Call center customer behavior simulation system based on deep learning

Abstract

The application discloses a call center customer behavior simulation system based on deep learning, relates to the technical field of call center customer behavior simulation, and performs behavior feature coding processing on a customer behavior sequence, extracts a multi-dimensional behavior feature vector to form a customer behavior feature set, and constructs a deep learning customer behavior prediction model according to the customer behavior feature set, and obtains a customer behavior transition probability matrix through training; obtains the running state parameters of a to-be-simulated call center and constructs a call center environment state model, couples the customer behavior transition probability matrix with the environment state model to generate a customer behavior simulation model; inputs an initial customer incoming event, generates a subsequent behavior event sequence according to the behavior transition probability, and forms a simulated customer behavior track; and according to the simulated customer behavior track, the number of times of repeated incoming calls of customers, the queue abandonment rate and the manual transfer rate are counted, so that a call center customer behavior simulation result is obtained, and the accuracy of call center operation analysis and service strategy evaluation is improved.

Description

Technical Field

[0001] This invention relates to the field of call center customer behavior simulation technology, and more specifically to a call center customer behavior simulation system based on deep learning. Background Technology

[0002] Call centers are crucial channels for enterprise customer service, marketing, and customer relationship management. With the development of intelligent customer service and data-driven operations, more and more call centers need to use customer behavior simulation to optimize agent scheduling, validate service strategies, and test marketing strategies. Current technologies often employ statistical rules, Markov chains, or simple probabilistic models to model customer behavior, such as simulating call duration, queuing, transfer to a human agent, and hang-up. However, real-world call center scenarios exhibit significant non-linearity, multi-stage decision-making, and strong context-dependent characteristics. For example, after long queuing times, customers may engage in complex behavioral chains such as repeat calls, cross-channel inquiries, or escalation of complaints. Traditional statistical models struggle to effectively characterize the deep relationships between these behaviors, leading to significant discrepancies between simulation results and actual business operations, thus affecting the accuracy of call center capacity planning and service strategy evaluation.

[0003] Furthermore, in certain high-value business scenarios, such as bank credit card centers, airline rebooking centers, or after-sales service centers during large e-commerce promotions, there exists a relatively subtle but significant "short-term repetitive inbound call behavior." This involves the same customer calling in multiple times within a short period and experiencing different service paths (e.g., first entering the IVR self-service process and then being transferred to a human operator, then calling again to directly request human assistance). This behavior is often influenced by multiple factors, including the customer's historical interaction trajectory, emotional changes, and system response delays. Existing technologies typically model only single inbound call events, lacking the ability to deeply model the evolution of behavior across inbound call sessions. This makes it difficult to realistically reproduce such fine-grained customer behavior patterns during simulations, leading to significant deviations in the simulation system under extreme or marginal scenarios. Therefore, how to construct a method that can characterize the multi-stage customer behavior dependencies and achieve highly realistic behavior simulation has become an urgent technical problem to be solved. Summary of the Invention

[0004] The purpose of this invention is to provide a call center customer behavior simulation system based on deep learning to address the shortcomings of the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a deep learning-based call center customer behavior simulation system, comprising: Historical interaction data acquisition module: Acquires historical interaction log data from the call center and performs session-level integration processing on the historical interaction log data to obtain a customer behavior sequence dataset; Behavioral feature encoding processing module: Performs behavioral feature encoding processing on the customer behavior sequence dataset, extracts the corresponding multi-dimensional behavioral feature vectors, and obtains a customer behavior feature set; Customer behavior prediction model construction module: Construct a deep learning customer behavior prediction model based on the customer behavior feature set, and train the deep learning customer behavior prediction model using the customer behavior feature set to obtain the customer behavior transition probability matrix; Call center environment status modeling module: acquires the operating status parameters of the call center to be simulated, and constructs a call center environment status model based on the operating status parameters; Behavior simulation model generation module: Couples the customer behavior transition probability matrix with the call center environment state model to generate a customer behavior simulation model; Customer behavior trajectory generation module: Input the initial customer call event into the customer behavior simulation model, and predict the subsequent behavior event sequence based on the customer behavior transfer probability matrix to generate a simulated customer behavior trajectory; Customer behavior statistical analysis module: Based on the simulated customer behavior trajectory, the module counts the number of repeat calls, queue abandonment rate, and manual transfer rate of the call center during the simulation period, thereby obtaining the call center customer behavior simulation results.

[0006] Preferably, in the customer behavior sequence dataset B, each customer behavior sequence Bi includes a set of customer behavior events arranged in chronological order. The behavior events include call-in time, IVR node dwell information, queuing time, human agent interaction status, hang-up method, and re-call interval features.

[0007] Preferably, the process of constructing a deep learning customer behavior prediction model based on the customer behavior feature set includes the following steps: Each behavioral sequence feature matrix in the customer behavior feature set is input into the sequence feature extraction network in the order of behavioral events to obtain the behavioral sequence latent feature vector. The latent feature vector of the behavior sequence is input into the attention weight calculation layer. By calculating the similarity between the latent feature vector of each behavior event and the overall feature vector of the sequence, a weighted behavior feature representation vector is generated. The weighted behavioral feature representation vector is input into a fully connected prediction network to calculate the behavior transfer probability, thereby obtaining the customer behavior transfer probability matrix. By using real behavior sequences from the customer behavior feature set as supervision labels, the customer behavior prediction model is iteratively trained to obtain a trained deep learning customer behavior prediction model.

[0008] Preferably, the process of obtaining the customer behavior switching probability matrix includes the following steps: Select behavioral sequence feature matrices from the customer behavior feature set according to the behavioral sequence number order, and construct the model training sample set; The training sample set of the model is input into the deep learning customer behavior prediction model, and the prediction probability vector of each behavior category is calculated through forward propagation. The cross-entropy loss value is calculated based on the difference between the behavior probability distribution and the supervised training labels, and the model parameters are iteratively updated using the adaptive moment estimation gradient descent algorithm. After completing multiple rounds of iterative training, the predicted probabilities of all training samples for each behavior category are statistically analyzed, and the probabilities are aggregated and calculated according to the transition relationship between behavior event categories, thus forming a customer behavior transition probability matrix.

[0009] Preferably, the process of constructing a call center environment state model includes the following steps: Collect the operation log data of the call center to be simulated within a preset statistical period to form a set of original operation status parameters; The original set of operating status parameters is subjected to time segmentation statistical processing. The preset statistical period is divided into multiple time intervals, and the average number of available seats, the average number of people queuing, and the average service time in each time interval are calculated to obtain the time interval operating status parameter vector. A queuing service relationship matrix is ​​constructed based on the operational status parameter vectors for each time interval. The matrix rows represent the status of different time intervals, and the matrix columns represent the transfer ratio of customers to three service paths: voice navigation process, queuing queue, and access to human service during that time interval. The queuing service relationship matrix and the time interval operation status parameter vector are jointly modeled to form a call center environment status model.

