WO2017140222A1 - 机器学习模型的建模方法及装置 - Google Patents

机器学习模型的建模方法及装置 Download PDF

Info

Publication number
WO2017140222A1
WO2017140222A1 PCT/CN2017/073023 CN2017073023W WO2017140222A1 WO 2017140222 A1 WO2017140222 A1 WO 2017140222A1 CN 2017073023 W CN2017073023 W CN 2017073023W WO 2017140222 A1 WO2017140222 A1 WO 2017140222A1
Authority
WO
WIPO (PCT)
Prior art keywords
machine learning
variable
initial target
target
target variable
Prior art date
Application number
PCT/CN2017/073023
Other languages
English (en)
French (fr)
Inventor
张柯
褚崴
施兴
谢树坤
谢锋
Original Assignee
阿里巴巴集团控股有限公司
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 阿里巴巴集团控股有限公司 filed Critical 阿里巴巴集团控股有限公司
Priority to JP2018542277A priority Critical patent/JP7102344B2/ja
Publication of WO2017140222A1 publication Critical patent/WO2017140222A1/zh
Priority to US15/999,073 priority patent/US20180374098A1/en

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
    • G06Q20/00—Payment architectures, schemes or protocols
    • G06Q20/38—Payment protocols; Details thereof
    • G06Q20/40—Authorisation, e.g. identification of payer or payee, verification of customer or shop credentials; Review and approval of payers, e.g. check credit lines or negative lists
    • G06Q20/401—Transaction verification
    • G06Q20/4016—Transaction verification involving fraud or risk level assessment in transaction processing
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
    • G06F17/10—Complex mathematical operations
    • G06F17/11—Complex mathematical operations for solving equations, e.g. nonlinear equations, general mathematical optimization problems
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
    • G06F17/10—Complex mathematical operations
    • G06F17/18—Complex mathematical operations for evaluating statistical data, e.g. average values, frequency distributions, probability functions, regression analysis
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • 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
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16Z—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS, NOT OTHERWISE PROVIDED FOR
    • G16Z99/00—Subject matter not provided for in other main groups of this subclass

