CN114708110A - Joint training method and device for continuous guarantee behavior prediction model and electronic equipment - Google Patents

Joint training method and device for continuous guarantee behavior prediction model and electronic equipment Download PDF

Info

Publication number
CN114708110A
CN114708110A CN202210284098.2A CN202210284098A CN114708110A CN 114708110 A CN114708110 A CN 114708110A CN 202210284098 A CN202210284098 A CN 202210284098A CN 114708110 A CN114708110 A CN 114708110A
Authority
CN
China
Prior art keywords
continuous
prediction model
training
behavior prediction
loss function
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202210284098.2A
Other languages
Chinese (zh)
Inventor
刘齐
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Ping An Health Insurance Company of China Ltd
Original Assignee
Ping An Health Insurance Company of China Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Ping An Health Insurance Company of China Ltd filed Critical Ping An Health Insurance Company of China Ltd
Priority to CN202210284098.2A priority Critical patent/CN114708110A/en
Publication of CN114708110A publication Critical patent/CN114708110A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q40/00—Finance; Insurance; Tax strategies; Processing of corporate or income taxes
    • G06Q40/08—Insurance
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q30/00—Commerce
    • G06Q30/02—Marketing; Price estimation or determination; Fundraising
    • G06Q30/0201—Market modelling; Market analysis; Collecting market data
    • G06Q30/0202—Market predictions or forecasting for commercial activities

Landscapes

  • Engineering & Computer Science (AREA)
  • Business, Economics & Management (AREA)
  • Accounting & Taxation (AREA)
  • Finance (AREA)
  • Theoretical Computer Science (AREA)
  • Development Economics (AREA)
  • Strategic Management (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • General Business, Economics & Management (AREA)
  • Software Systems (AREA)
  • Economics (AREA)
  • Marketing (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Data Mining & Analysis (AREA)
  • Artificial Intelligence (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Computation (AREA)
  • Medical Informatics (AREA)
  • Game Theory and Decision Science (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Technology Law (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The application discloses a joint training method and device for a continuous guarantee behavior prediction model and electronic equipment, wherein the method comprises the following steps: constructing a joint training network of the continuous insurance behavior prediction model, wherein the joint training network comprises a plurality of continuous insurance behavior prediction models, and each continuous insurance behavior prediction model corresponds to a specified time node; constructing a training sample set; and inputting the training sample set into a joint training network for training, and adjusting the parameters of each continuous guarantee behavior prediction model according to the result of the overall loss function to obtain the continuous guarantee behavior prediction model corresponding to each designated time node. According to the method, the 'knowledge' learned by the continuous-keeping behavior prediction models of the observation points is utilized to cooperate with each other and assist in training, the precision of the continuous-keeping behavior prediction models of the observation points is improved, a joint training method is adopted, the prediction probability spaces of the continuous-keeping behavior prediction models of different time nodes are aligned, business comparison is facilitated, and marketing strategies are formulated.