[0010] Preferably, the process of generating a customer behavior simulation model includes the following steps: The time interval mapping process is performed on each behavior transfer probability in the customer behavior transfer probability matrix. Based on the running status parameters of the corresponding time interval in the call center environment state model, the service path ratio vector corresponding to the current time interval is determined. The customer behavior transition probability matrix is ​​weighted according to the service path ratio vector to obtain the behavior transition probability matrix after environmental state adjustment. The behavior transition probability matrix after environmental state adjustment is used as the customer behavior state transition rule, and a behavior state transition table is constructed by combining the time interval running state parameters. A customer behavior simulation model is constructed based on the aforementioned behavior state transition table.

[0011] Preferably, the process of generating simulated customer behavior trajectories includes the following steps: Set the start time interval for customer behavior simulation and generate the initial customer call event based on the behavior state transition table; The initial customer call event is input into the customer behavior simulation model, and the set of the next behavior event transition probability corresponding to the current behavior event category is read from the behavior state transition table of the corresponding time interval. Based on the set of next behavior event transition probabilities, a probabilistic random sampling algorithm is used to determine the category of the next behavior event and generate the corresponding behavior event record; The next generated behavioral event is used as the new current behavioral event to continue the behavioral transfer calculation, and a sequence of behavioral events is generated step by step in chronological order to form a complete simulated customer behavior trajectory.

[0012] The technical effects and advantages provided by the present invention in the above technical solution are as follows: 1. This invention constructs a customer behavior sequence dataset and extracts multi-dimensional behavioral feature vectors. It further utilizes a deep learning customer behavior prediction model to model the evolution of customer behavior, enabling accurate characterization of behavioral relationships between voice navigation, queuing, human assistance, and repeat calls. Compared to traditional simulation methods based on statistical rules or simple probability models, this invention extracts temporal dependency features from customer behavior sequences using a long short-term memory neural network and combines this with an attention weight calculation method to assess the importance of key behavioral events. This more accurately reflects the dynamic transfer relationships of customer behavior across different interaction stages, improves the accuracy of customer behavior transfer probability calculation, and makes the generated customer behavior sequences closer to real call center operations.

[0013] 2. This invention couples a customer behavior transition probability matrix with a call center environment state model, enabling the dynamic reflection of the impact of changes in call center operational status on customer decision-making during the customer behavior evolution process. For example, when agent resources are scarce or the number of people in the queue increases, customers are more likely to abandon the queue or make repeated calls. By adjusting the behavior transition probability through environmental state parameters, dynamic simulation of customer behavior trajectories can be achieved, and key operational indicators such as the number of repeated calls, queue abandonment rate, and agent transfer rate can be further statistically obtained. This method can predict the impact of changes in call center operational status on customer behavior in advance without actual business trials, providing reliable data support for call center service process optimization, agent resource allocation, and customer experience improvement. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0015] Figure 1 This is a flowchart of the system modules of the present invention. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] For examples, please refer to Figure 1 As shown in this embodiment, the deep learning-based call center customer behavior simulation system includes: Historical interaction data acquisition module: Acquires historical interaction log data from the call center and performs session-level integration processing on the historical interaction log data to obtain a customer behavior sequence dataset; Behavioral feature encoding processing module: Performs behavioral feature encoding processing on the customer behavior sequence dataset, extracts the corresponding multi-dimensional behavioral feature vectors, and obtains a customer behavior feature set; Customer behavior prediction model construction module: Construct a deep learning customer behavior prediction model based on the customer behavior feature set, and train the deep learning customer behavior prediction model using the customer behavior feature set to obtain the customer behavior transition probability matrix; Call center environment status modeling module: acquires the operating status parameters of the call center to be simulated, and constructs a call center environment status model based on the operating status parameters; Behavior simulation model generation module: Couples the customer behavior transition probability matrix with the call center environment state model to generate a customer behavior simulation model; Customer behavior trajectory generation module: Input the initial customer call event into the customer behavior simulation model, and predict the subsequent behavior event sequence based on the customer behavior transfer probability matrix to generate a simulated customer behavior trajectory; Customer behavior statistical analysis module: Based on the simulated customer behavior trajectory, the module counts the number of repeat calls, queue abandonment rate, and manual transfer rate of the call center during the simulation period, thereby obtaining the call center customer behavior simulation results.

[0018] In this embodiment, historical interaction log data from the call center is acquired, and session-level integration processing is performed on the historical interaction log data to obtain a customer behavior sequence dataset. The specific method is as follows: In this embodiment, historical interaction log data of the call center is first acquired. This historical interaction log data originates from original interaction records made during the operation of the call center business. These original interaction records are automatically generated and stored by the call center platform during each customer call, queuing, entry into the voice navigation process, and interaction with a live agent. Each original interaction record includes at least the customer identification information, call timestamp, unique call identifier, voice navigation node identifier, node entry time, node exit time, queue entry time, queue exit time, live agent access time, call end time, and call end type information.

[0019] To ensure data consistency for subsequent behavioral modeling, the acquired historical call center interaction log data needs to be cleaned first. Specific processing methods include: deleting records with missing customer identification information; deleting records with abnormal call timestamps (defined as calls made more than 24 hours earlier than the system log generation time); deleting records with duplicate call unique identifiers and completely overlapping times; and sorting the remaining records chronologically according to their call timestamps to form standardized historical interaction log data.

[0020] After data cleaning, the standardized historical interaction log data undergoes session-level integration processing. The purpose of session-level integration is to combine multiple call behaviors of the same customer within a continuous time window into a single customer behavior sequence. Specifically, this is achieved as follows: First, all historical interaction log data is grouped according to customer identification information; for each group of customer identification data sets, they are then sorted in ascending order by call timestamp; subsequently, the time interval between two adjacent calls is calculated. When the time interval between two adjacent calls is less than or equal to a preset session time threshold, the two calls are classified as the same customer behavior sequence; when the time interval between two adjacent calls is greater than the preset session time threshold, they are classified as a new customer behavior sequence.

[0021] The preset session time threshold is 30 minutes, which is determined as follows: First, the time interval between two consecutive incoming calls from all customers in the historical interaction log data is counted; then, a histogram of time interval distribution is constructed; then, the cumulative probability function of the time interval distribution is calculated, and the time interval corresponding to the cumulative probability reaching 0.85 is selected as the session segmentation threshold; in the actual call center sample data, this time interval value is 30 minutes, so the session time threshold is set to 30 minutes.

