Missed call priority ranking and distributing method and system based on machine learning

Abstract

The invention relates to a telephone traffic system missed call priority ranking and real-time distribution method and system based on multi-feature machine learning, and the method comprises the following steps: 1) monitoring a missed call event that the ringing of an incoming call is finished and the incoming call is not answered, and generating a missed call record and a trigger signal; 2) obtaining multi-dimensional feature data according to the trigger signal; 3) performing data cleaning and standardization on the multi-dimensional feature data to obtain a structured feature vector; 4) taking the structured feature vector as an input quantity of a pre-trained machine learning model to obtain a priority score; and 5) according to the priority score and the real-time seat state data, executing an intelligent routing decision, and allocating the missed call information to the target seat in real time, thereby realizing real-time, accurate and active allocation of the high-value missed calls, and through a return visit feedback data optimization model, the high-value missed call identification accuracy and recall rate are significantly improved, and the customer loss risk is reduced.

Description

Technical Field

[0001] This invention relates to the field of call center call system technology, specifically to a method and system for prioritizing and real-time allocating missed calls based on multi-feature machine learning. It is applicable to intelligent priority assessment and accurate agent allocation for missed calls, thereby improving follow-up efficiency and customer retention rate. Background Technology

[0002] As a core channel for businesses to establish deep connections with customers, the quality of missed call handling in call centers directly determines the quality of customer experience and the efficiency of business value conversion. Currently, existing missed call handling solutions for call centers can be roughly categorized into three types. However, in practical applications, it has been found that many shortcomings still exist, failing to accurately meet business needs, as detailed below: (1) The first type is the basic sorting mode. These solutions typically generate a work order list using timestamps and push it to customer service agents in a first-in-first-out (FIFO) or manually randomized manner. This relies entirely on the agent's subjective judgment to determine the priority of follow-up calls, lacking both effective identification of the value of missed calls and a scientific sorting basis. For example, the publication CN106341560B, "A Method and System for Judging the Necessity of Call Center Missed Call Follow-up," attempts to determine the necessity of follow-up calls through weighted calculations based on missed call frequency, service content importance, and user category. However, it is essentially a static scoring system under fixed rules, lacking flexibility in weight allocation and failing to capture the dynamic changes and complex relationships in user behavior. For instance, when facing a customer with historically high spending but a recent sharp drop in call frequency, it is impossible to quantify the high churn risk inherent in this situation using a fixed weighted model. This results in a persistently high risk of missing high-value calls, directly leading to lost business opportunities or customer dissatisfaction.

[0003] (2) The second type is the fixed rule improvement mode. This type of solution sorts missed calls by pre-setting a single-dimensional scoring rule (such as assigning high scores to 400 number segments and marking "urgent" for the IVR "complaint" option). However, this method can only identify single-dimensional features and cannot capture complex user behavior patterns. Therefore, it lacks the ability to perceive dynamic risks, which leads to a significant reduction in the accuracy of missed call value judgment. For example, the publication number CN118780808B, "A predictive complaint processing decision support method and system based on machine learning".

[0004] (3) The third type is advanced systems with basic allocation functions. Although these systems are equipped with basic allocation mechanisms such as ACD (Automatic Call Distribution), the allocation rules are preset and fixed, and the allocation is based solely on the agent's idle status, completely ignoring the compatibility between missed call types (such as technical complaints and business inquiries) and agent capabilities.

[0005] Therefore, it can be seen that the above three types of solutions share the following three major common defects: First, the rules are rigid, resulting in low accuracy in value judgment. Whether it is a fixed weighted model, a single-dimensional scoring system, or a static allocation logic, none of them can adapt to the complex and ever-changing user behavior patterns and the composition of missed call value (for example, a customer with a history of high consumption but a recent sharp drop in call frequency has a much higher risk of churn than a customer who calls frequently but never consumes). It is difficult to quantify dynamic risks and potential value, resulting in insufficient accuracy in judging the priority of missed calls.