Definitions

  • the present invention relates to computer technology, and in particular, to a method and apparatus for modeling a machine learning model.
  • the machine learning model When using the machine learning model to judge the behavior pattern, common feature extraction is often performed for various specific behaviors belonging to a certain target behavior, so that the machine learning model is constructed according to these common characteristics. In determining whether a particular behavior belongs to the target behavior, the constructed machine learning model determines whether it is the target behavior based on whether it has these common characteristics.
  • a false transaction means that the seller user and/or the buyer user obtains the false product sales by fictitious or concealing the facts of the transaction, circumventing or maliciously using the credit history rules, interfering with or impeding the credit record order, and the like. Inappropriate interests such as shop ratings, credit scores, or product reviews. For example, there have been false transaction types such as swiping, speculating, cashing, and making fake order loans. In order to regulate online trading behavior, it is necessary to judge the behavior pattern of false transactions.
  • each type of false transaction has a variety of specific implementation methods, and the transaction behavior between various types of false transactions is quite different.
  • the extraction of common features cannot be adopted.
  • the way to construct a machine model for judging a false transaction therefore, the prior art employs a method in which each machine learning model makes a determination for a certain implementation form or a type of fraudulent transaction.
  • each machine learning model makes a determination for a certain implementation form or a type of fraudulent transaction.
  • the invention provides a modeling method and device for a machine learning model, which is used for constructing a machine learning model which can be applied to comprehensively determine a target behavior when a target behavior has multiple different types of implementation forms, thereby saving costs. Improve recognition efficiency.
  • a method of modeling a machine learning model comprising:
  • the machine learning sub-model corresponding to each intermediate target variable is respectively trained to obtain a probability value of each machine learning sub-model; wherein the machine learning sub-model is used to determine a false indication corresponding to the intermediate target variable according to the characteristic variable describing the transaction behavior Transaction Type;
  • a target machine learning model for determining a fraudulent transaction is established based on the target probability value and the feature variable.
  • a modeling apparatus for a machine learning model comprising:
  • Obtaining a module configured to obtain mutually exclusive intermediate target variables; wherein the intermediate target variable is used to indicate a type of the fake transaction;
  • a training module configured to respectively train a machine learning sub-model corresponding to each intermediate target variable to obtain a probability value of each machine learning sub-model; wherein the machine learning sub-model is used to determine a corresponding intermediate according to a characteristic variable describing a transaction behavior The type of false transaction indicated by the target variable;
  • a summation module configured to perform probability summation on probability values of each machine learning submodel to obtain a target probability value
  • a modeling module configured to establish a target machine learning model for determining a fake transaction according to the target probability value and the feature variable.
  • the method and device for modeling a machine learning model provided by an embodiment of the present invention, by training for a machine learning sub-model corresponding to each intermediate target variable, obtaining a probability value of each machine learning sub-model, and then learning each machine learning sub-model
  • the probability values are subjected to probability summation, and a target machine learning model for determining the target behavior is established based on the target probability value obtained by the summation and the feature variables for describing the transaction behavior.
  • each machine learning sub-model is used to determine a specific type of target behavior, probability values of the machine learning sub-models are probabilistically summed, thereby obtaining a probability that the sample belongs to at least one of a plurality of target behavior types, A machine learning model built on this probability can be used to make a comprehensive decision on the target behavior. For example, if this modeling method is applied to the application scenario of a false transaction, the constructed model can comprehensively determine the false transaction without the need to construct different implementation forms or types for the false transaction in the prior art. Multiple models achieve the goal of saving costs and improving the recognition efficiency of false transactions.
  • FIG. 1 is a schematic flowchart of a modeling method of a machine learning model according to Embodiment 1 of the present invention
  • FIG. 2 is a schematic flowchart of a modeling method of a machine learning model according to Embodiment 2 of the present invention
  • 3 is a schematic flow chart of reconstructing a target variable
  • FIG. 4 is a schematic structural diagram of a modeling device for a machine learning model according to Embodiment 3 of the present invention.
  • FIG. 5 is a schematic structural diagram of a device for modeling a machine learning model according to Embodiment 4 of the present invention.
  • FIG. 1 is a schematic flowchart of a method for modeling a machine learning model according to Embodiment 1 of the present invention.
  • the method provided in this embodiment may be used for determining a false transaction, that is, in this embodiment.
  • the target behavior mentioned may be a false transaction, and the method provided in this embodiment may also be used in other abnormal transactions, which is not limited in this embodiment, as shown in FIG. 1 , including:
  • Step 101 Train the machine learning sub-model corresponding to each intermediate target variable to obtain a probability value of each machine learning sub-model.
  • the machine learning sub-model is used to determine the target behavior type indicated by the corresponding intermediate target variable according to the characteristic variable describing the transaction behavior.
  • the target behavior has multiple implementation forms, and the implementation form with similar transaction behavior can be divided into one type, so that within each type, the transaction behavior is similar, but the transaction between each type Behaviors tend to be more variable.
  • false transactions have multiple implementation forms.
  • the trading behaviors of speculative and swipe orders are relatively similar. It can be divided into one type, and the trading behavior of cashing and fraudulent loans is quite different from each other and can be used as a type.
  • the compatible initial target variables may be merged according to the compatibility or mutual exclusion state between the initial target variables used to indicate the specific implementation form of the target behavior, and the mutual mutual interactions are obtained.
  • the intermediate target variable of the state If the transaction behaviors of the target behaviors of different implementation forms are different, the initial target variables corresponding to different implementation forms are mutually exclusive; if the transaction behaviors of the target behaviors of different implementation forms are less different, different implementations are implemented.
  • the initial target variables corresponding to the form are compatible.
  • the machine learning sub-model corresponding to each intermediate target variable is constructed.
  • the machine learning sub-model is a two-class model for determining whether the sample belongs to the target behavior type indicated by the corresponding intermediate target variable according to the characteristic variable describing the transaction behavior.
  • each machine learning sub-model is trained to obtain the probability values of each machine learning sub-model.
  • Step 102 Perform probability summation on probability values of each machine learning submodel to obtain a target probability value.
  • probabilistic summation of the probability values of the machine learning sub-models can obtain at least one of the plurality of target behavior types. The probability of one, the target probability value.
  • Step 103 Establish a target machine learning model for determining a target behavior according to the target probability value and the feature variable.
  • a target machine learning model for determining a target behavior is constructed.
  • the target machine learning model is a two-class model, the probability of the target machine learning model takes the target probability value, and the input of the target machine learning model is a characteristic variable describing the transaction behavior.
  • the output is a target variable for indicating whether it is a target behavior.
  • the value of the target variable is 0 or 1.
  • the probability values of the machine learning sub-models are obtained, and then the probability values of the machine learning sub-models are probability-summed, based on the summation.
  • the obtained target probability value and the feature variable used to describe the transaction behavior establish a target machine learning model for determining the target behavior.
  • the target behavior is a false transaction, so that each machine learning sub-model is used to determine the type of false transaction indicated by the corresponding intermediate target variable, and the probability value of each machine learning sub-model is obtained.
  • the probability of determining at least one of the plurality of types of fraudulent transactions can be obtained, and the model constructed based on the probability value can determine a plurality of types of fraudulent transactions, thereby realizing cost saving and improving false transactions. Identify the purpose of efficiency.
  • FIG. 2 is a schematic flowchart of a modeling method of a machine learning model according to Embodiment 2 of the present invention.
  • a machine learning model for comprehensively determining a false transaction is taken as an example to further illustrate an embodiment of the present invention.
  • the technical solution, as shown in FIG. 2, includes:
  • Step 201 Obtain a preset initial target variable and a feature variable.
  • each transaction record formed by a transaction that has occurred in history is recorded, and each transaction record includes three dimensions of transaction information, which are buyer transaction information, seller transaction information, and commodity transaction information.
  • each transaction record also includes information on whether or not it belongs to a specific implementation form of various fraudulent transactions.
  • the specific implementation forms of the false transaction include but are not limited to billing, cashing, fraudulent loans and speculation.
  • the user can extract parameters for describing the transaction information from the historical transaction data, set the feature variable x, and extract the parameter for describing the type of the fake transaction as the initial feature variable y.
  • the brush sheet can be used as the initial feature variable y 1
  • the cashout is taken as the initial feature variable y 2
  • the fraudulent loan is taken as the initial feature variable y 3
  • the speculative letter is taken as the initial feature variable y 4 .
  • the user can extract as much the parameter used to describe the transaction information as the feature variable, and extract the perfect transaction information to make the transaction behavior described by the feature variable. It is more accurate. Based on this, the machine learning model is more accurate when performing analysis operations such as classification.
  • Step 202 Obtain a mutually exclusive intermediate target variable according to the initial target variable.
  • the compatibility or mutual exclusion state between the initial target variables is determined, and the compatible initial target variables are combined according to the compatible or mutually exclusive states to obtain intermediate target variables that are mutually exclusive states.
  • Num ij represents the number of transaction records in the historical transaction data that are simultaneously defined as positive samples by the initial target variable y i and the initial target variable y j .
  • Num i represents the number of transaction records in the historical transaction data defined by the initial target variable y i as a positive sample
  • Num j represents the number of transaction records in the historical transaction data defined by the initial target variable y j as a positive sample.
  • the range of values of i and j is: 1 ⁇ i ⁇ N, 1 ⁇ j ⁇ N.
  • N is the total number of initial feature variables.
  • Mutually exclusive is more strict, that is, the smaller the impact of one initial target variable on the value of another initial target variable.
  • the positive sample here means that the transaction record belongs to the false transaction type indicated by the initial target variable, and vice versa.
  • Mutual exclusion means that the value of one initial target variable has little effect on the value of another initial target variable, and the blending means that the value of one initial target variable has a large influence on the value of another initial target variable. .
  • Each split set contains an initial target variable in the initial target variable pair and all elements in the split set except the initial target variable pair.
  • the split sets that have an inclusion relationship with each other are merged to obtain the target subset.
  • the initial target variables in the same target subset are merged as the intermediate target variable Y.
  • the initial target variable pair y1 and y2, the initial target variable pair y1 and y4, and the initial target variable pair are determined.