[0022] After the above session-level integration processing, a customer behavior sequence dataset B is obtained. The customer behavior sequence dataset B consists of multiple customer behavior sequences, each customer behavior sequence is denoted as Bi, where i is the customer behavior sequence number, and the value of i ranges from 1 to N, where N represents the total number of customer behavior sequences.

[0023] In the customer behavior sequence dataset B, each customer behavior sequence Bi includes a set of customer behavior events Bi arranged in chronological order. The set of customer behavior events Bi is represented as follows: Bi is composed of multiple behavior events arranged in chronological order, and its form is that Bi is equal to the sequence of behavior events ei1, ei2 and up to eik, where k represents the number of behavior events in the customer behavior sequence.

[0024] In this embodiment, each behavioral event eij corresponds to a specific call interaction process, where j is the position number of the behavioral event in the customer behavior sequence, and the value of j ranges from 1 to k. Each behavioral event eij contains the following behavioral characteristic information: First, call in time. Call in time is defined as the system record time when a client terminal initiates a call request to the call center platform, and its unit is a timestamp in seconds.

[0025] Second, voice navigation node dwell information. Voice navigation node dwell information indicates the node path the customer traverses during the voice navigation process and the duration of dwell time at each node. Specifically, it is calculated by subtracting the entry time from the exit time of the node, with the dwell time measured in seconds.

[0026] Third, queuing time. Queuing time is defined as the time between when a customer enters the queue and when a human agent connects to them. It is calculated by subtracting the queuing entry time from the queuing end time, and the result is in seconds.

[0027] Fourth, the status of the human agent interaction. The status of the human agent interaction indicates whether the customer has entered human service and the result of the interaction. The status values ​​include three types: not entered human service, entered human service and ended the call normally, and entered human service but the customer actively ended the call.

[0028] Fifth, hang-up method. The hang-up method describes how the call ends. It includes three types: customer-initiated hang-up, agent-ended call, and system timeout termination.

[0029] Sixth, the recall interval feature. The recall interval feature represents the time difference between the current behavioral event and the next behavioral event. It is calculated by subtracting the call time of the current behavioral event from the call time of the next behavioral event, in seconds. When the current behavioral event is the last behavioral event in the customer's behavioral sequence Bi, the recall interval feature is defined as 0.

[0030] Through the above processing, a customer behavior sequence dataset B is formed. In the customer behavior sequence dataset B, each customer behavior sequence Bi records the continuous call behavior events of the customer within a session cycle in chronological order, thereby providing basic data for subsequent customer behavior feature extraction and deep learning model training.

[0031] The customer behavior sequence dataset is subjected to behavioral feature encoding processing to extract the corresponding multidimensional behavioral feature vectors, thereby obtaining a customer behavior feature set.

[0032] In this embodiment, after obtaining the customer behavior sequence dataset B, it is necessary to perform behavioral feature encoding processing on the customer behavior sequence dataset B to transform the original behavioral event data into multi-dimensional behavioral feature vectors that can be used for deep learning training, thereby obtaining the customer behavior feature set F. The specific implementation process includes the following steps: First, feature standardization is performed on the behavioral events in the customer behavior sequence dataset B. Since behavioral events include time-based, path-based, and state-based features, and the value ranges of different features vary significantly, normalization is required for the numerical features. These numerical features include call-in time, dwell time at voice navigation nodes, queuing time, and re-call interval. The normalization process uses a minimum-maximum value normalization algorithm. The calculation method is as follows: for a given numerical feature, subtract the minimum value of that feature from the original value in the entire customer behavior sequence dataset B, and then divide the result by the difference between the maximum and minimum values ​​of that feature in the dataset. The normalized feature values ​​range from 0 to 1.

[0033] After normalizing the numerical features, the voice navigation node paths are sequence encoded. Since voice navigation nodes are discrete category information, node identifiers need to be converted into numerical vectors. Specifically, the process is as follows: First, all voice navigation node identifiers appearing in the customer behavior sequence dataset B are counted and sorted from highest to lowest frequency. Then, a unique node number is assigned to each voice navigation node, starting from 1 and incrementing. After obtaining the node numbers, a node embedding vector table is constructed. The node embedding vector table is set to 16 dimensions, meaning each voice navigation node corresponds to a vector of length 16. The node embedding vectors are initialized randomly, with random numbers ranging from -0.05 to +0.05. Subsequently, the node embedding vectors are updated during model training using a backpropagation algorithm, ensuring that the node embedding vectors reflect the semantic relationships between different voice navigation nodes.

[0034] Next, the interaction status and call termination methods of the human agent are encoded. The human agent interaction status includes three values: "not connected to human service," "connected to human service and ended the call normally," and "connected to human service but the customer actively terminated the call." The call termination methods include three values: "customer actively terminated," "agent ended the call," and "system timed out." One-hot encoding is used to encode these statuses. Specifically, each status is assigned a fixed binary position; when a status occurs, the corresponding position is set to 1, and the remaining positions are set to 0. For example, if there are three possible human agent interaction statuses, a status vector of length 3 is constructed; similarly, if there are three possible call termination methods, a call termination method vector of length 3 is constructed.

[0035] After obtaining the aforementioned coding features, all features from the same behavioral event are concatenated to form a behavioral event feature vector. Specifically, the normalized call-in time feature, normalized queuing time feature, normalized recall interval feature, voice navigation node embedding vector, human agent interaction state coding vector, and hang-up method coding vector are concatenated in a fixed order to form a multi-dimensional feature vector for a single behavioral event. This multi-dimensional feature vector has 24 dimensions, including 3-dimensional normalized time features, a 16-dimensional voice navigation node embedding vector, and 5-dimensional state coding features.

[0036] Subsequently, sequence feature construction processing is performed on the customer behavior sequences. For each customer behavior sequence Bi in the customer behavior sequence dataset B, the feature vectors of the behavior events are extracted sequentially according to the time order of occurrence, and then combined in order to form a behavior sequence feature matrix. If the customer behavior sequence Bi contains k behavior events, a k-row, 24-column behavior sequence feature matrix is ​​constructed, where each row represents the feature vector corresponding to a behavior event.

[0037] Since the number of behavioral events k may vary in different customer behavior sequences, the length of the behavior sequences needs to be uniformly processed to ensure consistent input dimensions for subsequent deep learning models. This implementation sets the maximum behavior sequence length to 20. When the number of behavioral events in a customer behavior sequence Bi is less than 20, a vector of all zeros is added to the end of the sequence until the sequence length reaches 20; when the number of behavioral events is greater than 20, the 20 most recent behavioral events are retained, and earlier behavioral events are deleted.

