Model training, prediction and work order allocation method, electronic equipment, medium and product

Abstract

The invention relates to a model training method, a model prediction method, a work order distribution method, electronic equipment, a medium and a product. According to the embodiment of the invention, the particle swarm is constructed based on the historical call data, and the particle swarm model is trained in combination with the preset loss function based on the historical work order samples represented by the particles in the particle swarm and the corresponding actual post-processing duration labels. The self-adaptive adjustment weight set by the preset loss function can be dynamically adjusted according to the real-time service load (queuing number and work order accumulation amount) during sample generation, so that the predicted post-processing duration output by the particle swarm model after training convergence is more accurate than that of a traditional method; therefore, the agent waiting time and the user queuing time can be reduced, and the work order processing efficiency of the agent is improved.

Description

Technical Field

[0001] This disclosure relates to the field of IT support technology, and in particular to a model training, prediction, and work order allocation method, electronic device, medium, and product. Background Technology

[0002] Post-call processing time refers to the time spent by agents in a call center system processing follow-up tasks after a call ends. Even when the call has ended and no customer is involved, the post-call processing time is still considered part of the interaction because agents are still processing customer inquiries, entering data, and filling out work orders.

[0003] Call post-processing time is a system-defined time limit for agents to process business after a call ends without entering an idle state, provided that the agent has been set to auto-answer. Once the post-processing time exceeds the set limit, the agent will automatically enter an idle state to prepare for another call. Currently, the post-processing time of call center business systems is primarily set uniformly through skill group or organizational dimensions.

[0004] However, this method of uniformly setting the post-call processing time is inefficient because the post-call processing time value is fixed and cannot adapt to changes in agent workload. Summary of the Invention

[0005] To address the aforementioned issues, this disclosure provides a model training, prediction, and work order allocation method, electronic device, medium, and product that can improve the work order processing efficiency of agents.

[0006] According to a first aspect of the present disclosure, a model training method is provided, the method comprising: Based on historical call data from the agent system, a particle swarm is constructed according to a preset time window. Each particle in the particle swarm is used to represent a training sample corresponding to a historical work order and the corresponding call post-processing duration. The features of the training sample include the call post-processing duration of the agent to which the target historical work order belongs, the average call post-processing duration of the work order type of the target historical work order, the call duration of the target historical work order, the number of people in the queue of the agent to which the target historical work order belongs, and the backlog of work orders of the agent to which the target historical work order belongs. The actual post-call processing time of the training samples corresponding to each particle in the particle swarm is used as the label of the training samples. The particle swarm model is trained based on a preset loss function. The preset loss function is determined based on the actual post-call processing time corresponding to each training sample, the predicted post-call processing time corresponding to each training sample, and the adaptive adjustment weight corresponding to each training sample. The adaptive adjustment weight of the target training sample is determined according to the real-time business load of the agent system corresponding to the target training sample. The real-time business load of the target training sample includes the current number of people in the queue and the current number of backlogged work orders of the agent system at the time of the target training sample. When the convergence condition is met, the particle swarm model is output to predict the post-processing time of each agent's real-time work order. The convergence condition includes: the preset loss function converges, the maximum number of iterations is reached, or the global optimal solution is found.

[0007] In one implementation, the predicted call post-processing duration corresponding to the target training sample is obtained based on the feature values ​​and corresponding feature weights of each feature in the target training sample.

[0008] In one embodiment, the training method for the particle swarm model includes: Based on real-time collected sample data, the predicted call post-processing time for the current iteration round is obtained; Based on the difference between the predicted call post-processing time of the current iteration round and the actual call post-processing time, the prediction error of the current iteration round is obtained. Based on the prediction error of the current iteration, the model parameters for the next iteration are adjusted. The model parameters include at least the global information feedback coefficient for call post-processing duration, which is used to adjust the balance between the model's global search capability and local search capability.

[0009] In one implementation, adjusting the model parameters for the next iteration based on the prediction error of the current iteration includes: Initialize the particle positions and particle velocities of the particle swarm, where the particle positions represent a solution to the post-call processing time and the particle velocities represent the iterative update step size of the post-call processing time. Update the learning acceleration factor for the next iteration. The learning acceleration factor includes an individual learning acceleration factor and a group learning acceleration factor. The individual learning acceleration factor represents the degree of dependence of the post-call processing time on the local optimum, and the group learning acceleration factor represents the degree of dependence of the post-call processing time on the global optimum. The particle update speed for the next iteration is adjusted based on the global information feedback coefficient of the call post-processing duration, wherein the global information feedback coefficient of the call post-processing duration is dynamically updated based on the difference between the global optimal prediction error and the prediction error of the current round. The particle position for the next iteration is updated based on the particle position and velocity of the current iteration until the preset loss function converges, thus obtaining the global optimal solution.

[0010] In one implementation, updating the learning acceleration factor for the next iteration includes: Based on the prediction error of the previous iteration and the prediction error of the current iteration, update the individual learning acceleration factor for the next iteration. The population learning acceleration factor for the next iteration is updated based on the global optimal prediction error and the prediction error of the current iteration.

[0011] According to a second aspect of the present disclosure, a method for predicting post-call processing duration is provided, the method comprising: Collect sample data corresponding to the target work order. The sample data includes: the post-call processing time of the agent to which the target historical work order belongs, the average post-call processing time of the work order type of the target historical work order, the call duration of the target work order, the number of people in the queue to which the target work order belongs, and the backlog of work orders to which the target work order belongs. The sample data corresponding to the target work order is input into the prediction model to obtain the predicted call processing duration corresponding to the target work order, wherein the prediction model is trained based on the method described in any one of the first aspects.