  • the reconstruction process of splitting and merging may be performed on the split set ⁇ y1, y2, y3, y4 ⁇
  • FIG. 3 is a flow diagram of reconstructing the target variable.
  • the obtained target subset is ⁇ y1, y2 ⁇ , ⁇ y2, y3 ⁇ , ⁇ y4 ⁇ .
  • the y1 and y2 phases are combined into Y1, the y2 and y3 phases are combined into Y2, and y4 is taken as Y3.
  • Step 203 Construct a machine learning sub-model corresponding to each intermediate target variable.
  • a two-class machine learning sub-model is constructed for each intermediate target variable, and the machine learning sub-model of the intermediate target variable is used to determine whether the sample is a positive sample of the intermediate target variable.
  • a machine learning submodel for each intermediate target variable can be The feature variables are filtered.
  • the characteristic variables of each machine learning sub-model after screening may be different. Only the same-direction feature variables are retained in each machine learning sub-model to avoid the inconsistency of the feature variables.
  • the specific screening process is: calculating a covariance between each feature variable and each initial target variable used to obtain the intermediate target variable, and filtering out characteristic variables that are inconsistent with the positive and negative covariance between the initial target variables.
  • the characteristic variables include X 1 , X 2 , ... X q ..., X n , where n is the total number of characteristic variables
  • the intermediate target variables include Y 1 , Y 2 , ... Y v ..., Y N ' , wherein , N' is the total number of intermediate target variables.
  • the initial target variables obtained by combining the intermediate target variables Y v are denoted as y s .
  • the formula is used for each of the feature variables X q in turn .
  • S is the number of initial target variables y s in which the intermediate target variable Y v is obtained
  • X qk is the feature variable X q in the historical transaction data.
  • the value in the transaction record, y sk is the value of the initial target variable y s in the kth transaction record in the historical transaction data.
  • the machine learning submodel M of the intermediate target variable Y is constructed, and the input of the machine learning submodel M is the filtered feature variable X, and the output is the intermediate target variable Y.
  • Step 204 Train the machine learning sub-model corresponding to each intermediate target variable to obtain a probability of the machine learning sub-model.
  • each transaction record in the historical transaction data is used as a training sample, and each machine learning sub-model is trained by using the training sample set formed by the historical transaction data to obtain the probability p v of the machine learning sub-model.
  • the weight of each initial target variable used to merge and obtain the intermediate target variable corresponding to the machine learning sub-model can be used in the historical transaction data for each transaction record.
  • the copy is made, and the copied historical transaction data is used as a training sample set. In this way, a training sample set for each machine learning submodel is constructed.
  • the weight is used to indicate the severity of the initial target variable, when the initial target variable is more serious, after the copy operation, the number of positive samples of the initial target variable is more in the training sample set, thereby training
  • the process can improve the simulation of training.
  • the predetermined combined to obtain respective initial target variable of the intermediate target variable Y v y of weights s weights are W 1, W 2, ... W s ... W S .
  • the weights W s of the initial target variables y s of the intermediate target variable Y v are merged and substituted into the formula.
  • the machine learning submodel corresponding to each intermediate target variable is trained by the copied training sample set, and the probabilities p 1 , p 2 , ... p v ..., p N ' of the machine learning submodel are obtained.
  • Step 205 Perform probability summation on the probability of the machine learning submodel to obtain a target probability value.
  • the probabilities p 1 , p 2 , ... p v ..., p N ' of the machine learning submodel are substituted into the formula Calculate the probability P of the machine learning model.
  • Step 206 Construct a machine learning model.
  • the machine learning model is a two-class model
  • the machine learning model has a probability P
  • the input is a feature variable x
  • the output is a target variable for indicating whether it is a false transaction.
  • the machine learning model constructed is used to determine whether the transaction behavior described by the input characteristic variable is a false transaction.
  • the sample is determined based on the machine learning model. If it is determined that the sample is a positive sample, the probability that the transaction indicated by the sample is a false transaction is high. If the sample is determined to be a negative sample, the transaction indicated by the sample is indicated. The possibility of a false transaction is small.
  • FIG. 4 is a schematic structural diagram of a device for modeling a machine learning model according to Embodiment 3 of the present invention. As shown in FIG. 4, the method includes: a training module 41, a summation module 42, and a modeling module 43.
  • the training module 41 is configured to separately train the machine learning sub-model corresponding to each intermediate target variable to obtain a probability value of each machine learning sub-model.
  • the machine learning sub-model is used to determine the target behavior type indicated by the corresponding intermediate target variable according to the characteristic variable describing the transaction behavior.
  • the summation module 42 is configured to perform probability summation on the probability values of the machine learning submodels to obtain the target probability values.
  • the summation module 42 is specifically configured to substitute the probability p v of each machine learning submodel into a formula. Calculate the probability P of the machine learning model.
  • N' is the number of machine learning submodels.
  • the modeling module 43 is configured to establish a target machine learning model for determining whether the target behavior is based on the target probability value and the feature variable.
  • each machine learns the probability values of the sub-models, and then probabilistic values of the machine learning sub-models are summed, and the target probability values obtained by the summation and the characteristic variables used to describe the transaction behavior are used to determine the target behavior.
  • Target machine learning model In the application scenario of false transaction judgment, the target behavior is a false transaction, so that each machine learning sub-model is used to determine the type of false transaction indicated by the corresponding intermediate target variable, and the probability value of each machine learning sub-model is obtained. And the probability of determining at least one of the plurality of types of fraudulent transactions can be obtained, and the model constructed based on the probability value can determine a plurality of types of fraudulent transactions, thereby realizing cost saving and improving false transactions. Identify the purpose of efficiency.
  • FIG. 5 is a schematic structural diagram of a modeling device for a machine learning model according to Embodiment 4 of the present invention. As shown in FIG. 5, on the basis of the modeling device of the machine learning model provided in FIG. 4, the modeling device further Including: obtaining module 44.
  • the obtaining module 44 is configured to combine the compatible initial target variables according to the compatible or mutually exclusive states between the initial target variables to obtain intermediate target variables that are mutually exclusive states; wherein the initial target Variables are used to indicate the implementation of the target behavior.
  • FIG. 6 is a schematic structural diagram of another apparatus for modeling a machine learning model according to Embodiment 4 of the present invention.
  • the obtaining module 44 further includes: a calculating unit 441, a combining unit 442, and a constructing Unit 443, split unit 444, merge unit 445, and determination unit 446.
  • a calculation unit 441 for using a formula according to Determine the compatibility or mutual exclusion state between the initial target variables.
  • Num ij represents the number of transaction records in which the initial target variable y i and the initial target variable y j are simultaneously defined as positive samples
  • Num i represents that the initial target variable y i is defined as a positive sample in the historical transaction data.
  • the number of transaction records, Num j represents the number of transaction records in which the initial target variable y j is defined as a positive sample in the historical transaction data, 1 ⁇ i ⁇ N, 1 ⁇ j ⁇ N, where N is the initial characteristic variable
  • T 1 and T 2 are preset thresholds, 0 ⁇ T 1 ⁇ 1, 0 ⁇ T 2 ⁇ 1.
  • the combining unit 442 is configured to form an initial target variable pair of initial target variables that are mutually exclusive between the two.
  • a building unit 443 is configured to construct a split set including each initial target variable.
  • a splitting unit 444 configured to sequentially perform, for each initial target variable pair, split the split set into two lower-level split sets according to the initial target variable pair, and perform splitting at each lower level
  • the diversity set includes an initial target variable of the initial target variable pair and a step of all elements other than the initial target variable pair in the split set that is split.
  • the split set of the next level is used to split according to the next initial target variable pair.
  • the merging unit 445 is configured to merge the split sets in which the inclusion relationship exists with each other to obtain the target subset.
  • the determining unit 446 is configured to merge the initial target variables in the same target subset as the intermediate target variable.
  • the machine learning sub-model is a linear model
  • the modeling device further includes: a covariance calculation module 45, a screening module 46, a calculation module 47, a replication module 48, and a sample module 49.
  • Covariance calculation module 45 a learning for each sub-model for the machine, calculation variable X q wherein each initial target covariance between the variables y s.
  • the initial target variable y s is used to merge to obtain an intermediate target variable corresponding to the machine learning sub-model.
  • Filter module 46 wherein if the variable X for the covariance between the sign of s are not the same, then filtering out the variable X q wherein each of Q and the initial target variable y, otherwise retaining the characteristic variable X q .
  • Weight calculation module 47 a record for each transaction, the initial and the target variable y s y s initial target variable weight W s, into the formula The number of copies CN of the transaction record is obtained.
  • the copying module 48 is configured to: for each machine learning sub-model, copy the transaction records in the historical transaction data according to the number of transaction record copies determined by each initial target variable y s weight W s ;
  • the initial target variable y s is used to merge to obtain an intermediate target variable corresponding to the machine learning sub-model;
  • the sample module 49 is configured to use the copied historical transaction data as a training sample of the machine learning sub-model.
  • the device provided in this embodiment is used to perform the method shown in FIG. 1 and FIG. 2, and the description of each functional module is described in the foregoing embodiment, and details are not described in this embodiment.
  • the probability values of the machine learning sub-models are obtained, and then the probability values of the machine learning sub-models are probability-summed, based on the summation.
  • the obtained target probability value and the characteristic variable used to describe the transaction behavior establish the purpose for determining the target behavior Standard machine learning model.
  • the target behavior is a false transaction, so that each machine learning sub-model is used to determine the type of false transaction indicated by the corresponding intermediate target variable, and the probability value of each machine learning sub-model is obtained.
  • the probability of determining at least one of the plurality of types of fraudulent transactions can be obtained, and the model constructed based on the probability value can determine a plurality of types of fraudulent transactions, thereby realizing cost saving and improving false transactions. Identify the purpose of efficiency.
  • the aforementioned program can be stored in a computer readable storage medium.
  • the program when executed, performs the steps including the foregoing method embodiments; and the foregoing storage medium includes various media that can store program codes, such as a ROM, a RAM, a magnetic disk, or an optical disk.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Software Systems (AREA)
  • Mathematical Physics (AREA)
  • General Engineering & Computer Science (AREA)
  • Artificial Intelligence (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Computation (AREA)
  • Business, Economics & Management (AREA)
  • Computing Systems (AREA)
  • Medical Informatics (AREA)
  • Pure & Applied Mathematics (AREA)
  • Mathematical Optimization (AREA)
  • Mathematical Analysis (AREA)
  • Computational Mathematics (AREA)
  • Accounting & Taxation (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Operations Research (AREA)
  • Algebra (AREA)
  • Databases & Information Systems (AREA)
  • Strategic Management (AREA)
  • Finance (AREA)
  • General Business, Economics & Management (AREA)
  • Computer Security & Cryptography (AREA)
  • Probability & Statistics with Applications (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

