Customer lead hierarchical follow-up method based on customer behavior data fusion and dynamic scoring

Abstract

The application discloses a customer clue hierarchical follow-up method based on customer behavior data fusion and dynamic scoring, and relates to the technical field of product hierarchical follow-up optimization. The specific implementation scheme is as follows: customer behavior data of all users of a product on a function is acquired to form a user behavior event table, a matrix is constructed based on the user behavior event table to generate a use clue score reflecting user experience and a development clue score reflecting technical defects, a comprehensive clue score is generated according to the use clue score and the development clue score, and the peak position of the kernel density function based on the comprehensive clue score is taken as an optimized target, and the score of each peak position is subjected to hierarchical optimization. Through the application, the dynamic scoring of multi-source behavior data on a function parameter and the accurate positioning and hierarchical optimization of function parameter optimization are realized.

Description

Technical Field

[0001] This application relates to the field of product tiered follow-up optimization technology, specifically to a customer lead tiered follow-up method based on customer behavior data fusion and dynamic scoring. Background Technology

[0002] This application relates to the field of electronic product user behavior data analysis and product optimization decision-making technology, and is particularly applicable to functional optimization scenarios for electronic products such as smartphones, tablets, and smart wearable devices. As electronic product functions become increasingly complex, user experience differences and the distribution of technical defects often exhibit significant group differentiation. Accurately identifying specific functions and formulating differentiated optimization strategies has become a key issue in product optimization decision-making.

[0003] In existing technologies, common practices related to product optimization include: feature activity analysis and retention analysis based on user behavior statistics, defect rate statistics based on device performance logs, and attempts to migrate sales lead scoring and tiered follow-up methods to the field of product optimization. For example, by collecting user operation data to calculate feature usage frequency and retention contribution, or by statistically analyzing technical indicators such as crash rate and lag rate to assess feature health, features are then ranked and tiered for follow-up based on their scores. These methods typically use "feature" as the smallest unit of analysis, processing subjective user experience and objective technical defects separately, and using fixed weights or simple weighted methods to comprehensively score them.

[0004] However, the aforementioned existing technologies have significant shortcomings: First, the granularity of the optimization targets is too coarse, making it impossible to assess the urgency of optimization for each parameter within a continuous parameter function, resulting in difficulty in accurately allocating optimization resources; second, there is a lack of a unified quantitative trade-off mechanism between subjective user experience (such as satisfaction and activity) and objective technical defects (such as crashes and lag), making it impossible to dynamically determine the focus based on actual optimization benefits; third, existing lead scoring methods are mainly geared towards sales conversion and fail to automatically identify areas of user demand differentiation or optimization urgency concentration using scoring distribution patterns, still requiring manual judgment on whether concentrated optimization or categorized optimization is needed. These shortcomings make it difficult for existing technologies to achieve precise optimization positioning and automatic hierarchical decision-making at the functional parameter level of electronic products, and the technical solution proposed in this application is precisely to solve these problems.

[0005] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] In view of this, the purpose of this application is to provide a customer lead stratification and follow-up method based on customer behavior data fusion and dynamic scoring, which can realize the accurate positioning and stratification of functional parameter dynamic scoring and functional parameter optimization by integrating multi-source behavioral data.

[0007] According to a first aspect of this application, a method for tiered follow-up of customer leads based on customer behavior data fusion and dynamic scoring is provided, including: Acquire all user operation data, device performance logs, and user feedback data of the product as customer behavior data, and preprocess the customer behavior data; classify and encode the preprocessed customer behavior data according to user identifier, function identifier, function parameter identifier, operation identifier, and device performance identifier to form a structured user behavior event table; Set the sliding window and step size, and obtain the user behavior event table within the sliding window; based on the event table, count the number of times each user uses each functional parameter, the usage duration, and the core operation completion rate, and construct a user-functional parameter usage feature matrix; calculate the activity factor and retention contribution factor of the functional parameters based on the usage feature matrix; calculate the association strength value based on the co-occurrence relationship of functional parameters within the session window, and construct a function-functional parameter co-occurrence matrix accordingly, and then calculate the synergy factor of the functional parameters; extract the satisfaction factor based on user feedback data; and fuse the activity factor, retention contribution factor, satisfaction factor, and synergy factor according to preset weights to generate a usage clue score. Based on device performance logs, the defect occurrence rate of each functional parameter triggering device performance type anomalies is statistically analyzed. The severity of defects is analyzed by constructing a function-development clue association table. A development clue score for each functional parameter is generated based on the defect occurrence rate and defect severity. The usage lead score is converted into an experience optimization score, and a comprehensive lead score is generated by weighting it with the development lead score. An exponential time decay weight is introduced to update the comprehensive lead score in each sliding window based on the update step value. Based on the current time window, kernel density estimation is performed on the comprehensive clue score to obtain the probability density function. Then, the local maximum value of the density function is detected as the peak position, and the position and score of each peak position are obtained. The position of the peak position is used as the optimization target, and the score corresponding to each peak position is used as the urgency level. The peak positions are then stratified into different urgency levels, and optimization strategies are formulated.

[0008] Furthermore, the operational data includes the number of times each function parameter is launched, the duration of use, and the operation completion rate; Device performance logs include CPU operating frequency, remaining memory capacity, remaining battery power, and product temperature; User feedback data includes app store reviews, customer service tickets, and satisfaction surveys; The preprocessing method is as follows: The acquired raw customer behavior data is cleaned by the Flink stream processing framework, which filters out invalid events with missing key codes, removes duplicate reports within a short period of time, and eliminates outliers. The structured user behavior event table includes user code, function code, function parameters, device performance log, event type, timestamp, and event value.

[0009] Furthermore, the steps for calculating the activity factor and retention contribution factor of the functional parameters are as follows: Constructing user-feature parameters using a feature matrix: Based on the user behavior event table in the sliding window, filter out the corresponding user identifier, function identifier and corresponding function parameter based on the code, and the event type is "start". Count the total number of the records to get the number of times the user uses the function parameter in the sliding window. Filter the usage duration of the event types for this function parameter, and sum the usage duration values ​​to get the total usage duration of this function parameter; The core operation completion rate is obtained by counting the number of times the core operation events of this function parameter occur and dividing the number of times the function parameter is started. Each user and function parameter combination is used to form a triplet, which consists of the number of times it is used, the duration of use, and the completion rate of the core operation. The user-function parameter usage feature matrix is ​​formed with the user as the row, the function parameter as the column, and the triplet as the element. The calculation method for the activity factor: The activity factor is calculated by dividing the average daily number of users who have used each function parameter by the sum of the average daily number of all users who have used the function. The calculation method for the retention contribution factor: For each function parameter, the number of users who used the function parameter on each day and continued to use it on the next day is added up daily to get the total number of retained users. Then, the total number of users who used the function parameter on each day is added up daily to get the total number of users. The total number of retained users is divided by the total number of users to get the user retention rate of the function parameter on the next day. The total number of users who used the function on each day within the sliding window and continued to use it on the next day is accumulated daily to obtain the total number of retained users. The total number of users who used the function on each day within the sliding window is accumulated daily to obtain the total number of active users. The total number of retained users is divided by the total number of active users to obtain the overall user retention rate on the next day.

[0010] The difference between the user day-one retention rate and the overall user day-one retention rate is then divided by the overall user day-one retention rate. The resulting value is the retention contribution factor of this function parameter within the sliding window.

[0011] Furthermore, the steps for calculating the synergistic factors of the functional parameters are as follows: Analyze the correlation strength value for each functional parameter; The session window can be divided within a sliding window by setting the session duration. Count the total number of session windows, the number of session windows containing each function parameter, and the number of session windows containing two function parameters simultaneously.

[0012] Based on the ratio of the number of sessions containing a single functional parameter to the total number of sessions in the session window, determine the individual probability value for each functional parameter; based on the ratio of the number of sessions containing two functional parameters to the total number of sessions, determine the joint probability value; after smoothing the session count, take the natural logarithm of the ratio of the joint probability value to the product of the two individual probability values ​​to obtain the association strength value. Construct a co-occurrence matrix of function-function parameters based on correlation strength values; Obtain the association strength values ​​of all function parameter pairs, retain only the function parameter pairs with an association strength greater than 0, and fill the matrix with the association strength values ​​greater than 0, with the function parameters as the rows and columns, to construct the function-function parameter co-occurrence matrix; Obtain the co-factor of the functional parameters; Based on the function-function parameter co-occurrence matrix, for each function parameter, the sum of its correlation strength values ​​with other function parameters is divided by the sum of the maximum value of the corresponding row sums of all function parameters and ε, and then normalized to obtain the co-occurrence factor of the function parameter.

[0013] Furthermore, the steps for extracting user satisfaction factors are as follows; We filter records related to each function parameter and scenario from user feedback data, determine whether each piece of feedback is positive or negative, count the number of positive feedback and the total number of feedback, and use the Laplace smoothing formula to extract the satisfaction factor.

[0014] Furthermore, the analytical method using clue scoring is as follows: Based on activity factors, retention contribution factors, satisfaction factors, and functional synergy factors, a usage clue score reflecting user experience is generated using a weighted approach.

[0015] Furthermore, the statistical method for defect incidence rate is as follows: Based on the current sliding window, extract the function parameter operation events from the user behavior event table that occur in the same user and the same session as the device performance abnormality and are located within the preset attribution time window. The event type is device performance type abnormality. For each functional parameter and each type of device performance anomaly, the total number of events that trigger that anomaly type for that functional parameter is counted as the numerator, and the total number of usage events for that functional parameter is counted as the denominator. The ratio of the two is the defect occurrence rate of that functional parameter in device performance. Build a functional-device anomaly association table; The defect rate of the function parameter across all device performance types is weighted and averaged according to a preset weight to obtain the total defect rate of the function parameter. The total defect rate is then limited to the range of 0 to 1. The function parameter is used as the row and the type of device performance abnormality is used as the column. Each defect rate is filled into the corresponding position to form a function-development clue association table. The method for judging abnormal equipment performance is as follows: For each metric in the device performance log, a normal range or safety threshold is preset. These metrics are monitored in real time within a sliding window, and when any metric exceeds the normal range, it is determined as a device performance anomaly event. The steps for analyzing defect severity based on the feature-development clue association table are as follows: Assign a severity weight to each type of device performance anomaly, and the sum of all severity weights is 1. Take the number of events for the function parameter in each anomaly type from the association table, multiply it by the corresponding severity weight, and sum them to obtain the weighted severity sum. Then sum the number of events for the function parameter in all anomaly types and add 1 as the denominator. Divide the weighted severity sum by the denominator to obtain the defect severity, and normalize the defect severity to the range of 0 to 1.

[0016] Furthermore, the method for generating clue scores is as follows; Subtract the total defect rate from 1 to get the probability that no defect occurred. Subtract the defect severity from 1 to get the probability that the severity was not reached. Multiply the probability that no defect occurred by the probability that the severity was not reached to get the probability that neither a defect occurred nor the severity was reached. Subtract this product from 1 to get the probability that at least one defect occurred and the severity was reached, and then multiply by 100.

[0017] Furthermore, the steps for generating a comprehensive clue score are as follows: After performing Min-Max normalization on the usage clue score, take 1 and subtract the normalized usage clue score to obtain the experience optimization score. Then, merge it with the development clue score according to the preset weight to generate the comprehensive clue score. The steps for updating the overall clue score are as follows: The time window is slid along the time axis according to the update step size, and the comprehensive clue score is recalculated for the slid time window; an exponential time decay weight is introduced into the comprehensive clue score sequence within the historical time window; the comprehensive clue scores of the current time window and the historical time window are weighted and averaged to obtain the final comprehensive clue score used for analysis. An exponential time decay weight is introduced into the comprehensive clue scoring sequence within the historical time window; The weighted average of the comprehensive clue scores from the current time window and historical time windows is used to obtain the final comprehensive clue score for analysis.

[0018] Furthermore, the method for obtaining the position and score of each peak is as follows: For each score in the comprehensive clue score, a Gaussian kernel function is placed centered on that score value. The vertical coordinates of the Gaussian kernel function corresponding to all data points at each position on the horizontal axis are added together to obtain a superimposed total curve. The superimposed total curve is normalized to obtain the probability density function. On the probability density curve, all local maxima are found. These local maxima are the peaks in the data distribution, and the comprehensive clue score corresponding to the peak is the score of each peak. The optimization strategy is as follows: A threshold for urgency is preset, and the comprehensive clue score is divided into three intervals. If the peak score is not lower than the first threshold, it is judged as high urgency; if the peak score is between the first threshold and the second threshold, it is judged as medium urgency; and if the peak score is lower than the second threshold, it is judged as low urgency.

[0019] One embodiment of the above application has the following advantages or beneficial effects: This application achieves automatic discovery and hierarchical follow-up of optimization clues at the functional parameter level of electronic products. By constructing a structured user behavior event table, it provides a foundation for the fusion of multi-source heterogeneous data under a unified framework for all subsequent analyses, solving the problem of difficulty in associating multi-source data in existing technologies. Based on this, a user-functional parameter usage feature matrix is ​​constructed, extracting activity factors, retention contribution factors, satisfaction factors, and functional synergy factors, and generating a weighted usage clue score. A functional-development clue association table is constructed based on device performance logs to analyze defect incidence and defect severity, generating a development clue score. Simultaneously, usage clue scores reflecting subjective user experience and scores reflecting objective technical defects are obtained. The development clue scoring system addresses the shortcomings of existing technologies that evaluate functionality from a single dimension. It transforms usage clue scores into experience optimization scores, which are then weighted and synthesized with development clue scores to form a comprehensive clue score. This comprehensive score is updated by introducing exponential time decay weights within a sliding window based on the update step size. This allows the comprehensive score to uniformly measure the optimization value of experience improvements and technical fixes while dynamically reflecting recent behavioral changes, improving the timeliness and accuracy of optimization clue scoring. By estimating the kernel density of the comprehensive clue score, it automatically detects local maxima of the density function as peaks, obtaining the number of peaks and their scores. Based on the peak positions and scores, it performs stratification and optimization strategy formulation. This process can automatically identify specific levels and their urgency within a functional parameter optimization demand set, thereby accurately guiding concentrated optimization of specific functional parameter levels or categorized optimization of differentiated groups. It overcomes the problems of existing technologies that use overall functionality as the granularity and require manual judgment of optimization direction, significantly improving the accuracy and decision-making efficiency of electronic product function optimization.

