Customer behavior prediction method and device in inbound call scenario
By using NAS technology to build customer characteristics and behavior prediction network models in call incoming scenarios, and automatically search and train models, the problem of low customer behavior prediction accuracy is solved, accurate prediction and appeal positioning of customer behavior is achieved, and the pressure of manual customer service is reduced , improves the efficiency of intelligent Q&A.
Patent Information
- Application Number
- CN202110404116.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-04-15
- Publication Date
- 2024-11-08
- Estimated Expiration
- 2041-04-15
AI Technical Summary
The accuracy of customer behavior prediction in the call-in scenario is low, and it is impossible to effectively screen customer characteristics, resulting in inaccurate positioning of customer demands and increasing the pressure on manual customer service.
Using the customer behavior prediction method based on NAS technology, by constructing customer feature models and behavior prediction network models, automating search and training models, extracting effective feature data, rapid feature screening and precise appeal positioning, and reducing the pressure of manual customer service.
It improves the efficiency of smart Q&A, reduces the number of incoming and transfers of customers and the pressure of manual customer service, and realizes accurate prediction and precise appeal positioning of customer behavior.
Smart Images

Figure CN113095879B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a method and device for predicting customer behavior in an incoming call scenario. Background Art
[0002] At present, the industry's intelligent technology is developing rapidly. The recognition capabilities of artificial intelligence and speech recognition are almost close to those of real-life scenes. Speech synthesis and video synthesis technologies are even more anthropomorphic. Major Internet companies around the world have entered the era of intelligence and have achieved intelligent transformation by building omni-channel, omni-media, and full-process intelligent service platforms. On the other hand, the development of new scenarios in the traditional machine learning field requires professional machine learning experts to spend a lot of energy on the construction and selection of algorithms and models, which sets a high threshold for other practitioners who are not professionals in the field.
[0003] As the field of machine learning develops to a new stage, NAS (Neural Architecture Search) technology emerges, and important steps such as model building become automated, so that machine learning models can be applied without too much intervention from professionals. NAS is an automated modeling method for deep neural network technology in the field of machine learning. AI learns to understand a large number of user questions while learning to record user behavior trajectories and portrait features, and then combines the two to establish a search and matching model. After the overall model architecture is confirmed, the overall model training is completed based on customer data.
[0004] The core function of the "predictive prediction" function in the inbound call scenario is to accurately predict the behavior information of the customer in the current call. However, the current inbound call scenario often uses the traditional rule matching method, has not yet effectively modeled customer characteristics, has a single user feature judgment, has low customer behavior prediction accuracy, cannot effectively screen customer behavior, and cannot accurately locate customer demands, resulting in a surge in the number of customer inbound calls transferred to manual services, and high pressure on inbound manual customer service. Summary of the invention
[0005] The main purpose of the embodiments of the present invention is to provide a customer behavior prediction method and device in an incoming call scenario, so as to effectively predict customer behavior, accurately locate customer demands, greatly improve the efficiency of intelligent question and answer, and reduce the pressure of manual customer service in customer incoming call scenarios.
[0006] In order to achieve the above object, an embodiment of the present invention provides a method for predicting customer behavior in an incoming call scenario, comprising:
[0007] Obtain a target customer call request, and extract the target customer incoming call information from the target customer call request;
[0008] Determine the target customer predicted behavior corresponding to the target customer incoming call information according to the target customer incoming call prediction rule, and push the corresponding voice question for the customer to choose according to the target customer predicted behavior; wherein, the step of creating the target customer incoming call prediction rule includes: constructing each target customer feature data according to each target customer incoming call information, inputting each target customer feature data into the customer behavior prediction network model created based on the customer feature model data and the customer behavior prediction network model architecture, obtaining each target customer predicted behavior, and creating the target customer incoming call prediction rule according to the corresponding relationship between each target customer incoming call information and each target customer predicted behavior; the customer behavior prediction network model architecture is constructed by customer feature training search data and customer feature verification search data.
[0009] The embodiment of the present invention further provides a customer behavior prediction device in an incoming call scenario, comprising:
[0010] The target customer incoming call information extraction module is used to obtain the target customer call request and extract the target customer incoming call information from the target customer call request;
[0011] A prediction behavior determination module is used to determine the target customer prediction behavior corresponding to the target customer incoming call information according to the target customer incoming call prediction rule, and push the corresponding voice question for the customer to choose according to the target customer predicted behavior; wherein the step of creating the target customer incoming call prediction rule includes: constructing each target customer feature data according to each target customer incoming call information, inputting each target customer feature data into a customer behavior prediction network model created based on customer feature model data and a customer behavior prediction network model architecture to obtain each target customer predicted behavior, and creating the target customer incoming call prediction rule according to the corresponding relationship between each target customer incoming call information and each target customer predicted behavior; the customer behavior prediction network model architecture is constructed through customer feature training search data and customer feature verification search data.
[0012] An embodiment of the present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, the steps of the customer behavior prediction method in the incoming call scenario are implemented.
[0013] An embodiment of the present invention further provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of the customer behavior prediction method in the incoming call scenario are implemented.