[0006] Secondly, the processing mode is passive and lacks timeliness. The existing systems have not established an active follow-up mechanism, and agents have to manually select follow-up targets from a massive number of work orders. This leads to serious delays in follow-up response, missing the best communication opportunity, and customers may turn to competitors or post negative reviews because they feel ignored.

[0007] Third, resource allocation is unbalanced and inefficient. The existing system lacks a precise matching mechanism between the value of missed calls and the capabilities of agents. This prevents high-value missed calls from receiving professional and appropriate service support, and also fails to optimize the allocation of human resources (for example, assigning a missed call from a potential high-value business customer to a novice agent handling ordinary inquiries, or assigning a missed call from a technical complaint to a sales agent). Ultimately, this affects overall service efficiency and business conversion benefits.

[0008] Therefore, how to construct a dynamic and flexible missed call value assessment system, a proactive and efficient follow-up triggering mechanism, and a precise and adaptable agent allocation scheme has always been a technical problem that needs to be solved by those skilled in the art. Summary of the Invention

[0009] The purpose of this invention is to address the shortcomings of existing technologies by providing a machine learning-based method and system for prioritizing and allocating missed calls. By collecting multi-dimensional feature data of missed call events (including the caller's historical behavior, current call context, and real-time agent status), and using a pre-trained machine learning model to accurately quantify and rank the priority of missed call events, intelligent routing decisions are executed to achieve real-time, accurate, and proactive allocation of high-value missed calls. This significantly improves follow-up response efficiency and customer satisfaction, effectively reduces the risk of high-value customer churn and negative word-of-mouth spread, and achieves optimal allocation of customer service resources and system self-evolution based on feedback.

[0010] The objective of this invention is achieved through the following approach: A machine learning-based method for prioritizing and allocating missed calls includes the following steps: 1) Monitor missed call events where the incoming call ends and is not answered, and generate a missed call record and a trigger signal containing a unique call identifier; 2) Based on the trigger signal, obtain multi-dimensional feature data, which includes at least the calling user's historical behavior data, current call context information, and real-time agent status data; 3) Perform data cleaning and standardization on multi-dimensional feature data to obtain structured feature vectors; 4) Use the structured feature vectors as input to a pre-trained machine learning model to obtain a priority score that represents the urgency and commercial value of missed calls; 5) Based on priority scores and real-time agent status data, execute intelligent routing decisions to allocate missed call information to target agents in real time.

[0011] Preferably, the calling user's historical behavior data includes at least one of the following: historical call frequency, average call duration, historical number of complaints, customer value level, and historical help-seeking problem type.

[0012] Preferably, the current call context information includes at least one of the following: the caller's number location, the call source, the IVR menu selection path, and the duration of this ringing.

[0013] Preferably, the real-time agent status data includes at least one of the following: currently available agent ID, skill group, proficiency in handling various types of issues, and historical average call duration.

[0014] Preferably, in step 5), the step of assigning missed call information to target agents in real time by performing intelligent routing decisions includes: 5-1) Set a priority threshold and compare the priority score with the priority threshold: If the priority score is less than the priority threshold, the real-time allocation process will not be triggered. If the priority score is greater than or equal to the priority threshold, the real-time allocation process is triggered, proceeding to step 5-2). 5-2) Based on the obtained real-time agent status data, select target agents from the available agents and push the missed call information to the target agent's terminal through the real-time communication interface.

[0015] Preferably, the real-time communication interface includes at least one of the following: a WEBSOCKET component, an SMS alarm interface, and an agent to-do list insertion interface.

[0016] Preferably, it also includes collecting the follow-up results of agents' missed call events in real time, generating feedback data, and using the feedback data to perform incremental learning or periodic retraining of the machine learning model.

