A method for intelligent dispatching of group call in call center

Abstract

This invention relates to the field of call center technology, specifically to a method for intelligent scheduling of group calls in call centers. The method includes: acquiring a pool of student data to be called and extracting multi-dimensional features to obtain student feature data; generating multiple student groups with different outbound call priorities based on the student feature data; dynamically generating a set of time windows for outbound calls on the current day based on the current timestamp and preset call blocking rules for the education and training industry; calculating the real-time connection probability value of each student at the current moment by combining the student groups and the time window set, and calculating the expected conversion value score based on the real-time connection probability value; calculating the agent capability score based on the real-time status data of the agent team; dynamically generating a group call task queue based on the conversion value score of each student in the student groups, and allocating the outbound call tasks in the group call task queue to matching agent terminals based on the agent capability score. This method can improve the outbound call conversion rate and operational compliance in the education and training industry.

Description

Technical Field

[0001] This invention relates to the field of call technology, and in particular to an intelligent scheduling method for group calls in call centers. Background Technology

[0002] With the rapid development of the education and training industry, call centers, as the core channel for communication between institutions and students, directly impact course conversion rates, renewal rates, and customer satisfaction through their outbound calling efficiency. Currently, call centers in the education and training industry mainly face the following technical challenges in outbound call scheduling: Existing technologies typically employ simple rules or random assignment for mass calling, failing to fully leverage learners' profiles and historical communication behaviors. This "one-size-fits-all" approach leads to the loss of highly interested learners due to a lack of timely outreach, while low-interest learners occupy a significant amount of agent resources, resulting in resource misallocation.

[0003] The education and training industry is subject to strict regulation, with clear legal and industry practice restrictions on outbound calling times. Existing dispatch systems often only set fixed no-call periods and lack the ability to dynamically generate outbound calling windows based on current date attributes and students' personalized preferences, which can easily lead to complaints and even legal risks.

[0004] Low agent-job fit and underutilized agent efficiency: Traditional outbound call task allocation is often based on simple polling or idle-first strategies, ignoring the real-time status of agents. Assigning high-value potential customers to agents with insufficient skills or poor performance directly reduces conversion success rates; conversely, having experienced agents handle low-value tasks results in wasted manpower. Summary of the Invention

[0005] The purpose of this invention is to provide a method for intelligent scheduling of group calls in call centers, which can improve the outbound call conversion rate and operational compliance in the education and training industry.

[0006] To achieve the above objectives, the present invention provides a method for intelligent scheduling of group calls in a call center, comprising: acquiring a pool of student data to be called and performing multi-dimensional feature extraction to obtain student feature data, wherein the features include student profiles and historical communication features; Based on the student characteristic data, cluster analysis is performed on the students using a student classification model to generate multiple student groups with different outbound call priorities. Based on the current timestamp and preset call blocking rules for the education and training industry, dynamically generate a set of time windows for outbound calls that can be made on the same day; By combining the student groups and time window sets, the real-time connection probability value of each student at the current moment is calculated, and the expected conversion value score is calculated based on the real-time connection probability value. Calculate agent capability scores based on real-time status data of the agent team; A group call task queue is dynamically generated based on the conversion value score of each student in the student group, and outbound call tasks in the group call task queue are assigned to matching agent terminals based on agent capability scores.

[0007] The method further includes, after dynamically generating a group call task queue based on the conversion value score of each student in the student group, and allocating outbound call tasks in the group call task queue to matching agent terminals based on agent capability scores, the method also includes, during the group call process, collecting outbound call connection status, student feedback information and agent service evaluation in real time, and dynamically adjusting the outbound call strategy and the priority of the student group.

[0008] The specific steps for obtaining the student data pool to be called and extracting multi-dimensional features to obtain student feature data include: Collect basic information and interaction data of potential outbound call students from the database of the customer relationship management system within a preset time range to build a student data pool to be called; Perform data cleaning on the raw data in the student data pool; Multidimensional feature extraction was performed on the cleaned data. The student profile extracted by multidimensional feature extraction includes the student's age, region, registration duration, current course and remaining course duration; historical communication features include historical communication duration, active time period and historical connection rate. The extracted multidimensional features are then subjected to data standardization processing to form structured student feature data.

[0009] The specific steps of generating multiple student groups with different outbound call priorities by performing cluster analysis on students based on the student feature data using a student classification model include: Principal component analysis was used to perform feature dimensionality reduction on the student feature data. Based on the dimensionality-reduced feature data, the student dataset is divided into a preset number of K initial cluster groups using the K-Means clustering algorithm. The clustering effect is evaluated by the silhouette coefficient to determine the optimal number of cluster groups. For each cluster, analyze its business meaning, identify the central features of each cluster, and assign a business label to each cluster based on the average performance of each cluster on key features such as "historical order conversion rate", "frequency of proactive inquiries in the past N days", "course expiration reminder time" and "historical connection rate". Generate student groups with different outbound call priorities based on business tags.

[0010] Specifically, generating student groups with different outbound call priorities based on business tags includes: designating groups with characteristics such as "high conversion rate", "frequent proactive inquiries", and "near course expiration" as high-intent groups and assigning them a first-level outbound call priority; designating groups with characteristics such as "medium conversion rate", "historical inquiry records but recent silence", and "medium connection rate" as groups to be activated and assigning them a second-level outbound call priority; and designating groups with characteristics such as "low conversion rate", "long-term no interaction", and "low historical connection rate" as low-intent groups and assigning them a third-level outbound call priority.

[0011] The specific steps for dynamically generating a set of time windows for outbound calls on the same day based on the current timestamp and preset call blocking rules for the education and training industry include: Get the current system timestamp, and parse it to get the weekday attribute of the current date, whether it is a public holiday, and the current specific time; Load the preset do-call rule library for the education and training industry, and load the preset do-call rules from the rule configuration center. The do-call rules include at least the legally mandated do-call periods, the industry-standard do-call periods, special rules for holidays, and personalized do-call rules for students. The daily 24-hour continuous timeline is superimposed on the aforementioned call ban rule base to calculate and remove all call ban periods, resulting in a set of continuous original callable time intervals for the day. The original time window is sliced ​​based on outbound call volume prediction. Historical outbound call data is used to predict the outbound call volume for each time period of the day. The original available outbound call time interval is sliced ​​according to the preset outbound call batch granularity to generate several discrete available outbound call time windows. Each time window has a start time and an end timestamp.

[0012] The specific steps of combining the student group and time window set to calculate the real-time connection probability value for each student at the current moment, and calculating the expected conversion value score based on the real-time connection probability value, include: Get the start and end times of the current outbound call window, and extract the time period attributes of the time window; Load the student's multidimensional feature data and group priority labels; The multidimensional feature data is input into a pre-trained connection probability prediction model, which is a binary classification model based on logistic regression, and outputs the real-time connection probability value of the student in the current time window. The expected conversion value base score of students is extracted from student characteristic data. The expected conversion value base score is obtained by comprehensively quantifying the student's historical order amount, number of renewals, course preference matching degree and current business opportunity stage. The real-time connection probability value is weighted and fused with the expected conversion value base score to calculate the student's expected conversion value score at the current moment.