[0038] It should be noted that, through the above processing steps, the customer behavior sequence dataset B can be transformed into a customer behavior feature set F. The customer behavior feature set F consists of multiple behavior sequence feature matrices, each corresponding to a customer behavior sequence Bi. These behavior sequence feature matrices represent the evolution of customer behavior within a single session and provide input data for subsequently building a customer behavior prediction model.

[0039] A deep learning customer behavior prediction model is constructed based on the customer behavior feature set, and the deep learning customer behavior prediction model is trained using the customer behavior feature set to obtain a customer behavior transition probability matrix.

[0040] In this embodiment, when constructing a deep learning customer behavior prediction model based on the customer behavior feature set F, the customer behavior sequence feature matrix is ​​first subjected to temporal feature extraction processing. Specifically, each behavior sequence feature matrix in the customer behavior feature set F is input into the sequence feature extraction network in the order of the occurrence time of the behavior events. The sequence feature extraction network is constructed using a long short-term memory neural network structure, which is used to learn the long-term and short-term dependencies between behavior events.

[0041] The sequence feature extraction network comprises two layers of Long Short-Term Memory (LSTM) neural networks, each containing 64 hidden neurons. The network input is a behavior sequence feature matrix of length 20 and dimension 24, where 20 represents the length of the behavior sequence and 24 represents the dimension of a single behavior event feature vector. The first LSM layer reads the behavior event feature vector step-by-step and calculates the current hidden state at each time step. The calculation process for the hidden state is as follows: first, the behavior event feature vector at the current time step is linearly weighted with the hidden state vector at the previous time step; then, information is filtered through gating units, which include an input gate, a forget gate, and an output gate. The input gate controls the proportion of current behavior event information entering the memory unit, the forget gate controls the proportion of information retained from the previous time step, and the output gate controls the output of the current hidden state. The hidden state sequence of the first layer network is obtained through the above calculations.

[0042] Subsequently, the hidden state sequence output by the first layer of the Long Short-Term Memory (LSTM) neural network is input into the second layer of the LSTM neural network for further temporal feature extraction, thereby obtaining the final behavioral sequence latent feature vector. The behavioral sequence latent feature vector is the hidden state vector output by the second layer of the LSTM neural network at the last time step, with a dimension of 64. This vector is used to represent the overall temporal dependency features of the customer behavior sequence.

[0043] After obtaining the implicit feature vector of the behavior sequence, the importance of each behavioral event feature in the behavior sequence is evaluated, thereby highlighting behavioral events that have a greater impact on the evolution of customer behavior. This implementation method achieves this process through an attention weight calculation layer.

[0044] The specific implementation is as follows: First, obtain the hidden state vector output by the second-layer Long Short-Term Memory neural network at each time step, denoted as the sequence of hidden feature vectors for behavioral events. Then, calculate the similarity between each hidden feature vector of a behavioral event and the hidden feature vector of the behavioral sequence. The similarity is calculated using vector dot product, which involves multiplying the values ​​at corresponding positions of the two vectors and summing them to obtain a similarity value.

[0045] After obtaining the similarity scores for all behavioral events, the similarity scores need to be normalized to obtain the importance weight of each behavioral event. The normalization process uses exponential normalization, which is calculated as follows: first, an exponential function is applied to each similarity score; then, each exponential result is divided by the sum of all exponential results to obtain the importance weight corresponding to each behavioral event. The sum of all weights is 1.

[0046] Subsequently, the latent feature vector of each behavioral event is multiplied by its corresponding importance weight, and all weighted latent feature vectors of behavioral events are summed to obtain a weighted behavioral feature representation vector. This weighted behavioral feature representation vector has 64 dimensions and is used to represent the importance of each behavioral event in the behavioral sequence and its comprehensive characteristics.

[0047] After obtaining the weighted behavioral feature representation vector, it is necessary to predict the probability of the customer's next behavioral event. This implementation uses a fully connected prediction network to calculate the behavior transition probability.

[0048] The fully connected prediction network comprises a two-layer neural network structure. The input to the first layer is a weighted behavioral feature representation vector with a dimension of 64, and the first layer has 128 neurons. The calculation process for this layer is as follows: the 64-dimensional input vector is multiplied by the weight matrix corresponding to the 128 neurons, and a bias vector is added to obtain a 128-dimensional intermediate feature vector; then, a linear rectified function is performed on the intermediate feature vector. The linear rectified function is calculated such that it outputs the original value when the input value is greater than 0, and outputs 0 when the input value is less than 0.

[0049] After completing the first layer of calculations, the resulting 128-dimensional intermediate feature vector is input into the second layer of the neural network for behavior category prediction. The number of neurons in the second layer of the neural network is equal to the number of behavior event categories. In this embodiment, the behavior event categories include five types: entering the voice navigation process, entering the queue, connecting to a live agent, the customer actively hanging up, and calling back. Therefore, the number of neurons in the second layer of the neural network is set to five.

[0050] Subsequently, the five values ​​output by the second-layer neural network are subjected to probability normalization. The probability normalization process uses exponential normalization, which involves first performing an exponential operation on each output value, then dividing each exponential result by the sum of the five exponential results to obtain the probability value corresponding to each behavior category. The sum of all probability values ​​is 1. This process yields the probability distribution of the corresponding behavior category in the customer behavior transition probability matrix.

[0051] After constructing the deep learning customer behavior prediction model structure, the model needs to be trained using real behavior sequences from the customer behavior feature set F to gradually converge the model parameters. The specific training process is as follows: First, the category of the next actual behavior event in the customer behavior sequence is used as the supervised training label. Then, the difference between the probability distribution of the predicted behavior categories and the actual behavior categories is calculated. The difference calculation uses the cross-entropy loss function. The calculation process of the cross-entropy loss function is as follows: First, the probability value corresponding to the actual behavior category is obtained; then, the logarithm of this probability value is taken and the negative value is obtained to obtain the loss value for a single sample.

[0052] After obtaining the loss value, the model parameters need to be updated using the gradient descent algorithm. This implementation uses the adaptive moment estimation gradient descent algorithm for parameter updates. This algorithm adjusts the parameter update magnitude by simultaneously calculating the first-order moment estimate and the second-order moment estimate of the gradient. The specific steps are as follows: First, calculate the gradient value of the current batch of training samples; then calculate the exponentially weighted average of the gradient as the first-order moment estimate; next, calculate the exponentially weighted average of the squared gradient as the second-order moment estimate; finally, update the model parameters based on the first-order moment estimate and the second-order moment estimate.

[0053] During model training, each calculation of all training samples is called a training epoch. After each training epoch, the overall average loss value is calculated. When the average loss value is less than the preset training threshold of 0.01, the training process stops, thus obtaining the trained deep learning customer behavior prediction model.