[0020] Other effects of the above-mentioned alternative methods will be described below in conjunction with specific embodiments. Attached Figure Description

[0021] The accompanying drawings are provided for a better understanding of this solution and do not constitute a limitation of this application. Wherein: Figure 1 This is a flowchart illustrating a customer lead stratification and follow-up method based on customer behavior data fusion and dynamic scoring, provided in an embodiment of this application. Figure 2 This is an example of an exponential time decay weighted curve provided in this application embodiment; Figure 3 This is a dynamic update graph of comprehensive clue scoring provided in the application embodiment; Figure 4 This is a schematic diagram of a comprehensive clue score distribution provided in an embodiment of this application; Figure 5 This application provides a schematic diagram illustrating the process of superimposing Gaussian kernel density estimates on comprehensive clue scores in an embodiment of the present application. Figure 6 This is a schematic diagram illustrating the detection of peak position and scoring of a probability density function, as provided in an embodiment of this application. Detailed Implementation

[0022] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0023] To facilitate understanding of this application, the embodiments of this application will be briefly described below: In the optimization of smartphone camera functions, different users often exhibit significant differences in their user experience and the distribution of technical defects at different focal lengths (such as ultra-wide-angle, wide-angle, and telephoto). Some users experience slow focusing and frequent image stabilization failures when shooting at telephoto focal lengths, while having a good experience at wide-angle focal lengths. This parameter-level differentiation makes it difficult for traditional analysis methods that focus on the entire camera function to pinpoint specific optimization targets. Existing product optimization methods typically involve statistically analyzing overall function usage activity, retention contribution, and crash rate, or analyzing user feedback and performance logs separately. However, they lack the means to integrate and score subjective user experience with objective technical defects at the function parameter level and to make tiered decisions, resulting in optimization resources not being accurately allocated to parameter ranges with prominent problems.

[0024] This application provides a method for hierarchical follow-up of optimization leads for electronic product functional parameters based on customer behavior data fusion and dynamic scoring. By collecting user operation data, device performance logs, user feedback, and scenario attribute data, a user-functional parameter usage feature matrix is ​​constructed to extract activity factors, retention contribution factors, satisfaction factors, and collaboration factors to generate usage lead scores. Simultaneously, based on a functional parameter-development lead correlation table, defect incidence and severity are analyzed to generate development lead scores. The usage lead scores are converted into experience optimization scores and weighted with the development lead scores to synthesize a comprehensive lead score. Kernel density estimation and peak detection are then performed on the comprehensive lead score to pinpoint the functional parameter levels requiring priority optimization, and optimization strategies are formulated hierarchically based on the peak scores. This method solves the problem of existing technologies being unable to accurately identify the urgency of functional parameter optimization levels for electronic products, achieving automated hierarchical follow-up of optimization leads.

[0025] See Figure 1 This is a flowchart illustrating a customer lead stratification and follow-up method based on customer behavior data fusion and dynamic scoring, provided in an embodiment of this application. Figure 1The execution subject of the method shown can be a combination of software and / or hardware, specifically, it can be one or more of various types of terminals, hardware systems, cloud computing, etc.

[0026] Figure 1 The customer lead stratification and follow-up method based on customer behavior data fusion and dynamic scoring, as shown, includes steps S101 to S105, as follows: S101: Obtain all user operation data, device performance logs, and user feedback data of the product as customer behavior data, and preprocess the customer behavior data; classify and encode the preprocessed customer behavior data according to user identifier, function identifier, function parameter identifier, operation identifier, and device performance identifier to form a structured user behavior event table.

[0027] This step simultaneously collects user operation data, device performance logs, and user feedback data. After uniformly classifying and encoding these multi-source heterogeneous data, a structured user behavior event table is formed. Its advantage lies in achieving complete integration of customer behavior data within a unified framework. This allows user subjective experience, objective device status, and scenario environment information to be correlated to the same functional parameter dimension, overcoming the shortcomings of existing technologies where fragmented data sources and inconsistent identifiers prevent cross-source correlation analysis. Compared to existing technologies, this step refines user operation data to the number of launches, usage duration, and operation completion rate, and incorporates multi-dimensional indicators such as CPU operating frequency, remaining memory capacity, remaining battery power, and product temperature into device performance logs. This enables subsequent evaluation not only of user behavior preferences but also direct correlation with device technical status, providing a primary basis for dual-track scoring of experience optimization and technical repair. Furthermore, the unified encoding and structured event table design makes functional parameter-level statistics possible, avoiding the coarse analysis caused by traditional methods that only consider overall functionality. This provides a basis for subsequent construction of user-functional parameter usage feature matrices, functional parameter-development clue correlation tables, and functional parameter co-occurrence matrices. The data foundation obtained through direct retrieval and aggregation ensures the accuracy of subsequent calculations of activity, retention contribution, synergy factor, satisfaction, and defect incidence and severity. Finally, the preprocessing process uses the Flink stream processing framework to clean the data, filtering out invalid events with missing key codes, removing duplicate reports in a short period, and eliminating outliers. This significantly reduces data noise and improves the credibility and stability of the scoring results. As a result, it provides high-quality data support for the dynamic updating of comprehensive clue scores, kernel density estimation, peak position, and hierarchical decision-making in subsequent steps. This enables the overall solution to accurately locate optimization clues at the functional parameter level and formulate differentiated strategies, avoiding misjudgments and resource waste caused by poor data quality or confusing labeling.

[0028] In this embodiment, the operational data used includes the number of times each function parameter is started, the usage duration, and the operation completion rate. Device performance logs include CPU operating frequency, remaining memory capacity, remaining battery power, and product temperature; User feedback data includes app store reviews, customer service tickets, and satisfaction surveys; The preprocessing method is as follows: The raw customer behavior data is cleaned by the Flink stream processing framework, which filters out invalid events with missing key codes, removes duplicate reports within a short period of time, and eliminates outliers.

[0029] The reason for choosing these parameters is that they can comprehensively depict the user's experience with the functional parameters of electronic products and the objective technical status of the products from different dimensions, providing the necessary quantitative basis for the subsequent generation of usage clue scores and development clue scores.

[0030] The launch count, usage duration, and completion rate in the operation data reflect the frequency of user use, depth of engagement, and quality of task completion for a specific function parameter, respectively. Launch count is a direct measure of the activity level of a function parameter, while usage duration reflects the actual level of user engagement at that parameter level; both together support the calculation of the activity factor. The completion rate reflects whether users can successfully complete the core operation (such as successfully taking a photo) using that function parameter. A low completion rate usually indicates usability issues or design flaws in the function parameter, thus helping to identify levels with poor user experience.

[0031] The CPU operating frequency, remaining memory capacity, remaining battery power, and product temperature in the device performance log are core objective indicators of the electronic product's operating status, directly reflecting the resource consumption and stability during the operation of functional parameters. Abnormally high CPU frequency or insufficient memory often leads to lag, while a sudden drop in remaining battery power and excessive temperature correspond to abnormal power consumption and overheating issues, respectively. These indicators are used to determine whether a device performance anomaly has occurred, and then to statistically analyze the defect incidence and severity at each functional parameter level. This provides factual evidence for development clue scoring, making the evaluation of technical defects more objective and reliable, independent of subjective user feedback.

[0032] App store reviews, customer service tickets, and satisfaction surveys are important sources of user subjective opinions in user feedback data, and these three sources complement each other: reviews cover general evaluations from a broad user base, tickets record specific problems encountered by users, and surveys provide structured satisfaction scores. By performing sentiment analysis on this feedback, we can extract user satisfaction factors for specific functional parameters, compensating for the inability of usage behavior data to directly express users' subjective feelings.

[0033] In summary, these parameters collectively cover three levels: "what the user did - how the user felt - whether the device was operating normally." They support the quantification of both experience optimization value and technical optimization value, providing complete data support for subsequent comprehensive lead scoring, peak detection, and functional parameter-level optimization stratification.

[0034] Furthermore, the Flink stream processing framework is used to clean the raw customer behavior data because Flink is a distributed stream processing engine capable of low-latency, high-throughput real-time computation on continuously generated user behavior event streams, perfectly meeting the needs of electronic products with large volumes of user behavior data and the requirement for dynamic updates to scores. The benefits of this preprocessing method are: First, filtering invalid events with missing user codes, function codes, or function parameter codes avoids unassigned dirty data during subsequent statistics, ensuring the integrity of the feature matrix and scoring results; second, removing short-term duplicate reports (e.g., only one report from the same user for the same event within 5 seconds) eliminates redundant data caused by network retries or repeated sending by the collection SDK, preventing the accumulation and distortion of metrics such as startup count and usage duration; third, removing outliers (e.g., single usage duration exceeding 24 hours, device temperature exceeding the physical reasonable range) eliminates the impact of sensor malfunctions or malicious data on the statistical results, improving the credibility of the scores. When performing the above cleaning process, the Flink stream processing framework pulls raw data streams from the Kafka message queue in real time. It first parses each record's fields and extracts key codes; records missing user codes, function codes, or function parameter codes are discarded. Then, it performs short-term deduplication based on user codes, function parameter codes, event types, and timestamps, merging duplicate events within a 5-second window into a single record. Next, it validates event values ​​according to preset reasonable ranges or thresholds; data exceeding these ranges, such as usage time, battery consumption, and temperature, are considered abnormal and removed. The cleaned, valid data is mapped to a unified code and written to Redis to maintain sliding window statistics for the past 30 days, and also written to Hive or ClickHouse to store all historical data. This supports both real-time updates of online ratings and offline full reconstruction of the feature matrix. Thus, the preprocessing stage provides a clean, consistent, and real-time usable data foundation for the subsequent construction of the user-function parameter usage feature matrix and the function parameter-development clue association table, enabling the entire optimization clue hierarchical follow-up method to accurately locate the optimization direction at the function parameter level while ensuring reliable data quality.

[0035] In this embodiment, the identification types include user identification, function identification, function parameter identification, operation identification, and device performance identification; The structured user behavior event table includes user code, function code, function parameters, device performance log, event type, timestamp, and event value.

[0036] This classification of identifier types is necessary because the analysis granularity of this application is refined to the functional parameter level, and it needs to simultaneously integrate user subjective behavior, objective device status, and external usage environment. Therefore, different identifiers must be used to distinguish "who," "on which function," "using which parameter," "what operation occurred," and "what device anomaly occurred," thereby supporting subsequent multi-dimensional, high-precision statistics and scoring. Among them, the user identifier is used to uniquely identify the user subject of the behavior; the function identifier is used to distinguish different functions provided by the product; the functional parameter identifier is used to discretize the specific level of continuous parameters as the analysis object; the operation identifier is used to distinguish different event types such as startup, usage duration, completion of core operations, and comments; and the device performance identifier is used to distinguish different performance anomaly types such as CPU operating frequency, remaining memory capacity, remaining battery power, and product temperature. The structured user behavior event table fields correspond to the above-mentioned identifier types, and also add two key fields: timestamp and event value. The user code and function code are associated with the user identifier and function identifier respectively. The function parameter field directly records the parameter level. The device performance log field records the specific values ​​of CPU, memory, power, temperature, etc. at the time the event occurs. The event type field corresponds to the operation identifier. The timestamp is used for time window division and time decay calculation. The event value is used for quantitative recording (such as number of startups, usage time in seconds, number of successful photos, etc.).

[0037] For example, when user A uses their phone to open the camera and switch to telephoto mode, and the device temperature reaches 42℃, a record is made in the user behavior event table: user code U001, function code F001 (camera), function parameter P004 (telephoto mode), device performance log {CPU frequency: 2.8GHz, remaining memory: 3GB, remaining battery: 65%, product temperature: 42℃}, event type "startup", timestamp 2026-08-22 20:15:30, event value 1. If a temperature anomaly occurs subsequently, another record is made with event type "temperature anomaly" and event value 1. This allows for subsequent statistical analysis based on any combination of dimensions such as user, function parameter, and device anomaly type. For example, it enables statistics on the number of telephoto mode startups, average usage time, and temperature anomaly occurrence rate. This provides direct data slicing capabilities for calculating activity factors, retention contribution factors, satisfaction factors, defect incidence rate, and defect severity, thereby supporting comprehensive clue scoring and peak detection to accurately pinpoint the optimization needs of specific function parameter levels.

[0038] S102: Set the sliding window and step size, and obtain the user behavior event table within the sliding window; based on the event table, count the number of times each user uses each function parameter, the usage duration, and the core operation completion rate, and construct a user-function parameter usage feature matrix; calculate the activity factor and retention contribution factor of the function parameter based on the usage feature matrix. The association strength value is calculated based on the co-occurrence relationship of functional parameters within the session window, and a function-functional parameter co-occurrence matrix is ​​constructed accordingly. Then, the synergy factor of the functional parameters is calculated. The satisfaction factor is extracted based on user feedback data. The activity factor, retention contribution factor, satisfaction factor and synergy factor are fused according to preset weights to generate a usage clue score.

