A multi-dimensional feature-based work order dynamic allocation method and system

Abstract

The application provides a kind of multi-dimensional feature-based work order dynamic allocation method and system, belongs to customer relationship management and artificial intelligence technical field.The method includes: obtaining the multi-dimensional dynamic characteristics of customer, work order and agent;Adopting graph neural network to model the heterogeneous relationship between entities;Define a composite reward function that considers business objectives such as immediate efficiency, customer long-term value and agent growth;Using reinforcement learning framework to learn and execute the optimal allocation strategy aimed at maximizing long-term cumulative rewards;And continuously monitor the system state, dynamically reallocate the assigned but not processed work orders.The application can improve matching accuracy, achieve multi-objective strategic optimization, and enhance system adaptability and timeliness.

Description

Technical Field

[0001] This invention relates to the fields of customer relationship management and artificial intelligence technology, and in particular to a method and system for dynamic work order allocation based on multi-dimensional features. Background Technology

[0002] In modern customer service centers, the ticketing system is central to handling customer requests, tracking issues, and managing service processes. The efficiency and accuracy of ticket allocation directly impact customer satisfaction, agent productivity, and overall operating costs. Traditional ticket allocation methods primarily include manual allocation and rule-based automated allocation. Manual allocation is highly subjective, inefficient, and struggles to handle large volumes of tickets; while rule-based automated allocation, such as polling or matching by skill tags, achieves automation, its rigid rules cannot handle complex scenarios or reflect subtle differences in agent capabilities.

[0003] In recent years, with the development of artificial intelligence technology, some intelligent allocation methods based on machine learning have emerged. For example, some technical solutions propose collecting multi-dimensional data, including customer relationship management data, and using methods such as reinforcement learning to dynamically learn the optimal allocation strategy.

[0004] However, existing technologies still have the following shortcomings: First, insufficient relationship modeling capabilities. While some existing intelligent methods can handle time-series data, they lack effective modeling tools for the complex and heterogeneous graph relationships formed between "customers-work orders-agents," making it difficult to uncover deep connections beyond static labels, such as the historical interaction performance between an agent and a specific customer group, resulting in limited matching accuracy. Second, singular optimization goals. The optimization goals of most systems pursue short-term operational indicators such as first-time resolution rate or average processing time. This singular goal orientation often sacrifices service quality, fails to maintain long-term customer value, and ignores work order allocation as a means to improve agent capabilities, lacking consideration for long-term strategic value such as "agent growth potential." Finally, static allocation strategies. Allocation decisions are usually passively triggered when a work order is generated. Once allocation is completed, even if the system state changes favorablely, such as a more suitable agent becoming idle, the system cannot dynamically adjust, missing opportunities for immediate optimization and lacking the ability to dynamically manage allocated but unprocessed tasks. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for dynamic work order allocation based on multi-dimensional features, so as to solve the problems in the existing work order allocation model that fails to make full use of the complex relationships between entities, has a single optimization objective, and has a rigid allocation strategy that cannot be dynamically adjusted.

[0006] To achieve the above objectives, this invention provides a method for dynamic work order allocation based on multi-dimensional features, comprising the following steps: Obtain multi-dimensional dynamic features: Obtain and integrate data from customer relationship management systems, contact center systems, and human resources systems to construct multi-dimensional dynamic feature vectors that include information on customers, work orders, and agents; Relationship modeling: Graph neural networks are used to model the heterogeneous relationships between customers, work orders, and agents in order to learn embedded representations that reflect the relationship information between entities; Multi-objective reward function definition: Define a composite reward function to evaluate allocation decisions, wherein the composite reward function comprehensively considers at least two business objectives among immediate resolution efficiency, service quality, long-term customer value, operating costs, and agent growth. Strategy learning and allocation decision: Using a reinforcement learning framework, based on the system state including the current work orders to be assigned and the status of available agents, an optimal allocation strategy aimed at maximizing the long-term cumulative reward defined by the composite reward function is learned and executed to make allocation decisions; Dynamic reallocation: Continuously monitor the system status and, when preset conditions are met, use the optimal allocation strategy to dynamically reallocate work orders that have been allocated but not yet processed.

[0007] Optionally, the multidimensional dynamic feature vector includes: a customer feature vector containing customer lifetime value or churn risk; a work order feature vector containing customer sentiment or problem complexity; and an agent feature vector containing dynamic workload, historical success rate, or skill level.

[0008] Optionally, the graph neural network is a graph attention network or a graph convolutional network.

[0009] Optionally, the reinforcement learning framework employs a proximal policy optimization algorithm or a deep Q-network algorithm for policy learning.

[0010] Optionally, the preset conditions include at least one of the following: the waiting time for a high-value customer's work order exceeds a threshold; an agent with a higher skill rating or a better match to the work order characteristics changes from "busy" to "idle"; or the shift of an assigned agent is expected to end before it completes processing.

[0011] Optionally, in the multi-objective reward function definition step, the agent growth objective is achieved in the following way: when the complexity of the work order assigned to the agent is higher than the agent's current skill level, but lower than the sum of the skill level and a preset complexity threshold, a positive reward is given.

