Incoming call to marketing scene data processing method and determining method and device, electronic equipment and storage medium

Abstract

The invention discloses an incoming call-to-marketing scene data processing method and device, an incoming call-to-marketing scene determination method and device, electronic equipment and a storage medium, and the method comprises the steps: carrying out the grouping according to historical behaviors in an incoming call-to-marketing scene, and enabling the historical behaviors to comprise a customer and a customer service; performing feature vector extraction based on a multi-field topology network according to a grouping result to obtain feature vectors of the customer and the customer service; and inputting the feature vectors of the customer and the customer service into a multi-target hybrid expert network for training, and outputting a matching rate of the customer and the customer service and a staging response rate of the customer. According to the invention, based on the incoming call-to-marketing scene, on one hand, the defect that the customer and the customer service have obviously different historical behavior characteristics in the incoming call scene is overcome, and on the other hand, the matching rate of the customer and the customer service and the staging response rate of the customer can be output at the same time.

Description

Technical Field

[0001] The present application relates to the technical fields of incoming call transfer to marketing scenarios and telephone customer service, and particularly relates to a data processing method, a determination method device, an electronic device, and a storage medium for an incoming call transfer to marketing scenario. Background Art

[0002] With the continuous development of technology and the improvement of the degree of informatization, telephone customer service plays an important role in banks and various financial service institutions. Its main tasks can be summarized into two major directions. On the one hand, the customer service needs to solve the current urgent problems of customers, such as customer complaints, customer card cancellation requirements, etc. Such requirements need to be quickly processed by the customer service in a short time to appease the strong emotions of customers and reduce the customer complaint rate. On the other hand, for customers with less intense demands and demands on the bill, the customer service is required to carry out financial business marketing for them, such as: bill installment, business handling recommendation, etc. Through the incoming call transfer to marketing service recommendation, appropriate financial services are provided for customers to achieve the purpose of reasonable marketing and increased business income.

[0003] In the related art, there are still situations such as customers and customer services with large differences in matching historical behaviors in telephone customer service. At the same time, due to the problem of large mobility of customer service employees, the traditional customer service connection allocation method is prone to allocate work orders to old employees, resulting in new customer services not being able to be allocated work orders, and the entire system falling into the dilemma of the "Matthew effect". Summary of the Invention

[0004] Embodiments of the present application provide a data processing method, a determination method device, an electronic device, and a storage medium for an incoming call transfer to marketing scenario, so as to reduce the complaint rate in the incoming call transfer to marketing scenario, improve the marketing success rate, and at the same time reduce the problems of mutual dependence, complex processes, and difficult maintenance brought by multiple model combinations in the incoming call transfer to marketing scenario.

[0005] Embodiments of the present application adopt the following technical solutions:

[0006] In a first aspect, an embodiment of the present application provides a data processing method for an incoming call transfer to marketing scenario, wherein the method includes:

[0007] Cluster according to historical behaviors in the incoming call transfer to marketing scenario, where the historical behaviors include customers and customer services;

[0008] Extract feature vectors based on a multi-domain topological network according to the clustering results to obtain feature vectors of customers and customer services; and

[0009] Input the feature vectors of the customers and customer services into a multi-objective hybrid expert network for training, and output the matching rate of the customers and customer services and the installment response rate of the customers.

[0010] In some embodiments, clustering is performed based on historical behaviors in the inbound call transfer to marketing scenario. The historical behaviors include those of customers and customer service representatives, and it includes:

[0011] Based on the historical behaviors in the inbound call transfer to marketing scenario, the customers are divided into those who have never made an inbound call, complaint types, non-complaint types, and balanced types, and the customer service representatives are divided into new employees, complaint types, marketing types, and all-round types, obtaining the clustering results.

[0012] In some embodiments, the multi-domain topological network includes a clustering-exclusive network and a central shared network. Feature vector extraction is performed based on the clustering results using the multi-domain topological network to obtain the feature vectors of customers and customer service representatives, and it includes:

[0013] According to the clustering results, data of different clusters are input into the corresponding exclusive networks, and all data including the customers and the customer service representatives are passed through the central shared FCN network based on the multi-domain topological network to divide the historical behaviors of the customers and customer service representatives into different domains;

[0014] During the process of feature vector extraction based on the multi-domain topological network, customers and customer service representatives with the same behaviors will be divided into FCN networks in the same domain for training according to their historical behaviors; and

[0015] Data including the customers and the customer service representatives are all passed through the central shared FCN network based on the multi-domain topological network. Finally, the vectors output by each domain network will perform a dot product inner product calculation with the shared FCN network through adaptive dynamic parameters to obtain the customer feature vectors and the customer service representative feature vectors.

[0016] In some embodiments, the multi-domain topological network specifically includes:

[0017] At the bottom layer, embedding sharing of the same ID between different domains is adopted,

[0018] The feature data of customers and customer service representatives are concatenated after being embedded into vectors, and the feature normalization results of each domain are obtained using the PN local normalization method;

[0019] According to the vector representation after PN normalization, it is input into a fully connected neural network. The fully connected neural network includes a central FCN network shared by all clustering data, as well as customer and customer service representative exclusive clustering neighborhood FCN networks;

[0020] After the output of the FCN network of each domain is added with the output of the auxiliary network, it is dynamically fused with the shared central FCN network through the topological network structure to obtain the customer feature vectors and the customer service representative feature vectors.

[0021] In some embodiments, the method further includes: setting adaptive fusion parameters in the multi-domain topology network:

[0022] The shared central network and each sub-group domain FCN network are fused through an adaptive parameter matrix with each domain FCN network. The fusion parameters are continuously updated and iterated during training. The shared FCN network updates parameters using all samples, and the FCN networks in the corresponding domains of the customer and the customer service update parameters using samples in the corresponding scenarios.

[0023] In some embodiments, the auxiliary network feature vectors in the multi-domain topology network include:

[0024] Design marked features for each domain, vectorize them through an auxiliary network, and then perform cumulative calculation with the output vector of the star-shaped topology network to obtain the final customer feature vector and the feature vector of the customer service.

[0025] In some embodiments, inputting the feature vectors of the customer and the customer service into a multi-objective hybrid expert network for training, and outputting the matching rate of the customer and the customer service and the installment response rate of the customer, includes:

[0026] Model the matching rate of the customer and the customer service and the installment response rate of the customer using a multi-objective hybrid expert network structure;

[0027] Divide the multi-expert network into a matching expert network and a response rate expert network. The matching expert network is used to receive the vector features of the customer and the customer service, and the response rate expert network is used to receive the vector features of the customer.

[0028] In some embodiments, the method further includes:

[0029] The gating model corresponding to the response rate FCN network only accepts the output vector of the response rate expert network;

[0030] The output vector of the gating model corresponding to the response rate FCN network is disconnected in the network by setting the weight to 0 in the fusion formula, so that the input value of the response rate FCN network only contains the feature of the customer's bill information;

[0031] In addition to using the customer vector and the customer service vector as input features, the original feature vector of the customer's historical incoming call behavior data is introduced as a supplementary input feature.

[0032] In some embodiments, the method further includes:

[0033] Execute a cold start mechanism for new employees in the customer service classification;

[0034] New employees obtain an activation probability value through the cold start strategy;

[0035] When the activation probability value is greater than the threshold, it is determined to select the current customer service for connection marketing, and the installment response rate of the customer is returned and displayed to the customer as a reference for whether to conduct installment marketing.

[0036] In some embodiments, the cold start strategy includes using a Bandit - like algorithm, and Thompson sampling is used as the algorithm for generating the probability value of a new customer service.

[0037] In a second aspect, an embodiment of the present application further provides a method for determining the installment marketing response rate in an incoming call transfer to marketing scenario. Among them, the data processing method described in the first aspect for the incoming call transfer to marketing scenario is used to determine the installment marketing response rate.

[0038] In a third aspect, an embodiment of the present application further provides a method for determining the matching rate in an incoming call transfer to marketing scenario. Among them, the data processing method described in the first aspect for the incoming call transfer to marketing scenario is used to determine the matching rate.

[0039] In a fourth aspect, an embodiment of the present application further provides a data processing device for an incoming call transfer to marketing scenario. Among them, the training device includes:

[0040] A clustering module, configured to perform clustering based on historical behaviors in the incoming call transfer to marketing scenario, where the historical behaviors include customers and customer services;

[0041] An extraction module, configured to extract feature vectors based on the clustering results using a multi - domain topology network to obtain the feature vectors of customers and customer services; and

[0042] An output module, configured to input the feature vectors of the customers and customer services into a multi - objective hybrid expert network for training, and output the matching rate of customers and customer services and the installment response rate of customers.