[0012] According to a third aspect of the present disclosure, a work order allocation method is provided, the method comprising: Collect sample data corresponding to the target work order. The sample data includes: the post-call processing time of the agent to which the target historical work order belongs, the average post-call processing time of the work order type of the target historical work order, the call duration of the target work order, the number of people in the queue to which the target work order belongs, and the backlog of work orders to which the target work order belongs. The sample data corresponding to the target work order is input into the prediction model to obtain the predicted call processing duration corresponding to the target work order, wherein the prediction model is trained based on the method described in any one of the first aspects; Based on the predicted call processing time corresponding to the target work order, a corresponding agent is assigned to the target work order.

[0013] According to a fourth aspect of the present disclosure, a model training apparatus is provided, the apparatus comprising: The construction module is used to construct a particle swarm based on the historical call data of the agent system according to a preset time window. Each particle in the particle swarm is used to represent the training sample and the corresponding call post-processing duration of a historical work order. The features of the training sample include the call post-processing duration of the agent to which the target historical work order belongs, the average call post-processing duration of the work order type of the target historical work order, the call duration of the target historical work order, the number of people in the queue of the agent to which the target historical work order belongs, and the work order backlog of the agent to which the target historical work order belongs. The training module is used to train the particle swarm model based on a preset loss function, using the actual post-call processing time of the training samples corresponding to each particle in the particle swarm as the label of the training samples. The preset loss function is determined based on the actual post-call processing time corresponding to each training sample, the predicted post-call processing time corresponding to each training sample, and the adaptive adjustment weight corresponding to each training sample. The adaptive adjustment weight of the target training sample is determined according to the real-time business load of the agent system corresponding to the target training sample. The real-time business load of the target training sample includes the current number of people in the queue and the current number of backlogged work orders of the agent system at the time of the target training sample. The output module is used to output the particle swarm model to predict the post-processing time of each agent's real-time work order when the convergence condition is met. The convergence condition includes: the preset loss function converges, the maximum number of iterations is reached, or the global optimal solution is found.

[0014] According to a fifth aspect of the present disclosure, an electronic device is provided, comprising: A memory on which computer programs are stored; A processor for executing the computer program in the memory to implement the method of any one of the first aspects, or to implement the steps of the method of the second or third aspect.

[0015] According to a sixth aspect of the present disclosure, a non-transitory computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, implements the method of any one of the first aspects, or implements the steps of the method of the second or third aspect.

[0016] According to a seventh aspect of the present disclosure, a computer program product is provided, comprising a computer program that, when executed by a processor, implements the method described in any one of the first aspects, or implements the steps of the method described in the second or third aspects.

[0017] The embodiments disclosed herein can achieve the following beneficial effects: This embodiment constructs a particle swarm optimization (PSO) model based on historical call data. The PSO model is trained using historical work order samples represented by the particles in the PSO and their corresponding actual post-processing time labels, combined with a preset loss function. Because the adaptively adjustable weights of the preset loss function can dynamically adjust according to the real-time business load (queue size and work order backlog) at the time of sample generation, the predicted post-processing time output by the converged PSO model is more accurate than traditional methods. Therefore, it can reduce agent waiting time and user queuing time, thereby improving agent work order processing efficiency.

[0018] Other features and advantages of this disclosure will be described in detail in the following detailed description section. Attached Figure Description

[0019] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart of a model training method provided in one embodiment of the present disclosure.

[0020] Figure 2 This is a flowchart of a model training method provided in one embodiment of the present disclosure.

[0021] Figure 3 This is a flowchart of a model training method provided in one embodiment of the present disclosure.

[0022] Figure 4 This is a flowchart of a model training method provided in one embodiment of the present disclosure.

[0023] Figure 5 This is a flowchart of a call post-processing duration prediction method provided in one embodiment of the present disclosure.

[0024] Figure 6 This is a flowchart of a work order allocation method provided in one embodiment of the present disclosure.

[0025] Figure 7 This is a schematic diagram illustrating how the model prediction error changes with the number of iterations, according to an embodiment of this disclosure.

[0026] Figure 8 This is a comparative schematic diagram of the effect of call post-processing duration prediction provided by an embodiment of this disclosure.

[0027] Figure 9 This is a block diagram of a model training device provided in one embodiment of the present disclosure.

[0028] Figure 10 This is a block diagram of an electronic device provided in one embodiment of the present disclosure. Detailed Implementation

[0029] The specific embodiments of this disclosure will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit this disclosure.

[0030] It should be understood that the term "comprising" and its variations as used herein are open-ended, meaning "including but not limited to". The term "based on" means "at least partially based on". The term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments". Definitions of other terms will be given in the following description.

[0031] It should be noted that the concepts of "first," "second," etc., mentioned in this disclosure are used only to distinguish different devices, modules, or units, and are not used to limit the order of functions performed by these devices, modules, or units or their interdependencies. The modifiers "a" and "a plurality of" mentioned in this disclosure are illustrative rather than restrictive, and those skilled in the art should understand that, unless explicitly stated in the context, they should be understood as "one or more." In the description of this disclosure, unless otherwise stated, "a plurality of" means two or more, and other quantifiers are similar; "at least one," "one or more," or similar expressions refer to any combination of these items, including any combination of single or multiple items.

[0032] Although operations or steps are described in a specific order in the accompanying drawings in the embodiments of this disclosure, it should not be construed as requiring these operations or steps to be performed in the specific order or serial order shown, or requiring all of the shown operations or steps to be performed to obtain the desired result. In the embodiments of this disclosure, these operations or steps may be performed serially; they may be performed in parallel; or a portion of these operations or steps may be performed.

[0033] The names of messages or information exchanged between multiple devices in the embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of these messages or information. It is understood that before using the technical solutions disclosed in the embodiments of this disclosure, users should be informed of the types, scope of use, and usage scenarios of the personal information involved in this disclosure in an appropriate manner in accordance with relevant laws and regulations, and user authorization should be obtained.