[0017] Preferably, a machine learning-based missed call priority ranking and allocation system includes an event triggering module, a data integration module, a real-time prediction module, an intelligent allocation module, and a model optimization module. The event triggering module is used to monitor missed call events where the incoming call ends and is not answered, and to trigger a signal in real time. The data integration module is used to receive trigger signals, initiate parallel queries to multiple data sources, clean and standardize the returned heterogeneous data, and output structured feature vectors. The real-time prediction module is used to load pre-trained machine learning models, receive structured feature vectors, and output priority scores. The intelligent allocation module is used to receive priority scores and real-time agent status data, execute intelligent routing decisions, and push missed call information to the target agent terminal. The model optimization module is used to connect to the agent desktop system, collect feedback data on return visit results, generate incremental update data packages for the model, and input the update data packages into the real-time prediction module to complete model optimization.

[0018] The beneficial effects of this invention are as follows: 1. This invention, through the synergistic application of multi-dimensional data integration and machine learning models, accurately captures complex nonlinear user behavior patterns, effectively solving the problem of low accuracy in identifying high-value missed calls. It realizes the transformation of missed call handling from passive waiting to proactive triggering, significantly improving the timeliness of follow-up response and reducing customer churn and negative reputation risks. At the same time, by establishing a precise matching mechanism between the value of missed calls and agent capabilities, it avoids the problem of resource mismatch, ensuring that high-value missed calls are matched with the optimal agent, significantly improving business conversion efficiency and customer satisfaction, and realizing the upgrade and iteration of customer service resources from experience-driven to data-driven.

[0019] 2. By clearly defining core historical data dimensions such as historical call frequency and average call duration, this invention provides key inputs for machine learning models to characterize the long-term value and demand characteristics of users. It can accurately quantify dynamic user risks (such as the risk of customer churn due to historically high consumption but a recent sharp drop in call frequency), effectively making up for the shortcomings of traditional fixed rules in capturing long-term behavioral trends. At the same time, by using data such as customer value level and historical help-seeking problem types, it can accurately define the differences between high-value customers and ordinary consultation users, ensuring that agent resources are tilted towards high-value business scenarios.

[0020] 3. By clearly identifying the caller's location and IVR menu selection path, this invention can capture the user's current core needs in real time. For example, calls selected by the IVR to "complaint" can be quickly marked as high urgency, avoiding the drawbacks of traditional systems that are slow to respond to immediate needs. It also supplements real-time scenario differences that cannot be covered by historical data (such as the ringing duration reflecting the user's willingness to wait, and the call source distinguishing between business and personal needs), making priority scoring more scenario-specific. At the same time, it can provide agents with key call context information in advance, helping them quickly locate problems, improve follow-up communication efficiency, and thus indirectly optimize customer experience.

[0021] 4. By collecting real-time agent status data, this invention can not only avoid resource mismatch problems, but also ensure that various missed call issues are handled efficiently by professionals by using data such as skill group affiliation, proficiency in handling various problems, and historical average call duration. At the same time, it can map the agent workload status in real time, avoid the imbalance of idle agent load, and improve the overall efficiency of agent resource utilization.

[0022] 5. This invention, by executing intelligent routing decisions, allocates missed call information to target agents in real time. This not only flexibly adapts to diverse business scenarios and avoids low-value missed calls occupying high-priority processing resources, ensuring agents focus on core needs, but also ensures accurate and targeted allocation through the logic of "screening target agents before pushing," reducing interference from invalid pop-up windows and improving agents' work focus. At the same time, the process-oriented allocation mechanism ensures consistency, avoids the randomness and subjectivity of manual allocation, and reduces the risk of lost business opportunities.

[0023] 6. By employing multiple real-time communication interfaces to notify target agents, this invention ensures that missed call information reaches target agent terminals quickly and stably, fully adapts to different agent work scenarios, improves information reach rate, and enhances system compatibility and practical feasibility.