[0043] In a fifth aspect, an embodiment of the present application further provides an electronic device, including: a processor; and a memory arranged to store computer - executable instructions, where the executable instructions, when executed, cause the processor to execute the above - mentioned method.

[0044] In a sixth aspect, an embodiment of the present application further provides a computer - readable storage medium. The computer - readable storage medium stores one or more programs, and when the one or more programs are executed by an electronic device including a plurality of application programs, the electronic device is caused to execute the above - mentioned method.

[0045] The above at least one technical solution adopted in the embodiments of the present application can achieve the following beneficial effects: clustering is performed according to historical behaviors in the incoming call transfer to marketing scenario, and then feature vectors of customers and customer service representatives are extracted based on the clustering results using a multi-domain topological network. Finally, the feature vectors of the customers and customer service representatives are input into a multi-objective hybrid expert network for training, and the matching rate between the customer and the customer service representative and the installment response rate of the customer are output. Through the above method, the obvious differences in historical behavior characteristics between customers and customer service representatives in the incoming call scenario are overcome. At the same time, the matching rate between the customer and the customer service representative and the installment response rate of the customer are output, enabling the model to "one-stop" output two target results, namely, the customer service representative - customer matching rate and the customer installment response rate.

[0046] In addition, by closely adhering to the incoming call transfer to marketing scenario, a new installment marketing process is constructed: a matching mechanism for customer service representatives and customers is proposed, and the vectors of customers and customer service representatives are refined and modeled using a multi-domain topological network. Then, using the feature vectors as inputs and leveraging the characteristics of the multi-objective hybrid expert network, the customer - customer service representative matching rate and the customer installment response rate are output simultaneously for modeling. The appropriate customer service representative for the customer is selected for connection, and the cold start mechanism is used to ensure that new customer service representatives can also obtain connection opportunities, maintaining the sustainability and closed-loop self-iteration of the system and the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] The drawings described herein are used to provide a further understanding of the present application and form a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation of the present application. In the drawings:

[0048] Figure 1 It is a schematic flowchart of the data processing method for the incoming call transfer to marketing scenario in the embodiments of the present application;

[0049] Figure 2 It is a schematic diagram of the historical behavior clustering process of customers and customer service representatives in the data processing method for the incoming call transfer to marketing scenario in the embodiments of the present application;

[0050] Figure 3 It is a schematic diagram of the multi-domain topological network structure diagram in the data processing method for the incoming call transfer to marketing scenario in the embodiments of the present application;

[0051] Figure 4 It is a schematic diagram of the multi-scenario topological network structure design in the data processing method for the incoming call transfer to marketing scenario in the embodiments of the present application;

[0052] Figure 5 It is a schematic diagram of the auxiliary network feature vector design in the data processing method for the incoming call transfer to marketing scenario in the embodiments of the present application;

[0053] Figure 6Schematic diagram of the multi-objective hybrid expert network structure in the data processing method for the inbound call transfer to marketing scenario in the embodiments of the present application;

[0054] Figure 7 Schematic diagram of the cold start mechanism design in the inbound call transfer to marketing scenario in the data processing method for the inbound call transfer to marketing scenario in the embodiments of the present application;

[0055] Figure 8 Schematic diagram of the structure of the data processing device for the inbound call transfer to marketing scenario in the embodiments of the present application;

[0056] Figure 9 Schematic diagram of the structure of an electronic device in the embodiments of the present application. Detailed implementation manners

[0057] To make the objectives, technical solutions and advantages of the present application clearer, the technical solutions of the present application will be clearly and completely described below in conjunction with the specific embodiments of the present application and the corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present application without creative efforts shall fall within the protection scope of the present application.

[0058] The technical terms involved in the embodiments of the present application are as follows:

[0059] Customer service inbound call transfer to marketing refers to the inbound call transfer to marketing scenario of the customer service system of a bank or financial institution. When a user dials a customer service phone number for help, the customer service staff conducts business marketing while solving the problem according to the user's needs and historical information. While improving customer service efficiency and customer satisfaction, it provides an important business promotion channel for banks and financial institutions.

[0060] User installment response rate generally refers to the probability value of handling installment business estimated by a credit card issuer based on the user's historical behavior and attributes when sending installment promotions to cardholders, which reflects the cardholder's interest and acceptance of the bank's installment service.

[0061] Multi-domain topology network is a special network structure in deep learning for dealing with multi-task / multi-scenario problems. Its main features include: shared topology structure, expert networks unique to tasks or scenarios, and the ability to transfer and share knowledge between different tasks / scenarios, thereby accelerating the learning speed and overall effect of each task / scenario.

[0062] Cold start mechanism generally refers to the cold start problem in a recommendation system. When the system lacks sufficient user behavior data or item data, it is difficult to provide accurate and personalized recommendations. This situation usually occurs in the following three cases:

[0063] (1) Cold start for new users: When a new user first uses the recommendation system, the system lacks the historical behavior data of this user, making it difficult to judge their preferences.

[0064] (2) Cold start for new items: When a new item or content is just added to the system and no user has interacted with it, it is difficult for the system to evaluate its popularity.

[0065] (3) System cold start: When the entire recommendation system is just launched, lacking the interaction data of all users and items, it is unable to provide effective recommendations.

[0066] The multi-objective hybrid expert network generally refers to the MMOE (Multi-gate Mixture-of-Experts) model. By introducing multiple gating models to dynamically adjust the fusion weights of multiple expert networks, it further improves the overall performance and flexibility of the network. By introducing multiple gate controllers, the MMOE model can more finely adjust the weights of each expert model, thereby improving the adaptability and generalization ability of the model.

[0067] For the gating model, the gating model structure is used to determine the fusion weights of multiple expert networks in different situations. It is usually a learner that dynamically adjusts the fusion weights of each expert network according to input features or context information, adaptively selects the appropriate expert network, and improves the performance and generalization ability of the overall model.

[0068] In the scenario of customer incoming calls, the customer service needs to handle user requests and judge whether to conduct business marketing during the precious call time, which requires high experience from the customer service. How to use big data and deep learning algorithms to help the customer service system allocate the most suitable customer service agent for connection and provide rich user business response rate labels for the customer service agent after connection to build an intelligent "matching + marketing" system and improve the customer incoming call experience and business marketing accuracy are two important issues urgently needed to be solved in the scenario of telephone customer service incoming calls.

[0069] Taking the telephone customer service of credit cards as an example, against the background that the growth rate of domestic credit cards has slowed down and entered the stock competition mode, the necessity for banks to conduct accurate demand insight and intelligent marketing for users has become prominent. As an important link for banks to contact users, the customer service call not only undertakes the task of solving user problems but also marketing credit card installment products, and has become one of the important channels for credit card marketing revenue contribution.

[0070] Business marketing in the scenario of telephone incoming calls is very different from traditional user response rate modeling, mainly reflected in three aspects:

[0071] (1) The incoming call scenario is a passive marketing scenario, with the characteristics of diverse customer intentions and stronger real-time nature, requiring higher complexity in modeling.

[0072] (2) In the inbound call scenario, the behavior types of customers / customer service representatives vary greatly, and refined modeling of features is required.

[0073] (3) Traditional customer response rate modeling only focuses on the unilateral characteristics of customers and does not consider the matching problem between customers and customer service representatives.

[0074] Therefore, taking the inbound call to marketing scenario of credit card customer service as an example, the following are the two main differences between the customer response rate prediction in this scenario and traditional response rate modeling:

[0075] (1) The introduction of the inbound call scenario automatically groups users according to whether they have an inbound call record recently (users with recent inbound calls, users without recent inbound calls). Different inbound call purposes and frequencies in the historical records also provide important information about the user's current bill pressure and repayment ability, which is directly related to the probability of installment application. Traditional response rate prediction models mostly use statistical machine learning methods and cannot efficiently utilize the historical inbound call behavior characteristics through model structure optimization, and do not have the conditions for refined modeling of the user's inbound call scenario.

[0076] (2) Marketing in the phone inbound call scenario is a passive process that needs to be considered from both the customer and customer service representative aspects. Its essence is more like a matching process of a recommendation system, that is, "recommending customer service representatives who are good at marketing to customers with a high willingness to apply for installments".