[0013] The specific steps for calculating the agent capability score based on the real-time status data of the agent team include: Real-time acquisition of multi-dimensional status signals from agent terminals, including current call duration, average call duration over historical periods, average conversion rate of the most recent N calls, number of consecutive unconnected calls, current working hours, and historical matching success rate of similar students. Based on the multidimensional state signals, a seat agent competence evaluation model is constructed, which includes the dimensions of skill proficiency, fatigue, emotional stability, and business matching. The average call duration, average conversion rate of the most recent N calls, and number of consecutive missed calls in the historical period are weighted by a time decay factor to obtain real-time quantitative scores for each dimension. The real-time quantified scores of each dimension are input into a preset weight allocation matrix to calculate the initial capability score of the agent. The initial ability score is dynamically corrected based on the current working hours and the rest interval threshold in the preset call ban rules for the education and training industry, to obtain the final agent ability score.

[0014] The specific steps of dynamically generating a group call task queue based on the conversion value score of each student in the student group, and allocating outbound call tasks in the group call task queue to matching agent terminals based on agent capability scores, include: In real time, all student groups are traversed, the expected conversion value score of each student is extracted, and all students to be called are globally sorted in descending order of expected conversion value score to generate an initial priority queue. Based on a preset set of time windows, a subset of students who can be reached within the current time window is extracted from the initial priority queue to construct a dynamic group call task queue. The system acquires the agent capability scores of all online agent terminals in real time and divides the agents into high-capability agent groups, medium-capability agent groups, and basic agent groups based on the scores. The high-value student tasks with expected conversion value in the dynamic group call task queue, which are at the top of the preset proportion, are forcibly mapped to the high-capacity agent group; the remaining student tasks are distributed to the medium-capacity agent group and the basic agent group according to the load balancing strategy. When an agent in a high-capacity agent group is detected to be idle, a task is immediately retrieved from the head of the dynamic group call task queue and sent to that agent terminal, and the queue status is updated in real time.

[0015] The specific steps for collecting outbound call connection status and student feedback information in real time during group calling, and dynamically adjusting the priority of student groups, include: Establish a real-time signaling monitoring channel to synchronously capture the underlying signaling status code of each outbound call task, and classify the outbound call connection status into valid connection, no answer, busy tone, and disconnected number. When a valid connection is detected, the call content is transcribed in real time to extract the student's feedback information, which includes explicit intention keywords, implicit emotional fluctuation values, specific objection tags, and promised follow-up actions. Based on the student feedback information, the real-time connection rate, effective dialogue rate, and positive feedback percentage of each student group within the current window are statistically analyzed. The dynamic priority correction coefficient for each student group is calculated based on the real-time connection rate, effective dialogue rate, and positive feedback percentage. Using the dynamic priority correction coefficient, the comprehensive priority score of each student group is updated in real time, and the task queue reordering mechanism is triggered.

[0016] This invention presents an intelligent scheduling method for group calls in call centers. By constructing a multi-dimensional feature extraction system, utilizing principal component analysis for dimensionality reduction and K-Means clustering algorithm, combined with business meaning analysis, it scientifically divides students into groups with different priorities such as "high intent," "awaiting activation," and "low intent." This method eliminates blind outbound calls, ensuring that key students with high conversion rates, high-frequency inquiries, and those nearing course expiration are prioritized for outreach, thereby significantly improving expected conversion value scores and increasing overall conversion and renewal rates. This invention innovatively introduces a dynamic time window generation mechanism, which not only loads legally mandated no-call periods and industry practices but also incorporates the current date's weekday attribute, holiday characteristics, and student-specific no-call rules. Through superimposed calculations, no-call periods are eliminated and sliced. This ensures that all outbound calls are executed within compliant and high-probability connection periods, effectively avoiding complaints of violations and harassment. Simultaneously, it utilizes historical data to predict task volume, optimizing the utilization of time resources.

[0017] This invention breaks through the traditional simple allocation model and establishes a seat capacity evaluation model that includes skill proficiency, fatigue level, emotional stability, and business matching. By introducing a time decay factor, it quantifies seat status in real time and classifies seats according to their scores. The system forcibly maps high-value student tasks to high-capability seat groups, ensuring that "good resources are used where they are most needed." This not only increases the probability of closing deals with high-intent students but also protects the work experience of ordinary seats through load balancing, maximizing overall team efficiency.

[0018] This invention establishes a real-time signaling monitoring and speech-to-text analysis channel during group calls, enabling real-time capture of connection status and extraction of explicit / implicit feedback from learners (such as keywords of intent and emotional fluctuations). Based on metrics such as real-time connection rate and effective dialogue rate, a dynamic priority correction coefficient is calculated, instantly updating learner group priorities and reordering the task queue. This closed-loop mechanism of "execution-feedback-adjustment" allows outbound calling strategies to dynamically evolve with real-time conditions, rapidly responding to market changes and ensuring continuously optimized outbound calling results. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0020] Figure 1 This is a flowchart of a call center group call intelligent scheduling method according to the present invention.

[0021] Figure 2 This is a flowchart of the present invention for obtaining student feature data by acquiring a pool of student data to be called out and performing multi-dimensional feature extraction.

[0022] Figure 3 This invention is a flowchart of how the present invention generates multiple student groups with different outbound call priorities by performing cluster analysis on students through a student classification model based on the student feature data.

[0023] Figure 4 This is a flowchart of the present invention, which dynamically generates a set of time windows for outbound calls on the same day based on the current timestamp and preset call blocking rules for the education and training industry.

[0024] Figure 5 This invention is a flowchart that combines the student group and time window set to calculate the real-time connection probability value of each student at the current moment, and calculates the expected conversion value score based on the real-time connection probability value.

[0025] Figure 6 This is a flowchart of the present invention for calculating agent capability scores based on real-time status data of the agent team.

[0026] Figure 7 This is a flowchart of the present invention that dynamically generates a group call task queue based on the conversion value score of each student in the student group, and assigns outbound call tasks in the group call task queue to matching agent terminals based on agent capability scores.

[0027] Figure 8 This invention is a flowchart of a process for real-time collection of outbound call connection status and student feedback information during group calling, and for dynamically adjusting the priority of student groups. Detailed Implementation

[0028] The embodiments of the present invention are described in detail below. Examples of the embodiments are shown in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, but should not be construed as limiting the present invention.