Description

Joint training method and device for continuous guarantee behavior prediction model and electronic equipment
Technical Field
The application relates to the technical field of artificial intelligence, in particular to a joint training method and device for a continuous insurance behavior prediction model and electronic equipment.
Background
With the development of big data and artificial intelligence, the accurate marketing model is widely applied. In the insurance industry, in some conversion and retention scenarios, the distribution of the characteristics of the clients and the joint distribution of the characteristics and target variables during the conversion period may change greatly due to the long whole conversion period. In order to obtain better prediction accuracy, a feasible method is to select several typical observation points, construct a data set based on different observation points, and respectively establish a model. However, this method has many disadvantages, for example, because the model is independently modeled based on different observation points, for some samples with little change of client characteristics, several models may predict results with large deviation; if no post-processing is added, the prediction probability spaces of the models are not aligned to a uniform dimension; and each model does not utilize the mutual cooperation and auxiliary training of the knowledge learned by other models.
Disclosure of Invention
Aiming at the situations, the embodiment of the application provides a joint training method and device for a persistence behavior prediction model and electronic equipment, a cross-time point joint modeling method is adopted, the idea of knowledge distillation is used for reference, consistency loss is introduced into a final loss function of a joint training network, the models are promoted to learn each other, and the final model precision is improved.
In a first aspect, an embodiment of the present application provides a joint training method for a continuation behavior prediction model, including:
constructing a joint training network of a continuous activity prediction model, wherein the joint training network comprises a plurality of continuous activity prediction models, each continuous activity prediction model corresponds to a specified time node, and an overall loss function of the joint training network is jointly constructed according to the loss and consistency loss of each continuous activity prediction model;
constructing a training sample set, wherein any sample in the training sample set comprises user data and corresponding labels of a client corresponding to a plurality of designated time nodes;
inputting the training sample set into the joint training network for training, and adjusting the parameters of each continuous activity prediction model according to the result of the overall loss function to obtain the continuous activity prediction model corresponding to each designated time node.
In a second aspect, an embodiment of the present application further provides a joint training apparatus for a continuous assurance behavior prediction model, including:
the model construction unit is used for constructing a joint training network of the continuous-keeping behavior prediction model, the joint training network comprises a plurality of continuous-keeping behavior prediction models, each continuous-keeping behavior prediction model corresponds to a specified time node, and the whole loss function of the joint training network is constructed jointly according to the loss and consistency loss of each continuous-keeping behavior prediction model;
the system comprises a sample construction unit, a data analysis unit and a data analysis unit, wherein the sample construction unit is used for constructing a training sample set, and any sample in the training sample set comprises user data and corresponding labels of a client at a plurality of designated time nodes;
and the training unit is used for inputting the training sample set into the joint training network for training, and adjusting the parameters of each continuous keeping behavior prediction model according to the result of the overall loss function to obtain the continuous keeping behavior prediction model corresponding to each designated time node.
In a third aspect, an embodiment of the present application further provides an electronic device, including: a processor; and a memory arranged to store computer executable instructions that, when executed, cause the processor to perform any of the methods described above.
In a fourth aspect, this application embodiment also provides a computer-readable storage medium storing one or more programs which, when executed by an electronic device including a plurality of application programs, cause the electronic device to perform any of the methods described above.
The embodiment of the application adopts at least one technical scheme which can achieve the following beneficial effects:
the method comprises the steps of constructing a joint training network of a continuous activity prediction model and a training sample set, wherein the joint training network comprises a plurality of continuous activity prediction models, each continuous activity prediction model corresponds to an appointed time node, a consistency loss function is designed, the loss function of each continuous activity prediction model and the consistency loss function jointly form an overall loss function of the joint training network, the overall loss function is used as a reference and is used as a basis for adjusting the parameter adjustment of the continuous activity prediction model, and the training sample set is adopted to carry out overall training on the continuous activity prediction model joint training network to obtain the continuous activity prediction model of each appointed time node. The method uses the thought of knowledge distillation for reference, adopts a cross-time point combined modeling method, utilizes mutual cooperation and auxiliary training of knowledge learned by the continuous-keeping behavior prediction model of each observation point, and improves the precision of the continuous-keeping behavior prediction model of each observation point, thereby realizing that the model can output more consistent prediction for samples with little change of client characteristics; by adopting a joint training method, the prediction probability spaces of the continuous-keeping behavior prediction models of different time nodes are aligned, so that the services are compared with each other conveniently, and a marketing strategy is formulated.
Drawings
The accompanying drawings, which are included to provide a further understanding of the application and are incorporated in and constitute a part of this application, illustrate embodiment(s) of the application and together with the description serve to explain the application and not to limit the application. In the drawings:
FIG. 1 illustrates a flow diagram of a method of joint training of an continuance behavior prediction model according to an embodiment of the present application;
FIG. 2 illustrates a schematic structural diagram of a joint training network of an continuance behavior prediction model according to another embodiment of the present application;
FIG. 3 illustrates a schematic structural diagram of a joint training apparatus for an continuance behavior prediction model according to an embodiment of the present application;
fig. 4 is a schematic structural diagram of an electronic device in an embodiment of the present application.
Detailed Description
In order to make the objects, technical solutions and advantages of the present application more apparent, the technical solutions of the present application will be described in detail and completely with reference to the following specific embodiments of the present application and the accompanying drawings. It should be apparent that the described embodiments are only some of the embodiments of the present application, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present application.
The technical solutions provided by the embodiments of the present application are described in detail below with reference to the accompanying drawings.
The idea of the application lies in that a cross-time point combined modeling method is adopted, the idea of knowledge distillation is used for reference, consistency loss is introduced into a final loss function for the same client, models are promoted to learn each other, and the final model precision is improved.
Fig. 1 is a flowchart illustrating a method for jointly training a continuous-activities prediction model according to an embodiment of the present application, and as can be seen from fig. 1, the present application at least includes steps S110 to S130:
step S110: the method comprises the steps of constructing a joint training network of a continuous activity prediction model, wherein the joint training network comprises a plurality of continuous activity prediction models, each continuous activity prediction model corresponds to a designated time node, and an overall loss function of the joint training network is jointly constructed according to the loss and consistency loss of each continuous activity prediction model.
With the popularity of insurance businesses, more and more individuals or companies choose to purchase insurance products to provide more assurance of future uncertainty. For an insurance company, after a user (a user who has purchased an insurance product) has expired, the insurance company may want to continue a guarantee for the user, and even if the user has a desire to continue the guarantee, the insurance company may want to intervene with a behavior of the user to enable the user to continue the guarantee.
For example, in the case of medical insurance, the medical insurance is continuously maintained for one year, after the insurance product is expired, the insurance is invalid, the user continues to maintain the insurance, and the insurance company provides the insurance service for the next year for the user. For insurance companies, it is desirable that the user be able to keep as long as possible year after year. Generally, the insurance is effective when the continuous insurance action is continued within a period of time before and after the insurance expires, but in order to improve the conversion and the retention of the user, the insurance company reminds the user of continuous insurance for a period of time before the insurance product expires, but a part of users still do not carry out continuous insurance, and at the moment, the insurance company intervenes in the action of the part of users who do not continue the insurance, especially the users who have certain continuous insurance willingness, so that the accurate prediction of the continuous insurance action or willingness of the user has great significance for the insurance company.
Due to the specificity of insurance, the time period for converting a potential user into a transaction user, or for carrying out the retention process by a user in insurance is long, such as 3-4 months, even half a year. During this conversion period, the client feature distribution, the joint distribution of the features and the target variables may be greatly changed. In the prior art, in order to achieve better prediction accuracy, several (usually 2) typical time observation points are selected, such as 60 days before insurance expires, 30 days before insurance expires, and the like; the method comprises the steps of constructing a sample data set based on different observation points, establishing prediction models of the observation points at different times, respectively training the constructed models of the nodes at different times to obtain prediction models of the observation points at the same time, and predicting the behavior of a user which is not converted at different time points.
However, the method has many disadvantages, for example, because the model is independently modeled based on different observation points, for some samples with little change of client characteristics, several models may predict results with large deviation; and the prediction probability spaces of several models are not aligned to a uniform dimension, for example, one model predicts a probability of 80% of sustainable behavior and the other model predicts 60% of sustainable behavior, but 80% and 60% are not comparative due to the different dimensions between different models; in addition, the 'knowledge' utilized by each model is isolated, and the accuracy of model prediction is low.
In contrast, the knowledge learned by the model can be mutually transferred by adopting a cross-time point combined modeling method, by using the idea of knowledge distillation as reference and restricting the consistency of the distribution of the prediction results of the model at each time point. In the final loss function, not only is the loss function used when the model of each time node is trained independently considered, but also consistency loss is introduced, so that the models are promoted to learn each other, and the final model precision is improved.
Specifically, firstly, a joint training network of the continuous activity prediction model is constructed, the joint training network comprises a plurality of continuous activity prediction models, each continuous activity prediction model corresponds to a specified time node, wherein an overall loss function of the joint training network is constructed according to the loss function and the consistency loss function of each continuous activity prediction model, and the weight of the consistency loss function is adjusted by adopting a specified parameter (alpha in formula 1).
Taking the continuous activity prediction model as two examples, fig. 2 shows a schematic structure diagram of a joint training network of the continuous activity prediction model according to an embodiment of the present application, and as can be seen from fig. 2, the joint training network 200 includes a continuous activity prediction model of a first time node, which is denoted as a T1 model 210, and a continuous activity prediction model of a second time node, which is denoted as a T2 model 220, and loss functions of the T1 model 210 and the T2 model 220 and the consistency loss function together form a loss function of the joint training network 200 of the continuous activity prediction model. The T1 model 210 and the T2 model 220 correspond to different designated time nodes, for example, the T1 model 210 corresponds to 60 days before the policy expires, and the T2 model 220 corresponds to the policy expires. When a given time node is designated as appropriate, it is desirable to have a renewal behavior prediction model for predicting when a user's behavior will occur.
In some embodiments of the present application, the consistency loss function is determined based on the weight of the consistency loss function, the temperature hyperparameter, and the like, and the concept of the design of the loss function is as follows: by using knowledge distillation thought for reference, the knowledge learned by the model can be mutually transferred by restricting the consistency of the distribution of the model prediction results at each time point.
Step S120: and constructing a training sample set, wherein any sample in the training sample set comprises user data and corresponding labels of a client at a plurality of designated time nodes.
When constructing the training sample set, the data and labels of the clients needing to be collected at a plurality of designated time nodes are in one-to-one correspondence with the designated time nodes in the step S110.
Assume that the customer in the sample is denoted as customer A, assume that there are 2 designated time nodes, i.e., observation points, one of which is a policy toThe other is 60 days before the expiration of the policy. When a training sample set is constructed, respectively acquiring data of client A60 days before the insurance policy is due, and marking the data as X1And the corresponding label y1(ii) a Acquiring data of client A when the policy is due, and marking the data as X2And the corresponding label y2For a training sample, can be noted as (X)1，X2)->(y1，y2). It should be noted that, for different customers, the expiration time of the customer's policy is taken as a reference, for example, the policy of customer a is expired No. 4.1, and the collected data are data of two time points, i.e., forward reckoning for 60 days No. 4.1 and data of No. 4.1; for client B, the policy expiration is 7.1, the collected data are forward estimated for 60 days under 7.1, and the two time points of 7.1, i.e. each sample is based on the policy expiration, and do not affect each other.
Wherein, in some embodiments of the present application, the renewal behavior prediction model is a binary model, y1And y2The specific value of (2) is 0 or 1, 0 means that renewal is not performed, and 1 means that renewal is performed. A plurality of samples form a training sample set, in the application, the number of the samples in the training sample set is not limited, and usually tens of thousands of samples can achieve a better training effect, and the more training samples, the better, such as hundreds of thousands of samples.
Step S130: inputting the training sample set into the joint training network for training, and adjusting the parameters of each continuous activity prediction model according to the result of the overall loss function to obtain the continuous activity prediction model corresponding to each designated time node.
After the joint training network of the continuous activity prediction model and the training sample set are obtained through construction, the training sample set can be input into the joint training network of the continuous activity prediction model for joint training, and the continuous activity prediction model of each appointed time node can be obtained.