[0024] 7. Experimental verification shows that the system and method described in this invention can significantly improve the accuracy of high-value missed call identification from 35% to 68%, a relative improvement of up to 94%. This indicates that agent resources can accurately and efficiently focus on truly high-value missed call handling scenarios, significantly improving time utilization efficiency and work value output. The high-value missed call recall rate has increased from 46.7% to 90.7%, a relative improvement of 94%, effectively reducing the probability of high-value customers being lost due to missed calls not being responded to in a timely manner, and significantly reducing customer churn risk and potential business losses. The comprehensive evaluation index F1-Score has increased from 0.40 to 0.78, a relative improvement of 95%. That is, the solution of this invention can effectively balance and optimize the identification accuracy and the comprehensive coverage of high-value missed calls, and has outstanding comprehensive performance advantages. Attached Figure Description

[0025] Figure 1 This is a flowchart of the present invention; Figure 2 This is a schematic diagram illustrating the workflow of each module in an embodiment of the present invention. Detailed Implementation

[0026] like Figures 1 to 2 As shown, a machine learning-based missed call priority ranking and allocation system includes an event triggering module, a data integration module, a real-time prediction module, an intelligent allocation module, and a model optimization module. Each module interacts and collaborates through internal APIs and data interfaces, as detailed below: The event triggering module monitors missed calls (calls that have ended ringing but not been answered) and generates trigger signals in real time. The module's hardware relies on the event monitoring interface of the call center system, while the software is developed in Java and deployed on the call center server. Its core function is to monitor missed call events in real time and generate trigger signals containing a unique call ID. After a call ends, the call system generates a call detail record (CDR) containing fields such as call time (a mandatory value), ring time, answer time, hang-up time, and call duration. Missed calls are determined by the business system's built-in logic (ring time is not empty and answer time is empty). Once a missed call occurs, an API interface is called to trigger the scoring process. Long-term experimental verification shows that this module has an average mean time between failures (MTBF) of ≥99.9%, demonstrating high reliability.

[0027] The data integration module receives trigger signals, initiates parallel queries to multiple data sources, and cleans and standardizes the returned heterogeneous data, outputting structured feature vectors. This module is custom-developed based on Spring Boot + MyBatis (e.g., Spring Boot 2.6.3 + MyBatis 3.5.3), supports connections to multiple databases such as MySQL and Oracle, and is configured with calling credentials for the CRM system REST API and the agent status system REST API. It incorporates core logic for data cleaning (e.g., filling null values ​​with the mean) and feature standardization (e.g., using Min-Max normalization for numerical features and one-hot encoding or label encoding for categorical features), ultimately outputting structured feature vectors in JSON format.

[0028] The real-time prediction module loads pre-trained machine learning models, receives structured feature vectors, and outputs priority scores. Deployed on a GPU server (configured with an NVIDIA Tesla T4), it uses TensorFlowServing 2.11 as its model service framework, supports high-concurrency prediction requests of ≥100 per second, and includes a built-in model health check mechanism that automatically triggers alarms and logs when the model prediction error exceeds 5%.

[0029] The intelligent allocation module receives priority scores and real-time agent status data, executes intelligent routing decisions, and pushes missed call information to the target agent's terminal in real time. Developed using the Python Flask framework, this module comprises three core components: a scoring threshold judgment logic, an agent matching algorithm based on a weighted scoring model, and a communication interface with the agent's desktop client. The communication interface supports the WEBSOCKET protocol, and testing shows a push success rate of ≥99.5%, enabling immediate delivery of missed call information.

[0030] The model optimization module connects to the agent desktop system, collects feedback data on follow-up visits, generates incremental update data packages for the model, and inputs these update data packages into the real-time prediction module to complete model optimization. This module can connect to the agent desktop system's MySQL database, uses Apache Airflow to schedule incremental learning tasks, supports model training log recording, version management, and historical parameter backtracking, and can use agent follow-up results (customer satisfaction, problem resolution, conversion rate) as labeled data to feed back into the machine learning model for iterative model development.