[0034] First, the application scenarios of this disclosure will be explained. Addressing the problems existing in the prior art, embodiments of this disclosure employ an improved particle swarm optimization algorithm to dynamically predict the post-call processing time of customer service agents based on the real-time business load of the customer service system. This improves the accuracy of post-call processing time prediction, reduces agent waiting time and user queuing time, thereby improving the efficiency of agent work order processing. The following description, in conjunction with specific embodiments, will further illustrate this disclosure.

[0035] Figure 1 This is a flowchart illustrating a model training method provided in one embodiment of this disclosure. Figure 1 As shown in the embodiments of this disclosure, a model training method is provided, which may include the following steps: In step S110, a particle swarm is constructed based on the historical call data of the agent system according to a preset time window.

[0036] In the particle swarm, each particle is used to represent a training sample corresponding to a historical work order and the corresponding call post-processing duration. The features of the training sample include the call post-processing duration of the agent to which the target historical work order belongs, the average call post-processing duration of the work order type of the target historical work order, the call duration of the target historical work order, the number of people in the queue to which the target historical work order belongs, and the work order backlog to which the target historical work order belongs.

[0037] In this step, a particle swarm can be constructed based on historical call data from the agent system, according to a preset time window. Each particle in the swarm represents a training sample corresponding to a historical work order and its corresponding post-call processing time. The characteristics of the training sample corresponding to the target historical work order may include the post-call processing time of the agent to which the target historical work order belongs, the average post-call processing time for the work order type of the target historical work order, the call duration of the target historical work order, the number of people queuing for the agent to which the target historical work order belongs, and the backlog of work orders for the agent to which the target historical work order belongs.

[0038] For example, the post-call processing time of the agent to which the target historical work order belongs can be obtained from the call center database, including the post-processing time required for each agent to process the work order over a period of time. Historical data can provide a benchmark for the efficiency and speed of each agent in processing work orders, extracted through specific fields such as agent ID, date and time. The average post-call processing time of the target historical work order type reflects the average post-processing time of work orders of a specific business type, calculated by reading historical data of work orders of the same business type from the business system database.

[0039] For example, for complaint-type work orders, it's necessary to collect the post-processing time for all complaint-type work orders and then calculate the average. This data reflects the processing complexity and time required for different types of work orders. The call duration of the target historical work order can be obtained in real-time through the call center's automated call allocation system, recording the total time from when the agent receives the ringing event to when the agent receives the hang-up event. This data is a crucial factor affecting post-processing time, as longer calls generally require more time to process related tasks. The number of people in the queue for the agent associated with the target historical work order can be obtained in real-time through the call center's queue management system, reflecting the number of callers currently waiting to be answered, typically including queue length and waiting time. The number of people in the queue is dynamic and needs to be updated in real-time to reflect the current operational status. The backlog of work orders for the agent associated with the target historical work order can also be obtained in real-time through the call center's automated call allocation system.

[0040] Historical call data often contains noise, missing values, and outliers, requiring preprocessing to ensure data quality. First, the raw historical call data is cleaned to remove duplicates and handle missing and outliers. Then, the feature data is normalized to eliminate the influence of different units of measurement, ensuring that all features related to call post-processing time participate in subsequent model training and optimization under a unified unit of measurement. Next, training samples are constructed from the processed historical call data according to a specific time window. Each sample consists of input features (such as historical post-processing time, average post-processing time for each work order type, call duration, number of people in the queue, and work order backlog) and the corresponding actual post-processing time. The processed data is then used for particle swarm optimization model training and prediction.

[0041] In step S120, the actual post-call processing time of the training samples corresponding to each particle in the particle swarm is used as the label of the training samples. The particle swarm model is trained based on a preset loss function. The preset loss function is determined based on the actual post-call processing time corresponding to each training sample, the predicted post-call processing time corresponding to each training sample, and the adaptive adjustment weight corresponding to each training sample. The adaptive adjustment weight of the target training sample is determined according to the real-time business load of the agent system corresponding to the target training sample. The real-time business load of the target training sample includes the current number of people in the queue and the current number of backlogged work orders of the agent system at the time of the target training sample.

[0042] In this step, the actual call post-processing time of the training sample corresponding to each particle in the particle swarm is used as the label of the training sample. The particle swarm model is trained based on a preset loss function, which is determined based on the actual call post-processing time corresponding to each training sample, the predicted call post-processing time corresponding to each training sample, and the adaptive adjustment weight corresponding to each training sample.

[0043] The adaptive adjustment weights of the target training samples are determined based on the real-time business load of the agent system corresponding to the target training samples. The real-time business load of the target training samples includes the current number of people in the queue and the current backlog of work orders at the current time of the target training sample. For example, a preset loss function is used. It can be obtained from the following formula: Formula 1 Where n is the number of work order samples, and n is not zero. To adaptively adjust the weights, Let i be the actual post-processing time for the i-th sample. Let t be the post-processing time for the prediction of the i-th sample, and t be the t-th iteration.

[0044] Adaptive weight adjustment It can be obtained from the following formula: Formula 2 in, For the real-time business load of the agent system in the t-th iteration, As an adjustment factor, it can typically range from 0.1 to 1.0, depending on the system's requirements for real-time data sensitivity.

[0045] In step S130, when the convergence condition is met, the particle swarm model is output to predict the post-processing time of each agent's real-time work order. The convergence condition includes: the preset loss function converges, the maximum number of iterations is reached, or a global optimal solution is found.

[0046] In this step, when the convergence conditions are met, a particle swarm model is output to predict the post-processing time of each agent's real-time work order. The convergence conditions include: convergence of the preset loss function, reaching the maximum number of iterations, or finding the global optimum.

