Voice line intelligent recommendation method and system

Abstract

The invention relates to the field of voice communication, and discloses a voice line intelligent recommendation method, which comprises the following steps: collecting historical call data, real-time communication data, line quality data and user feedback data of a user; cleaning and standardizing the collected data to enable the data to meet model input requirements; extracting related characteristics of the user and the line; the invention also discloses a voice line intelligent recommendation system, and the system comprises a data collection module which is used for collecting the historical call data, the real-time communication data, the line quality data and the user feedback data of the user. Compared with the prior art, by means of the multi-dimensional data comprehensive scoring model, dynamic adjustment of the recommendation strategy and accurate feature attribution analysis, the defects of an existing recommendation system in the aspects of adapting to user demand changes, optimizing line selection in real time and improving recommendation accuracy are overcome, and therefore the self-adaptive capacity and the recommendation effect of the system are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of voice communication, and specifically to an intelligent voice line recommendation method and system. Background Art

[0002] In the modern communication field, the selection and management of voice lines have become an important link in multiple industries such as telecommunications operators, enterprise call centers, multinational enterprises, and online education platforms. With the continuous development of communication networks and the diversification of user needs, how to effectively manage voice lines, optimize line recommendations, and ensure call quality and reduce communication costs have become important challenges faced by major communication service providers and enterprises.

[0003] Traditional voice line selection methods usually rely on manual experience and decision-making models based on simple rules. Operators manually select appropriate lines according to certain criteria, such as line connection rate, delay, and network bandwidth. Although this method could meet the needs in the early stage, with the increase in communication demands and the complexity of the network environment, manual decision-making gradually exposes problems such as low efficiency, poor accuracy, and poor flexibility. Especially in large-scale business operation scenarios, manually selecting lines not only consumes a large amount of time and resources but is also easily affected by human factors, resulting in the inability to effectively improve the system efficiency.

[0004] In addition, in the existing technologies, many line recommendation systems still only focus on a single line quality index, such as connection rate, packet loss rate, and delay, without fully considering multi-dimensional data analysis. This simple quality assessment method cannot comprehensively reflect the actual performance of the line. Especially in the case of diverse user needs and complex network environments, traditional systems are difficult to dynamically adjust the recommendation strategy, resulting in poor line recommendation effects. For example, during peak hours or when the network fluctuates, the existing systems cannot re-evaluate and recommend lines according to the real-time network status, resulting in insufficient optimization of line selection.

[0005] At the same time, although some voice line recommendation systems have started to introduce machine learning technologies in recent years, most systems are still limited to rule-based models or shallow learning models. The disadvantages of these systems are that they fail to fully utilize multi-level feature information and do not have dynamic learning capabilities. Especially when facing complex user needs and ever-changing communication networks, these models fail to adjust strategies in a timely manner, resulting in inflexible and non-real-time recommendation results. In addition, existing systems often do not have an effective feedback mechanism to optimize real-time line recommendations based on user feedback data. Users' needs and experiences change over time, and many existing systems lack flexible adaptive capabilities and cannot adjust the recommendation strategy and line score in a timely manner according to user feedback, resulting in difficulties in improving recommendation accuracy and user satisfaction.

[0006] In summary, the voice line recommendation systems in the prior art mainly have the following problems: First, they rely on manual experience and static rules, lacking automation and intelligence; second, the evaluation of line quality is too simple, failing to comprehensively consider multi-dimensional feature information; third, the existing machine learning methods fail to achieve sufficient real-time performance and adaptability, and cannot be quickly adjusted according to the dynamically changing network environment and user requirements. Therefore, the prior art cannot meet the requirements of modern communication systems for efficient, accurate, and intelligent voice line recommendations, and there is an urgent need for a more flexible, efficient, and intelligent recommendation method to solve these problems. Summary of the Invention

[0007] In view of the deficiencies of the prior art, the present invention provides a voice line intelligent recommendation method and system, which solves the problem that traditional voice line recommendation systems cannot optimize line selection and recommendation strategies in real time under dynamic network environments and diverse user requirements.

[0008] To achieve the above objectives, the present invention is realized through the following technical solutions: A voice line intelligent recommendation method, including the following steps: Collect historical call data, real-time communication data, line quality data, and user feedback data of users; Clean and standardize the collected data to make it meet the model input requirements; Extract relevant features of users and lines, including but not limited to call duration, call frequency, and line quality, and perform necessary dimensionality reduction processing; Based on the feature data of the lines, use feature attribution analysis to calculate the contribution degree of each line to the connection rate, and obtain the contribution value of each line; Based on the feature data and contribution degree of the lines, use collaborative filtering, deep learning, or reinforcement learning algorithm models to score each line; According to the scoring results of each line, sort the lines and select the line with the highest score as the recommendation result; Dynamically adjust the recommendation strategy based on real-time feedback data and user requirements, so as to optimize the recommendation effect.

[0009] Preferably, the calculation formula for the contribution value is: where g i represents the contribution degree of line i to the change in the overall connection rate, b i represents the change value of line i to the overall connection rate of the product, represents the change value of the overall connection rate of all lines used by the outbound product, represents the change value of the overall connection rate of the product generated by line i, respectively represent the usage ratio and connection rate of line i in the current cycle respectively represent the usage ratio and connection rate of line i in the previous cycle, B represents the sum of the changes in connection rates for all lines, and ∈ is a smoothing factor.

[0010] Preferably, the feature attribution analysis step includes: When B > 0, if g i > 0, it means that line i helps to improve the overall connection rate of the product, and then the connection rate eigenvalue of this line will be rewarded in model training; When B > 0, if g i < 0, it means that line i is not conducive to improving the overall connection rate of the product, and then the connection rate eigenvalue of this line will be punished in model training; When B < 0, if g i > 0, it means that line i is not conducive to improving the overall connection rate of the product, and then the connection rate eigenvalue of this line will be punished in model training; When B < 0, if g i < 0, it means that line i is not conducive to improving the overall connection rate of the product, and then the connection rate eigenvalue of this line will be punished in model training.

[0011] Preferably, the step of scoring each line using a collaborative filtering, deep learning, or reinforcement learning algorithm model includes: Use user-based collaborative filtering algorithm or item-based collaborative filtering algorithm to calculate the similarity of users, and recommend voice lines according to the choices of similar users; Use a deep learning model to analyze the non-linear relationship between the call behavior of users and line characteristics, so as to make accurate line recommendations; Use a reinforcement learning algorithm to dynamically adjust the line recommendation strategy and update the model according to the instant feedback of users.

[0012] Preferably, the recommendation step includes: When a user requests, real-time obtain the current location of the user and the real-time information of the call destination; Input the real-time data into the trained recommendation model, and calculate the recommendation scores of each line; Sort all lines according to the recommendation scores, and select several lines with the highest scores for recommendation.