[0031] For details regarding the above functional modules, please refer to Table 1: Table 1

[0032] In this embodiment, the above-mentioned missed call priority ranking and allocation system is deployed as a whole in the hardware and software environment of the call center, and the specific configuration is as follows: The hardware environment includes a call center server (Intel Xeon Gold 6248 CPU, 64GB RAM), a GPU prediction server (Intel Xeon Gold 6248 CPU, 64GB RAM, NVIDIA Tesla T4 GPU), and agent terminals (Intel Core i5-12400 CPU, 16GB RAM). All hardware devices are interconnected via Gigabit Ethernet to ensure data transmission speed and stability. Software environment: The operating system is CentOS 7.4, the database is MySQL 8.24 (supporting master-slave replication to ensure data reliability), the CRM system is Salesforce CRM, the call center system is Avaya Aura, the model training framework is Python 3.9 + XGBoost 1.7, and the model service framework is TensorFlow Serving 2.11.

[0033] The method for accurately prioritizing and intelligently matching missed call events with the optimal agent using the aforementioned missed call priority ranking and allocation system includes the following steps: 1) Missed Call Event Triggering and Signal Generation: Configure an event monitoring script in the Avaya Aura call management system to continuously monitor for specific events such as "ringing ended and no answer". When the call detail record (CDR) generated by the call management system meets the missed call judgment condition of "ringing time not empty and answering time empty", the event triggering module generates a missed call record in real time containing core information such as call time, caller ID, and ringing duration. At the same time, a unique call identifier (Call ID, format example "YYYYMMDDHHMMSSXXXX", where XXXX is a random sequence, such as "202405201430251234") is generated. The trigger signal containing the Call ID is sent to the data integration module via an HTTP POST request. The interface address is http: / / data-integration:8080 / trigger, and the interface request timeout is set to 500 milliseconds. This step takes ≤1 second, which can effectively ensure that missed call processing starts immediately and avoid the loss of high-value customers due to response delays.

[0034] 2) Upon receiving the trigger signal, the data integration module immediately initiates parallel queries against multiple data sources, including the CRM system, historical call database, and real-time agent status system, based on the unique call identifier (Call ID). (Thread pool technology is used to achieve multi-task parallelism, avoiding efficiency losses caused by serial queries.) This comprehensively acquires multi-dimensional feature data, specifically including the caller's historical behavior data, current call context information, and real-time agent status data. The specific implementation process is as follows: ① Call the CRM system API (address: "https: / / salesforce-api.com / v1 / customer", token: "abc123") and historical call database to query the caller's historical behavior data, including at least one of the following: historical call frequency (times / week) in the past 3 months, average call duration (minutes), number of historical complaints, customer value level (e.g., Platinum / VIP / Regular), and types of historical assistance issues (e.g., technical failure / billing inquiry / complaints and suggestions), etc. ② Query the current call log to obtain the current call context information, including at least one of the following: the caller's location (province / city), the source of the call (mobile / landline / 400 number), the IVR menu selection path (such as "main menu - bill inquiry - objection complaint"), and the duration of this ringing (seconds); ③ Call the agent status management system API (address http: / / agent-status:8081 / agent) to query real-time agent status data, including at least one of the following: currently available agent ID, skill group (such as technical support, top sales), proficiency in handling various types of issues (e.g., quantified on a scale of 1-5, with 5 being the highest), average call duration over the past month, and the number of tasks currently pending by the agent.

[0035] This step utilizes a parallel query mechanism across multiple data sources to ensure that the entire data integration process (including subsequent cleaning and standardization) takes ≤3 seconds. This process enables the comprehensive collection of multi-dimensional feature data, providing complete and high-quality data support for subsequent model predictions, while also guaranteeing the timeliness and accuracy of the feature data.