[0039] This step obtains a user behavior event table within a sliding window by setting a sliding window and step size. Based on this, a user-feature parameter usage feature matrix is ​​constructed. The usage frequency, total usage time, and core operation completion rate of each feature parameter are statistically analyzed. Then, activity factors, retention contribution factors, collaboration factors, and satisfaction factors are extracted, and finally, a weighted usage clue score is generated. The advantage of this step is that it refines the analysis granularity from the traditional overall function to the feature parameter level. It can accurately reflect the user's actual usage intensity, depth of investment, task completion quality, and user subjective satisfaction at different parameter levels. At the same time, it uses the correlation strength value to construct a feature parameter co-occurrence matrix to explore the collaborative usage relationship between different parameter levels. This ensures that the score not only considers the performance of a single parameter but also takes into account the ecological linkage value between parameters. Compared to existing technologies, which typically use the entire function as the smallest unit of analysis, this approach fails to distinguish the differences in user experience across different parameter levels within the same function. This can easily lead to a blurring of optimization directions by mixing the needs of different user groups together. In contrast, this step uses feature matrices at the function parameter level and multi-factor weighted scoring to accurately characterize the user experience level of each parameter level, providing the necessary data foundation for subsequent peak detection and identification of the urgency of optimization between parameter levels. In this technical solution, the contribution of this step to the overall solution is mainly reflected in the following aspects: the sliding window and step size mechanism ensure the timeliness and continuity of data statistics, enabling the score to be dynamically updated according to changes in user behavior, and providing a window sequence basis for the exponential time decay weighting in subsequent steps; the user-function parameter feature matrix provides a structured data source for the calculation of retention contribution factors and activity factors, ensuring the objectivity and accuracy of subsequent usage clue scoring; the synergistic factors obtained from the association strength co-occurrence matrix supplement the correlation information between function parameters, avoiding the one-sidedness of relying solely on the frequency of use of a single parameter while ignoring the linkage of usage scenarios; and the satisfaction factor extracts subjective evaluations from user feedback, complementing behavioral statistics, enabling usage clue scoring to more comprehensively reflect the user's real experience. Overall, S102 provides an alignment benchmark for the user experience side of developing clue scoring in S103, and also lays the foundation for the comprehensive clue scoring in S104 and the kernel density estimation and peak stratification in S105, which is a key data processing link for achieving precise optimization and positioning at the function parameter level.

[0040] More specifically, the sliding window and step size should be determined based on the changing cycle of electronic product function parameter usage behavior, statistical stability requirements, and the timeliness requirements of business decisions. The sliding window is the time range used to statistically analyze user behavior, and the step size is the interval at which the window slides along the time axis. The following factors should be considered when setting the window: the periodicity and volatility of user behavior. For example, electronic product function usage has obvious daily cycles and weekend effects; a window that is too short will be affected by daily fluctuations, while a window that is too long will introduce outdated data. The response speed requirements for function parameter optimization decisions; if it is necessary to quickly capture new versions or new issues, the window should not be too long, and the step size should not be too large. Whether the statistical sample size is sufficient; a window that is too short will lead to sparse usage samples for some function parameters, making the estimation of indicators such as retention rate and defect incidence rate unstable. In this embodiment, the sliding window is set to 30 days because 30 days can cover at least one complete usage cycle for electronic product users, and is sufficient to accumulate enough usage events and feedback samples, while ensuring the stability and comparability of statistical indicators. It also complements the exponential time decay mechanism in the subsequent comprehensive clue scoring, balancing statistical reliability and sensitivity to recent changes. The step size is set to 1 day, which allows the score to be updated daily, enabling timely detection of changes in the urgency of optimizing functional parameters, while keeping the computational cost under control.

[0041] This step extracts satisfaction factors from user feedback data in the user behavior event table and weights them together with activity factors, retention contribution factors, and functional synergy factors to generate a usage clue score. Its advantage lies in achieving a multi-dimensional fusion of subjective user evaluation and objective behavioral data. This allows the usage clue score to reflect not only the user's actual usage frequency, retention value, and the synergistic relationship between functional parameters, but also to directly reflect the user's subjective satisfaction with the functional parameters. Compared to existing technologies, which often rely solely on single-dimensional behavioral statistics (such as usage frequency or retention rate) or analyze user feedback alone, the former struggles to explain situations where users "used it but didn't like it," while the latter fails to quantify user behavioral input and functional ecosystem value, easily leading to a one-sided and distorted assessment of the functional parameter experience level. This step, however, uses a weighted combination of four factors to simultaneously characterize the experience from four perspectives: "how much did the user use?", "did they stay after using it?", "were they satisfied with the experience?", and "how well did it synergize with other parameters," significantly improving the accuracy and interpretability of the usage clue score. In this technical solution, the positive impact of this step on the overall solution is mainly reflected in the fact that the introduction of the satisfaction factor compensates for the inability of pure behavioral statistics to reflect users' subjective preferences, making the usage clue scoring closer to the real user experience. The combination of activity, retention contribution, collaboration factors and satisfaction factors makes the scoring more discriminative at the functional parameter level, and can identify differentiated scenarios such as "high activity but low satisfaction" or "good retention contribution but poor collaboration". This provides a reliable user experience benchmark for the comparison of subsequent development clue scores and the synthesis of comprehensive clue scores. After transformation, the usage clue score will serve as an important component of the comprehensive clue score, directly affecting the position and height of the peak in the kernel density estimation, thereby determining the stratification results of the optimized clues. Therefore, the factor design in this step directly ensures that the final optimization strategy can accurately target the functional parameter levels that users are truly dissatisfied with, avoiding the allocation of optimization resources to parameters that users do not care about, and improving the accuracy of the overall solution decision and the efficiency of resource utilization.

[0042] In this embodiment, the steps for calculating the activity factor and retention contribution factor of the functional parameters are as follows: Constructing user-feature parameters using a feature matrix: Based on the user behavior event table in the sliding window, filter out the corresponding user identifier, function identifier and corresponding function parameter based on the code, and the event type is "start". Count the total number of the records to get the number of times the user uses the function parameter in the sliding window. Filter the usage duration of the event types for this function parameter, and sum the usage duration values ​​to get the total usage duration of this function parameter; The core operation completion rate is obtained by counting the number of times the core operation events of this function parameter occur and dividing the number of times the function parameter is started. For each user and function parameter combination, the number of times it is used, the duration of use, and the completion rate of the core operation are combined into a triplet. With users as rows, function parameters as columns, and triplets as elements, a user-function parameter usage feature matrix is ​​formed.

[0043] The user-function parameter usage feature matrix is ​​constructed by combining the number of uses, usage duration, and core operation completion rate for each user and function parameter combination. This is because these three parameters characterize the user's actual usage behavior for a specific function parameter level from three different dimensions: usage frequency, usage depth, and usage quality. Using any one of these indicators alone cannot fully reflect the user's actual experience and value of that function parameter. The number of uses represents the number of times a user activates the function parameter within the sliding window, reflecting the intensity of the user's demand for and activity level of the function parameter. The higher the number of uses, the more frequently the user actively triggers the parameter level, which is the core basis for calculating the activity factor. The usage duration represents the cumulative time a user spends on the function parameter, reflecting the depth of the user's investment in the function parameter. It can compensate for the two behavioral patterns that the number of uses cannot distinguish: "many activations but only short stays each time" and "few activations but deep use each time", avoiding misjudgment of the activity factor by only looking at the number of uses. The core operation completion rate represents the proportion of users who complete the preset core operations (such as successfully taking a photo or completing the charging to the target level) on the function parameter. It reflects whether the function parameter truly helps the user achieve the usage goal. It is a key indicator for measuring the experience from a quality perspective. A low completion rate often indicates that the parameter level has problems such as operational obstacles, insufficient performance, or unreasonable design. By combining these three parameters into a triplet and constructing a matrix with users as rows and functional parameters as columns, the behavioral profile of each user at each functional parameter level can be stored in a structured manner. Subsequently, the average number of uses, average usage time, and average completion rate of that parameter level among all users can be directly aggregated by functional parameter column, thereby deriving the activity factor and retention contribution factor. At the same time, this matrix also preserves individual difference information at the user level, providing data slicing and aggregation convenience for segmented statistics by scenario and user group, as well as subsequent peak position analysis. It is a key data hub connecting the original behavioral events and functional parameter level scores, ensuring that the subsequent usage clue scores can accurately reflect the real experience level of different parameter levels, thereby supporting the precise optimization and positioning of functional parameter levels in this solution.

[0044] The calculation method for the activity factor: The activity factor is calculated by dividing the average daily number of users who have used each function parameter by the sum of the average daily number of all users who have used the function.

[0045] The activity factor is calculated based on the user count for each functional parameter column in the matrix. The specific steps are as follows: For a given functional parameter, count the number of rows where the triplet usage count is greater than 0. This gives the total number of unique users who used that functional parameter within the window. Divide this number by the number of days in the window to obtain the average daily user count. Then, divide this average daily user count by the sum of the average daily user counts for all functional parameters under that function. The resulting ratio is the activity factor for that functional parameter. This factor reflects the relative usage popularity of that functional parameter within the entire function; a higher value indicates a higher proportion of users choosing that parameter level.

[0046] The activity factor is calculated by counting the average daily number of users who have used each functional parameter and dividing that average daily number of users by the sum of the average daily number of users for all functional parameters under that function. The advantages are: using the average daily number of users instead of the cumulative total number of users eliminates the impact of repeated use within the sliding window and changes in window length on the statistical results, making the popularity of different functional parameters comparable on the same time scale; normalizing the average daily number of users for a single functional parameter to the percentage of the sum of the average daily number of users for all functional parameters within that function directly reflects the relative usage popularity of that functional parameter within the entire function, avoiding the incomparability between different functions due to differences in the overall user scale when directly using absolute user numbers. This percentage reflects the distribution of user preferences for parameter levels within the function, helping to identify which parameter levels are actively used by users and which have low usage rates, thus providing a basis for subsequent optimization priority decisions.

[0047] For example, suppose the camera function has five focal length parameters: ultra-wide-angle, wide-angle, standard, telephoto, and super telephoto. Within a 30-day sliding window, the average daily number of users for each focal length is: ultra-wide-angle 200, wide-angle 500, standard 800, telephoto 150, and super telephoto 60. The sum of the average daily number of users for all focal lengths is 200 + 500 + 800 + 150 + 60 = 1710. At this point, the activity factor for the telephoto focal length is approximately 150 ÷ ​​1710 ≈ 0.088, while the activity factor for the standard focal length is approximately 800 ÷ 1710 ≈ 0.468. This indicates that the standard focal length is the most frequently used focal length on cameras, while the telephoto focal length is relatively less frequently used. If the telephoto focal length has both a high development clue score and a low experience score, it means that although its usage frequency is low, there are significant issues that may require focused optimization. On the other hand, while the standard focal length has high activity, if its experience score is high and its defect score is low, it does not need to be prioritized. Calculating the relative proportion of activity factors in this way can more rationally guide the allocation of optimization resources, avoiding the problem of ignoring differences in the internal parameter structure of functions simply based on the absolute number of users.

[0048] The calculation method for the retention contribution factor: For each function parameter, the number of users who used the function parameter on each day and continued to use it on the next day is added up daily to get the total number of retained users. Then, the total number of users who used the function parameter on each day is added up daily to get the total number of users. The total number of retained users is divided by the total number of users to get the user retention rate of the function parameter on the next day. The total number of users who used the function on each day within the sliding window and continued to use it on the next day is accumulated daily to obtain the total number of retained users. The total number of users who used the function on each day within the sliding window is accumulated daily to obtain the total number of active users. The total number of retained users is divided by the total number of active users to obtain the overall user retention rate on the next day. The retention contribution factor of this function parameter within the sliding window is obtained by subtracting the overall user retention rate from the user retention rate of this function parameter on the next day, and then dividing the difference by the sum of the overall user retention rate on the next day and ε.

[0049] Here, ε is a very small positive number used to prevent the denominator from being zero or too small, which would lead to numerical instability.

[0050] When calculating the retention contribution factor by combining the feature matrix of user-feature parameters and the time information of the user behavior event table, the user set for a specific feature parameter is selected based on the usage frequency of that parameter column in the feature matrix being greater than 0. Then, the user behavior event table is used to count whether these users used the feature parameter each day within the window and whether they still had any activity (i.e., active) the following day. The number of users who "used the feature parameter on the day and remained active the next day" is accumulated daily as the numerator, and the total number of users who "used the feature parameter on the day" is accumulated daily as the denominator. Dividing the two yields the next-day retention rate for that feature parameter. Next, the user sets of all feature parameters under that feature are merged and deduplicated to obtain the overall user set for the feature. The next-day retention rate of the overall user set for the feature is also calculated daily. Finally, the next-day retention rate of users using that feature parameter is subtracted from the next-day retention rate of the overall user set for the feature, and the difference is divided by the next-day retention rate of the overall user set for the feature. The resulting value is the retention contribution factor for that feature parameter. This factor represents the percentage increase or decrease in user retention relative to the overall retention level of the feature. A positive value indicates a positive contribution to user retention, while a negative value indicates a drag.

[0051] The retention contribution factor is obtained by comparing the user day-one retention rate of a feature parameter with the overall user day-one retention rate of the feature and dividing the difference by the overall retention rate. The purpose of this is to measure the additional contribution or drag on user retention of the feature parameter relative to the overall level of the feature. The advantages of this method are as follows: First, using a relative increase ratio rather than an absolute difference can eliminate the influence of the overall retention level of the feature itself, making the retention contributions between different feature parameters and between different features comparable. For example, a 5% increase in retention rate for both features has drastically different meanings in scenarios with an overall retention rate of 40% and 80%. The relative ratio can more fairly reflect the true contribution of the parameter level. Second, using the overall user retention rate of the feature as a benchmark allows for comparison of the retention performance of different parameter levels under the same feature within the same user group and similar usage background. This isolates the overall brand effect of the feature or differences in user groups, highlighting the impact of the parameter level's own characteristics on whether users continue to use the feature. Third, calculating the next-day retention rate by accumulating data day by day within a sliding window can integrate data from multiple days, smooth out daily fluctuations, and avoid drastic jumps in retention rate due to abnormal user behavior on individual days, thereby obtaining a stable and reliable average retention level within the window period.