[0012] As a further aspect of the present invention, the above-mentioned work order dynamic allocation method based on multi-dimensional features further includes: after allocating the work order to a selected agent, pushing auxiliary information related to the work order to the terminal of the selected agent.

[0013] This invention also provides a dynamic work order allocation system based on multi-dimensional features, comprising: The data acquisition module is used to acquire and integrate data from the customer relationship management system, contact center system and human resources system to construct a multi-dimensional dynamic feature vector containing information on customers, work orders and agents, and to obtain the real-time availability status of agents; The graph neural network modeling module is used to model the heterogeneous relationships between customers, work orders, and agents using graph neural networks, so as to learn an embedded representation that reflects the relationship information between entities; The reinforcement learning decision-making module is used for: Define a composite reward function that takes into account at least two of the following business objectives: immediate resolution efficiency, service quality, long-term customer value, operating costs, and agent growth. Receive the embedded representation generated by the graph neural network modeling module; Using a reinforcement learning framework, based on the system state including the embedded representation and the real-time availability state of the agents, an optimal allocation strategy aimed at maximizing the long-term cumulative reward defined by the composite reward function is learned and executed to make allocation decisions; The system monitoring module continuously monitors the system status and dynamically reallocates assigned but unprocessed work orders when preset conditions are met.

[0014] Optionally, the graph neural network modeling module is configured to employ a graph attention network or a graph convolutional network.

[0015] Optionally, the reinforcement learning decision module is configured to use a proximal policy optimization algorithm or a deep Q-network algorithm for policy learning.

[0016] Compared with the prior art, the present invention has the following beneficial effects: 1. By using graph neural networks to model the heterogeneous relationship between "customer-work order-agent", we can uncover implicit matching relationships that go beyond traditional skill tags, thereby assigning work orders to agents with higher matching degrees and improving the accuracy of matching.

[0017] 2. By introducing a multi-objective composite reward function that includes "long-term customer value" and "salesperson growth potential", work order allocation is upgraded from passive task matching to proactive, multi-objective dynamic resource optimization. This can ensure short-term operational efficiency while taking into account long-term strategic goals such as maintaining high-value customers and developing employees.

[0018] 3. By establishing a continuous monitoring and dynamic reallocation mechanism, the system is given the ability to make real-time adjustments after environmental changes. It can urgently allocate suitable resources for high-value customers who have been waiting for too long, or proactively optimize the task queue when the agent status changes, thereby seizing fleeting optimization opportunities and improving the overall service level. Attached Figure Description

[0019] Figure 1 A schematic diagram of the architecture and business process of a work order dynamic allocation system based on multi-dimensional features provided in an embodiment of the present invention; Figure 2 A flowchart illustrating a work order dynamic allocation method based on multi-dimensional features provided in an embodiment of the present invention; Figure 3 This is a signaling interaction timing diagram of the work order dynamic allocation process in one embodiment of the present invention; Figure 4 This is a schematic diagram of the functional module structure of a work order dynamic allocation system based on multi-dimensional features, provided in an embodiment of the present invention.

[0020] The module consists of: 10. Data acquisition module; 20. Feature engineering module; 30. Graph neural network modeling module; 40. Reinforcement learning decision module; 50. System monitoring module; 100. Data layer; S101. Multidimensional dynamic feature acquisition; S102. Graph neural network embedding learning; S103. Reinforcement learning policy learning and execution; S104. Monitoring and dynamic reallocation check; 200. Model training and decision layer; 300. Execution and feedback layer. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it.

[0022] Example 1 This embodiment provides a basic implementation scheme for a work order dynamic allocation method and system based on multi-dimensional features. It aims to fully demonstrate the entire process from data acquisition, feature modeling, intelligent decision-making to dynamic optimization, so as to achieve accurate, efficient and multi-objective driven management of work order allocation.

[0023] Please see Figure 1This paper illustrates the overall architecture and business process of a dynamic work order allocation system according to an embodiment of the present invention. Logically, the system architecture can be divided into a data layer 100, a model training and decision-making layer 200, and an execution and feedback layer 300. The data layer 100, as the data foundation, may include customer relationship management systems, contact center systems (such as Computer Telephony Integration (CTI) and Automatic Call Distribution (ACD)), and human resource systems. The model training and decision-making layer 200 is the core of this technical solution, responsible for executing the computation and decision-making logic of intelligent allocation. It may include key steps such as multi-dimensional dynamic feature acquisition S101, graph neural network embedding learning S102, reinforcement learning strategy learning and execution S103, and monitoring and dynamic reallocation checking S104. The execution and feedback layer 300 is responsible for executing allocation actions and collecting results to feed back to the model layer, thereby forming a continuously optimized closed loop.

[0024] Please see Figure 4 The diagram illustrates the system functional module structure for implementing the above method. This system can be embodied as a work order intelligent allocation server deployed on a server or cloud platform, which mainly includes: a data acquisition module 10, a feature engineering module 20, a graph neural network modeling module 30, a reinforcement learning decision-making module 40, and a system monitoring module 50. All modules work collaboratively to execute the method proposed in this invention.

[0025] Combination Figure 2 The method flowchart shown in this embodiment may include the following specific steps: Step S201: Obtain multidimensional dynamic features.