[0047] This embodiment constructs a particle swarm optimization (PSO) model based on historical call data. The PSO model is trained using historical work order samples represented by the particles in the PSO and their corresponding actual post-processing time labels, combined with a preset loss function. Because the adaptively adjustable weights of the preset loss function can dynamically adjust according to the real-time business load (queue size and work order backlog) at the time of sample generation, the predicted post-processing time output by the converged PSO model is more accurate than traditional methods. Therefore, it can reduce agent waiting time and user queuing time, thereby improving agent work order processing efficiency.

[0048] In one possible implementation, the predicted call post-processing duration corresponding to the target training sample is obtained based on the feature values ​​and corresponding feature weights of each feature in the target training sample. For example, the predicted call post-processing duration corresponding to the target training sample... It can be obtained from the following formula: Formula 3 in, Let j be the j-th feature value of the work order sample. is the feature weight corresponding to the j-th feature value of the work order sample, and m is the number of feature values ​​of the sample. In this embodiment, m is 5.

[0049] Figure 2 This is a flowchart illustrating a model training method provided in one embodiment of this disclosure. Figure 2 As shown in the embodiments of this disclosure, a training method for the particle swarm model is provided, which may include the following steps: In step S210, the predicted call post-processing time for the current iteration round is obtained based on the real-time collected sample data.

[0050] In this step, the predicted call post-processing duration for the current iteration round is obtained based on the real-time collected sample data. For example, the method for obtaining the predicted call post-processing duration for the current iteration round can refer to Formula 3.

[0051] In step S220, the prediction error of the current iteration round is obtained based on the difference between the predicted call post-processing time of the current iteration round and the actual call post-processing time.

[0052] In this step, the prediction error for the current iteration round is obtained based on the difference between the predicted call post-processing time and the actual call post-processing time in the current iteration round. For example, the prediction error in the current iteration round. It can be obtained from the following formula: Formula 4 in, Let i be the current actual post-call processing time for the i-th work order sample. Let n be the predicted post-call processing time for the current iteration round of the i-th work order sample, where n is the number of work order samples and is not zero.

[0053] In step S230, the model parameters for the next iteration are adjusted based on the prediction error of the current iteration. The model parameters include at least the global information feedback coefficient for call post-processing duration, which is used to adjust the balance between the model's global search capability and local search capability.

[0054] In this step, the prediction error of the current iteration round is used as a basis. Adjust the model parameters for the next iteration. The model parameters should include at least the global information feedback coefficient γ for the post-call processing duration. The global information feedback coefficient γ for the post-call processing duration is used to adjust the balance between the model's global search capability and local search capability.

[0055] Figure 3 This is a flowchart illustrating a model training method provided in one embodiment of this disclosure. Figure 3 As shown, adjusting the model parameters for the next iteration based on the prediction error of the current iteration may include the following steps: In step S310, the particle positions and particle velocities of the particle swarm are initialized. The particle positions represent a solution for the post-call processing time, and the particle velocities represent the iterative update step size of the post-call processing time.

[0056] In this step, the particle positions and velocities of the particle swarm are initialized. The particle position represents a solution for the call post-processing duration, and the particle velocity represents the iterative update step size of the call post-processing duration. For example, according to Equation 3, each predicted call post-processing duration corresponds to a feature weight vector, meaning each particle position can also correspond to a solution for the feature weight vector. Therefore, the weight vector of the i-th particle in the initial iteration can be initialized first. , For example, the weight vector of the i-th particle in the initial iteration It can be represented as The initial feature weights can be randomly set within the interval [0,1] or set based on business experience.

[0057] The initial velocity vector of the i-th particle It can be represented as , Typically, the value is in the range of [-0.1, 0.1] or [-0.05, 0.05] to ensure smooth adjustment of feature weights in the initial stage. Particle velocity is used to control the feature weight update magnitude, i.e., the iteration update step size.

[0058] Initial prediction call post-processing time for the i-th particle It can be obtained from the following formula: Formula 5 in, ~ Let be the feature value of the i-th particle (work order sample). This prediction result serves as the initial reference for algorithm iteration, ensuring that the model can be optimized from a reasonable starting point under different business scenarios.

[0059] In step S320, the learning acceleration factor for the next iteration is updated. The learning acceleration factor includes an individual learning acceleration factor and a group learning acceleration factor. The individual learning acceleration factor represents the degree of dependence of the post-call processing time on the local optimum, and the group learning acceleration factor represents the degree of dependence of the post-call processing time on the global optimum.

[0060] In this step, the learning acceleration factor for the next iteration is updated. The learning acceleration factor includes an individual learning acceleration factor and a group learning acceleration factor. The individual learning acceleration factor represents the dependence of the post-call processing time on the local optimum, while the group learning acceleration factor represents the dependence of the post-call processing time on the global optimum. For example, the individual learning acceleration factor for the next iteration can be updated based on the prediction error of the previous iteration and the prediction error of the current iteration. The group learning acceleration factor for the next iteration can be updated based on the global optimum prediction error and the prediction error of the current iteration.

[0061] In step S330, the particle update speed of the next iteration is adjusted according to the global information feedback coefficient of the call post-processing duration, wherein the global information feedback coefficient of the call post-processing duration is dynamically updated according to the difference between the global optimal prediction error and the prediction error of the current round.