[0077] When the inventor was researching, it was found that for the installment response rate modeling on the marketing side, traditional machine learning methods are generally used as the modeling solution. By using various regression prediction algorithms, decision tree regression, xgboost models, neural networks and other algorithms, the response rate model is constructed. Then, for users with a relatively high response rate, the marketing list method is adopted and delivered to the business department for active marketing of installments through manual or system marketing channels.

[0078] With the increase in the complexity of the marketing scenario and the improvement of the marketing accuracy requirement, the installment response rate model has gradually evolved from the structure of a "single response rate model" into a framework of multi-model combination. To better utilize the historical behavior data of customers and improve the marketing accuracy of the business, the "clustering model + recommendation algorithm" method is adopted for intelligent marketing of the business, and the response probability of customers for different marketing services is predicted. Its essence is to group customers through the clustering model, and for customers in different groups, use the clustering features and combine with the marketing service features to adopt the recommendation algorithm for response rate prediction.

[0079] Based on the framework of traditional single response rate models and multi-model combinations, "end-to-end" deep models have many advantages in one-stop modeling. For example, automatic feature representation through deep networks reduces feature engineering processing, network design schemes with stronger scalability, and the combination of multiple models into one model can reduce model maintenance costs and system coupling while achieving one-stop services from feature extraction to model prediction. Therefore, in the recommendation matching scenarios of Internet companies, deep model frameworks are often used to design unique network structures according to business scenarios to complete the modeling and prediction processes of recommendation matching.

[0080] Traditional installment response rate models conduct active marketing through customer lists. The real-time requirements for the entire marketing activity are not high, and only the installment willingness of target customers is modeled. Usually, traditional statistical models are used. For example, in some solutions, feature engineering and LightGBM models are used to estimate the installment application response rate for each bill, and then marketing recommendations are made to users with high response rates. In the scenario of incoming call transfer to marketing, this solution only considers customer-side information and does not consider the matching relationship between customers and customer service representatives.

[0081] In other installment marketing solutions, long short-term neural networks are used to finely model customer historical behaviors, eliminating complex and difficult manual feature engineering. For example, in some solutions, long short-term neural networks are used to process various customer behavior sequences to predict the current installment response rate of customers. This solution is still an active marketing solution for modeling customer-side information and does not consider the matching problem between customers and incoming call customer service representatives in the passive marketing scenario of incoming call transfer to marketing.

[0082] Compared with traditional business marketing, there are two participants, customers and incoming call customer service representatives, in the scenario of incoming call transfer to marketing. At the same time, incoming calls belong to the scenario of passive call answering, and incoming call customers have more demands and more complex historical behaviors. Precise matching and response rate modeling need to be carried out from both the customer and customer service sides, which is more similar to the scenario of a recommendation system, that is, selecting customers with installment willingness and matching them with customer service representatives who are good at marketing.

[0083] In some solutions, multiple recommendation algorithm models are used to process various features of users separately, and finally multiple models are integrated for comprehensive scoring and recommendation. This type of method still uses traditional machine learning algorithms for modeling, cannot customize the modeling of important features according to the characteristics of the scenario, lacks flexibility, and increases the system maintenance cost and potential risks by using the multi-model combination method.

[0084] In some other solutions, a gating structure and a multi-expert network are used to model different behaviors of users, aiming to capture the unique interests of users and make the recommendation results more personalized. However, in the inbound-to-marketing scenario, in addition to paying attention to the customer behavior characteristics, it is also necessary to pay attention to the historical connection situation of the customer service, such as the favorable rate of the customer service's historical processed work orders, whether the customer service is good at handling marketing needs, etc. Therefore, it is necessary to model the historical connection behavior and marketing performance of the customer service side to match the suitable customer service with the customer when the call comes in.

[0085] The prediction of the installment response rate in the inbound-to-marketing scenario is very different from the traditional active business marketing due to the passivity, uncertainty of the inbound scenario and the addition of customer service personnel, and is more similar to the product recommendation in the Internet. When a customer makes a call to initiate an inbound call, the customer service system needs to select a customer service to answer the call, and judge whether to conduct installment marketing for the customer according to factors such as the current inbound call request and the bill situation of the customer. This is very similar to the situation in the recommendation scenario where when a user clicks a certain button on the APP, the favorite products are recommended to the user.

[0086] Aiming at the deficiencies of the single response rate model or the combined model and the end-to-end model, in the embodiments of the present application, based on the multi-domain expert network and the mixture of experts network, the historical inbound call behaviors of customers are fully utilized, combined with the historical connection characteristics of the customer service, and a set of inbound-to-marketing business engines based on the "end-to-end" deep matching recommendation network and the cold start mechanism are customized for the inbound-to-marketing scenario. Specifically, a multi-part network structure is designed in combination with the characteristics of the inbound-to-marketing scenario:

[0087] In the embodiments of the present application, customers are grouped according to the frequency and purpose of historical inbound call behaviors, and different groups of customers are respectively modeled by using a multi-domain expert network to obtain more accurate customer vector representations. Similarly, on the customer service side, based on the mixture of experts network, the network structure is innovated to fit the scenario, and expert networks with different input vectors are designed for different types of customers / customer services, so as to achieve the two goals of simultaneously outputting the customer-customer service matching rate and the customer installment response rate by one model. For the two different prediction goals, exclusive network structures and customer / customer service input features are designed for the gating structure to achieve the "accurate two-way matching" of the customer-customer service and the accurate prediction of the customer installment response rate.

[0088] In the embodiments of the present application, in the inbound call transfer to marketing scenario, a model is used to achieve accurate matching between customers and customer service representatives during inbound calls and the accuracy and independence of estimating the customer installment response rate, helping customer service representatives better conduct inbound marketing activities, ultimately increasing the success rate of installment processing and reducing the customer complaint rate. At the same time, in order to ensure that new customer service representatives have opportunities during inbound call connections, a cold start mechanism is designed to guarantee the positive cycle of the marketing system. For the specific credit card installment marketing framework in the inbound call transfer to marketing scenario, using deep learning algorithms, starting from the historical behaviors of both the customer side and the customer service representative side, a model structure suitable for this scenario is customized, and a mechanism that can perform closed-loop self-iteration is formulated to achieve the three main goals in the inbound call transfer to marketing scenario:

[0089] (1) Use a multi-domain topology network to model customers and customer service representatives with significantly different historical behaviors, obtaining better vector representations of customers and customer service representatives.

[0090] (2) Utilize the characteristics of a multi-objective mixture-of-experts network to simultaneously output the matching degree between the customer and the customer service representative and the installment response probability of the customer, achieving the purpose of having two different results output by one model.

[0091] (3) Use a closed-loop self-iteration mechanism to adopt a cold start strategy for new customer service representatives, ensuring the entire positive cycle iteration of the model and preventing it from falling into the "Matthew effect".

[0092] The following will, with reference to the accompanying drawings, elaborate in detail on the technical solutions provided by the embodiments of the present application.

[0093] The embodiments of the present application provide a data processing method for the inbound call transfer to marketing scenario, as Figure 1 shown, providing a schematic flowchart of the data processing method for the inbound call transfer to marketing scenario in the embodiments of the present application. The method at least includes the following steps S110 to S130:

[0094] Step S110: Cluster according to historical behaviors in the inbound call transfer to marketing scenario, where the historical behaviors include customers and customer service representatives.

[0095] In the inbound call transfer to marketing scenario, cluster the historical behaviors of customers and customer service representatives. After the clustering process, customers and customer service representatives with significantly different historical behaviors are respectively assigned to different cluster categories.

[0096] It should be noted that before the inbound call behavior of the customer occurs, it is impossible to know who the inbound customer is, nor can it be determined which customer service representative will answer the call. Therefore, only the historical behaviors of customers and customer service representatives can be analyzed to extract valuable information from them. As the participating entities, customers and customer service representatives respectively have different historical behavior distributions.

[0097] Step S120: Based on the clustering results, extract feature vectors based on the multi-domain topological network to obtain the feature vectors of customers and customer service representatives.

[0098] To refine the modeling of the historical behaviors of different types of customers and customer service representatives, a multi-domain topological network structure is adopted. By using the differences in the historical behavior distributions of customers and customer service representatives, they are divided into different domains. Through sharing the underlying domain, domain-specific networks, and adaptive mechanism methods in the multi-domain topological network, the vector representations generated in different domains are dynamically parameter-fused with the shared domain vectors, thereby obtaining more accurate vector representations of customers and customer service representatives. In this way, while extracting the generalities in the extraction scenario, the personalized information of customers and customer service representatives can be retained. Based on the clustering results, extract feature vectors based on the multi-domain topological network, and use the feature vectors of the customers and customer service representatives as the subsequent network input.