[0029] First embodiment: Please see Figures 1-8 This invention provides a method for intelligent scheduling of group calls in a call center, comprising: S101 acquires the student data pool to be called out and performs multi-dimensional feature extraction to obtain student feature data, including student profile and historical communication features. The specific steps include: S201 collects basic information and interaction data of potential outbound call students from the database of the customer relationship management system within a preset time range, and builds a student data pool to be called. Break down data silos by using the core database of an enterprise-level CRM system (such as MySQL, Oracle, or the distributed data warehouse Hive) as the data source, locking in the student information storage table, and dynamically setting preset time ranges and time granularities of days / weeks / months based on business needs to complete the collection of potential outbound student data. The time range can be configured according to business needs, such as students who registered but did not complete a transaction in the past 30 days, or enrolled students with less than 7 days of course remaining.

[0030] The collected data includes two categories: basic information and interactive data. Basic information includes static attributes such as student ID, name, contact information, and registration time. Interactive data includes historical call records, SMS sending records, APP login logs, page browsing history, and past work order records. Using student ID as the primary key, the collected data is correlated and aggregated to form a "Waiting-for-Outbound-Calls Student Wide Table," thus constructing a student data pool for those awaiting outbound calls.

[0031] S202 performs data cleaning on the raw data in the student data pool; Multiple deduplication is performed based on the student's unique identifier and mobile phone number to merge redundant records of the same student generated on different channels; then invalid contact methods such as empty numbers, incorrectly formatted mobile phone numbers, and test data are removed, and logically conflicting data such as negative registration duration and remaining course duration exceeding the total class hours are identified and removed.

[0032] For numerical features (such as age), fill with the mean, median, or cluster-based prediction; for categorical features (such as location), fill with the mode or mark them as "unknown" category. Then, convert the date and time to the ISO8601 standard and convert the area code to the national standard administrative division code to ensure data consistency.

[0033] S203 performs multi-dimensional feature extraction on the cleaned data. The student profile extracted by multi-dimensional feature extraction includes the student's age, region, registration duration, current course and remaining course duration; historical communication features include historical communication duration, active time period and historical connection rate. The cleaned raw data is transformed into high-value feature vectors that are machine-understandable and business-interpretable, constructing a panoramic view of trainees that integrates static profiles and dynamic behaviors. The extracted features are divided into two categories: trainee profile features and historical communication features. Student profile features: These are static / semi-static dimensions, including demographic features (age is derived from birth date and age group labels are refined), geographical features (province, city, and district are analyzed and city level and economic development level labels are derived), life cycle features (registration duration is calculated and the student's stage is determined), and academic status features (information on currently enrolled courses is extracted and the remaining course duration is calculated as a key basis for urgency scoring). Historical communication characteristics: This belongs to the dynamic behavior dimension and includes communication depth characteristics (statistical analysis of historical communication duration to reflect the degree of cooperation and intention of students), time preference characteristics (analysis of the distribution of successful historical calls to identify students' active periods and generate optimal contact time window labels), and response probability characteristics (calculation of historical connection rate, and construction of student anti-interference index and answering tendency score by combining behaviors such as rejection and hanging up).

[0034] S204 performs data standardization processing on the extracted multidimensional features to form structured student feature data.

[0035] To eliminate differences in the dimensions of different features and accelerate model convergence, unstructured or semi-structured data is transformed into a standardized matrix format. The specific processing method is as follows: Normalization: For bounded numerical features such as connection rate, map them to the interval [0, 1]; Standardization: For characteristics such as age and communication duration that follow a normal distribution, zero mean and unit variance are applied to eliminate the influence of dimensions; After completing the above processing, the feature vectors are assembled into a standard JSON object or a DataFrame matrix, and metadata tags such as data update timestamps and feature version numbers are added. Finally, structured student feature data is output, which can be directly stored in the feature store or pushed to the outbound call scheduling engine in real time to provide input for intelligent routing matching and predictive outbound call models.

[0036] S102 Based on the student characteristic data, cluster analysis is performed on the students through a student classification model to generate multiple student groups with different outbound call priorities; By employing feature reduction, unsupervised clustering, business meaning analysis, and priority labeling, students are divided into different outbound call priority groups, providing a basis for formulating differentiated outbound call strategies. Specific steps include: S301 uses principal component analysis to perform feature dimensionality reduction on student feature data; The covariance matrix of standardized student feature data is calculated to quantify the linear correlation between feature variables; eigenvalue decomposition is performed on the covariance matrix to calculate the contribution rate of each principal component; a cumulative variance contribution rate threshold of 85%-95% is set to automatically select the top n principal components and retain the core variation information of the data; the original high-dimensional feature data is projected onto the selected low-dimensional principal component space to generate a dimension-reduced feature vector matrix, removing noise interference and making the cluster structure of the student group clearer.

[0037] Based on the dimensionality-reduced feature data, S302 uses the K-Means clustering algorithm to divide the student dataset into a preset number of K initial cluster groups, and evaluates the clustering effect through the silhouette coefficient to determine the optimal number of cluster groups; The K-Means++ algorithm is used to optimize the selection of initial cluster centers, avoiding the local optimum problem caused by random initialization in traditional K-Means and accelerating algorithm convergence. The Euclidean or cosine distance between each student sample and the current K cluster centers is calculated, and the sample is assigned to the group belonging to the nearest cluster center. The centroids of each group are recalculated, and this process is repeated until the centroid positions show no significant change or the maximum number of iterations is reached. Clustering operations are performed on a preset range of K=3 to K=10, using the silhouette coefficient as the evaluation metric to calculate the difference between sample cohesion and separation. The K value corresponding to the maximum silhouette coefficient is selected as the optimal number of clusters, ensuring the highest similarity within groups and the greatest difference between groups. K initial cluster groups are generated, each containing the corresponding student ID and dimensionality-reduced feature coordinates.

[0038] S303 performs business meaning analysis on each cluster group, analyzes the characteristic center point of each cluster group, and assigns business labels to each group based on the average performance of each group on key characteristics such as "historical order conversion rate", "frequency of proactive inquiries in the past N days", "course expiration reminder time" and "historical connection rate". The mathematical clustering results are transformed into business-understandable user profiles. Mean analysis of key business metrics is used to assign clear commercial meanings to each group. Then, the centroids of each cluster are inversely mapped from the principal component space back to the original feature space to obtain the typical performance of the group in terms of the original business features. For each group, the arithmetic mean of four core business metrics—historical conversion rate, frequency of proactive inquiries in the past N days, course expiration reminder time, and historical connection rate—is calculated, where N can be dynamically adjusted according to the business cycle. Based on the average performance of the metrics, descriptive tags are generated for each group using the business expert rule base, such as "high conversion - high activity - near expiration", "medium conversion - inactive - long term", and "low conversion - low connection - churn risk".

[0039] S304 generates student groups with different outbound call priorities based on business tags.

[0040] Specifically, groups with characteristics such as "high conversion rate", "frequent proactive inquiries", and "near course expiration" are designated as high-intent groups and given first-level outbound call priority; groups with characteristics such as "medium conversion rate", "historical inquiry records but recent silence", and "medium connection rate" are designated as groups to be activated and given second-level outbound call priority; and groups with characteristics such as "low conversion rate", "long-term no interaction", and "low historical connection rate" are designated as low-intent groups and given third-level outbound call priority.