After the training process of one round is finished, parameters of each model in the whole network are adjusted according to the result of the whole loss function obtained in the training process of the round, the updated model parameters are used for the next round of training until the preset training requirement is met, if the preset training requirement is iteration 20000 times, the whole training process is finished after the iteration number is met, and the final continuous guarantee behavior prediction model of each time node is obtained.
According to the method, a plurality of continuous guarantee behavior prediction models can be obtained through one-time training, and each continuous guarantee behavior prediction model corresponds to one appointed time node. In the prediction, the conversion probability of the customer is predicted based on different observation points by using the corresponding model and the user data of the corresponding observation point.
The method shown in fig. 1 shows that a joint training network of the continuous activity prediction model and a training sample set are constructed in the application, the joint training network comprises a plurality of continuous activity prediction models, each continuous activity prediction model corresponds to an appointed time node, a consistency loss function is designed, the loss function of each continuous activity prediction model and the consistency loss function jointly form an overall loss function of the joint training network, the overall loss function is used as a reference and is used as a basis for adjusting parameter adjustment of the continuous activity prediction model, and the training sample set is used for carrying out overall training on the continuous activity prediction model joint training network to obtain the continuous activity prediction model of each appointed time node. The method uses the thought of knowledge distillation for reference, adopts a cross-time point combined modeling method, utilizes mutual cooperation and auxiliary training of knowledge learned by the continuous-keeping behavior prediction model of each observation point, and improves the precision of the continuous-keeping behavior prediction model of each observation point, thereby realizing that the model can output more consistent prediction for samples with little change of client characteristics; by adopting a joint training method, the prediction probability spaces of the continuous guarantee behavior prediction models of different time nodes are aligned, so that the services are compared with each other conveniently, and a marketing strategy is formulated.
In some embodiments of the present application, in the above method, the number of the designated time nodes is plural, and the overall loss function L (θ) is:
Figure BDA0003559370480000071
wherein the content of the first and second substances,
Figure BDA0003559370480000072
if each continuous guarantee behavior prediction model is a binary classification model, then
Figure BDA0003559370480000073
If each of the continuous guarantee behavior prediction models is a multi-classification model, then
Figure BDA0003559370480000074
Figure BDA0003559370480000075
Characterizing a loss of each of the renewal behavior prediction models;
n represents the number of training samples, a represents the weight of the consistency loss function, T represents the temperature over-parameter,
Figure BDA0003559370480000076
indicating whether the current sample belongs to the task,
Figure BDA0003559370480000077
represents a sample xjThe true value of (a) of (b),
Figure BDA0003559370480000078
and predicting the output content of the model for each continuous keeping behavior.
In the actual scene, the scenes in which the binary models are used are relatively wide, and therefore, the binary models are generally used
Figure BDA0003559370480000079
Wherein, if the number of the designated time nodes is two, the overall loss function L (θ) is:
Figure BDA00035593704800000710
wherein the content of the first and second substances,
Figure BDA00035593704800000711
Figure BDA00035593704800000712
Figure BDA00035593704800000713
respectively represent samples x1，x2The true value of (a) of (b),
Figure BDA00035593704800000714
the output content of the prediction model for each renewal behavior is as above with other parameters.
In some embodiments of the present application, in the above method, the plurality of designated time nodes comprises: the policy expiration time, a first specified time before the policy expiration, and a second specified time after the policy expiration.
In some embodiments of the present application, in the above method, the building a joint training network of the continuous activity prediction model includes: determining a plurality of designated time nodes; and determining the number of the continuous maintenance behavior prediction models and the expression of the overall loss function according to the number of the appointed time nodes, and forming a joint training network of the continuous maintenance behavior prediction models.
That is, when a time node is set in order to use a model for predicting the behavior of a user at the time node, how many continuous behavior prediction models are set in order to predict the behavior of several time nodes. Generally, renewal occurs before and after the expiration of the policy for a period of time during which the policy is valid, and thus, this period of time is a period of high incidence of user behavior prediction and intervention, such as 60 days before the expiration of the policy, and 60 days after the expiration of the policy. Therefore, the appointed time node can be set at the policy expiration time, the first appointed time before the policy expiration, and the second appointed time after the policy expiration, and the obtained multiple continuous-keeping behavior prediction models can be respectively used for predicting the behaviors of the user at the first appointed time before the policy expiration, the policy expiration time, and the second appointed time after the policy expiration, so as to meet the application requirements.
In some embodiments of the present application, in the method, the structures of the continuous activity prediction models are the same, and the structures of the continuous activity prediction models respectively include: the embedding layer connects MLP (multi-layer perceptron) models. However, the present invention is not limited to this, and an appropriate model structure may be determined by more specific scene characteristics.
The sensing device comprises an embedding layer, a multi-layer Perceptron (MLP) layer, an Artificial Neural Network (ANN) layer and a plurality of hidden layers, wherein the embedding layer is used for converting training samples into input vectors, the ANN layer is also called an Artificial Neural Network (MLP), the multiple hidden layers can be arranged in the middle of the MLP layer except for an input and output layer, the simplest MLP layer only comprises one hidden layer, namely a three-layer structure, and the multiple layers of Perceptron layer are all connected with one another. The bottom layer of the multilayer perceptron is an input layer, the middle layer is a hidden layer, and the last layer is an output layer. The MLP model in the prior art can be directly cited, and can also be modified according to requirements, and the application is not limited.
In some embodiments of the present application, in the above method, the constructing the training sample set comprises: for a training sample, acquiring user information and user behaviors of the training sample at a plurality of designated times, and determining user data and corresponding labels of a user corresponding to nodes at a plurality of designated times according to the user information and the user behaviors; and circularly executing the steps of acquiring the user information and the user behaviors of one training sample at a plurality of designated times, and determining the user data and the corresponding labels of the user corresponding to the nodes at a plurality of designated times according to the user information and the user behaviors until a designated number of training samples are obtained to form the training sample set.
In the training phase, a sample is constructed by selecting the policy for which the renewal period (during which the customer can decide whether to renew the policy) has ended, for which we know whether it has been renewed and therefore can be used to construct the Y label; and for the characteristics, selecting behavior data before the corresponding observation point to construct.