提供了一种机器学习模型的建模方法及装置，通过针对每一个中间目标变量所对应的机器学习子模型进行训练，获得各机器学习子模型的概率值，进而将各机器学习子模型的概率值进行概率求和，基于求和所获得的目标概率值以及用于描述交易行为的特征变量建立用于判定目标行为的目标机器学习模型。在虚假交易判断的应用场景中，目标行为即为虚假交易，从而每一个机器学习子模型用于判定对应的中间目标变量所指示的虚假交易类型，对各机器学习子模型的概率值进行概率求和就能够获得用于判定属于多种虚假交易类型中的至少一种的概率，基于这一概率值所构建的模型便能够对于多种虚假交易类型进行判定，实现了节省成本，提高虚假交易的识别效率的目的。

Description

机器学习模型的建模方法及装置
本申请要求2016年02月19日递交的申请号为201610094664.8、发明名称为“机器学习模型的建模方法及装置”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本发明涉及计算机技术，尤其涉及一种机器学习模型的建模方法及装置。
背景技术
在利用机器学习模型进行行为模式的判定时，往往采用针对同属于某一种目标行为的各种具体行为进行共性特征提取，从而依据这些共性特征构建机器学习模型。在判定某一个具体行为是否属于该目标行为时，所构建的机器学习模型依据其是否具有这些共性特征，来确定是否为该目标行为。
以网络中的虚假交易为例，虚假交易是指卖方用户和/或买方用户通过虚构或隐瞒交易事实、规避或恶意利用信用记录规则、干扰或妨害信用记录秩序等不正当方式获取虚假的商品销量、店铺评分、信用积分或商品评论等不当利益的行为。例如现在已经出现了刷单、炒信、套现和制造虚假订单贷款等虚假交易类型，为了规范网络交易行为，需要对虚假交易这种行为模式进行判定。
由于虚假交易有着不同的类型，每一种类型的虚假交易又有着多种具体的实现方式，同时各种类型虚假交易之间的交易行为差异较大，在现有技术中，无法采用提取共性特征的方式构建用于判断虚假交易的机器模型，因此，现有技术中采用了每一个机器学习模型针对某一种实现形式或者一个类型的虚假交易进行判定的方式。在利用机器学习模型识别虚假交易时，便需要建立多个模型，需要耗费大量成本，识别效率较低。
发明内容
本发明提供一种机器学习模型的建模方法及装置，用于构建一个能够适用于在目标行为具有多种不同类型的实现形式时，对目标行为进行综合判定的机器学习模型，从而节省成本，提高识别效率。
为达到上述目的，本发明的实施例采用如下技术方案：
第一方面，提供了一种机器学习模型的建模方法，包括：
获得互斥的各中间目标变量；其中，所述中间目标变量用于指示虚假交易的类型；
分别对各中间目标变量对应的机器学习子模型进行训练，获得各机器学习子模型的概率值；其中，所述机器学习子模型用于根据描述交易行为的特征变量判定对应中间目标变量指示的虚假交易类型；
对各机器学习子模型的概率值进行概率求和，获得目标概率值；
根据所述目标概率值和所述特征变量，建立用于判定虚假交易的目标机器学习模型。
第二方面，提供了一种机器学习模型的建模装置，包括：
获得模块，用于获得互斥的各中间目标变量；其中，所述中间目标变量用于指示虚假交易的类型；
训练模块，用于分别对各中间目标变量对应的机器学习子模型进行训练，获得各机器学习子模型的概率值；其中，所述机器学习子模型用于根据描述交易行为的特征变量判定对应中间目标变量指示的虚假交易类型；
求和模块，用于对各机器学习子模型的概率值进行概率求和，获得目标概率值；
建模模块，用于根据所述目标概率值和所述特征变量，建立用于判定虚假交易的目标机器学习模型。
本发明实施例提供的机器学习模型的建模方法及装置，通过针对每一个中间目标变量所对应的机器学习子模型进行训练，获得各机器学习子模型的概率值，进而将各机器学习子模型的概率值进行概率求和，基于求和所获得的目标概率值以及用于描述交易行为的特征变量建立用于判定目标行为的目标机器学习模型。由于每一个机器学习子模型用于判定目标行为的某一特定类型，对各机器学习子模型的概率值进行概率求和，从而获得样本属于多种目标行为类型中的至少一种类型的概率，基于这一概率所构建的机器学习模型可以用于对目标行为进行综合判定。举例来说，若将这种建模方法应用于虚假交易的应用场景中，所构建的模型便能够对于虚假交易进行综合判定，而无需现有技术中需要针对虚假交易的不同实现形式或类型构建多个模型，实现了节省成本，提高虚假交易的识别效率的目的。
上述说明仅是本发明技术方案的概述，为了能够更清楚了解本发明的技术手段，而可依照说明书的内容予以实施，并且为了让本发明的上述和其它目的、特征和优点能够更明显易懂，以下特举本发明的具体实施方式。
附图说明
通过阅读下文优选实施方式的详细描述，各种其他的优点和益处对于本领域普通技术人员将变得清楚明了。附图仅用于示出优选实施方式的目的，而并不认为是对本发明的限制。而且在整个附图中，用相同的参考符号表示相同的部件。在附图中：
图1为本发明实施例一提供的一种机器学习模型的建模方法的流程示意图；
图2为本发明实施例二提供的一种机器学习模型的建模方法的流程示意图；
图3为重构目标变量的流程示意图；
图4为本发明实施例三提供的一种机器学习模型的建模装置的结构示意图；
图5为本发明实施例四提供的一种机器学习模型的建模装置的结构示意图。
具体实施方式
下面将参照附图更详细地描述本公开的示例性实施例。虽然附图中显示了本公开的示例性实施例，然而应当理解，可以以各种形式实现本公开而不应被这里阐述的实施例所限制。相反，提供这些实施例是为了能够更透彻地理解本公开，并且能够将本公开的范围完整的传达给本领域的技术人员。
下面结合附图对本发明实施例提供的机器学习模型的建模方法及装置进行详细描述。
实施例一
图1为本发明实施例一提供的一种机器学习模型的建模方法的流程示意图，本实施例所提供的方法，可以用于进行虚假交易的判断中，也就是说，本实施例中所提及的目标行为可以为虚假交易，本实施例所提供的方法还可以用于其他异常交易中，本实施例中对此不做限定，如图1所示，包括：
步骤101、分别对各中间目标变量对应的机器学习子模型进行训练，获得各机器学习子模型的概率值。
其中，机器学习子模型用于根据描述交易行为的特征变量判定对应中间目标变量指示的目标行为类型。
具体地，目标行为有着多种实现形式，可以将交易行为相似的实现形式划分为一种类型，从而使得在每一种类型的内部，交易行为是相似的，但每一种类型之间的交易行为往往具有较大的差异性。例如，在虚假交易判定的应用场景中，虚假交易有着多种实现形式，在刷单、套现、骗贷和炒信的实现形式中，炒信和刷单的交易行为相似性较高， 可以划分为一个类型，而套现和骗贷的交易行为则与之具有较大差异，可以分别作为一个类型。
在进行目标行为的类别划分时，可以根据用于指示目标行为的具体实现形式的初始目标变量之间的相容或互斥状态，对相容的初始目标变量进行合并，获得相互之间为互斥状态的中间目标变量。若不同实现形式的目标行为的交易行为差异性较大，则不同实现形式所对应的初始目标变量之间为相斥的；若不同实现形式的目标行为的交易行为差异性较小，则不同实现形式所对应的初始目标变量之间为相容的。
构建各中间目标变量对应的机器学习子模型，机器学习子模型是一个二分类模型，用于根据描述交易行为的特征变量判定样本是否属于对应中间目标变量指示的目标行为类型。利用训练样本，对各机器学习子模型进行训练，获得各机器学习子模型的概率值。
步骤102、对各机器学习子模型的概率值进行概率求和，获得目标概率值。
由于每一个机器学习子模型用于判定对应的中间目标变量所指示的目标行为类型，对各机器学习子模型的概率值进行概率求和就能够获得用于判定属于多种目标行为类型中的至少一种的概率，即目标概率值。