[0054] When constructing the model training data, supervised learning training samples need to be generated from the customer behavior feature set F. Specifically, the behavior sequence feature matrix is ​​read sequentially according to the behavior sequence number, and the first 19 behavior event feature vectors from the feature matrix of each behavior sequence are extracted as model input data. Each behavior event feature vector has a dimension of 24, therefore the input data constitutes a 19-row, 24-column matrix.

[0055] Subsequently, the behavior category corresponding to the 20th behavior event in the behavior sequence is used as the supervised training label. In this way, the model can predict the category of the 20th behavior event based on the first 19 behavior events. By processing all behavior sequences in the above manner, the model training sample set can be constructed.

[0056] After obtaining the model training sample set, the model training sample set is input into the deep learning customer behavior prediction model for forward propagation calculation. The forward propagation calculation process includes sequence feature extraction calculation, attention weight calculation, and behavior probability prediction calculation.

[0057] The specific calculation method is as follows: First, the 19 input behavioral event feature vectors are sequentially input into a long short-term memory neural network for temporal feature extraction, thereby obtaining the behavioral sequence latent feature vector; then, the importance weight of each behavioral event is calculated through the attention weight calculation layer, and a weighted behavioral feature representation vector is generated; finally, the weighted behavioral feature representation vector is input into a fully connected prediction network to obtain the behavioral category prediction value.

[0058] After obtaining the predicted values ​​for each behavior category, the predicted values ​​are normalized to obtain the behavior probability distribution. This probability distribution represents the model's prediction of the probability of occurrence for each behavior category.

[0059] After obtaining the behavior probability distribution, the model parameters need to be updated based on the difference between the prediction results and the supervised training labels. This implementation calculates the prediction error using the cross-entropy loss function.

[0060] The specific process is as follows: First, determine the behavior category corresponding to the supervised training label; then, read the predicted probability value corresponding to that category from the behavior probability distribution; next, perform a logarithmic operation on the predicted probability value and take its negative value to obtain the loss value of a single training sample. Then, average all the loss values ​​in a training batch to obtain the average loss value for that batch.

[0061] After obtaining the average loss value, the model parameters are updated using the adaptive moment estimation gradient descent algorithm. The initial learning rate of this algorithm is set to 0.001. During each parameter update, the parameter update magnitude is calculated based on the current first-moment and second-moment estimates of the gradient, and the network weight parameters are adjusted according to the calculation results to form new behavior prediction model parameters.

[0062] After completing multiple rounds of model training, it is necessary to statistically analyze the transition probability relationships between customer behaviors based on the model prediction results, thereby forming a customer behavior transition probability matrix.

[0063] The specific implementation method is as follows: First, the trained deep learning customer behavior prediction model is used to predict the behavior of all training samples, and the probability distribution of the behavior category corresponding to each sample is recorded. Then, statistical calculations are performed according to the transition relationship between behavior event categories.

[0064] The behavioral event categories include five types: entering the voice navigation process, entering the queue, connecting to a live agent, customer-initiated hanging up, and re-calling. For each behavioral category, the probability of transitioning to the next behavioral category when it is the current behavioral event is calculated. The calculation method is as follows: first, all predicted probability values ​​are accumulated; then, the sum of the probabilities corresponding to a certain behavioral category is divided by the sum of all probability values ​​to obtain the transition probability between behavioral categories.

[0065] Through the above statistical calculation process, a 5-row, 5-column customer behavior transition probability matrix is ​​formed. Each row of the matrix represents the current behavior event category, each column represents the next behavior event category, and the values ​​in the matrix represent the corresponding behavior transition probabilities. This customer behavior transition probability matrix is ​​used for behavior path generation in subsequent customer behavior simulation processes.

[0066] Obtain the operating status parameters of the call center to be simulated, and construct a call center environment status model based on the operating status parameters.

[0067] In this embodiment, the operational status parameters of the call center to be simulated are first obtained. Specifically, this is achieved by collecting operational log data generated by the call center within a preset statistical period. The operational log data originates from the business operation records automatically recorded by the call center business platform during customer inbound calls, entry into the voice navigation process, entry into the queuing queue, and access to human assistance. Each operational log entry includes a timestamp, a unique customer call identifier, an agent access identifier, a voice navigation node number, queuing entry time, queuing end time, and the start and end times of human assistance.

[0068] The preset statistical period is set to 7 consecutive days. This setting is based on the fact that call center business volume usually exhibits stable cyclical changes on a weekly basis. Therefore, a 7-day statistical period can fully cover the business fluctuations on weekdays and non-working days. Subsequently, operational status parameters are extracted from the operational log data. These operational status parameters include the number of agents, queue length, voice navigation process path, and human service duration records.

[0069] The number of agent seats is obtained by counting the number of agents in "answerable state" at each time point; the queue length is obtained by counting the number of customers in the queue at the same time point; the voice navigation process path is obtained by recording the order in which customers jump between voice navigation nodes; and the manual service duration is calculated by subtracting the manual service start time from the manual service end time, in seconds. The above processing forms the original set of operating status parameters.

[0070] After obtaining the original set of operating state parameters, it is necessary to perform time-segmented statistical processing to obtain stable time interval operating state characteristics.

[0071] The specific implementation method is as follows: First, the preset statistical period is divided into multiple time intervals according to fixed time intervals. In this implementation method, 24 hours is divided into 48 time intervals, each with a length of 30 minutes. Then, all operation log data is assigned to the corresponding time interval according to the timestamp.

[0072] Within each time interval, calculate the average number of available seats, the average number of people in the queue, and the average service time. The average number of available seats is calculated by summing the available seats across all sampling points within the time interval and dividing by the total number of sampling points. The average number of people in the queue is calculated by averaging the number of customers in the queue within the time interval. The average service time is calculated by summing the total human service time within the time interval and dividing by the number of service records to obtain the average service time, expressed in seconds.

[0073] Through the above statistical calculations, the operational status parameter vector for each time interval can be obtained. The operational status parameter vector for each time interval consists of three values, representing the average number of available seats, the average number of people in the queue, and the average service time, respectively.

[0074] After obtaining the operating status parameter vectors for each time interval, it is necessary to construct a queuing service relationship matrix to describe the distribution of service paths for customers under different operating states.

[0075] The specific implementation method is as follows: First, determine the three types of service paths for customers during the call center interaction process: entering the voice navigation process, entering the queue, and accessing human service. Then, analyze the service path distribution of customer call events within each time interval.

[0076] The statistical method is as follows: First, the total number of customer calls within the time interval is counted; then, the number of calls entering the voice navigation process, the number of calls entering the queue, and the number of calls successfully connected to human assistance are counted separately. Next, the transfer ratio for each of the three service paths is calculated. The transfer ratio is calculated by dividing the number of calls for a specific service path by the total number of calls within the time interval to obtain the ratio for that path.