Acquiring user information and user behaviors of a training sample at a plurality of specified times, wherein the user information comprises but is not limited to identity, age, working property and income condition; the user behavior includes, but is not limited to, whether to purchase insurance, purchase insurance seeds, premium amount, etc. According to the user information and the user behavior, determining user data and corresponding labels corresponding to the user at a plurality of designated time nodes, in a simple way, the identity, the age, the working property, the income condition, the insurance type purchase, the premium and the insurance amount can be used as the user data of the user, and whether to purchase insurance can be used for labeling the training sample.
In the prediction phase, when the policy satisfies the corresponding time point, the corresponding model is used for prediction. In some embodiments of the present application, in the method, the inputting the training sample set into the joint training network for training, and adjusting parameters of each of the continuous behavior prediction models according to a result of the overall loss function to obtain a continuous behavior prediction model corresponding to each designated time node includes: sampling is carried out on the training sample set to obtain a target training sample; dividing a target training sample into a plurality of input vectors corresponding to each designated time node; respectively inputting the input vector of each appointed time node into a corresponding continuous guarantee behavior prediction model; determining the value of the overall loss function according to the output content of each continuous guarantee behavior prediction model; and updating the parameters of each continuous guarantee behavior prediction model according to the value of the overall loss function.
In some embodiments of the application, each training sample includes feature data of a client at different designated time nodes, and the feature data that does not pass through the designated time nodes is input into the corresponding model during training, assuming that the data of sample a includes (X) of the first designated time node1，Y1) And a second time node (X)2，Y2) (X) of the first designated time node1，Y1) Recording as a first input vector, the (X) of the second time node2，Y2) And (2) recording as a second input vector, assuming that a first designated time node corresponds to the T1 model 210 (figure 2) and a second designated time node corresponds to the T2 model 220 (figure 2), during training, inputting the first input vector into the T1 model 210 and inputting the second input vector into the T2 model 220, then enabling the output contents of the T1 model 210 and the T2 model 220 to enter an overall loss function, calculating and determining the value of the overall loss function, and updating the parameters of each continuous maintenance behavior prediction model according to the value of the overall loss function, so that one round of training is completed. And circularly executing the steps until the preset precision requirement is met, and obtaining the continuous keeping behavior prediction model of each appointed time node.
It should be noted that, through one-time joint training, a plurality of continuous guarantee behavior prediction models are obtained, each model corresponds to a time node, and which model is adopted when the behavior of the time node of the user is to be predicted, and after the training is completed, the continuous guarantee behavior prediction models are used independently in the prediction process.
Fig. 3 is a schematic structural diagram of a joint training apparatus of a continuous activity prediction model according to an embodiment of the present application, and as can be seen from fig. 3, the apparatus 300 includes:
the model construction unit 310 is configured to construct a joint training network of the continuous activity prediction models, where the joint training network includes a plurality of continuous activity prediction models, and each continuous activity prediction model corresponds to a specified time node, where an overall loss function of the joint training network is constructed jointly according to the loss and the consistency loss of each continuous activity prediction model;
a sample construction unit 320, configured to construct a training sample set, where any sample in the training sample set includes user data and corresponding labels that correspond to a client at a plurality of designated time nodes;
and the training unit 330 is configured to input the training sample set into the joint training network for training, and adjust parameters of each renewal behavior prediction model according to the result of the overall loss function to obtain a renewal behavior prediction model corresponding to each designated time node.
In some embodiments of the present application, in the above apparatus, the number of the designated time nodes is plural, and the overall loss function L (θ) is:
Figure BDA0003559370480000101
wherein the content of the first and second substances,
Figure BDA0003559370480000102
if each continuous guarantee behavior prediction model is a binary classification model, then
Figure BDA0003559370480000103
If each of the continuous guarantee behavior prediction models is a multi-classification model, then
Figure BDA0003559370480000104
Figure BDA0003559370480000105
Characterizing a loss of each of the renewal behavior prediction models;
n represents the number of training samples, a represents the weight of the consistency loss function, T represents the temperature over-parameter,
Figure BDA0003559370480000106
indicating whether the current sample belongs to the task or not,
Figure BDA0003559370480000107
represents a sample xjThe true value of (a) of (b),
Figure BDA0003559370480000108
and predicting the output content of the model for each continuous keeping behavior.
In some embodiments of the present application, in the above apparatus, the number of the designated time nodes is two, and the overall loss function L (θ) is:
Figure BDA0003559370480000109
wherein the content of the first and second substances,
Figure BDA00035593704800001010
Figure BDA00035593704800001011
Figure BDA00035593704800001012
respectively represent samples x1，x2The true value of (a) of (b),
Figure BDA00035593704800001013
and predicting the output content of the model for each continuous keeping behavior.
In some embodiments of the present application, in the above apparatus, the plurality of designated time nodes includes: the policy expiration time, a first specified time length before the policy expiration and a second specified time length after the policy expiration; a model building unit 310 for determining a plurality of specified time nodes; and determining the number of the continuous maintenance behavior prediction models and the expression of the overall loss function according to the number of the appointed time nodes, and forming a joint training network of the continuous maintenance behavior prediction models.
In some embodiments of the present application, in the above apparatus, the sample constructing unit 320 is configured to, for a training sample, obtain user information and user behaviors of the training sample at multiple specified times, and determine, according to the user information and the user behaviors, user data and corresponding labels that the user corresponds to nodes at multiple specified times; and circularly executing the steps of acquiring the user information and the user behaviors of the training sample at a plurality of designated time, and determining the user data and the corresponding labels of the user at a plurality of designated time nodes according to the user information and the user behaviors until a designated number of training samples are obtained to form the training sample set.
In some embodiments of the present application, in the above apparatus, the training unit 330 is configured to sample in the training sample set to obtain a target training sample; dividing a target training sample into a plurality of input vectors corresponding to each designated time node; respectively inputting the input vectors of all the appointed time nodes into the corresponding persistence behavior prediction models; determining the value of the overall loss function according to the output content of each continuous guarantee behavior prediction model; and updating the parameters of each continuous keeping behavior prediction model according to the value of the overall loss function.
In some embodiments of the present application, in the apparatus, the structures of the continuous guarantee behavior prediction models are the same, and the structures of the continuous guarantee behavior prediction models respectively include: the embedding layer connects the MLP models.
It should be noted that, the joint training device of the continuous activity prediction model can implement the joint training method of the continuous activity prediction model one by one, and details are not repeated here.