[0099] Step S130: Input the feature vectors of the customers and customer service representatives into a multi-objective hybrid expert network for training, and output the matching rate of the customers and customer service representatives and the installment response rate of the customers.

[0100] The multi-objective hybrid expert network MMOE (Multi-gate Mixture-of-Experts) model is a deep learning model architecture specifically designed for multi-task learning (MTL). Input the feature vectors of the customers and customer service representatives into a multi-objective hybrid expert network (optimized) for training, so as to output the estimated results of the customer response rate and the estimated results of the matching rate of the customers and customer service representatives.

[0101] Through the above method, a new installment marketing process is constructed based on the inbound-to-marketing scenario. That is, a matching mechanism for customer service representatives and customers is proposed, and the vectors of customers and customer service representatives are refined through a multi-domain topological network. And using the feature vectors as input, taking advantage of the characteristics of the multi-objective hybrid expert network, simultaneously outputting the customer-customer service representative matching rate and the customer installment response rate for modeling, and selecting the appropriate customer service representative to answer the call for the customer.

[0102] Through the above method, to cope with the large historical behavior differences between customers and customer service representatives in the inbound scenario, refine the modeling of vector representations. First, perform domain separation on the historical behaviors of customers and customer service representatives, then apply the multi-domain topological network to separately model the customers and customer service representatives in different domains, and finally fuse them through a parameter matrix. While ensuring the generalization of vector representations, the unique characteristics of each domain are retained. The purpose of refining the modeling of customer service representatives and customers with large differences is achieved.

[0103] Through the above method, to simultaneously output the customer-customer service matching probability and the marketing response rate in an "end-to-end" manner for a model, a multi-objective mixture-of-experts network is adopted. By using a mixture-of-experts network and a gating structure, the input vectors of the two output networks are made different, and the two tasks learn their respective knowledge from each other, while improving the matching accuracy between customers and customer service representatives and the marketing response rate of customers.

[0104] Different from the common combined model modeling methods in the related art, through the above method, a "one-stop" modeling solution is adopted, that is, a multi-objective model is used to estimate two results at one time, reducing the problems of mutual dependence, complex processes, and difficult maintenance caused by the combination of multiple models.

[0105] Different from the installment response rate modeling method in the related art, through the above method, combined with the characteristics of the incoming call scenario, a customer-customer service matching mechanism is introduced, and the customer and customer service vectors are refined through a multi-domain topology network; the customer-customer service matching rate and the customer installment response rate are modeled at one time through a multi-objective mixture model, so as to reduce the complaint rate and improve the marketing success rate.

[0106] In the embodiment of the present application, closely following the incoming call to marketing scenario, a set of matching and marketing processes in the incoming call to marketing scenario is innovatively constructed: combined with the characteristics of the incoming call scenario, a customer-customer service matching mechanism is introduced, and the customer and customer service vectors are refined through a multi-domain topology network; the customer-customer service matching rate and the customer installment response rate are modeled at one time through a multi-objective mixture model, so as to reduce the complaint rate and improve the marketing success rate.

[0107] In an embodiment of the present application, the clustering according to the historical behavior in the incoming call to marketing scenario, where the historical behavior includes customers and customer service representatives, includes: dividing the customers into never-called, complaint-type, non-complaint-type, and balanced types according to the historical behavior in the incoming call to marketing scenario, and dividing the customer service representatives into new employees, complaint-handling types, marketing types, and all-round types, to obtain the clustering result.

[0108] As Figure 2 shown, the incoming call to marketing scenario belongs to a passive marketing scenario. Before the incoming call behavior of the customer occurs, it is impossible to know who the incoming customer is, nor can it be judged which customer service representative will answer the call. Therefore, only the historical behavior of customers and customer service representatives can be analyzed to extract valuable information from it.

[0109] Customers and customer service representatives, as the participating subjects, have different historical behavior distributions respectively. For example, the historical incoming call frequency of customers, the main purpose of customers' incoming calls (complaint or others). The historical behavior of customer service representatives is also relatively diverse, such as pure new customer service representatives (new employees without any experience), customer service representatives handling complaints, marketing recommendation customer service representatives, etc.

[0110] As Figure 2As shown, there are significant distribution differences in the historical behaviors of different types of customers and customer service representatives. Before modeling, analysis is required followed by clustering. By analyzing their historical behaviors, in the embodiments of this application, customers are divided into four categories: never called in, complaint type, non-complaint type, and balanced type; customer service representatives are divided into four categories: completely new customer service representatives (new employees), complaint type, marketing type, and all-round type.

[0111] After the clustering process as Figure 2 shown, customers and customer service representatives with significant differences in historical behaviors are respectively assigned to different clusters. Among them, completely new customer service representatives and completely new customers are relatively unique types of entities. They have no task historical data in the phone call-in system, and a special "cold start" mechanism will be adopted for them later to ensure that such entities can also obtain marketing and connection opportunities, so as to ensure that the entire model can perform closed-loop self-update and iteration.

[0112] In an embodiment of this application, the multi-domain topology network includes a cluster-exclusive network and a central shared network. Feature vectors of customers and customer service representatives are extracted based on the multi-domain topology network according to the clustering results, including: according to the clustering results, inputting data of different clusters into the corresponding exclusive networks, and passing all data including the customers and the customer service representatives through the central shared FCN network based on the multi-domain topology network to divide the historical behaviors of the customers and customer service representatives into different domains; during the process of extracting feature vectors based on the multi-domain topology network, customers and customer service representatives with the same behavior will be divided into the FCN network in the same domain for training according to their historical behaviors; and data including the customers and the customer service representatives will all pass through the central shared FCN network based on the multi-domain topology network. Finally, the vectors output by each domain network will perform vector dot product inner product calculation with the shared FCN network through adaptive dynamic parameters to obtain the customer feature vectors and the customer service representative feature vectors.

[0113] As Figure 4 shown, during the model training process, customers and customer service representatives will be divided into the FCN network in the same domain for training according to their historical behaviors. At the same time, all data will pass through the central shared FCN network. Finally, the vectors output by each domain network will perform vector dot product inner product calculation with the shared FCN network through adaptive dynamic parameters, and finally obtain a vector representation that takes into account the commonalities in the call-in scenario and is personalized within each domain.

[0114] For data modeling with large differences, if traditional statistical learning methods are used for modeling, it is impossible to flexibly design the model structure to capture the differences between different user behaviors. If traditional neural networks are used for modeling, it will cause the network parameters to tend to customers or customer service representatives with a larger sample size, resulting in the common "Matthew effect" in recommendation systems, where the behaviors of niche customers or customer service representatives are submerged, and the model becomes too generalized and loses its personalized capabilities. The basic principle of the multi-domain topology network adopted in the embodiments of this application is as Figure 3 shown. To refine the modeling of the historical behaviors of different types of customers and customer service representatives, in the embodiments of this application, the multi-domain star topology network structure proposed in the paper "One Model to Serve All: Star Topology Adaptive Recommender for Multi-Domain CTR Prediction" published by Alibaba in 2021 is used. By using the differences in the historical behavior distributions of customers and customer service representatives, they are divided into different domains. Through sharing the underlying domain, domain-specific networks, and adaptive mechanism methods, the vector representations generated in different domains are dynamically parameter-fused with the shared domain vector to obtain more accurate customer and customer service representative vector representations, aiming to extract the generalities in the scenario while retaining the personalized information of customers and customer service representatives.

[0115] Based on the multi-domain topology network, the feature vectors of customers / customer service representatives are finally obtained as the input for the subsequent network to predict the customer-customer service representative matching rate and the customer response rate.

[0116] Refined modeling of customer service representatives and customers with large differences. To address the large historical behavior differences between customers and customer service representatives in the inbound scenario, refined modeling of vector representations. First, the historical behaviors of customers and customer service representatives are operated in different domains, and the multi-domain topology network is applied to separately model customers and customer service representatives in different domains, and finally fused through a parameter matrix. While ensuring the generalization of vector representations, the unique characteristics of each domain are retained.

[0117] In the embodiments of this application, based on the multi-domain topology network, separate modeling is performed on customers and customer service representatives with large historical behavior differences in the inbound scenario, and fusion is carried out through a parameter matrix. While learning the commonalities of the samples, the unique information of customers and customer service representatives within each domain is retained, making the vector representations of the two more accurate.