[0014] The customer behavior prediction method and device in the incoming call scenario of the embodiment of the present invention extracts the target customer's incoming call information from the target customer's call request, determines the target customer's predicted behavior corresponding to the target customer's incoming call information according to the target customer's incoming call prediction rule, and pushes the corresponding voice question for the customer to choose. It can effectively predict customer behavior, accurately locate customer demands, greatly improve the efficiency of intelligent question and answer, and reduce the pressure of manual customer service in the customer's incoming call scenario. BRIEF DESCRIPTION OF THE DRAWINGS
[0015] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0016] Figure 1 is a flow chart of a method for predicting customer behavior in an incoming call scenario in an embodiment of the present invention;
[0017] Figure 2 is a flow chart of constructing a customer behavior prediction network model architecture in an embodiment of the present invention;
[0018] Figure 3 is a flowchart of constructing a customer behavior prediction network model architecture in another embodiment of the present invention;
[0019] Figure 4 is a schematic structural diagram of an initial unit in an embodiment of the present invention;
[0020] Figure 5 is a schematic structural diagram of an initial unit under various structural parameters in an embodiment of the present invention;
[0021] Figure 6 is a schematic diagram of the structure of the unit under various structural parameters after multiple iterations in an embodiment of the present invention;
[0022] Figure 7 is a schematic diagram of the structure of a unit under the target structural parameters in an embodiment of the present invention;
[0023] Figure 8 is a flow chart of creating a customer behavior prediction network model in an embodiment of the present invention;
[0024] Fig. 9 is a structural block diagram of a customer behavior prediction device in an incoming call scenario in an embodiment of the present invention;
[0025] Fig.10 It is a structural block diagram of a computer device in an embodiment of the present invention. DETAILED DESCRIPTION
[0026] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0027] Those skilled in the art will appreciate that the embodiments of the present invention may be implemented as a system, device, apparatus, method or computer program product. Therefore, the present disclosure may be specifically implemented in the following forms, namely: complete hardware, complete software (including firmware, resident software, microcode, etc.), or a combination of hardware and software.
[0028] In view of the low accuracy of customer behavior prediction in the prior art, the inability to effectively screen customer behavior and accurately locate customer demands, which leads to a surge in the number of customers transferred to manual services in the incoming call scenario and great pressure on the manual customer service. The embodiment of the present invention provides a customer behavior prediction method in the incoming call scenario, which performs feature selection based on the two dimensions of basic customer information and incoming call information in the incoming call scenario, and uses NAS technology to select an automated model, and trains the automatically searched model. Finally, the trained model is used to perform rapid feature screening in the customer group that is about to enter the incoming call scenario, thereby effectively predicting customer behavior and accurately locating customer demands. The efficiency of intelligent question and answer is greatly improved by actively asking questions, and the number of customers transferred to manual services and the pressure on the manual customer service in the incoming call scenario are reduced. The present invention is described in detail below in conjunction with the accompanying drawings.
[0029] Figure 1 is a flow chart of a method for predicting customer behavior in an incoming call scenario according to an embodiment of the present invention. Figure 1 As shown in FIG. 1 , the customer behavior prediction method in the inbound call scenario includes:
[0030] S101: Obtain a target customer call request, and extract target customer incoming call information from the target customer call request.
[0031] S102: Determine the target customer predicted behavior corresponding to the target customer incoming call information according to the target customer incoming call prediction rule, and push the corresponding voice question for the customer to select according to the target customer predicted behavior.
[0032] The target customer inbound call prediction rules in the embodiment of the present invention have been pre-created before receiving the inbound call request, including the inbound call information of each target customer and the corresponding target customer predicted behavior. For example, when the customer's predicted behavior is credit card business, multiple voice questions related to credit card business are automatically pushed to the customer for selection when the customer calls in. Later, the customer can automatically reply according to the voice questions selected by the customer, which can reduce the number of manual transfers in the customer's inbound call scene and the pressure on the inbound manual customer service.
[0033] Among them, the step of creating target customer incoming call prediction rules includes: constructing each target customer's characteristic data based on each target customer's incoming call information, inputting each target customer's characteristic data into a customer behavior prediction network model created based on customer characteristic model data and a customer behavior prediction network model architecture to obtain each target customer's predicted behavior, and creating target customer incoming call prediction rules based on the correspondence between each target customer's incoming call information and each target customer's predicted behavior; the customer behavior prediction network model architecture is constructed through customer characteristic training search data and customer characteristic verification search data.
[0034] Figure 1 The execution subject of the customer behavior prediction method in the inbound call scenario shown can be a computer. Figure 1 As can be seen from the process shown, the customer behavior prediction method in the incoming call scenario of the embodiment of the present invention extracts the target customer incoming call information from the target customer call request, determines the target customer predicted behavior corresponding to the target customer incoming call information according to the target customer incoming call prediction rule, and pushes the corresponding voice question for the customer to choose. This can effectively predict customer behavior, accurately locate customer demands, greatly improve the efficiency of intelligent question and answer, and reduce the pressure on manual customer service in the customer incoming call scenario.
[0035] Figure 2 It is a flow chart of constructing a customer behavior prediction network model architecture in an embodiment of the present invention. Figure 3 FIG. 1 is a flowchart of constructing a customer behavior prediction network model architecture in another embodiment of the present invention. Figure 2-Figure 3 As shown in the figure, the customer behavior prediction network model architecture includes:
[0036] Perform the following iterative processing:
[0037] S201: Determine a training loss function according to customer feature training search data, actual customer behavior search data corresponding to the customer feature training search data, various structural parameters and weights corresponding to various structural parameters.
[0038] In one embodiment, S201 includes: determining customer behavior training search data based on customer feature training search data, each structural parameter and the weight corresponding to each structural parameter; determining a training loss function based on the customer behavior training search data and the corresponding customer behavior actual search data.
[0039] S202: Determine a verification loss function according to the customer feature verification search data, the customer behavior actual search data corresponding to the customer feature verification search data, each structural parameter and the weight corresponding to each structural parameter.
[0040] In one embodiment, S202 includes: determining customer behavior verification search data based on customer feature verification search data, each structural parameter and the weight corresponding to each structural parameter; determining a verification loss function based on the customer behavior verification search data and the corresponding customer behavior actual search data.