Fig. 4 is a schematic structural diagram of an electronic device according to an embodiment of the present application. Referring to fig. 4, at a hardware level, the electronic device includes a processor, and optionally further includes an internal bus, a network interface, and a memory. The Memory may include a Memory, such as a Random-Access Memory (RAM), and may further include a non-volatile Memory, such as at least 1 disk Memory. Of course, the electronic device may also include hardware required for other services.
The processor, the network interface, and the memory may be connected to each other via an internal bus, which may be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus, an EISA (Extended Industry Standard Architecture) bus, or the like. The bus may be divided into an address bus, a data bus, a control bus, etc. For ease of illustration, only one double-headed arrow is shown in FIG. 4, but that does not indicate only one bus or one type of bus.
And the memory is used for storing programs. In particular, the program may include program code comprising computer operating instructions. The memory may include both memory and non-volatile storage and provides instructions and data to the processor.
The processor reads the corresponding computer program from the nonvolatile memory into the memory and then runs the computer program to form the combined training device of the continuous-keeping behavior prediction model on the logic level. The processor is used for executing the program stored in the memory and is specifically used for executing the following operations:
constructing a joint training network of a continuous activity prediction model, wherein the joint training network comprises a plurality of continuous activity prediction models, each continuous activity prediction model corresponds to a specified time node, and an overall loss function of the joint training network is jointly constructed according to the loss and consistency loss of each continuous activity prediction model;
constructing a training sample set, wherein any sample in the training sample set comprises user data and corresponding labels of a client corresponding to a plurality of designated time nodes;
inputting the training sample set into the joint training network for training, and adjusting the parameters of each continuous activity prediction model according to the result of the overall loss function to obtain the continuous activity prediction model corresponding to each designated time node.
The method performed by the joint training device of the continuous preserving behavior prediction model disclosed in the embodiment of fig. 3 of the present application can be applied to or implemented by a processor. The processor may be an integrated circuit chip having signal processing capabilities. In implementation, the steps of the above method may be performed by integrated logic circuits of hardware in a processor or by instructions in the form of software. The Processor may be a general-purpose Processor, including a Central Processing Unit (CPU), a Network Processor (NP), and the like; but also Digital Signal Processors (DSPs), Application Specific Integrated Circuits (ASICs), Field Programmable Gate Arrays (FPGAs) or other Programmable logic devices, discrete Gate or transistor logic devices, discrete hardware components. The various methods, steps, and logic blocks disclosed in the embodiments of the present application may be implemented or performed. A general purpose processor may be a microprocessor or the processor may be any conventional processor or the like. The steps of the method disclosed in connection with the embodiments of the present application may be directly implemented by a hardware decoding processor, or implemented by a combination of hardware and software modules in the decoding processor. The software module may be located in ram, flash memory, rom, prom, or eprom, registers, etc. storage media as is well known in the art. The storage medium is located in a memory, and a processor reads information in the memory and completes the steps of the method in combination with hardware of the processor.
The electronic device may further execute the method executed by the joint training device of the continuous activity prediction model in fig. 3, and implement the functions of the joint training device of the continuous activity prediction model in the embodiment shown in fig. 3, which are not described herein again in this embodiment of the present application.
An embodiment of the present application further provides a computer-readable storage medium storing one or more programs, where the one or more programs include instructions, which, when executed by an electronic device including a plurality of application programs, enable the electronic device to perform the method performed by the joint training apparatus for the continuation behavior prediction model in the embodiment shown in fig. 3, and are specifically configured to perform:
constructing a joint training network of a continuous activity prediction model, wherein the joint training network comprises a plurality of continuous activity prediction models, each continuous activity prediction model corresponds to a specified time node, and an overall loss function of the joint training network is jointly constructed according to the loss and consistency loss of each continuous activity prediction model;
constructing a training sample set, wherein any sample in the training sample set comprises user data and corresponding labels of a client corresponding to a plurality of designated time nodes;
and inputting the training sample set into the joint training network for training, and adjusting the parameters of each continuous guarantee behavior prediction model according to the result of the overall loss function to obtain the continuous guarantee behavior prediction model corresponding to each designated time node.
As will be appreciated by one skilled in the art, embodiments of the present application may be provided as a method, system, or computer program product. Accordingly, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present application may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, and the like) having computer-usable program code embodied therein.
The present application is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the application. It will be understood that each flow and/or block of the flowchart illustrations and/or block diagrams, and combinations of flows and/or blocks in the flowchart illustrations and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.
These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means which implement the function specified in the flowchart flow or flows and/or block diagram block or blocks.
These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.
In a typical configuration, a computing device includes one or more processors (CPUs), input/output interfaces, network interfaces, and memory.
The memory may include forms of volatile memory in a computer readable medium, Random Access Memory (RAM) and/or non-volatile memory, such as Read Only Memory (ROM) or flash memory (flash RAM). Memory is an example of a computer-readable medium.
Computer-readable media, including both non-transitory and non-transitory, removable and non-removable media, may implement information storage by any method or technology. The information may be computer readable instructions, data structures, modules of a program, 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 disc read only memory (CD-ROM), Digital Versatile Discs (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices, or any other non-transmission medium that can be used to store information that can be accessed by a computing device. As defined herein, a computer readable medium does not include a transitory computer readable medium such as a modulated data signal and a carrier wave.
It should also be noted that the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising an … …" does not exclude the presence of other like elements in a process, method, article, or apparatus that comprises the element.
As will be appreciated by one skilled in the art, embodiments of the present application may be provided as a method, system, or computer program product. Accordingly, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present application may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, and the like) having computer-usable program code embodied therein.
The above description is only an example of the present application and is not intended to limit the present application. Various modifications and changes may occur to those skilled in the art. Any modification, equivalent replacement, improvement or the like made within the spirit and principle of the present application shall be included in the scope of the claims of the present application.