[0026] In step S201, the data acquisition module 10 and the feature engineering module 20 work together to acquire multidimensional dynamic features. Specifically, the data acquisition module 10 acquires raw data in real time or near real time from multiple heterogeneous business systems constituting the data layer 100 through application programming interfaces or other data integration methods.

[0027] For example, the data acquisition module 10 may perform the following operations: Obtain static and dynamic information about customers from the customer relationship management system. Static information may include basic customer profiles, industry, and contracted products; dynamic information may include the customer's lifecycle stage, customer lifetime value score calculated based on historical consumption behavior, and customer churn risk score predicted based on data such as recent interaction frequency and service request type.

[0028] Retrieve ticket information and agent status from the contact center system. Ticket information may include the original request text submitted by the customer via telephone, online chat, email, etc., the urgency of the ticket, and the product line it belongs to. Agent status is dynamic and includes "idle," "on call," "following up," and "offline," as well as the length of the currently assigned but unprocessed ticket queue.

[0029] Obtain agent information from the human resources system, such as their skill level assessment (e.g., junior, intermediate, senior), languages ​​they know, salary level, and historical performance appraisal results.

[0030] After acquiring the raw data, the feature engineering module 20 processes and transforms the acquired data to construct a multidimensional dynamic feature vector that can be used by the model. This process may include: Construction of customer feature vectors: Numerical features such as customer lifetime value score (e.g., a floating-point number from 0 to 10) and churn risk score (e.g., a percentage from 0% to 100%) can be used directly or after normalization.

[0031] Construction of work order feature vectors: Natural language processing techniques are applied to the text content of work orders. For example, sentiment analysis models can be used to extract customer sentiment scores (such as continuous values ​​from -1 to 1, representing the intensity of sentiment from negative to positive) from customer descriptions; text classification or keyword extraction techniques can also be used to assess the technical complexity of the problem (e.g., quantifying it into levels from 1 to 10). For classification features such as channel type, one-hot encoding can be used for transformation.

[0032] Construction of Agent Feature Vector: This agent feature vector can include static, semi-dynamic, and fully dynamic features. Static features include skill levels provided by the human resources system; semi-dynamic features include the historical success rate of agents' interactions with specific customer groups (such as VIP customers) or specific problem types (such as refund issues), which can be periodically calculated and updated based on data from the past period (such as three months); fully dynamic features are the agent's real-time workload saturation, which reflects the agent's current level of busyness.

[0033] In this embodiment, the saturation can be quantified using the following formula: , For example, if an agent currently has 3 work orders being processed, with an average processing time of 20 minutes, and a working day is 8 hours (480 minutes), then their current saturation level is (3 20) / 480=12.5%. Understandably, this indicator can more accurately reflect the potential reception capacity of the agents compared to the simple "busy / idle" status.

[0034] Step S202: Construct a heterogeneous graph and model it using a graph neural network.

[0035] In step S202, the graph neural network modeling module 30 constructs a heterogeneous graph and uses a graph neural network for modeling. To delve deeper into the complex relationships between "customer-work order-sales agent" that go beyond surface labels, this embodiment constructs a heterogeneous graph to represent the relationships between them.

[0036] The process of constructing a heterogeneous graph is as follows: Nodes: The graph contains three types of nodes: customer nodes, work order nodes, and agent nodes. Each node is associated with a feature vector constructed in step S201.

[0037] Edges: Edges represent the relationships between nodes. For example, a customer "creating" a work order corresponds to an edge of type "customer-work order"; an agent "processing" a work order corresponds to an edge of type "agent-work order"; an agent's history of "serving" a customer corresponds to an edge of type "agent-customer". Furthermore, edges can be constructed based on attribute similarity. For example, if two work orders both involve "refunds", a similarity edge of "work order-work order" can be created.

[0038] After the heterogeneous graph is constructed, the graph neural network modeling module 30 can use the graph attention network GAT to model the heterogeneous graph. This process corresponds to Figure 1 The GNN embedding learning in S102. The characteristic of the Graph Attention Network (GAT) is that, when aggregating neighbor node information, it can assign different weights to different neighbor nodes through an attention mechanism. In the work order assignment scenario, this means that the graph neural network modeling module 30 can learn which historically relevant customer or agent information is more important for a specific work order. For example, when searching for an agent for a VIP customer's work order, the GAT model may automatically assign higher attention weights to agent nodes that have historically successfully served that VIP customer.

[0039] Through the propagation and aggregation of the multi-layer graph attention network (GAT), the graph neural network modeling module 30 ultimately learns a low-dimensional, dense embedding representation for each node (customer, work order, agent) in the heterogeneous graph. This embedding representation, as a vectorized expression, not only contains the node's own feature information but also implies its contextual relationship information in the heterogeneous graph structure, thus forming an implicit, data-driven deep profile.

[0040] Step S203: Construct a reinforcement learning environment and define a multi-objective reward function.

[0041] Step S203 involves constructing a reinforcement learning environment and defining a multi-objective reward function, which is implemented by the reinforcement learning decision module 40. This invention models the work order allocation problem as a reinforcement learning process.