[0077] The above calculations yield a service path proportion vector for each time interval. These vectors are then arranged according to the time intervals to construct a queuing service relationship matrix. Each row of the matrix represents a time interval state, and each column represents the proportion of paths entering the voice navigation process, the proportion of paths entering the queue, and the proportion of paths accessing human assistance.

[0078] After obtaining the queuing service relationship matrix and the time interval operation status parameter vector, it is necessary to construct a call center environment state model to describe the impact of the call center operation environment on customer behavior paths.

[0079] The specific implementation is as follows: First, the operational status parameter vector for each time interval is combined with the service path ratio vector for the corresponding time interval to form environmental status sample data. Each environmental status sample data contains three operational status parameters and three service path ratio parameters, for a total of six parameters. Then, a multi-layer feedforward neural network is used to construct a call center environmental status model. This neural network includes an input layer, a hidden layer, and an output layer. The input layer has 3 neurons, corresponding to the three operational status parameters: average available agent seats, average number of people in the queue, and average service time. The hidden layer has 32 neurons, used to extract the non-linear relationships between the operational status parameters. The output layer has 3 neurons, corresponding to the predicted values ​​of the three types of service path ratios.

[0080] During model training, the operating state parameter vector is used as input data, and the service path ratio vector corresponding to the time interval is used as supervision labels. The squared error between the predicted and actual ratio values ​​is calculated, and the neural network weights are iteratively updated using the gradient descent algorithm. Training stops when the mean squared error of all training samples falls below 0.005, thus obtaining the trained call center environment state model. This model is used to describe the changes in customer service paths under different operating state parameter conditions in a call center.

[0081] The customer behavior transition probability matrix is ​​coupled with the call center environment state model to generate a customer behavior simulation model.

[0082] In this embodiment, after obtaining the customer behavior transfer probability matrix and the call center environment state model, a time interval mapping process is first required to establish a correspondence between the customer behavior transfer probability and the call center operating status. Specifically, the time granularity of the customer behavior simulation is first determined. In this embodiment, the simulation time granularity is set to 30 minutes and divided into multiple time intervals according to continuous time sequence. Then, the operating status parameter vector corresponding to the time interval in the call center environment state model is read. The operating status parameter vector includes three numerical parameters: average number of available agents, average number of people in the queue, and average service time.

[0083] After obtaining the operational status parameter vector, the service path proportion vector corresponding to that time interval is calculated using the call center environment status model. The service path proportion vector contains three proportion values, representing the proportion of customers entering the voice navigation process path, the proportion entering the queuing queue path, and the proportion accessing human assistance path during that operational state. Through this time interval mapping process, a correlation is established between the behavior categories in the customer behavior transition probability matrix and the service path proportions of the corresponding time intervals, thus providing foundational data for subsequent probability adjustments.

[0084] After completing the time interval mapping process, the customer behavior transition probability matrix needs to be weighted according to the service path proportion vector to reflect the impact of different operating states on the customer behavior path. Specifically, the implementation involves first reading the behavior transition probability values ​​from the customer behavior transition probability matrix. This matrix is ​​a 5x5 matrix, where each row represents the current behavior event category, each column represents the next behavior event category, and the values ​​in the matrix represent the corresponding behavior transition probability.

[0085] Subsequently, the corresponding probabilities in the matrix are weighted according to the service path proportion vector. For example, if the next action event belongs to the voice navigation process path, the original action transition probability is multiplied by the voice navigation process path proportion; if the next action event belongs to the queuing queue path, the original action transition probability is multiplied by the queuing queue path proportion; if the next action event belongs to the human service access path, the original action transition probability is multiplied by the human service path proportion.

[0086] After completing the product calculation, all weighted behavior transition probabilities need to be normalized. The normalization process is as follows: first, calculate the sum of all weighted probabilities within the same row for the same behavior event; then, divide each weighted probability by this sum to obtain a new behavior transition probability value. This process yields the behavior transition probability matrix adjusted for the environmental state.

[0087] After obtaining the behavior transition probability matrix after environmental state adjustment, a behavior state transition table needs to be constructed to clarify the transition relationship of customer behavior events under different operating states. The specific implementation method is as follows: First, the behavior transition probability matrix after environmental state adjustment is read sequentially according to time intervals, and this matrix is ​​associated and stored with the corresponding time interval operating state parameter vectors.

[0088] A behavior state transition table is then established. This table contains four types of information fields: time interval number, current behavior event category, next behavior event category, and corresponding behavior transition probability. For each time interval, it is necessary to traverse all rows and columns of the behavior transition probability matrix and record the transition probability between each pair of behavior event categories in the behavior state transition table.

[0089] By using the above-described traversal method, a complete behavioral state transition table can be formed. This table reflects the transition probability relationships between various customer behavioral events under different time intervals and operating conditions.

[0090] After obtaining the behavior state transition table, a customer behavior simulation model needs to be built based on this table to dynamically generate customer behavior sequences. The specific implementation is as follows: First, a simulation time period is set, and the corresponding behavior state transition table data is read sequentially according to 30-minute time intervals. During the simulation, when a current behavior event occurs, its next behavior event needs to be determined according to the behavior state transition table. The specific selection method is a probabilistic random sampling algorithm. The algorithm is implemented as follows: First, all the transition probabilities of the next behavior events corresponding to the current behavior event are read, and the cumulative probability value is calculated according to the behavior category order; then, a random number between 0 and 1 is generated; the random number is then compared with the cumulative probability interval. When the random number falls into a certain cumulative probability interval, the corresponding next behavior event category is determined. After determining the next behavior event category, this behavior event is used as the new current behavior event for further transition calculations, and the above probabilistic random sampling process is repeated to gradually generate a complete customer behavior sequence. The customer behavior simulation model built in this way can simulate the evolution of customer behavior under different operating state parameters, providing a simulation data foundation for call center business operation analysis.

[0091] An initial customer call event is input into the customer behavior simulation model, and the subsequent behavioral event sequence is predicted based on the customer behavior transfer probability matrix, thereby generating a simulated customer behavior trajectory.

[0092] In this implementation, after constructing the customer behavior simulation model, it is necessary to first generate an initial customer call event as the starting point for generating the customer behavior trajectory. Specifically, the starting time interval for the customer behavior simulation is first set. This implementation divides the call center's operating time into continuous time intervals of 30 minutes and numbers them sequentially. For example, the first time interval represents 0:00 to 0:30, the second time interval represents 0:30 to 1:00, and so on.