[0118] In one embodiment of the present application, the multi-domain topology network specifically includes: at the bottom layer, the embedding sharing with the same ID between different domains is adopted. The feature data of the customer and the customer service are spliced after being embedded into vectors, and the feature normalization results of each domain are obtained by using the PN local normalization method; according to the vector representation after PN normalization, it is input into a fully connected neural network, and the fully connected neural network includes a central FCN network shared by all clustering data, as well as a customer and customer service exclusive clustering neighborhood FCN network; after the output of the FCN network of each domain is added to the output of the auxiliary network, it is dynamically fused with the shared central FCN network through the topology network structure, the customer feature vector and the feature vector of the customer service.

[0119] As Figure 4 shown, specifically, the structural design of the multi-domain topology network consists of the following parts:

[0120] First, at the bottom layer, the embedding sharing with the same ID between different domains (cross-domain sharing) is adopted, which can save memory and improve efficiency.

[0121] Secondly, the features of the customer and the customer service are spliced after being embedded into vectors, and the feature normalization results of each domain are obtained by using the local normalization method (partitioned normalization, PN algorithm).

[0122] Then, the vector representation after PN normalization is input into a fully connected neural network (FCN network), including a shared FCN network and an FCN network corresponding to the customer / customer service domain, namely shared FCN and domain-specific FCN.

[0123] It should be noted that for the output of the FCN network of each domain added to the output of the auxiliary network, it is dynamically fused with the shared FCN network through the topology network structure to generate customer and customer service vector representations that reflect both domain characteristics and generalization characteristics.

[0124] In one embodiment of the present application, the method further includes: setting adaptive fusion parameters in the multi-domain topology network: the shared central network and the FCN networks of each clustering domain are fused through an adaptive parameter matrix, and the fusion parameters are obtained by continuous training and iterative update. The shared FCN network updates the parameters using all samples, and the FCN networks corresponding to the customer and customer service domains update the parameters using samples in the corresponding scenarios.

[0125] The shared FCN network and the FCN networks of each domain are fused through an adaptive parameter matrix:

[0126] Among them, the fusion parameters are obtained through continuous training and iterative updates. The multi-network fusion formula is as follows:

[0127]

[0128] Among them, the fully convolutional network shared across all scenarios (shared FCN) is denoted as W, B

[0129] The parameters of the fully convolutional network for each scenario (domain-specific FCN) are denoted as W p , B p (p represents the domain)

[0130] The parameters of each domain are element-wise added according to the W and b of the shared FCN and the domain-specific FCN to obtain the vector representations of customers or customer service representatives in different domains.

[0131] In addition, the shared FCN updates its parameters using all samples, and the domain-specific FCN only updates its parameters using the samples corresponding to the scenario.

[0132] In an embodiment of the present application, the auxiliary network feature vectors in the multi-domain topology network include: designing marker features for each domain, vectorizing them through the auxiliary network, and then performing cumulative calculation with the output vector of the star-shaped topology network to obtain the final customer feature vector and the feature vector of the customer service representative.

[0133] As Figure 5 shown, in order to enhance the distinguishability of vectors in each domain, in the embodiment of the present application, marker features are specifically designed for each domain, vectorized through the auxiliary network, and then cumulatively calculated with the output vector of the star-shaped topology network to obtain the final vector representation of the customer / customer service representative.

[0134] In an embodiment of the present application, the feature vectors of the customer and the customer service representative are input into a multi-objective hybrid expert network for training, and the matching rate of the customer and the customer service representative and the installment response rate of the customer are output, including: modeling the matching rate of the customer and the customer service representative and the installment response rate of the customer using a multi-objective hybrid expert network structure; dividing the multi-expert network into a matching expert network and a response rate expert network, where the matching expert network is used to receive the vector features of the customer and the customer service representative, and the response rate expert network is used to receive the vector features of the customer.

[0135] To accurately capture the matching relationship between customers and customer service representatives, and at the same time be able to model the installment response rate of customers unilaterally, so as to achieve the goal of completing multiple predicted values with one model, in the embodiments of this application, a multi-objective hybrid expert network is used as the multi-objective output modeling structure for the customer-customer service matching probability and the customer installment response rate.

[0136] The multi-objective hybrid expert network MMOE (Multi-gate Mixture-of-Experts) model is a deep learning model architecture designed specifically for multi-task learning (MTL). It realizes knowledge sharing between tasks through multiple expert networks (Experts) and gating networks (Gates), and at the same time allows each task to independently select the expert network suitable for its needs.

[0137] The MMOE model was proposed by Google in the paper "Modeling Task Relationships in Multi-task Learning with Multi-gate Mixture-of-Experts", aiming to effectively handle the relationships between tasks in multi-task learning. Its core idea is to combine the outputs of multiple expert networks through multiple gating networks, and each task selects the output of the expert network most suitable for its needs through its unique gating network. This design can not only share knowledge between tasks but also maintain the independence of tasks.

[0138] It can be understood that as the recommendation task scenario becomes more complex, by customizing the design of network structure modules that fit the business scenario, key features within the scenario can be effectively utilized to strengthen the vector representation of users and products by the deep network, and improve the accuracy of recommendation matching. For example, in the recommendation scenario to solve the multi-objective problem, that is, one model simultaneously estimates two related objectives, the multi-task multi-expert hybrid network MMOE (Multi-gate Mixture-of-Experts) is used for modeling. MMOE is a deep learning model for multi-task learning. It effectively shares and allocates underlying features by introducing a multi-gating mechanism to improve the target estimation effect of each task, and it has extensive applications in fields such as recommendation systems and prediction of advertising click-through rates.

[0139] In the recommended scenario, to solve the multi-objective problem, that is, a model simultaneously estimates two correlated objectives, a multi-task and multi-expert hybrid network MMOE (Multi-gate Mixture-of-Experts) is used for modeling. MMOE is a deep learning model for multi-task learning. It effectively shares and allocates underlying features by introducing a multi-gate mechanism to improve the objective estimation effect of each task. It has extensive applications in fields such as recommendation systems and advertisement click-through rate prediction. As the recommendation task scenario becomes more complex, by customizing the design of network structure modules that fit the business scenario, key features within the scenario can be effectively utilized, strengthening the vector representation of users and commodities by the deep network, and improving the accuracy of recommendation matching.

[0140] Multi-objective hybrid expert network design: Simultaneously output the customer-customer service matching probability and the customer installment response rate.

[0141] To achieve the "end-to-end" simultaneous output of the customer-customer service matching probability and the marketing response rate by a single model, in the embodiments of this application, a multi-objective hybrid expert network is adopted. By using a multi-expert network and a gating structure, the input vectors of the two output networks are made different from each other, and the two tasks learn their respective knowledge from each other, simultaneously improving the matching accuracy of customers and customer service and the marketing response rate of customers. That is, based on the multi-objective hybrid expert network, a unique network structure and gating structure are custom-designed, achieving the independence of the customer response rate model estimation and the accuracy of the customer-customer service matching target. The solution in the embodiments of this application adopts a "one-stop" modeling solution, that is, a multi-objective model estimates two results at one time, reducing the problems of mutual dependence, complex processes, and difficult maintenance brought by the combination of multiple models.

[0142] In an embodiment of this application, the method further includes: The gating model corresponding to the response rate FCN network only accepts the output vector of the response rate expert network; the output vector of the gating model corresponding to the response rate FCN network is disconnected in the network by setting the weight to 0 in the fusion formula, so that the input value of the response rate FCN network only contains the bill information features of the customer; in addition to using the customer vector and the customer service vector as input features, the original feature vector of the customer's historical incoming call behavior data is introduced as supplementary input features.

[0143] As Figure 6 shown, the MMOE network structure adopted in the embodiments of this application models two objectives: (1) the customer-customer service matching rate; (2) the installment response rate of the customer's bill.

[0144] For these two objectives and the incoming call to marketing scenario, in the embodiments of this application, an innovative transformation of the traditional MMOE network structure suitable for the marketing scenario is carried out, mainly including the following two points:

[0145] To ensure the independent integrity of the customer response rate prediction, that is, the customer installment response rate prediction is not affected by the customer service characteristics, this solution divides the multi-expert network into two types: matching experts and response rate experts. Among them, the matching experts receive the vector features of customers and customer service, and the response rate experts only receive the vector features of customers, avoiding the influence of the customer service vector on the installment response rate prediction from the input features of the expert network. The specific formula is as follows:

[0146] y k =h k (f k (x))

[0147] Where y k represents k prediction tasks, corresponding to y1 (customer-customer service matching rate prediction) and y2 (customer installment response rate prediction) in the embodiments of this application. h k is the upper-layer FCN network corresponding to k prediction tasks, and h1 and h2 are the matching FCN network and the response rate FCN network respectively.