[0036] 3) Data Cleaning and Standardization: The heterogeneous data (including numerical, categorical, and textual data) returned by the query in step 2) is standardized according to the built-in logic of the data integration module. For example, data cleaning is first completed by removing null values ​​and correcting outliers; then, categorical features (such as converting customer level "Platinum" to the value "5" and parsing the IVR path JSON as feature encoding) are labeled or one-hot encoded, and numerical features (such as call duration and ringing duration) are normalized using Min-Max (mapped to the [0,1] interval). Finally, a structured feature vector is generated, such as a structured feature vector containing 20-30 features, and output to the real-time prediction module in JSON format via an HTTP POST request. The interface address is http: / / prediction:8501 / v1 / models / misscall:predict. This step takes ≤3 seconds to ensure the data quality and format consistency of the input model.

[0037] 4) Real-time priority score prediction: In order to accurately quantify the urgency and commercial value of missed calls, the structured feature vector generated in step 3) is input into a pre-trained machine learning model (in this embodiment, the gradient boosting tree model LightGBM or XGBoost) for real-time prediction to obtain a priority score that represents the urgency and commercial value of missed calls.

[0038] Specifically, the construction process of this machine learning model is divided into two stages: offline training and online deployment, as detailed below: A. Offline training phase: ① Data collection: The caller data (including customer level and consumption records), missed call records (including call ID, ringing duration, IVR path, and caller's location) for the past year were exported from the CRM system, call system, and agent system, respectively. The follow-up results (including whether the customer was lost, whether they were converted, whether they were converted into business opportunities, and whether the problem was resolved) were collected in total, totaling 120,000 valid missed call samples. ② Feature Engineering: Extract 25 features from the collected sample data, including historical call frequency (number of calls in the past 30 days), average call duration in the past 3 months, cumulative number of historical complaints, customer rating (1-5 points), IVR path coding (e.g., "complaint" path coding is 5), ringing duration, agent skill group matching degree, etc.; perform one-hot coding on categorical features (e.g., call source, problem type), and perform Min-Max normalization on numerical features (e.g., call duration, ringing duration) to ensure uniformity of units among features; ③ Model Training: The 120,000 missed call samples were divided into training and testing sets in a 7:3 ratio. The supervision labels Y were "whether the customer churned within 30 days" and "whether the customer converted into a business opportunity." The XGBoost algorithm was used for model training, with parameters set as learning_rate=0.1, max_depth=6, and n_estimators=200. Parameters were optimized using 5-fold cross-validation to determine the optimal parameter combination. After training, the accuracy on the training set reached 94.2%, and the accuracy on the testing set reached 92.5%. The model file "xgb_misscall_model.bin" was generated. B. Online model deployment: Start TensorFlowServing on the GPU prediction server, load the trained model file, and configure the model inference batch size to 32 to ensure prediction response speed. After receiving the feature vector sent by the data integration module, perform real-time calculation and output a priority score P between 0 and 1 (the larger the P value, the higher the urgency of the missed call and the greater its commercial value). Return the score to the intelligent allocation module, and the response interface address is http: / / intelligent-allocation:8082 / score. This step takes ≤500 milliseconds to predict, which meets the real-time processing requirements.