步骤103、根据目标概率值和特征变量，建立用于判定目标行为的目标机器学习模型。
具体地，构建用于判定目标行为的目标机器学习模型，目标机器学习模型是一个二分类模型，目标机器学习模型的概率取目标概率值，目标机器学习模型的输入为描述交易行为的特征变量，输出为用于指示是否为目标行为的目标变量，该目标变量的取值为0或1，当根据样本的特征变量判定出目标变量取值为1时，该样本为正样本，即属于目标行为，否则不为目标行为。
本实施例中，通过针对每一个中间目标变量所对应的机器学习子模型进行训练，获得各机器学习子模型的概率值，进而将各机器学习子模型的概率值进行概率求和，基于求和所获得的目标概率值以及用于描述交易行为的特征变量建立用于判定目标行为的目标机器学习模型。在虚假交易判断的应用场景中，目标行为即为虚假交易，从而每一个机器学习子模型用于判定对应的中间目标变量所指示的虚假交易类型，对各机器学习子模型的概率值进行概率求和就能够获得用于判定属于多种虚假交易类型中的至少一种的概率，基于这一概率值所构建的模型便能够对于多种虚假交易类型进行判定，实现了节省成本，提高虚假交易的识别效率的目的。
实施例二
图2为本发明实施例二提供的一种机器学习模型的建模方法的流程示意图，本实施例中，以构建对虚假交易进行综合判定的机器学习模型为例，进一步说明本发明实施例的技术方案，如图2所示，该方法包括：
步骤201、获得预设的初始目标变量和特征变量。
具体的，在历史交易数据中记录有历史发生过的交易所形成的交易记录，在每一条交易记录中包括三个维度的交易信息，分别为买家交易信息、卖家交易信息和商品交易信息。另外，每一条交易记录还包括是否属于各种虚假交易的具体实现形式的信息，虚假交易具体实现形式包括但不限于刷单、套现、骗贷和炒信。
用户可以从历史交易数据中，提取用于描述交易信息的参数，设定为特征变量x，以及提取用于描述虚假交易类型的参数设定为初始特征变量y。
例如：可以将刷单作为初始特征变量y1，将套现作为初始特征变量y2，将骗贷作为初始特征变量y3，将炒信作为初始特征变量y4。
由于历史信息中包含有大量的参数，在进行特征变量设定时，用户可以尽量多地提取用于描述交易信息的参数作为特征变量，通过提取完善的交易信息，使得特征变量所描述的交易行为也就更加准确。基于此所建立的机器学习模型在进行分类等分析运算时，所获得的结果也就更加准确。
步骤202、根据初始目标变量获得互斥的中间目标变量。
具体地，确定初始目标变量之间的相容或互斥状态，根据相容或互斥状态，对相容的初始目标变量进行合并，获得相互之间为互斥状态的中间目标变量。
首先，根据公式
Figure PCTCN2017073023-appb-000001
确定初始目标变量之间的相容或互斥状态，将存在互斥状态的两初始目标变量作为一个初始目标变量对。其中，Numij表示历史交易数据中被初始目标变量yi和初始目标变量yj同时定义为正样本的交易记录的条数。Numi表示历史交易数据中被初始目标变量yi定义为正样本的交易记录的条数，Numj表示历史交易数据中被初始目标变量yj定义为正样本的交易记录的条数。i和j的取值范围为：1≤i≤N，1≤j≤N。N为初始特征变量的总数。H＝1时两初始目标变量之间为互斥，H＝0时两初始目标变量之间为相融。另外，上述公式中的0.2仅是一种优选的阈值，在实际使用中，可以选择其他取值，阈值的取值越小，则当H＝1时所判定出的 两初始目标变量之间的互斥就更加严格，也就是说，一个初始目标变量对另一初始目标变量的取值所构成的影响越小。
需要说明的是，正样本在这里是指交易记录属于初始目标变量所指示的虚假交易类型，反之为负样本。互斥在这里是指一个初始目标变量的取值对另一个初始目标变量的取值构成影响小，而相融是指一个初始目标变量的取值对另一个初始目标变量的取值构成影响大。
其次，构建一个拆分集合，集合中包括全部的初始目标变量。进而，重复执行根据一个初始目标变量对，将拆分集合拆分为两个下一级的拆分集合的步骤，其中，下一级的拆分集合用于根据下一个初始目标变量对进行拆分，直至全部的初始目标变量对执行完毕。在每一个拆分集合中包含有初始目标变量对中的一个初始目标变量，以及被拆分集合中除初始目标变量对之外的全部元素。对相互之间存在包含关系的拆分集合进行合并，获得目标子集。将同一目标子集中的初始目标变量合并，作为中间目标变量Y。
例如：若初始目标变量为y1，y2，y3，y4，经过计算，若确定存在互斥关系的为初始目标变量对y1和y2、初始目标变量对y1和y4、初始目标变量对y2和y4、初始目标变量对y3和y4，则可以据此对拆分集合{y1，y2，y3，y4}执行进行拆分和合并的重构过程，图3为重构目标变量的流程示意图，如图3所示，获得的目标子集为{y1，y2}、{y2，y3}、{y4}。将y1和y2相合并为Y1，将y2和y3相合并为Y2，将y4作为Y3。
步骤203、构建各中间目标变量对应的机器学习子模型。
具体地，针对每一个中间目标变量构建一个二分类的机器学习子模型，该中间目标变量的机器学习子模型用于确定样本是否为该中间目标变量的正样本。
在机器学习子模型为线性模型的基础上，为了在对机器学习子模型进行训练的过程中，提高机器学习子模型的性能，减少训练噪声，可以针对每一个中间目标变量的机器学习子模型，进行特征变量的筛选，筛选后的每一个机器学习子模型的特征变量有可能是不相同的，在每个机器学习子模型中仅保留同向的特征变量，避免由于特征变量的同向性不一致所导致的训练噪声。具体筛选过程为：计算各特征变量与用于合并获得该中间目标变量的各初始目标变量之间的协方差，筛选掉与这些初始目标变量之间的协方差正负向不一致的特征变量。
例如：特征变量包括X1、X2、…Xq…、Xn，其中，n为特征变量的总个数，中间目标变量包括Y1、Y2、…Yv…、YN’，其中，N’为中间目标变量的总个数。
合并获得中间目标变量Yv的各初始目标变量记为ys。在中间目标变量Yv的机器学习子模型中，依次针对每一个特征变量Xq，采用公式
Figure PCTCN2017073023-appb-000002
计算与各初始目标变量ys之间的协方差，若计算出的各协方差特征变量Covq1，Covq2，……Covq S正负号相同，则保留该特征变量Xq，否则筛选掉该特征变量Xq。
其中，1≤q≤n，1≤s≤S，S为合并获得该中间目标变量Yv的各初始目标变量ys的个数，Xqk为特征变量Xq在历史交易数据中第k条交易记录中的取值，ysk为初始目标变量ys在历史交易数据中第k条交易记录中的取值，
Figure PCTCN2017073023-appb-000003
为特征变量Xq在历史交易数据中的平均值，
Figure PCTCN2017073023-appb-000004
为各目标初始目标变量ys在历史交易数据中的平均值。
构建中间目标变量Y的机器学习子模型M，机器学习子模型M的输入为筛选后的特征变量X，输出为中间目标变量Y。
步骤204、对各中间目标变量对应的机器学习子模型进行训练，获得机器学习子模型的概率。
具体地，将历史交易数据中的每一条交易记录作为一个训练样本，利用历史交易数据所构成的训练样本集，分别对各机器学习子模型进行训练，获得机器学习子模型的概率pv。
为了使得对于机器学习子模型进行训练的模拟性更好，可以根据用于合并获得机器学习子模型所对应的中间目标变量的各初始目标变量的权重，在历史交易数据中，对每一条交易记录进行复制，将复制后的历史交易数据作为训练样本集。采用该方式，构建每个机器学习子模型的训练样本集。
由于权重是用于指示初始目标变量的严重程度，从而当初始目标变量越严重时，经过复制的操作后，在训练样本集中，该初始目标变量的正样本的数量也就越多，从而在训练过程中能够提高训练的模拟性。
例如：