[0148] As Figure 6 shown, to achieve that the response rate prediction is not affected by the matching rate prediction and customer service characteristics, the expert network in the embodiments of this application is divided into two categories, and the response rate experts among them only use the customer bill characteristics as the input. The following formula:

[0149] y 匹配率 =h 匹配率 (f 匹配率 (X 客户向量&客服向量 ))

[0150] y 响应率 =h 响应率 (f 响应率 (X 客户向量 ))

[0151] Where, X 客户向量&客服向量 =Concat(X 客户向量 ,X 客服向量 ), which is obtained by splicing the customer-customer service bilateral vector features.

[0152] In the design of the gating structure, the gating structure corresponding to the response rate FCN network only accepts the output vector of the response rate expert. In the fusion formula of the gating structure, the output vector of the matching expert is forcibly disconnected in the network by setting the weight to 0 (that is, Figure 6 the a connection part in

[0153]

[0154] Through these two customized network structure designs, it is ensured that the prediction of the customer response rate will not be interfered by the customer service vector; the customer-customer service vector matching can take into account bilateral features, and the potential differences in the matching process of different customers and customer services are learned through multiple expert networks, so as to achieve the purpose of accurately predicting the matching rate.

[0155] Please continue to refer to Figure 6 , in the multi-objective hybrid network MMOE, the gating structure is the core structure that controls how multiple expert networks are aggregated and fused into a vector and output to the upper-layer FCN network. To strengthen the matching modeling of customer-customer service, in the embodiments of the present application, the input features of the gating network for the customer-customer service accurate matching task are carefully designed. To closely adhere to the goal of the "matching" task, the gating structure of the accurate matching FCN network, in addition to using the customer vector and the customer service vector as input features, also specifically introduces the original feature vector of the customer's historical incoming call behavior as a supplementary input feature (i.e., Figure 6 part a in

[0156] G k1 = g 匹配 (X input ) = g 匹配 Concat(X 客户向量 , X 客服向量 , X 客户历史呼入行为原始特征 )

[0157] Among them, G represents the gating structure network, and k1 represents the gating network corresponding to the FCN network for matching the customer-customer service matching rate.

[0158] Through the design of the input features, the accurate matching of customers and customer services is carried out. Through the analysis of the feature importance, the optimal features for the matching relationship in customers and customer services are selected as the input vectors of the gating structure, and the user features irrelevant to the bilateral matching are discarded, such as: most of the features related to bill information. While reducing the model complexity, the matching efficiency of customer-customer service is improved. That is, from the overall consideration of the incoming call scenario, a cold start mechanism is introduced for the customer service side, which ensures the iterative update during the turnover of customer service personnel, creates a closed-loop link for the marketing of the entire scenario and the continuous iteration of the model, enables the model to continuously obtain new sample data using the cold start mechanism, and ensures the long-term effectiveness of the model in this scenario.

[0159] In an embodiment of the present application, the method further includes: executing a cold start mechanism for new employees in the customer service classification; the new employees obtain an activation probability value through the cold start strategy; when the activation probability value is greater than the threshold, it is determined to select the current customer service for connection marketing, and the installment response rate of the customer is returned and displayed to the customer as a reference for whether to conduct installment marketing.

[0160] AsFigure 7 As shown, first, to ensure the customer service experience, the cold start mechanism only takes effect on the side of pure new customer service representatives. When a customer calls in, the new and old customer service representatives will adopt two different strategies. For regular customer service representatives, the personalized matching algorithm mentioned above is used for accurate customer-customer service representative matching and customer response rate prediction. For pure new customer service representatives, the cold start mechanism in Figure 7 is used for processing:

[0161] S1. The new customer service representative will obtain an activation probability value through the cold start strategy. When it is greater than the threshold, it is determined to select this customer service representative for connection and marketing.

[0162] S2. At the same time, the cold start module will call the personalized matching algorithm, but only return the installment response rate of the customer and display it to the customer as a reference for whether to conduct installment marketing.

[0163] S3. After the incoming call connection is completed, the customer service representative will score the provided installment response rate label, and the scoring result will be used as the weight of the sample and participate in the continuous iterative upgrade of the model.

[0164] It can be understood that the Matthew Effect refers to a situation of "the strong getting stronger and the weak getting weaker" in a certain scenario, which is widely applicable in various fields, and the same is true for the incoming call transfer to marketing scenario. Specifically, the model prediction results tend to match customer service representatives who are good at handling marketing to customers who are enthusiastic about installment payments, so as to maximize the success rate of handling this installment. However, such an approach will cause the entire system to fall into the dilemma of local optimum, that is, the model frequently matches experienced customer service representatives to customers, and the range of customer service representatives participating in marketing is getting smaller and smaller. Pure new customer service representatives cannot get opportunities in the entire matching process, resulting in a decline in the overall efficiency of the customer service system. Due to the relatively large mobility of customer service staff, the "Matthew Effect" will cause the entire system to rely only on a few "hot" customer service representatives. When these employees leave or other situations occur, the entire incoming call transfer to marketing scenario will fall into a "paralyzed" state. At the same time, the gradually increasing "Matthew Effect" makes the samples used for continuous iteration of the model more and more limited (a small number of "hot" customer service representatives), ultimately affecting the overall revenue and sustainable development in the entire scenario.

[0165] Adopt the cold start mechanism design to reduce the Matthew Effect and build a sustainable iterative model and system. By introducing the cold start mechanism in the recommendation system, it is ensured that pure new customer service representatives can obtain connection opportunities, enabling the entire system to operate in a virtuous cycle. The model can continuously obtain new sample data and is not affected by the "Matthew Effect".

[0166] In an embodiment of the present application, the cold start strategy includes adopting a Bandit-like algorithm, and Thompson sampling is used as the algorithm for generating the probability value of new customer service representatives.

[0167] For the cold start strategy, in the embodiments of the present application, a Bandit-like algorithm is adopted to explore and exploit new customer service representatives (exploration-exploitation).

[0168] It should be noted that the Bandit algorithm, also known as the Multi-Armed Bandit Algorithm, originally stems from the slot machine problem, that is, under the limitation of a finite number of coin tosses, how to select among multiple slot machines (each with a different reward distribution) to maximize the overall reward of this game. Corresponding to the incoming call scenario, it is about how to match suitable candidates for new customer service representatives to minimize the complaint rate and maximize the marketing success rate.

[0169] In addition, in the embodiments of the present application, Thompson Sampling is selected as the algorithm for generating the probability value of new customer service representatives.

[0170] It should be noted that the Thompson Sampling algorithm is a Bayesian method for solving the multi-armed bandit problem. It uses Bayesian inference to determine the current best choice by sampling the reward distribution of each choice. Assuming there are k choices, and the reward distribution of each choice i is an unknown probability distribution, in the embodiments of the present application, it is hoped to maximize the cumulative reward value through this method.

[0171] In the embodiments of the present application, a method for determining the installment marketing response rate in the incoming call to marketing scenario is also provided. Among them, the data processing method of the incoming call to marketing scenario is adopted to determine the installment marketing response rate. The data processing method of the incoming call to marketing scenario includes

[0172] Cluster according to the historical behaviors in the incoming call to marketing scenario, where the historical behaviors include customers and customer service representatives;

[0173] Extract feature vectors based on the multi-domain topology network according to the clustering results to obtain the feature vectors of customers and customer service representatives; and

[0174] Input the feature vectors of the customers and customer service representatives into a multi-objective hybrid expert network for training, and output the matching rate of customers and customer service representatives and the installment response rate of customers.

[0175] In the embodiments of the present application, a method for determining the matching rate in the incoming call to marketing scenario is also provided. Among them, the data processing method of the incoming call to marketing scenario is adopted to determine the matching rate.

[0176] The data processing method of the incoming call to marketing scenario includes

[0177] Cluster according to the historical behaviors in the incoming call to marketing scenario, where the historical behaviors include customers and customer service representatives;

[0178] Based on the results of clustering, feature vectors of customers and customer service representatives are extracted based on a multi-domain topological network; and

[0179] The feature vectors of the customers and customer service representatives are input into a multi-objective hybrid expert network for training, and the matching rate of the customers and customer service representatives and the installment response rate of the customers are output.