[0052] More specifically, for example, suppose the overall user retention rate for the camera function within a 30-day sliding window is 40%. For the telephoto lens, the statistics show a total of 1000 users who used the telephoto lens within 30 days, with 520 still active the next day, resulting in a 52% retention rate for the telephoto lens. For the standard lens, the total number of users is 5000, with 1900 still active the next day, resulting in a 38% retention rate. Therefore, the retention contribution factor for the telephoto lens is (52% - 40%) ÷ (40% + 1%) = 29.27%, and the retention contribution factor for the standard lens is (38% - 40%) ÷ (40% + 1%) = -4.88%. This indicates that user retention rates for telephoto lenses are significantly higher than the overall camera performance, suggesting a positive contribution to user retention, likely because it meets the essential needs of professional users. Conversely, user retention rates for standard lenses are slightly lower than the overall average, indicating that while standard lenses have a large user base, they haven't effectively encouraged continued use, potentially due to insufficient user engagement. Through this calculation, the system can identify high-value parameter settings with fewer users but higher retention rates, and parameter settings with many users but lower retention rates that require improvement. This provides a more precise basis for optimizing resource allocation, rather than simply relying on absolute values ​​of usage frequency or retention rate.

[0053] In this embodiment, the steps for calculating the synergy factor of the functional parameters are as follows: Analyze the correlation strength value for each functional parameter; The session window can be divided within a sliding window by setting the session duration. Count the total number of session windows, the number of session windows containing each function parameter, and the number of session windows containing two function parameters simultaneously.

[0054] Based on the ratio of the number of sessions containing a single functional parameter to the total number of sessions in the session window, the individual probability value of each functional parameter is determined; based on the ratio of the number of sessions containing two functional parameters to the total number of sessions, the joint probability value is determined; after smoothing the session count, the association strength value is obtained by taking the natural logarithm of the ratio of the joint probability value to the product of the two individual probability values.

[0055] The association strength value is a statistic used to measure the strength of the association between two events. It is calculated by dividing the joint probability of the two events by the product of their individual probabilities, and then taking the natural logarithm. A association strength value greater than 0 indicates that the probability of the two events occurring simultaneously is higher than their probability of occurring simultaneously under independent random conditions, meaning there is a positive cooperative relationship between them. A association strength value equal to 0 indicates that the two events are independent. A association strength value less than 0 indicates that the two events tend to be mutually exclusive. In this scheme, the association strength value is used to measure the degree of association between two functional parameters when they are used sequentially or jointly by the same user within the same session window, thereby determining whether there is a cooperative tendency in the use of functional parameters.

[0056] Analyzing the correlation strength values ​​of each functional parameter and constructing a co-occurrence matrix for functional parameters can uncover hidden accompanying usage patterns between different functional parameters, compensating for the shortcomings of activity factors and retention contribution factors that only focus on the internal performance of a single parameter. Its advantage lies in identifying which functional parameters have strong correlations. For example, users often open the photo album to view photos immediately after using the camera's telephoto mode; this correlation is captured through positive correlation strength values, helping to understand user operation paths and the functional ecosystem. As a component of usage cue scoring, the synergy factor reflects the ability of a functional parameter to drive the use of other functional parameters. Parameters with high synergy factors, even if their own usage frequency is low, may still have a significant impact on the overall experience and should be given attention during optimization. Only retaining functional parameter pairs with correlation strength greater than 0 to construct the co-occurrence matrix can filter out irrelevant or mutually exclusive noise relationships, making the synergy factor more interpretable and stable, and avoiding a large number of negative values ​​interfering with the scoring.

[0057] For example, within a 30-day sliding window, there are 1000 session windows. The camera telephoto mode appears in 150 sessions, the album function appears in 300 sessions, and both functions appear simultaneously in 80 sessions. Therefore, the individual probability of the telephoto mode is P(telephoto) = 150 / 1000 = 0.15, the individual probability of the album function is P(album) = 300 / 1000 = 0.3, and the joint probability of both is P(telephoto, album) = 80 / 1000 = 0.08. The correlation strength value is ln(0.08 / (0.15×0.3)) = ln(1.778)≈0.575, which is greater than 0. This indicates a positive synergistic relationship between the camera's telephoto mode and the album function. The probability of a user opening the album to view photos after taking photos with telephoto lens is significantly higher than at random levels. Based on this positive correlation strength value, the element corresponding to the album function in the co-occurrence matrix of functional parameters for telephoto mode is 0.575. This value will be included in the summation of the correlation strength value for telephoto mode, thereby increasing its synergistic factor. This result shows that telephoto mode is not only important for its own user experience but also forms a linkage with the album function. Optimizing the quality of telephoto shooting and the viewing process may simultaneously improve album user retention, providing a correlation basis for the formulation of optimization strategies that goes beyond the perspective of a single parameter.

[0058] Furthermore, by setting session time, when dividing the session window within the sliding window, the specific method for setting the session time is usually to define a threshold for the time interval between adjacent events. When the time interval between two consecutive behavioral events of the same user exceeds this threshold, it is considered that the previous session ends and the next session begins; otherwise, the two events belong to the same session. Factors to consider when setting the session time threshold include: First, the usage characteristics of the product functions. For example, continuous functions such as camera, charging, and audio / video playback may last for a relatively long time with a single use, while instantaneous functions such as settings and notification viewing have a very short usage time. Therefore, the threshold needs to be determined in conjunction with the typical usage duration of the analyzed function parameters. Then, the analysis purpose should be considered. This solution analyzes the collaborative usage relationship between function parameters. If the threshold is too short, what should be a continuous use will be divided into multiple sessions, thus underestimating the co-occurrence relationship of function parameters in the same usage context. If the threshold is too long, irrelevant functions from different user usage contexts will be mixed into the same session, causing the correlation strength value to be inflated by irrelevant co-occurrence, reducing the accuracy of the collaborative relationship. Therefore, the session time threshold is usually selected between a few minutes and tens of minutes. For example, in this solution, a new session is divided when the interval between adjacent events exceeds 30 minutes. This value takes into account the single usage duration of most electronic product function parameters, the continuity of user operations, and the sparsity of data, and can reflect the positive collaborative relationship between function parameters relatively stably. At the same time, it is also convenient to calculate and determine directly through event timestamps in engineering.

[0059] Construct a co-occurrence matrix of function-function parameters based on correlation strength values; Obtain the association strength values ​​of all function parameter pairs, retain only the function parameter pairs with an association strength greater than 0, and fill the matrix with the association strength values ​​greater than 0, with the function parameters as the rows and columns, to construct the function-function parameter co-occurrence matrix.

[0060] The co-occurrence matrix of functions and parameters based on association strength values ​​is constructed because this matrix can centrally express the complex synergistic relationships between functional parameters in a structured numerical form, providing a direct and clear data foundation for subsequent calculations of the synergistic factor of each functional parameter. The specific advantages of this approach are that by arranging the association strength values ​​between pairs of functional parameters by row and column, it is easy to see at a glance whether there is a positive synergistic relationship and its strength between any two functional parameters, avoiding the tedious process of searching through scattered session statistics one by one. When constructing the matrix, only pairs of functional parameters with an association strength greater than 0 are retained, effectively filtering out independent or mutually exclusive parameter relationships. This ensures that the matrix stores only positive associations with actual synergistic significance, reducing noise interference and guaranteeing the accuracy and stability of subsequent synergistic factor calculations. Based on the matrix, the association strength values ​​of each functional parameter's row or column can be easily summed, and then normalized by dividing by the maximum sum among all functional parameters to obtain the synergistic factor of that functional parameter. The calculation process is simple and interpretable. For example, if the correlation strength between the telephoto mode and the album function is 0.575, and the correlation strength between the wide-angle mode and the album function is -0.2, only the positive correlation strength value of the former is retained when constructing the matrix, while the latter is set to 0 or ignored. In this way, when calculating the synergy factor of the telephoto mode, only parameters such as the album that have a positive correlation are accumulated, avoiding the negative correlation between the wide-angle and the album function from lowering the synergy score of the telephoto mode. This allows the synergy factor to truly reflect the positive linkage ability of the parameter with other parameters, providing reliable support for the rationality of using cue scoring and the targeting of optimization strategies.

[0061] Obtain the co-factor of the functional parameters; Based on the function-function parameter co-occurrence matrix, for each function parameter, the sum of its correlation strength values ​​with other function parameters is divided by the sum of the maximum value of the corresponding row sums of all function parameters and ε, and then normalized to obtain the co-occurrence factor of the function parameter.

[0062] The purpose of obtaining the synergy factor for functional parameters is to quantify the relative strength of the positive synergy between that functional parameter and other functional parameters within the entire functional ecosystem. This means how many other parameters the functional parameter is positively correlated with and how strong that correlation is. A higher value indicates that the parameter is more likely to interact with other parameters during use, and thus contributes more to the overall user experience ecosystem. This approach compensates for the shortcomings of activity factors and retention contribution factors, which only focus on the usage frequency and retention performance of a single parameter. It supplements the correlation information between parameters from a functional ecosystem perspective, making the usage clue scoring more comprehensive. It can identify functional parameters that are "not frequently used themselves but have high synergy value." For example, although the telephoto lens has fewer users, it is strongly correlated with parameters such as the photo album and cloud backup. Optimizing it may improve the user experience of multiple related parameters. Therefore, when resources are limited, such parameters can be prioritized. Through normalization, the synergy factor becomes comparable across different functions, facilitating its unified participation in subsequent weighted synthesis of usage clue scoring.

[0063] When obtaining synergistic factors based on the function-function parameter co-occurrence matrix, firstly, sum all positive correlation strength values ​​in the row of each function parameter in the matrix to obtain the total correlation strength of that function parameter; then, find the maximum value of the total correlation strength among all function parameters; finally, divide the total correlation strength of that function parameter by this maximum value, and the result is the normalized synergistic factor, with a value range between 0 and 1.

[0064] More specifically, for example, when the correlation strength between the telephoto lens and the album is 0.575, the correlation strength with cloud backup is 0.32, and the positive correlation strength with other parameters is 0, then the total correlation strength of the telephoto lens is 0.895; the correlation strength between the wide-angle lens and the album is 0.2, and the positive correlation strength with other parameters is 0, then the total correlation strength of the wide-angle lens is 0.2; when the total correlation strength of all functional parameters is the largest at 0.895 for the telephoto lens, then the co-occurrence factor of the telephoto lens is 0.895 ÷ (0.895 + 0.001) ≈ 1, and the co-occurrence factor of the wide-angle lens is 0.2 ÷ (0.895 + 0.001) ≈ 0.223. This indicates that the telephoto mode has strong interrelationships within the functional ecosystem and tends to be used in conjunction with parameters such as the photo album and cloud backup. Optimizing the telephoto mode may simultaneously affect the user experience of the photo album and cloud backup. Therefore, its synergy factor is high and should be reflected in the usage clue scoring to help the system more reasonably evaluate the optimization value of this functional parameter.

[0065] In this embodiment, the step of extracting the user satisfaction factor is as follows: We filter records related to each function parameter and scenario from user feedback data, determine whether each piece of feedback is positive or negative, count the number of positive feedback and the total number of feedback, and use the Laplace smoothing formula to extract the satisfaction factor.

[0066] Extracting satisfaction factors from user feedback data in user behavior event tables is crucial because users' subjective evaluations directly reflect their genuine feelings about functional parameters. This compensates for the shortcomings of behavioral metrics such as usage frequency, usage duration, and core operation completion rate in conveying the dimension of "user satisfaction." Behavioral data can only indicate how many times a user used the app, how long they used it, and whether they completed the operation, but it cannot determine whether the user felt happy or disappointed during the process. In contrast, feedback data such as app store reviews, customer service tickets, and satisfaction questionnaires represent users' proactive opinions and contain clear emotional inclinations. Therefore, extracting satisfaction factors from these sources allows usage clue scoring to more closely reflect the user's actual experience, improving the effectiveness of scoring and the targeting of optimization decisions. The specific implementation steps are as follows: First, filter records whose event type belongs to the user feedback category from the structured user behavior event table, such as "app store reviews," "customer service tickets," and "satisfaction surveys." Then, filter feedback related to the target function parameters based on the function codes and function parameter fields in the records. Next, determine the sentiment polarity of each feedback item to identify whether it is positive or negative. The determination method can use a preset sentiment dictionary matching or rule recognition. For example, positive words such as "easy to use," "clear," "satisfactory," and "smooth" are considered positive, while negative words such as "laggy," "blurry," "difficult to use," and "overheating" are considered negative. For satisfaction surveys, the rating is directly determined as positive based on whether it is higher than a set threshold. Next, the number of positive feedbacks and the total number of feedbacks for this function parameter within the sliding window are counted. The total number of feedbacks includes positive, negative, and neutral feedback. Finally, the satisfaction factor is calculated using a proportional formula with Laplace smoothing. That is, the satisfaction factor is equal to the ratio of the number of positive feedbacks plus 1 to the total number of feedbacks plus 2. The numerator is increased by 1 and the denominator is increased by 2 to avoid extreme values ​​of 0 or 1 when the number of feedbacks is small, making the satisfaction estimate more robust.

[0067] Furthermore, Laplace smoothing is a smoothing technique whose core idea is to artificially add a small positive number to the statistical count to avoid extreme cases where the probability estimate is 0 or 1 due to insufficient sample size or the absence of a certain type of event. Its basic form typically involves adding a constant (such as 1) to the numerator and a multiple of that constant (such as the number of categories) to the denominator, thus slightly shifting the estimation result towards a uniform distribution and improving the robustness of the estimate.