对于中间目标变量Yv的机器学习子模型构建训练样本集时，预先确定合并获得该中间目标变量Yv的各初始目标变量ys的权重分别为W1、W2、…Ws…WS。针对每一条交易记录，将合并获得该中间目标变量Yv的各初始目标变量ys的权重Ws，代入公式
Figure PCTCN2017073023-appb-000005
若该训练样本为初始目标变量ys正样本，则ys＝1，若该训练样本为初始目标变量ys 负样本，则ys＝0，获得该样本的复制份数CN。对每一条训练样本复制对应的CN份，从而构建一个训练样本集。
进而，利用该复制后的训练样本集分别对各中间目标变量对应的机器学习子模型进行训练，获得机器学习子模型的概率p1、p2、…pv…、pN’。
步骤205、对机器学习子模型的概率进行概率求和，获得目标概率值。
具体地，将机器学习子模型的概率p1、p2、…pv…、pN’代入公式
Figure PCTCN2017073023-appb-000006
计算获得机器学习模型的概率P。
步骤206、构建机器学习模型。
具体地，该机器学习模型为二分类模型，机器学习模型概率为P，输入为特征变量x，输出为用于指示是否为虚假交易的目标变量。所构建的机器学习模型用于判定输入的特征变量所描述的交易行为是否属于虚假交易。基于该机器学习模型对样本进行判定，若判定出该样本为正样本则说明样本所指示的交易为虚假交易的可能性很高，若判定出该样本为负样本，则说明样本所指示的交易为虚假交易的可能性很小。
实施例三
图4为本发明实施例三提供的一种机器学习模型的建模装置的结构示意图，如图4所示，包括：训练模块41、求和模块42和建模模块43。
训练模块41，用于分别对各中间目标变量对应的机器学习子模型进行训练，获得各机器学习子模型的概率值。
其中，机器学习子模型用于根据描述交易行为的特征变量判定对应中间目标变量指示的目标行为类型。
求和模块42，用于对各机器学习子模型的概率值进行概率求和，获得目标概率值。
具体地，求和模块42，具体用于将各机器学习子模型的概率pv代入公式
Figure PCTCN2017073023-appb-000007
计算获得机器学习模型的概率P。
其中，N’为机器学习子模型的个数。
建模模块43，用于根据目标概率值和特征变量，建立用于判定是否为目标行为的目标机器学习模型。
本实施例中，通过针对每一个中间目标变量所对应的机器学习子模型进行训练，获 得各机器学习子模型的概率值，进而将各机器学习子模型的概率值进行概率求和，基于求和所获得的目标概率值以及用于描述交易行为的特征变量建立用于判定目标行为的目标机器学习模型。在虚假交易判断的应用场景中，目标行为即为虚假交易，从而每一个机器学习子模型用于判定对应的中间目标变量所指示的虚假交易类型，对各机器学习子模型的概率值进行概率求和就能够获得用于判定属于多种虚假交易类型中的至少一种的概率，基于这一概率值所构建的模型便能够对于多种虚假交易类型进行判定，实现了节省成本，提高虚假交易的识别效率的目的。
实施例四
图5为本发明实施例四提供的一种机器学习模型的建模装置的结构示意图，如图5所示，在图4所提供的机器学习模型的建模装置的基础上，建模装置还包括：获得模块44。
获得模块44，用于根据各初始目标变量之间的相容或互斥状态，对相容的初始目标变量进行合并，获得相互之间为互斥状态的中间目标变量；其中，所述初始目标变量用于指示目标行为的实现形式。
具体地，图6为本发明实施例四提供的另一种机器学习模型的建模装置的结构示意图，作为一种可能的实现形式，获得模块44进一步包括：计算单元441、组合单元442、构建单元443、拆分单元444、合并单元445和确定单元446。
计算单元441，用于根据公式
Figure PCTCN2017073023-appb-000008
确定初始目标变量之间的相容或互斥状态。
其中，Numij表示历史交易数据中初始目标变量yi和初始目标变量yj同时定义为正样本的交易记录的条数，Numi表示所述历史交易数据中初始目标变量yi定义为正样本的交易记录的条数，Numj表示所述历史交易数据中初始目标变量yj定义为正样本的交易记录的条数，1≤i≤N，1≤j≤N，N为初始特征变量的总数，H＝1时两初始目标变量之间为互斥，H＝0时两初始目标变量之间为相融。
T1和T2为预设阈值，0＜T1＜1，0＜T2＜1。优选的，T1＝T2＝0.2。
组合单元442，用于将两两之间为互斥状态的初始目标变量组成初始目标变量对。
构建单元443，用于构建包括各初始目标变量的拆分集合。
拆分单元444，用于依次针对每一个初始目标变量对，执行根据所述初始目标变量对，将拆分集合拆分为两个下一级的拆分集合，在每一个下一级的拆分集合中包含有所述初始目标变量对中的一个初始目标变量，以及被拆分的拆分集合中除所述初始目标变量对之外的全部元素的步骤。
其中，所述下一级的拆分集合用于根据下一个初始目标变量对进行拆分。
合并单元445，用于对相互之间存在包含关系的拆分集合进行合并，获得目标子集。
确定单元446，用于将同一目标子集中的初始目标变量合并，作为所述中间目标变量。
进一步，机器学习子模型为线性模型，建模装置，还包括：协方差计算模块45、筛选模块46、计算模块47、复制模块48和样本模块49。
协方差计算模块45，用于针对每一个机器学习子模型，计算特征变量Xq与各初始目标变量ys之间的协方差。
其中，所述初始目标变量ys用于合并获得所述机器学习子模型所对应的中间目标变量。
筛选模块46，用于若所述特征变量Xq与各初始目标变量ys之间的协方差正负号不相同，则筛选掉所述特征变量Xq，否则，保留所述特征变量Xq。
计算模块47，用于针对每一条交易记录，将初始目标变量ys及所述初始目标变量ys的权重Ws，代入公式
Figure PCTCN2017073023-appb-000009
获得所述交易记录的复制条数CN。
其中，所述交易记录为所述初始目标变量ys正样本时，ys＝1，否则，ys＝0；S为所述初始目标变量ys的个数。
复制模块48，用于针对每一个机器学习子模型，根据各初始目标变量ys权重Ws所确定出的交易记录复制条数，对所述历史交易数据中的交易记录进行复制；其中，所述初始目标变量ys用于合并获得所述机器学习子模型所对应的中间目标变量；
样本模块49，用于将复制后的历史交易数据作为所述机器学习子模型的训练样本。
本实施例所提供的装置，用于执行如图1和2所示的方法，各功能模块的描述请参见前述实施例，本实施例中不再赘述。
本实施例中，通过针对每一个中间目标变量所对应的机器学习子模型进行训练，获得各机器学习子模型的概率值，进而将各机器学习子模型的概率值进行概率求和，基于求和所获得的目标概率值以及用于描述交易行为的特征变量建立用于判定目标行为的目 标机器学习模型。在虚假交易判断的应用场景中，目标行为即为虚假交易，从而每一个机器学习子模型用于判定对应的中间目标变量所指示的虚假交易类型，对各机器学习子模型的概率值进行概率求和就能够获得用于判定属于多种虚假交易类型中的至少一种的概率，基于这一概率值所构建的模型便能够对于多种虚假交易类型进行判定，实现了节省成本，提高虚假交易的识别效率的目的。
本领域普通技术人员可以理解：实现上述各方法实施例的全部或部分步骤可以通过程序指令相关的硬件来完成。前述的程序可以存储于一计算机可读取存储介质中。该程序在执行时，执行包括上述各方法实施例的步骤；而前述的存储介质包括：ROM、RAM、磁碟或者光盘等各种可以存储程序代码的介质。
最后应说明的是：以上各实施例仅用以说明本发明的技术方案，而非对其限制；尽管参照前述各实施例对本发明进行了详细的说明，本领域的普通技术人员应当理解：其依然可以对前述各实施例所记载的技术方案进行修改，或者对其中部分或者全部技术特征进行等同替换；而这些修改或者替换，并不使相应技术方案的本质脱离本发明各实施例技术方案的范围。