[0042] State: The state of the environment is a comprehensive description of the current snapshot of the system. At any decision time t, the state St may include: the set of embedded representations of all currently pending work orders, and the set of embedded representations of all currently available (e.g., in an "idle" state) agents and their real-time dynamic features (such as real-time saturation).

[0043] Action: An action At is defined as a pair that performs one work order assignment, that is, selecting a work order i to be assigned and an available agent j to form a pair (i, j). The action space is all possible combinations of work orders and available agents.

[0044] Reward Function: To achieve comprehensive optimization of multiple business objectives, this embodiment designs a composite reward function R. When an action (i.e., an allocation decision) is executed, the system will calculate a reward value based on the business results generated at some future time (such as when a work order is closed).

[0045] This composite reward function can be expressed as a weighted sum of multiple sub-reward items: , in: It is used to characterize the efficiency of immediate resolution; for example, it can be inversely proportional to the work order processing time, with higher rewards for faster processing.

[0046] Used to characterize service quality, it is usually derived from customer satisfaction ratings after a work order is closed; the higher the rating, the higher the reward.

[0047] Maintenance measures are used to characterize long-term customer value. For example, if a high-value customer's work order is resolved efficiently and with high quality, an additional positive reward is given; conversely, if it leads to a customer complaint, a negative reward is given.

[0048] Used to characterize operating costs, it is inversely proportional to the salary level of the agent handling the work order; using a lower-cost agent to complete the task can earn a positive reward.

[0049] Used to characterize agent growth potential, it aims to encourage the assignment of challenging tasks to agents to promote their skill development. These are the weighting coefficients for each reward, and these weights can be adjusted by the business administrator according to the business strategy at different times.

[0050] Step S204: Generate assignment / reassignment actions based on the current state and learning strategy.

[0051] In step S204, the reinforcement learning decision module 40 generates an assignment or reallocation action based on the current state and the learned policy. This process corresponds to... Figure 1 The reinforcement learning policy learning and execution step S103 is described in this embodiment. In this example, the Proximal Policy Optimization (PPO) algorithm can be used to train a policy network. It is understood that PPO, as a policy gradient algorithm, has the characteristics of stable training and high sample efficiency. The policy network is a deep neural network whose input is the current environmental state St (i.e., the embedding representation and dynamic features of work orders and agents), and its output is the probability distribution of executing each possible action At in that state. During the training phase, the reinforcement learning agent collects data (state, action, reward) through continuous interaction with the environment (which can be a real business system or an offline simulator), and updates the parameters of the policy network using the PPO algorithm. The goal is to enable the policy network to generate actions that maximize long-term cumulative rewards. After training, during actual allocation, the system inputs the current state into the trained policy network. The policy network then assigns a preference score or probability to each possible pairing (work order, agent). The system can choose the action with the highest probability or sample based on the probability to make the final allocation decision.

[0052] Step S205: Perform the action.

[0053] The reinforcement learning decision module 40 sends the generated allocation decision (e.g., "assign work order T123 to agent A007") to the corresponding component (such as the ACD system) in the execution and feedback layer 300, which then performs the actual operation of pushing the work order to the agent terminal.

[0054] Step S206: Collect execution results and provide feedback to optimize the model.

[0055] After a work order is processed and closed, the execution and feedback layer 300 collects various outcome metrics for that work order, such as actual processing time, customer satisfaction rating, and whether the issue was resolved on the first contact. These results are used to calculate the reward R defined in step S203. This complete experience sample is stored in the experience replay pool for subsequent iterative training of the PPO algorithm, thereby enabling the policy model to continuously evolve.

[0056] To illustrate the working process of this embodiment in detail, the following will be combined with... Figure 3 The signaling interaction sequence diagram is illustrated using an exemplary scenario. Suppose Mr. Li, a VIP customer of a financial company, submitted a work order via a mobile app regarding his inability to redeem a complex wealth management product.

[0057] The user (Mr. Li) submits a work order to the work order allocation system through the client (interaction 301).

[0058] After receiving a new work order, the work order allocation system triggers an allocation process internally (interaction 302).

[0059] The data acquisition module 10 is activated and sends a request to the customer relationship management system / human resources system to enhance the data (interaction 303) to obtain Mr. Li's customer value (e.g., 9.5 / 10), churn risk (e.g., 80%), and the latest status and profiles of all available agents.

[0060] The customer relationship management system / human resources system returned the requested data (interaction 304).

[0061] The feature engineering module 20 analyzes the work order text and determines that the problem is highly complex (e.g., 9 / 10) and the customer is anxious (e.g., -0.8).

[0062] The system enters the GNN / RL decision-making process (interaction 305). The graph neural network modeling module 30 utilizes a pre-trained GAT model, combining the features of the new work order and the current agent, to generate embedded representations of all relevant nodes. By analyzing historical interaction graphs, the GAT model may discover that Agent A (with an intermediate skill level), while not a top expert in financial products, has historically had a very high success rate in handling complex complaints from clients with similar backgrounds to Mr. Li (such as high-net-worth individuals with cautious personalities), forming an implicit profile of a "high-net-worth client communication expert." Agent B, while possessing the highest skill level in financial products, has historical interaction data showing a tendency to trigger secondary complaints when handling emotionally agitated clients.