[0068] In the calculation of the satisfaction factor in this scheme, the Laplace smoothing formula is used: Satisfaction Factor = (Number of Positive Feedback + 1) / (Total Number of Feedback + 2). Adding 1 to the numerator and 2 to the denominator is equivalent to considering both the number of positive feedback and the total number of feedback as 0 when there is no feedback. In this case, the satisfaction factor is (0+1) / (0+2) = 0.5, meaning the initial estimate is neutral. When the number of feedbacks is small, this smoothing can avoid extreme values ​​such as obtaining a satisfaction score of 1.0 due to only one positive feedback or 0 due to the absence of positive feedback. This makes the satisfaction estimate closer to reality without being overly conservative, thus providing a more stable input for subsequent use of clue scoring.

[0069] More specifically, for example, if a certain feature parameter has only one positive feedback and no negative feedback, without smoothing, the satisfaction score would be 1.0, while with smoothing it would be (1+1) / (1+2)=0.667, better reflecting the uncertainty when the sample is insufficient. The satisfaction factor obtained through the above steps takes a value between 0 and 1. The closer the value is to 1, the higher the user's satisfaction with the feature parameter. This factor will be weighted together with the activity factor, retention contribution factor, and feature synergy factor to generate a usage clue score, thus ensuring that the usage clue score includes both user behavioral input and user subjective evaluation, providing a more comprehensive quantitative basis for experience-side evaluation for subsequent comprehensive clue scoring and peak stratification.

[0070] In this embodiment, the analysis method using clue scoring is as follows: Based on activity factors, retention contribution factors, satisfaction factors, and functional synergy factors, a usage clue score reflecting user experience is generated using a weighted approach.

[0071] The use of clue scoring, which combines activity factors, retention contribution factors, satisfaction factors, and functional synergy factors using a weighted approach, is because a single factor can only reflect one aspect of the user experience, while the true level of user experience often needs to be judged comprehensively from multiple dimensions such as usage popularity, retention value, subjective satisfaction, and ecosystem synergy.

[0072] The activity factor reflects the frequency and intensity of users' use of the function parameter; the retention contribution factor reflects the contribution of the parameter to users' continued use; the satisfaction factor directly reflects users' subjective evaluation; and the function synergy factor reflects the ecological value of the parameter in conjunction with other parameters. The four complement each other and together constitute a complete characterization of the user experience of the function parameter.

[0073] The advantage of using a weighted synthesis method is that it can unify indicators with different dimensions and meanings into a comparable scoring system, which facilitates the subsequent synthesis of comprehensive lead scores with development lead scores and the analysis of peak positions. The adjustable weights make the scoring model flexible and can highlight the importance of different factors according to different product stages, different functional characteristics or different optimization goals. For example, more attention is paid to retention contribution in the user growth stage and more attention is paid to satisfaction in the experience optimization stage.

[0074] Determining the weight of each factor requires consideration of multiple factors. First, there are the business optimization goals. If the core issue of the current product is user churn, the weight of the retention contribution factor should be increased; if the core issue is poor reputation, the weight of the satisfaction factor should be increased. Second, there is the data quality and reliability. Factors with large data fluctuations or sparse samples should have their weights appropriately reduced to avoid unstable impacts on the overall score. Third, there is the factor's discriminative power and importance. This can be achieved by statistically analyzing the correlation between each factor and the final optimization effect or retention metric through historical data. The weight is determined based on the strength of the correlation; factors with higher correlation have greater weights, and the sum of all factor weights equals 1, ensuring that the lead score remains stable between 0 and 100. In practice, initial weights can be set first, such as 0.35 for activity, 0.25 for retention contribution, 0.25 for satisfaction, and 0.15 for functional synergy. Then, as optimization data accumulates, the weights are dynamically adjusted based on the relationship between each factor and actual retention improvement or changes in user feedback after each optimization. This allows the lead score to gradually approach the actual user experience level, thereby improving the accuracy of the entire lead stratification decision-making process.

[0075] S103: Based on device performance logs, calculate the defect occurrence rate of each functional parameter that triggers device performance type anomalies, analyze defect severity by constructing a function-development clue association table, and generate a development clue score for each functional parameter based on the defect occurrence rate and defect severity.

[0076] This step involves building a functional parameter-development clue association table based on device performance logs. It calculates the probability that each functional parameter will trigger abnormal device performance types such as abnormal CPU operating frequency, insufficient memory, sudden drop in power, and excessive temperature. It also analyzes the defect occurrence rate and defect severity, and finally generates a development clue score for each functional parameter. Its advantage lies in quantifying the prevalence and severity of technical defects based on the objective device status, and accurately locating the stability issues of specific functional parameter levels, rather than just staying at the overall functional level. Compared to existing technologies, which typically only count the overall crash rate or lag rate of a function, or rely solely on user feedback to indirectly infer technical problems, this approach fails to differentiate between defects at different parameter levels within the same function, and struggles to simultaneously measure the frequency of defects and the severity of each defect's impact on user experience. In contrast, this step correlates device performance anomalies by type and functional parameter level, and calculates the defect occurrence rate (number of anomalies divided by number of startup events) and defect severity (weighted average based on preset severity weights). This allows the development lead scoring to reflect both the probability of a problem occurring at a specific parameter level and the impact on users once a problem occurs, significantly improving the granularity and accuracy of technical defect evaluation. In this technical solution, the positive impact of this step on the overall solution is mainly reflected in the following aspects: the development clue score and the user clue score obtained in S102 form a dual-track quantification of subjective experience and objective technology, providing technical input for the weighted synthesis of experience optimization score and development clue score in S104. This allows the comprehensive clue score to simultaneously reflect "how well users find it to use" and "how stable the device is," avoiding the one-sidedness of relying solely on user feedback, which may miss technical hidden dangers, or relying solely on performance logs, which may ignore experience requirements. The analysis of the defect incidence and severity at the functional parameter level enables subsequent peak detection to identify parameter levels where technical defects are concentrated and frequently occur, such as the telephoto level or the 65W fast charging level, thereby supporting the priority of development and repair for these levels and improving the efficiency of optimization resource investment. Thirdly, by setting abnormal thresholds for device performance indicators and monitoring them in real time, the development clue score has the ability to be dynamically updated, which can promptly discover stability issues that occur in new versions or new scenarios, providing timely quantitative basis for closed-loop optimization and strategy adjustment. Therefore, S103 is the objective technical data foundation for achieving precise optimization and positioning at the functional parameter level, supporting comprehensive clue scoring and peak-level hierarchical decision-making, and plays an irreplaceable complementary and reinforcing role in the entire solution.

[0077] In this embodiment, the statistical method for the defect incidence rate is as follows: Based on the current sliding window, extract the function parameter operation events from the user behavior event table that occur in the same user and the same session as the device performance abnormality and are located within the preset attribution time window. The event type is device performance type abnormality. For each functional parameter and each type of device performance anomaly, the total number of events that trigger that anomaly type for that functional parameter is counted as the numerator, and the total number of usage events for that functional parameter is counted as the denominator. The ratio of the two is the defect occurrence rate of that functional parameter in device performance. Build a functional-device anomaly association table; The total defect rate of the functional parameter is obtained by weighting the defect rate of the functional parameter across all device performance types according to a preset weight, and the total defect rate is limited to the range of 0 to 1. The method for judging abnormal equipment performance is as follows: For each metric in the device performance log, a normal range or safety threshold is preset. These metrics are monitored in real time within a sliding window, and when any metric exceeds the normal range, it is determined as a device performance anomaly event.

[0078] The defect incidence rate refers to the probability of a defect occurring in a specific type of equipment performance anomaly for a given functional parameter. It is calculated by dividing the total number of events that trigger this type of equipment performance anomaly by the total number of activation events for that functional parameter. This represents the average proportion of that type of equipment performance anomaly that occurs each time the functional parameter is activated and used; the closer the value is to 1, the more frequent the anomaly. The advantage of this calculation is that it normalizes the anomaly frequency with the number of uses, allowing direct comparisons between functional parameters with different usage frequencies. It avoids misleading results due to variations in absolute anomaly counts caused by different usage volumes, thus accurately reflecting the prevalence of defects in a specific technical dimension for that functional parameter, providing a basis for subsequent technical defect assessments.

[0079] The total defect rate refers to the overall probability that a functional parameter triggers any device performance abnormality, representing the comprehensive technical stability level of that functional parameter. This total defect rate, as a defect occurrence rate factor in the development clue scoring, is combined with defect severity to synthesize the development clue score, ensuring that the technical defect evaluation considers both prevalence and severity, thus improving the rationality and accuracy of the development clue scoring.

[0080] Taking the telephoto mode of a camera as an example. Assume that within a 30-day sliding window, the total number of activation events for the telephoto mode is 10,000. Records associated with the telephoto mode and classified as device performance anomalies are extracted from the user behavior event table. Among these, 200 are overheating anomalies and 100 are CPU overclocking anomalies.

[0081] The defect rate for the long-zoom mode in the case of overheating anomaly is 200 ÷ 10000 = 0.02, or 2%; the defect rate for the long-zoom mode in the case of CPU overfrequency anomaly is 100 ÷ 10000 = 0.01, or 1%.

[0082] The defect incidence rate means that, on average, the probability of an overheating anomaly and the probability of a CPU overclocking anomaly occur each time the telephoto mode is activated are 2%. The advantage of this calculation is that it normalizes the number of abnormal events with the number of activations, allowing for a fair comparison between functional parameters with different usage frequencies. It avoids mistakenly assuming that a parameter is more severely defective simply because it is used more often, thus accurately reflecting the prevalence of defects for that parameter in specific anomaly types.

[0083] Based on the data from the above embodiments, the defect rate for the telephoto lens in cases of overheating anomalies is 0.02, and the defect rate in cases of CPU overclocking anomalies is 0.01. The correct total defect rate should be calculated using the union probability formula: Total Defect Rate = 1 - ∏(1 - Defect Rate). Substituting the values: Total defect rate = 1 - (1 - 0.02) × (1 - 0.01) = 1 - 0.98 × 0.99 = 1 - 0.9702 = 0.0298, or 2.98%.

[0084] The total defect rate represents the probability that any device performance malfunction (overheating or CPU overclocking) will occur each time the telephoto lens is activated. The advantage of this calculation is that the result is always between 0 and 1, and its physical meaning is clear—that is, the probability of at least one malfunction occurring. This avoids the problem of a simple summation leading to a total probability exceeding 1, and more accurately reflects the overall technical defect risk of the functional parameters. This total defect rate will be used as a defect occurrence rate factor in the weighted calculation of the development clue scoring.

[0085] The steps for analyzing defect severity based on the feature-development clue association table are as follows: Assign a severity weight to each type of device performance anomaly, and the sum of all severity weights is 1. Take the number of events for the function parameter in each anomaly type from the association table, multiply it by the corresponding severity weight, and sum them to obtain the weighted severity sum. Then sum the number of events for the function parameter in all anomaly types and add 1 as the denominator. Divide the weighted severity sum by the denominator to obtain the defect severity, and normalize the defect severity to the range of 0 to 1.

[0086] When analyzing defect severity based on the functional parameter-development clue association table, the weighted calculation is performed directly using the event counts corresponding to each anomaly type within each functional parameter row of the table. The reason for relying on this association table is that it aggregates device performance anomaly records scattered across the user behavior event table by functional parameter and anomaly type, retaining the event count and defect occurrence rate for each anomaly type. Subsequent defect severity calculations do not require rescanning the event table to quickly retrieve the necessary counts, ensuring both computational efficiency and ensuring that the defect occurrence rate and defect severity use the same data source, avoiding inconsistencies between the two statistical analyses. The structure of the association table clearly defines the correspondence between functional parameters and each anomaly type, providing a direct data foundation for multi-dimensional defect analysis.

[0087] When calculating defect severity, a severity weight is pre-assigned for each type of device performance anomaly. For example, overheating has a weight of 0.4, CPU overclocking has a weight of 0.3, insufficient memory has a weight of 0.2, and sudden battery drop has a weight of 0.1. The higher the weight, the greater the impact of the anomaly on user experience and device safety. The number of events triggered by the function parameter for each anomaly type is retrieved from the association table. The number of events for each anomaly type is multiplied by the corresponding severity weight, and then summed to obtain the weighted severity total. Simultaneously, the number of events for this function parameter across all anomaly types is summed and 1 is added to the denominator. The weighted severity sum is then divided by this denominator to obtain the defect severity, which is then normalized to the range of 0 to 1.

[0088] Within a 30-day sliding window, the number of device performance anomaly events recorded in the Function Parameters-Development Clues Association Table for this function parameter is as follows: overheating 20 times, CPU overclocking 30 times, insufficient memory 10 times, and sudden power drop 5 times. Severity weights are pre-set for each anomaly type, and the sum of the weights is 1. For example, overheating has a weight of 0.4, CPU overclocking has a weight of 0.3, insufficient memory has a weight of 0.2, and sudden power drop has a weight of 0.1.

[0089] First, calculate the weighted severity sum: 20 × 0.4 = 8, 30 × 0.3 = 9, 10 × 0.2 = 2, 5 × 0.1 = 0.5. Add the four results together: 8 + 9 + 2 + 0.5 = 19.5, which gives a weighted severity total of 19.5.

[0090] Then, the total number of events for this function parameter across all exception types is calculated as follows: 20 + 30 + 10 + 5 = 65.

[0091] Divide the weighted severity sum by the total number of events and add 1: 19.5 ÷ (65 + 1) = 19.5 ÷ 66 ≈ 0.2955, so the defect severity is approximately 0.2955.

[0092] The purpose of this normalization is to unify the defect severity to a dimensionless range of 0 to 1, so that it can be combined with the defect occurrence rate, which is also in the range of 0 to 1, on a unified scale to form a development clue score, while ensuring that the defect severity between different functional parameters can be directly compared.