[0041] Before executing S201 and S202, the following steps are also included:
[0042] 1. Extract effective feature data based on various data such as customer information, call details and result feedback in the incoming call scenario of the intelligent service platform, perform feature selection from two dimensions: customer basic data and incoming call data. Extract customer gender, customer age, region, education level, nature of unit, marital status, occupation, administrative level, annual income, whether or not a wealth card user, whether or not a gold card user, etc. from the customer basic data as customer feature data; extract product type, consultation type, call duration, call time, interaction rounds, number of calls on the day, whether or not transferred to manual, etc. from the incoming call data as customer feature data.
[0043] 2. Use One-Hot Binary encoding to process customer feature data. For example, for the three-domain data X = [Date = Wednesday, Gender = Male, TaskType = Financial], Date has 7 values, which can be compiled into a 7-dimensional binary vector, where only Wednesday is 1 and the others are 0, because it has only one eigenvalue; Gender has two dimensions, one of which is 1; there are 10 TaskTypes in total, and the value of Financial products (Financial) is 1, and the others are 0. Then the converted X is:
[0044] X=[0,0,1,0,0,0,0|0,1|0,0,1,0,..0].
[0045] Finally, all customer feature data will be encoded by One-Hot Binary to obtain a high-dimensional sparse vector. The time from 00:00 to 24:00 is divided into one-hour dimensions, and all customer feature data are distinguished by the time dimension, so as to establish a two-dimensional feature matrix X as the input of the model. Among them, the customer feature data is used to build the customer behavior prediction network model architecture and the customer behavior prediction network model, and the target customer feature data is used to input the created customer behavior prediction network model to predict the target customer prediction behavior corresponding to all target customers. Therefore, the time period of the customer feature data is earlier than the target customer feature data, and the number of customer feature data is less than the number of target customer feature data.
[0046] 3. Divide the customer feature data into customer feature training search data and customer feature verification search data used to build the customer behavior prediction network model architecture, and customer feature training model data, customer feature verification model data, and customer feature test model data used to create the customer behavior prediction network model. Among them, 15% of the customer feature data is customer feature training search data, 5% of the customer feature data is customer feature verification search data, 60% of the customer feature data is customer feature training model data, 10% of the customer feature data is customer feature verification model data, and 10% of the customer feature data is customer feature test model data.
[0047] Taking S201 as an example, the customer behavior training search data can be determined through the following steps:
[0048] 1. Define a cell unit, which can be seen as a directed acyclic graph with N nodes. The edges between the nodes represent possible operations (such as 3×3 convolution). The operations between the nodes are unknown at initialization.
[0049] Figure 4 Schematic diagram of the structure of the initial unit in the embodiment of the present invention. Figure 5 It is a schematic diagram of the structure of the initial unit under various structural parameters in the embodiment of the present invention. Figure 6 It is a schematic diagram of the structure of the unit under various structural parameters after multiple iterations in an embodiment of the present invention. Figure 7 Schematic diagram of the structure of the unit under the target structure parameters in the embodiment of the present invention. Figure 4-Figure 7 As shown, a unit is a directed acyclic graph (ordered sequence) consisting of N nodes. Each (i-th) node x (i) is a hidden expression (i.e., feature graph), and the directed edge (i, j) represents the specific operation o (i,j) , used to convert the j-th node x (j) Transform to the i-th node x (i) . The unit is composed of input nodes, intermediate nodes, output nodes and edges, and each unit has two input nodes and one output node. For the convolutional network, the two input nodes are the outputs of the first two layers of units respectively; each intermediate node is converted by its predecessor node through edge operation and then summed; the output node is obtained by combining and summing all the intermediate nodes. The edge represents the calculation operation (such as 3×3 convolution, corresponding to the structural parameter). In the process of converging to the structure, all the edges between the two nodes will exist and participate in the training, and finally weighted average. The calculation operation (structural parameter) in the present invention can be the following eight types: 3×3 separable convolution, 5×5 separable convolution, 3×3 dilated convolution, 5×5 dilated convolution, 3×3 maximum pooling, 3×3 average pooling, identity and zero connection.
[0050] 2. If Figure 5As shown in the figure, the discrete search space is continuous, and each edge is the superposition of the softmax weights of all sub-operations. The output of the unit is obtained by merging all the parent nodes. Each parent node can be calculated as follows:
[0051]
[0052] Among them, x (i) is the i-th node, x (j) is the jth node, o (i,j) To change the jth node x (j) Transform to the i-th node x (i) When j = 0, the initial x (0) The search data is trained for customer features, and the output nodes in the unit are trained for customer behavior.
[0053] Assume O is a set of candidate operations, each of which represents a function o(*) acting on x. (i) In order to make the discrete search space continuous, the type of specific operation is selected to be continuous as the softmax function of all possible operations:
[0054]
[0055] in, represents the set of softmax functions for all possible operations on directed edges (i, j), represents the operation mixture weight (structural parameter) of the directed edge (i, j), and o(x) represents a single specific type of operation. After the discrete space is continuous, the structure search task becomes: learning a set of continuous variables After the search is completed, a discrete structure is obtained by replacing each mixed operator with the operation with the highest probability, namely:
[0056]
[0057] In specific implementation, L train represents the training loss function, L val It can be seen from S201 and S202 that the training loss function and the verification loss function are both related to the structural parameters The goal of structured search is to find an optimal structural parameter. So that it can minimize the loss function of the validation set And find the optimal weight w* to minimize the training set loss, that is: Finally, the bidirectional optimization problem can be expressed by the following formula:
[0058]
[0059]
[0060] in, express and The validation loss function under represents w and The training loss function under , w represents the weight, Represents the structural parameters, represents the optimal structural parameter, and w* represents the optimal weight.
[0061] S203: Determine whether the current number of iterations reaches a preset number of structure iterations.