[0041] The criteria for determining the priority of outbound calls (high intent / high value groups) are to simultaneously meet or weightedly meet the following characteristics: "high historical conversion rate" (indicating strong willingness to pay), "high frequency of proactive inquiries" (indicating strong current demand), and "approaching course expiration" (indicating urgent scenarios such as renewal).

[0042] Place it at the top of the outbound call queue to ensure priority calls are made during peak hours (such as 10:00-11:30 AM and 2:30-4:30 PM).

[0043] Resource matching involves assigning the task to a senior top salesperson or a dedicated consultant for follow-up.

[0044] The sales strategy employs high-conversion scripts oriented towards "closing an order / renewing subscriptions," emphasizing limited-time offers and a sense of urgency regarding the course.

[0045] The criteria for determining the priority of secondary outbound calls (pending activation / potential groups) are: characterized by "medium conversion rate", "historical consultation records but recent silence" (potential for wake-up call), and "medium connection rate".

[0046] The strategy is defined as the "potential reactivation group." Outreach is scheduled during breaks in primary group calls or off-peak hours. Individual sales staff or AI bots are assigned for initial screening. The communication strategy employs a "care / research" oriented activation script, re-establishing connections and uncovering potential needs through offering materials, course trials, etc.

[0047] The criteria for determining the priority of the third-level outbound call (low intention / long tail group) are characterized by "low historical order conversion rate", "long-term no interaction" (exceeding the set threshold, such as 90 days) and "low historical connection rate" (multiple rejections or empty numbers).

[0048] The strategy is defined as a "long-tail maintenance group." Outbound calls are made in batches only during system downtime or specific marketing events (such as major promotions) to avoid consuming valuable human resources and lines. The calls are mainly handled by a fully automated AI outbound calling system, or sent via SMS / WeChat template messages. A "wide net" notification style is used. If the call is not answered or is explicitly rejected multiple times, the call is automatically moved to the "public pool" or the outbound calls are suspended to prevent excessive disturbance from damaging the brand reputation.

[0049] S103 dynamically generates a set of time windows for outbound calls that can be made on the same day based on the current timestamp and preset call blocking rules for the education and training industry; The specific steps include: S401 obtains the current system timestamp, and parses it to get the weekday attribute of the current date, whether it is a statutory holiday, and the current specific time; Obtain the current system timestamp accurate to milliseconds and synchronize it to the national standard time source to avoid compliance risks caused by local clock deviations. At the same time, complete the following time attribute parsing: Analyze the weekday attribute of the current date to distinguish between weekdays and rest days; Call the national statutory holiday interface or the built-in calendar library to determine whether the current date is a statutory holiday, a workday with adjusted working hours, or a special commemorative day; Extract the current time, minute, and second to determine the call blocking threshold in real time.

[0050] The above parsing results are encapsulated into a standard object and used as input parameters for the subsequent rule engine.

[0051] S402 loads the preset do-call rule library for the education and training industry and loads the preset do-call rules from the rule configuration center. The do-call rules include at least the legally mandated do-call periods, the industry-standard do-call periods, special rules for holidays, and personalized do-call rules for students. The latest version of the call blocking policy is pulled in real time from the rule configuration center, supporting hot updates without restarting the service. The call blocking rules cover four dimensions and establish a priority arbitration mechanism: Legally mandated no-call periods: In accordance with regulations such as the "Regulations on the Management of Short Message Services," daily mandatory no-call periods are set, such as 12:00-14:00 on weekdays and 21:00-09:00 the next day. Industry-standard no-call periods: Based on the characteristics of the education and training industry, special no-call zones are set during school hours, exam weeks, etc. Special rules for holidays: A "quiet period" strategy will be implemented for long holidays such as Spring Festival and National Day, allowing only low-frequency access during specific periods; Personalized call blocking rules for trainees: Read the global "Do Not Disturb" list and the trainee's personal preference settings to form a personalized blacklist time set.

[0052] Rule priority follows the order of statutory > industry > personalized > default. When rules conflict, the strictest restriction period will be adopted according to the principle of "strictest rule".

[0053] S403 overlays the daily 24-hour continuous timeline with the aforementioned call ban rule base, removes all call ban periods, and obtains the set of original callable time intervals for the day. Construct a continuous linear timeline from 00:00:00 to 23:59:59 on the current day; convert the call blocking rules into a set of specific time intervals, convert the relative time rules into absolute timestamps, and perform a merging operation on overlapping call blocking intervals to form mutually exclusive call blocking time blocks; perform set difference operation to remove all call blocking periods and generate a set of continuous original callable time intervals for the current day.

[0054] S404 slices the original time window based on outbound call volume prediction. Use the historical outbound call database to extract outbound call data from the past 30 to 90 days that have the same weekday attributes and similar holiday scenarios.

[0055] Using time series forecasting algorithms (such as Prophet and ARIMA) or machine learning regression models (such as XGBoost), the expected total number of outbound calls and peak concurrent demand at the hourly level are predicted based on outbound call data. Combining the current number of agents on duty with the expected average efficiency per agent, the maximum theoretical outbound call capacity for each time period is calculated.

[0056] S405 calls historical outbound call data to predict the outbound call volume for each time period of the day, and slices the original outbound callable time interval according to the preset outbound call batch granularity to generate several discrete outbound callable time windows, each time window having a start time and an end timestamp.

[0057] The continuous raw intervals are divided into executable, monitorable, and dynamically adjustable discrete task units to achieve fine-grained traffic control. Based on business management needs, preset outbound call batch granularity is set, typically 15 minutes, 30 minutes, or 1 hour.

[0058] During periods of stable task volume forecasts, the original outbound call intervals are divided equally into segments with a fixed time granularity (e.g., 30 minutes). During "golden hours" with high predicted task volume or high connection rates, the segment granularity is reduced (e.g., 15 minutes) to increase scheduling frequency and flexibility; during off-peak hours, the segment granularity is increased to reduce system overhead. Ensure that the estimated task volume within each segment window does not exceed the maximum capacity of the agents during that period to prevent queue backlog.

[0059] Finally, the final set of discrete outbound time windows is generated, and each window object contains: Window_ID: A unique identifier.

[0060] Start_Timestamp / End_Timestamp: Precise start and end timestamps.

[0061] Priority_Level: Window priority based on predicted call completion rate.

[0062] Max_Call_Volume: This window suggests the maximum number of outbound calls allowed.

[0063] Status: The initial status is marked as "to be activated".

[0064] S104 combines the student group and time window set to calculate the real-time connection probability value of each student at the current moment, and calculates the expected conversion value score based on the real-time connection probability value. The specific steps include: S501 obtains the start and end times of the current outbound call window and extracts the time period attributes of the window. Read the set of currently available outbound call time windows generated by S103, and accurately obtain the Start_Time (start time stamp) and End_Time (end time stamp) of the current execution window.