[0039] 5) Based on the priority score obtained in step 4) and the real-time agent status data obtained in step 2), the intelligent allocation module performs intelligent routing decisions to allocate missed call information to target agents in real time, so as to achieve accurate matching of high-value missed calls with optimal agents. Specific steps include: 5-1) Set a priority threshold in the intelligent allocation module (in this embodiment, the priority threshold is 0.8, which can be dynamically adjusted according to scenarios such as peak business periods and agent resource load), and compare the priority score with the priority threshold: If the priority score P < priority threshold 0.8, it is judged as a low-value missed call, and the real-time allocation process is not triggered. The missed call record is stored in the ordinary missed call work order pool and processed in order when the agent is available. If the priority score P ≥ priority threshold 0.8, the real-time allocation process is immediately triggered, proceeding to step 5-2). 5-2) Based on the real-time agent status data obtained in step 2), call the agent matching algorithm (skill group matching weight 60% + processing proficiency weight 40%) to filter target agents from the currently available agents; for example, for "technical fault type missed calls", prioritize matching available agents with "technical support skill group and proficiency ≥ 4 points" (example agent ID is "Agent001"). The missed call information is then pushed to the target agent's terminal via a real-time communication interface. Specific push methods include: Send a pop-up message to the target agent's desktop client (communication address "ws: / / agent-desktop:8083 / agent001") via the WEBSOCKET component. For example, the message content is "Missed Call ID: 202405201430251234, Customer: Zhang San, Customer Level: Platinum VIP, Problem Type: Technical Fault, Priority Score: 0.85, Caller's Location: XX City, Call Source: Mobile Phone, IVR Selection Path: Main Menu - Technical Support - Fault Reporting, Please Call Back Immediately". Alarm text messages can also be sent to the agent's work phone via the SMS alarm interface, or the missed call task can be directly inserted at the top of the agent's to-do list to ensure that the agent is aware of it immediately. The pushed missed call information and customer summary should include core information such as missed call call ID, customer name, customer level, problem type, priority score, caller's location, call source, and IVR selection path, which can help agents grasp the key points of communication in advance; this step takes ≤2 seconds in total to ensure immediate response to high-value missed calls.

[0040] After the agent completes the follow-up call, in order to continuously improve the prediction accuracy and generalization ability of the machine learning model, the model optimization module performs iterative optimization operations: First, the model optimization module collects real-time feedback data on the return call results from the MySQL database of the agent desktop system, including missed call call ID, whether the call was successfully connected, customer satisfaction score (1-5 points), whether the problem was resolved (yes / no), whether it was converted into a business opportunity (yes / no), whether the customer was lost within 30 days (yes / no), agent handling evaluation, etc. Then, configure the Apache Airflow scheduling task to automatically extract the feedback data of the previous day at 2:00 AM every day, generate an incremental training dataset (to ensure data timeliness), and call the partial_fit interface of the XGBoost algorithm to incrementally update the parameters of the model in the real-time prediction module, avoiding the resource consumption of full training. On the 1st of each month, the system automatically extracts all feedback data for a complete model retraining, generates a new model file, and replaces the old model in the real-time prediction module. This ensures that the model's prediction accuracy continues to improve as business data accumulates, enabling the system to evolve itself.

[0041] To verify the superiority of the system and method of the present invention, comparative experiments were conducted, as follows: a. Experimental Data: The experiment was conducted based on 10,000 anonymized historical missed call records from a large e-commerce customer service center. Each record contained multi-dimensional features such as customer level, historical call frequency, IVR selected path, ringing duration, caller's location, and number of historical complaints. Missed calls that were successfully retained after follow-up or resulted in high-value order conversions were marked as high-value missed calls (1,500 records in total, accounting for 15%), serving as the true label (gold standard) for the experiment.

[0042] b. Resource constraints: Simulating the situation where agent resources are limited in actual business scenarios, only 20% of missed calls (i.e., 2000 calls) will be prioritized for follow-up.

[0043] c. The object of comparison: Rule system (traditional fixed rule solution): The missed calls are scored and ranked according to the industry's typical fixed rules. The specific rules are set as follows: VIP customer level adds 50 points, IVR menu selection of "complaint" path adds 30 points, historical complaint record adds 20 points, ringing time exceeds 20 seconds adds 10 points. Finally, the missed calls are ranked according to the total score. The system of this invention adopts the scheme described in this embodiment, uses the XGBoost (v1.6.2) algorithm, trains a machine learning model based on all 25 multi-dimensional features, the model outputs the priority probability score (value range 0-1) for each missed call, and sorts the missed calls according to the probability score.

[0044] d. Evaluation criteria: Precision, recall, F1-Score (a comprehensive evaluation indicator), and the estimated amount of customer losses to be recovered are used as the core evaluation indicators.