Claims (18)

  1. 一种机器学习模型的建模方法，其特征在于，包括：
    分别对各中间目标变量对应的机器学习子模型进行训练，获得各机器学习子模型的概率值；其中，所述机器学习子模型用于根据描述交易行为的特征变量判定对应中间目标变量指示的目标行为类型；
    对各机器学习子模型的概率值进行概率求和，获得目标概率值；
    根据所述目标概率值和所述特征变量，建立用于判定是否为目标行为的目标机器学习模型。
  2. 根据权利要求1所述的建模方法，其特征在于，所述分别对各中间目标变量对应的机器学习子模型进行训练之前，包括：
    根据各初始目标变量之间的相容或互斥状态，对相容的初始目标变量进行合并，获得相互之间为互斥状态的中间目标变量；其中，所述初始目标变量用于指示目标行为的实现形式。
  3. 根据权利要求2所述的建模方法，其特征在于，所述根据各初始目标变量之间的相容或互斥状态，对相容的初始目标变量进行合并，获得相互之间为互斥状态的中间目标变量，包括：
    将两两之间为互斥状态的初始目标变量组成初始目标变量对；
    构建包括各初始目标变量的拆分集合；
    依次针对每一个初始目标变量对，执行根据所述初始目标变量对，将拆分集合拆分为两个下一级的拆分集合的步骤，在每一个下一级的拆分集合中包含有所述初始目标变量对中的一个初始目标变量，以及被拆分的拆分集合中除所述初始目标变量对之外的全部元素；其中，所述下一级的拆分集合用于根据下一个初始目标变量对进行拆分；
    对相互之间存在包含关系的拆分集合进行合并，获得目标子集；
    将同一目标子集中的初始目标变量合并，作为所述中间目标变量。
  4. 根据权利要求2所述的建模方法，其特征在于，所述根据各初始目标变量之间的相容或互斥状态，对相容的初始目标变量进行合并，获得相互之间为互斥状态的中间目标变量之前，还包括：
    根据公式
    Figure PCTCN2017073023-appb-100001
    确定初始目标变量之 间的相容或互斥状态；其中，Numij表示历史交易数据中初始目标变量yi和初始目标变量yj同时定义为正样本的交易记录的条数，Numi表示所述历史交易数据中初始目标变量yi定义为正样本的交易记录的条数，Numj表示所述历史交易数据中初始目标变量yj定义为正样本的交易记录的条数，1≤i≤N，1≤j≤N，N为初始特征变量的总数，H＝1时两初始目标变量之间为互斥，H＝0时两初始目标变量之间为相融，T1和T2为预设阈值，0＜T1＜1，0＜T2＜1。
  5. 根据权利要求2所述的建模方法，其特征在于，所述机器学习子模型为线性模型，所述分别对各中间目标变量对应的机器学习子模型进行训练，获得各机器学习子模型的概率之前，还包括：
    针对每一个机器学习子模型，计算特征变量Xq与各初始目标变量ys之间的协方差；其中，所述初始目标变量ys用于合并获得所述机器学习子模型所对应的中间目标变量；
    若所述特征变量Xq与各初始目标变量ys之间的协方差正负号不相同，则筛选掉所述特征变量Xq，否则，保留所述特征变量Xq。
  6. 根据权利要求2所述的建模方法，其特征在于，所述分别对各中间目标变量对应的机器学习子模型进行训练，获得各机器学习子模型的概率之前，还包括：
    针对每一个机器学习子模型，根据各初始目标变量ys权重Ws所确定出的交易记录复制条数，对所述历史交易数据中的交易记录进行复制；其中，所述初始目标变量ys用于合并获得所述机器学习子模型所对应的中间目标变量；
    将复制后的历史交易数据作为所述机器学习子模型的训练样本。
  7. 根据权利要求6所述的建模方法，其特征在于，所述对所述历史交易数据中的交易记录进行复制之前，还包括：
    针对每一条交易记录，将初始目标变量ys及所述初始目标变量ys的权重Ws，代入公式
    Figure PCTCN2017073023-appb-100002
    获得所述交易记录的复制条数CN；其中，所述交易记录为所述初始目标变量ys正样本时，ys＝1，否则，ys＝0；S为所述初始目标变量ys的个数。
  8. 根据权利要求1所述的建模方法，其特征在于，所述对各机器学习子模型的概率值进行概率求和，获得目标概率值，包括：
    将各机器学习子模型的概率pv代入公式
    Figure PCTCN2017073023-appb-100003
    计算获得机器学习模型的概率P；其中，N’为机器学习子模型的个数。
  9. 根据权利要求1-8任一项所述的建模方法，其特征在于，所述目标行为为虚假交易。
  10. 一种机器学习模型的建模装置，其特征在于，包括：
    训练模块，用于分别对各中间目标变量对应的机器学习子模型进行训练，获得各机器学习子模型的概率值；其中，所述机器学习子模型用于根据描述交易行为的特征变量判定对应中间目标变量指示的目标行为类型；
    求和模块，用于对各机器学习子模型的概率值进行概率求和，获得目标概率值；
    建模模块，用于根据所述目标概率值和所述特征变量，建立用于判定是否为目标行为的目标机器学习模型。
  11. 根据权利要求10所述的建模装置，其特征在于，所述装置，还包括：
    获得模块，用于根据各初始目标变量之间的相容或互斥状态，对相容的初始目标变量进行合并，获得相互之间为互斥状态的中间目标变量；其中，所述初始目标变量用于指示目标行为的实现形式。
  12. 根据权利要求11所述的建模装置，其特征在于，所述获得模块，包括：
    组合单元，用于将两两之间为互斥状态的初始目标变量组成初始目标变量对；
    构建单元，用于构建包括各初始目标变量的拆分集合；
    拆分单元，用于依次针对每一个初始目标变量对，执行根据所述初始目标变量对，将拆分集合拆分为两个下一级的拆分集合，在每一个下一级的拆分集合中包含有所述初始目标变量对中的一个初始目标变量，以及被拆分的拆分集合中除所述初始目标变量对之外的全部元素的步骤；其中，所述下一级的拆分集合用于根据下一个初始目标变量对进行拆分；
    合并单元，用于对相互之间存在包含关系的拆分集合进行合并，获得目标子集；
    确定单元，用于将同一目标子集中的初始目标变量合并，作为所述中间目标变量。
  13. 根据权利要求11所述的建模装置，其特征在于，所述获得模块，还包括：
    计算单元，用于根据公式
    Figure PCTCN2017073023-appb-100004
    确定初始目标变量之间的相容或互斥状态；其中，Numij表示历史交易数据中初始目标变量yi和初始目标变量yj同时定义为正样本的交易记录的条数，Numi表示所述历史交易数据中初始目标变量yi定义为正样本的交易记录的条数，Numj表示所述历史交易数据中初始目标变 量yj定义为正样本的交易记录的条数，1≤i≤N，1≤j≤N，N为初始特征变量的总数，H＝1时两初始目标变量之间为互斥，H＝0时两初始目标变量之间为相融，T1和T2为预设阈值，0＜T1＜1，0＜T2＜1。
  14. 根据权利要求11所述的建模装置，其特征在于，所述机器学习子模型为线性模型，所述装置，还包括：
    协方差计算模块，用于针对每一个机器学习子模型，计算特征变量Xq与各初始目标变量ys之间的协方差；其中，所述初始目标变量ys用于合并获得所述机器学习子模型所对应的中间目标变量；
    筛选模块，用于若所述特征变量Xq与各初始目标变量ys之间的协方差正负号不相同，则筛选掉所述特征变量Xq，否则，保留所述特征变量Xq。
  15. 根据权利要求11所述的建模装置，其特征在于，所述装置，还包括：
    复制模块，用于针对每一个机器学习子模型，根据各初始目标变量ys权重Ws所确定出的交易记录复制条数，对所述历史交易数据中的交易记录进行复制；其中，所述初始目标变量ys用于合并获得所述机器学习子模型所对应的中间目标变量；
    样本模块，用于将复制后的历史交易数据作为所述机器学习子模型的训练样本。
  16. 根据权利要求15所述的建模装置，其特征在于，所述装置，还包括：
    计算模块，用于针对每一条交易记录，将初始目标变量ys及所述初始目标变量ys的权重Ws，代入公式
    Figure PCTCN2017073023-appb-100005
    获得所述交易记录的复制条数CN；其中，所述交易记录为所述初始目标变量ys正样本时，ys＝1，否则，ys＝0；S为所述初始目标变量ys的个数。
  17. 根据权利要求10所述的建模装置，其特征在于，
    所述求和模块，具体用于将各机器学习子模型的概率pv代入公式
    Figure PCTCN2017073023-appb-100006
    计算获得机器学习模型的概率P；其中，N’为机器学习子模型的个数。
  18. 根据权利要求10-17任一项所述的建模装置，其特征在于，所述目标行为为虚假交易。
PCT/CN2017/073023 2016-02-19 2017-02-07 机器学习模型的建模方法及装置 WO2017140222A1 (zh)

Priority Applications (2)

Application Number Priority Date Filing Date Title
JP2018542277A JP7102344B2 (ja) 2016-02-19 2017-02-07 機械学習モデルのモデリング方法及びデバイス
US15/999,073 US20180374098A1 (en) 2016-02-19 2018-08-17 Modeling method and device for machine learning model

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201610094664.8A CN107103171B (zh) 2016-02-19 2016-02-19 机器学习模型的建模方法及装置
CN201610094664.8 2016-02-19

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US15/999,073 Continuation US20180374098A1 (en) 2016-02-19 2018-08-17 Modeling method and device for machine learning model

Publications (1)

Publication Number Publication Date
WO2017140222A1 true WO2017140222A1 (zh) 2017-08-24

Family

ID=59624727

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2017/073023 WO2017140222A1 (zh) 2016-02-19 2017-02-07 机器学习模型的建模方法及装置

Country Status (5)

Country Link
US (1) US20180374098A1 (zh)
JP (1) JP7102344B2 (zh)
CN (1) CN107103171B (zh)
TW (1) TWI789345B (zh)
WO (1) WO2017140222A1 (zh)

Cited By (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20180374098A1 (en) * 2016-02-19 2018-12-27 Alibaba Group Holding Limited Modeling method and device for machine learning model
CN111080360A (zh) * 2019-12-13 2020-04-28 中诚信征信有限公司 行为预测方法、模型训练方法、装置、服务器及存储介质
CN111626762A (zh) * 2019-02-28 2020-09-04 富士通株式会社 计算机可读记录介质、确定方法以及信息处理设备
US11155555B2 (en) 2015-09-23 2021-10-26 Janssen Pharmaceutica Nv Compounds
CN113705824A (zh) * 2021-01-23 2021-11-26 深圳市玄羽科技有限公司 一种用于构建机器学习建模过程的系统
US11542247B2 (en) 2015-09-23 2023-01-03 Janssen Pharmaceutica Nv Bi-heteroaryl substitute 1,4-benzodiazepines and uses thereof for the treatment of cancer
CN116205301A (zh) * 2023-01-31 2023-06-02 苏州浪潮智能科技有限公司 基于量子机器学习的训练框架构建方法、装置、系统

Families Citing this family (16)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107423883B (zh) * 2017-06-15 2020-04-07 创新先进技术有限公司 待处理业务的风险识别方法及装置、电子设备
CN109426701B (zh) * 2017-08-30 2022-04-05 西门子（中国）有限公司 数据模型的运行方法、运行系统和存储介质
CN108228706A (zh) * 2017-11-23 2018-06-29 中国银联股份有限公司 用于识别异常交易社团的方法和装置
CN109165249B (zh) * 2018-08-07 2020-08-04 阿里巴巴集团控股有限公司 数据处理模型构建方法、装置、服务器和用户端
US11567964B2 (en) * 2018-08-31 2023-01-31 Eligible, Inc. Feature selection for artificial intelligence in healthcare management
CN109325193B (zh) * 2018-10-16 2021-02-26 杭州安恒信息技术股份有限公司 基于机器学习的waf正常流量建模方法以及装置
CN109934709A (zh) * 2018-11-05 2019-06-25 阿里巴巴集团控股有限公司 基于区块链的数据处理方法、装置和服务器
US20200159690A1 (en) * 2018-11-16 2020-05-21 Sap Se Applying scoring systems using an auto-machine learning classification approach
US11593811B2 (en) * 2019-02-05 2023-02-28 International Business Machines Corporation Fraud detection based on community change analysis using a machine learning model
US11574360B2 (en) * 2019-02-05 2023-02-07 International Business Machines Corporation Fraud detection based on community change analysis
CN110263938B (zh) * 2019-06-19 2021-07-23 北京百度网讯科技有限公司 用于生成信息的方法和装置
CN110991650A (zh) * 2019-11-25 2020-04-10 第四范式（北京）技术有限公司 训练养卡识别模型、识别养卡行为的方法及装置
CN111860865B (zh) * 2020-07-23 2022-07-19 中国工商银行股份有限公司 模型构建和分析的方法、装置、电子设备和介质
CN112465626B (zh) * 2020-11-24 2023-08-29 平安科技（深圳）有限公司 基于客户端分类聚合的联合风险评估方法及相关设备
CN113177597A (zh) * 2021-04-30 2021-07-27 平安国际融资租赁有限公司 模型训练数据确定方法、检测模型训练方法、装置及设备
JP7552893B2 (ja) 2021-05-25 2024-09-18 日本電気株式会社 不正検知システム、不正検知方法および不正検知プログラム