[0093] Subsequently, the probability of a customer inbound call occurring within this time interval is calculated based on the call center environment state model. The inbound call probability is obtained by statistically analyzing the average number of inbound calls during this time interval from historical operation log data. Specifically, the calculation method is as follows: first, the total number of customer inbound calls within this time interval is counted within a preset statistical period, and the results are summed; then, the total number of inbound calls is divided by the number of days in the statistical period to obtain the average number of inbound calls. An initial customer inbound call event is then generated based on the average number of inbound calls.

[0094] When generating an initial customer call-in event, it is necessary to construct event attribute information. The initial customer call-in event includes three types of information fields: call-in time, voice navigation node number, and current behavior event category. The call-in time is obtained by generating random time points within the current time interval; the voice navigation node number is determined by reading the node number that appears most frequently in the voice navigation process path; and the current behavior event category is uniformly set to the customer call-in event category, thus forming the initial customer call-in event record.

[0095] After generating the initial customer call event, the event needs to be input into the customer behavior simulation model, and the candidate set of the next behavior event needs to be determined according to the behavior state transition table. Specifically, the current behavior event category in the initial customer call event is read first, and the time interval number of that behavior event is determined.

[0096] The system then retrieves the set of records corresponding to the time interval number and the current behavior event category from the behavior state transition table. This set of records contains multiple next behavior event categories and their corresponding behavior transition probabilities. The behavior transition probabilities are derived from the behavior transition probability matrix after environmental state adjustment, and their values ​​represent the probability that the current behavior event will transition to the next behavior event.

[0097] After reading the record set, a set of transition probabilities for the next behavioral event is formed according to the order of behavioral event categories. This set contains multiple behavioral event categories and their corresponding probability values, and serves as the input data for calculating the selection of the next behavioral event.

[0098] After obtaining the set of transition probabilities for the next action event, it is necessary to determine the category of the next action event. This implementation uses a probabilistic random sampling algorithm for action event selection. The probabilistic random sampling algorithm is used to randomly select an action event category according to a probability distribution.

[0099] The specific implementation steps of the algorithm are as follows: First, the probability values ​​in the probability set are accumulated according to the order of the behavior event categories. The accumulation calculation method is as follows: the probability value of the first behavior category is taken as the first accumulated probability; the probability value of the second behavior category is added to the first accumulated probability to obtain the second accumulated probability; the same method is used to calculate until the last accumulated probability is 1.

[0100] A random number between 0 and 1 is then generated. This random number is generated using a pseudo-random number generation algorithm, which employs a linear congruential algorithm. The linear congruential algorithm works by multiplying the previous random number by a fixed multiplier, adding a fixed increment, and then performing a remainder operation with a preset modulus to obtain a new random value. This new value is then divided by the modulus to obtain a decimal between 0 and 1.

[0101] After obtaining the random number, it is compared with the cumulative probability interval. When the random number falls into a certain cumulative probability interval, the behavioral event category corresponding to that interval is determined as the next behavioral event category. Subsequently, the corresponding behavioral event record is generated, recording its behavioral event category and occurrence time.

[0102] After generating the next behavioral event, this behavioral event needs to be treated as the new current behavioral event, and the behavioral transfer calculation continues to be performed, thereby gradually generating a complete customer behavior trajectory.

[0103] The specific implementation is as follows: First, read the behavior event category and occurrence time of the newly generated behavior event, and determine its time interval number based on the occurrence time. Then, search the behavior state transition table for the time interval and the set of next behavior event transition probabilities corresponding to the current behavior event category.

[0104] After obtaining a new set of probabilities, the probabilistic random sampling algorithm is repeatedly executed to determine the new next behavioral event category and generate a new behavioral event record. By continuously repeating the above behavioral transition process, a sequence of behavioral events arranged in chronological order can be formed.

[0105] To control the termination conditions of the simulation process, this implementation method sets two termination conditions. The first termination condition is that the behavior event category is a customer-initiated hang-up event; the second termination condition is that the number of behavior events reaches a preset maximum behavior sequence length of 20. When either termination condition is met, the behavior transfer calculation stops, and all currently generated behavior events are arranged in chronological order to form a complete simulated customer behavior trajectory. This simulated customer behavior trajectory is used for subsequent call center operation performance analysis and service strategy evaluation.

[0106] Based on the simulated customer behavior trajectory, the call center counts the number of repeat calls, queue abandonment rate, and manual transfer rate during the simulation period, thereby obtaining the call center customer behavior simulation results.

[0107] In this embodiment, after generating simulated customer behavior trajectories, statistical analysis needs to be performed on these trajectories to obtain the number of repeat inbound calls, queue abandonment rate, and human transfer rate within the simulation period, thereby forming the call center customer behavior simulation results. The specific implementation process includes the following steps.

[0108] First, the simulated customer behavior trajectory is processed through trajectory analysis. This trajectory consists of multiple behavioral events arranged chronologically. Each event includes the event category, the time of occurrence, and the corresponding voice navigation node number. The trajectory analysis is implemented by reading each behavioral event record sequentially according to its occurrence time, and grouping all events according to customer identification information to obtain the complete set of behavioral trajectories for the corresponding customer. Each customer behavior trajectory represents a sequence of all call behavior events that occurred for the same customer within the simulation period.

[0109] After completing the trajectory analysis, the number of repeat incoming calls is statistically calculated. Repeat incoming call behavior is defined as the same customer making another call within a preset time threshold after an initial incoming call. In this implementation, the preset time threshold is set to 1800 seconds. This time threshold is determined as follows: first, the time interval between two adjacent incoming calls is statistically analyzed in the historical interaction log data; then, the probability distribution of all time interval values ​​is calculated; finally, the time interval corresponding to a cumulative probability of 0.80 is selected as the repeat incoming call judgment threshold. During the statistical process, incoming calls in each customer behavior trajectory are read sequentially, and the time difference between two adjacent incoming calls is calculated. When the time difference is less than or equal to 1800 seconds, the behavior is counted as a repeat incoming call event. After traversing and statistically analyzing all customer behavior trajectories, the number of repeat incoming calls within the simulation period can be obtained.

[0110] Queue abandonment rate is statistically calculated. Queue abandonment is defined as the act of a customer actively hanging up after entering the queue without being connected to a live agent. The statistical process first involves counting the total number of events involving entering the queue in the simulated customer behavior trajectory; then, the number of events where a customer actively hangs up after entering the queue is counted. The queue abandonment rate is calculated by dividing the number of queue abandonment events by the total number of events involving entering the queue. The queue abandonment rate is a decimal between 0 and 1, representing the proportion of customers who abandon service during the queuing phase.