Claims (10)

1. A joint training method of a continuous guarantee behavior prediction model is characterized by comprising the following steps:
constructing a joint training network of a continuous activity prediction model, wherein the joint training network comprises a plurality of continuous activity prediction models, each continuous activity prediction model corresponds to a specified time node, and an overall loss function of the joint training network is jointly constructed according to the loss and consistency loss of each continuous activity prediction model;
constructing a training sample set, wherein any sample in the training sample set comprises user data and corresponding labels of a client corresponding to a plurality of designated time nodes;
inputting the training sample set into the joint training network for training, and adjusting the parameters of each continuous activity prediction model according to the result of the overall loss function to obtain the continuous activity prediction model corresponding to each designated time node.
2. The method of claim 1, wherein the number of the designated time nodes is plural, and the overall loss function L (θ) is:
Figure FDA0003559370470000011
wherein the content of the first and second substances,
Figure FDA0003559370470000012
if each continuous guarantee behavior prediction model is a binary classification model, then
Figure FDA0003559370470000013
If each of the continuous guarantee behavior prediction models is a multi-classification model, then
Figure FDA0003559370470000014
Figure FDA0003559370470000015
Characterizing a loss of each of the renewal behavior prediction models;
n represents the number of training samples, a represents the weight of the consistency loss function, T represents the temperature over-parameter,
Figure FDA0003559370470000016
indicating whether the current sample belongs to the task,
Figure FDA0003559370470000017
represents a sample xjThe true value of (a) of (b),
Figure FDA0003559370470000018
and predicting the output content of the model for each continuous keeping behavior.
3. The method of claim 2, wherein the number of the designated time nodes is two, and wherein the global loss function L (θ) is:
Figure FDA0003559370470000019
wherein, the first and the second end of the pipe are connected with each other,
Figure FDA00035593704700000110
Figure FDA00035593704700000111
Figure FDA00035593704700000112
respectively represent samples x1，x2The true value of (a) of (b),
Figure FDA00035593704700000113
and predicting the output content of the model for each continuous guarantee behavior.
4. The method of claim 1, wherein the plurality of designated time nodes comprises: the policy expiration time, a first specified duration before the policy expiration, and a second specified duration after the policy expiration;
the joint training network for constructing the continuous activity prediction model comprises the following steps:
determining a plurality of designated time nodes;
and determining the number of the continuous and continuous behavior prediction models and the expression of the overall loss function according to the number of the specified time nodes to form a joint training network of the continuous and continuous behavior prediction models.
5. The method of claim 1, wherein constructing the training sample set comprises:
for a training sample, acquiring user information and user behaviors of the training sample at a plurality of designated times, and determining user data and corresponding labels of a user corresponding to nodes at a plurality of designated times according to the user information and the user behaviors;
and circularly executing the steps of acquiring the user information and the user behaviors of one training sample at a plurality of designated times, and determining the user data and the corresponding labels of the user corresponding to the nodes at a plurality of designated times according to the user information and the user behaviors until a designated number of training samples are obtained to form the training sample set.
6. The method of claim 1, wherein the inputting the training sample set into the joint training network for training and adjusting parameters of each of the continuous behavior prediction models according to the result of the global loss function to obtain a continuous behavior prediction model corresponding to each designated time node comprises:
sampling is carried out on the training sample set to obtain a target training sample;
dividing a target training sample into a plurality of input vectors corresponding to each designated time node;
respectively inputting the input vector of each appointed time node into a corresponding continuous guarantee behavior prediction model;
determining the value of the overall loss function according to the output content of each continuous guarantee behavior prediction model;
and updating the parameters of each continuous keeping behavior prediction model according to the value of the overall loss function.
7. The method according to any one of claims 1 to 6, wherein the structures of the continuous activity prediction models are the same, and the structures of the continuous activity prediction models respectively comprise: the embedding layer connects the MLP models.
8. A joint training apparatus for a continuation of life behavior prediction model, comprising:
the model construction unit is used for constructing a joint training network of the continuous insurance behavior prediction model, the joint training network comprises a plurality of continuous insurance behavior prediction models, each continuous insurance behavior prediction model corresponds to a specified time node, and the whole loss function of the joint training network is constructed together according to the loss and consistency loss of each continuous insurance behavior prediction model;
the system comprises a sample construction unit, a data analysis unit and a data analysis unit, wherein the sample construction unit is used for constructing a training sample set, and any sample in the training sample set comprises user data and corresponding labels corresponding to a client at a plurality of designated time nodes;
and the training unit is used for inputting the training sample set into the joint training network for training, and adjusting the parameters of each continuous keeping behavior prediction model according to the result of the overall loss function to obtain the continuous keeping behavior prediction model corresponding to each designated time node.
9. An electronic device, comprising:
a processor; and
a memory arranged to store computer executable instructions that, when executed, cause the processor to perform the method of claims 1 to 7.
10. A computer readable storage medium storing one or more programs which, when executed by an electronic device comprising a plurality of application programs, cause the electronic device to perform the method of claims 1-7.
CN202210284098.2A 2022-03-22 2022-03-22 Joint training method and device for continuous guarantee behavior prediction model and electronic equipment Pending CN114708110A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210284098.2A CN114708110A (en) 2022-03-22 2022-03-22 Joint training method and device for continuous guarantee behavior prediction model and electronic equipment