[0065] Multidimensional mapping of time period attributes: Basic attributes: Analyze the hourly time period (such as "morning peak 09:00-10:00", "after lunch break 14:30-15:30", "evening prime time 19:00-20:00").

[0066] Business Attributes: Mark the special significance of this period in the education and training industry (such as "after-school pick-up and drop-off period", "parents' free time period", "after-school tutoring period").

[0067] Environmental attributes: Combine calendar information to mark whether it is a weekday, weekend, the eve of a holiday, or a special marketing period (such as "Double 11 pre-sale period").

[0068] Feature vectorization: The above attributes are converted into numerical values ​​or One-Hot encoded vectors that the model can recognize, which are then used as the time dimension input for subsequent probability prediction.

[0069] S502 loads the student's multidimensional feature data and group priority labels; Retrieve multidimensional feature data of students generated by S101 from the feature repository, including demographic information (age, region), academic status (remaining course time, subjects enrolled), etc., and inject the cluster analysis results generated by S102 in real time. Focus on loading group priority labels (such as "high intention group" and "group to be activated") and the statistical characteristics of the group (such as average connection rate of the group and average conversion rate of the group).

[0070] Real-time behavior capture: If a real-time data stream with millisecond-level latency is available, supplement it with the student's most recent behavior trajectory (e.g., "just browsed the renewal page 5 minutes ago") to enhance the timeliness of the features.

[0071] S503 inputs the multidimensional feature data into a pre-trained connection probability prediction model, which is a binary classification model based on logistic regression, and outputs the real-time connection probability value of the student in the current time window. We invoked a pre-trained and deployed binary logistic regression model. This model was chosen because it has extremely high inference speed and good probability calibration capability when dealing with linearly separable features, making it suitable for high-concurrency real-time scenarios.

[0072] The model is trained based on massive historical outbound call logs. Positive samples are "successfully connected," and negative samples are "not connected / rejected / phone switched off." Training features cover student profiles, historical communication habits, time window attributes, and interaction features.

[0073] The time attribute vector of S501 is concatenated with the student feature vector of S502 and then input into the model.

[0074] The model calculates the output value using the Sigmoid function, resulting in a floating-point number between 0 and 1, representing the real-time connection probability. For example, an output of 0.85 indicates that the student has an 85% probability of answering the call at the current moment.

[0075] S504 extracts the expected conversion value base score of students from student characteristic data. The expected conversion value base score is obtained by comprehensively quantifying the student's historical order amount, number of renewals, course preference matching degree and current business opportunity stage. The following core factors were extracted from the student characteristic data and then standardized: Historical GMV: Total amount paid by students in the past, reflecting their payment ability and trust level (weighted at 30%).

[0076] Renewal Count: Historical renewal frequency, reflecting loyalty and long-term retention potential (weighted at 20%).

[0077] Course Preference Match (Pref): The cosine similarity between the currently recommended courses and the student's historical interest tags, reflecting the degree of fit of needs (weighted at 25%).

[0078] The current business opportunity stage is scored based on the position in the sales funnel (e.g., "Strong Intent" = 1.0, "Initial Contact" = 0.4, "Lost" = 0.1) (weighting 25%).

[0079] Comprehensive quantification formula: Score base =w1⋅Norm(GMV)+w2⋅Norm(Count)+w3⋅Norm(Pref)+w4⋅Stage Where w i The weighting coefficients configured for the business logic, and Norm is the normalization function.

[0080] S505 performs a weighted fusion of the real-time connection probability value and the expected conversion value base score to calculate the student's expected conversion value score at the current moment.

[0081] By organically combining the two dimensions of "whether the call can be connected" and "whether it's worth making the call," a final decision score guiding outbound call ranking is calculated. Expected Value Theory is used for weighted fusion. The core logic is: a student's actual expected value = the probability of them answering × the value that can be generated after they answer.

[0082] S105 calculates agent capability scores based on the real-time status data of the agent team; The specific steps include: S601 collects multi-dimensional status signals from the agent terminal in real time. The multi-dimensional status signals include the agent's current call duration, the average call duration over a historical period, the average conversion rate of the most recent N calls, the number of consecutive unconnected calls, the current working time, and the historical matching success rate of similar students. Establish a high-frequency, low-latency data perception layer to capture key indicators in real time from the CTI system, CRM system, and agent desktop monitoring system.

[0083] Call behavior signals: Current call duration: Real-time monitoring of the duration of the current call to determine whether it has become a pointless conversation or an excessively long communication session.

[0084] Historical average call duration: Statistics on the average processing time over the past 1-4 hours, reflecting the current pace of communication efficiency.

[0085] Performance conversion signal: Average conversion rate of the most recent N calls: Using a sliding window mechanism (such as the most recent 20 calls), the percentage of successful orders or agreements is calculated to keenly capture the current "feel" and sales momentum of the agents.

[0086] Number of consecutive missed calls: This counts the number of times a call is rejected, hung up, or the number is no longer in service, serving as a key indicator of the frustration experienced by the agent.

[0087] Fatigue and load signals: Current working hours: Calculated as the cumulative online time since starting work today (including short breaks).

[0088] Continuous work segments: Identify whether the work is in a state of continuous operation without rest for a long period of time.

[0089] Matching performance signal: Historical success rate of matching similar students: Based on the group characteristics of the current student to be assigned (such as "high-interest K12 parents"), retrieve the historical success rate of the agent in handling such students to assess the matching degree of their specialized skills.

[0090] Based on the multi-dimensional state signals, S602 constructs a seat operator competence evaluation model that includes the dimensions of skill proficiency, fatigue, emotional stability, and business matching. By mapping discrete raw signals to four core capability dimensions with business implications, a structured evaluation framework is constructed.

[0091] Skill proficiency dimension: Input sources: historical matching success rate of similar students, and average call duration (efficiency item) for historical time periods.

[0092] Definition: Measures the agent's proficiency and communication efficiency in handling specific types of business.

[0093] Fatigue level dimension: Input sources: current working hours, continuous working segments, current call duration (if abnormally long).

[0094] Definition: Quantifying the physiological fatigue level of agents; excessive fatigue will lead to decreased attention and a harsh tone of voice.

[0095] Emotional stability dimension: Input sources: Number of consecutive unanswered calls, volatility of conversion rate of the most recent N calls.

[0096] Definition: This assessment evaluates the psychological resilience and emotional state of agents after experiencing consecutive rejections. The more consecutive unanswered calls, the lower the emotional stability score.

[0097] Business fit dimension: Input source: Historical matching success rate of similar students (targeted weighting).

[0098] Definition: This measure is the degree to which the skills of the sales representatives match the characteristics of the trainees to be assigned, reflecting the suitability of "specialized personnel for specialized tasks".