[0093] Furthermore, the severity weighting should be based on the degree of impact of device performance anomalies on user experience and device security. Generally, the greater the impact on user experience and the more likely the anomaly type is to cause device damage or forced interruption, the higher its severity weight should be set. For example, overheating may cause the device to force downclock, automatically shut down, or even damage the hardware, which has a serious impact on users, so its weight can be set to 0.4; CPU overclocking may cause lag and unresponsive applications, which has a significant impact, so its weight can be set to 0.3; insufficient memory may cause applications to be killed or cause lag when switching between applications, so its weight can be set to 0.2; although a sudden drop in battery power affects battery life, it generally does not immediately interrupt use, so its weight can be set to 0.1. When setting the weighting, the following factors need to be considered: First, the degree to which the anomaly disrupts the core user experience. If the anomaly directly causes functionality to become unavailable or the device to restart, its weighting should be significantly increased. Second, the potential damage the anomaly may cause to the long-term reliability of the device. For example, high temperatures and overcurrents accelerate hardware aging, so their weighting should be higher than that of general performance fluctuations. Third, the intensity perceptible to the user. Stuttering, overheating, and power loss that users can clearly perceive have a greater impact on the experience than background performance fluctuations, so their weighting should be increased accordingly. Fourth, the ease and cost of anomaly recovery. The slower the recovery and the more user intervention required, the higher the weighting. At the same time, the weighting should be adjusted differentiated based on product type and functional characteristics. For example, for products where battery life is a core selling point, the weighting of sudden battery drop anomalies should be appropriately increased; for high-performance devices with prominent overheating issues, the weighting of temperature anomalies should be appropriately increased.

[0094] This method of calculating defect severity is used because different anomaly types have significantly different impacts on users, and cannot be simply averaged based on the number of occurrences. For example, an overheating event might force the device to reduce its frequency or even shut down, while a memory shortage might only cause minor lag. Simply averaging by the number of occurrences would underestimate the harm of overheating anomalies. Weighted averaging gives higher-severity anomalies a greater proportion of the results, more accurately reflecting the severity level of functional parameter defects. Adding 1 to the denominator avoids division by zero when no functional parameter anomalies occur, and also serves to smooth the results, preventing extreme results when the sample size is extremely small.

[0095] Defect severity represents the average severity of a device's performance anomaly each time that functional parameter fails. A higher value indicates a more severe impact on users and the device if the functional parameter malfunctions. The advantage of this calculation is that it quantifies the prevalence and severity of defects separately. Defect severity, along with defect occurrence rate, constitutes two dimensions of development lead scoring, allowing it to simultaneously answer the questions of "how frequent are defects" and "how severe are defects?" This avoids focusing solely on defect occurrence rate while ignoring high-risk anomalies, and also avoids focusing solely on the number of anomalies while ignoring the differences in severity among different anomalies. This provides a more comprehensive assessment of the technical stability of functional parameters and offers a precise basis for formulating optimization strategies.

[0096] In this embodiment, the method for generating development clue scores is as follows; Subtract the total defect rate from 1 to get the probability that no defect occurred. Subtract the defect severity from 1 to get the probability that the severity was not reached. Multiply the probability that no defect occurred by the probability that the severity was not reached to get the probability that neither a defect occurred nor the severity was reached. Subtract this product from 1 to get the probability that at least one defect occurred and the severity was reached, and then multiply by 100. The calculation formula is:

[0097] in, To develop lead scoring, For defect severity, This represents the total defect rate.

[0098] This method treats defect occurrence rate and defect severity as two independent probabilistic events, calculating the combined probability of "at least one defect occurring and reaching the corresponding severity level," thus generating a development clue score. The development clue score is a quantitative assessment of the overall risk level of each functional parameter in terms of technical defects, represented by a value from 0 to 1. This score represents the overall risk level of device performance anomalies occurring for that functional parameter within the current sliding window; the higher the value, the more unstable the functional parameter is, and the more priority the development team needs to prioritize technical fixes.

[0099] The core logic of this method is that the defect occurrence rate represents the frequency of defect occurrences, while the defect severity represents the impact on user experience and device security once a defect occurs. Both together determine the technical risk of functional parameters; the absence of either one makes it impossible to fully characterize the risk level. Therefore, the development lead scoring should measure the joint probability of "both a defect occurring and a defect being severe," rather than a simple addition of the two or a linear combination with fixed weights.

[0100] The main advantage of this method is that the score has a clear probabilistic meaning, representing the degree of comprehensive defect risk faced by functional parameters during use. This avoids the distortion problem of linear weighting under extreme combinations. For example, when the defect incidence rate is extremely high but the severity is extremely low, linear weighting may give a moderately high score, while the joint probability method will give a more reasonable risk value. When both are high, the score quickly approaches 1, which can effectively highlight the high-risk parameter level. Therefore, using the joint risk probability method to generate development clue scores can more scientifically quantify the technical defect risk of functional parameters and provide reliable technical input for subsequent comprehensive clue scoring and peak position, and stratification.

[0101] Furthermore, the usage lead scores and development lead scores in this application are not subjective scores directly given by users or experts in the traditional sense. Instead, they are quantitative evaluation values ​​obtained by fusing and calculating multi-dimensional data using statistical methods. This is because the essence of scoring is to assign a comparable numerical value to the measured object to reflect its level or degree. The usage lead score is a weighted synthesis of activity factors, retention contribution factors, satisfaction factors, and functional synergy factors, compressing the four dimensions of user usage frequency, retention value, subjective satisfaction, and ecosystem synergy into a 0-1 value to reflect the user experience of functional parameters. The development lead score is a comprehensive risk value of technical defects in functional parameters calculated by combining defect incidence rate and defect severity using a joint risk probability method, which is also a 0-1 score. The reason for not using traditional manual scoring or simple rating methods is that this application needs to process massive amounts of user-generated behavioral data, device performance logs, and feedback data. Traditional scoring methods cannot cover such a large-scale, real-time, dynamically changing scenario, and manual scoring is highly subjective, inconsistent, and costly. In contrast, this application calculates scores using deterministic statistical formulas, which can automatically, objectively, and interpretably fuse multi-source data into a unified value. Furthermore, the scoring results can be updated in real time with the sliding window, meeting the needs for precise optimization and positioning of electronic product functional parameters. Therefore, the scoring in this application is a broad data-driven quantitative evaluation. Although the calculation method differs from traditional manual scoring, it is better able to support subsequent peak detection, hierarchical decision-making, and closed-loop optimization, demonstrating significant technical advantages.

[0102] S104: Convert usage clue scores into experience optimization scores, and generate a comprehensive clue score by weighting it with development clue scores. Based on the update step size value, introduce an exponential time decay weight to update the comprehensive clue score in each sliding window.

[0103] This step transforms usage lead scores into experience optimization scores, which are then weighted and synthesized with development lead scores to form a comprehensive lead score. The comprehensive lead score within the sliding window is then updated using an exponential time decay weight based on the update step size. Its advantage lies in unifying two originally opposing quantitative indicators—user experience and technical defect indicators—into a single "optimization value" direction. This allows the comprehensive lead score to directly reflect the urgency of comprehensive functional parameter optimization, solving the problem of direct comparison and integration of experience and defect indicators in existing technologies. Compared to existing technologies, which typically separately calculate user satisfaction or crash rate, or simply add them with fixed weights, not only lack dynamic basis for weight setting but also fail to reflect the impact of recent behavioral changes on optimization needs, this step inverts the usage lead score to obtain an experience optimization score (higher score for worse experience), and then weights and synthesizes it with the development lead score. This ensures that the two types of scores are aligned in direction and can dynamically adjust their weights based on historical optimization benefits, thus more accurately reflecting the relative value of experience optimization and technical fixes. Meanwhile, by introducing exponential time decay weights into the comprehensive clue scores within historical time windows based on the update step size, recent scores contribute more to current optimization decisions, improving the responsiveness of scores to product and user behavior changes and avoiding lag and misjudgment caused by fixed-window equal-weight statistics. In this technical solution, the positive impact of this step on the overall solution is mainly reflected in the following: the comprehensive clue score merges user experience and technical defects into a single comparable optimization value, providing direct input for kernel density estimation and peak detection in S105. This allows the position and height of the peak to simultaneously reflect the combined urgency of experience and technical issues. The dynamic weighting mechanism enables the comprehensive clue score to adaptively adjust according to actual optimization benefits, ensuring that subsequent tiered decisions always focus on the parameter levels that contribute most to retention and satisfaction improvement. The exponential time decay update makes the comprehensive clue score timely, enabling timely capture of newly emerging problems or the improvement effects of optimizations, supporting closed-loop iteration and dynamic adjustment of strategies. Therefore, S104 is a key bridge connecting dual-track scoring and peak-level tiered decision-making, realizing the transformation from multi-source data to unified optimization value and laying the foundation for precise optimization positioning at the functional parameter level.

[0104] In this embodiment, the steps for generating a comprehensive clue score are as follows: After performing Min-Max normalization on the usage clue score, take 1 and subtract the normalized usage clue score to obtain the experience optimization score. Then, merge it with the development clue score according to the preset weight to generate the comprehensive clue score.

[0105] The core purpose of this method is to unify the direction of the two scores by subtracting the maximum value of the usage lead score from the total usage lead score. A higher usage lead score indicates a better user experience, while a higher development lead score indicates more severe technical defects. Since they are in opposite directions, they cannot be directly added together. However, by subtracting the usage lead score from 1, we obtain the experience optimization score. A higher experience optimization score indicates a worse user experience and a greater need for optimization. Thus, both the experience optimization score and the development lead score become indicators of "the higher the score, the greater the need for optimization," with consistent direction and values ​​ranging from 0 to 1. Therefore, a weighted composite score can be obtained. The advantage of this method is that it avoids the weighting cancellation or logical confusion caused by opposite directions, giving the composite lead score a clear physical meaning while preserving the linear relationship of the original scores. It directly maps experience optimization needs and technical repair needs to the same optimization value scale, facilitating subsequent weight adjustments based on actual business benefits. The comprehensive clue score represents the urgency of optimizing a functional parameter within the current sliding window. A higher score indicates a more severe problem with that parameter in terms of user experience or technical stability, requiring priority resource allocation for optimization. Therefore, this score forms the basis for subsequent kernel density estimation and peak stratification; peak identification directly pinpoints specific functional parameter levels with prominent optimization needs.

[0106] The weighting of experience optimization score and development lead score in the overall lead scoring can be determined using a dynamic adjustment method based on historical optimization benefits. Specifically, firstly, the average change in retention rate the day after each unit change in experience optimization score and development lead score is calculated using historical data to obtain the experience optimization benefit and development optimization benefit. Then, both are normalized after taking their positive values ​​to obtain their respective weights. If historical data is insufficient, initial weights can be set, for example, experience optimization weight of 0.5 and development optimization weight of 0.5, and gradually adjusted as optimization data accumulates. When determining weights, the following factors need to be considered: business optimization goals. If the main problem of the current product is user churn or poor reputation, the weight of experience optimization can be appropriately increased. If the main problem is technical stability issues such as crashes or lag, the weight of development optimization can be increased. The usage characteristics of functional parameters. For functional parameters where technical stability is critical (such as charging and performance modes), the weight of development optimization can be appropriately increased. For functional parameters where interactive experience is the primary concern (such as camera focal length and sound effect modes), the weight of experience optimization can be appropriately increased. Through the above dynamic weight determination method, the comprehensive lead score can more accurately reflect the actual optimization priority of each functional parameter under the current product stage and business goals, providing a more reliable basis for subsequent peak stratification and optimization decisions.

[0107] The steps for updating the overall clue score are as follows: The time window is slid along the time axis according to the update step size, and the comprehensive clue score is recalculated for the slid time window. An exponential time decay weight is introduced into the comprehensive clue score sequence within the historical time window. The comprehensive clue scores of the current time window and the historical time window are weighted and averaged to obtain the final comprehensive clue score used for analysis.

[0108] The purpose of updating the comprehensive clue score using this method is to ensure that the final comprehensive clue score used for analysis retains some historical information to smooth out random fluctuations while highlighting the true impact of recent behavioral changes on optimization needs, thereby improving the accuracy and timeliness of peak detection and hierarchical decision-making. Specifically, when the comprehensive clue score is recalculated based solely on the current sliding window, the score may fluctuate significantly due to sudden user behavior or device anomalies on individual dates, leading to unstable peaks in subsequent kernel density estimation and even generating misleading optimization clues. By sliding forward along the time axis and retaining the score sequence of multiple historical time windows, and then introducing exponential time decay weights, the score of the window closer to the current moment contributes more to the final result, while the score of the older window contributes less. This suppresses short-term noise while keeping the score sensitive to new product problems or the improvement effects of optimization, avoiding the lag caused by simple equal-weighted averaging. The determination of the historical time length mainly considers the following aspects: First, the selection of decay weights. The decay weights determine the rate of weight decay. They are typically set based on the iteration cycle of product function parameters and the rate of change in user behavior. For example, with a decay weight of 0.05 / day, the half-life is approximately 14 days. After about 30 days, the weight decays to approximately 0.25, which effectively balances smoothness and timeliness. Second, the requirement for statistical stability. Too few historical windows result in insufficient smoothing, while too many introduce outdated information. Therefore, a time span covering 2 to 3 half-lives is generally chosen as the historical sequence length, ensuring the weight decays to a sufficiently small value without being completely negligible. Furthermore, the actual rhythm of the business scenario must be considered. For example, when the function parameter version update cycle is short, the historical length can be appropriately shortened to reflect new version changes more quickly, while during version stability periods, the historical length can be appropriately extended to enhance scoring stability. In summary, the historical time length can typically be set to the same 30 days as the sliding window, with the window scoring sequence retained daily in steps. Combined with an exponential time decay weight for weighted averaging, a good balance can be achieved between timeliness, stability, and computational cost.