[0062] In this step, the update rate of particle i in the next iteration is adjusted based on the global information feedback coefficient γ of the post-call processing duration. Among them, the global information feedback coefficient γ of the post-call processing time is based on the global optimal prediction error. The difference between the prediction error and the current round Dynamic updates. For example, the update rate of particle i in the next iteration. It can be obtained from the following formula: Formula 6 in, This is the inertia weight, typically ranging from 0.6 to 0.9. Let be the update rate of particle i in the k-th iteration. Let be the update rate of particle i in the (k+1)th iteration. For the updated individual learning accelerator factor, For the updated group learning accelerator, For the local information feedback coefficient of post-call processing time, This is the global information feedback coefficient for post-call processing time. This represents the historical optimal solution (local optimal solution) for particle i, i.e., the historical optimal post-call processing time for the agent. The global optimal solution for particle swarm optimization is the system's globally optimal prediction of the post-call processing time. This represents the current position of particle i, i.e., the current predicted post-call processing time.

[0063] The global information feedback coefficient γ for post-call processing duration is based on the global optimal prediction error. Prediction error compared to the current round The formula for dynamically updating the difference is as follows: Formula 7 in, This is the adjustment coefficient for post-call processing time, typically ranging from 0.05 to 0.2. For updating the global information feedback coefficient of post-call processing time, For the unupdated global information feedback coefficient of call post-processing duration, where n is the number of particles. This represents the optimal prediction error for the nth particle during the iteration process.

[0064] The global information feedback coefficient is dynamically adjusted by using the difference between the global optimal prediction error and the prediction error of the current round. This allows particles to achieve a balance between relying on their own local experience and the global optimal experience of the swarm, thereby optimizing the prediction accuracy of post-processing time. During busy periods, such as peak call times or a surge in work orders, the dependence of particles on the swarm's historical optimal prediction parameters is increased (by increasing...). This allows the model to approach the overall optimal prediction result more quickly, thereby helping the scheduling system to rationally allocate agent resources and reduce the risk of work order backlog; while when agent operations are relatively stable and call volume is relatively stable, it enhances the adaptability of particles to their own historical prediction experience (reducing...). This allows the model to fine-tune the prediction parameters for individual agents or specific work order types, improving the accuracy of post-processing time prediction for individual agents, thereby optimizing scheduling, improving service quality and customer satisfaction.

[0065] In step S340, the particle position for the next iteration is updated based on the particle position and particle velocity of the current iteration until the preset loss function converges, thus obtaining the global optimal solution.

[0066] In this step, based on the particle position of the current iteration round and updated particle velocities Update particle positions for the next iteration This continues until the preset loss function converges, yielding the global optimum. For example, the particle position in the next iteration... It can be obtained from the following formula: Formula 8 in, It can be obtained from formula 6.

[0067] After each iteration, the local and global optima are updated by comparing the current call post-processing duration prediction error with the historical best prediction error (local optimum), ensuring that the model always optimizes in the direction of minimizing prediction error. If the current call post-processing duration prediction error is less than the historical best prediction error, the current call post-processing duration prediction error is updated to the current local optimum. If the current local optimum is less than the current global optimum, the current local optimum is updated to the global optimum, until the preset loss function converges and the global optimum is obtained.

[0068] In one possible implementation, iteration can be stopped when a preset maximum number of iterations is reached, and the current globally optimal solution can be output as the global optimal solution. This can prevent excessive iteration and reduce algorithm efficiency.

[0069] Figure 4 This is a flowchart illustrating a model training method provided in one embodiment of this disclosure. Figure 4 As shown in the embodiments of this disclosure, updating the learning acceleration factor for the next iteration may include the following steps: In step S410, the individual learning acceleration factor for the next iteration is updated based on the prediction error of the previous iteration and the prediction error of the current iteration.

[0070] In this step, the prediction error from the previous iteration is used as a basis. Prediction error with the current iteration round Update the individual learning acceleration factor for the next iteration. For example, the individual learning acceleration factor for the next iteration. It can be obtained from the following formula: Formula 9 in, For the unupdated individual learning accelerator factor, This is the adjustment coefficient for the individual learning acceleration factor, with a value ranging from 0.1 to 0.5.

[0071] Formula 10 in, Let i be the current actual post-call processing time for the i-th work order sample (particle). Let n be the predicted post-call processing time of the i-th work order sample (particle) in the previous iteration, and n be the number of work order samples (particles), which is not zero.

[0072] In step S420, the group learning acceleration factor for the next iteration is updated based on the global optimal prediction error and the prediction error of the current iteration.

[0073] In this step, based on the global optimal prediction error Prediction error with the current iteration round Update the group learning acceleration factor for the next iteration. For example, the group learning acceleration factor for the next iteration. It can be obtained from the following formula: Formula 11 in, For the unupdated group learning accelerator, This is the adjustment coefficient for the group learning acceleration factor, with a value ranging from 0.1 to 0.5.

[0074] The individual learning acceleration factor C1 represents the degree to which the predicted call post-processing time depends on the local (individual) optimal solution, while the group learning acceleration factor C2 represents the degree to which the predicted call post-processing time depends on the global (group) optimal solution. When the prediction error (fitness) changes little, increasing C1 strengthens the model's dependence on the individual's historical optimal solution, improving local search ability and prediction accuracy. When the prediction error (fitness) changes significantly, increasing C2 strengthens the model's dependence on the global optimal solution, thereby improving global search ability and prediction accuracy.

[0075] Figure 5 This is a flowchart illustrating a method for predicting post-call processing time according to an embodiment of this disclosure. Figure 5 As shown in the embodiments of this disclosure, a method for predicting post-call processing duration is provided, the method including the following steps: In step S510, sample data corresponding to the target work order is collected. The sample data includes: the post-call processing time of the agent to which the target historical work order belongs, the average post-call processing time of the work order type of the target historical work order, the call duration of the target work order, the number of people in the queue to which the target work order belongs, and the backlog of work orders to which the target work order belongs.

[0076] In this step, sample data corresponding to the target work order is collected. The sample data includes: the post-call processing time of the agent to which the target historical work order belongs, the average post-call processing time of the work order type of the target historical work order, the call duration of the target work order, the number of people in the queue to which the target work order belongs, and the backlog of work orders to which the target work order belongs.