[0063] The reinforcement learning decision module 40 receives these embedded representations as states. Since the current business strategy sets a very high weight w3 for maintaining high-value customers, the PPO policy network, after evaluation, determines that assigning the work order to agent A will yield a higher long-term expected reward, as this effectively reduces the risk of customer churn.

[0064] Therefore, the system ultimately decides to assign the work order to agent A and push the work order to agent A's terminal (interaction 306).

[0065] As an optional enhancement, while pushing the work order, the system can also automatically retrieve redemption failure case analysis, standard processing procedures, and key communication tips for anxious customers related to the financial product from the knowledge base, and push them to Agent A's terminal, thus realizing the combination of allocation decision and auxiliary information push.

[0066] Through the close coordination of the above steps, this embodiment realizes a dynamic work order allocation scheme that can understand deep relationships, balance multiple business objectives, and has the ability to learn and evolve.

[0067] Example 2 This embodiment illustrates that the technical solution proposed in this invention has flexibility and scalability in model selection, and its core idea is not limited to a specific graph neural network or reinforcement learning algorithm. This embodiment will demonstrate that using a different model combination than in Embodiment 1, namely the graph convolutional network (GCN) and the deep Q-network (DQN) algorithm, can also achieve the technical solution of this invention.

[0068] In this embodiment, the overall system architecture, functional modules, and method flow can be kept consistent with those of Embodiment 1. The main difference lies in the specific implementation of the graph neural network modeling module 30 and the reinforcement learning decision module 40.

[0069] 1. Variant Implementation of Relationship Modeling: Using a Graph Convolutional Network (GCN) In Example 1, the graph neural network modeling module 30 uses a graph attention network (GAT). As an optional implementation, this embodiment can use a graph convolutional network (GCN) to complete the GNN embedding learning in step S202.

[0070] Graph Convolutional Networks (GCNs) are another widely used graph neural network model. Their core idea is to generate a new feature representation for a node in the next layer by weighting the feature vectors of all its first-order neighbors (the weights are usually related to the node's degree) and then applying a non-linear activation function. By stacking multiple layers of GCNs, a node can aggregate information from its multi-order neighbors.

[0071] Compared to GAT, Graph Convolutional Networks (GCNs) typically do not distinguish the importance of neighbors when aggregating neighbor information; instead, they apply equal or degree-based simple weighting. This makes the computation process of GCNs relatively simple, with fewer model parameters, and training speed is generally faster. Although it may not be able to capture the differentiated importance of different neighbors as finely as GAT, GCNs can still learn embedding representations that reflect graph structure and node features very effectively in many scenarios.

[0072] Therefore, the graph neural network modeling module 30 can be configured to use GCN for modeling. After the constructed heterogeneous graph of "customer-work order-agent" and node features are input into the GCN model, after multi-layer convolution operations, an embedding representation rich in contextual information can be generated for each customer, work order, and agent node. These embedding representations are then passed to the reinforcement learning decision module 40 for subsequent allocation decisions.

[0073] 2. Variant Implementation of Dynamic Allocation Decision: In Example 1, the reinforcement learning decision module 40 employs the policy-based proximal policy optimization (PPO) algorithm. As an optional implementation, this example can use the value-based deep Q-network (DQN) algorithm to complete the policy learning and allocation decision in step S204. DQN combines deep neural networks with the classic Q-Learning algorithm. Its core idea is to train a neural network (called a Q-network) to approximate the state-action value function Q(s,a). This value function predicts the sum of future cumulative rewards that can be obtained by performing action a in state s and following the current optimal policy thereafter. It's important to note that applying the DQN algorithm typically requires discretizing the action space. In a work order assignment scenario, the action space is all pairs of (work orders, available seats). If the number of work orders to be assigned is N and the number of available seats is M, then the size of the action space is N×M. The DQN workflow may include: Q-Network Construction: The input to a Q-network is the environment state s (e.g., composed of embedded representations of a specific work order and a specific agent), and the output is the Q-value for performing the assigned action in that state. Alternatively, it can be designed so that the input is the state s and the output is a list of Q-values ​​for all possible actions.

[0074] Policy Execution: During decision-making, for each work order i to be assigned, the system iterates through all available agents j. For each pairing (i,j), its corresponding state is input into the trained Q-network to obtain the Q-value Q(s_ij,a_ij). The system then selects the agent j that maximizes the Q-value. To assign work order i, i.e. j = argmax_j Q(s_ij, a_ij).

[0075] Model Training: DQN is trained using techniques such as experience replay and target networks. Samples (state, action, reward, next state) generated by the agent's interaction with the environment are stored in a replay buffer. During training, a batch of samples is randomly drawn from the buffer, and the weights of the Q-network are updated using temporal difference error, so that its predicted Q-value gets closer and closer to the true future cumulative reward.

[0076] By replacing the algorithmic core of the reinforcement learning decision module 40 with DQN, the system is also able to learn an allocation strategy aimed at maximizing long-term multi-objective composite rewards.