[0109] Please see Figure 2This is an exponential time decay weight curve provided in the embodiments of this application. The figure shows the curve relationship of the exponential time decay weight changing with the time interval. The figure intuitively verifies the rationality and effectiveness of introducing the exponential time decay mechanism in the comprehensive clue score update of this scheme. The horizontal axis represents the time interval (days) from the window to the current time, and the vertical axis represents the weight value (0~1). The following technical effects can be observed from the figure: (1) Decay characteristics: The weight starts from 1.0 and decays exponentially with the increase of the time interval. It drops to 0.5 (half-life) at about 14 days and to about 0.22 at 30 days. This decay law makes the contribution of the recent score to the current optimization decision much greater than that of the long-term score. (2) Timeliness guarantee: The decay weight enables the scheme to maintain sensitivity to newly emerging problems while suppressing short-term noise. When new technical defects or user experience problems appear in the functional parameters, they can be significantly reflected in the comprehensive score within about 2-3 half-lives (14-42 days). (3) Balance: The decay curve effectively balances the two objectives of "smoothing noise" and "timely response". Compared to simple equal-weighted averages, exponential decay updates avoid the interference of outdated data on current decisions; compared to using only the current window data, they can smooth out accidental fluctuations through historical information.

[0110] An exponential time decay weight is introduced into the comprehensive clue score within the historical time window, making recent scores have a greater impact on current decisions. Taking focal length settings as an example, when the comprehensive score of the current window is updated, the score of windows closer to the current time has a higher weight, and the weight of windows further away is lower, thus smoothing out noise while maintaining sensitivity to newly emerging issues. The curve in the figure visually illustrates this exponential decay law.

[0111] Please see Figure 3 This is a dynamic update graph of comprehensive clue scoring provided in an embodiment of this application. Figure 3 The dynamic update process of the comprehensive cue score for the focal length parameter (85mm telephoto) in 15 consecutive sliding windows is shown. The black dashed line is the original score, and the red solid line is the score after exponential time decay weighted smoothing. The following technical effects can be observed from the figure: (1) Noise suppression: The original score shows obvious fluctuations around 0.76 (fluctuation range of about ±0.06), while the curve after exponential time decay smoothing is smoother, effectively suppressing the random fluctuations of a single window. (2) Trend tracking: Although the smoothed curve is more stable, it can still track the overall trend change of the original score. When the score shows a continuous rise or fall, the smoothed curve can keep up with the trend within 2-3 windows. (3) Improved comparability: The smoothed score is more comparable, providing a stable input for subsequent peak detection. If the original score is used directly for kernel density estimation, random fluctuations may lead to unstable peaks. This figure verifies the technical effect of the exponential time decay update mechanism in improving the stability and timeliness of the score from the perspective of actual data.

[0112] Taking the focal length parameter (85mm telephoto) as an example, the original score fluctuated around 0.76, and the smoothed curve stabilized around 0.76. This indicates that the time decay mechanism can effectively suppress single-window fluctuations, making the scores more comparable and providing a stable input for subsequent peak position detection.

[0113] S105: Based on the current time window, perform kernel density estimation on the comprehensive clue score to obtain the probability density function. Then, detect the local maximum value of the density function as the peak position, and obtain the position and score of each peak position. Based on the position of the peak position as the optimization target, the score of each peak position is used as the urgency level, and the peak positions are stratified into different urgency levels. An optimization strategy is then formulated.

[0114] This step estimates the kernel density of the comprehensive clue scores and identifies local maxima of the density function as peaks. Then, based on the comprehensive clue scores corresponding to the peak positions, it classifies the urgency and formulates optimization strategies. Its advantage lies in automatically identifying the clustering centers and differentiation characteristics of functional parameter optimization needs in the score distribution, without requiring manual pre-setting of clusters or reliance on empirical thresholds. Compared to existing technologies, which typically only rank overall functional scores or classify priorities based on fixed thresholds, this approach fails to reveal the distribution pattern and group differentiation of optimization urgency among different parameter levels within the same function, easily mixing parameter levels with significantly different needs. This step, however, uses kernel density estimation to transform discrete scores into continuous probability density curves. By detecting peaks, it can intuitively identify which parameter levels are concentrated in low optimization need regions and which are concentrated in high optimization need regions. The number of peaks reflects the degree of differentiation, and the peak score directly quantifies the optimization urgency of the group. This allows the optimization target to be precise from the vague functional level to the specific parameter level, and it can automatically distinguish between concentrated optimization and categorical optimization scenarios. In this technical solution, the positive impact of this step on the overall solution is mainly reflected in the fact that it takes over the comprehensive clue score output by S104 and transforms the numerical values ​​into actionable optimization decisions. This is a key link connecting the scoring analysis with actual optimization actions. The position of the peak directly corresponds to the urgency of the optimization of the functional parameter level, providing an objective basis for subsequent stratification. For example, the parameter level covered by the peak with a high score is identified as a high-priority optimization object. When multiple peaks are detected, it indicates that there is a significant differentiation in the optimization needs of different parameter level groups. At this time, centralized optimization or classified optimization strategies can be formulated for each peak, such as repairing technical defects or providing parameter mode selection for high-urgency peaks, thereby avoiding a one-size-fits-all approach to the entire function. This method automatically determines the optimization level based on the statistical distribution pattern, reducing the subjectivity and workload of manual judgment, improving the accuracy and response speed of optimization resource allocation, and enabling the overall solution to achieve a closed-loop process from data collection, dual-track scoring, comprehensive synthesis to automatic stratified follow-up, significantly improving the scientificity and efficiency of electronic product functional parameter optimization decisions.

[0115] In this embodiment, the method for obtaining the position and score of each peak is as follows: For each score in the comprehensive clue score, a Gaussian kernel function is placed centered on that score. The ordinate values ​​of the Gaussian kernel function corresponding to all data points at each position on the horizontal axis are added together to obtain a superimposed total curve. The superimposed total curve is normalized to obtain the probability density function. On the probability density curve, all local maxima are found. These local maxima are the peaks in the data distribution, and the comprehensive clue score corresponding to each peak is the score of each peak.

[0116] For each score in the comprehensive clue scoring, a Gaussian kernel function is placed centered on that score. The Gaussian kernel function is a bell-shaped curve, with its highest point at the center, gradually decreasing towards both sides. The width of the curve is determined by the bandwidth; a larger bandwidth results in a smoother curve, while a smaller bandwidth makes the curve closer to the original data points. The ordinate values ​​of the Gaussian kernel functions corresponding to all scores are superimposed on the horizontal axis to obtain a superimposed overall curve. This overall curve forms higher peaks in areas with dense scores and lower peaks in areas with sparse scores. The superimposed overall curve is normalized so that the total area under the curve equals 1, thus obtaining the probability density function. On the probability density curve, all local maxima are found, i.e., the locations where the density values ​​in the left and right neighborhoods of each point are lower than that point. These local maxima are the peaks in the data distribution, and the horizontal coordinate of each peak is the comprehensive clue score, which is the score for each peak. Simultaneously, to avoid small fluctuations caused by noise being misjudged as peaks, a filtering rule can be set to retain only local maxima whose peak height is not lower than a certain percentage of the highest peak as significant peaks.

[0117] Furthermore, bandwidth is a key parameter in the Gaussian kernel function. A larger bandwidth results in a wider Gaussian kernel curve and a smoother probability density curve obtained from kernel density estimation, but this may mask details and true peak positions in the rating distribution. Conversely, a smaller bandwidth makes the curve closer to the original rating points, revealing more local features, but it is also prone to generating numerous small fluctuations caused by noise, leading to the detection of spurious peaks. The bandwidth is typically determined using the Silverman rule: bandwidth equals 1.06 multiplied by the standard deviation of the rating values, then multiplied by the negative fifth power of the sample size. This allows for automatic adjustment of the bandwidth based on the data dispersion and sample size, achieving a balance between undersmoothing and oversmoothing.

[0118] The method obtains the position and score of each peak. Its advantage lies in the fact that kernel density estimation transforms discrete score values ​​into continuous probability density curves, which can intuitively present the clustering center and dispersion of score values ​​without the need for manual pre-setting of the number of clusters or thresholds. The position of the peak directly reflects the concentration level of the optimization needs of a group of functional parameters. The higher the peak score, the higher the urgency of optimization of the group of functional parameters, and the lower the peak score, the weaker the optimization needs. The number of peaks can reveal whether there is a differentiation in optimization needs among functional parameters. For example, detecting two peaks indicates the existence of two obviously different optimization urgency groups, providing an objective basis for formulating centralized optimization or classified optimization strategies. Fourth, the method automatically determines the optimization target and urgency based on the distribution pattern of the data itself, reducing the subjectivity and workload of manual judgment, and making optimization decisions more scientific, accurate and efficient.

[0119] Furthermore, when finding all local maxima on the probability density curve, further filtering can be performed. For example, a minimum peak height threshold can be set, and a relative threshold method can be used. That is, candidate peaks whose peak height is not lower than a certain percentage (e.g., 10%) of the highest peak height can be retained, and candidate peaks lower than this percentage can be regarded as noise peaks and eliminated. Furthermore, in order to avoid the same real peak being split into multiple adjacent peaks due to local fluctuations, a minimum peak spacing can be set. When the scoring positions of two candidate peaks are too close (e.g., less than a certain multiple of the bandwidth or less than 5 points) and there is a very small dip between them, the lower peak is merged into the higher peak.

[0120] Please see Figure 4 This is a schematic diagram of a comprehensive clue score distribution provided in an embodiment of this application. The following technical effects can be observed from the figure: (1) Obvious scoring gradient: The comprehensive clue score gradually increases from 0.18 for F1 to 0.90 for F8, showing an obvious gradient distribution. This indicates that the scheme can effectively distinguish the optimization urgency of different focal length levels. (2) Prominent differentiation characteristics: The scoring distribution clearly shows three clustering areas - F1-F2 are concentrated in the low score area (0.18-0.22), F3-F5 are concentrated in the medium score area (0.40-0.50), and F6-F8 are concentrated in the high score area (0.76-0.90). (3) Effective threshold stratification: The blue bars (low score level) and red bars (high score level) are naturally divided on both sides of the threshold line, indicating that the comprehensive clue score can effectively support automated urgency stratification. The figure clearly shows the differentiation ability of the comprehensive clue score at the functional parameter level, verifying the effectiveness of the dual-track scoring fusion method.

[0121] Please see Figures 5-6The application provides a schematic diagram of a comprehensive clue score distribution, a schematic diagram of a process of superimposing Gaussian kernel density estimation on comprehensive clue scores, and a schematic diagram of peak position and score detection of probability density function provided in the application.

[0122] As shown in the figure Figure 5 The horizontal axis represents the overall clue score, which ranges from 0 to 1 after normalization; the vertical axis represents the probability density. The light blue curves in the graph represent the Gaussian kernel function corresponding to each data point (focal length score). Each kernel function is bell-shaped, with its complete shape untrunculated. The thick black line represents the probability density function obtained by superimposing and normalizing all kernel functions. The red dots represent the original data points, i.e., the overall score values ​​for the eight focal lengths.

[0123] The eight score values ​​are located at 0.18, 0.22, 0.40, 0.45, 0.50, 0.76, 0.82, and 0.90. A Gaussian kernel function (light blue bell-shaped curve) is placed at each score value with a bandwidth of 0.06. All kernel functions are superimposed and normalized to obtain the probability density function (dark blue curve). This function forms a higher density in densely populated score regions (e.g., 0.2-0.40) and a lower density in sparse regions. Normalization ensures the area under the curve is 1, and the horizontal axis remains within the 0-1 range, facilitating peak detection.

[0124] Probability density curve characteristics: The probability density function after superposition and normalization forms a significant peak in the dense scoring area and a lower density in the sparse scoring area. This curve transforms discrete scoring values ​​into a continuous probability distribution. (3) Automatic clustering capability: Kernel density estimation does not require manual preset of the number of clusters or thresholds, and automatically identifies the scoring cluster centers based entirely on the distribution pattern of the data itself. Compared with traditional clustering methods such as K-Means, this method does not rely on the selection of initial values, and the results are more stable and reliable. This figure shows the core process of kernel density estimation transforming discrete scores into continuous probability density, verifying the technical advantages of this method in automatically discovering the characteristics of scoring distribution.

[0125] like Figure 6 As shown, the horizontal axis represents the comprehensive clue score, ranging from 0 to 1; the vertical axis represents the probability density. The black curve is the probability density function, the red diamond marks the detected peaks, and the dashed lines indicate the comprehensive score values ​​corresponding to the peaks. Three peaks are marked in the figure, approximately located at 0.22, 0.45, and 0.82.

[0126] Peak position scores directly indicate the urgency of optimization for that group: high-resolution peaks (0.82) correspond to telephoto lenses, indicating high urgency; low-resolution peaks (0.22) correspond to wide-angle lenses, indicating low urgency. Based on this, urgency levels can be stratified: focal lengths covered by high-resolution peaks are prioritized for optimization, with focused optimization strategies developed, such as fixing technical defects like slow autofocus and image stabilization failures in telephoto lenses, or optimizing the user experience of telephoto mode. Mid- and low-resolution peaks are either left untreated or continue to be monitored.

[0127] While comprehensive clue scoring can quantify the optimization urgency of each functional parameter, directly ranking them by score and setting a threshold to select optimization targets faces problems such as subjective threshold selection and inability to identify natural clustering structures and differentiation among parameters. Kernel density estimation and peak detection aim to solve these problems. Through kernel density estimation, discrete score values ​​are transformed into continuous probability density curves. Regions of dense scores naturally form peaks, which represent the optimization urgency centers of a group of functional parameters with similar scores. The number of peaks directly reflects whether there is differentiation in the optimization needs of functional parameters: a single peak indicates that all parameters have similar urgency and are suitable for concentrated optimization; multiple peaks indicate the existence of significantly different optimization urgency groups that require classified optimization. The horizontal axis score value corresponding to the peak quantifies the urgency of the group, thus enabling automatic stratification without the need to manually set a threshold for "optimization only after a certain score". Therefore, kernel density estimation and peak detection are not a reuse of the comprehensive score, but rather an automatic identification of the group affiliation, differentiation degree and optimization strategy type of the optimization target by utilizing the overall shape of the score distribution. This shifts the optimization decision-making from relying on manual sorting and subjective thresholds to objective stratification based on the data distribution shape, which not only improves the accuracy and stability of the decision-making, but also significantly reduces the cost and bias of manual intervention.

[0128] The optimization strategy is as follows: A threshold for urgency is preset, and the comprehensive clue score is divided into three intervals. If the peak score is not lower than the first threshold, it is judged as high urgency; if the peak score is between the first threshold and the second threshold, it is judged as medium urgency; and if the peak score is lower than the second threshold, it is judged as low urgency.

[0129] After determining the specific optimization targets as the functional parameter levels covered by the peak positions, the comprehensive clue score of each peak position is used as the quantitative basis for the urgency of optimization for that group, and the urgency level is stratified and optimization strategies are formulated according to the following methods.

[0130] The urgency threshold is preset, and the comprehensive clue score is divided into three intervals. If the peak score is not lower than the first threshold, it is judged as high urgency; if the peak score is between the first threshold and the second threshold, it is judged as medium urgency; and if the peak score is lower than the second threshold, it is judged as low urgency. Furthermore, the first threshold and the second threshold can be set to 0.70 and 0.50, respectively. The setting of the first threshold and the second threshold needs to take into account the overall distribution level of the functional parameter comprehensive clue score, the sufficiency of optimization resources, the user-perceived experience difference threshold, and the product stage goals. Specifically, the first threshold (e.g., 0.70) is used to define high urgency. When setting it, it is necessary to consider whether functional parameters with a comprehensive lead score above this value have indeed caused strong user dissatisfaction or technical defects so severe that immediate intervention is required, and whether the number of high-urgency parameter levels matches the available R&D resources. If the threshold is too low, too many high-urgency objects will be created, and resources will be dispersed; if it is too high, truly serious problems may be missed. The second threshold (e.g., 0.50) is used to distinguish between medium and low urgency. When setting it, it is necessary to consider whether functional parameters with scores below this value can be considered to have weak optimization needs and can be observed temporarily. This threshold should be consistent with the theoretical median level or average distribution of comprehensive lead scores to avoid a large number of parameter levels being classified as low urgency and left unprocessed for a long time. Finally, the threshold should be dynamically adjusted during system operation based on changes in peak score distribution and optimization feedback effects to ensure that the urgency level stratification always meets actual business needs.

[0131] Secondly, count the number of peaks. If only one peak is detected, it indicates that the optimization needs of all functional parameters are concentrated around that peak. In this case, if the peak score reaches a high urgency level, all parameter levels covered by that peak are listed as high-priority optimization targets, and major R&D resources are allocated for unified improvement. If the peak score is only at a low urgency level, it is not urgent to handle it for now, and monitoring continues. If multiple peaks are detected, it indicates that the optimization needs of different functional parameter groups have diverged. In this case, it is necessary to handle them separately according to the urgency level of each peak: for peaks with a high urgency level, the parameter levels they cover are listed as high-priority optimization targets; for peaks with a low urgency level, the parameter levels they cover are not processed for now. When there are both high-urgency and low-urgency peaks among the multiple peaks, a classification optimization strategy can be adopted. That is, separate optimization plans are formulated for the parameter levels covered by the high-urgency peaks, such as splitting them into independent modes or versions for users to choose from, instead of optimizing all parameter levels uniformly.

[0132] Using the above method, S105 can transform peak position scores into clear urgency levels and differentiated optimization strategies, achieving precise optimization positioning at the functional parameter level, avoiding a one-size-fits-all approach to the entire function, while ensuring that optimization resources are prioritized for the parameter groups with the most prominent problems.

[0133] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0134] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0135] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for tiered follow-up of customer leads based on customer behavior data fusion and dynamic scoring, characterized in that, include: Acquire all user operation data of the product's functions, device performance logs, and user feedback data as customer behavior data, and preprocess the customer behavior data; The preprocessed customer behavior data is classified and coded according to user identifier, function identifier, function parameter identifier, operation identifier and equipment performance identifier to form a structured user behavior event table; Set the sliding window and step size, and obtain the user behavior event table within the sliding window; based on the event table, count the number of times each user uses each function parameter, the usage duration, and the core operation completion rate, and construct a user-function parameter usage feature matrix; calculate the activity factor and retention contribution factor of the function parameter based on the usage feature matrix; calculate the association strength value based on the co-occurrence relationship of the function parameters within the session window, and construct a function-function parameter co-occurrence matrix accordingly, and then calculate the synergy factor of the function parameter; Satisfaction factors are extracted based on user feedback data; activity factors, retention contribution factors, satisfaction factors, and collaboration factors are integrated according to preset weights to generate usage clue scores; Based on device performance logs, the defect occurrence rate of each functional parameter triggering device performance type anomalies is statistically analyzed. The severity of defects is analyzed by constructing a function-development clue association table. A development clue score for each functional parameter is generated based on the defect occurrence rate and defect severity. The usage lead score is converted into an experience optimization score, and a comprehensive lead score is generated by weighting it with the development lead score. An exponential time decay weight is introduced to update the comprehensive lead score in each sliding window based on the update step value. Based on the current time window, kernel density estimation is performed on the comprehensive clue score to obtain the probability density function. Then, the local maximum value of the density function is detected as the peak position, and the position and score of each peak position are obtained. The position of the peak position is used as the optimization target, and the score corresponding to each peak position is used as the urgency level. The peak positions are then stratified into different urgency levels, and optimization strategies are formulated.

2. The method according to claim 1, characterized in that, Operational data includes the number of times each function parameter is launched, the duration of use, and the operation completion rate. Device performance logs include CPU operating frequency, remaining memory capacity, remaining battery power, and product temperature; User feedback data includes app store reviews, customer service tickets, and satisfaction surveys; The preprocessing method is as follows: The acquired raw customer behavior data is cleaned by the Flink stream processing framework, which filters out invalid events with missing key codes, removes duplicate reports within a short period of time, and eliminates outliers. The structured user behavior event table includes user code, function code, function parameters, device performance log, event type, timestamp, and event value.

3. The method according to claim 1, characterized in that, The steps for calculating the activity factor and retention contribution factor of the functional parameters are as follows: Constructing user-feature parameters using a feature matrix: Based on the user behavior event table in the sliding window, filter out the corresponding user identifier, function identifier and corresponding function parameter based on the code, and the event type is "start". Count the total number of the records to get the number of times the user uses the function parameter in the sliding window. Filter the usage duration of the event types for this function parameter, and sum the usage duration values ​​to get the total usage duration of this function parameter; The core operation completion rate is obtained by counting the number of times the core operation events of this function parameter occur and dividing the number of times the function parameter is started. Each user and function parameter combination is used to form a triplet, which consists of the number of times it is used, the duration of use, and the completion rate of the core operation. The user-function parameter usage feature matrix is ​​formed with the user as the row, the function parameter as the column, and the triplet as the element. The calculation method for the activity factor: The activity factor is calculated by dividing the average daily number of users who have used each function parameter by the sum of the average daily number of all users who have used the function. The calculation method for the retention contribution factor: For each function parameter, the number of users who used the function parameter on each day and continued to use it on the next day is added up daily to get the total number of retained users. Then, the total number of users who used the function parameter on each day is added up daily to get the total number of users. The total number of retained users is divided by the total number of users to get the user retention rate of the function parameter on the next day. The total number of users who used the function on each day within the sliding window and continued to use it on the next day is accumulated daily to obtain the total number of retained users. The total number of users who used the function on each day within the sliding window is accumulated daily to obtain the total number of active users. The total number of retained users is divided by the total number of active users to obtain the overall user retention rate on the next day. The difference between the user day-one retention rate and the overall user day-one retention rate is then divided by the overall user day-one retention rate. The resulting value is the retention contribution factor of this function parameter within the sliding window.

4. The method according to claim 1, characterized in that, The steps for calculating the synergistic factor of the functional parameters are as follows: Analyze the correlation strength value for each functional parameter; The session window can be divided within a sliding window by setting the session duration. Count the total number of session windows, the number of session windows containing each function parameter, and the number of session windows containing two function parameters simultaneously; Based on the ratio of the number of sessions containing a single functional parameter to the total number of sessions in the session window, determine the individual probability value of each functional parameter; The joint probability value is determined based on the ratio of the number of sessions that simultaneously include two functional parameters to the total number of sessions. After smoothing the session count, the association strength value is obtained by taking the natural logarithm of the ratio of the joint probability value to the product of the two individual probability values. Construct a co-occurrence matrix of function-function parameters based on correlation strength values; Obtain the association strength values ​​of all function parameter pairs, retain only the function parameter pairs with an association strength greater than 0, and fill the matrix with the association strength values ​​greater than 0, with the function parameters as the rows and columns, to construct the function-function parameter co-occurrence matrix; Obtain the co-factor of the functional parameters; Based on the function-function parameter co-occurrence matrix, for each function parameter, the sum of its correlation strength values ​​with other function parameters is divided by the sum of the maximum value of the corresponding row sums of all function parameters and ε, and then normalized to obtain the co-occurrence factor of the function parameter.

5. The method according to claim 1, characterized in that, The steps to extract user satisfaction factors are as follows: We filter records related to each function parameter and scenario from user feedback data, determine whether each piece of feedback is positive or negative, count the number of positive feedback and the total number of feedback, and use the Laplace smoothing formula to extract the satisfaction factor.

6. The method according to claim 1, characterized in that, The analysis method using clue scoring is as follows: Based on activity factors, retention contribution factors, satisfaction factors, and functional synergy factors, a usage clue score reflecting user experience is generated using a weighted approach.

7. The method according to claim 1, characterized in that, The statistical method for defect incidence rate is as follows: Based on the current sliding window, extract the function parameter operation events from the user behavior event table that occur in the same user and the same session as the device performance abnormality and are located within the preset attribution time window. The event type is device performance type abnormality. For each functional parameter and each type of device performance anomaly, the total number of events that trigger that anomaly type for that functional parameter is counted as the numerator, and the total number of usage events for that functional parameter is counted as the denominator. The ratio of the two is the defect occurrence rate of that functional parameter in device performance. Build a functional-device anomaly association table; The defect rate of the function parameter across all device performance types is weighted and averaged according to a preset weight to obtain the total defect rate of the function parameter. The total defect rate is then limited to the range of 0 to 1. The function parameter is used as the row and the type of device performance abnormality is used as the column. Each defect rate is filled into the corresponding position to form a function-development clue association table. The method for judging abnormal equipment performance is as follows: For each metric in the device performance log, a normal range or safety threshold is preset. These metrics are monitored in real time within a sliding window, and when any metric exceeds the normal range, it is determined as a device performance anomaly event. The steps for analyzing defect severity based on the feature-development clue association table are as follows: Assign a severity weight to each type of device performance anomaly, and the sum of all severity weights is 1. Take the number of events for the function parameter in each anomaly type from the association table, multiply it by the corresponding severity weight, and sum them to obtain the weighted severity sum. Then sum the number of events for the function parameter in all anomaly types and add 1 as the denominator. Divide the weighted severity sum by the denominator to obtain the defect severity, and normalize the defect severity to the range of 0 to 1.

8. The method according to claim 1, characterized in that, The method for generating clue scores is as follows; Subtract the total defect rate from 1 to get the probability that no defect occurred. Subtract the defect severity from 1 to get the probability that the severity was not reached. Multiply the probability that no defect occurred by the probability that the severity was not reached to get the probability that neither a defect occurred nor the severity was reached. Subtract this product from 1 to get the probability that at least one defect occurred and the severity was reached, and then multiply by 100.

9. The method according to claim 1, characterized in that, The steps to generate a comprehensive clue score are as follows: After performing Min-Max normalization on the usage clue score, take 1 and subtract the normalized usage clue score to obtain the experience optimization score. Then, merge it with the development clue score according to the preset weight to generate the comprehensive clue score. The steps for updating the overall clue score are as follows: The time window is slid along the time axis according to the update step size, and the comprehensive clue score is recalculated after the sliding time window; an exponential time decay weight is introduced into the comprehensive clue score sequence within the historical time window; The combined clue scores of the current time window and the historical time windows are weighted and averaged to obtain the final combined clue score used for analysis. An exponential time decay weight is introduced into the comprehensive clue scoring sequence within the historical time window; The weighted average of the comprehensive clue scores from the current time window and historical time windows is used to obtain the final comprehensive clue score for analysis.

10. The method according to claim 1, characterized in that, The method for obtaining the position and score of each peak is as follows: For each score in the comprehensive clue score, a Gaussian kernel function is placed centered on that score value. The vertical coordinates of the Gaussian kernel function corresponding to all data points at each position on the horizontal axis are added together to obtain a superimposed total curve. The superimposed total curve is normalized to obtain the probability density function. On the probability density curve, all local maxima are found. These local maxima are the peaks in the data distribution, and the comprehensive clue score corresponding to the peak is the score of each peak. The optimization strategy is as follows: A threshold for urgency is preset, and the comprehensive clue score is divided into three intervals. If the peak score is not lower than the first threshold, it is judged as high urgency; if the peak score is between the first threshold and the second threshold, it is judged as medium urgency; and if the peak score is lower than the second threshold, it is judged as low urgency.

Images