[0099] S603 uses a time decay factor to weight the average call duration of the historical period, the average conversion rate of the most recent N calls, and the number of consecutive missed calls to obtain real-time quantitative scores for each dimension. Introducing the "recency effect" gives higher weight to recent performance in the rating, ensuring that the rating reflects the current status of the agent in real time, rather than being diluted by long-standing historical data.

[0100] The weighted processing logic is based on the average call duration over historical periods: the more recent the call duration, the higher the weight, quickly reflecting changes in the agent's current speaking speed and communication rhythm. Each call within the sliding window is assigned a decreasing weight, with the success or failure of the most recent call having the greatest impact on the score, rapidly amplifying signals of "winning streaks" or "losing streaks." The number of consecutive missed calls serves as a negative indicator; the more frequent and recent the missed calls, the heavier the deduction in the emotional stability dimension, exhibiting a non-linear downward trend. S604 inputs the real-time quantified scores of each dimension into a preset weight allocation matrix to calculate the initial capability score of the agent. The importance of each dimension is flexibly adjusted based on the current business strategy to calculate a comprehensive initial capability score. Multiple preset weight configuration schemes can be dynamically switched according to operating periods or business objectives.

[0101] Standard Mode: Skills (30%) + Matching (30%) + Mood (20%) + Fatigue (20%).

[0102] In a push-up mode (such as a month-end sales push): increase the weighting of skills and compatibility, and tolerate fatigue to a certain extent.

[0103] Care models (such as newbie protection): Increase the weight of emotional stability to avoid high stress leading to attrition.

[0104] S605 dynamically corrects the initial ability score based on the current working hours and the rest interval threshold in the preset call ban rules for the education and training industry, to obtain the final agent ability score.

[0105] By introducing both hard constraints and soft adjustment mechanisms, we ensure that the scoring results not only meet the compliance requirements of the education and training industry, but also reflect the human touch of management.

[0106] Read the pre-set no-call rules for the education and training industry and the rest intervals stipulated by labor law in S103 (such as "a 15-minute break is required after 2 hours of continuous work"). If the current working time is greater than the maximum continuous working threshold and the time since the last break is less than the minimum rest interval, a forced correction mechanism is triggered.

[0107] Regardless of how high the initial score is, the final score will be set to 0 or a very low value (such as 1), and the agent will be marked as "forced to rest" in the routing system, preventing the assignment of new tasks until the rest conditions are met.

[0108] S106 dynamically generates a group call task queue based on the conversion value score of each student in the student group, and assigns the outbound call tasks in the group call task queue to the matching agent terminal based on the agent capability score.

[0109] The specific steps include: The S701 iterates through all student groups in real time, extracts the expected conversion value score of each student, and sorts all students to be called globally according to the expected conversion value score from high to low, generating an initial priority queue. Real-time scanning of all student groups generated in S102 (high intent, pending activation, low intent, etc.), extracting the expected conversion value score of each student calculated in S104.

[0110] Strictly ranked by expected conversion value score, students are sorted in descending order. For students with the same score, auxiliary ranking factors are introduced, such as "course expiration urgency" (higher urgency, higher priority), "historical unreachable time" (longer unreachable time, higher priority), or "random disturbance factor" (to prevent deadlock), to ensure the uniqueness and fairness of the ranking results. All ranked student IDs and their associated metadata (score, group tags, recommended script ID, etc.) are loaded into a memory cache to generate an initial priority queue. This queue is a dynamically changing list that is automatically rearranged as student scores are updated in real time.

[0111] S702 constructs a dynamic group call task queue by extracting a subset of reachable students within the current time window from the initial priority queue based on a preset set of time windows. Read the set of available outbound call time windows generated by S103 for the current day to obtain the valid window range to which the current system time belongs. Traverse the initial priority queue and check the personalized call blocking rules for each student and the general time constraints of their group. Remove all student records that are currently in a call blocking period, have not yet reached their scheduled outbound call time, or have been marked as "outbound calls completed today". Extract the remaining student records that meet the requirements of the current time window to form a dynamic group call task queue.

[0112] S703 obtains the agent capability scores of all online agent terminals in real time, and divides the agents into high-capability agent group, medium-capability agent group and basic agent group according to the scores. Based on the real-time agent capability score calculated by S105, online agent resources are finely classified and a differentiated service capability pool is constructed.

[0113] The system polls all agent terminals in the "Ready" or "Account After Processing (ACW)" state, indicating an impending end, every millisecond to obtain their latest agent capability score. Based on preset dynamic thresholds or percentile rankings, agents are divided into three tiers: Elite Group: Ranked in the top 20% (or score > 85). Characteristics: High skill proficiency, emotionally stable, and high recent conversion rate. Primarily responsible for challenging tasks.

[0114] Standard Group: Scores in the middle 50% (or 60 points < score ≤ 85 points). Characteristics: Stable performance, possesses basic business skills. Responsible for the main batch outbound call tasks.

[0115] Junior Group: Ratings in the bottom 30% (or scores ≤ 60). Characteristics: Newcomers, high fatigue levels, or recent performance fluctuations. Assigned low-difficulty tasks or cleaning-related tasks.

[0116] Status labeling: Each agent is labeled with a real-time group, and the real-time idle number of agents in each group is maintained in the routing engine.

[0117] S704 forcibly maps high-value student tasks with expected conversion value in the dynamic group call task queue, which are at the top of a preset proportion, to the high-capacity agent group; and distributes the remaining student tasks to the medium-capacity agent group and the basic agent group according to the load balancing strategy. Forced mapping of high-value tasks: Threshold setting: Define a threshold for "high-value learners" (e.g., agent competency score > 90 points, or tasks in the top 5% of the queue). These high-value tasks will be locked and can only be assigned to high-competency agent groups. If all members of the high-competency group are busy, the task will be suspended at the head of the queue and will never be downgraded to be assigned to the intermediate / basic groups.

[0118] The scope of load balancing for remaining tasks is defined as the remaining student tasks (medium-low value) excluding high-value tasks. Weighted Round-Robin or Least Connections algorithms are used to assign tasks to medium-capacity agent groups and basic agent groups.

[0119] The allocation ratio is dynamically adjusted based on the current queue length and average processing time of each group. If the medium-capability group experiences severe backlog, lower-difficulty tasks (such as notifications) can be appropriately diverted to the basic group, and vice versa. A real-time Task_ID -> Agent_Group mapping table is generated to guide subsequent specific actions.

[0120] When the S705 detects that the status of an agent in a high-capacity agent group has become idle, it immediately retrieves the task from the head of the dynamic group call task queue and sends it to the agent terminal, and updates the queue status in real time.

[0121] A zero-latency task triggering mechanism is implemented to ensure that agents receive the optimal task as soon as they are available, and the system status is updated in real time to prevent duplicate assignment.

[0122] Idle Event Listener: Deploy an event listener to monitor agent status change messages in the CTI system in real time. Once an agent in a high-capacity agent group is detected to have changed to "idle", the routing logic is immediately triggered.