[0062] S204: When the current number of iterations reaches the preset number of structural iterations, the maximum value among the structural parameters is determined as the target structural parameter, and the customer behavior prediction network model architecture is constructed according to the target structural parameter and the corresponding weights.
[0063] The preset number of structure iterations is 100. The present invention randomly performs 4 trainings based on random seeds, each training is iterated 100 times based on customer feature training search data and customer feature verification search data, and the module structure (unit) with the best performance is selected from the 4 search trainings as the final module structure.
[0064] In specific implementation, the target structure parameters can be determined by the following formula:
[0065]
[0066] like Figure 6 As shown, the multiple edges between the nodes correspond to multiple structural parameters, and the thickness of the lines represents the size of the structural parameters. Figure 7 Only the thickest line is retained, which is equivalent to determining the maximum value among the structural parameters as the target structural parameter.
[0067] Constructing a customer behavior prediction network model architecture according to target structure parameters and corresponding weights includes: constructing target units according to target structure parameters and corresponding weights, stacking multiple units according to a preset number of network layers to form a final customer behavior prediction network model architecture.
[0068] In specific implementation, models of different sizes can be quickly and automatically searched and customized for different scales of data and feature dimensions. The ultimate goal of automated network structure search is to first train a cell (small module, target unit), and then connect the cells to form a large network. The hyperparameter layers (number of network layers, number of network modules) can control how many cells are connected. For example, layers = 20 means that there are 20 cells connected front and back. The number of network channels in each cell (module) controls the size of the search model.
[0069] According to the basic data scale of customer feature training search data and customer feature verification search data, the scale of the customer behavior prediction network model architecture can be controlled by the number of network modules and the number of network channels of each network module. For scenarios with large data, the number of network modules and the number of network channels can be enlarged to automatically search for a relatively large-scale model to improve model accuracy. On the contrary, for scenarios with small data, the number of network modules and the number of network channels can be reduced to increase the speed of model automated search, and finally search for a customer behavior prediction network model architecture that matches the data scale.
[0070] Taking a network architecture with 20 layers and 36 network channels as an example, network training requires 600 iterations based on customer feature training search data and customer feature verification search data, and the initial input data batch size is 96.
[0071] S205: When the current number of iterations does not reach the preset number of structural iterations, the weights corresponding to the structural parameters are updated according to the training loss function, and the structural parameters are updated according to the verification loss function and the updated weights, and the iterative process is continued.
[0072] In specific implementation, the structural parameters and weights corresponding to the edges on the sub-operation mixing probability can be updated by joint bidirectional optimization. and weights w are mutually optimized using a gradient descent step.
[0073] For example, in the kth iteration, given the current structure (structure parameters in the kth iteration) Move the weight w in the k-1th iteration in the direction that reduces the training loss k-1 To get the weight w in the kth iteration k . Then keep the weight w k Keep the structure parameters unchanged and update them so that they can minimize the validation loss function (after performing a gradient descent):
[0074]
[0075] Among them, ξ is the learning rate of the gradient descent, w k is the weight in the kth iteration, is the structural parameter in the k-1th iteration, w k and The validation loss function is given by .
[0076] pass The structural gradient can be obtained as follows:
[0077]
[0078] in, Represents the weight of the model after a single-step iteration, for w' and The validation loss function under . is computationally expensive and can be approximated by the following formula, where ε is a tiny scalar parameter:
[0079]
[0080]
[0081]
[0082] The process of constructing the customer behavior prediction network model architecture of the present invention is not to search in a discrete set of candidate structures, but to approximately change the search space to a continuous space, so that iterative training optimization of the continuous space can be performed by gradient descent, which can effectively shorten the time of automated network search and save a lot of hardware resources.
[0083] Figure 8 : is a flow chart of creating a customer behavior prediction network model in an embodiment of the present invention. Figure 8 As shown, creating a customer behavior prediction network model includes:
[0084] Perform the following iterative processing:
[0085] S301: Determine customer behavior model data according to customer feature model data, target structure parameters and corresponding weights.
[0086] Among them, the customer feature model data in the step of creating a customer behavior prediction network model is customer feature training model data.
[0087] S302: Determine a model loss function according to the customer behavior model data and the corresponding customer behavior actual data.
[0088] S303: Determine whether the current number of iterations reaches the preset number of model iterations.
[0089] S304: When the current number of iterations reaches the preset number of model iterations, a customer behavior prediction network model is constructed according to the target structure parameters and corresponding weights.
[0090] S305: When the current number of iterations does not reach the preset number of model iterations, the weights corresponding to the target structure parameters are updated according to the model loss function, and the iterative process continues.
[0091] After creating a customer behavior prediction network model, you can also verify the customer behavior prediction network model through customer feature verification model data, and test the customer behavior prediction network model through customer feature test model data. Customer feature verification model data can avoid overfitting during network training. If the accuracy of the customer feature training model data improves, but the accuracy of the customer feature verification model data remains the same or decreases, it means that the neural network is overfitting and training should be stopped. Customer feature test model data is used to test the final user behavior prediction network model to confirm the actual prediction ability of the network, and can predict the hit success rate of customer behavior in the inbound call scenario.
[0092] The specific process of the embodiment of the present invention is as follows:
[0093] 1. Determine the customer behavior training search data based on the customer feature training search data, each structural parameter and the weight corresponding to each structural parameter; determine the training loss function based on the customer behavior training search data and the corresponding customer behavior actual search data.
[0094] 2. Determine the customer behavior verification search data based on the customer feature verification search data, each structural parameter and the weight corresponding to each structural parameter; determine the verification loss function based on the customer behavior verification search data and the corresponding customer behavior actual search data.