[0077] In summary, this embodiment, by using GCN and DQN as alternatives to GAT and PPO respectively, aims to demonstrate that the core technical concepts of this invention—namely, "modeling entity relationships using graph neural networks" and "performing multi-objective dynamic decision-making using a reinforcement learning framework"—are universal methodologies. This shows that any graph neural network model capable of effectively learning graph embedding representations, and any reinforcement learning algorithm capable of handling complex state and action spaces and optimizing long-term rewards, can be applied to the technical framework of this invention, thus demonstrating the versatility of the invention's solution.

[0078] Example 3 This embodiment aims to illustrate the flexibility and strategic value of the multi-objective composite reward function in the method of the present invention, especially how to guide the work order allocation system to serve the specific business strategies of enterprises at different stages of development by adjusting the weights of different objectives in the reward function.

[0079] This embodiment follows the system architecture and model selection described in Embodiment 1, namely, it uses a combination of the Graph Attention Network (GAT) and the Proximal Policy Optimization (PPO) algorithm. The focus of this embodiment is the composite reward function defined in step S203: , And how to strategically configure weighting coefficients This is to achieve different business objectives. Understandably, the weighting coefficients can be set by the system administrator (e.g., through the strategy configuration interface), thereby enabling flexible adjustments to business strategies without modifying the underlying code.

[0080] Scenario A: Cost Control First Strategy In a scenario (e.g., a company during a quarter or a period of stable business), its primary strategic objective is to control operating costs and maximize profits. In this case, the system administrator can configure the weights in the reward function as follows: (Cost-related weight): Increase it, for example, set it to 0.5.

[0081] (Efficiency-related weight): Keep it at a high value, such as 0.3, because efficiency also indirectly affects cost.

[0082] (Quality and Customer Value Weighting): Maintain a moderate level, for example, 0.1 for each, to ensure that service quality does not decline significantly.

[0083] (Seat growth weight): Lower it, for example, set it to 0 or a very small value.

[0084] With this weighting configuration, the reinforcement learning agent, when learning strategies, will discover those that bring high returns. The rewarded action is a "good" action. Since the calculation method is inversely proportional to the cost of the selected agent (such as the salary level), the system will exhibit the following tendency when making allocation decisions: for a work order, if multiple agents have the skills to meet the processing requirements, the system will preferentially select the agent with a lower salary level among them. Even if assigning the task to an experienced, senior agent with a higher salary can complete the task slightly faster, as long as that junior agent can also complete the task within an acceptable time, the system will choose the latter in order to obtain a higher cost reward. In this way, the entire work order allocation system becomes an automated tool for implementing the cost control strategy.

[0085] Scenario B: New Employee Rapid Growth Strategy In another scenario (for example, when an enterprise is in a period of rapid expansion and has recruited a large number of new agents), the main contradiction of the enterprise may be the insufficient skills of new employees, and the overall service capability of the team needs to be improved. At this time, the strategic goal is transformed into enabling new employees to grow up as soon as possible. To this end, the system administrator can configure the weight of the reward function to the "growth first" mode: (Agent growth weight): significantly increased, for example, set to 0.6.

[0086] : correspondingly reduced, for example, accounting for a total weight of 0.4.

[0087] Among them, the quantification of the agent growth reward is the key to implementing this strategy. This embodiment provides a specific, operable quantification method. First, the system needs to quantitatively rate the complexity of the work order and the skill level of the agent, for example, both using a scale of 1-10. The work order complexity C_ticket comes from the evaluation of the feature engineering module 20, and the current skill level S_agent of the agent comes from the human resources system.

[0088] The logic of the growth reward is: when the work order assigned to the agent constitutes an "achievable challenge" for the agent, a positive reward is given. Specifically, when the complexity of the work order is higher than the current skill level of the agent but does not exceed the preset complexity threshold T_c (for example, T_c=2), the assignment is considered to be conducive to growth.

[0089] Therefore, the calculation rule can be defined as: if S_agent<C_ticket ≤ (S_agent + T_c), a higher positive reward is given, for example, R_gro = +10. Otherwise, R_gro = 0.

[0090] For example: A new agent with a current skill level of S_agent of 3.

[0091] A work order A has a complexity C_ticket of 4. Since 3 < 4 ≤ (3 + 2), this work order is a good learning opportunity for the agent. If the system assigns work order A to him, he will receive a growth bonus of +10.

[0092] A work order B has a complexity C_ticket of 2. This task is too simple and has no growth value, so R_gro is 0.

[0093] A work order C has a complexity of C_ticket of 6. Since 6 > (3 + 2), the task may be too difficult for the new agent, which may lead to customer complaints and agent frustration. Therefore, no growth bonus is given, and R_gro is 0.

[0094] With a "growth-first" weighting, reinforcement learning agents will strive to achieve high performance when making decisions. Rewards. Therefore, the system proactively and intelligently filters and assigns new employees work orders that are slightly above their current capabilities and appropriately challenging. Even if assigning work order A to a senior agent with a skill level of 8 would allow for faster completion (and thus higher rewards), the system will still prioritize this type of task. ), but because The weight is much higher than The system's final calculations will show that the long-term total reward for assigning work orders to new agents (which takes into account the value of talent development) is higher.