[0013] Preferably, the reinforcement learning algorithm uses a Q-learning model and realizes the optimization of the dynamic recommendation strategy through the following steps: At each recommendation, use the current state of the user as the input to generate a state s; For each candidate line, select action a, choose the line with the maximum expected return, and update the corresponding Q value; According to the user's real-time feedback data and environmental changes, adjust the recommendation strategy so that the recommendation results can be continuously optimized.

[0014] Preferably, the dynamic adjustment of the recommendation strategy is achieved through the following steps: During the recommendation process, adjust the scoring weights of the lines in real time according to the user's feedback data and environmental information; According to the real-time feedback data, including the connection rate and call quality, update the score of each line and adjust the output result of the recommendation model; Use a feedback mechanism based on reinforcement learning, combine the changes in user needs and network environment, and optimize the line recommendation strategy in real time to achieve optimal line selection and improve the overall recommendation effect; Dynamically update the feature model according to real-time data, and continuously optimize the model weights using historical data to further improve the accuracy and stability of system recommendations.

[0015] The present invention also provides a voice line intelligent recommendation system, including: A data collection module for collecting the user's historical call data, real-time communication data, line quality data, and user feedback data; A data preprocessing module for cleaning and standardizing the collected data; A feature engineering module for extracting relevant features of users and lines and performing necessary dimensionality reduction processing; A feature attribution analysis module for calculating the contribution degree of each line to the connection rate and performing reward or punishment processing on the line features; A recommendation algorithm module for scoring each line based on line features and contribution degrees using collaborative filtering, deep learning, or reinforcement learning algorithm models; A recommendation generation module for sorting the lines according to the scores and selecting the line with the highest score as the recommendation result; An adaptive optimization module for dynamically adjusting the recommendation strategy based on real-time feedback data and user needs to optimize the recommendation effect.

[0016] The present invention provides a voice line intelligent recommendation method and system. It has the following beneficial effects: 1. The present invention adopts a comprehensive scoring model based on multi-dimensional data, combines algorithms such as collaborative filtering, deep learning, and reinforcement learning to intelligently score the lines and optimize the recommendation strategy. Through a comprehensive analysis of the user's historical data, real-time communication data, and line quality data, the system can provide personalized and accurate line recommendations. Compared with the existing recommendation methods that rely on simple rules or static algorithms, the present invention solves the deficiencies of the existing recommendation systems that cannot dynamically adapt to the changes in user needs and network environment fluctuations, thereby greatly improving the accuracy of recommendations and user satisfaction.

[0017] 2. By dynamically adjusting the recommendation strategy and optimizing the line score based on real-time feedback data, the present invention can timely adjust the recommendation results when the user needs change or the network condition fluctuates. Compared with the relatively fixed recommendation mechanisms in the prior art, the present invention provides a flexible and intelligent adjustment mechanism to ensure that the system can optimize the recommendation results according to real-time feedback, solves the deficiencies of traditional recommendation systems that cannot flexibly respond to complex and changing user needs, and improves the adaptive ability and long-term benefits of the system.

[0018] 3. The present invention adopts an accurate mathematical model in the feature attribution analysis link, which can quantify the contribution degree of each line to the connection rate and sort the line priorities according to the contribution degree. The application of this technical solution makes the line recommendation more scientific and reasonable, avoids the limitations of simple judgments based only on user behavior or basic features in traditional methods, and through this technology, the system can deeply analyze the actual benefits and effects of each line, thereby improving the recommendation accuracy and solving the problems of low efficiency and unstable recommendation effects in previous recommendation systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 is a schematic diagram of the method flow of the present invention; Figure 2 is a schematic diagram of the system flow of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0020] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0021] Please refer to the attached Figure 1 , the embodiments of the present invention provide a method for intelligent recommendation of voice lines, including the following steps: S1. Collect the user's historical call data, real-time communication data, line quality data, and user feedback data; In the implementation process of the intelligent recommendation system for voice lines, data collection is the foundation of the system. By collecting multi-dimensional data from different sources, necessary information input can be provided for subsequent recommendation models. The quality of the data directly affects the accuracy of model training and recommendation results. Therefore, the design of this step must ensure the integrity and effectiveness of the data.

[0022] In this embodiment, the system collects the user's historical call data, real-time communication data, line quality data, and user feedback data through various means. These data sources include, but are not limited to, call records of telecommunications operators, user interaction records of online education platforms, call data of enterprise call centers, etc. In actual applications, the system obtains these data in real time by accessing different data source interfaces and performs preliminary preprocessing to provide support for subsequent feature engineering and model training.

[0023] Specifically, the user's historical call data includes information such as the user's call duration, call frequency, call time period, call destination, etc. These information can reflect the user's call behavior and preferences. For example, some users may make calls frequently during specific time periods or frequently call users in certain regions. By collecting this data, the system can understand the user's call habits and needs and provide a basis for subsequent line recommendations.

[0024] The real-time communication data includes the network environment information of the current user, such as the network quality, latency, packet loss rate, etc. of the current call. These real-time data can reflect the quality status of the current line and further help the system make more accurate judgments during recommendations. For example, when the system detects that the latency of a certain line is high, it can consider excluding this line from the recommendation results.

[0025] In addition, the line quality data mainly includes the physical performance indicators of the line, such as packet loss rate, latency, silent rate, etc. By collecting these data, the system can evaluate the quality of different lines and provide more accurate line evaluation information for model training. Especially in the scenario of multi-line selection, the line quality data is crucial and affects the accuracy of the final recommendation results.

[0026] The user feedback data refers to the feedback information provided by the user during the use process, such as the user's satisfaction with the recommended line, the connection rate of the line, etc. These data can help the system optimize the recommendation model, achieve dynamic adjustment and self-improvement. The system can continuously adjust the recommendation strategy according to the real-time changes of user feedback to optimize the line recommendation effect.

[0027] In practical applications, the data collection module synchronizes data through interfaces with multiple data sources to ensure real-time updates and accuracy of the data. For different types of data, the system will perform appropriate processing according to the characteristics of the data to ensure that all data can be used for subsequent analysis and modeling.

[0028] For example, the user's historical call data may include information such as call duration and call frequency. These are not just single data items but are time-sequential. To effectively analyze the data, the system will group and statistically analyze these data by time period, calculating the average call duration and call frequency within each time period to provide a basis for subsequent feature extraction.

[0029] In one possible implementation, the system adopts different data collection strategies to meet the requirements in different scenarios. For example, in the scenario of a telecommunications operator, the system can directly obtain the user's call records and line quality information by accessing the telecommunications network data interface; in the scenario of an enterprise call center, the system may collect the user's real-time call data and feedback information by integrating the internal call system.