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103064987A (zh) * 2013-01-31 2013-04-24 五八同城信息技术有限公司 一种虚假交易信息识别方法
CN104636912A (zh) * 2015-02-13 2015-05-20 银联智惠信息服务（上海）有限公司 信用卡套现识别方法和装置
CN104834918A (zh) * 2015-05-20 2015-08-12 中国科学院上海高等研究院 一种基于高斯过程分类器的人体行为识别方法
US20150363791A1 (en) * 2014-01-10 2015-12-17 Hybrid Application Security Ltd. Business action based fraud detection system and method

Family Cites Families (19)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP4226754B2 (ja) 2000-03-09 2009-02-18 富士電機システムズ株式会社 ニューラルネットワークの最適化学習方法
KR100442835B1 (ko) 2002-08-13 2004-08-02 삼성전자주식회사 인공 신경망을 이용한 얼굴 인식 방법 및 장치
JP2004265190A (ja) 2003-03-03 2004-09-24 Japan Energy Electronic Materials Inc 階層型ニューラルネットワークの学習方法、そのプログラム及びそのプログラムを記録した記録媒体
JP5142135B2 (ja) 2007-11-13 2013-02-13 インターナショナル・ビジネス・マシーンズ・コーポレーション データを分類する技術
JP5072102B2 (ja) 2008-05-12 2012-11-14 パナソニック株式会社 年齢推定方法及び年齢推定装置
CN102467726B (zh) * 2010-11-04 2015-07-29 阿里巴巴集团控股有限公司 一种基于网上交易平台的数据处理方法和装置
US20160223554A1 (en) * 2011-08-05 2016-08-04 Nodality, Inc. Methods for diagnosis, prognosis and methods of treatment
US9916538B2 (en) * 2012-09-15 2018-03-13 Z Advanced Computing, Inc. Method and system for feature detection
JP5835802B2 (ja) 2012-01-26 2015-12-24 日本電信電話株式会社 購買予測装置、方法、及びプログラム
CN103106365B (zh) * 2013-01-25 2015-11-25 中国科学院软件研究所 一种移动终端上的恶意应用软件的检测方法
US20140279379A1 (en) * 2013-03-14 2014-09-18 Rami Mahdi First party fraud detection system
US20140279745A1 (en) * 2013-03-14 2014-09-18 Sm4rt Predictive Systems Classification based on prediction of accuracy of multiple data models
CN104679777B (zh) * 2013-12-02 2018-05-18 中国银联股份有限公司 一种用于检测欺诈交易的方法及系统
WO2015130928A1 (en) * 2014-02-26 2015-09-03 Nancy Packes, Inc. Real estate evaluating platform methods, apparatuses, and media
CN104933053A (zh) * 2014-03-18 2015-09-23 中国银联股份有限公司 非平衡类数据的分类
CN103914064B (zh) * 2014-04-01 2016-06-08 浙江大学 基于多分类器和d-s证据融合的工业过程故障诊断方法
CN105022845A (zh) * 2015-08-26 2015-11-04 苏州大学张家港工业技术研究院 一种基于特征子空间的新闻分类方法及系统
US20170147941A1 (en) * 2015-11-23 2017-05-25 Alexander Bauer Subspace projection of multi-dimensional unsupervised machine learning models
CN107103171B (zh) * 2016-02-19 2020-09-25 阿里巴巴集团控股有限公司 机器学习模型的建模方法及装置

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103064987A (zh) * 2013-01-31 2013-04-24 五八同城信息技术有限公司 一种虚假交易信息识别方法
US20150363791A1 (en) * 2014-01-10 2015-12-17 Hybrid Application Security Ltd. Business action based fraud detection system and method
CN104636912A (zh) * 2015-02-13 2015-05-20 银联智惠信息服务（上海）有限公司 信用卡套现识别方法和装置
CN104834918A (zh) * 2015-05-20 2015-08-12 中国科学院上海高等研究院 一种基于高斯过程分类器的人体行为识别方法

Cited By (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US11155555B2 (en) 2015-09-23 2021-10-26 Janssen Pharmaceutica Nv Compounds
US11542247B2 (en) 2015-09-23 2023-01-03 Janssen Pharmaceutica Nv Bi-heteroaryl substitute 1,4-benzodiazepines and uses thereof for the treatment of cancer
US20180374098A1 (en) * 2016-02-19 2018-12-27 Alibaba Group Holding Limited Modeling method and device for machine learning model
CN111626762A (zh) * 2019-02-28 2020-09-04 富士通株式会社 计算机可读记录介质、确定方法以及信息处理设备
CN111080360A (zh) * 2019-12-13 2020-04-28 中诚信征信有限公司 行为预测方法、模型训练方法、装置、服务器及存储介质
CN111080360B (zh) * 2019-12-13 2023-12-01 中诚信征信有限公司 行为预测方法、模型训练方法、装置、服务器及存储介质
CN113705824A (zh) * 2021-01-23 2021-11-26 深圳市玄羽科技有限公司 一种用于构建机器学习建模过程的系统
CN116205301A (zh) * 2023-01-31 2023-06-02 苏州浪潮智能科技有限公司 基于量子机器学习的训练框架构建方法、装置、系统

Also Published As

Publication number Publication date
CN107103171A (zh) 2017-08-29
US20180374098A1 (en) 2018-12-27
JP2019511037A (ja) 2019-04-18
CN107103171B (zh) 2020-09-25
JP7102344B2 (ja) 2022-07-19
TWI789345B (zh) 2023-01-11
TW201734844A (zh) 2017-10-01

Similar Documents

Publication Publication Date Title
WO2017140222A1 (zh) 机器学习模型的建模方法及装置
KR102061987B1 (ko) 위험 평가 방법 및 시스템
TWI673669B (zh) 評價模型的建模方法及裝置
CN107193876B (zh) 一种基于最近邻knn算法的缺失数据填补方法
CN108960833B (zh) 一种基于异构金融特征的异常交易识别方法，设备及存储介质
CN104915879B (zh) 基于金融数据的社会关系挖掘的方法及装置
CN109300029A (zh) 借贷欺诈检测模型训练方法、借贷欺诈检测方法及装置
CN107392217B (zh) 计算机实现的信息处理方法及装置
CN111325619A (zh) 一种基于联合学习的信用卡欺诈检测模型更新方法及装置
CN110084609B (zh) 一种基于表征学习的交易欺诈行为深度检测方法
CN110930038A (zh) 一种贷款需求识别方法、装置、终端及存储介质
Sasank et al. Credit card fraud detection using various classification and sampling techniques: a comparative study
CN110634060A (zh) 一种用户信用风险的评估方法、系统、装置及存储介质
CN112884480A (zh) 异常交易识别模型的构造方法、装置、计算机设备和介质
CN106874286B (zh) 一种筛选用户特征的方法及装置
Bui et al. A clustering-based shrink autoencoder for detecting anomalies in intrusion detection systems
CN112446505B (zh) 一种元学习建模方法及装置、电子设备、存储介质
CN110570301B (zh) 风险识别方法、装置、设备及介质
Caplescu et al. Will they repay their debt? Identification of borrowers likely to be charged off
CN113554099A (zh) 一种识别异常商户的方法及装置
CN113344581A (zh) 业务数据处理方法及装置
JP2021022199A5 (zh)
Guo et al. Evaluation and Optimization of Machine Learning Algorithms in Personalized Marketing
CN110782342B (zh) 基于二分类模型验证新渠道特征工程正确性的方法及装置
Nasution et al. Credit Risk Detection in Peer-to-Peer Lending Using CatBoost

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 17752654

Country of ref document: EP

Kind code of ref document: A1

ENP Entry into the national phase

Ref document number: 2018542277

Country of ref document: JP

Kind code of ref document: A

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 17752654

Country of ref document: EP

Kind code of ref document: A1