[0077] In step S520, the sample data corresponding to the target work order is input into the prediction model to obtain the predicted call processing time corresponding to the target work order.

[0078] In this step, the sample data corresponding to the target work order is input into the prediction model to obtain the predicted call processing time corresponding to the target work order. The prediction model is trained based on any of the above model training methods.

[0079] Figure 6 This is a flowchart illustrating a work order allocation method according to an embodiment of this disclosure. Figure 6 As shown in the figure, this disclosure provides a work order allocation method, which may include the following steps: In step S610, sample data corresponding to the target work order is collected. The sample data includes: the post-call processing time of the agent to which the target historical work order belongs, the average post-call processing time of the work order type of the target historical work order, the call duration of the target work order, the number of people in the queue to which the target work order belongs, and the backlog of work orders to which the target work order belongs.

[0080] In step S620, the sample data corresponding to the target work order is input into the prediction model to obtain the predicted call processing time corresponding to the target work order.

[0081] The implementation methods of the above steps S610 to S620 are the same as those of steps S510 to S520, and will not be repeated here.

[0082] In step S630, based on the predicted call processing time corresponding to the target work order, a corresponding agent is assigned to the target work order.

[0083] In this step, the agent's current status changes from "in a call" to "in a processing state" when the call is connected. The processing state duration is the predicted post-call processing time. After the predicted post-call processing time ends, the agent enters an idle state, waiting for the next call to be assigned. The call center then assigns a new work order to the customer service agent based on the predicted post-call processing time. The predicted post-call processing time is obtained based on the post-call processing time prediction method described above.

[0084] Figure 7 This is a schematic diagram illustrating the change of model prediction error with the number of iterations according to an embodiment of this disclosure. Through post-processing time optimization training using a call post-processing time algorithm, the change of the optimal individual fitness (prediction error) with the number of evolutionary (iterative) generations is shown below. Figure 7 As shown in the figure, the improved algorithm's fitness underwent three main stages during the optimization of prediction post-processing time: Significant fluctuations occurred initially: At the outset, the fitness value for call post-processing duration was high (indicating significant prediction errors or resource waste), and the fluctuations were substantial. This means that in the initial stage, the algorithm's weighting of historical and real-time business data loads for each agent was unstable, potentially leading to large deviations in call post-processing duration predictions or that the resource scheduling scheme did not fully utilize historical experience.

[0085] Rapid Decrease and Convergence: The fitness value for post-call processing time decreases rapidly in this stage, indicating that the algorithm has found more effective feature weights in the short term. After combining agent characteristics, work order type, and real-time call data, the system quickly improves the accuracy of post-call processing time prediction.

[0086] The amplitude of oscillations decreases in the later stages: As the iteration progresses, the fitness curve of the post-call processing time fluctuates slightly within a relatively low range, indicating that the algorithm has approached the global optimum or suboptimal solution and only needs to be fine-tuned.

[0087] In summary, the figure demonstrates that by using an improved particle swarm optimization algorithm for call post-processing duration, the system can find an effective prediction scheme for call post-processing duration within a shorter number of iterations, and then steadily fine-tune it in the later stages, enabling the prediction to maintain high accuracy and robustness in the business environment, thereby improving overall operational efficiency and customer satisfaction.

[0088] Figure 8 This is a comparative schematic diagram illustrating the effect of predicting post-call processing time according to an embodiment of this disclosure. For example... Figure 8 As shown, there is a significant gap between traditional methods and actual observation data; while the original particle swarm optimization algorithm improves the prediction effect to some extent, it still has some deviation from the actual post-processing time data.

[0089] In contrast, the improved and optimized particle swarm optimization (PSO) algorithm for predicting call post-processing duration shows a high degree of consistency with the actual data in both overall trend and numerical distribution. The error between the predicted values ​​and the actual data is significantly reduced, demonstrating better prediction accuracy and stability. This method not only significantly improves the accuracy of call post-processing duration prediction but also provides more reliable data support for scheduling decisions and resource allocation in actual business operations, thereby contributing to improved overall operational efficiency of customer service centers.

[0090] Figure 9 This is a block diagram of a model training apparatus provided in one embodiment of this disclosure. Figure 9 As shown, this disclosure provides a model training device 900, which may include the following modules: The construction module 910 is used to construct a particle swarm based on the historical call data of the agent system according to a preset time window. Each particle in the particle swarm is used to represent the training sample and the corresponding call post-processing duration of a historical work order. The features of the training sample include the call post-processing duration of the agent to which the target historical work order belongs, the average call post-processing duration of the work order type of the target historical work order, the call duration of the target historical work order, the number of people in the queue of the agent to which the target historical work order belongs, and the work order backlog of the agent to which the target historical work order belongs.

[0091] The training module 920 is used to train a particle swarm model based on a preset loss function, using the actual post-call processing time of the training samples corresponding to each particle in the particle swarm as the label of the training samples. The preset loss function is determined based on the actual post-call processing time corresponding to each training sample, the predicted post-call processing time corresponding to each training sample, and the adaptive adjustment weight corresponding to each training sample. The adaptive adjustment weight of the target training sample is determined according to the real-time business load of the agent system corresponding to the target training sample. The real-time business load of the target training sample includes the current number of people in the queue and the current number of backlogged work orders of the agent system at the time of the target training sample.

[0092] The output module 930 is used to output the particle swarm model to predict the post-processing time of each agent's real-time work order when the convergence condition is met. The convergence condition includes: the preset loss function converges, the maximum number of iterations is reached, or the global optimal solution is found.