[0045] By sorting 10,000 missed calls using two systems, the top 2,000 missed calls were selected for performance evaluation, and the results are shown in Table 2 below: Table 2

[0046] In summary, under the same agent resource constraints (prioritizing follow-up on 20% of missed calls), the overall performance and business value of the system of this invention far surpass those of traditional fixed-rule systems: its accuracy and recall rates both achieve a significant improvement of 94%, the F1-Score comprehensive evaluation index improves by 95%, and the expected amount of customer loss recovered increases by 94% simultaneously. This fully demonstrates that the missed call priority ranking and real-time allocation mechanism constructed by the multi-feature machine learning model of this invention can accurately capture complex business patterns and dynamic user behavior characteristics that fixed rules cannot identify, efficiently filter high-value missed calls, significantly optimize agent resource allocation efficiency, effectively reduce the risk of high-value customer churn, and improve business conversion efficiency.

[0047] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications made to the present invention by those skilled in the art without departing from the spirit of the present invention shall fall within the protection scope of the present invention.

Claims

1. A method for machine learning based prioritization and assignment of missed calls, the method comprising: Includes the following steps: 1) Monitor missed call events where the incoming call ends and is not answered, and generate a missed call record and a trigger signal containing a unique call identifier; 2) Based on the trigger signal, obtain multi-dimensional feature data, which includes at least the calling user's historical behavior data, current call context information, and real-time agent status data; 3) Perform data cleaning and standardization on multi-dimensional feature data to obtain structured feature vectors; 4) Use the structured feature vectors as input to a pre-trained machine learning model to obtain a priority score that represents the urgency and commercial value of missed calls. 5) Based on priority scores and real-time agent status data, execute intelligent routing decisions to allocate missed call information to target agents in real time.

2. The method according to claim 1, characterized in that, The historical behavior data of the calling user shall include at least one of the following: historical call frequency, average call duration, number of historical complaints, customer value level, and types of historical help requests.

3. The method according to claim 1, characterized in that, The current call context information includes at least one of the following: the caller's number location, the call source, the IVR menu selection path, and the duration of this ringing.

4. The method according to claim 1, characterized in that, The real-time agent status data includes at least one of the following: currently available agent ID, skill group, proficiency in handling various types of issues, and historical average call duration.

5. The method according to claim 1, characterized in that, Step 5), which involves assigning missed call information to target agents in real time by executing intelligent routing decisions, includes the following steps: 5-1) Set a priority threshold and compare the priority score with the priority threshold: If the priority score is less than the priority threshold, the real-time allocation process will not be triggered. If the priority score is greater than or equal to the priority threshold, the real-time allocation process is triggered, proceeding to step 5-2). 5-2) Based on the obtained real-time agent status data, select target agents from the available agents and push the missed call information to the target agent's terminal through the real-time communication interface.

6. The method according to claim 1, characterized in that, The real-time communication interface includes at least one of the following: WEBSOCKET component, SMS alarm interface, and agent to-do list insertion interface.

7. The method according to claim 1, characterized in that, It also includes real-time collection of follow-up results from agents regarding missed calls, generating feedback data, and using this feedback data to incrementally learn or periodically retrain machine learning models.

8. A machine learning-based missed call priority ranking and allocation system, characterized in that, It includes an event triggering module, a data integration module, a real-time prediction module, an intelligent allocation module, and a model optimization module; The event triggering module is used to monitor missed call events where the incoming call ends and is not answered, and to trigger a signal in real time. The data integration module is used to receive trigger signals, initiate parallel queries to multiple data sources, clean and standardize the returned heterogeneous data, and output structured feature vectors. The real-time prediction module is used to load pre-trained machine learning models, receive structured feature vectors, and output priority scores. The intelligent allocation module is used to receive priority scores and real-time agent status data, execute intelligent routing decisions, and push missed call information to the target agent terminal. The model optimization module is used to collect feedback data from follow-up visits, generate incremental update data packages for the model, and input the update data packages into the real-time prediction module to complete model optimization.

Images