ID, emotional mutation timestamp, identified technical root cause identifier, evidence vector (i.e., the observed symptom vector that led to the diagnosis), attribution confidence score (i.e., the similarity score between the observed symptom vector and the matching matrix row vectors), and suggested remedial action code.

[0118] A telecommunications service quality assessment system based on user sentiment analysis includes:

[0119] The client-side digital sensor deployment module is responsible for instantiating and managing multiple functionally specific digital peripheral sensors (such as fluency harm sensors and interaction entropy increase sensors) in the client application of telecommunications services. This module is also responsible for listening to harmful technical events captured by the sensors and encoding the events according to the quadruple in step 1 to generate a first data stream and a second data stream carrying nanosecond-level timestamps.

[0120] The hierarchical cross-modal attention analysis module receives the first data stream encoded and transmitted by the digital sensor deployment module. This module incorporates the hierarchical cross-modal attention model described in step 2. It performs temporal modeling and cross-modal fusion on the sensor signals in the first data stream, generating an emotional trajectory representing the user's emotional state changes in a high-dimensional vector space. Next, the module analyzes the geometric shape of the emotional trajectory (calculating curvature) and, following the method in step 3, identifies the state-space transitions occurring within a preset time window, determining the emotional abrupt change events and their timing.

[0121] The causal attribution matrix matching engine responds to the mood abrupt change events determined by the hierarchical cross-modal attention analysis module. This engine receives the time point information of the mood abrupt change and, based on the backtracking time window in step 4, extracts the corresponding technical event signals from the second data stream transmitted from the client, combining them into an observed symptom vector. The engine incorporates a causal attribution matrix and a learning distance metric neural network. During matching, the engine calculates the non-Euclidean distance between the observed symptom vector and each row vector (representing the technical root cause) in the causal attribution matrix using the learning distance metric neural network. It determines the technical root cause corresponding to the matrix row closest to the observed symptom vector in terms of learning distance, serving as the evaluation result for telecommunications service quality issues.

[0122] Diagnostic profile generation module: Based on the technical root causes determined by the causal attribution matrix matching engine and related evidence vectors, confidence scores, and other information, a structured business quality diagnostic profile is generated according to the structure in step 6.

[0123] The causal attribution matrix matching engine further includes a learning distance metric neural network; when performing a match, the engine uses the neural network to calculate the non-Euclidean distance between the observed symptom vector and each row vector in the causal attribution matrix to determine the technical root that best matches the symptom.

[0124] Example 1: Implemented exactly according to the steps. This includes deploying smoothness, interaction entropy increase, and other receptors; using a hierarchical cross-modal attention model to analyze emotional trajectories and detect abrupt changes; constructing symptom vectors based on a second data stream with a backtracking time window; matching the learned distance metric model with the causal attribution matrix; and generating a complete structured diagnostic profile.

[0125] Parameter settings: The fluency sensor check frequency is dynamically adjusted according to device performance (100Hz for high-end devices, 60Hz for low-end devices); the interaction entropy increase sensor information entropy calculation sliding window is 30 seconds; the interaction entropy growth rate adaptive threshold is dynamically adjusted according to the user's historical behavior, with an initial value of 0.5 bits / second; the emotion trajectory analysis time window is 5 seconds; the emotion trajectory curvature threshold is calibrated based on historical normal user data; the backtracking time window is 60 seconds.

[0126] The observed symptom vector dimension covers all deployed receptor types and their key impulse intensity dimensions, including: number of stutters, average stutter duration, maximum event queue length, peak value of interaction entropy increase, number of invalid operations, number of network timeouts, number of page load failures, number of application crashes, and duration of abnormal power consumption.

[0127] The causal attribution matrix is ​​trained on massive amounts of historical fault data and contains 500 known technical root causes.

[0128] The distance metric model is a three-layer feedforward neural network with 128 neurons in each layer, using the ReLU activation function.

[0129] Example 2: Compared to Example 1, the sentiment analysis part is simplified. Only the raw impulse intensity of the fluency and interaction entropy increase receptors is used as input. Instead of going through a complex cross-modal attention model, these impulse intensity sequences are directly mapped to the sentiment trajectory. The other steps are the same as in Example 1.

[0130] Parameter settings: The sentiment analysis model is simplified by directly using LSTM or GRU to process each modality sequence, and then simply concatenating or summing the hidden states of each modality to obtain the sentiment trajectory.