[0093] In one possible implementation, the predicted call post-processing duration corresponding to the target training sample is obtained based on the feature values ​​and corresponding feature weights of each feature in the target training sample.

[0094] In one possible implementation, the training module 920 is further configured to: Based on real-time collected sample data, the predicted call post-processing time for the current iteration round is obtained; Based on the difference between the predicted call post-processing time of the current iteration round and the actual call post-processing time, the prediction error of the current iteration round is obtained. Based on the prediction error of the current iteration, the model parameters for the next iteration are adjusted. The model parameters include at least the global information feedback coefficient for call post-processing duration, which is used to adjust the balance between the model's global search capability and local search capability.

[0095] In one possible implementation, the training module 920 is further configured to: Initialize the particle positions and particle velocities of the particle swarm, where the particle positions represent a solution to the post-call processing time and the particle velocities represent the iterative update step size of the post-call processing time. Update the learning acceleration factor for the next iteration. The learning acceleration factor includes an individual learning acceleration factor and a group learning acceleration factor. The individual learning acceleration factor represents the degree of dependence of the post-call processing time on the local optimum, and the group learning acceleration factor represents the degree of dependence of the post-call processing time on the global optimum. The particle update speed for the next iteration is adjusted based on the global information feedback coefficient of the call post-processing duration, wherein the global information feedback coefficient of the call post-processing duration is dynamically updated based on the difference between the global optimal prediction error and the prediction error of the current round. The particle position for the next iteration is updated based on the particle position and velocity of the current iteration until the preset loss function converges, thus obtaining the global optimal solution.

[0096] In one possible implementation, the training module 920 is further configured to: Based on the prediction error of the previous iteration and the prediction error of the current iteration, update the individual learning acceleration factor for the next iteration. The population learning acceleration factor for the next iteration is updated based on the global optimal prediction error and the prediction error of the current iteration.

[0097] In one possible implementation, the output module 930 is further configured to: Collect sample data corresponding to the target work order. The sample data includes: the post-call processing time of the agent to which the target historical work order belongs, the average post-call processing time of the work order type of the target historical work order, the call duration of the target work order, the number of people in the queue to which the target work order belongs, and the backlog of work orders to which the target work order belongs. The sample data corresponding to the target work order is input into the prediction model to obtain the predicted call processing duration corresponding to the target work order, wherein the prediction model is trained based on the method of any one of claims 1-5.

[0098] In one possible implementation, the device 900 further includes a distribution module for: Collect sample data corresponding to the target work order. The sample data includes: the post-call processing time of the agent to which the target historical work order belongs, the average post-call processing time of the work order type of the target historical work order, the call duration of the target work order, the number of people in the queue to which the target work order belongs, and the backlog of work orders to which the target work order belongs. The sample data corresponding to the target work order is input into the prediction model to obtain the predicted call processing duration corresponding to the target work order, wherein the prediction model is trained based on the method of any one of claims 1-5; Based on the predicted call processing time corresponding to the target work order, a corresponding agent is assigned to the target work order.

[0099] Regarding the apparatus in the above embodiments, the specific manner in which each module performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated upon here.

[0100] This embodiment of the disclosure constructs a particle swarm based on historical call data using a construction module. A training module trains the particle swarm model using historical work order samples represented by particles in the particle swarm and their corresponding actual post-processing time labels, combined with a preset loss function. This model is used to predict the post-processing time of real-time work orders for agents. An output module outputs the predicted call post-processing time. Because the adaptive adjustment weights set by the preset loss function can dynamically adjust according to the real-time business load (queue size and work order backlog) at the time of sample generation, the predicted post-processing time output by the converged particle swarm model is more accurate than traditional methods. Therefore, it can reduce agent waiting time and user queuing time, thereby improving the work order processing efficiency of agents.

[0101] Figure 10 This is a block diagram of an electronic device provided in one embodiment of this disclosure. For example... Figure 10 As shown, the electronic device 1000 may include: a processor 1001 and a memory 1002. The electronic device 1000 may also include one or more of a multimedia component 1003, an input / output (I / O) interface 1004, and a communication component 1005.

[0102] The processor 1001 controls the overall operation of the electronic device 1000 to complete all or part of the steps in the aforementioned model training method. The memory 1002 stores various types of data to support the operation of the electronic device 1000. This data may include, for example, instructions for any application or method operating on the electronic device 1000, and application-related data such as contact data, sent and received messages, images, audio, video, etc. The memory 1002 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. Multimedia component 1003 may include a screen and an audio component. The screen may be, for example, a touchscreen, and the audio component is used to output and / or input audio signals. For example, the audio component may include a microphone for receiving external audio signals. The received audio signals may be further stored in memory 1002 or transmitted via communication component 1005. The audio component also includes at least one speaker for outputting audio signals. I / O interface 1004 provides an interface between processor 1001 and other interface modules, such as a keyboard, mouse, buttons, etc. These buttons may be virtual or physical buttons. Communication component 1005 is used for wired or wireless communication between the electronic device 1000 and other devices. Wireless communication, such as Wi-Fi, Bluetooth, Near Field Communication (NFC), 2G, 3G, 4G, NB-IoT, eMTC, or other 5G technologies, or combinations thereof, is not limited here. Therefore, the corresponding communication component 1005 may include: a Wi-Fi module, a Bluetooth module, an NFC module, etc.

[0103] In an exemplary embodiment, the electronic device 1000 may be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to perform the model training method described above, the call post-processing duration prediction method described above, or the work order allocation method described above.

[0104] In another exemplary embodiment, a computer-readable storage medium including program instructions is also provided, which, when executed by a processor, implement the steps of the model training method described above. For example, the computer-readable storage medium may be the memory 1002 including program instructions, which may be executed by the processor 1001 of the electronic device 1000 to complete the model training method, the post-call processing duration prediction method, or the work order allocation method described above.