[0095] 3. When the current number of iterations reaches the preset number of structural iterations, the maximum value of each structural parameter is determined as the target structural parameter, and the customer behavior prediction network model architecture is constructed according to the target structural parameter and the corresponding weight. Otherwise, the weight corresponding to each structural parameter is updated according to the training loss function, and each structural parameter is updated according to the verification loss function and the updated weight, and then return to step 1.
[0096] 4. Determine the customer behavior model data based on the customer feature model data, target structure parameters and corresponding weights.
[0097] 5. Determine the model loss function based on the customer behavior model data and the corresponding actual customer behavior data.
[0098] 6. When the current number of iterations reaches the preset number of model iterations, a customer behavior prediction network model is constructed based on the target structure parameters and the corresponding weights. Otherwise, the weights corresponding to the target structure parameters are updated according to the model loss function, and the process returns to step 4.
[0099] 7. Construct characteristic data of each target customer based on the incoming call information of each target customer, input the characteristic data of each target customer into the customer behavior prediction network model, and obtain the predicted behavior of each target customer.
[0100] 8. Create target customer incoming call prediction rules based on the correspondence between each target customer's incoming call information and each target customer's predicted behavior.
[0101] 9. Obtain the target customer's call request, and extract the target customer's incoming call information from the target customer's call request.
[0102] 10. Determine the target customer predicted behavior corresponding to the target customer's incoming call information based on the target customer's incoming call prediction rules, and push the corresponding voice questions for the customer to choose based on the target customer's predicted behavior.
[0103] In summary, in order to solve the problems of low customer behavior prediction accuracy in the incoming call scenario, inability to effectively screen customer features, inability to accurately locate customer demands, resulting in a surge in the number of customer incoming calls transferred to manual services, and high pressure on manual inbound customer service, the present invention provides a customer behavior prediction method based on the incoming call scenario, which extracts valid feature data based on multiple data such as customer information, call details, and result feedback in the incoming call scenario of the intelligent service platform, extracts valid feature information and establishes relevant training data sets, verification data sets, and test data sets for automated model selection, training, and prediction. In summary, the present invention has the following three advantages:
[0104] 1. It can automatically search and build models based on training data and verification data, saving time and energy in manually building models.
[0105] 2. For data of different scales and feature dimensions, it can adaptively and quickly automatically search and customize models of different scales.
[0106] 3. Through model training and optimization, rapid behavior prediction can be performed in the customer base in the inbound call scenario, so as to effectively screen customer characteristics and accurately locate customer demands. Then, by actively asking questions, the efficiency of intelligent question and answer can be greatly improved, reducing the number of manual transfers in the customer inbound call scenario and the pressure on inbound manual customer service.
[0107] Based on the same inventive concept, an embodiment of the present invention also provides a customer behavior prediction device in an incoming call scenario. Since the principle of solving the problem by the device is similar to the customer behavior prediction method in the incoming call scenario, the implementation of the device can refer to the implementation of the method, and the repeated parts will not be repeated.
[0108] Fig. 9 is a structural block diagram of a customer behavior prediction device in an incoming call scenario according to an embodiment of the present invention. Fig. 9 As shown, the customer behavior prediction device in the incoming call scenario includes:
[0109] The target customer incoming call information extraction module is used to obtain the target customer call request and extract the target customer incoming call information from the target customer call request;
[0110] A prediction behavior determination module is used to determine the target customer prediction behavior corresponding to the target customer incoming call information according to the target customer incoming call prediction rule, and push the corresponding voice question for the customer to choose according to the target customer predicted behavior; wherein the step of creating the target customer incoming call prediction rule includes: constructing each target customer feature data according to each target customer incoming call information, inputting each target customer feature data into a customer behavior prediction network model created based on customer feature model data and a customer behavior prediction network model architecture to obtain each target customer predicted behavior, and creating the target customer incoming call prediction rule according to the corresponding relationship between each target customer incoming call information and each target customer predicted behavior; the customer behavior prediction network model architecture is constructed through customer feature training search data and customer feature verification search data.
[0111] In one embodiment, a network model architecture building module is further included, which is used to perform the following iterative processing:
[0112] Determine a training loss function based on customer feature training search data, actual customer behavior search data corresponding to the customer feature training search data, each structural parameter and a weight corresponding to each structural parameter;
[0113] Determine a verification loss function based on customer feature verification search data, actual customer behavior search data corresponding to the customer feature verification search data, each structural parameter and a weight corresponding to each structural parameter;
[0114] When the current number of iterations reaches the preset number of structural iterations, the maximum value of each structural parameter is determined as the target structural parameter, and the customer behavior prediction network model architecture is constructed according to the target structural parameters and the corresponding weights. Otherwise, the weights corresponding to each structural parameter are updated according to the training loss function, and each structural parameter is updated according to the verification loss function and the updated weights, and the iterative processing continues.
[0115] In one embodiment, the network model architecture building module is specifically used to:
[0116] Determine the customer behavior training search data according to the customer feature training search data, each structural parameter and the weight corresponding to each structural parameter;
[0117] Determine a training loss function based on customer behavior training search data and corresponding customer behavior actual search data;
[0118] Determine the customer behavior verification search data based on the customer feature verification search data, each structural parameter and the weight corresponding to each structural parameter;
[0119] The verification loss function is determined based on the customer behavior verification search data and the corresponding customer behavior actual search data.
[0120] In one embodiment, a network model creation module is further included, which is used to perform the following iterative processing:
[0121] Determine customer behavior model data based on customer feature model data, target structure parameters and corresponding weights;
[0122] Determine the model loss function based on the customer behavior model data and the corresponding actual customer behavior data;
[0123] When the current number of iterations reaches the preset number of model iterations, a customer behavior prediction network model is constructed according to the target structure parameters and the corresponding weights. Otherwise, the weights corresponding to the target structure parameters are updated according to the model loss function, and the iterative processing continues.