[0131] Comparative Example 1: No user sentiment analysis is involved. It relies solely on preset technical indicator thresholds. For example, if the server-side interface response time exceeds 2 seconds, or the client crash rate exceeds 1%, an alarm is triggered. The diagnostic process mainly relies on manually reviewing logs or a preset rule engine.

[0132] Parameter settings: Server interface response time threshold 2000 milliseconds; Client crash rate threshold 1%; Network latency threshold 500 milliseconds; Page load time threshold 5 seconds; No automatic attribution mechanism.

[0133] Comparative Example 2: Analyze user behavior sequences within specific business processes, using statistical methods (such as Hidden Markov Models and sequence anomaly detection algorithms) to detect significant deviations from normal behavior patterns. If abnormal behavior is detected, it is marked as a potential problem. Emotions are not directly quantified or causal attribution matrices are used.

[0134] Parameter settings: The behavior sequence modeling algorithm adopts a hidden Markov model, and the anomaly detection threshold is calibrated based on the likelihood of historical normal user behavior. There is no automatic attribution mechanism.

[0135] Comparative Example 3: A simple statistical analysis of the frequency of different technical events (lag, timeout, etc.) within a certain time window is performed. If the frequency of a certain event exceeds a preset threshold, an alarm is triggered. This approach does not consider the correlation between events, analyze user sentiment, or perform causal attribution.

[0136] Parameter settings: Lag event frequency threshold exceeds 5 times per minute, network timeout event frequency threshold exceeds 3 times per minute, no automatic attribution mechanism.

[0137] Several typical telecommunications service scenarios were selected for simulation:

[0138] Scenario A: A user is making an online top-up. During the process, a brief network fluctuation causes the page to load slowly, and the client interface experiences slight lag.

[0139] Scenario B: A user attempts to change their plan, but the process is complex, causing the user to repeatedly go back and click again, resulting in confusion.

[0140] Scenario C: When a user is watching a video, the video playback is stuttering due to a CDN node failure. At the same time, the user frequently switches between different resolutions and performs fast forward / rewind operations.

[0141] Scenario D: When a user is using an application in the background, a background task bug causes abnormal battery drain.

[0142] Simulate the above scenario and record the actual technical problems and user behaviors, as detailed in Table 1:

[0143]

[0144] Example 1, by combining user emotions, multiple technical events, and causal attribution, exhibits the highest recall, precision, and diagnostic accuracy across various scenarios, with the lowest false positive rate and short diagnosis time. This is the optimal implementation of the present invention.

[0145] Example 2, while simplifying the emotion analysis model, can still effectively detect and diagnose problems, but its sensitivity to sudden emotional changes and diagnostic accuracy in complex scenarios are slightly lower than that of Example 1.

[0146] Compared to traditional alarms, Comparative Example 1 can only detect issues where technical indicators have reached hard thresholds. It is powerless to address situations where user experience has been compromised but technical indicators have not yet deteriorated significantly. Furthermore, diagnosis relies on manual intervention, which is inefficient.

[0147] Comparative Example 2 (abnormal behavior) can detect problems encountered by users in the process, but it is difficult to distinguish whether the problem is a process design issue or an underlying technical failure, and it is also impossible to directly locate the technical root cause.

[0148] Comparative Example 3 (simple statistics) is prone to generating a large number of false alarms because it only focuses on the frequency of events and does not consider the correlation between events or their impact on user experience.

[0149] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any ordinary changes and substitutions made by those skilled in the art within the scope of the technical solutions of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for assessing the quality of telecommunications services based on user sentiment analysis, characterized in that: Includes the following steps: S1. Deploy multiple digital peripheral sensors in the client application. The sensors listen for and encode harmful technical events related to negative user experiences in real time to generate a first data stream and a second data stream carrying timestamps with nanosecond precision. S2. Using a hierarchical cross-modal attention model, process the first data stream to generate an emotional trajectory that represents the changes in the user's emotional state in a high-dimensional vector space. S3. By analyzing the geometric shape of the emotional trajectory, identify the state space transitions that occur within a preset time window to determine the emotional mutation events and their timing. S4. In response to identifying a mood mutation event, combine the second data stream within the retrospective time window prior to the event into an observed symptom vector; S5. Match the observed symptom vectors with the preset causal attribution matrix; S6. Based on the matching results, determine the technical root cause corresponding to the matrix row that is closest to the observed symptom vector in terms of learning distance, and use it as the assessment result of the telecommunications service quality problem, and generate a structured diagnostic profile.