[0123] Retrieve task details from the head of the dynamic group call task queue (i.e., the task with the highest current score that is not currently locked). At the microsecond instant of clicking "issue," re-verify whether the student has just been reached by another agent or if their status has changed (to prevent concurrent conflicts). Send an outbound call instruction (including student number, pop-up information, recommended script, and value score prompt) to the agent's terminal via the softphone interface.

[0124] After a task is successfully assigned, the student is immediately removed from the dynamic queue or marked as "Calling," and the occupied agent ID is recorded. If there are no available agents in the high-capability group, the system automatically switches to the idle event of the listening / basic group and assigns the task according to the allocation logic in S704. After the agent ends the call, the result (connected / not connected / order completed) is sent back, triggering the recalculation of scores in S104 and S105, which in turn causes dynamic adjustments to the queue sorting and agent grouping, forming a complete adaptive closed loop.

[0125] During the group call process, S107 collects the outbound call connection status, student feedback information, and agent service evaluation in real time, and dynamically adjusts the outbound call strategy and the priority of the student group.

[0126] The specific steps include: S801 establishes a real-time signaling monitoring channel, synchronously captures the underlying signaling status code of each outbound call task, and classifies the outbound call connection status into valid connection, no answer, busy tone, and disconnected number. By utilizing the SIP protocol stack or CTI middleware, a low-latency real-time signaling monitoring channel is established to synchronously capture the underlying signaling interaction logs of each outbound call task.

[0127] Parse the standard SIP status codes returned by the operator (such as 200 OK, 486 Busy, 404 Not Found, 480 Temporarily Unavailable, etc.) and their proprietary extension codes. Combine ASR (Automatic Speech Recognition) pre-silence detection and ringback tone duration analysis to eliminate false call connections (such as machine answers and ringback tone timeouts). Normalize the complex signaling results into four standard status categories: Valid connection: A call is considered valid if a real person answers and the call duration exceeds a preset threshold (e.g., >5 seconds).

[0128] No answer: The ringing timeout period has passed without being answered or there is no voice interaction before the user hangs up.

[0129] Busy tone: A busy tone signal is detected or a 486 / 503 status code is returned, indicating that the user's line is busy.

[0130] Service suspended due to invalid number: An invalid number prompt tone is detected or a 404 / 604 status code is returned, marking it as invalid data.

[0131] When S802 detects a valid connection, it performs real-time speech transcription of the call content to extract the student's feedback information, which includes explicit intention keywords, implicit emotional fluctuation values, specific objection tags, and promised follow-up actions. Upon detecting a "valid connection," the low-latency streaming speech recognition engine is immediately activated to convert two-way speech into a text stream in real time, supporting dialect recognition and noise filtering. Through keyword matching and intent recognition models, high-value intent words such as "how much," "when is the class," "want to try it out," and "think about it" are extracted to determine the student's purchase intention level (H / M / L). Using acoustic feature analysis (pitch, speech rate, volume) combined with text sentiment analysis, the student's emotional score (-1.0 to +1.0) is calculated. The system also identifies the emotional states behind micro-expressions such as anger, anxiety, excitement, and hesitation.

[0132] The system automatically categorizes student reasons for refusal, such as "too expensive," "time conflict," "already enrolled in another institution," and "child not interested," forming structured Object_Tags. It identifies mutually agreed-upon actions, such as "adding each other on WeChat," "sending materials," "scheduling a Wednesday follow-up visit," and "attending a public class," automatically generating CRM to-do tasks. The extracted results are encapsulated into an Interaction_Feedback object and associated with the current call log and student ID.

[0133] S803 calculates the real-time connection rate, effective dialogue rate, and positive feedback percentage of each student group within the current window based on the student feedback information. S804 calculates the dynamic priority correction coefficient for each student group based on the real-time connection rate, effective dialogue rate, and positive feedback ratio. The dynamic priority correction coefficient K for each student group is calculated based on the real-time connection rate, effective dialogue rate, and positive feedback ratio, after being weighted according to different factors. adjust This transforms multidimensional performance indicators into actionable mathematical coefficients to guide the prioritization of tasks.

[0134] If all three indicators are 20% above the baseline, then K adjust The upper limit can be increased to 1.5, significantly boosting the weight.

[0135] If a high percentage of "numbers that are not in service" or "strong negative sentiment" occur consecutively, the Kadjust lower limit will drop to 0.5 or even 0.1, resulting in a forced demotion.

[0136] Circuit breaker protection: If the complaint rate or negative sentiment value of a group exceeds a threshold within a short period of time, the circuit breaker will be automatically triggered (K). adjust =0, suspend outbound calls for this group.

[0137] S805 uses the dynamic priority correction coefficient to update the comprehensive priority score of each student group in real time and triggers the task queue reordering mechanism.

[0138] Read the raw static priority scores for each group.

[0139] Apply dynamic score adjustment dynamic =Score static ×K adjust .

[0140] Introduce a time decay factor to ensure that the latest feedback has the greatest impact on the score.

[0141] Once the new score is calculated dynamic Immediately send a reordering signal to the task scheduling engine in S106. Without interrupting the current call, the scheduling engine recalculates the expected conversion value score of all students belonging to the group in the outbound call queue (S104 logic reuse) and adjusts their positions in the dynamic group call task queue.

[0142] The above-disclosed embodiments are merely one or more preferred embodiments of this application and should not be construed as limiting the scope of this application. Those skilled in the art can understand that all or part of the processes for implementing the above embodiments and equivalent changes made in accordance with the claims of this application still fall within the scope of this application.

Claims

1. A method for intelligent dispatching of group calls in a call center, characterized in that, include: Obtain the student data pool to be called out and perform multi-dimensional feature extraction to obtain student feature data, including student profile and historical communication features; Based on the student characteristic data, cluster analysis is performed on the students using a student classification model to generate multiple student groups with different outbound call priorities. Based on the current timestamp and preset call blocking rules for the education and training industry, dynamically generate a set of time windows for outbound calls that can be made on the same day; By combining the student groups and time window sets, the real-time connection probability value of each student at the current moment is calculated, and the expected conversion value score is calculated based on the real-time connection probability value. Calculate agent capability scores based on real-time status data of the agent team; A group call task queue is dynamically generated based on the conversion value score of each student in the student group, and outbound call tasks in the group call task queue are assigned to matching agent terminals based on agent capability scores.

2. The intelligent dispatching method for group calls in a call center as described in claim 1, characterized in that, After dynamically generating a group call task queue based on the conversion value score of each student in the student group, and allocating the outbound call tasks in the group call task queue to the matching agent terminals based on the agent's ability score, the method further includes collecting outbound call connection status, student feedback information and agent service evaluation in real time during the group call process, and dynamically adjusting the outbound call strategy and the priority of the student group.