[0180] The embodiment of the present application also provides a data processing device 800 for the inbound call transfer to marketing scenario, as Figure 8 shown, which provides a schematic structural diagram of the data processing device for the inbound call transfer to marketing scenario in the embodiment of the present application. The data processing device 800 for the inbound call transfer to marketing scenario at least includes: a clustering module 810, an extraction module 820, and an output module 830, where:[[]]

[0181] In an embodiment of the present application, the clustering module 810 is specifically configured to: perform clustering according to historical behaviors in the inbound call transfer to marketing scenario, and the historical behaviors include customers and customer service representatives.

[0182] In the inbound call transfer to marketing scenario, the historical behaviors of customers and customer service representatives are clustered. After the clustering process, customers and customer service representatives with significantly different historical behaviors are respectively assigned to different cluster categories.

[0183] It should be noted that before the inbound call behavior of a customer occurs, it is impossible to know who the inbound customer is, nor can it be determined which customer service representative will answer the call. Therefore, only the historical behaviors of customers and customer service representatives can be analyzed to extract valuable information from them. As the participating parties, customers and customer service representatives have different historical behavior distributions respectively.

[0184] In an embodiment of the present application, the extraction module 820 is specifically configured to: extract feature vectors of customers and customer service representatives based on a multi-domain topological network according to the clustering results.

[0185] To refine the modeling of the historical behaviors of different types of customers and customer service representatives, a multi-domain topological network structure is adopted. By using the differences in the historical behavior distributions of customers and customer service representatives, they are divided into different domains. Through sharing the underlying domain, domain-specific network, and adaptive mechanism method in the multi-domain topological network, the vector representations generated in different domains are dynamically parameter-fused with the shared domain vectors, so as to obtain more accurate vector representations of customers and customer service representatives. In this way, while extracting the generalities in the scenario, the personalized information of customers and customer service representatives can be retained. Feature vectors are extracted based on the multi-domain topological network according to the clustering results, and the feature vectors of the customers and customer service representatives are used as the subsequent network input.

[0186] In one embodiment of the present application, the output module 830 is specifically configured to: input the feature vectors of the customer and the customer service into a multi-objective hybrid expert network for training, and output the matching rate between the customer and the customer service and the installment response rate of the customer.

[0187] The multi-objective hybrid expert network MMOE (Multi-gate Mixture-of-Experts) model is a deep learning model architecture designed specifically for multi-task learning (MTL). Input the feature vectors of the customer and the customer service into a multi-objective hybrid expert network (optimized) for training, so as to output the estimated result of the customer response rate and the estimated result of the matching rate between the customer and the customer service.

[0188] In one embodiment of the present application, the clustering module 810 is further configured to:

[0189] According to the historical behaviors in the inbound call transfer to marketing scenario, divide the customers into never-called, complaint type, non-complaint type, and balanced type, and divide the customer services into new employees, complaint type, marketing type, and all-round type, to obtain the clustering result.

[0190] In one embodiment of the present application, the multi-domain topology network includes a clustering exclusive network and a central shared network. The extraction module 820 is further configured to:

[0191] According to the clustering result, input the data of different clusters into the corresponding exclusive networks, and input all the data including the customer and the customer service through the central shared FCN network based on the multi-domain topology network, and divide the historical behaviors of the customer and the customer service into different domains;

[0192] During the process of feature vector extraction based on the multi-domain topology network, customers and customer services with the same behavior will be divided into the FCN networks in the same domain according to their historical behaviors for training; and

[0193] The data including the customer and the customer service are all input through the central shared FCN network based on the multi-domain topology network. Finally, the vectors output by each domain network will perform a dot product inner product calculation with the shared FCN network through adaptive dynamic parameters to obtain the customer feature vector and the customer service feature vector.

[0194] In one embodiment of the present application, the multi-domain topology network specifically includes:

[0195] At the bottom layer, the embedding sharing of the same ID between different domains is adopted.

[0196] The characteristic data of the customer and the customer service are concatenated after being embedded into vectors, and the feature normalization results of each domain are obtained by using the PN local normalization method;

[0197] According to the vector representation after PN normalization, it is input into a fully connected neural network, and the fully connected neural network includes a central FCN network shared by all clustering data, as well as a customer and customer service exclusive clustering neighborhood FCN network;

[0198] After the output of the FCN network of each domain is added with the output of the auxiliary network, it is dynamically fused with the shared central FCN network through a topological network structure, the customer feature vector and the feature vector of the customer service.

[0199] In an embodiment of the present application, the extraction module 820 is further configured to:

[0200] Set the adaptive fusion parameters in the multi-domain topological network:

[0201] The shared central network and each clustering domain FCN network are fused through an adaptive parameter matrix with each domain FCN network. The fusion parameters are continuously updated and iterated during training. The shared FCN network updates the parameters using all samples, and the FCN networks corresponding to the customer and the customer service in the corresponding domain update the parameters using the samples in the corresponding scenario.

[0202] In an embodiment of the present application, the extraction module 820 is further configured to:

[0203] Based on the auxiliary network feature vectors in the multi-domain topological network, including:

[0204] Design a marked feature for each domain, vectorize it through the auxiliary network, and perform cumulative calculation with the output vector of the star-shaped topological network to obtain the final customer feature vector and the customer service feature vector.

[0205] In an embodiment of the present application, the output module 830 is further configured to:

[0206] Model the matching rate of the customer and the customer service and the installment response rate of the customer by using a multi-objective hybrid expert network structure;

[0207] The multi-expert network is divided into a matching expert network and a response rate expert network. The matching expert network is used to receive the vector features of the customer and the customer service, and the response rate expert network is used to receive the vector features of the customer.

[0208] In an embodiment of the present application, the output module 830 is further configured to:

[0209] The gating model corresponding to the response rate FCN network only accepts the output vector of the response rate expert network;

[0210] The output vector of the gating model corresponding to the response rate FCN network is disconnected in the network by setting the weight to 0 in the fusion formula, so that the input value of the response rate FCN network only contains the billing information features of the customer;

[0211] In addition to using the customer vector and the customer service vector as input features, the original feature vector of the customer's historical incoming call behavior data is also introduced as supplementary input features.

[0212] In an embodiment of the present application, it further includes: a cold start module, which is used for:

[0213] Execute the cold start mechanism for new employees in the customer service classification;

[0214] New employees obtain the activation probability value through the cold start strategy;

[0215] When the activation probability value is greater than the threshold, it is determined to select the current customer service for connection marketing, and the installment response rate of the customer is returned and displayed to the customer as a reference for whether to conduct installment marketing.

[0216] In an embodiment of the present application, the cold start strategy includes using a Bandit algorithm, and Thompson sampling is used as the algorithm for generating the probability value of a new customer service.

[0217] It can be understood that the above data processing device for the incoming call to marketing scenario can implement each step of the data processing method for the incoming call to marketing scenario provided in the foregoing embodiments. The relevant explanations about the data processing method for the incoming call to marketing scenario are applicable to the data processing device for the incoming call to marketing scenario, and will not be elaborated here.

[0218] Figure 9 It is a schematic structural diagram of an electronic device according to an embodiment of the present application. Please refer to Figure 9 , at the hardware level, the electronic device includes a processor, and optionally also includes an internal bus, a network interface, and a memory. Among them, the memory may include a memory, such as a high-speed random access memory (Random-Access Memory, RAM), and may also include a non-volatile memory, such as at least one disk memory, etc. Of course, the electronic device may also include other hardware required for other services.

[0219] The processor, network interface, and memory can be interconnected via an internal bus, which can be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus, an EISA (Extended Industry Standard Architecture) bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. For the sake of representation, Figure 9 only a bidirectional arrow is used in

[0220] Memory, which is used to store programs. Specifically, the program can include program code, and the program code includes computer operation instructions. The memory can include a memory and a non-volatile memory, and provides instructions and data to the processor.

[0221] The processor reads the corresponding computer program from the non-volatile memory into the memory and then runs it, forming a data processing device for the incoming call transfer to marketing scenario at the logical level. The processor executes the program stored in the memory and is specifically used to perform the following operations:

[0222] Cluster according to the historical behaviors in the incoming call transfer to marketing scenario, where the historical behaviors include customers and customer service representatives;

[0223] Extract feature vectors based on a multi-domain topology network according to the clustering results to obtain the feature vectors of customers and customer service representatives; and

[0224] Input the feature vectors of the customers and customer service representatives into a multi-objective hybrid expert network for training, and output the matching rate of customers and customer service representatives and the installment response rate of customers.