2. The telecommunications service quality assessment method based on user sentiment analysis according to claim 1, characterized in that, Digital peripheral receptors include: The smoothness damage sensor checks the state of the UI rendering thread at a frequency higher than the visual refresh rate of the user interface through the microtask scheduler. The sensor is activated when the rendering thread is continuously blocked or the event queue length exceeds a preset threshold. The interactive entropy increase sensor monitors the sequence of user actions in a specific business process and calculates the rate of increase of information entropy in the sequence in real time. When the rate of increase of entropy exceeds a preset threshold, the sensor is activated to quantify the user's level of confusion and frustration.

3. The telecommunications service quality assessment method based on user sentiment analysis according to claim 2, characterized in that, The coding steps for harmful technical incidents include: When any digital peripheral receptor is activated, a structured noxious stimulus signal is generated; The data structure of the stimulus signal is a quadruple, which includes a nanosecond-level timestamp, a unique receptor ID, a pulse intensity vector, and a local context summary ID. The pulse intensity vector is a vector of at least two values ​​used to quantify multiple dimensions of severity in this harmful technical incident.

4. The telecommunications service quality assessment method based on user sentiment analysis according to claim 3, characterized in that, The structure of the hierarchical cross-modal attention model includes: The intramodal temporal layer contains at least two parallel gated recurrent unit networks, which perform temporal modeling on the speech acoustic feature sequence and the user operation behavior sequence, respectively, to extract the temporal dynamic features of each modality. The cross-modal attention layer receives the output of the intramodal temporal layer and learns and fuses cross-modal cooperative change patterns by calculating the attention weights of the hidden state sequence of one modality to the hidden state sequence of another modality.

5. The telecommunications service quality assessment method based on user sentiment analysis according to claim 4, characterized in that, The steps for analyzing the geometric shape of emotional trajectories include: The emotional trajectory is viewed as a parametric curve in a high-dimensional space; Calculate the average curvature of the curve within a preset time window; A state space transition is determined when the average curvature exceeds a curvature threshold calibrated based on historical data.

6. The telecommunications service quality assessment method based on user sentiment analysis according to claim 5, characterized in that, In the causal attribution matrix, each row corresponds to a specific, locatable technical root cause identifier. The technical root cause identifiers point to fault types including slow database queries, content delivery network node failures, and timeouts of specific third-party interfaces. Each column of the matrix corresponds to one dimension of the observed symptom vector; The values ​​in the matrix represent the strength of the association between a particular technological root cause and a particular combination of symptoms.

7. The telecommunications service quality assessment method based on user sentiment analysis according to claim 6, characterized in that, The observed symptom vectors are matched with the causal attribution matrix using a learned distance metric model to calculate the semantic distance between the observed symptom vectors and each row vector in the causal attribution matrix.

8. The telecommunications service quality assessment method based on user sentiment analysis according to claim 1, characterized in that, The generated structured diagnostic profile includes a diagnostic session ID, a timestamp of emotional shifts, a identified technical root cause identifier, and an evidence vector.

9. The telecommunications service quality assessment system based on user sentiment analysis according to any one of claims 1-8, characterized in that, include: A client-side digital sensor deployment module is used to instantiate and manage multiple functionally specific digital peripheral sensors and to encode harmful technical events. The hierarchical cross-modal attention analysis module is used to receive and process the first data stream encoded by the receptors to generate emotional state trajectories and detect their state space transitions. The causal attribution matrix matching engine is used to respond to state space transitions, combine the second data stream into an observed symptom vector, and match it with the built-in causal attribution matrix to pinpoint the root cause of the technology. The diagnostic profile generation module is used to generate a structured business quality diagnostic profile based on the technical root causes.

10. The telecommunications service quality assessment system based on user sentiment analysis according to claim 9, characterized in that, The causal attribution matrix matching engine includes a learning distance metric neural network. When performing a match, the engine uses the neural network to calculate the non-Euclidean distance between the observed symptom vector and each row vector in the causal attribution matrix to determine the technical root that best matches the symptom.

Images