[0124] In summary, the customer behavior prediction device in the incoming call scenario of the embodiment of the present invention extracts the target customer incoming call information from the target customer call request, determines the target customer predicted behavior corresponding to the target customer incoming call information according to the target customer incoming call prediction rule, and pushes the corresponding voice question for the customer to choose, which can effectively predict customer behavior, accurately locate customer demands, greatly improve the efficiency of intelligent question and answer, and reduce the pressure of manual customer service in the customer incoming call scenario.
[0125] The embodiment of the present invention also provides a specific implementation of a computer device that can implement all the steps in the customer behavior prediction method in the incoming call scenario in the above embodiment. Fig.10 is a block diagram of a computer device in an embodiment of the present invention, see Fig.10 , the computer equipment specifically includes the following contents:
[0126] Processor (processor) 1001 and memory (memory) 1002.
[0127] The processor 1001 is used to call the computer program in the memory 1002. When the processor executes the computer program, all steps in the customer behavior prediction method in the incoming call scenario in the above embodiment are implemented. For example, when the processor executes the computer program, the following steps are implemented:
[0128] Obtain a target customer call request, and extract the target customer incoming call information from the target customer call request;
[0129] Determine the target customer predicted behavior corresponding to the target customer incoming call information according to the target customer incoming call prediction rule, and push the corresponding voice question for the customer to choose according to the target customer predicted behavior; wherein, the step of creating the target customer incoming call prediction rule includes: constructing each target customer feature data according to each target customer incoming call information, inputting each target customer feature data into the customer behavior prediction network model created based on the customer feature model data and the customer behavior prediction network model architecture, obtaining each target customer predicted behavior, and creating the target customer incoming call prediction rule according to the corresponding relationship between each target customer incoming call information and each target customer predicted behavior; the customer behavior prediction network model architecture is constructed by customer feature training search data and customer feature verification search data.
[0130] In summary, the computer device of the embodiment of the present invention extracts the target customer's incoming call information from the target customer's call request, determines the target customer's predicted behavior corresponding to the target customer's incoming call information according to the target customer's incoming call prediction rule, and pushes the corresponding voice question for the customer to choose. This can effectively predict customer behavior, accurately locate customer demands, greatly improve the efficiency of intelligent question and answer, and reduce the pressure on manual customer service in customer incoming call scenarios.
[0131] The embodiment of the present invention also provides a computer-readable storage medium capable of implementing all the steps of the customer behavior prediction method in the incoming call scenario in the above embodiment. The computer-readable storage medium stores a computer program. When the computer program is executed by a processor, all the steps of the customer behavior prediction method in the incoming call scenario in the above embodiment are implemented. For example, when the processor executes the computer program, the following steps are implemented:
[0132] Obtain a target customer call request, and extract the target customer incoming call information from the target customer call request;
[0133] Determine the target customer predicted behavior corresponding to the target customer incoming call information according to the target customer incoming call prediction rule, and push the corresponding voice question for the customer to choose according to the target customer predicted behavior; wherein, the step of creating the target customer incoming call prediction rule includes: constructing each target customer feature data according to each target customer incoming call information, inputting each target customer feature data into the customer behavior prediction network model created based on the customer feature model data and the customer behavior prediction network model architecture, obtaining each target customer predicted behavior, and creating the target customer incoming call prediction rule according to the corresponding relationship between each target customer incoming call information and each target customer predicted behavior; the customer behavior prediction network model architecture is constructed by customer feature training search data and customer feature verification search data.
[0134] In summary, the computer-readable storage medium of the embodiment of the present invention extracts the target customer's incoming call information from the target customer's call request, determines the target customer's predicted behavior corresponding to the target customer's incoming call information according to the target customer's incoming call prediction rule, and pushes the corresponding voice question for the customer to choose, which can effectively predict customer behavior, accurately locate customer demands, greatly improve the efficiency of intelligent question and answer, and reduce the pressure of manual customer service in customer incoming call scenarios.
[0135] The specific embodiments described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.
[0136] Those skilled in the art may also understand that the various illustrative logical blocks, units, and steps listed in the embodiments of the present invention may be implemented by electronic hardware, computer software, or a combination of the two. In order to clearly demonstrate the interchangeability of hardware and software, the various illustrative components, units, and steps described above have generally described their functions. Whether such functions are implemented by hardware or software depends on the specific application and the design requirements of the entire system. Those skilled in the art may use various methods to implement the described functions for each specific application, but such implementation should not be understood as exceeding the scope of protection of the embodiments of the present invention.
[0137] The various illustrative logic blocks, or units, or devices described in the embodiments of the present invention can be implemented or operated by a general-purpose processor, a digital signal processor, an application-specific integrated circuit (ASIC), a field programmable gate array or other programmable logic device, a discrete gate or transistor logic, a discrete hardware component, or any combination of the above. The general-purpose processor can be a microprocessor, and optionally, the general-purpose processor can also be any conventional processor, controller, microcontroller or state machine. The processor can also be implemented by a combination of computing devices, such as a digital signal processor and a microprocessor, a plurality of microprocessors, one or more microprocessors combined with a digital signal processor core, or any other similar configuration.
[0138] The steps of the method or algorithm described in the embodiments of the present invention can be directly embedded in hardware, a software module executed by a processor, or a combination of the two. The software module can be stored in a RAM memory, a flash memory, a ROM memory, an EPROM memory, an EEPROM memory, a register, a hard disk, a removable disk, a CD-ROM, or other storage media of any form in the art. Exemplarily, the storage medium can be connected to the processor so that the processor can read information from the storage medium and can write information to the storage medium. Optionally, the storage medium can also be integrated into the processor. The processor and the storage medium can be arranged in an ASIC, and the ASIC can be arranged in a user terminal. Optionally, the processor and the storage medium can also be arranged in different components in the user terminal.