[0225] The above as in this application Figure 1The method executed by the data processing device for the incoming call transfer to marketing scenario disclosed in the illustrated embodiment can be applied to or implemented by a processor. The processor may be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the above method can be completed by the integrated logic circuit in the hardware of the processor or the instructions in the form of software. The above processor may be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it may also be a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components. It can implement or execute the various methods, steps, and logic block diagrams disclosed in the embodiments of the present application. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc. The steps of the method disclosed in combination with the embodiments of the present application can be directly embodied as being executed and completed by the hardware decoding processor, or executed and completed by a combination of the hardware and software modules in the decoding processor. The software module may be located in a mature storage medium in the art such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory, or an electrically erasable programmable memory, a register, etc. This storage medium is located in the memory, and the processor reads the information in the memory and combines its hardware to complete the steps of the above method.

[0226] The electronic device can also execute Figure 1 the method executed by the data processing device for the incoming call transfer to marketing scenario in Figure 1 the illustrated embodiment, and implement the functions of the data processing device for the incoming call transfer to marketing scenario in

[0227] Embodiments of the present application also propose a computer-readable storage medium that stores one or more programs. The one or more programs include instructions that, when executed by an electronic device including multiple application programs, can enable the electronic device to execute Figure 1 the method executed by the data processing device for the incoming call transfer to marketing scenario in the illustrated embodiment, and specifically used to execute:

[0228] Cluster according to the historical behavior in the incoming call transfer to marketing scenario, where the historical behavior includes customers and customer service representatives;

[0229] Extract feature vectors based on the multi-domain topology network according to the clustering results to obtain the feature vectors of customers and customer service representatives; and

[0230] Input the feature vectors of the customer and the customer service into the training based on the multi-objective hybrid expert network, and output the matching rate of the customer and the customer service and the installment response rate of the customer.

[0231] Those skilled in the art should understand that the embodiments of the present invention can be provided as a method, a system, or a computer program product. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0232] The present invention is described with reference to the flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to the embodiments of the present invention. It should be understood that each flow and / or block in the flowchart and / or block diagram, as well as the combination of flows and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate for implementing in the process Figure 1 one process or multiple processes and / or blocks Figure 1 a device for the functions specified in one block or multiple blocks.

[0233] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory generate a manufactured article including an instruction device, and the instruction device implements in the process Figure 1 one process or multiple processes and / or blocks Figure 1 a device for the functions specified in one block or multiple blocks.

[0234] These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process, and thus the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in Figure 1 one process or multiple processes and / or blocks Figure 1 a device for the functions specified in one block or multiple blocks.

[0235] In a typical configuration, a computing device includes one or more processors (CPUs), an input / output interface, a network interface, and a memory.

[0236] The memory may include non-permanent storage in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM. The memory is an example of a computer-readable medium.

[0237] Computer readable media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. Information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer readable media does not include temporary computer readable media (transitory media), such as modulated data signals and carrier waves.

[0238] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.

[0239] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0240] The above is only an embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the scope of the claims of the present application.

Claims

1. A data processing method for an incoming call-to-marketing scenario, wherein: The method comprises: Grouping is performed based on historical behaviors in the call-to-marketing scenario, where the historical behaviors include customers and customer service staff; According to the clustering results, feature vectors are extracted based on the multi-domain topological network to obtain feature vectors of customers and customer service; and The feature vectors of the customer and the customer service are input into a multi-objective hybrid expert network for training, and the matching rate of the customer and the customer service and the installment response rate of the customer are output.

2. The method of claim 1, wherein: The grouping is performed based on historical behaviors in the call-to-marketing scenario, where the historical behaviors include customers and customer service, including: According to the historical behaviors in the inbound call to marketing scenario, the customers are divided into never calling, complaint type, non-complaint type and balance type, and the customer service staff are divided into new employees, complaint type, marketing type and all-round type, and the grouping results are obtained.

3. The method of claim 2, wherein: The multi-domain topology network includes a cluster-specific network and a central shared network. The feature vector extraction is performed based on the multi-domain topology network according to the clustering result to obtain the feature vectors of the customer and the customer service, including: According to the grouping results, the data of different groups are input into the corresponding exclusive network, and all the data including the customers and the customer service are passed through the central shared FCN network based on the multi-domain topology network, and the historical behaviors of the customers and customer service are divided into different domains; In the process of extracting feature vectors based on multi-domain topological networks, customers and customer service representatives with the same behavior will be divided into FCN networks in the same domain for training based on their historical behaviors; and The data of the customer and the customer service all pass through the central shared FCN network based on the multi-domain topology network. Finally, the vectors output by the networks in each domain will be calculated by vector dot multiplication and inner product with the shared FCN network through adaptive dynamic parameters to obtain the feature vector of the customer and the feature vector of the customer service.

4. The method of claim 3, wherein: The multi-domain topology network specifically includes: At the bottom layer, embeddings with the same ID are shared between different fields. The feature data of customers and customer service are concatenated after being embedded into vectors, and the PN local normalization method is used to obtain the feature normalization results of each domain; According to the vector representation after PN normalization, it is input into a fully connected neural network, wherein the fully connected neural network includes a central FCN network shared by all group data and a customer and customer service exclusive group neighborhood FCN network; After the FCN network of each field is added with the output of the auxiliary network, the customer feature vector and the customer service feature vector are dynamically merged with the shared central FCN network through the topological network structure.

5. The method according to claim 3 or 4, further comprising: The adaptive fusion parameter setting in the multi-domain topology network is as follows: The shared central network and the FCN networks in each clustering field are fused with the FCN networks in each field through an adaptive parameter matrix. The fusion parameters are obtained by continuous updating and iteration of training. The shared FCN network uses all samples to update parameters, and the FCN networks in the corresponding fields of customers and customer service use samples of corresponding scenarios to update parameters.

6. The method according to claim 3 or 4, wherein the auxiliary network feature vector based on the multi-domain topology network comprises: Marking features are designed for each field, vectorized through the auxiliary network, and accumulated with the output vector of the star topology network to obtain the final customer feature vector and customer service feature vector.

7. The method of claim 1, wherein: The feature vectors of the customer and the customer service are input into a multi-objective hybrid expert network for training, and the matching rate between the customer and the customer service and the installment response rate of the customer are output, including: The matching rate between customers and customer service and the installment response rate of customers are modeled based on a multi-objective hybrid expert network structure. The multi-expert network is divided into a matching expert network and a response rate expert network. The matching expert network is used to accept vector features of customers and customer service, and the response rate expert network is used to accept vector features of customers.

8. The method of claim 7, further comprising: The gating model corresponding to the response rate FCN network only accepts the output vector of the response rate expert network; The output vector of the gating model corresponding to the response rate FCN network is disconnected in the network by resetting the weight to 0 in the fusion formula, so that the input value of the response rate FCN network only contains the customer's bill information features; In addition to using customer vectors and customer service vectors as input features, the original feature vectors of customer historical inbound call behavior data are also introduced as supplementary input features.

9. The method of claim 2, further comprising: For new employees in the customer service category, a cold start mechanism is implemented; The new employee obtains an activation probability value through the cold start strategy; When the activation probability value is greater than a threshold, it is determined that the current customer service is selected for call marketing, and the customer's installment response rate is returned and displayed to the customer as a reference for whether to conduct installment marketing.

10. The method of claim 9, wherein: The cold start strategy includes adopting a Bandit-type algorithm and using Thompson sampling as an algorithm for generating a new customer service probability value.

11. A method for determining the response rate of phased marketing in a call-to-marketing scenario, wherein: The data processing method for the incoming call-to-marketing scenario as described in any one of claims 1 to 10 is used to determine the installment marketing response rate.

12. A method for determining the matching rate of an incoming call-to-marketing scenario, wherein: The matching rate is determined by using the data processing method for the incoming call-to-marketing scenario as described in any one of claims 1 to 10.

13. A data processing device for an incoming call-to-marketing scenario, wherein: The training device comprises: A grouping module, used to group calls according to historical behaviors in the call-to-marketing scenario, where the historical behaviors include customers and customer service staff; An extraction module is used to extract feature vectors based on the multi-domain topological network according to the clustering results to obtain feature vectors of customers and customer service; and The output module is used to input the feature vectors of the customer and the customer service into a multi-objective hybrid expert network for training, and output the matching rate of the customer and the customer service and the installment response rate of the customer.

14. An electronic device comprising: processor; as well as A memory arranged to store computer executable instructions, which when executed cause the processor to perform the method of any one of claims 1 to 10.

15. A computer-readable storage medium storing one or more programs, which, when executed by an electronic device including a plurality of application programs, causes the electronic device to execute any one of the methods of claims 1 to 10.

Images