[0095] As can be seen from the comparison of the above scenarios, the multi-objective reward function mechanism of this invention is not only a static technical model, but also a flexible and configurable business strategy execution engine. It enables the underlying operation of work order allocation to maintain dynamic consistency with the strategic intentions of senior management, thereby helping to enhance the system's business value and strategic adaptability.

[0096] Example 4 This embodiment illustrates the dynamic redistribution mechanism in the method of the present invention, and details its diverse triggering conditions and added value in actual business scenarios. Dynamic redistribution is a key feature that distinguishes the present invention from traditional static redistribution methods. It gives the system the ability to actively seek better solutions based on continuous changes in the environmental state after redistribution is completed.

[0097] This mechanism is mainly implemented collaboratively by the system monitoring module 50 and the reinforcement learning decision-making module 40, corresponding to Figure 1The monitoring and dynamic reallocation check S104 and the RL policy learning and execution S103 are called again. The system monitoring module 50, as an independent background process, continuously monitors the key state variables of the system. When it detects that the preset reallocation triggering conditions are met, it will activate the reinforcement learning decision module 40 to re-determine one or more work orders that have been assigned but have not yet started processing.

[0098] This embodiment will describe in detail several typical preset triggering conditions.

[0099] Triggering condition 1: The waiting time for a high-value customer's work order exceeds a preset threshold. In customer service centers, response speed to high-value customers is a key indicator of service level. This system allows for different service level agreements (SLAs) for different customer tiers; for example, the processing time for VIP customer work orders should not exceed 5 minutes. The workflow may include: Monitoring: The system monitoring module 50 continuously scans the status of all work orders in the pending queue. For a work order T1 associated with a high-value customer, it records the waiting time (wait_time) after the work order was assigned.

[0100] Trigger: Suppose work order T1 was initially assigned to agent A, but agent A is handling a lengthy call that has not yet ended. When the system monitoring module 50 detects that the wait_time of work order T1 exceeds a preset threshold T_wait_vip (e.g., 5 minutes), the dynamic reallocation condition is met.

[0101] Decision-making and execution: A reallocation is triggered. At this point, the system initiates a new decision-making process, which proceeds as follows: Figure 3 The reallocation interaction 307 is triggered by the system monitoring module 50. The system will request data again (interaction 308) to obtain the latest status of all current agents and receive the returned new data (interaction 309). The reinforcement learning decision module 40 is invoked to re-execute the decision (interaction 310). It will re-evaluate the long-term reward of assigning work order T1 to all currently available agents (including agent B, which may have just become idle). If an available agent B is found, work order T1 will be immediately reallocated to the currently idle expert agent B. Although this will result in a switch, it will yield a higher reward because it can immediately stop the waiting of high-value customers and avoid the risk of customer churn. If the reward is significantly higher than maintaining the status quo, the system will generate a redistribution decision.

[0102] Execution: The system withdraws work order T1 from agent A's queue and pushes the reassigned work order to agent B's terminal (interaction 311).

[0103] Triggering condition two: The best-matched agent changes from "busy" to "available". Optimal matching is dynamic. At the moment a work order is generated, the optimal agent may be busy. This mechanism can seize optimization opportunities when that agent becomes idle. Its operation may include: Initial Assignment: A work order T2 regarding a complex network fault is generated. At the time of assignment, network expert agent C is busy. Based on the suboptimal choice at the time, the system assigns work order T2 to the more versatile general agent D, and work order T2 enters agent D's pending queue.

[0104] Status Change and Monitoring: Ten minutes later, expert C completed his previous task, and his status changed to idle. The system monitoring module 50 captured this status change in real time.

[0105] Triggering and Decision-Making: After detecting that agent C has become idle, the system monitoring module 50 checks whether there are any work orders in the current queue that highly match agent C's "implicit profile" (provided by the GNN model) or explicit skill tag. It finds that work order T2 is such an example, so the reallocation condition is met. The reinforcement learning decision-making module 40 is activated to perform a "hypothesis analysis": If T2 is reallocated from D to C, the expected overall reward (higher solution efficiency) is... and customer satisfaction Whether the reward for maintaining the status quo can be covered. Since expert C has a strong historical track record in handling such problems, the model determines that the reallocation is more likely to be favorable.

[0106] Execution: The system removes work order T2 from the queue of agent D and reassigns it to expert agent C.

[0107] Triggering condition 3: The shifts with assigned seats are expected to end before their processing is completed. This mechanism aims to proactively optimize the continuity of resource utilization and avoid service interruptions or poor handover due to agents leaving get off work.

[0108] Its working process may include: Prediction and Monitoring: Agent E's off-duty time is 6 PM. The system monitoring module 50 not only knows the agents' shift schedule, but also combines the characteristics of work orders (such as complexity and type) and the agents' historical processing data to predict the approximate processing time required for each work order. Assuming that at 5:30 PM, there is an assigned but not yet started work order T3 in Agent E's queue, the system predicts its processing time (predict_duration) to be 45 minutes.

[0109] Trigger: System monitoring module 50 performs a prospective check and discovers... The current time plus `predict_duration(5:30+0:45=6:15)` is later than `shift_end_time(6:00)`. This means that if agent E handles this work order, there is a risk that it may not be completed on time, requiring work order handover, which increases communication costs and the risk of errors. This conflict of "estimated completion time exceeding shift end time" constitutes the trigger condition for reassignment.