[0139] In one or more exemplary designs, the above functions described in the embodiments of the present invention can be implemented in hardware, software, firmware or any combination of the three. If implemented in software, these functions can be stored on a computer-readable medium, or transmitted in the form of one or more instructions or codes on a computer-readable medium. Computer-readable media include computer storage media and communication media that facilitate the transfer of computer programs from one place to another. The storage medium can be any available medium that can be accessed by any general or special computer. For example, such computer-readable media can include but are not limited to RAM, ROM, EEPROM, CD-ROM or other optical disk storage, disk storage or other magnetic storage devices, or any other medium that can be used to carry or store program codes in the form of instructions or data structures and other forms that can be read by general or special computers, or general or special processors. In addition, any connection can be appropriately defined as a computer-readable medium, for example, if the software is transmitted from a website site, server or other remote resource through a coaxial cable, fiber optic cable, twisted pair, digital subscriber line (DSL) or wirelessly, such as infrared, wireless and microwave, it is also included in the defined computer-readable medium. The disk and disc include compact disk, laser disk, optical disk, DVD, floppy disk and blue-ray disk. Disks usually copy data magnetically, while discs usually copy data optically with lasers. The above combination can also be included in computer readable media.
Claims
1. A customer behavior prediction method in an inbound call scenario, characterized in that: include: Obtaining a target customer call request, and extracting target customer incoming call information from the target customer call request; Determine the target customer predicted behavior corresponding to the target customer incoming call information according to the target customer incoming call prediction rule, and push the corresponding voice question according to the target customer predicted behavior for the customer to choose; wherein, the step of creating the target customer incoming call prediction rule includes: constructing each target customer feature data according to each target customer incoming call information, inputting each target customer feature data into a customer behavior prediction network model created based on customer feature model data and a customer behavior prediction network model architecture to obtain each target customer predicted behavior, and creating the target customer incoming call prediction rule according to the corresponding relationship between each target customer incoming call information and each target customer predicted behavior; the customer behavior prediction network model architecture is constructed by customer feature training search data and customer feature verification search data; Building a customer behavior prediction network model architecture includes: Perform the following iterative processing: Determine a training loss function according to the customer feature training search data, the customer behavior actual search data corresponding to the customer feature training search data, each structural parameter and the weight corresponding to each structural parameter; Determine a verification loss function according to the customer feature verification search data, the customer behavior actual search data corresponding to the customer feature verification search data, each structural parameter and the weight corresponding to each structural parameter; When the current number of iterations reaches the preset number of structural iterations, the maximum value among the structural parameters is determined as the target structural parameter, and the customer behavior prediction network model architecture is constructed according to the target structural parameter and the corresponding weight, otherwise the weight corresponding to each structural parameter is updated according to the training loss function, and each structural parameter is updated according to the verification loss function and the updated weight, and the iterative process is continued; Determining a training loss function according to the customer feature training search data, the customer behavior actual search data corresponding to the customer feature training search data, each structural parameter and the weight corresponding to each structural parameter includes: Determine customer behavior training search data according to the customer feature training search data, each structural parameter and the weight corresponding to each structural parameter; Determine a training loss function based on the customer behavior training search data and the corresponding customer behavior actual search data; Determining a verification loss function according to the customer feature verification search data, the customer behavior actual search data corresponding to the customer feature verification search data, each structural parameter and the weight corresponding to each structural parameter includes: Determine the customer behavior verification search data according to the customer feature verification search data, each structural parameter and the weight corresponding to each structural parameter; Determine a verification loss function based on the customer behavior verification search data and the corresponding customer behavior actual search data; A unit is defined as a directed acyclic graph consisting of N nodes, where the i-th node x (i) It is a feature graph, and the directed edge (i, j) represents the specific operation o (i,j) , used to convert the j-th node x (j) Transform to the i-th node x (i) ; The unit is composed of input nodes, intermediate nodes, output nodes and edges. Each unit has two input nodes and one output node. The two input nodes are the outputs of the first two layers of units respectively. Each intermediate node is obtained by converting its predecessor node through edge operations and then summing it up. The output node is obtained by combining and summing all the intermediate nodes. The edges represent calculation operations. In the process of converging to the structure, all the edges between the two nodes will exist and participate in the training, and finally the weighted average is calculated. The discrete search space is made continuous, and each edge is the superposition of the softmax weights of all sub-operations; the output of the unit is obtained by merging all the parent nodes, and each parent node can be calculated as follows: Among them, x (i) is the i-th node, x (j) is the jth node, o (i,j) To change the jth node x (j) Transform to the i-th node x (i) structural parameters; when j = 0, the initial x (0) The search data is trained for customer characteristics, and the output nodes in the unit are trained for customer behavior. O is a set of candidate operations, each of which represents a function o(*) acting on x. (i) ; In order to make the discrete search space continuous, the type of specific operation is selected to be continuous as the softmax function of all possible operations: in, represents the set of softmax functions for all possible operations on directed edges (i, j), represents the structural parameter of the directed edge (i, j), and o(x) represents a single specific type of operation; after the discrete space is continuous, the structured search task becomes: learning a set of continuous variables After the search is finished, each mixed operator is replaced by the operation with the highest probability to obtain a discrete structure, namely: The goal of structured search is to find an optimal structure parameter So that it can minimize the loss function of the validation set And find the optimal weight w* to minimize the training set loss, that is: The final bidirectional optimization problem can be expressed by the following formula: in, express and The validation loss function under represents w and The training loss function under , w represents the weight, Represents the structural parameters, represents the optimal structural parameter, and w* represents the optimal weight.