3. The intelligent dispatching method for group calls in a call center as described in claim 2, characterized in that, The specific steps for obtaining the student data pool to be called and performing multi-dimensional feature extraction to obtain student feature data include: Collect basic information and interaction data of potential outbound call students from the database of the customer relationship management system within a preset time range to build a student data pool to be called; Perform data cleaning on the raw data in the student data pool; Multidimensional feature extraction was performed on the cleaned data. The student profile extracted by multidimensional feature extraction includes the student's age, region, registration duration, current course and remaining course duration; historical communication features include historical communication duration, active time period and historical connection rate. The extracted multidimensional features are then subjected to data standardization processing to form structured student feature data.

4. The intelligent dispatching method for group calls in a call center as described in claim 3, characterized in that, The specific steps of generating multiple student groups with different outbound call priorities by performing cluster analysis on students based on the student feature data using a student classification model include: Principal component analysis was used to perform feature dimensionality reduction on the student feature data. Based on the dimensionality-reduced feature data, the student dataset is divided into a preset number of K initial cluster groups using the K-Means clustering algorithm. The clustering effect is evaluated by the silhouette coefficient to determine the optimal number of cluster groups. For each cluster, analyze its business meaning, identify the central features of each cluster, and assign a business label to each cluster based on the average performance of each cluster on key features such as "historical conversion rate", "frequency of proactive inquiries in the past N days", "course expiration reminder time" and "historical connection rate". Generate student groups with different outbound call priorities based on business tags.

5. The intelligent dispatching method for group calls in a call center as described in claim 4, characterized in that, The process of generating student groups with different outbound call priorities based on business tags specifically includes: designating groups with characteristics such as "high conversion rate", "frequent proactive inquiries", and "near course expiration" as high-intent groups and assigning them a first-level outbound call priority; designating groups with characteristics such as "medium conversion rate", "historical inquiry records but recent silence", and "medium connection rate" as groups to be activated and assigning them a second-level outbound call priority; and designating groups with characteristics such as "low conversion rate", "long-term no interaction", and "low historical connection rate" as low-intent groups and assigning them a third-level outbound call priority.

6. The intelligent dispatching method for group calls in a call center as described in claim 5, characterized in that, The specific steps for dynamically generating a set of time windows for outbound calls on the same day based on the current timestamp and preset no-call rules for the education and training industry include: Get the current system timestamp, and parse it to get the weekday attribute of the current date, whether it is a public holiday, and the current specific time; Load the preset do-call rule library for the education and training industry, and load the preset do-call rules from the rule configuration center. The do-call rules include at least the legally mandated do-call periods, the industry-standard do-call periods, special rules for holidays, and personalized do-call rules for students. The daily 24-hour continuous timeline is superimposed on the aforementioned call ban rule base to calculate and remove all call ban periods, resulting in a set of continuous original callable time intervals for the day. The original time window is sliced ​​based on outbound call volume prediction. Historical outbound call data is used to predict the outbound call volume for each time period of the day. The original available outbound call time interval is sliced ​​according to the preset outbound call batch granularity to generate several discrete available outbound call time windows. Each time window has a start time and an end timestamp.

7. The intelligent dispatching method for group calls in a call center as described in claim 6, characterized in that, The specific steps for calculating the real-time connection probability value for each student at the current moment by combining the student group and time window set, and calculating the expected conversion value score based on the real-time connection probability value, include: Get the start and end times of the current outbound call window, and extract the time period attributes of the time window; Load the student's multidimensional feature data and group priority labels; The multidimensional feature data is input into a pre-trained connection probability prediction model, which is a binary classification model based on logistic regression, and outputs the real-time connection probability value of the student in the current time window. The expected conversion value base score of students is extracted from student characteristic data. The expected conversion value base score is obtained by comprehensively quantifying the student's historical order amount, number of renewals, course preference matching degree and current business opportunity stage. The real-time connection probability value is weighted and fused with the expected conversion value base score to calculate the student's expected conversion value score at the current moment.

8. The intelligent dispatching method for group calls in a call center as described in claim 7, characterized in that, The specific steps for calculating the agent capability score based on the real-time status data of the agent team include: Real-time acquisition of multi-dimensional status signals from agent terminals, including current call duration, average call duration over historical periods, average conversion rate of the most recent N calls, number of consecutive unconnected calls, current working hours, and historical matching success rate of similar students. Based on the multidimensional state signals, a seat agent competence evaluation model is constructed, which includes the dimensions of skill proficiency, fatigue, emotional stability, and business matching. The average call duration, average conversion rate of the most recent N calls, and number of consecutive missed calls in the historical period are weighted by a time decay factor to obtain real-time quantitative scores for each dimension. The real-time quantified scores of each dimension are input into a preset weight allocation matrix to calculate the initial capability score of the agent. The initial ability score is dynamically corrected based on the current working hours and the rest interval threshold in the preset call ban rules for the education and training industry, to obtain the final agent ability score.

9. The intelligent dispatching method for group calls in a call center as described in claim 8, characterized in that, The specific steps of dynamically generating a group call task queue based on the conversion value score of each student in the student group, and allocating outbound call tasks in the group call task queue to matching agent terminals based on agent capability scores, include: In real time, all student groups are traversed, the expected conversion value score of each student is extracted, and all students to be called are globally sorted in descending order of expected conversion value score to generate an initial priority queue. Based on a preset set of time windows, a subset of students who can be reached within the current time window is extracted from the initial priority queue to construct a dynamic group call task queue. The system acquires the agent capability scores of all online agent terminals in real time and divides the agents into high-capability agent groups, medium-capability agent groups, and basic agent groups based on the scores. The high-value student tasks with expected conversion value in the dynamic group call task queue, which are at the top of the preset proportion, are forcibly mapped to the high-capacity agent group; the remaining student tasks are distributed to the medium-capacity agent group and the basic agent group according to the load balancing strategy. When an agent in a high-capacity agent group is detected to be idle, a task is immediately retrieved from the head of the dynamic group call task queue and sent to that agent terminal, and the queue status is updated in real time.

10. The intelligent dispatching method for group calls in a call center as described in claim 9, characterized in that, The specific steps for collecting outbound call connection status and student feedback information in real time during group calling, and dynamically adjusting the priority of student groups, include: Establish a real-time signaling monitoring channel to synchronously capture the underlying signaling status code of each outbound call task, and classify the outbound call connection status into valid connection, no answer, busy tone, and disconnected number. When a valid connection is detected, the call content is transcribed in real time to extract the student's feedback information, which includes explicit intention keywords, implicit emotional fluctuation values, specific objection tags, and promised follow-up actions. Based on the student feedback information, the real-time connection rate, effective dialogue rate, and positive feedback percentage of each student group within the current window are statistically analyzed. The dynamic priority correction coefficient for each student group is calculated based on the real-time connection rate, effective dialogue rate, and positive feedback percentage. Using the dynamic priority correction coefficient, the comprehensive priority score of each student group is updated in real time, and the task queue reordering mechanism is triggered.

Images