[0110] Decision-making and execution: The reinforcement learning decision module 40 is invoked to seek a better scheduling scheme. One possible decision is to remove work order T3 from the queue of agent E and reassign it to agent F, whose shift has ample time. Simultaneously, to make full use of agent E's last 30 minutes of work time, the system may look for a short work order T4 with a predicted processing time of less than 30 minutes from the allocation pool and assign it to agent E.

[0111] Through the above specific scenarios, this embodiment demonstrates that the dynamic reallocation mechanism effectively enhances the system's adaptability and timeliness. It transforms work order allocation from a "one-off" static decision into a continuous, dynamic optimization process, enabling the system to proactively respond to various unexpected state changes and seize every fleeting optimization opportunity, thereby achieving optimal allocation of service resources at a higher level.

[0112] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for dynamic work order allocation based on multi-dimensional features, characterized in that, Includes the following steps: Obtain multi-dimensional dynamic features: Obtain and integrate data from customer relationship management systems, contact center systems, and human resources systems to construct multi-dimensional dynamic feature vectors that include information on customers, work orders, and agents; Relationship modeling: Graph neural networks are used to model the heterogeneous relationships between customers, work orders, and agents in order to learn embedded representations that reflect the relationship information between entities; Multi-objective reward function definition: Define a composite reward function to evaluate allocation decisions, wherein the composite reward function comprehensively considers at least two business objectives among immediate resolution efficiency, service quality, long-term customer value, operating costs, and agent growth. Strategy learning and allocation decision: Using a reinforcement learning framework, based on the system state including the current work orders to be assigned and the status of available agents, an optimal allocation strategy aimed at maximizing the long-term cumulative reward defined by the composite reward function is learned and executed to make allocation decisions; Dynamic reallocation: Continuously monitor the system status and, when preset conditions are met, use the optimal allocation strategy to dynamically reallocate work orders that have been allocated but not yet processed.

2. The work order dynamic allocation method based on multi-dimensional features according to claim 1, characterized in that, The multidimensional dynamic feature vector includes: Customer feature vectors that include customer lifetime value or churn risk; The work order feature vector includes customer sentiment or problem complexity; And agent feature vectors that include dynamic workload, historical success rate, or skill level.

3. The work order dynamic allocation method based on multi-dimensional features according to claim 1, characterized in that, In the relationship modeling step, the graph neural network is either a graph attention network (GAT) or a graph convolutional network (GCN).

4. The method for dynamic work order allocation based on multi-dimensional features according to claim 1, characterized in that, In the policy learning and allocation decision-making steps, the reinforcement learning framework employs the Proximal Policy Optimization (PPO) algorithm or the Deep Q-Network (DQN) algorithm for policy learning.

5. The work order dynamic allocation method based on multi-dimensional features according to claim 2, characterized in that, In the dynamic reallocation step, the preset conditions include at least one of the following: The wait time for work orders from high-value customers exceeds the preset threshold; Agents with higher skill ratings or better matching work order characteristics will change from busy to idle. The shifts with assigned seats are expected to end before their processing is completed.

6. The method for dynamic work order allocation based on multi-dimensional features according to claim 1, characterized in that, Also includes: In the multi-objective reward function definition step, the agent growth objective is achieved in the following way: when the complexity of the work order assigned to the agent is higher than the agent's current skill level, but lower than the sum of the current skill level and the preset complexity threshold, a positive reward is given.

7. The method for dynamic work order allocation based on multi-dimensional features according to claim 1, characterized in that, After assigning or dynamically reassigning a work order to a selected agent, auxiliary information related to the work order is pushed to the terminal of the selected agent.

8. A work order dynamic allocation system based on multi-dimensional features, characterized in that, include: The data acquisition module is used to acquire and integrate data from the customer relationship management system, contact center system and human resources system to construct a multi-dimensional dynamic feature vector containing information on customers, work orders and agents, and to obtain the real-time availability status of agents; The graph neural network modeling module is used to model the heterogeneous relationships between customers, work orders, and agents using graph neural networks, so as to learn an embedded representation that reflects the relationship information between entities; The reinforcement learning decision-making module is used for: Define a composite reward function that takes into account at least two of the following business objectives: immediate resolution efficiency, service quality, long-term customer value, operating costs, and agent growth. Receive the embedded representation generated by the graph neural network modeling module; Using a reinforcement learning framework, based on the system state including the embedded representation and the real-time availability state of the agents, an optimal allocation strategy aimed at maximizing the long-term cumulative reward defined by the composite reward function is learned and executed to make allocation decisions; The system monitoring module continuously monitors the system status and dynamically reallocates assigned but unprocessed work orders when preset conditions are met.

9. A work order dynamic allocation system based on multi-dimensional features according to claim 8, characterized in that, The graph neural network modeling module is configured to use either a graph attention network (GAT) or a graph convolutional network (GCN).

10. A work order dynamic allocation system based on multi-dimensional features according to claim 8, characterized in that, The reinforcement learning decision module is configured to use either the Proximal Policy Optimization (PPO) algorithm or the Deep Q Network (DQN) algorithm for policy learning.

Images