2. The customer behavior prediction method in the inbound call scenario according to claim 1, characterized in that: Creating a customer behavior prediction network model involves: Perform the following iterative processing: Determine customer behavior model data according to the customer feature model data, the target structure parameters and corresponding weights; Determine a model loss function according to the customer behavior model data and the corresponding customer behavior actual data; When the current number of iterations reaches the preset number of model iterations, a customer behavior prediction network model is constructed according to the target structure parameters and the corresponding weights, otherwise the weights corresponding to the target structure parameters are updated according to the model loss function, and the iterative processing continues.
3. A customer behavior prediction device in an incoming call scenario, characterized in that: include: A target customer incoming call information extraction module is used to obtain target customer call requests and extract target customer incoming call information from the target customer call requests; A prediction behavior determination module is used to determine the target customer prediction behavior corresponding to the target customer incoming call information according to the target customer incoming call prediction rule, and push the corresponding voice question according to the target customer predicted behavior for the customer to choose; wherein the step of creating the target customer incoming call prediction rule includes: constructing each target customer feature data according to each target customer incoming call information, inputting each target customer feature data into a customer behavior prediction network model created based on customer feature model data and a customer behavior prediction network model architecture to obtain each target customer predicted behavior, and creating the target customer incoming call prediction rule according to the corresponding relationship between each target customer incoming call information and each target customer predicted behavior; the customer behavior prediction network model architecture is constructed by customer feature training search data and customer feature verification search data; The customer behavior prediction device in the incoming call scenario also includes a network model architecture building module, which is used to perform the following iterative processing: Determine a training loss function according to the customer feature training search data, the customer behavior actual search data corresponding to the customer feature training search data, each structural parameter and the weight corresponding to each structural parameter; Determine a verification loss function according to the customer feature verification search data, the customer behavior actual search data corresponding to the customer feature verification search data, each structural parameter and the weight corresponding to each structural parameter; When the current number of iterations reaches the preset number of structural iterations, the maximum value among the structural parameters is determined as the target structural parameter, and the customer behavior prediction network model architecture is constructed according to the target structural parameter and the corresponding weight, otherwise the weight corresponding to each structural parameter is updated according to the training loss function, and each structural parameter is updated according to the verification loss function and the updated weight, and the iterative process is continued; The network model architecture building module is specifically used for: Determine customer behavior training search data according to the customer feature training search data, each structural parameter and the weight corresponding to each structural parameter; Determine a training loss function based on the customer behavior training search data and the corresponding customer behavior actual search data; Determine the customer behavior verification search data according to the customer feature verification search data, each structural parameter and the weight corresponding to each structural parameter; Determine a verification loss function based on the customer behavior verification search data and the corresponding customer behavior actual search data; A unit is defined as a directed acyclic graph consisting of N nodes, where the i-th node x (i) It is a feature graph, and the directed edge (i, j) represents the specific operation o (i,j) , used to convert the j-th node x (j) Transform to the i-th node x (i) ; The unit is composed of input nodes, intermediate nodes, output nodes and edges. Each unit has two input nodes and one output node. The two input nodes are the outputs of the first two layers of units respectively. Each intermediate node is obtained by converting its predecessor node through edge operations and then summing it up. The output node is obtained by combining and summing all the intermediate nodes. The edges represent calculation operations. In the process of converging to the structure, all the edges between the two nodes will exist and participate in the training, and finally the weighted average is calculated. The discrete search space is made continuous, and each edge is the superposition of the softmax weights of all sub-operations; the output of the unit is obtained by merging all the parent nodes, and each parent node can be calculated as follows: Among them, x (i) is the i-th node, x (j) is the jth node, o (i,j) To change the jth node x (j) Transform to the i-th node x (i) structural parameters; when j = 0, the initial x (0) The search data is trained for customer characteristics, and the output nodes in the unit are trained for customer behavior. O is a set of candidate operations, each of which represents a function o(*) acting on x. (i) ; In order to make the discrete search space continuous, the type of specific operation is selected to be continuous as the softmax function of all possible operations: in, represents the set of softmax functions for all possible operations on directed edges (i, j), represents the structural parameter of the directed edge (i, j), and o(x) represents a single specific type of operation; after the discrete space is continuous, the structured search task becomes: learning a set of continuous variables After the search is finished, each mixed operator is replaced by the operation with the highest probability to obtain a discrete structure, namely: The goal of structured search is to find an optimal structure parameter So that it can minimize the loss function of the validation set And find the optimal weight w* to minimize the training set loss, that is: The final bidirectional optimization problem can be expressed by the following formula: in, express and The validation loss function under represents w and The training loss function under , w represents the weight, Represents the structural parameters, represents the optimal structural parameter, and w* represents the optimal weight.
4. The customer behavior prediction device in the incoming call scenario according to claim 3, characterized in that: It also includes a network model creation module that performs the following iterative processing: Determine customer behavior model data according to the customer feature model data, the target structure parameters and corresponding weights; Determine a model loss function according to the customer behavior model data and the corresponding customer behavior actual data; When the current number of iterations reaches the preset number of model iterations, a customer behavior prediction network model is constructed according to the target structure parameters and the corresponding weights, otherwise the weights corresponding to the target structure parameters are updated according to the model loss function, and the iterative processing continues.
5. A computer device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that: When the processor executes the computer program, the steps of the customer behavior prediction method in the incoming call scenario described in any one of claims 1 to 2 are implemented.
6. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method for predicting customer behavior in an incoming call scenario according to any one of claims 1 to 2 are implemented.
Citation Information
Patent Citations
User voice call processing method, device and system based on big data
CN107155010A
Intelligent robot service method, electronic device and readable storage medium of computer
CN107688667A
Power customer demand prediction analysis method and system
CN110110881A
Customer service assisting method and device
CN110909148A