[0105] In another exemplary embodiment, a computer program product is also provided, the computer program product comprising a computer program executable by a programmable device, the computer program having code portions for executing the above-described model training method, or executing the above-described call post-processing duration prediction method, or executing the above-described work order allocation method when executed by the programmable device.

[0106] The preferred embodiments of this disclosure have been described in detail above with reference to the accompanying drawings. However, this disclosure is not limited to the specific details of the above embodiments. Within the scope of the technical concept of this disclosure, various simple modifications can be made to the technical solutions of this disclosure, and these simple modifications all fall within the protection scope of this disclosure.

[0107] It should also be noted that the various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction. In order to avoid unnecessary repetition, this disclosure will not describe the various possible combinations separately.

[0108] Furthermore, various different embodiments of this disclosure can be combined in any way, as long as they do not violate the spirit of this disclosure, they should also be regarded as the content disclosed in this disclosure.

Claims

1. A model training method, characterized in that, The method includes: Based on historical call data from the agent system, a particle swarm is constructed according to a preset time window. Each particle in the particle swarm is used to represent a training sample corresponding to a historical work order and the corresponding call post-processing duration. The features of the training sample include the call post-processing duration of the agent to which the target historical work order belongs, the average call post-processing duration of the work order type of the target historical work order, the call duration of the target historical work order, the number of people in the queue of the agent to which the target historical work order belongs, and the backlog of work orders of the agent to which the target historical work order belongs. The actual post-call processing time of the training samples corresponding to each particle in the particle swarm is used as the label of the training samples. The particle swarm model is trained based on a preset loss function. The preset loss function is determined based on the actual post-call processing time corresponding to each training sample, the predicted post-call processing time corresponding to each training sample, and the adaptive adjustment weight corresponding to each training sample. The adaptive adjustment weight of the target training sample is determined according to the real-time business load of the agent system corresponding to the target training sample. The real-time business load of the target training sample includes the current number of people in the queue and the current number of backlogged work orders of the agent system at the time of the target training sample. When the convergence condition is met, the particle swarm model is output to predict the post-processing time of each agent's real-time work order. The convergence condition includes: the preset loss function converges, the maximum number of iterations is reached, or the global optimal solution is found.

2. The method according to claim 1, characterized in that, The predicted post-call processing time corresponding to the target training sample is obtained based on the feature values ​​and corresponding feature weights of each feature in the target training sample.

3. The method according to claim 1, characterized in that, The training method for the particle swarm optimization model includes: Based on real-time collected sample data, the predicted call post-processing time for the current iteration round is obtained; Based on the difference between the predicted call post-processing time of the current iteration round and the actual call post-processing time, the prediction error of the current iteration round is obtained. Based on the prediction error of the current iteration, the model parameters for the next iteration are adjusted. The model parameters include at least the global information feedback coefficient for call post-processing duration, which is used to adjust the balance between the model's global search capability and local search capability.

4. The method according to claim 3, characterized in that, The step of adjusting the model parameters for the next iteration based on the prediction error of the current iteration includes: Initialize the particle positions and particle velocities of the particle swarm, where the particle positions represent a solution to the post-call processing time and the particle velocities represent the iterative update step size of the post-call processing time. Update the learning acceleration factor for the next iteration. The learning acceleration factor includes an individual learning acceleration factor and a group learning acceleration factor. The individual learning acceleration factor represents the degree of dependence of the post-call processing time on the local optimum, and the group learning acceleration factor represents the degree of dependence of the post-call processing time on the global optimum. The particle update speed for the next iteration is adjusted based on the global information feedback coefficient of the call post-processing duration, wherein the global information feedback coefficient of the call post-processing duration is dynamically updated based on the difference between the global optimal prediction error and the prediction error of the current round. The particle position for the next iteration is updated based on the particle position and velocity of the current iteration until the preset loss function converges, thus obtaining the global optimal solution.

5. The method according to claim 4, characterized in that, The update of the learning acceleration factor for the next iteration includes: Based on the prediction error of the previous iteration and the prediction error of the current iteration, update the individual learning acceleration factor for the next iteration. The population learning acceleration factor for the next iteration is updated based on the global optimal prediction error and the prediction error of the current iteration.

6. A method for predicting post-call processing duration, characterized in that, The method includes: Collect sample data corresponding to the target work order. The sample data includes: the post-call processing time of the agent to which the target historical work order belongs, the average post-call processing time of the work order type of the target historical work order, the call duration of the target work order, the number of people in the queue to which the target work order belongs, and the backlog of work orders to which the target work order belongs. The sample data corresponding to the target work order is input into the prediction model to obtain the predicted call processing duration corresponding to the target work order, wherein the prediction model is trained based on the method of any one of claims 1-5.

7. A work order allocation method, characterized in that, The method includes: Collect sample data corresponding to the target work order. The sample data includes: the post-call processing time of the agent to which the target historical work order belongs, the average post-call processing time of the work order type of the target historical work order, the call duration of the target work order, the number of people in the queue to which the target work order belongs, and the backlog of work orders to which the target work order belongs. The sample data corresponding to the target work order is input into the prediction model to obtain the predicted call processing duration corresponding to the target work order, wherein the prediction model is trained based on the method of any one of claims 1-5; Based on the predicted call processing time corresponding to the target work order, a corresponding agent is assigned to the target work order.

8. An electronic device, characterized in that, include: A memory on which computer programs are stored; A processor for executing the computer program in the memory to implement the steps of the method as claimed in any one of claims 1-5, or to implement the steps of the method as claimed in claim 6 or 7.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the method as described in any one of claims 1-5, or the steps of the method as described in claim 6 or 7.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1-5, or the steps of the method as described in claim 6 or 7.

Images