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210284098.2A CN114708110A (en) 2022-03-22 2022-03-22 Joint training method and device for continuous guarantee behavior prediction model and electronic equipment

Publications (1)

Publication Number Publication Date
CN114708110A true CN114708110A (en) 2022-07-05

Family

ID=82168090

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210284098.2A Pending CN114708110A (en) 2022-03-22 2022-03-22 Joint training method and device for continuous guarantee behavior prediction model and electronic equipment

Country Status (1)

Country Link
CN (1) CN114708110A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116993663A (en) * 2023-06-12 2023-11-03 阿里巴巴（中国）有限公司 Image processing method and training method of image processing model

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116993663A (en) * 2023-06-12 2023-11-03 阿里巴巴（中国）有限公司 Image processing method and training method of image processing model
CN116993663B (en) * 2023-06-12 2024-04-30 阿里巴巴（中国）有限公司 Image processing method and training method of image processing model

Similar Documents

Publication Publication Date Title
US11281969B1 (en) Artificial intelligence system combining state space models and neural networks for time series forecasting
Nijkamp et al. Comparative modelling of interregional transport flows: Applications to multimodal European freight transport
CN110686633B (en) Landslide displacement prediction method and device and electronic equipment
CN112381216B (en) Training and predicting method and device for mixed graph neural network model
Peng et al. Lifelong property price prediction: A case study for the toronto real estate market
CN111080304A (en) Credible relationship identification method, device and equipment
Brahimi et al. Modelling on Car‐Sharing Serial Prediction Based on Machine Learning and Deep Learning
CN114708110A (en) Joint training method and device for continuous guarantee behavior prediction model and electronic equipment
CN111160566A (en) Sample generation method and device, computer readable storage medium and computer equipment
CN114969191A (en) Data analysis method, system and device based on big data and storage medium
CN113837635A (en) Risk detection processing method, device and equipment
Shah et al. A neoteric technique using ARIMA-LSTM for time series analysis on stock market forecasting
Bin et al. Peer-dependence valuation model for real estate appraisal
CN116257798A (en) Click rate prediction model training and click rate prediction method, system and equipment
CN116029422A (en) Behavior sequence prediction method and device, electronic equipment and medium
CN115526402A (en) Index prediction method, index prediction device, electronic equipment and storage medium
CN114925919A (en) Service resource processing method and device, computer equipment and storage medium
CN110060146B (en) Data transfer method and related equipment
CN109299321B (en) Method and device for recommending songs
CN115640336B (en) Business big data mining method, system and cloud platform
CN113393330B (en) Financial wind control management system based on block chain
CN112949882A (en) Load prediction method and device and electronic equipment
Huang A demonstration of reliability of the interior-outer-set model
CN117808040B (en) Method and device for predicting low forgetting hot events based on brain map
CN117891811B (en) Customer data acquisition and analysis method and device and cloud server

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination