WO2018201646A1 - 模型分析方法、装置、及计算机可读存储介质 - Google Patents

模型分析方法、装置、及计算机可读存储介质 Download PDF

Info

Publication number
WO2018201646A1
WO2018201646A1 PCT/CN2017/100042 CN2017100042W WO2018201646A1 WO 2018201646 A1 WO2018201646 A1 WO 2018201646A1 CN 2017100042 W CN2017100042 W CN 2017100042W WO 2018201646 A1 WO2018201646 A1 WO 2018201646A1
Authority
WO
WIPO (PCT)
Prior art keywords
customer information
model
predetermined
proportion
information sample
Prior art date
Application number
PCT/CN2017/100042
Other languages
English (en)
French (fr)
Inventor
陈依云
Original Assignee
平安科技（深圳）有限公司
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 平安科技（深圳）有限公司 filed Critical 平安科技（深圳）有限公司
Priority to SG11201900264PA priority Critical patent/SG11201900264PA/en
Priority to US16/084,242 priority patent/US11507963B2/en
Priority to KR1020187024854A priority patent/KR20190021189A/ko
Priority to EP17899229.3A priority patent/EP3428854A4/en
Priority to AU2017408798A priority patent/AU2017408798A1/en
Priority to JP2018534667A priority patent/JP2019519821A/ja
Publication of WO2018201646A1 publication Critical patent/WO2018201646A1/zh

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
    • G06Q30/00—Commerce
    • G06Q30/02—Marketing; Price estimation or determination; Fundraising
    • G06Q30/0201—Market modelling; Market analysis; Collecting market data
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/10—Geometric CAD
    • 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
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • G06N20/20—Ensemble learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • 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
    • G06Q10/00—Administration; Management
    • G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/048—Activation functions
    • 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
    • G06Q10/00—Administration; Management
    • G06Q10/10—Office automation; Time management
    • 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
    • G06Q40/03—Credit; Loans; Processing thereof
    • 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

Definitions

  • the present invention relates to the field of computer technologies, and in particular, to a model analysis method, apparatus, and computer readable storage medium.
  • the main object of the present invention is to provide a model analysis method, apparatus, and computer readable storage medium, which are intended to improve the accuracy of prediction results.
  • a first aspect of the present invention provides a model analysis method, the method comprising the following steps:
  • a second aspect of the present invention provides a model analysis apparatus including a processing device and a storage device, the storage device storing a model analysis program, the model analysis program being executed by the processing device The following steps are implemented:
  • a third aspect of the present application provides a computer readable storage medium having stored thereon at least one computer readable instruction executable by a processing device to:
  • the model analysis method, device and computer readable storage medium provided by the invention train a plurality of predetermined models through a preset number of customer information samples, and combine the trained multiple models into a composite model, and receive the After analyzing the customer information, the combined customer model is used to analyze the customer information to be analyzed. Because it combines multiple models to analyze and predict using the combined composite model, it can combine the advantages of different models, and the accuracy of the prediction results is effectively improved compared with the single model prediction.
  • FIG. 1 is a schematic flow chart of an embodiment of a model analysis method according to the present invention.
  • FIG. 2 is a schematic diagram of a refinement process of step S10 in FIG. 1;
  • FIG. 3 is a schematic diagram of an embodiment of a model analysis device of the present invention.
  • the present invention provides a model analysis method.
  • FIG. 1 is a schematic flow chart of an embodiment of a model analysis method according to the present invention.
  • the model analysis method comprises:
  • Step S10 training a plurality of predetermined models based on a preset number of customer information samples
  • a plurality of predetermined models are trained based on a preset number (for example, 100,000) of customer information samples.
  • the customer information in the customer information sample includes, but is not limited to, gender, age, contact information, home address, work unit, credit history record, insurance product information held, insurance behavior habits, historical claims information, etc.
  • Some insurance product information includes but is not limited to insurance-type insurance products, income-based insurance products, short-term insurance products, life-long insurance products, etc., and the behavior of insurance is a customer for a period of time (such as the last one or three years)
  • the product that holds the longest time in the insurance product held, or holds the product with the largest proportion represents that the customer's insurance behavior is the product. For example, if a customer holds more than a predetermined percentage (for example, 60%) of the insurance products is a security product, the insurance behavior on behalf of the customer is a preference insurance product.
  • Models that need to be trained in advance include, but are not limited to, Decision Tree models, linear regression models, Logistic regression models, Neural Networks (NN) models, and the like.
  • the decision tree model is a simple but widely used classifier. By constructing a decision tree through training data, it is possible to classify unknown data efficiently.
  • the decision tree model is readable, descriptive, and helpful for manual analysis; and efficient.
  • Linear regression model can be In linear regression or multiple linear regression models, unary linear regression is a major influencing factor as an independent variable to explain the variation of dependent variables. In the study of real problems, the variation of dependent variables is often affected by several important factors. It is necessary to use two or more influencing factors as independent variables to explain the change of the dependent variable. This is the multiple regression and multiple regression.
  • Logistic regression model is a commonly used machine learning model in the industry, which is used to estimate the possibility of a certain thing, such as the possibility of predicting the type of customer insured or insured in this embodiment.
  • the neural network model is a complex network system formed by a large number of simple processing units (called neurons) interconnected to each other. It reflects many basic features of human brain function and is a highly complex nonlinear dynamic learning system. .
  • Neural networks have massively parallel, distributed storage and processing, self-organizing, adaptive, and self-learning capabilities, and are particularly well-suited for handling inaccurate and ambiguous information processing problems that require many factors and conditions to be considered simultaneously. For example, after training the neural network model, the trained neural network model can be used to predict the probability of the customer's insurance or the probability of the type of insurance.
  • step S20 the trained multiple models are combined into a composite model according to a predetermined combination rule, and after receiving the customer information to be analyzed, the customer information to be analyzed is input into the composite model to output an analysis result.
  • a plurality of predetermined models trained based on a preset number of customer information samples such as a decision tree model, a linear regression model, a logistic regression model, a neural network model, and the like, are combined into a composite model according to a predetermined combination rule. For example, according to the characteristics and advantages of different models, and the characteristics of the customer information to be analyzed, the corresponding weights are set for different models. If the relationship between the dependent variable and the target variable is a simple linear relationship, then the composite model can be combined. The weight of the linear regression model is increased to improve the speed and efficiency of the prediction; if there are many dependent variables and complex analysis is needed, the weight of the neural network model can be improved when the composite model is combined to improve the accuracy of the prediction.
  • the customer information to be analyzed may be input into the composite model after receiving the customer information to be analyzed, so as to integrate multiple models.
  • Decision tree model, linear regression model, logistic regression model, neural network model and other advantages and different judgment angles are used to analyze and predict the customer information to be analyzed, so as to output more accurate analysis and prediction results.
  • a plurality of predetermined models are trained through a preset number of customer information samples, and the plurality of trained models are combined into a composite model.
  • the combined composite model is used to treat the Analyze customer information for analysis. Because it combines multiple models to analyze and predict using the combined composite model, it can combine the advantages of different models, and the accuracy of the prediction results is effectively improved compared with the single model prediction.
  • the predetermined number of models is N, N is a natural number greater than 2, the i-th predetermined model is denoted as Fi, and i is a positive integer less than or equal to N, Combine the trained models into a composite model according to predetermined combination rules:
  • the composite model (1/N) * F1 + (1/N) * F2 + ... + (1/N) * FN.
  • the plurality of models are averaged to Combining and forming a composite model can balance the influence of each model to balance the prediction results of each model, and obtain the most reasonable prediction results when the prediction results of each model are not much different.
  • the predetermined number of models is N, N is a natural number greater than 2, the i-th predetermined model is denoted as Fi, and i is a positive integer less than or equal to N, Combine the trained models into a composite model according to predetermined combination rules:
  • the N-th roots are obtained for the results analyzed by the respective models, and combined to form a composite model. Since the prediction results of each model have a great influence on the prediction results of the final combined model, the role of each model in the combined model can be highlighted, and the analysis and prediction of each model in the combined model can be maximized.
  • the analysis results of various aspects determine the prediction results of the final combined model and improve the accuracy of the prediction.
  • the predetermined number of models is N, N is a natural number greater than 2, the i-th predetermined model is denoted as Fi, and i is a positive integer less than or equal to N, Combine the trained models into a composite model according to predetermined combination rules:
  • the composite model N / (1/F1 + 1 / F2 + ... + 1 / FN).
  • the three models of logistic regression, decision tree, and neural network may be combined to form a combined model, and the performance of the combined model on the verification set is analyzed.
  • Table 1 lists the degree of elevation of the logistic regression, combined 1 to 6 models predicted on the validation set.
  • the depths 1-99 in Table 1 represent the corresponding logistic regression, and the combination 1 to 6 models in the prediction set on the verification set have a score of 1%-99%, and the experimental data in Table 1 shows that the depth is 1
  • the average performance of the combined model is 4.5% higher than that of the logistic regression; at a depth of 5, the average performance of the combined model is 5.3% higher than that of the logistic regression; at a depth of 10, the average performance of the combined model is higher than that of the logistic regression. 1.9%.
  • the prediction effect of the model is better than that of the single logistic regression model. That is, the combined model is more accurate than the single model prediction, which effectively improves the accuracy of the prediction results. degree.
  • step S10 may include:
  • Step S101 in the process of training a predetermined model, after each training, input each customer information sample into the currently trained model to determine a customer information sample of the model analysis error;
  • Step S102 calculating whether the ratio of the number of customer information samples of the model analysis error to the number of all customer information samples is less than a preset threshold
  • Step S103 if the proportion of the number of customer information samples in which the model analysis error accounts for the number of all customer information samples is less than a preset threshold, the predetermined model training ends;
  • Step S104 If the ratio of the number of sampled customer information samples to the total number of customer information samples is greater than or equal to a preset threshold, the customer information sample with the model analysis error is added to the total customer information sample according to the preset proportional increase rate. The proportion of the customer information samples belonging to the same type, and reducing the proportion of the correct customer information samples in the total customer information sample according to the preset ratio reduction, and returning to the above step S101.
  • a plurality of predetermined models are trained based on a preset number of customer information samples.
  • the accuracy of the current model is analyzed and judged after each training. For example, each customer information sample can be separately input into the currently trained model for analysis and prediction. If the model analyzes the wrong number of customer information samples and the proportion of all customer information samples is less than a preset threshold (for example, 5%), it indicates that the current model has a higher accuracy rate, and then the training is ended, and the current model is used as the model. Trained models.
  • the model analyzes the wrong number of customer information samples in the proportion of all customer information samples is greater than or equal to the preset threshold (for example, 5%), it indicates that the accuracy of the current model is low, and the increase rate according to the preset ratio (for example, 1%) in the total customer information sample, the proportion of the customer information sample belonging to the same type of model analysis error is added, and the proportion of the customer information sample of the same type is reduced according to the preset ratio (for example, 1%) in the total customer information sample.
  • the preset threshold for example, 5%
  • the reduction model analyzes the proportion of the correct customer information sample (for example, if the total customer information sample, the model analyzes the correct customer information sample percentage is 80%, after reduction, it is 79%), and the adjusted customer information On the basis of the sample, the model will continue to be learned and trained until the accuracy of the model meets the requirements.
  • the accuracy of the current model can be analyzed and judged after each training in the process of training a predetermined model, only the accuracy of the model meets the requirements to end the training, and each pre-use for the combination is guaranteed. Determine the high accuracy of the model. Moreover, when it is judged that the accuracy of the current model is not up to the requirement, the proportion of the customer information sample in the total customer information sample that is incorrectly analyzed by the model belongs to the same type of customer information sample, so that the customer is easily analyzed for the model. The information sample type is used to focus on learning and training the model, which is more targeted and improves the efficiency and speed of model training.
  • a specific description is made on whether a customer has a claim.
  • the usual predictive model is such that there is a business goal, such as predicting whether the customer will make a claim within the next six months, and the probability of claim is How big.
  • Predictors are the selection of data indicators that affect the target scalar, including: gender, age information, information on insurance products held (such as insurance-type insurance products, income-based insurance products, short-term insurance products, life-long insurance products, etc.), insurance Behavioral habits (for example, if a customer holds more than a predetermined percentage of the insurance products is a guaranteed product, the customer's insurance behavior is a preference for insurance products), historical claims information, and so on.
  • the decision tree model can be established based on the customer's predictor variables and target variables.
  • the decision tree model gives the probability of each customer's claim, if the threshold is set to 0.5, when the decision is made.
  • the tree model predicts that the probability of customer claims is greater than 0.5, it is considered that the customer will make claims within the next six months.
  • the decision tree model predicts that the probability of customer claims is less than 0.5, the customer is considered to be in the next six months. No claims will occur within.
  • the tree model analyzes the wrong customer information sample.
  • the invention further provides a model analysis device.
  • Figure 3 is a schematic illustration of an embodiment of a model analysis device of the present invention.
  • the model analysis device 2 may be a PC (Personal Computer), or may be a terminal device such as a smart phone, a tablet computer, an e-book reader, or a portable computer.
  • PC Personal Computer
  • the model analysis device 2 may be a PC (Personal Computer), or may be a terminal device such as a smart phone, a tablet computer, an e-book reader, or a portable computer.
  • the model analysis device 2 includes a storage device 11, a processing device 12, a communication bus 13, and a network interface 14.
  • the storage device 11 includes at least one type of readable storage medium including a flash memory, a hard disk, a multimedia card, a card type memory (for example, SD or DX memory, etc.), a magnetic memory, a magnetic disk, an optical disk, and the like. .
  • the storage device 11 may in some embodiments be an internal storage unit of a social network based user keyword extraction device, such as a hard disk of the model analysis device.
  • the storage device 11 may also be an external storage device of the model analysis device in other embodiments, such as a plug-in hard disk equipped on the model analysis device, a smart memory card (SMC), and a secure digital (Secure Digital, SD). ) cards, flash cards, etc.
  • the storage device 11 may also include both an internal storage unit of the model analysis device and an external storage device.
  • the storage device 11 can be used not only for storing application software installed in the model analysis device and various types of data, such as model analysis codes, but also for temporarily storing data that has been output or is to be output.
  • Processing device 12 may be a Central Processing Unit (CPU), controller, microcontroller, microprocessor, or other data processing chip for running program code or processing stored in memory 11. Data, such as executing a model analysis program.
  • CPU Central Processing Unit
  • controller microcontroller
  • microprocessor or other data processing chip for running program code or processing stored in memory 11.
  • Data such as executing a model analysis program.
  • the network interface 13 can optionally include a standard wired interface, a wireless interface (such as a WI-FI interface), and is typically used to establish a communication connection between the device and other electronic devices.
  • a standard wired interface such as a WI-FI interface
  • Communication bus 14 is used to implement connection communication between these components.
  • Figure 3 shows only the model analysis device with components 11-14, but it should be understood that not all illustrated components may be implemented and that more or fewer components may be implemented instead.
  • the device may further include a user interface
  • the user interface may include a display device
  • an input unit such as a keyboard
  • the optional user interface may further include a standard wired interface and a wireless interface.
  • the display device may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, an OLED (Organic Light-Emitting Diode) touch sensor, or the like.
  • the display device may also be appropriately referred to as a display screen or a display unit for displaying information processed in the model analysis device and a user interface for displaying the visualization.
  • a model analysis program is stored in the storage device 11; when the processing device 12 executes the model analysis program stored in the storage device 11, the following steps are implemented:
  • a plurality of predetermined models are trained based on a preset number (for example, 100,000) of customer information samples.
  • the customer information in the customer information sample includes, but is not limited to, gender, age, contact information, home address, work unit, credit history record, insurance product information held, insurance behavior habits, historical claims information, etc.
  • Some insurance product information includes but is not limited to protection Insurance products, income insurance products, short-term insurance products, life-long insurance products, etc., the behavior of insurance behavior is the most held by a customer in a period of time (such as the last one or three years) A long product, or the largest share of the product, represents the customer's insurance behavior is the product. For example, if a customer holds more than a predetermined percentage (for example, 60%) of the insurance products is a security product, the insurance behavior on behalf of the customer is a preference insurance product.
  • a predetermined percentage for example, 60%
  • Models that need to be trained in advance include, but are not limited to, Decision Tree, Linear Regression Model, Logistic regression model, Neural Networks (NN) model, and the like.
  • the decision tree model is a simple but widely used classifier. By constructing a decision tree through training data, it is possible to classify unknown data efficiently.
  • the decision tree model is readable, descriptive, and helpful for manual analysis; and efficient.
  • the linear regression model can be a linear regression or multiple linear regression model.
  • the linear regression is a major influencing factor as an independent variable to explain the variation of the dependent variable. In the study of real problems, the variation of the dependent variable is often affected by several important factors.
  • Logistic regression model is a commonly used machine learning model in the industry, which is used to estimate the possibility of a certain thing, such as the possibility of predicting the type of customer insured or insured in this embodiment.
  • the neural network model is a complex network system formed by a large number of simple processing units (called neurons) interconnected to each other. It reflects many basic features of human brain function and is a highly complex nonlinear dynamic learning system. .
  • Neural networks have massively parallel, distributed storage and processing, self-organizing, adaptive, and self-learning capabilities, and are particularly well-suited for handling inaccurate and ambiguous information processing problems that require many factors and conditions to be considered simultaneously.
  • the trained neural network model can be used to predict the probability of the customer's insurance or the probability of the type of insurance.
  • a plurality of predetermined models trained based on a preset number of customer information samples such as a decision tree model, a linear regression model, a logistic regression model, a neural network model, and the like, are combined into a composite model according to a predetermined combination rule. For example, according to the characteristics and advantages of different models, and the characteristics of the customer information to be analyzed, the corresponding weights are set for different models. If the relationship between the dependent variable and the target variable is a simple linear relationship, then the composite model can be combined. The weight of the linear regression model is increased to improve the speed and efficiency of the prediction; if there are many dependent variables and complex analysis is needed, the weight of the neural network model can be improved when the composite model is combined to improve the accuracy of the prediction.
  • the customer information to be analyzed may be input into the composite model after receiving the customer information to be analyzed, so as to integrate multiple models.
  • Decision tree model, linear regression model, logistic regression model, neural network model and other advantages and different judgment angles to analyze and predict the customer information to be analyzed, so that the output is more accurate Analysis and prediction results.
  • a plurality of predetermined models are trained through a preset number of customer information samples, and the plurality of trained models are combined into a composite model.
  • the combined composite model is used to treat the Analyze customer information for analysis. Because it combines multiple models to analyze and predict using the combined composite model, it can combine the advantages of different models, and the accuracy of the prediction results is effectively improved compared with the single model prediction.
  • the predetermined number of models is N, N is a natural number greater than 2, the i-th predetermined model is denoted as Fi, and i is a positive integer less than or equal to N, Combine the trained models into a composite model according to predetermined combination rules:
  • the composite model (1/N) * F1 + (1/N) * F2 + ... + (1/N) * FN.
  • the predetermined number of models is N, N is a natural number greater than 2, the i-th predetermined model is denoted as Fi, and i is a positive integer less than or equal to N, Combine the trained models into a composite model according to predetermined combination rules:
  • the N-th roots are obtained for the results analyzed by the respective models, and combined to form a composite model. Since the prediction results of each model have a great influence on the prediction results of the final combined model, the role of each model in the combined model can be highlighted, and the analysis and prediction of each model in the combined model can be maximized.
  • the analysis results of various aspects determine the prediction results of the final combined model and improve the accuracy of the prediction.
  • the predetermined number of models is N, N is a natural number greater than 2, the i-th predetermined model is denoted as Fi, and i is a positive integer less than or equal to N, Combine the trained models into a composite model according to predetermined combination rules:
  • the composite model N / (1/F1 + 1 / F2 + ... + 1 / FN).
  • the three models of logistic regression, decision tree, and neural network may be combined to form a combined model, and the performance of the combined model on the verification set is analyzed.
  • Table 2 lists the degree of elevation of the logistic regression, combined 1 to 6 models predicted on the validation set.
  • the depths 1-99 in Table 2 represent the corresponding logistic regression, the combination 1 to 6 models in the prediction set on the verification set, the score is 1%-99% of the sample, the experimental data in Table 2 shows that the depth is 1
  • the average performance of the combined model is 4.5% higher than that of the logistic regression; at a depth of 5, the average performance of the combined model is 5.3% higher than that of the logistic regression; at a depth of 10, the average performance of the combined model is higher than that of the logistic regression. 1.9%.
  • the prediction effect of the model is better than that of the single logistic regression model. That is, the combined model is more effective than the single model prediction, which improves the accuracy of the prediction results.
  • step A further includes the following steps: C. In training a predetermined model, after each training, input each customer information sample into the currently trained model to determine The model analyzes the wrong customer information sample;
  • the proportion of the sampled customer information sample to the total number of customer information samples is greater than or equal to the preset threshold, the customer information sample added to the total customer information sample and the model analysis error is added according to the preset proportional increase.
  • the accuracy of the current model is analyzed after each training. Judging, if each customer information sample can be separately input into the currently trained model for analysis and prediction, if the model analyzes the wrong number of customer information samples, the proportion of the total number of customer information samples is less than a preset threshold (for example, 5%). , indicating that the accuracy of the current model is high, then the training is ended, and the current model is used as the trained model.
  • a preset threshold for example, 5%
  • the model analyzes the wrong number of customer information samples in the proportion of all customer information samples is greater than or equal to the preset threshold (for example, 5%), it indicates that the accuracy of the current model is low, and the increase rate according to the preset ratio (for example, 1%) in the total customer information sample, the proportion of the customer information sample belonging to the same type of model analysis error is added, and the proportion of the customer information sample of the same type is reduced according to the preset ratio (for example, 1%) in the total customer information sample.
  • the preset threshold for example, 5%
  • the reduction model analyzes the proportion of the correct customer information sample (for example, if the total customer information sample, the model analyzes the correct customer information sample percentage is 80%, after reduction, it is 79%), and the adjusted customer information On the basis of the sample, the model will continue to be learned and trained until the accuracy of the model meets the requirements.
  • the accuracy of the current model can be analyzed and judged after each training in the process of training a predetermined model, only the accuracy of the model meets the requirements to end the training, and each pre-use for the combination is guaranteed. Determine the high accuracy of the model. Moreover, when it is judged that the accuracy of the current model is not up to the requirement, the proportion of the customer information sample in the total customer information sample that is incorrectly analyzed by the model belongs to the same type of customer information sample, so that the customer is easily analyzed for the model. The information sample type is used to focus on learning and training the model, which is more targeted and improves the efficiency and speed of model training.
  • a specific description is made on whether a customer has a claim.
  • the usual predictive model is such that there is a business goal, such as predicting whether the customer will make a claim within the next six months, and the probability of claim is How big.
  • Predictors are the selection of data indicators that affect the target scalar, including: gender, age information, information on insurance products held (such as insurance-type insurance products, income-based insurance products, short-term insurance products, life-long insurance products, etc.), insurance Behavioral habits (for example, if a customer holds more than a predetermined percentage of the insurance products is a guaranteed product, the insurance behavior on behalf of the customer is a preference-protected insurance product), historical claims letter Interest and so on.
  • the decision tree model can be established based on the customer's predictor variables and target variables.
  • the decision tree model gives the probability of each customer's claim, if the threshold is set to 0.5, when the decision is made.
  • the tree model predicts that the probability of customer claims is greater than 0.5, it is considered that the customer will make claims within the next six months.
  • the decision tree model predicts that the probability of customer claims is less than 0.5, the customer is considered to be in the next six months. No claims will occur within.
  • the tree model analyzes the wrong customer information sample.
  • the model analysis program may also be divided into one or more modules, and one or more modules are stored in the memory 11 and are composed of one or more processors (this embodiment is The processor 12) is executed to complete the invention, and the term "module" as used herein refers to a series of computer program instructions that are capable of performing a particular function.
  • FIG. 4 it is a functional block diagram of a model analysis program in an embodiment of the model analysis device of the present invention.
  • the model analysis program may be divided into a training module 10 and an analysis module 20, where The functions or operational steps implemented by the analysis module 10 and the analysis module 20 are similar to the above, and are not described in detail herein, by way of example, for example:
  • the training module 10 is configured to train a plurality of predetermined models based on a preset number of customer information samples
  • the analysis module 20 is configured to combine the trained multiple models into a composite model according to a predetermined combination rule, and after receiving the customer information to be analyzed, input the customer information to be analyzed into the composite model to output the analysis result.
  • embodiments of the present invention also provide a computer readable storage medium having stored thereon at least one computer readable storage instruction executable by a processing device to:
  • the foregoing embodiment method can be implemented by means of software plus a necessary general hardware platform, and of course, It is implemented by hardware, but in many cases the former is a better implementation.
  • the technical solution of the present invention which is essential or contributes to the prior art, may be embodied in the form of a software product stored in a storage medium (such as ROM/RAM, disk,
  • the optical disc includes a number of instructions for causing a terminal device (which may be a cell phone, a computer, a server, an air conditioner, or a network device, etc.) to perform the methods described in various embodiments of the present invention.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Software Systems (AREA)
  • Business, Economics & Management (AREA)
  • Data Mining & Analysis (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Artificial Intelligence (AREA)
  • Strategic Management (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Biomedical Technology (AREA)
  • Computational Linguistics (AREA)
  • Health & Medical Sciences (AREA)
  • Biophysics (AREA)
  • Economics (AREA)
  • Development Economics (AREA)
  • Human Resources & Organizations (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Marketing (AREA)
  • General Business, Economics & Management (AREA)
  • Game Theory and Decision Science (AREA)
  • Finance (AREA)
  • Accounting & Taxation (AREA)
  • Medical Informatics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Operations Research (AREA)
  • Tourism & Hospitality (AREA)
  • Quality & Reliability (AREA)
  • Geometry (AREA)
  • Computational Mathematics (AREA)
  • Mathematical Analysis (AREA)
  • Mathematical Optimization (AREA)
  • Pure & Applied Mathematics (AREA)
  • Computer Hardware Design (AREA)

Abstract

一种模型分析方法、装置、及计算机可读存储介质，该方法包括：基于预设数量的客户信息样本，训练多种预先确定的模型（S10）；将训练的多种模型按照预先确定的组合规则组合成复合模型，并在收到待分析的客户信息后，将该待分析的客户信息输入所述复合模型以输出分析结果（S20）。该方法由于是对多种模型进行组合来利用组合的复合模型进行分析、预测，能结合不同模型的优点，相比于单一模型预测，有效提高了预测结果的精准度。

Description

模型分析方法、装置、及计算机可读存储介质
本申请申明享有2017年5月5日递交的申请号为201710311997.6、名称为“模型分析方法及装置”的中国专利申请的优先权，该中国专利申请的整体内容以参考的方式结合在本申请中。
技术领域
本发明涉及计算机技术领域，尤其涉及一种模型分析方法、装置、及计算机可读存储介质。
背景技术
目前，在金融、保险等领域的数据挖掘预测项目中，业界通常采用单一模型来进行特定目标事件(例如，保险理赔事件)的预测，而众所周知，不同类型的模型对于目标事件的解释角度和侧重点会有所不同，因此，采用单一模型带来的预测结果的精准度有很大的局限性，预测错误率较高。
发明内容
本发明的主要目的在于提供一种模型分析方法、装置、及计算机可读存储介质，旨在提高预测结果的精准度。
为实现上述目的，本发明第一方面提供的一种模型分析方法，所述方法包括以下步骤：
A、基于预设数量的客户信息样本，训练多种预先确定的模型；
B、将训练的多种模型按照预先确定的组合规则组合成复合模型，并在收到待分析的客户信息后，将该待分析的客户信息输入所述复合模型以输出分析结果。
此外，为实现上述目的，本发明第二方面提供一种模型分析装置，所述模型分析装置包括处理设备、存储设备，该存储设备存储有模型分析程序，该模型分析程序被所述处理设备执行时实现如下步骤：
A、基于预设数量的客户信息样本，训练多种预先确定的模型；
B、将训练的多种模型按照预先确定的组合规则组合成复合模型，并在收到待分析的客户信息后，将该待分析的客户信息输入所述复合模型以输出分析结果。
本申请第三方面提供一种计算机可读存储介质，其上存储有至少一个可被处理设备执行以实现以下操作的计算机可读指令：
A、基于预设数量的客户信息样本，训练多种预先确定的模型；
B、将训练的多种模型按照预先确定的组合规则组合成复合模型，并在收到待分析的客户信息后，将该待分析的客户信息输入所述复合模型以输出分 析结果。
本发明提出的模型分析方法、装置、及计算机可读存储介质，通过预设数量的客户信息样本训练出多种预先确定的模型，并将训练的多种模型组合成复合模型，在收到待分析的客户信息后，利用组合的复合模型对该待分析的客户信息进行分析。由于是对多种模型进行组合来利用组合的复合模型进行分析、预测，能结合不同模型的优点，相比于单一模型预测，有效提高了预测结果的精准度。
附图说明
图1为本发明模型分析方法一实施例的流程示意图；
图2为图1中步骤S10的细化流程示意图；
图3为本发明模型分析装置一实施例的示意图。
本发明目的的实现、功能特点及优点将结合实施例，参照附图做进一步说明。
具体实施方式
为了使本发明所要解决的技术问题、技术方案及有益效果更加清楚、明白，以下结合附图和实施例，对本发明进行进一步详细说明。应当理解，此处所描述的具体实施例仅仅用以解释本发明，并不用于限定本发明。
本发明提供一种模型分析方法。
参照图1，图1为本发明模型分析方法一实施例的流程示意图。
在一实施例中，该模型分析方法包括：
步骤S10，基于预设数量的客户信息样本，训练多种预先确定的模型；
本实施例中，基于预设数量(例如，10万)的客户信息样本，训练多种预先确定的模型。例如，所述客户信息样本中的客户信息包括但不限于性别、年龄、联系方式、家庭住址、工作单位、征信记录、持有的保险产品信息、投保行为习惯、历史理赔信息等等，持有的保险产品信息包括但不限于保障型保险产品、收益型保险产品、短期型保险产品、终身型保险产品等等，投保行为习惯为一个客户在一段时间(如最近1年或3年)内持有的保险产品中持有时间最长的产品，或持有占比最大的产品，则代表该客户的投保行为习惯是该产品。例如，若一个客户持有的保险产品中超过预设比例(例如，60％)的产品是保障型产品，则代表该客户的投保行为习惯是偏好保障型保险产品。
预先确定需进行训练的模型包括但不限于决策树(Decision Tree)模型、线性回归模型、逻辑回归(Logistic regression)模型、神经网络(Neural Networks,NN)模型等。其中，决策树模型是一种简单但是广泛使用的分类器，通过训练数据构建决策树，可以高效的对未知的数据进行分类。决策树模型可读性好，具有描述性，有助于人工分析；且效率高。线性回归模型可以为 一元线性回归或多元线性回归模型，一元线性回归是一个主要影响因素作为自变量来解释因变量的变化，而在现实问题研究中，因变量的变化往往受几个重要因素的影响，此时就需要用两个或两个以上的影响因素作为自变量来解释因变量的变化，这就是多元回归亦称多重回归。当多个自变量与因变量之间是线性关系时，所进行的回归分析就是多元性回归。逻辑回归模型是当前业界比较常用的机器学习模型，用于估计某种事物的可能性，如本实施例中预测客户投保或投保类型的可能性。神经网络模型是由大量的、简单的处理单元(称为神经元)广泛地互相连接而形成的复杂网络系统，它反映了人脑功能的许多基本特征，是一个高度复杂的非线性动力学习系统。神经网络具有大规模并行、分布式存储和处理、自组织、自适应和自学能力，特别适合处理需要同时考虑许多因素和条件的、不精确和模糊的信息处理问题。例如，对神经网络模型进行训练后，可利用训练好的神经网络模型预测客户投保的概率或投保类型的概率等等。
步骤S20，将训练的多种模型按照预先确定的组合规则组合成复合模型，并在收到待分析的客户信息后，将该待分析的客户信息输入所述复合模型以输出分析结果。
将基于预设数量的客户信息样本训练出的多种预先确定的模型如决策树模型、线性回归模型、逻辑回归模型、神经网络模型等按照预先确定的组合规则组合成复合模型。例如，可根据不同模型的特点及优势，并综合待分析的客户信息的特点，为不同模型设置相应的权重，如若因变量与目标变量的关系只是简单的线性关系，则可在组合复合模型时提高线性回归模型的权重，以提高预测的速度及效率；若因变量较多，且需要进行复杂的分析，则可在组合复合模型时提高神经网络模型的权重，以提高预测的准确度。在将训练的多种模型按照预先确定的组合规则组合成复合模型之后，即可在收到待分析的客户信息后，将该待分析的客户信息输入所述复合模型，以综合多种模型如决策树模型、线性回归模型、逻辑回归模型、神经网络模型等的优势及不同判断角度来对该待分析的客户信息进行分析、预测，从而输出更加精准的分析、预测结果。
本实施例通过预设数量的客户信息样本训练出多种预先确定的模型，并将训练的多种模型组合成复合模型，在收到待分析的客户信息后，利用组合的复合模型对该待分析的客户信息进行分析。由于是对多种模型进行组合来利用组合的复合模型进行分析、预测，能结合不同模型的优点，相比于单一模型预测，有效提高了预测结果的精准度。
进一步地，在其他实施例中，所述预先确定的模型的数量为N，N为大于2的自然数，第i个预先确定的模型记为Fi，i为小于或者等于N的正整数，所述将训练的多种模型按照预先确定的组合规则组合成复合模型为：
所述复合模型＝(1/N)*F1+(1/N)*F2+……+(1/N)*FN。
本实施例中，在对训练的多种模型进行组合时，将多种模型进行平均以 组合形成复合模型，能均衡地考虑各个模型的影响，以平衡各个模型的预测结果，在各个模型的预测结果相差不大的情况下，得到最合理的预测结果。
进一步地，在其他实施例中，所述预先确定的模型的数量为N，N为大于2的自然数，第i个预先确定的模型记为Fi，i为小于或者等于N的正整数，所述将训练的多种模型按照预先确定的组合规则组合成复合模型为：
所述复合模型＝POWER(F1,1/N)*POWER(F2,1/N)*……*POWER(FN,1/N)，其中，POWER(Fi,1/N)是对Fi模型分析出的结果求N次方根。
本实施例中，在对训练的多种模型进行组合时，对各个模型分析出的结果求N次方根，并进行组合，以形成复合模型。由于每一模型的预测结果对最终组合模型的预测结果影响很大，能突出每一模型在组合模型中的作用，能最大化的发挥出每一模型在组合模型中的分析、预测作用，基于各个方面的分析结果来决定最终组合模型的预测结果，提高预测的精准度。
进一步地，在其他实施例中，所述预先确定的模型的数量为N，N为大于2的自然数，第i个预先确定的模型记为Fi，i为小于或者等于N的正整数，所述将训练的多种模型按照预先确定的组合规则组合成复合模型为：
所述复合模型＝N/(1/F1+1/F2+……+1/FN)。
本实施例中，在对训练的多种模型进行组合时，所述复合模型＝N/(1/F1+1/F2+……+1/FN)，即所述复合模型＝NF1F2*…FN/(F2F3*…FN+F1F3*…FN+……F1F2*…F(N-1))，在考虑每一模型的预测结果对最终组合模型的预测结果的影响作用基础上，还综合考虑对多个模型进行不同组合后的预测结果对最终组合模型的预测结果的影响，以得到最合理的预测结果，进一步提高预测的精准度。
进一步地，在一种具体实施方式中，可对逻辑回归、决策树、神经网络三种模型进行组合形成组合模型，并分析该组合模型在验证集上预测精准度的表现。针对逻辑回归、决策树、神经网络三种模型的组合，提供如下六种组合方式：组合一:等权平均法(如上述实施例中的所述复合模型＝(1/N)*F1+(1/N)*F2+……+(1/N)*FN)；组合二：三个模型概率加权平均法；组合三：加权几何组合平均模型；组合四：加权调和几何平均组合模型(如上述实施例中的所述复合模型＝POWER(F1,1/N)*POWER(F2,1/N)*……*POWER(FN,1/N))；组合五：预测误差平方和倒数法；组合六：简单加权平均方法。如下表1所示，表1中列出了逻辑回归、组合1～组合6模型在验证集上预测的提升度。表1中的深度1-99分别代表对应逻辑回归、组合1～组合6模型在验证集上预测结果中评分处于1％-99％的样本，通过表1中的实验数据显示，在深度为1时，组合模型的平均表现要比逻辑回归提升4.5％；在深度为5时，组合模型的平均表现要比逻辑回归提升5.3％；在深度为10时，组合模型的平均表现要比逻辑回归提升1.9％。综上所述，通过对逻辑回归、决策树、神经网络三个模型进行组合，其模型的预测效果要优于单一逻辑回归模型的表现。即组合模型相比于单一模型预测，有效提高了预测结果的精准 度。
表1
深度 组合1 组合2 组合3 组合4 组合5 组合6 逻辑回归
1 5.62 5.61 5.65 5.63 5.62 5.55 5.37
5 4.148 4.142 4.126 4.078 4.148 4.104 3.916
10 3.41 3.409 3.414 3.411 3.409 3.43 3.349
15 3.003 3.002 2.997 2.987 3.002 2.997 2.953
20 2.697 2.7 2.702 2.686 2.697 2.69 2.642
25 2.466 2.462 2.460 2.458 2.466 2.450 2.403
30 2.26 2.257 2.263 2.258 2.260 2.258 2.217
35 2.109 2.108 2.106 2.098 2.109 2.103 2.060
40 1.969 1.971 1.970 1.962 1.969 1.960 1.928
45 1.842 1.843 1.839 1.830 1.842 1.832 1.807
50 1.725 1.725 1.727 1.72 1.725 1.722 1.701
55 1.627 1.625 1.629 1.627 1.627 1.624 1.606
60 1.540 1.538 1.541 1.544 1.540 1.536 1.523
65 1.459 1.457 1.459 1.461 1.459 1.454 1.449
70 1.384 1.382 1.385 1.384 1.384 1.381 1.374
75 1.310 1.309 1.311 1.312 1.310 1.308 1.306
80 1.24 1.240 1.241 1.241 1.24 1.238 1.238
85 1.173 1.173 1.173 1.172 1.173 1.172 1.172
90 1.110 1.110 1.109 1.108 1.110 1.110 1.110
95 1.052 1.052 1.051 1.050 1.052 1.052 1.052
99 1.010 1.010 1.009 1.008 1.010 1.010 1.010
进一步地，如图2所示，上述步骤S10可以包括：
步骤S101，在训练一种预先确定的模型过程中，每训练一次后，将各个客户信息样本分别输入当前训练的模型中以确定出模型分析错误的客户信息样本；
步骤S102，计算出模型分析错误的客户信息样本数量占所有客户信息样本数量的比例是否小于预设阈值；
步骤S103，若模型分析错误的客户信息样本数量占所有客户信息样本数量的比例小于预设阈值，则该预先确定的模型训练结束；
步骤S104，若模型分析错误的客户信息样本数量占所有客户信息样本数量的比例大于或者等于预设阈值，则按照预设的比例增加幅度在总客户信息样本中增加与模型分析错误的客户信息样本属于同一类型的客户信息样本的比重，并按照预设的比例减少幅度在总客户信息样本中减少模型分析正确的客户信息样本的比重，并返回执行上述步骤S101。
本实施例中，在基于预设数量的客户信息样本训练多种预先确定的模型 的过程中，对一种预先确定的模型进行训练时，每训练一次后均对当前该模型的准确率进行分析判断，如可将各个客户信息样本分别输入当前训练的该模型中进行分析、预测，若该模型分析错误的客户信息样本数量占所有客户信息样本数量的比例小于预设阈值(例如，5％)，则说明当前该模型的准确率较高，则结束训练，以当前该模型作为训练好的模型。若该模型分析错误的客户信息样本数量占所有客户信息样本数量的比例大于或者等于预设阈值(例如，5％)，则说明当前该模型的准确率较低，则按照预设的比例增加幅度(例如，1％)在总客户信息样本中增加与模型分析错误的客户信息样本属于同一类型的客户信息样本的比重，并按照预设的比例减少幅度(例如，1％)在总客户信息样本中减少模型分析正确的客户信息样本的比重(例如，若总客户信息样本中，模型分析正确的客户信息样本占比为80％，减少后，则为79％)，并在调整后的客户信息样本基础上继续对该模型进行学习、训练，直至该模型的准确率达到要求。
由于能在训练一种预先确定的模型过程中每训练一次后均对当前该模型的准确率进行分析判断，只有该模型的准确率达到要求才结束训练，保证了用于组合的每一种预先确定的模型的高准确率。而且，在判断当前该模型的准确率达不到要求时，增加总客户信息样本中与该模型分析错误的客户信息样本属于同一类型的客户信息样本的比重，以针对该模型容易分析错误的客户信息样本类型来对该模型进行重点学习、训练，针对性更强，提高了模型训练效率及速度。
例如，在一种实施方式中，以客户是否发生理赔为例进行具体说明：通常的预测模型是这样的，有一个业务目标，如预测客户是否在未来六个月内发生理赔，理赔的概率有多大。定义目标变量为Y：即客户是否发生理赔，Y是二值变量，Y＝1为理赔，Y＝0表示不发生理赔。预测变量是影响目标标量的数据指标的选取，包括：性别、年龄信息、持有的保险产品信息(如保障型保险产品、收益型保险产品、短期型保险产品、终身型保险产品等)、投保行为习惯(例如，若一个客户持有的保险产品中超过预设比例的产品是保障型产品，代表该客户的投保行为习惯是偏好保障型保险产品)、历史理赔信息等。可以根据客户的预测变量和目标变量建立决策树模型，一旦建立决策树模型，给定某个客户信息，该决策树模型会给出每个客户理赔的概率，假如设置阈值为0.5，当该决策树模型预测出客户理赔的概率大于0.5时，则认为客户在接下来六个月内会发生理赔，当该决策树模型预测出客户理赔的概率小于0.5时，则认为客户在接下来六个月内不发生理赔。将该决策树模型预测出客户是否理赔的变量设为hat(Y)，并将hat(Y)与客户真实的理赔情况Y进行对比。如果hat(Y)＝Y，则认为这个样本被正确学习，如果hat(Y)不等于Y，则认为这个样本被错误学习，这样就可以确定出所有错误学习的客户信息样本，也即该决策树模型分析错误的客户信息样本。
本发明进一步提供一种模型分析装置。
参照图3，图3为本发明模型分析装置一实施例的示意图。
在一实施例中，该模型分析装置2中，可以是PC(Personal Computer，个人电脑)，也可以是智能手机、平板电脑、电子书阅读器、便携计算机等终端设备。
该模型分析装置2包括存储设备11、处理设备12，通信总线13，以及网络接口14。
其中，存储设备11至少包括一种类型的可读存储介质，所述可读存储介质包括闪存、硬盘、多媒体卡、卡型存储器(例如，SD或DX存储器等)、磁性存储器、磁盘、光盘等。存储设备11在一些实施例中可以是基于社交网络的用户关键词提取装置的内部存储单元，例如该模型分析装置的硬盘。存储设备11在另一些实施例中也可以是模型分析装置的外部存储设备，例如模型分析装置上配备的插接式硬盘，智能存储卡(Smart Media Card,SMC)，安全数字(Secure Digital,SD)卡，闪存卡(Flash Card)等。进一步地，存储设备11还可以既包括模型分析装置的内部存储单元也包括外部存储设备。存储设备11不仅可以用于存储安装于模型分析装置的应用软件及各类数据，例如模型分析代码等，还可以用于暂时地存储已经输出或者将要输出的数据。
处理设备12在一些实施例中可以是一中央处理器(Central Processing Unit,CPU)、控制器、微控制器、微处理器或其他数据处理芯片，用于运行存储器11中存储的程序代码或处理数据，例如执行模型分析程序等。
网络接口13可选的可以包括标准的有线接口、无线接口(如WI-FI接口)，通常用于在该装置与其他电子设备之间建立通信连接。
通信总线14用于实现这些组件之间的连接通信。
图3仅示出了具有组件11-14的模型分析装置，但是应理解的是，并不要求实施所有示出的组件，可以替代的实施更多或者更少的组件。
可选地，该装置还可以包括用户接口，用户接口可以包括显示设备(Display)、输入单元比如键盘(Keyboard)，可选的用户接口还可以包括标准的有线接口、无线接口。可选地，在一些实施例中，显示设备可以是LED显示器、液晶显示器、触控式液晶显示器以及OLED(Organic Light-Emitting Diode，有机发光二极管)触摸器等。其中，显示设备也可以适当的称为显示屏或显示单元，用于显示模型分析装置中处理的信息以及用于显示可视化的用户界面。
在图3所示的装置实施例中，存储设备11中存储有模型分析程序；处理设备12执行存储设备11中存储的模型分析程序时实现如下步骤：
A、基于预设数量的客户信息样本，训练多种预先确定的模型；
本实施例中，基于预设数量(例如，10万)的客户信息样本，训练多种预先确定的模型。例如，所述客户信息样本中的客户信息包括但不限于性别、年龄、联系方式、家庭住址、工作单位、征信记录、持有的保险产品信息、投保行为习惯、历史理赔信息等等，持有的保险产品信息包括但不限于保障 型保险产品、收益型保险产品、短期型保险产品、终身型保险产品等等，投保行为习惯为一个客户在一段时间(如最近1年或3年)内持有的保险产品中持有时间最长的产品，或持有占比最大的产品，则代表该客户的投保行为习惯是该产品。例如，若一个客户持有的保险产品中超过预设比例(例如，60％)的产品是保障型产品，则代表该客户的投保行为习惯是偏好保障型保险产品。
预先确定需进行训练的模型包括但不限于决策树模型(Decision Tree)、线性回归模型、逻辑回归(Logistic regression)模型、神经网络(Neural Networks,NN)模型等。其中，决策树模型是一种简单但是广泛使用的分类器，通过训练数据构建决策树，可以高效的对未知的数据进行分类。决策树模型可读性好，具有描述性，有助于人工分析；且效率高。线性回归模型可以为一元线性回归或多元线性回归模型，一元线性回归是一个主要影响因素作为自变量来解释因变量的变化，而在现实问题研究中，因变量的变化往往受几个重要因素的影响，此时就需要用两个或两个以上的影响因素作为自变量来解释因变量的变化，这就是多元回归亦称多重回归。当多个自变量与因变量之间是线性关系时，所进行的回归分析就是多元性回归。逻辑回归模型是当前业界比较常用的机器学习模型，用于估计某种事物的可能性，如本实施例中预测客户投保或投保类型的可能性。神经网络模型是由大量的、简单的处理单元(称为神经元)广泛地互相连接而形成的复杂网络系统，它反映了人脑功能的许多基本特征，是一个高度复杂的非线性动力学习系统。神经网络具有大规模并行、分布式存储和处理、自组织、自适应和自学能力，特别适合处理需要同时考虑许多因素和条件的、不精确和模糊的信息处理问题。例如，对神经网络模型进行训练后，可利用训练好的神经网络模型预测客户投保的概率或投保类型的概率等等。
B、将训练的多种模型按照预先确定的组合规则组合成复合模型，并在收到待分析的客户信息后，将该待分析的客户信息输入所述复合模型以输出分析结果。
将基于预设数量的客户信息样本训练出的多种预先确定的模型如决策树模型、线性回归模型、逻辑回归模型、神经网络模型等按照预先确定的组合规则组合成复合模型。例如，可根据不同模型的特点及优势，并综合待分析的客户信息的特点，为不同模型设置相应的权重，如若因变量与目标变量的关系只是简单的线性关系，则可在组合复合模型时提高线性回归模型的权重，以提高预测的速度及效率；若因变量较多，且需要进行复杂的分析，则可在组合复合模型时提高神经网络模型的权重，以提高预测的准确度。在将训练的多种模型按照预先确定的组合规则组合成复合模型之后，即可在收到待分析的客户信息后，将该待分析的客户信息输入所述复合模型，以综合多种模型如决策树模型、线性回归模型、逻辑回归模型、神经网络模型等的优势及不同判断角度来对该待分析的客户信息进行分析、预测，从而输出更加精准 的分析、预测结果。
本实施例通过预设数量的客户信息样本训练出多种预先确定的模型，并将训练的多种模型组合成复合模型，在收到待分析的客户信息后，利用组合的复合模型对该待分析的客户信息进行分析。由于是对多种模型进行组合来利用组合的复合模型进行分析、预测，能结合不同模型的优点，相比于单一模型预测，有效提高了预测结果的精准度。
进一步地，在其他实施例中，所述预先确定的模型的数量为N，N为大于2的自然数，第i个预先确定的模型记为Fi，i为小于或者等于N的正整数，所述将训练的多种模型按照预先确定的组合规则组合成复合模型为：
所述复合模型＝(1/N)*F1+(1/N)*F2+……+(1/N)*FN。
本实施例中，在对训练的多种模型进行组合时，将多种模型进行平均以组合形成复合模型，能均衡地考虑各个模型的影响，以平衡各个模型的预测结果，在各个模型的预测结果相差不大的情况下，得到最合理的预测结果。
进一步地，在其他实施例中，所述预先确定的模型的数量为N，N为大于2的自然数，第i个预先确定的模型记为Fi，i为小于或者等于N的正整数，所述将训练的多种模型按照预先确定的组合规则组合成复合模型为：
所述复合模型＝POWER(F1,1/N)*POWER(F2,1/N)*……*POWER(FN,1/N)，其中，POWER(Fi,1/N)是对Fi模型分析出的结果求N次方根。
本实施例中，在对训练的多种模型进行组合时，对各个模型分析出的结果求N次方根，并进行组合，以形成复合模型。由于每一模型的预测结果对最终组合模型的预测结果影响很大，能突出每一模型在组合模型中的作用，能最大化的发挥出每一模型在组合模型中的分析、预测作用，基于各个方面的分析结果来决定最终组合模型的预测结果，提高预测的精准度。
进一步地，在其他实施例中，所述预先确定的模型的数量为N，N为大于2的自然数，第i个预先确定的模型记为Fi，i为小于或者等于N的正整数，所述将训练的多种模型按照预先确定的组合规则组合成复合模型为：
所述复合模型＝N/(1/F1+1/F2+……+1/FN)。
本实施例中，在对训练的多种模型进行组合时，所述复合模型＝N/(1/F1+1/F2+……+1/FN)，即所述复合模型＝NF1F2*…FN/(F2F3*…FN+F1F3*…FN+……F1F2*…F(N-1))，在考虑每一模型的预测结果对最终组合模型的预测结果的影响作用基础上，还综合考虑对多个模型进行不同组合后的预测结果对最终组合模型的预测结果的影响，以得到最合理的预测结果，进一步提高预测的精准度。
进一步地，在一种具体实施方式中，可对逻辑回归、决策树、神经网络三种模型进行组合形成组合模型，并分析该组合模型在验证集上预测精准度的表现。针对逻辑回归、决策树、神经网络三种模型的组合，提供如下六种组合方式：组合一:等权平均法(如上述实施例中的所述复合模型＝(1/N)*F1+(1/N)*F2+……+(1/N)*FN)；组合二：三个模型概率加权平均法；组合 三：加权几何组合平均模型；组合四：加权调和几何平均组合模型(如上述实施例中的所述复合模型＝POWER(F1,1/N)*POWER(F2,1/N)*……*POWER(FN,1/N))；组合五：预测误差平方和倒数法；组合六：简单加权平均方法。如下表2所示，表2中列出了逻辑回归、组合1～组合6模型在验证集上预测的提升度。表2中的深度1-99分别代表对应逻辑回归、组合1～组合6模型在验证集上预测结果中评分处于1％-99％的样本，通过表2中的实验数据显示，在深度为1时，组合模型的平均表现要比逻辑回归提升4.5％；在深度为5时，组合模型的平均表现要比逻辑回归提升5.3％；在深度为10时，组合模型的平均表现要比逻辑回归提升1.9％。综上所述，通过对逻辑回归、决策树、神经网络三个模型进行组合，其模型的预测效果要优于单一逻辑回归模型的表现。即组合模型相比于单一模型预测，有效提高了预测结果的精准度。
表2
深度 组合1 组合2 组合3 组合4 组合5 组合6 逻辑回归
1 5.62 5.61 5.65 5.63 5.62 5.55 5.37
5 4.148 4.142 4.126 4.078 4.148 4.104 3.916
10 3.41 3.409 3.414 3.411 3.409 3.43 3.349
15 3.003 3.002 2.997 2.987 3.002 2.997 2.953
20 2.697 2.7 2.702 2.686 2.697 2.69 2.642
25 2.466 2.462 2.460 2.458 2.466 2.450 2.403
30 2.26 2.257 2.263 2.258 2.260 2.258 2.217
35 2.109 2.108 2.106 2.098 2.109 2.103 2.060
40 1.969 1.971 1.970 1.962 1.969 1.960 1.928
45 1.842 1.843 1.839 1.830 1.842 1.832 1.807
50 1.725 1.725 1.727 1.72 1.725 1.722 1.701
55 1.627 1.625 1.629 1.627 1.627 1.624 1.606
60 1.540 1.538 1.541 1.544 1.540 1.536 1.523
65 1.459 1.457 1.459 1.461 1.459 1.454 1.449
70 1.384 1.382 1.385 1.384 1.384 1.381 1.374
75 1.310 1.309 1.311 1.312 1.310 1.308 1.306
80 1.24 1.240 1.241 1.241 1.24 1.238 1.238
85 1.173 1.173 1.173 1.172 1.173 1.172 1.172
90 1.110 1.110 1.109 1.108 1.110 1.110 1.110
95 1.052 1.052 1.051 1.050 1.052 1.052 1.052
99 1.010 1.010 1.009 1.008 1.010 1.010 1.010
进一步地，在一实施例中，上述步骤A还包括如下步骤：C、在训练一种预先确定的模型过程中，每训练一次后，将各个客户信息样本分别输入当前训练的模型中以确定出模型分析错误的客户信息样本；
D、计算出模型分析错误的客户信息样本数量占所有客户信息样本数量的比例是否小于预设阈值；
E、若模型分析错误的客户信息样本数量占所有客户信息样本数量的比例小于预设阈值，则结束该预先确定的模型训练；
F、若模型分析错误的客户信息样本数量占所有客户信息样本数量的比例大于或者等于预设阈值，则按照预设的比例增加幅度在总客户信息样本中增加与模型分析错误的客户信息样本属于同一类型的客户信息样本的比重，并按照预设的比例减少幅度在总客户信息样本中减少模型分析正确的客户信息样本的比重，并重新执行上述步骤C、D、E、F。
本实施例中，在基于预设数量的客户信息样本训练多种预先确定的模型的过程中，对一种预先确定的模型进行训练时，每训练一次后均对当前该模型的准确率进行分析判断，如可将各个客户信息样本分别输入当前训练的该模型中进行分析、预测，若该模型分析错误的客户信息样本数量占所有客户信息样本数量的比例小于预设阈值(例如，5％)，则说明当前该模型的准确率较高，则结束训练，以当前该模型作为训练好的模型。若该模型分析错误的客户信息样本数量占所有客户信息样本数量的比例大于或者等于预设阈值(例如，5％)，则说明当前该模型的准确率较低，则按照预设的比例增加幅度(例如，1％)在总客户信息样本中增加与模型分析错误的客户信息样本属于同一类型的客户信息样本的比重，并按照预设的比例减少幅度(例如，1％)在总客户信息样本中减少模型分析正确的客户信息样本的比重(例如，若总客户信息样本中，模型分析正确的客户信息样本占比为80％，减少后，则为79％)，并在调整后的客户信息样本基础上继续对该模型进行学习、训练，直至该模型的准确率达到要求。
由于能在训练一种预先确定的模型过程中每训练一次后均对当前该模型的准确率进行分析判断，只有该模型的准确率达到要求才结束训练，保证了用于组合的每一种预先确定的模型的高准确率。而且，在判断当前该模型的准确率达不到要求时，增加总客户信息样本中与该模型分析错误的客户信息样本属于同一类型的客户信息样本的比重，以针对该模型容易分析错误的客户信息样本类型来对该模型进行重点学习、训练，针对性更强，提高了模型训练效率及速度。
例如，在一种实施方式中，以客户是否发生理赔为例进行具体说明：通常的预测模型是这样的，有一个业务目标，如预测客户是否在未来六个月内发生理赔，理赔的概率有多大。定义目标变量为Y：即客户是否发生理赔，Y是二值变量，Y＝1为理赔，Y＝0表示不发生理赔。预测变量是影响目标标量的数据指标的选取，包括：性别、年龄信息、持有的保险产品信息(如保障型保险产品、收益型保险产品、短期型保险产品、终身型保险产品等)、投保行为习惯(例如，若一个客户持有的保险产品中超过预设比例的产品是保障型产品，代表该客户的投保行为习惯是偏好保障型保险产品)、历史理赔信 息等。可以根据客户的预测变量和目标变量建立决策树模型，一旦建立决策树模型，给定某个客户信息，该决策树模型会给出每个客户理赔的概率，假如设置阈值为0.5，当该决策树模型预测出客户理赔的概率大于0.5时，则认为客户在接下来六个月内会发生理赔，当该决策树模型预测出客户理赔的概率小于0.5时，则认为客户在接下来六个月内不发生理赔。将该决策树模型预测出客户是否理赔的变量设为hat(Y)，并将hat(Y)与客户真实的理赔情况Y进行对比。如果hat(Y)＝Y，则认为这个样本被正确学习，如果hat(Y)不等于Y，则认为这个样本被错误学习，这样就可以确定出所有错误学习的客户信息样本，也即该决策树模型分析错误的客户信息样本。
可选地，在其他的实施例中，模型分析程序还可以被分割为一个或者多个模块，一个或者多个模块被存储于存储器11中，并由一个或多个处理器(本实施例为处理器12)所执行，以完成本发明，本发明所称的模块是指能够完成特定功能的一系列计算机程序指令段。例如，参照图4所示，为本发明模型分析装置一实施例中的模型分析程序的功能模块示意图，该实施例中，模型分析程序可以被分割为训练模块10、以及分析模块20，其中，分析模块10与分析模块20所实现的功能或操作步骤均与上文类似，此处不再详述，示例性地，例如其中：
训练模块10用于基于预设数量的客户信息样本，训练多种预先确定的模型；
分析模块20用于将训练的多种模型按照预先确定的组合规则组合成复合模型，并在收到待分析的客户信息后，将该待分析的客户信息输入所述复合模型以输出分析结果。
此外，本发明实施例还提出一种计算机可读存储介质，其上存储有至少一个可被处理设备执行以实现以下操作的计算机可读存储指令：
A、基于预设数量的客户信息样本，训练多种预先确定的模型；
B、将训练的多种模型按照预先确定的组合规则组合成复合模型，并在收到待分析的客户信息后，将该待分析的客户信息输入所述复合模型以输出分析结果。
需要说明的是，本发明计算机可读存储介质具体实施方式与上述模型分析装置和模型分析方法各实施例基本相同，在此不作累述。
需要说明的是，在本文中，术语“包括”、“包含”或者其任何其他变体意在涵盖非排他性的包含，从而使得包括一系列要素的过程、方法、物品或者装置不仅包括那些要素，而且还包括没有明确列出的其他要素，或者是还包括为这种过程、方法、物品或者装置所固有的要素。在没有更多限制的情况下，由语句“包括一个……”限定的要素，并不排除在包括该要素的过程、方法、物品或者装置中还存在另外的相同要素。
通过以上的实施方式的描述，本领域的技术人员可以清楚地了解到上述实施例方法可借助软件加必需的通用硬件平台的方式来实现，当然也可以通 过硬件来实现，但很多情况下前者是更佳的实施方式。基于这样的理解，本发明的技术方案本质上或者说对现有技术做出贡献的部分可以以软件产品的形式体现出来，该计算机软件产品存储在一个存储介质(如ROM/RAM、磁碟、光盘)中，包括若干指令用以使得一台终端设备(可以是手机，计算机，服务器，空调器，或者网络设备等)执行本发明各个实施例所述的方法。
以上参照附图说明了本发明的优选实施例，并非因此局限本发明的权利范围。上述本发明实施例序号仅仅为了描述，不代表实施例的优劣。另外，虽然在流程图中示出了逻辑顺序，但是在某些情况下，可以以不同于此处的顺序执行所示出或描述的步骤。
本领域技术人员不脱离本发明的范围和实质，可以有多种变型方案实现本发明，比如作为一个实施例的特征可用于另一实施例而得到又一实施例。凡在运用本发明的技术构思之内所作的任何修改、等同替换和改进，均应在本发明的权利范围之内。

Claims (20)

  1. 一种模型分析方法，其特征在于，所述方法包括以下步骤：
    A、基于预设数量的客户信息样本，训练多种预先确定的模型；
    B、将训练的多种模型按照预先确定的组合规则组合成复合模型，并在收到待分析的客户信息后，将该待分析的客户信息输入所述复合模型以输出分析结果。
  2. 如权利要求1所述的模型分析方法，其特征在于，所述基于预设数量的客户信息样本，训练多种预先确定的模型的步骤包括：
    C、在训练一种预先确定的模型过程中，每训练一次后，将各个客户信息样本分别输入当前训练的模型中以确定出模型分析错误的客户信息样本；
    D、计算出模型分析错误的客户信息样本数量占所有客户信息样本数量的比例是否小于预设阈值；
    E、若模型分析错误的客户信息样本数量占所有客户信息样本数量的比例小于预设阈值，则该预先确定的模型训练结束；
    F、若模型分析错误的客户信息样本数量占所有客户信息样本数量的比例大于或者等于预设阈值，则按照预设的比例增加幅度在总客户信息样本中增加与模型分析错误的客户信息样本属于同一类型的客户信息样本的比重，并按照预设的比例减少幅度在总客户信息样本中减少模型分析正确的客户信息样本的比重，并重新执行上述步骤C、D、E、F。
  3. 如权利要求1所述的模型分析方法，其特征在于，所述预先确定的模型的数量为N，N为大于2的自然数，第i个预先确定的模型记为Fi，i为小于或者等于N的正整数，所述将训练的多种模型按照预先确定的组合规则组合成复合模型为：
    所述复合模型＝(1/N)*F1+(1/N)*F2+……+(1/N)*FN。
  4. 如权利要求3所述的模型分析方法，其特征在于，所述基于预设数量的客户信息样本，训练多种预先确定的模型的步骤包括：
    C、在训练一种预先确定的模型过程中，每训练一次后，将各个客户信息样本分别输入当前训练的模型中以确定出模型分析错误的客户信息样本；
    D、计算出模型分析错误的客户信息样本数量占所有客户信息样本数量的比例是否小于预设阈值；
    E、若模型分析错误的客户信息样本数量占所有客户信息样本数量的比例小于预设阈值，则该预先确定的模型训练结束；
    F、若模型分析错误的客户信息样本数量占所有客户信息样本数量的比例大于或者等于预设阈值，则按照预设的比例增加幅度在总客户信息样本中增加与模型分析错误的客户信息样本属于同一类型的客户信息样本的比重，并按照预设的比例减少幅度在总客户信息样本中减少模型分析正确的客户信息样本的比重，并重新执行上述步骤C、D、E、F。
  5. 如权利要求1所述的模型分析方法，其特征在于，所述预先确定的模型的数量为N，N为大于2的自然数，第i个预先确定的模型记为Fi，i为小于或者等于N的正整数，所述将训练的多种模型按照预先确定的组合规则组合成复合模型为：
    所述复合模型＝POWER(F1,1/N)*POWER(F2,1/N)*……*POWER(FN,1/N)，其中，POWER(Fi,1/N)是对Fi模型分析出的结果求N次方根。
  6. 如权利要求5所述的模型分析方法，其特征在于，所述基于预设数量的客户信息样本，训练多种预先确定的模型的步骤包括：
    C、在训练一种预先确定的模型过程中，每训练一次后，将各个客户信息样本分别输入当前训练的模型中以确定出模型分析错误的客户信息样本；
    D、计算出模型分析错误的客户信息样本数量占所有客户信息样本数量的比例是否小于预设阈值；
    E、若模型分析错误的客户信息样本数量占所有客户信息样本数量的比例小于预设阈值，则该预先确定的模型训练结束；
    F、若模型分析错误的客户信息样本数量占所有客户信息样本数量的比例大于或者等于预设阈值，则按照预设的比例增加幅度在总客户信息样本中增加与模型分析错误的客户信息样本属于同一类型的客户信息样本的比重，并按照预设的比例减少幅度在总客户信息样本中减少模型分析正确的客户信息样本的比重，并重新执行上述步骤C、D、E、F。
  7. 如权利要求1所述的模型分析方法，其特征在于，所述预先确定的模型的数量为N，N为大于2的自然数，第i个预先确定的模型记为Fi，i为小于或者等于N的正整数，所述将训练的多种模型按照预先确定的组合规则组合成复合模型为：
    所述复合模型＝N/(1/F1+1/F2+……+1/FN)。
  8. 如权利要求7所述的模型分析方法，其特征在于，所述基于预设数量的客户信息样本，训练多种预先确定的模型的步骤包括：
    C、在训练一种预先确定的模型过程中，每训练一次后，将各个客户信息样本分别输入当前训练的模型中以确定出模型分析错误的客户信息样本；
    D、计算出模型分析错误的客户信息样本数量占所有客户信息样本数量的比例是否小于预设阈值；
    E、若模型分析错误的客户信息样本数量占所有客户信息样本数量的比例小于预设阈值，则该预先确定的模型训练结束；
    F、若模型分析错误的客户信息样本数量占所有客户信息样本数量的比例大于或者等于预设阈值，则按照预设的比例增加幅度在总客户信息样本中增加与模型分析错误的客户信息样本属于同一类型的客户信息样本的比重，并按照预设的比例减少幅度在总客户信息样本中减少模型分析正确的客户信息样本的比重，并重新执行上述步骤C、D、E、F。
  9. 一种模型分析装置，包括处理设备、存储设备，该存储设备存储有模 型分析程序，该模型存储程序被所述处理设备执行时实现如下步骤：
    A、基于预设数量的客户信息样本，训练多种预先确定的模型；
    B、将训练的多种模型按照预先确定的组合规则组合成复合模型，并在收到待分析的客户信息后，将该待分析的客户信息输入所述复合模型以输出分析结果。
  10. 如权利要求9所述的模型分析装置，其特征在于，所述基于预设数量的客户信息样本，训练多种预先确定的模型的步骤包括：
    C、在训练一种预先确定的模型过程中，每训练一次后，将各个客户信息样本分别输入当前训练的模型中以确定出模型分析错误的客户信息样本；
    D、计算出模型分析错误的客户信息样本数量占所有客户信息样本数量的比例是否小于预设阈值；
    E、若模型分析错误的客户信息样本数量占所有客户信息样本数量的比例小于预设阈值，则该预先确定的模型训练结束；
    F、若模型分析错误的客户信息样本数量占所有客户信息样本数量的比例大于或者等于预设阈值，则按照预设的比例增加幅度在总客户信息样本中增加与模型分析错误的客户信息样本属于同一类型的客户信息样本的比重，并按照预设的比例减少幅度在总客户信息样本中减少模型分析正确的客户信息样本的比重，并重新执行上述步骤C、D、E、F。
  11. 如权利要求9所述的模型分析装置，其特征在于，所述预先确定的模型的数量为N，N为大于2的自然数，第i个预先确定的模型记为Fi，i为小于或者等于N的正整数，所述将训练的多种模型按照预先确定的组合规则组合成复合模型为：
    所述复合模型＝(1/N)*F1+(1/N)*F2+……+(1/N)*FN。
  12. 如权利要求11所述的模型分析装置，其特征在于，所述基于预设数量的客户信息样本，训练多种预先确定的模型的步骤包括：
    C、在训练一种预先确定的模型过程中，每训练一次后，将各个客户信息样本分别输入当前训练的模型中以确定出模型分析错误的客户信息样本；
    D、计算出模型分析错误的客户信息样本数量占所有客户信息样本数量的比例是否小于预设阈值；
    E、若模型分析错误的客户信息样本数量占所有客户信息样本数量的比例小于预设阈值，则该预先确定的模型训练结束；
    F、若模型分析错误的客户信息样本数量占所有客户信息样本数量的比例大于或者等于预设阈值，则按照预设的比例增加幅度在总客户信息样本中增加与模型分析错误的客户信息样本属于同一类型的客户信息样本的比重，并按照预设的比例减少幅度在总客户信息样本中减少模型分析正确的客户信息样本的比重，并重新执行上述步骤C、D、E、F。13、如权利要求9所述的模型分析装置，其特征在于，所述预先确定的模型的数量为N，N为大于2的自然数，第i个预先确定的模型记为Fi，i为小于或者等于N的正整数，所 述将训练的多种模型按照预先确定的组合规则组合成复合模型为：
    所述复合模型＝POWER(F1,1/N)*POWER(F2,1/N)*……*POWER(FN,1/N)，其中，POWER(Fi,1/N)是对Fi模型分析出的结果求N次方根。
  13. 如权利要求9所述的模型分析装置，其特征在于，所述预先确定的模型的数量为N，N为大于2的自然数，第i个预先确定的模型记为Fi，i为小于或者等于N的正整数，所述将训练的多种模型按照预先确定的组合规则组合成复合模型为：
    所述复合模型＝POWER(F1,1/N)*POWER(F2,1/N)*……*POWER(FN,1/N)，其中，POWER(Fi,1/N)是对Fi模型分析出的结果求N次方根。
  14. 如权利要求13所示的模型分析装置，其特征在于，所述基于预设数量的客户信息样本，训练多种预先确定的模型的步骤包括：
    C、在训练一种预先确定的模型过程中，每训练一次后，将各个客户信息样本分别输入当前训练的模型中以确定出模型分析错误的客户信息样本；
    D、计算出模型分析错误的客户信息样本数量占所有客户信息样本数量的比例是否小于预设阈值；
    E、若模型分析错误的客户信息样本数量占所有客户信息样本数量的比例小于预设阈值，则该预先确定的模型训练结束；
    F、若模型分析错误的客户信息样本数量占所有客户信息样本数量的比例大于或者等于预设阈值，则按照预设的比例增加幅度在总客户信息样本中增加与模型分析错误的客户信息样本属于同一类型的客户信息样本的比重，并按照预设的比例减少幅度在总客户信息样本中减少模型分析正确的客户信息样本的比重，并重新执行上述步骤C、D、E、F。15、如权利要求9所述的模型分析装置，其特征在于，所述预先确定的模型的数量为N，N为大于2的自然数，第i个预先确定的模型记为Fi，i为小于或者等于N的正整数，所述将训练的多种模型按照预先确定的组合规则组合成复合模型为：
    所述复合模型＝N/(1/F1+1/F2+……+1/FN)。
  15. 如权利要求9所述的模型分析装置，其特征在于，所述预先确定的模型的数量为N，N为大于2的自然数，第i个预先确定的模型记为Fi，i为小于或者等于N的正整数，所述将训练的多种模型按照预先确定的组合规则组合成复合模型为：
    所述复合模型＝N/(1/F1+1/F2+……+1/FN)。
  16. 如权利要求15所述的模型分析装置，其特征在于，所述基于预设数量的客户信息样本，训练多种预先确定的模型的步骤包括：
    C、在训练一种预先确定的模型过程中，每训练一次后，将各个客户信息样本分别输入当前训练的模型中以确定出模型分析错误的客户信息样本；
    D、计算出模型分析错误的客户信息样本数量占所有客户信息样本数量的比例是否小于预设阈值；
    E、若模型分析错误的客户信息样本数量占所有客户信息样本数量的比例 小于预设阈值，则该预先确定的模型训练结束；
    F、若模型分析错误的客户信息样本数量占所有客户信息样本数量的比例大于或者等于预设阈值，则按照预设的比例增加幅度在总客户信息样本中增加与模型分析错误的客户信息样本属于同一类型的客户信息样本的比重，并按照预设的比例减少幅度在总客户信息样本中减少模型分析正确的客户信息样本的比重，并重新执行上述步骤C、D、E、F。
  17. 一种计算机可读存储介质，其上存储有至少一个可被处理设备执行以实现以下操作的计算机可读指令：
    A、基于预设数量的客户信息样本，训练多种预先确定的模型；
    B、将训练的多种模型按照预先确定的组合规则组合成复合模型，并在收到待分析的客户信息后，将该待分析的客户信息输入所述复合模型以输出分析结果。
  18. 如权利要求17所述的计算机可读存储介质，其特征在于，所述预先确定的模型的数量为N，N为大于2的自然数，第i个预先确定的模型记为Fi，i为小于或者等于N的正整数，所述将训练的多种模型按照预先确定的组合规则组合成复合模型为：
    所述复合模型＝(1/N)*F1+(1/N)*F2+……+(1/N)*FN。
  19. 如权利要求17所述的计算机可读存储介质，其特征在于，所述预先确定的模型的数量为N，N为大于2的自然数，第i个预先确定的模型记为Fi，i为小于或者等于N的正整数，所述将训练的多种模型按照预先确定的组合规则组合成复合模型为：
    所述复合模型＝POWER(F1,1/N)*POWER(F2,1/N)*……*POWER(FN,1/N)，其中，POWER(Fi,1/N)是对Fi模型分析出的结果求N次方根。
  20. 如权利要求17所述的计算机可读存储介质，其特征在于，所述预先确定的模型的数量为N，N为大于2的自然数，第i个预先确定的模型记为Fi，i为小于或者等于N的正整数，所述将训练的多种模型按照预先确定的组合规则组合成复合模型为：
    所述复合模型＝N/(1/F1+1/F2+……+1/FN)。
PCT/CN2017/100042 2017-05-05 2017-08-31 模型分析方法、装置、及计算机可读存储介质 WO2018201646A1 (zh)

Priority Applications (6)

Application Number Priority Date Filing Date Title
SG11201900264PA SG11201900264PA (en) 2017-05-05 2017-08-31 Method and device of analysis based on model, and computer readable storage medium
US16/084,242 US11507963B2 (en) 2017-05-05 2017-08-31 Method and device of analysis based on model, and computer readable storage medium
KR1020187024854A KR20190021189A (ko) 2017-05-05 2017-08-31 모델 분석 방법, 장치 및 컴퓨터 판독 가능한 저장매체
EP17899229.3A EP3428854A4 (en) 2017-05-05 2017-08-31 MODEL ANALYSIS METHOD AND DEVICE, AND COMPUTER READABLE INFORMATION MEDIUM
AU2017408798A AU2017408798A1 (en) 2017-05-05 2017-08-31 Method and device of analysis based on model, and computer readable storage medium
JP2018534667A JP2019519821A (ja) 2017-05-05 2017-08-31 モデル分析方法、装置、及びコンピュータ読み取り可能な記憶媒体

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201710311997.6A CN107766929B (zh) 2017-05-05 2017-05-05 模型分析方法及装置
CN201710311997.6 2017-05-05

Publications (1)

Publication Number Publication Date
WO2018201646A1 true WO2018201646A1 (zh) 2018-11-08

Family

ID=61265240

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2017/100042 WO2018201646A1 (zh) 2017-05-05 2017-08-31 模型分析方法、装置、及计算机可读存储介质

Country Status (8)

Country Link
US (1) US11507963B2 (zh)
EP (1) EP3428854A4 (zh)
JP (1) JP2019519821A (zh)
KR (1) KR20190021189A (zh)
CN (1) CN107766929B (zh)
AU (1) AU2017408798A1 (zh)
SG (1) SG11201900264PA (zh)
WO (1) WO2018201646A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US11651380B1 (en) * 2022-03-30 2023-05-16 Intuit Inc. Real-time propensity prediction using an ensemble of long-term and short-term user behavior models

Families Citing this family (12)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109492191B (zh) * 2018-09-17 2024-05-07 平安科技（深圳）有限公司 计算投保概率的方法、装置、计算机设备和存储介质
US11124185B2 (en) * 2018-11-13 2021-09-21 Zoox, Inc. Perception collision avoidance
JP7167668B2 (ja) * 2018-11-30 2022-11-09 コニカミノルタ株式会社 学習方法、学習装置、プログラムおよび記録媒体
US10977738B2 (en) * 2018-12-27 2021-04-13 Futurity Group, Inc. Systems, methods, and platforms for automated quality management and identification of errors, omissions and/or deviations in coordinating services and/or payments responsive to requests for coverage under a policy
CN109816158A (zh) * 2019-01-04 2019-05-28 平安科技（深圳）有限公司 预测模型的组合方法、装置、设备及可读存储介质
CN110321944A (zh) * 2019-06-26 2019-10-11 华中科技大学 一种基于接触网画质评估的深度神经网络模型的构建方法
CN111402068B (zh) * 2020-02-22 2023-07-25 中国平安财产保险股份有限公司 基于大数据的保费数据分析方法、装置及存储介质
CN111489134B (zh) * 2020-04-13 2024-05-10 深圳前海微众银行股份有限公司 数据模型的构建方法、装置、设备及计算机可读存储介质
CN111581615A (zh) * 2020-05-08 2020-08-25 南京大创师智能科技有限公司 一种向个人提供人工智能平台的方法及系统
US20220138632A1 (en) * 2020-10-29 2022-05-05 Accenture Global Solutions Limited Rule-based calibration of an artificial intelligence model
CN113457167A (zh) * 2021-06-29 2021-10-01 网易（杭州）网络有限公司 用户分类网络的训练方法、用户分类方法及装置
CN113487182B (zh) * 2021-07-06 2024-04-26 新奥新智科技有限公司 设备健康状态评估方法、装置、计算机设备和介质

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102270309A (zh) * 2011-07-27 2011-12-07 华北电力大学 一种基于集成学习的短期电力负荷预测方法
CN105719002A (zh) * 2016-01-18 2016-06-29 重庆大学 一种基于组合预测的风电机组状态参数异常辨识方法
CN105894138A (zh) * 2016-06-28 2016-08-24 山东大学 一种制造业发货量的最优加权组合预测方法
CN106157132A (zh) * 2016-06-20 2016-11-23 中国工商银行股份有限公司 信用风险监控系统及方法
US20170061330A1 (en) * 2015-08-31 2017-03-02 International Business Machines Corporation Method, system and computer program product for learning classification model
CN106548210A (zh) * 2016-10-31 2017-03-29 腾讯科技（深圳）有限公司 机器学习模型训练方法及装置

Family Cites Families (20)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CA2938561C (en) * 2004-02-27 2019-09-03 Accenture Global Services Limited System for individualized customer interaction
US7958063B2 (en) * 2004-11-11 2011-06-07 Trustees Of Columbia University In The City Of New York Methods and systems for identifying and localizing objects based on features of the objects that are mapped to a vector
US7496549B2 (en) * 2005-05-26 2009-02-24 Yahoo! Inc. Matching pursuit approach to sparse Gaussian process regression
US8719073B1 (en) * 2005-08-25 2014-05-06 Hewlett-Packard Development Company, L.P. Producing a measure regarding cases associated with an issue after one or more events have occurred
US7937350B2 (en) * 2007-02-14 2011-05-03 International Business Machines Corporation Method, system and program product for determining a time for retraining a data mining model
US20090240539A1 (en) * 2008-03-21 2009-09-24 Microsoft Corporation Machine learning system for a task brokerage system
US8762180B2 (en) * 2009-08-25 2014-06-24 Accenture Global Services Limited Claims analytics engine
CN102184453A (zh) * 2011-05-16 2011-09-14 上海电气集团股份有限公司 基于模糊神经网络和支持向量机的风电功率组合预测方法
WO2013012898A2 (en) * 2011-07-19 2013-01-24 Causata Inc. Distributed scalable incrementally updated models in decisioning systems
CN102508907B (zh) * 2011-11-11 2013-11-20 北京航空航天大学 一种基于训练集优化的推荐系统的动态推荐方法
CN102663513B (zh) * 2012-03-13 2016-04-20 华北电力大学 利用灰色关联度分析的风电场功率组合预测建模方法
US9070046B2 (en) * 2012-10-17 2015-06-30 Microsoft Technology Licensing, Llc Learning-based image webpage index selection
US20150032598A1 (en) * 2013-07-26 2015-01-29 Corelogic Solutions, Llc System and method for generating a natural hazard credit model
CN103679287A (zh) * 2013-12-05 2014-03-26 王海燕 一种组合型电力负荷预测方法
US10289962B2 (en) * 2014-06-06 2019-05-14 Google Llc Training distilled machine learning models
US9697469B2 (en) 2014-08-13 2017-07-04 Andrew McMahon Method and system for generating and aggregating models based on disparate data from insurance, financial services, and public industries
JP6239486B2 (ja) 2014-11-05 2017-11-29 株式会社東芝 予測モデル作成方法
US9280740B1 (en) 2015-01-12 2016-03-08 Bank Of America Corporation Transforming predictive models
CN106355140B (zh) * 2016-08-22 2018-03-02 平安科技（深圳）有限公司 车辆检测的方法及装置
US10671932B1 (en) * 2017-01-27 2020-06-02 Intuit Inc. Software application selection models integration

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102270309A (zh) * 2011-07-27 2011-12-07 华北电力大学 一种基于集成学习的短期电力负荷预测方法
US20170061330A1 (en) * 2015-08-31 2017-03-02 International Business Machines Corporation Method, system and computer program product for learning classification model
CN105719002A (zh) * 2016-01-18 2016-06-29 重庆大学 一种基于组合预测的风电机组状态参数异常辨识方法
CN106157132A (zh) * 2016-06-20 2016-11-23 中国工商银行股份有限公司 信用风险监控系统及方法
CN105894138A (zh) * 2016-06-28 2016-08-24 山东大学 一种制造业发货量的最优加权组合预测方法
CN106548210A (zh) * 2016-10-31 2017-03-29 腾讯科技（深圳）有限公司 机器学习模型训练方法及装置

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
See also references of EP3428854A4 *

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US11651380B1 (en) * 2022-03-30 2023-05-16 Intuit Inc. Real-time propensity prediction using an ensemble of long-term and short-term user behavior models

Also Published As

Publication number Publication date
EP3428854A1 (en) 2019-01-16
EP3428854A4 (en) 2020-05-13
CN107766929B (zh) 2019-05-24
CN107766929A (zh) 2018-03-06
US20210117989A1 (en) 2021-04-22
AU2017408798A1 (en) 2018-11-22
SG11201900264PA (en) 2019-02-27
JP2019519821A (ja) 2019-07-11
US11507963B2 (en) 2022-11-22
KR20190021189A (ko) 2019-03-05

Similar Documents

Publication Publication Date Title
WO2018201646A1 (zh) 模型分析方法、装置、及计算机可读存储介质
WO2019169756A1 (zh) 产品推荐方法、装置及存储介质
Wang et al. Methods for correcting inference based on outcomes predicted by machine learning
Chen et al. Feature selection with redundancy-complementariness dispersion
US11704500B2 (en) Techniques to add smart device information to machine learning for increased context
US20170308678A1 (en) Disease prediction system using open source data
CN111210335B (zh) 用户风险识别方法、装置及电子设备
US20130246290A1 (en) Machine-Assisted Legal Assessments
Zhao et al. Price trend prediction of stock market using outlier data mining algorithm
US10387800B2 (en) System and methods for generating reason codes for ensemble computer models
US11373760B2 (en) False detection rate control with null-hypothesis
CN112070545B (zh) 用于优化信息触达的方法、装置、介质和电子设备
CN112988840A (zh) 一种时间序列预测方法、装置、设备和存储介质
TWI677830B (zh) 模型中關鍵變量的探測方法及裝置
US11928683B2 (en) Feature encoding in online application origination (OAO) service for a fraud prevention system
CN113886697A (zh) 基于聚类算法的活动推荐方法、装置、设备及存储介质
CN110796379B (zh) 业务渠道的风险评估方法、装置、设备及存储介质
CN110598995B (zh) 智能客户评级方法、装置及计算机可读存储介质
US11704598B2 (en) Machine-learning techniques for evaluating suitability of candidate datasets for target applications
WO2019218517A1 (zh) 服务器、文本数据的处理方法及存储介质
CN114925275A (zh) 产品推荐方法、装置、计算机设备及存储介质
US9892462B1 (en) Heuristic model for improving the underwriting process
US11645555B2 (en) Feature selection using Sobolev Independence Criterion
CN116630059A (zh) 基于人工智能的流失预测方法、装置、设备及存储介质
CN116796140A (zh) 基于人工智能的异常分析方法、装置、设备及存储介质

Legal Events

Date Code Title Description
ENP Entry into the national phase

Ref document number: 2018534667

Country of ref document: JP

Kind code of ref document: A

ENP Entry into the national phase

Ref document number: 20187024854

Country of ref document: KR

Kind code of ref document: A

WWE Wipo information: entry into national phase

Ref document number: 2017899229

Country of ref document: EP

ENP Entry into the national phase

Ref document number: 2017899229

Country of ref document: EP

Effective date: 20180912

ENP Entry into the national phase

Ref document number: 2017408798

Country of ref document: AU

Date of ref document: 20170831

Kind code of ref document: A

NENP Non-entry into the national phase

Ref country code: DE