[0030] As an option, the system can initially screen the collected data through artificial intelligence algorithms to filter out outliers or missing values to ensure that the data entering subsequent processing is valid. For missing data items, the system can adopt certain filling strategies, such as reasonably inferring and filling based on the user's historical call behavior, or using data from surrounding time periods for inference and filling. This processing method can effectively avoid the negative impact of data missing on the recommendation results.

[0031] In specific implementation, the system can collect and update data from different sources in real time by configuring different data collection interfaces. By collaborating with telecommunications operators, enterprise systems, etc., the system can quickly obtain real-time data, thereby achieving more accurate line recommendations.

[0032] These data not only have reference value within the current period but can also be used for retrospective analysis of historical data, providing rich training data for subsequent model training. For example, historical data can help the system understand call habits and the change rules of line quality in different time periods, providing support for subsequent feature engineering and model training.

[0033] In short, data collection is the first step of the voice line intelligent recommendation system, directly affecting the accuracy of the system's recommendation effect. By comprehensively and real-time collecting data from different sources, the system can ensure strong support for subsequent model training and optimization of recommendation strategies, providing sufficient data guarantee for the final line recommendation results.

[0034] S2. Clean and standardize the collected data to meet the model input requirements; In the intelligent voice line recommendation system, data cleaning and standardization are important steps to ensure data quality and model accuracy. By cleaning and standardizing the collected raw data, it is possible to effectively remove noise, outliers, and invalid data in the data, providing accurate data input for subsequent feature extraction and model training. Therefore, this step is closely linked to the aforementioned data collection step to ensure the consistency and reliability of the data from collection to the final model input.

[0035] In this embodiment, first, clean the collected user historical call data, real-time communication data, line quality data, and user feedback data. The data cleaning steps include removing duplicate data, filling in missing data, and processing abnormal data. Specifically, the system automatically cleans the data by analyzing problems such as missing values, duplicate values, and invalid values in the data. For missing values, the system can fill them based on the user's historical data or use other common interpolation methods (such as mean filling, interpolation method, etc.) to ensure the integrity of the data. For duplicate values, the system detects and removes duplicate data by comparing the unique identifiers of the records (such as user ID, timestamp, etc.).

[0036] For example, when the system detects duplicate call records in a user's historical call data, it will identify the duplicates based on the timestamp or call ID and automatically delete one of the records, retaining only the unique data entry. In a possible implementation, for some cases where line quality data is missing, the system can infer and fill in the missing data based on the quality data of other lines, or use mean filling, interpolation methods to fill in the missing quality metric values, such as packet loss rate and latency.

[0037] After completing the data cleaning, the system will standardize all numerical data. The purpose of standardization is to convert data with different dimensions to a unified scale range, so as to ensure that the subsequent model has the same weight and influence on different features. The standardization operation usually uses the following formula: where x ′ represents the standardized value, x is the original data, μ is the mean of this data set, and σ is the standard deviation of the data set. In this way, the system can convert data with different dimensions (such as call duration, latency, packet loss rate, etc.) into standardized values, eliminating the bias caused by different dimensions.

[0038] Generally, the standardization process can make the model training unaffected by the data dimension, ensuring that all data plays an equal role in the model training. As an option, if some data causes information loss or abnormal fluctuations after standardization, the system can use other normalization methods, such as min-max normalization, etc., to appropriately process the data.

[0039] Specifically, for data containing categorical variables (such as user location, call destination, etc.), the system will use methods such as one-hot encoding to convert categorical variables into numerical variables. This process ensures that all input data can be adapted to subsequent machine learning models, especially algorithm models with strict requirements for data input formats, such as neural networks.

[0040] In the implementation process, for some data with temporal properties, such as call duration, call frequency, etc., the system not only standardizes these data but also processes them based on time series characteristics. For example, for user historical call data, the system may consider the call pattern changes in each time period and use a sliding window method to smooth the data, thereby reducing the volatility in the data and retaining important trend information.

[0041] After the data is standardized, the system will also perform a quality check on the processed data to ensure that all data meets the input requirements. If some data still has anomalies or does not meet expectations, the system will perform cleaning processing again until the data set fully meets the model input requirements.

[0042] After data cleaning and standardization are completed, the processed data will enter the subsequent feature extraction step. In this step, the system extracts features related to user needs and line quality based on the cleaned data, providing the basic input for subsequent model training. This processing process ensures that the data not only has good quality but also provides appropriate feature expressions for model training.

[0043] For example, call duration data may need to be weighted according to the user's active period to highlight the influence of highly active users during feature extraction. For line quality data, the system may further refine the regionalized data according to the network environment differences in different regions to ensure that data in different regions can be properly processed and analyzed.

[0044] In a possible implementation, for the processing of user feedback data, the system uses natural language processing (NLP) technology to perform sentiment analysis and semantic extraction on the user's feedback content, converting unstructured feedback information into analyzable structured data. This processing method can better utilize the user's direct feedback information to further optimize the recommendation strategy.

[0045] In summary, data cleaning and standardization are one of the key steps in the intelligent recommendation system for voice lines. Through the refined processing of the original data, the system can ensure the high quality, accuracy, and consistency of the data, thereby providing a reliable basis for subsequent feature extraction, model training, and recommendation strategy optimization.

[0046] S3. Extract relevant features of users and lines, including but not limited to call duration, call frequency, and line quality, and perform necessary dimensionality reduction processing; In the intelligent recommendation system for voice lines, feature extraction is a crucial step, which provides high-quality input data for subsequent model training. By extracting relevant features of users and lines, the system can capture important information affecting line recommendations, thereby providing an accurate basis for the recommendation model. Feature extraction not only helps improve the accuracy of the model but also reduces the computational complexity through dimensionality reduction processing, enhancing the recommendation efficiency. Therefore, this step is a key link after data cleaning and standardization processing and is directly related to the effectiveness of subsequent steps.

[0047] In this embodiment, feature extraction mainly includes the extraction of user behavior features and line quality features. First, the system extracts a series of behavior features from the user's historical call data, such as call duration, call frequency, call time period, and call destination. By analyzing these features, the system can understand the user's call habits and preferences. For example, some users may prefer to make calls in the morning or late at night, and some users may be more inclined to call users in certain specific regions. Such information can provide valuable clues for the recommendation algorithm.

[0048] Specifically, call duration is one of the most common and important indicators in user behavior features and is usually used to measure the user's call activity. The system can calculate the average call duration of each user within a specific time period or calculate the total call duration of each user over a past period. This feature can reflect the user's call frequency and importance, helping the system identify high-frequency users and thus giving priority to recommending lines with better network quality.

[0049] In addition, call frequency is also an important feature, which can reflect the intensity of the user's call demand. The system will analyze the number of calls made by each user over a past period to obtain the user's call frequency feature.

[0050] As an option, if some users have a high call frequency within a specific time period, this data will help optimize line recommendations in subsequent steps. Especially during high-load periods, the system can preferentially select lines with higher quality for recommendation.

[0051] Next, the system will also extract the quality characteristics of the lines, which usually include network quality metrics such as latency, packet loss rate, and silence rate. The quality metrics of each line can directly affect the user's call experience. Therefore, these characteristics are crucial for the recommendation system. By comprehensively evaluating these metrics, the system can understand the performance of each line under different network conditions and thus recommend the best line for the user.

[0052] In a possible implementation, to ensure the accuracy of feature extraction, the system will perform time series analysis on the line quality data. For example, the system will classify the latency and packet loss rate of the line by time period, calculate the average value of each time period, and then obtain the change trend of the line quality in different time periods. This approach helps to identify lines with poor quality during certain specific periods and avoid selecting these lines in the recommendation.

[0053] After extracting the relevant features, the system will perform necessary dimensionality reduction on these data. Generally, high-dimensional data may contain redundant or irrelevant information, which not only increases the computational complexity of model training but also may lead to overfitting problems. Therefore, the system uses dimensionality reduction techniques (such as principal component analysis, PCA) to reduce the data dimension. The purpose of dimensionality reduction is to reduce the computational complexity and improve the robustness of the model while retaining important information.

[0054] Specifically, during the dimensionality reduction process, the system will calculate the correlation between each feature and screen out the principal components that best represent the user's behavior and line quality. These principal components will become the core features of the subsequent model input, ensuring that the model can make full use of meaningful features during training and thus reducing unnecessary noise.

[0055] As an option, the system can also combine other dimensionality reduction methods, such as factor analysis or t-SNE, etc., to further improve the accuracy and efficiency of feature extraction.

[0056] In practical applications, choosing the appropriate dimensionality reduction technique will help to better handle complex high-dimensional data and optimize the expression form of features.

[0057] Through this series of feature extraction and dimensionality reduction processes, the system can extract the most representative features from a large amount of raw data, greatly improving the training efficiency and prediction accuracy of the model. At the same time, the feature extraction and dimensionality reduction also ensure that the subsequent model evaluation and recommendation steps have efficient data support, laying a foundation for the successful operation of the intelligent recommendation system.

[0058] In summary, the feature extraction step plays a crucial role in the intelligent voice line recommendation system. By extracting key features from user behavior and line quality data and combining dimensionality reduction processing, the system can reduce data noise, optimize recommendation inputs, and improve the accuracy and reliability of recommendation results.

[0059] S4. Based on the feature data of the line, use feature attribution analysis to calculate the contribution degree of each line to the connection rate, and obtain the contribution value of each line; In the intelligent voice line recommendation system, feature attribution analysis is an important step to deeply evaluate the performance of each line. By calculating the contribution degree of each line to the connection rate, the system can quantify the effects of different lines and preferentially select the lines that contribute more to the improvement of the connection rate in the subsequent recommendation process. Feature attribution analysis provides decision support for the model to ensure that the recommendation results can maximize the overall effect of the product.

[0060] In this embodiment, the system first calculates the contribution degree of each line to the connection rate according to the feature data of each line. The core of this step is to use the feature attribution analysis method, which can analyze the usage amount, connection rate, and other relevant feature data of the line to evaluate the impact of each line on the change of the overall connection rate of the product. During the calculation process, the system needs to evaluate the influence of each line according to the change of the connection rate in different periods.

[0061] In this process, the system calculates through the following formula: where g i represents the contribution degree of line i to the change of the overall connection rate, b i represents the change value of line i to the overall connection rate of the product, represents the change value of the overall connection rate of all lines used by the outbound product, represents the change value of the overall connection rate of the product generated by line i, respectively represent the usage proportion and connection rate of line i in the current period, respectively represent the usage proportion and connection rate of line i in the previous period, B represents the sum of the changes in the connection rate of all lines, and ∈ is a smoothing factor.

[0062] When B≠0, the calculation result of ∈ is extremely small and does not change the calculation logic of the original formula; When B = 0, the denominator of ∈ is not zero, ensuring the feasibility of the calculation, and ∈ is small enough not to affect the overall trend of the calculus calculation.

[0063] To prevent the contribution degree g from being caused by the too small or even zero historical usage amount B of the line during the calculation of the contribution value iThe value is abnormally high, so the system introduces a smoothing factor ∈ in the contribution value formula. The smoothing factor ∈ is a dynamic adjustment parameter, and its value is determined based on the distribution characteristics of the line's historical usage B, ensuring that the denominator B+∈ is always greater than zero to avoid abnormal values.

[0064] The design principle of the smoothing factor is: when the value of B is large, the influence of ∈ on the contribution value tends to be ineffective, and the contribution value is mainly affected by the historical usage B; when B is close to 0, the system limits g by ∈ i Over-amplification of the formula is maintained stable; for cold start routes with B=0, the system makes cold start recommendations through collaborative filtering or business strategies, and then introduces the contribution value model after the route is actually used.

[0065] The value of ∈ can be dynamically adjusted, for example: ∈ = β·std(B) Among them, β is a hyperparameter and std(B) represents the standard deviation of B in historical data to ensure that ∈ is adaptively adjusted.

[0066] Indicates the usage ratio of a line in a cycle. There are two specific measurement methods: Based on the proportion of call duration: the proportion of the line's cumulative call duration in the period, that is: Based on the proportion of call times: the proportion of calls made by the line in the period, that is: In different application scenarios, different measurement methods can be selected: If you are concerned about the impact of user call duration on the overall connection rate, you should use the call duration ratio as the

[0067] If you are concerned about the impact of the call frequency of different lines on the overall connection rate, you should use the call frequency ratio as the

[0068] In some cases, the two can be considered together, that is, different weights are given to call duration and call number, for example: In general, the contribution of each line g i It will be adjusted according to the changes in its connection rate and usage ratio in different cycles. Through this formula, the system can calculate the proportion of each line in the change of connection rate and judge the contribution value of the line accordingly. i The larger the value, the stronger the effect of the line on improving the overall connection rate of the product, and the system will give priority to these lines in subsequent recommendation strategies.

[0069] The value of the weight α is determined based on data statistical analysis and business requirements. The main basis is the correlation between call duration, call frequency and call connection rate, as well as the actual requirements of the business scenario.

[0070] 1. Business scenarios with call duration as the main influencing factor In this scenario, historical data statistical analysis shows that there is a high correlation between call duration and call connection rate. For example, the Pearson correlation coefficient is usually above 0.8, indicating a high positive correlation between the two. Therefore, in order to make full use of this high correlation, the weight α takes a value of 0.7 - 0.9 in the model calculation.

[0071] Reason: Since users with longer call durations are usually more willing to answer calls, call duration is an important predictor of call connection rate. A high correlation (>0.8) indicates that call duration can well explain the change of call connection rate and is suitable as the main factor.

[0072] Through experimental analysis, when the value range of different α values is between 0.7 - 0.9, the calculation results can better fit the business requirements.

[0073] 2. Business scenarios with call frequency as the main influencing factor In this scenario, the impact of call frequency on call connection rate is relatively weak. Historical data statistical analysis shows that the Pearson correlation coefficient between call frequency and call connection rate is usually less than 0.4, indicating a low correlation between the two. Therefore, in the model calculation, the weight α takes a value of 0.3 - 0.5.

[0074] Reason: Since users with a high number of calls are not necessarily more willing to answer calls, the explanatory power of call frequency is weak. A low correlation (<0.4) means that call frequency cannot well predict the call connection rate, and the weight should not be too high. The α value between 0.3 - 0.5 can better balance the calculation results, making it consider the impact of call frequency without over - relying on it.

[0075] 3. Hybrid business scenarios In the hybrid business scenario, that is, when both call duration and call frequency jointly affect the call connection rate, the weight α needs to be dynamically adjusted according to specific circumstances to ensure the optimal calculation results.

[0076] Method: Machine learning regression analysis: Learn historical data through regression models (such as linear regression, Lasso regression, decision tree regression, etc.) to find the optimal weight allocation of call duration and call frequency.

[0077] Dynamic optimization: According to data changes, use adaptive algorithms to adjust α, for example, find the optimal solution through gradient descent or Bayesian optimization.

[0078] Regular adjustment: Based on the changes in business requirements and data distribution, retrain the model at regular intervals to ensure the optimality of α.

[0079] Summary: When the call duration is the main influencing factor, α takes a value of 0.7 - 0.9 (because of its high correlation with the connection rate). When the number of calls is the main influencing factor, α takes a value of 0.3 - 0.5 (because of its low correlation with the connection rate). In a mixed business scenario, α is dynamically adjusted through machine learning to ensure the optimal calculation result.

[0080] When B > 0, if g i > 0, it means that line i helps to improve the overall connection rate of the product, and then the connection rate eigenvalue of this line will be rewarded in the model training; When B > 0, if g i < 0, it means that line i is not conducive to improving the overall connection rate of the product, and then the connection rate eigenvalue of this line will be punished in the model training; When B < 0, if g i > 0, it means that line i is not conducive to improving the overall connection rate of the product, and then the connection rate eigenvalue of this line will be punished in the model training; When B < 0, if g i < 0, it means that line i is not conducive to improving the overall connection rate of the product, and then the connection rate eigenvalue of this line will be punished in the model training.

[0081] As an option, the system can also further adjust the contribution calculation according to different business requirements. For example, in certain specific scenarios, the system may weight the quality characteristics of the lines. If a certain line is of poor quality but has a high usage frequency, the system may give a corresponding adjusted weight to its contribution to avoid recommending low-quality lines to users during the recommendation process.

[0082] Specifically, when calculating the contribution of each line, the system will conduct attribution analysis based on the usage proportion and connection rate difference of the line, combined with historical cycle data. This process ensures that the system can identify the optimal line in a dynamically changing network environment in real time, especially when the user's needs change rapidly, and can quickly adapt and make adjustments. In addition, when calculating the contribution, the interaction between lines also needs to be considered to avoid simple summation based on independent calculations.

[0083] In practical applications, voice lines may share bandwidth or hardware resources, resulting in mutual influence between different lines. Therefore, when calculating the overall contribution, the contributions of each line cannot be simply added directly, but a non-linear influence factor w i needs to be introduced for weighting: Among them, w i represents the dynamic impact weight of the line, which can be determined by factors such as the bandwidth occupancy of the line, the network load situation, and historical stability, so as to correct the contribution degree of each line and avoid errors caused by simple summation. Its calculation formula is as follows: w i = ρ1·S bandwidth,i + ρ2·S load,i + ρ3·S stability,i ; Among them: S bandwidth,i —— Bandwidth occupancy score: S load,i —— Network load score: S stability,i —— Historical stability score: Among them: ρ1, ρ2, ρ3 are adjustable hyperparameters that control the influence of different factors on the weight, and are used to control the contributions of packet loss rate, latency, and jitter to stability.

[0084] In addition, in order to improve the accuracy of contribution degree calculation, considering the synergy relationship between lines and correcting the overall contribution degree B, the system can further introduce a synergy coefficient C ij , which is used to measure the influence between line i and line j: Among them: C ij > 0 indicates that line i and line j share resources and may affect each other's quality. For example, when sharing bandwidth or network resources, a change in the load of one line may affect the quality of the other line; C ij < 0 indicates that line i and line j may form a complementary relationship. For example, under load balancing of different paths, the two lines may provide complementary support in some situations, improving the stability of the system; C ij = 0 indicates that the two do not affect each other.

[0085] This coefficient can be dynamically calculated based on historical network traffic data, packet loss rate and other indicators through machine learning models or correlation analysis methods, so as to more accurately reflect the true contribution degree of the line. These methods can accurately reflect the mutual influence between lines according to the actual network state and traffic changes, so as to make a more accurate correction to the overall contribution degree.

[0086] The correlation analysis method can judge whether two lines share resources or have a complementary relationship by calculating the correlation between lines.

[0087] Use the Pearson correlation coefficient or Spearman rank correlation coefficient to measure the correlation of quality indicators such as traffic and latency between lines: Among them, Cov(r i , r j ) is the covariance of the connection rates of line i and line j, and are the corresponding standard deviations.

[0088] If the correlation of C ij is positive, there may be a relationship of resource sharing between the two lines; if it is negative, they may have a complementary relationship.

[0089] Machine learning models (such as regression analysis, neural networks, etc.) can predict the synergy coefficient C ij based on a large amount of historical data and dynamically adjust it according to the real-time network environment.

[0090] Use machine learning models, such as regression analysis or neural networks, to predict the synergy coefficient C ij : Input: data such as historical traffic, bandwidth, packet loss rate, delay, etc.

[0091] Output: Predicted C ij .

[0092] The deep learning model can dynamically adjust the synergy coefficient according to historical data and real-time feedback to ensure the flexibility and accuracy of the calculation.

[0093] In a possible implementation, the system can use a reinforcement learning algorithm to dynamically adjust the contribution calculation. For example, when the system detects an abnormal fluctuation in the connection rate of a certain line during a specific period, it can dynamically update the estimation of the contribution according to the reinforcement learning algorithm to cope with the impact brought by the line quality fluctuation. This method can improve the robustness of the system in an unstable network environment and ensure more accurate recommendation results.

[0094] The reward function of the reinforcement learning model can be improved so that it not only optimizes the improvement of the connection rate but also considers the competition relationship between lines: Reward - improvement of connection rate - λ · competition loss, where λ is a trade-off factor; The calculation method of the competition loss can be based on indicators such as line load, packet loss rate, and bandwidth occupancy to reflect whether the use of a certain line has a negative impact on other lines.

[0095] For the evaluation of multiple lines, the system may adopt techniques such as ensemble learning to further weight and fuse the contribution degrees of each line. For example, the system can combine the calculation results of the contribution degrees of multiple models (such as collaborative filtering, neural network, decision tree, etc.) to comprehensively evaluate the overall impact of each line, so as to obtain a more robust contribution degree evaluation and reduce the bias brought by single-model evaluation. In this way, the final recommendation is not only based on a single calculation method, but combines the results of multiple algorithms, improving the generalization ability of the model.

[0096] By calculating the contribution degree of each line, the feature attribution analysis in this embodiment provides key input data for the model. These data will be used in the subsequent scoring and recommendation processes to ensure that the finally recommended lines not only meet the user's needs, but also can maximize the connection rate and overall effect of the product.

[0097] To more accurately identify the line contribution degree, in the feature attribution analysis, it is possible to: Introduce time series models (such as LSTM, Prophet) to predict line stability; Consider the dependency relationship between lines (such as graph neural network GNN modeling bandwidth sharing); Combine real-time monitoring data for dynamic weight adjustment.

[0098] S5. Based on the feature data and contribution degree of the line, use collaborative filtering, deep learning or reinforcement learning algorithm models to score each line; In the scoring session, the system calculates the scoring values of each line according to the line feature data and contribution degree calculation results obtained in the previous steps, in combination with historical user behaviors and line quality indicators. The scoring model needs to adapt to different scenarios to ensure the adaptability and stability of the recommendation. For this purpose, this embodiment introduces collaborative filtering, deep learning and reinforcement learning methods to model the similarity between lines, the non-linear relationship of features and dynamic user feedback respectively, and optimize the line scoring from different dimensions.

[0099] In this embodiment, the collaborative filtering algorithm is used for scoring calculation. Generally, collaborative filtering is divided into two methods: user-based and item-based. User-based collaborative filtering calculates the similarity between different users, finds users with similar behavior patterns, and predicts the score of a target user for a certain line. The similarity between users can be calculated by the Pearson correlation coefficient: where r u,i represents the score of user u for line i, is the average score of user u, L is the set of lines rated by the user, Sim(u, v) represents the similarity between user u and user v, and r v,i represents the score value of user v for line i, Represents the average rating of user v. User-based collaborative filtering is applicable to scenarios with rich historical behavior data and can effectively capture user preference patterns.

[0100] After calculating the user similarity, the predicted rating of target user u for line i is calculated as follows: As an option, item-based collaborative filtering can be used for rating calculation. This method analyzes the similarity between lines and infers the correlation relationship of line ratings. The similarity between lines is calculated by cosine similarity: where U represents the set of all users, r u,i represents the rating of user u for line i, Sim(i,j) represents the similarity between line i and line j, and r u,j represents the rating value of user u for line j. Item-based collaborative filtering is applicable to scenarios where the line quality is relatively stable and can assist in optimizing line ratings.

[0101] After calculating the similarity between lines, the predicted rating S item (u,i) of target user u for line i is calculated as follows: Finally, the final collaborative filtering rating is obtained by weighted summation of the user-based and item-based rating calculation results: where λ1 and λ2 are weight parameters, which are optimized according to the data distribution.

[0102] In some embodiments, the system uses a deep learning model for rating calculation. Deep learning can model the non-linear relationship between line features and user preferences and improve the accuracy of ratings. Specifically, the system uses a deep neural network (DNN) to process the multi-dimensional features of lines, and the inputs include but are not limited to: Line quality features (packet loss rate, latency, jitter); The contribution degree of the line (impact on the overall connection rate); The historical communication data of the user; The stability index of the line in different scenarios.

[0103] The basic calculation method of the deep neural network is as follows: h (l) = σ(W (l) h (l-1) + b (l) ), where h (l) is the output of the neurons in the l-th layer, W (l) is the weight matrix of this layer, and b (l)is the bias term, and σ is the activation function. Generally, the system uses ReLU as the activation function to improve the training convergence speed.

[0104] The final score S DL is obtained from the output of the last layer of the neural network: S DL = f DNN (X), where X is the input feature vector and f DNN is the function expression of the neural network.

[0105] In a possible implementation, the score calculation is optimized using reinforcement learning. Reinforcement learning can dynamically adjust the route score based on user feedback to improve the real-time performance and adaptability of the score calculation. Specifically, the system uses Q-learning for score optimization, and the core formula is as follows: where Q(s,a) is the score value for performing action a in state s, r is the immediate reward, representing the user's real-time feedback on the route, α is the learning rate, controlling the score update rate, γ is the discount factor, affecting the weight of future rewards, is the optimal score estimate value for the next state. Reinforcement learning is suitable for scenarios where user needs change continuously and can effectively improve the dynamic optimization ability of the score.

[0106] The final score calculation of reinforcement learning is as follows: S RL = Q(s,a); As an option, the score calculation can adopt a multi-algorithm fusion method. Specifically, the system performs a weighted combination of collaborative filtering, deep learning, and reinforcement learning scores to improve the stability and adaptability of the score. The final score calculation is as follows: S = w1S CF + w2S DL + w3S RL ; where S CF is the collaborative filtering score, S DL is the deep learning score, S RL is the reinforcement learning score, and w1, w2, w3 are the weight coefficients of the corresponding features, which are adjusted according to the data characteristics.

[0107] Generally: When the data is less, the collaborative filtering weight is higher (w1 > w2, w3).

[0108] After data accumulation, deep learning dominates (w2 > w1, w3).

[0109] When the user's behavior changes, the proportion of reinforcement learning increases (w3 increases).

[0110] In some embodiments, the system dynamically adjusts the priority of the lines based on the scoring calculation results. The lines with higher scores have their priorities increased, while the lines with lower scores have their priorities decreased to ensure the accuracy of recommendations and the optimization effect.

[0111] S6. Sort the lines according to the scoring results of each line, and select the line with the highest score as the recommended result; in the intelligent voice line recommendation system, the scoring and sorting steps are the key links to ensure that the system can recommend the optimal line for the user. By scoring each line, the system can quantify the matching degree between the quality of the line and the user's needs, so as to sort the priorities of different lines. In this process, the system combines the results of the foregoing feature extraction, feature attribution analysis, and algorithm scoring, and recommends the line that best meets the user's needs to the user.

[0112] In this embodiment, the system first sorts all candidate lines according to the scoring results of each line. Specifically, the system sorts the recommended scores of each line according to the output scores of collaborative filtering, deep learning, or reinforcement learning algorithms. These scores reflect the adaptability, performance, and matching degree of each line to the user's needs. During the sorting process, the system further adjusts the sorting strategy according to the user's personalized needs, such as call duration, call frequency, line quality, etc., to ensure that the finally recommended line can meet the user's expectations.

[0113] Generally, the sorting of the scoring results is carried out in descending order. The system starts from the line with the highest score and gradually screens out the line that best meets the current user's needs according to the score. For each user, the system dynamically calculates a comprehensive score based on historical call data, real-time communication data, and user feedback data to reflect the fitness of the line to the user.

[0114] Specifically, assume that the system calculates the score S of each line based on the foregoing algorithms (such as collaborative filtering, deep learning, or reinforcement learning) i , where S i is the score value of line i, indicating the adaptability of this line to the current user's needs. The system sorts the scores of all lines to obtain the sorting results S1, S2,..., S n , where S1 is the line with the highest score, and S n is the line with the lowest score. The sorted lines will be used in the next recommendation process.

[0115] In some cases, as an option, the system may also set a threshold and only recommend routes with scores higher than a certain set threshold. This threshold can be dynamically adjusted based on the system's historical recommendation effectiveness or user feedback to ensure the quality of the recommendation results. For example, if the score of a certain route is lower than the set threshold, the system can consider the recommendation suitability of this route to be poor and choose not to recommend it. This method ensures the high quality of the recommendation results and avoids recommending routes that do not meet the user's needs to the user.

[0116] In a possible implementation, the system can also weight the sorting strategy according to different business scenarios or the needs of specific users. For example, if a user has high requirements for the quality of certain routes, the system can increase the scoring weight of route quality and give priority to recommending routes with higher quality scores. On the contrary, if the user pays more attention to other factors such as price or call duration, the system can adjust the scoring weight of these factors to provide recommendation results that better meet the user's personalized needs.

[0117] Specifically, assuming the user's demand priority is: quality > call duration > price, the system will weight the scores of each feature according to this priority. The final score S of each route i will be calculated by weighting the quality score, call duration score, and price score of the route. The weighting formula is as follows: S i = w1·Q i + w2·T i + w3·P i where: S i is the final score of route i; Q i is the quality score of route i; T i is the call duration score of route i; P i is the price score of route i; w1, w2, w3 are the weight coefficients of the corresponding features, and w1 + w2 + w3 = 1.

[0118] Through this weighting strategy, the system can adjust the sorting of the recommendation results according to the preferences of different users, ensuring a high degree of fit between the recommendation results and the user's needs.

[0119] In a possible implementation, the system can also introduce the user's historical behavior data to adjust the scoring and sorting strategy. For example, for users who have frequently selected low-latency routes, the system can increase the weight of latency-related features in the scoring, so that the system gives priority to recommending routes with lower latency when sorting. This method helps to optimize personalized recommendations and makes the recommendation results more accurately conform to the user's historical preferences.

[0120] In summary, through the sorting process based on the scoring results, the system can provide the most suitable voice line for users. By sorting the scoring results of collaborative filtering, deep learning, or reinforcement learning algorithms and adjusting the sorting strategy according to the personalized needs of users, the system can effectively improve the accuracy of the recommendation results and user satisfaction. In addition, based on the strategy of weighted scoring and dynamic adjustment, the system can optimize according to the needs and preferences of different users, so as to provide more accurate and personalized recommendation results.

[0121] S7. Dynamically adjust the recommendation strategy based on real-time feedback data and user needs to optimize the recommendation effect.

[0122] In the intelligent voice line recommendation system, the dynamic adjustment of the recommendation strategy is a key link to ensure that the recommendation system can respond to changes in user needs and network conditions in real time. Through the acquisition and analysis of real-time feedback data, the system can continuously optimize the recommendation strategy and improve the accuracy and real-time adaptability of the recommendation. Especially in the case of changing user needs and complex network environments, being able to adjust the recommendation strategy in a timely manner will significantly improve the efficiency of the system and user satisfaction.

[0123] In this embodiment, the system dynamically adjusts the recommendation strategy according to the real-time feedback data of the user and the changes in the current network environment. Specifically, after the user selects a line, the system will collect the user's feedback information and adjust the feature weights and scoring mechanism in the recommendation algorithm according to the feedback. This adjustment not only depends on the user's direct feedback on the recommended line, but also includes a comprehensive analysis of information such as user behavior, call quality, and line usage.

[0124] Generally, the real-time feedback data includes the user's satisfaction with the recommended line, feedback on call quality, line connection rate, etc. The system will update the score of each line according to this feedback data to ensure that the recommendation algorithm reflects the latest needs of the user. For example, if the user gives poor feedback on the quality of a certain line, the system will reduce the score of this line and give priority to recommending other lines that better meet the user's needs.

[0125] As an option, the system can also make dynamic adjustments in combination with the user's historical preferences. For example, if a certain user has a long-term preference for high-quality lines, the system will further strengthen the recommendation weight of high-quality lines according to the user's historical behavior data to ensure that the recommendation results better meet the personal needs of the user. This adjustment based on personalized historical behavior can significantly improve the accuracy of the recommendation and enhance the user experience.

[0126] Specifically, based on the real-time feedback data, the system can use the following dynamic adjustment formula: S i (t) = S i (t - 1)+α·ΔS i (t); Among them, S i (t) represents the score of line i at time t; S i (t - 1) is the score of line i at the previous time point t - 1, and ΔS i (t) is the score change based on user feedback, reflecting the impact of user feedback on the score of this line; α is an adjustment factor that controls the sensitivity of the score adjustment to user feedback.

[0127] In a possible implementation, the system assigns different weights to different feedback data according to different feedback types (such as call quality, connection rate, etc.). For example, if a certain line has a high latency and the user feedbacks poor call quality, the system will pay more attention to the latency data and make a relatively large adjustment to the line score. For other types of feedback, such as price or call time selection, the score adjustment range is relatively small.

[0128] As another option, the system can adopt a reinforcement learning feedback mechanism to achieve dynamic adjustment. The reinforcement learning algorithm can better handle complex demand changes by adjusting the recommendation strategy according to the user's immediate feedback. After recommending each line, the system records the user's feedback data in real time and adjusts the line score through the reinforcement learning model. The Q - learning model can be applied in this process to dynamically update the score of each line. The specific update formula is: Among them, Q(s,a) is the score value of executing action a in state s, r is the immediate reward, representing the user's real - time feedback on the line, α is the learning rate that controls the score update rate, γ is the discount factor that affects the weight of future rewards, is the optimal score estimate value for the next state.

[0129] This reinforcement learning mechanism can automatically adjust the recommendation strategy according to the changes in user behavior, making the system more flexible and efficient in a constantly changing environment.

[0130] In specific implementation, the dynamic adjustment of the system is not only based on immediate feedback but also takes into account long - term user behavior changes. For example, if a certain user has been inclined to choose lines with lower latency for a long time, the system will gradually increase the recommendation intensity of lines with lower latency, even if the user's immediate feedback may not have a great impact on the score of this line. This adjustment based on long - term behavior makes the recommendation results more stable and personalized.

[0131] In summary, through the dynamic adjustment of real-time feedback data and user requirements, the system can continuously optimize the recommendation strategy to ensure that the recommendation results are highly consistent with user needs. This process, through reinforcement learning, scoring mechanism adjustment, and historical data analysis, ensures the flexibility, accuracy, and real-time nature of the recommendation system, greatly improving the user experience and enhancing the system's adaptive ability.

[0132] The intelligent voice line recommendation system described below can be correspondingly referred to with the intelligent voice line recommendation method described above.

[0133] An intelligent voice line recommendation system, comprising: A data collection module, configured to collect historical call data, real-time communication data, line quality data, and user feedback data of users; A data preprocessing module, configured to clean and standardize the collected data; A feature engineering module, configured to extract relevant features of users and lines, and perform necessary dimensionality reduction processing; A feature attribution analysis module, configured to calculate the contribution degree of each line to the connection rate, and perform reward or punishment processing on the line features; A recommendation algorithm module, configured to score each line based on line features and contribution degrees, using collaborative filtering, deep learning, or reinforcement learning algorithm models; A recommendation generation module, configured to sort the lines according to the scores, and select the line with the highest score as the recommendation result; An adaptive optimization module, configured to dynamically adjust the recommendation strategy based on real-time feedback data and user requirements, and optimize the recommendation effect. The system of this embodiment can be used to execute the method embodiment above, and its principle and technical effect are similar, so details are not described here.

[0134] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principle and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent recommendation of voice lines, characterized in that, It includes the following steps: Collect the user's historical call data, real-time communication data, line quality data, and user feedback data; Clean and standardize the collected data to make it meet the model input requirements; Extract relevant features of the user and the line, including but not limited to call duration, call frequency, and line quality, and perform necessary dimensionality reduction processing; Based on the characteristic data of the line, use feature attribution analysis to calculate the contribution degree of each line to the connection rate, and obtain the contribution value of each line; Based on the characteristic data and contribution degree of the line, use collaborative filtering, deep learning, or reinforcement learning algorithm models to score each line; According to the scoring results of each line, sort the lines and select the line with the highest score as the recommended result; Dynamically adjust the recommendation strategy based on real-time feedback data and user needs, so as to optimize the recommendation effect.

2. The intelligent recommendation method for voice lines according to claim 1, wherein The calculation formula of the contribution value is: Among them, g i represents the contribution degree of line i to the change in the overall connection rate, b i represents the change value of line i to the overall connection rate of the product, represents the change value of the overall connection rate of all lines used by the outbound product, represents the change value of the overall connection rate of the product generated by line i, respectively represent the usage ratio and connection rate of line i in the current period, respectively represent the usage ratio and connection rate of line i in the previous period. B represents the total sum of the changes in the connection rate of all lines, and ∈ is the smoothing factor.

3. The intelligent recommendation method for voice lines according to claim 1, characterized in that, The feature attribution analysis steps include: When B > 0, if g i > 0, it means that line i helps to improve the overall connection rate of the product. Next, in model training, a reward will be given to the connection rate eigenvalue of this line; When B > 0, if g i < 0, it means that line i is not conducive to improving the overall connection rate of the product. Next, in the model training, the eigenvalue of the connection rate of this line will be penalized; When B < 0, if g i > 0, it means that line i is not conducive to improving the overall connection rate of the product. Next, in model training, a penalty is imposed on the connection rate eigenvalue of this line; When B < 0, if g i < 0, it means that line i is not conducive to improving the overall connection rate of the product. Next, the connection rate eigenvalue of this line will be penalized during model training.

4. The intelligent recommendation method for voice lines according to claim 1, wherein The steps of using collaborative filtering, deep learning, or reinforcement learning algorithm models to score each line include: Use user-based collaborative filtering algorithm or item-based collaborative filtering algorithm to calculate the similarity of users, and recommend voice lines according to the choices of similar users; Use a deep learning model to analyze the non-linear relationship between the user's call behavior and line characteristics, so as to make accurate line recommendations; Use a reinforcement learning algorithm to dynamically adjust the line recommendation strategy and update the model according to the user's immediate feedback.

5. The intelligent recommendation method for voice lines according to claim 1, characterized in that The recommendation steps include: When the user requests, obtain the real-time information of the user's current location and call destination in real time; Input the real-time data into the trained recommendation model to calculate the recommendation scores of each line; Sort all lines according to the recommendation scores, and select several lines with the highest scores for recommendation.

6. The intelligent recommendation method for a voice line according to claim 1, wherein The reinforcement learning algorithm uses the Q-learning model to optimize the dynamic recommendation strategy through the following steps: At each recommendation, generate a state s with the user's current state as the input; Select an action a for each candidate line, select the line with the maximum expected return, and update the corresponding Q value; According to the user's real-time feedback data and environmental changes, adjust the recommendation strategy so that the recommendation results can be continuously optimized.

7. The intelligent recommendation method for voice lines according to claim 1, characterized in that The dynamic adjustment of the recommendation strategy is achieved through the following steps: During the recommendation process, adjust the scoring weights of the lines in real time according to the user's feedback data and environmental information; According to the real-time feedback data, including connection rate and call quality, update the score of each line and adjust the output result of the recommendation model; Use a reinforcement learning-based feedback mechanism to combine the changes in user needs and network environment, and optimize the line recommendation strategy in real time to achieve the optimal line selection and improve the overall recommendation effect; Dynamically update the feature model according to real-time data, and continuously optimize the model weights using historical data to further improve the accuracy and stability of system recommendations.

8. Voice line intelligent recommendation system, characterized in that It includes: A data collection module for collecting the user's historical call data, real-time communication data, line quality data, and user feedback data; A data preprocessing module for cleaning and standardizing the collected data; A feature engineering module, which is used to extract relevant features of users and lines and perform necessary dimensionality reduction processing; A feature attribution analysis module, which is used to calculate the contribution of each line to the connection rate and perform reward or punishment processing on line features; A recommendation algorithm module, which is used to score each line based on line features and contribution degrees by using collaborative filtering, deep learning or reinforcement learning algorithm models; A recommendation generation module, which is used to sort the lines according to the scores and select the line with the highest score as the recommendation result; An adaptive optimization module, which is used to dynamically adjust the recommendation strategy and optimize the recommendation effect based on real-time feedback data and user requirements.

Images