[0111] The manual transfer rate was statistically calculated. Manual transfer behavior is defined as the act of a customer successfully connecting to a human agent after the voice navigation process has ended. In the statistical process, the total number of behavioral events entering the voice navigation process in the simulated customer behavior trajectory was first counted; then, the number of behavioral events that successfully entered the human agent service stage after the voice navigation process was counted. The manual transfer rate was calculated by dividing the number of events that successfully connected to a human agent by the total number of events entering the voice navigation process.

[0112] After completing the above statistical calculations, the statistical results need to be summarized during a simulation period. The specific implementation method is as follows: First, all statistical results are filtered for time range according to the preset simulation period; then, the statistical results of all customer behavior trajectories within the simulation period are summarized to obtain the three indicators: total number of repeat inbound calls, queue abandonment rate, and manual transfer rate.

[0113] Finally, the number of repeat calls, queue abandonment rate, and human transfer rate are output as the results of the call center customer behavior simulation. The number of repeat calls reflects the stability of the customer service experience, the queue abandonment rate reflects the level of waiting pressure, and the human transfer rate reflects the effectiveness of the voice navigation process. Through the above statistical analysis, the customer behavior characteristics of the call center under simulated operating conditions can be quantitatively evaluated, thus providing data for optimizing call center service strategies.

[0114] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A deep learning-based call center customer behavior simulation system, characterized in that: include: Historical interaction data acquisition module: Acquires historical interaction log data from the call center and performs session-level integration processing on the historical interaction log data to obtain a customer behavior sequence dataset; Behavioral feature encoding processing module: Performs behavioral feature encoding processing on the customer behavior sequence dataset, extracts the corresponding multi-dimensional behavioral feature vectors, and obtains a customer behavior feature set; Customer behavior prediction model construction module: Construct a deep learning customer behavior prediction model based on the customer behavior feature set, and train the deep learning customer behavior prediction model using the customer behavior feature set to obtain the customer behavior transition probability matrix; Call center environment status modeling module: acquires the operating status parameters of the call center to be simulated, and constructs a call center environment status model based on the operating status parameters; Behavior simulation model generation module: Couples the customer behavior transition probability matrix with the call center environment state model to generate a customer behavior simulation model; Customer behavior trajectory generation module: Input the initial customer call event into the customer behavior simulation model, and predict the subsequent behavior event sequence based on the customer behavior transfer probability matrix to generate a simulated customer behavior trajectory; Customer behavior statistical analysis module: Based on the simulated customer behavior trajectory, the module counts the number of repeat calls, queue abandonment rate, and manual transfer rate of the call center during the simulation period, thereby obtaining the call center customer behavior simulation results.

2. The deep learning-based call center customer behavior simulation system according to claim 1, characterized in that: In the customer behavior sequence dataset B, each customer behavior sequence Bi includes a set of customer behavior events arranged in chronological order. The behavior events include call time, IVR node dwell information, queuing time, human agent interaction status, hang-up method, and re-call interval features.

3. The deep learning-based call center customer behavior simulation system according to claim 1, characterized in that: The process of constructing a deep learning customer behavior prediction model based on the customer behavior feature set includes the following steps: Each behavioral sequence feature matrix in the customer behavior feature set is input into the sequence feature extraction network in the order of behavioral events to obtain the behavioral sequence latent feature vector. The latent feature vector of the behavior sequence is input into the attention weight calculation layer. By calculating the similarity between the latent feature vector of each behavior event and the overall feature vector of the sequence, a weighted behavior feature representation vector is generated. The weighted behavioral feature representation vector is input into a fully connected prediction network to calculate the behavior transfer probability, thereby obtaining the customer behavior transfer probability matrix. By using real behavior sequences from the customer behavior feature set as supervision labels, the customer behavior prediction model is iteratively trained to obtain a trained deep learning customer behavior prediction model.

4. The deep learning-based call center customer behavior simulation system according to claim 3, characterized in that: The process of obtaining the customer behavior transition probability matrix includes the following steps: Select behavioral sequence feature matrices from the customer behavior feature set according to the behavioral sequence number order, and construct the model training sample set; The training sample set of the model is input into the deep learning customer behavior prediction model, and the prediction probability vector of each behavior category is calculated through forward propagation. The cross-entropy loss value is calculated based on the difference between the behavior probability distribution and the supervised training labels, and the model parameters are iteratively updated using the adaptive moment estimation gradient descent algorithm. After completing multiple rounds of iterative training, the predicted probabilities of all training samples for each behavior category are statistically analyzed, and the probabilities are aggregated and calculated according to the transition relationship between behavior event categories, thus forming a customer behavior transition probability matrix.

5. The deep learning-based call center customer behavior simulation system according to claim 1, characterized in that: The process of building a call center environment state model includes the following steps: Collect the operation log data of the call center to be simulated within a preset statistical period to form a set of original operation status parameters; The original set of operating status parameters is subjected to time segmentation statistical processing. The preset statistical period is divided into multiple time intervals, and the average number of available seats, the average number of people queuing, and the average service time in each time interval are calculated to obtain the time interval operating status parameter vector. A queuing service relationship matrix is ​​constructed based on the operational status parameter vectors for each time interval. The matrix rows represent the status of different time intervals, and the matrix columns represent the transfer ratio of customers to three service paths: voice navigation process, queuing queue, and access to human service during that time interval. The queuing service relationship matrix and the time interval operation status parameter vector are jointly modeled to form a call center environment status model.

6. The deep learning-based call center customer behavior simulation system according to claim 1, characterized in that: The process of generating a customer behavior simulation model includes the following steps: The time interval mapping process is performed on each behavior transfer probability in the customer behavior transfer probability matrix. Based on the running status parameters of the corresponding time interval in the call center environment state model, the service path ratio vector corresponding to the current time interval is determined. The customer behavior transition probability matrix is ​​weighted according to the service path ratio vector to obtain the behavior transition probability matrix after environmental state adjustment. The behavior transition probability matrix after environmental state adjustment is used as the customer behavior state transition rule, and a behavior state transition table is constructed by combining the time interval running state parameters. A customer behavior simulation model is constructed based on the aforementioned behavior state transition table.

7. The deep learning-based call center customer behavior simulation system according to claim 1, characterized in that: The process of generating simulated customer behavior trajectories includes the following steps: Set the start time interval for customer behavior simulation and generate the initial customer call event based on the behavior state transition table; The initial customer call event is input into the customer behavior simulation model, and the set of the next behavior event transition probability corresponding to the current behavior event category is read from the behavior state transition table of the corresponding time interval. Based on the set of next behavior event transition probabilities, a probabilistic random sampling algorithm is used to determine the category of the next behavior event and generate the corresponding behavior event record; The next generated behavioral event is used as the new current behavioral event to continue the behavioral transfer calculation, and a sequence of behavioral events is generated step by step in chronological order to form a complete simulated customer behavior trajectory.

Images