WO2020238229A1 - 交易特征生成模型的训练、交易特征的生成方法和装置 - Google Patents

交易特征生成模型的训练、交易特征的生成方法和装置 Download PDF

Info

Publication number
WO2020238229A1
WO2020238229A1 PCT/CN2020/070847 CN2020070847W WO2020238229A1 WO 2020238229 A1 WO2020238229 A1 WO 2020238229A1 CN 2020070847 W CN2020070847 W CN 2020070847W WO 2020238229 A1 WO2020238229 A1 WO 2020238229A1
Authority
WO
WIPO (PCT)
Prior art keywords
feature
transaction
original
sample
features
Prior art date
Application number
PCT/CN2020/070847
Other languages
English (en)
French (fr)
Inventor
李龙飞
Original Assignee
创新先进技术有限公司
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 创新先进技术有限公司 filed Critical 创新先进技术有限公司
Priority to US16/803,150 priority Critical patent/US11210673B2/en
Publication of WO2020238229A1 publication Critical patent/WO2020238229A1/zh

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00Pattern recognition
    • G06F18/20Analysing
    • G06F18/21Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00Pattern recognition
    • G06F18/20Analysing
    • G06F18/24Classification techniques

Definitions

  • This specification relates to the field of machine learning technology, and in particular to a method and device for training transaction feature generation models and transaction feature generation models.
  • this specification provides a method and device for training transaction feature generation model and transaction feature generation model.
  • a method for generating transaction characteristics, the transaction characteristics being used to identify illegal transactions comprising:
  • Target data set including several pieces of transaction data
  • a training method for a transaction feature generation model including:
  • sample data set including several pieces of sample transaction data with transaction tags, and the transaction tags are used to mark whether the corresponding sample transaction data is an illegal transaction
  • a feature generation method includes:
  • the target data set including several pieces of data
  • a device for generating transaction characteristics, the transaction characteristics being used to identify illegal transactions comprising:
  • a data set acquisition unit to acquire a target data set, the target data set including several pieces of transaction data
  • a feature acquisition unit which acquires several original features of the transaction data, and determines one or more combinations of the original features
  • the feature combination unit determines the feature vector of the new feature obtained by combining the original features based on each combination mode
  • a feature prediction unit which inputs the feature vector into a trained transaction feature generation model, and outputs a prediction result of the new feature
  • the feature generation unit selects several new features whose prediction results meet the specified conditions as the transaction features generated for the target data set.
  • a training device for a transaction feature generation model including:
  • the sample acquisition unit acquires a sample data set, the sample data set includes several pieces of sample transaction data with transaction tags, and the transaction tags are used to mark whether the corresponding sample transaction data is an illegal transaction;
  • a feature obtaining unit which obtains several original features of the sample transaction data, and determines one or more combinations of the original features
  • the feature combination unit determines the feature vector of the new feature obtained by combining the original features based on each combination mode
  • a difference calculation unit which calculates the difference between the new feature and the transaction tag as the feature tag of the new feature
  • the model training unit trains the transaction feature generation model according to the feature vector of the new feature and the feature label.
  • Memory for storing machine executable instructions
  • the processor is prompted to:
  • Target data set including several pieces of transaction data
  • a training device for a transaction feature generation model including:
  • Memory for storing machine executable instructions
  • the processor is prompted to:
  • sample data set including several pieces of sample transaction data with transaction tags, and the transaction tags are used to mark whether the corresponding sample transaction data is an illegal transaction
  • this embodiment can combine the original features of the data to obtain several new features, and then use the trained feature generation model to predict the new features, and select several new features whose prediction results meet the specified conditions.
  • the feature is used as a newly generated feature for subsequent data prediction, so as to realize the automatic generation of the feature, which greatly improves the efficiency of feature generation.
  • Fig. 1 is a schematic flowchart of a training method for a transaction feature generation model shown in an exemplary embodiment of this specification.
  • Fig. 2 is a schematic flowchart of a method for generating transaction features shown in an exemplary embodiment of the present specification.
  • Fig. 3 is a schematic flowchart of a feature generation method shown in an exemplary embodiment of this specification.
  • Fig. 4 is a schematic structural diagram of an apparatus for generating transaction features shown in an exemplary embodiment of this specification.
  • Fig. 5 is a block diagram of a device for generating transaction features shown in an exemplary embodiment of the present specification.
  • Fig. 6 is a schematic structural diagram of a training device for a transaction feature generation model shown in an exemplary embodiment of this specification.
  • Fig. 7 is a block diagram of a training device for a transaction feature generation model shown in an exemplary embodiment of this specification.
  • first, second, third, etc. may be used in this specification to describe various information, the information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other.
  • first information may also be referred to as second information, and similarly, the second information may also be referred to as first information.
  • word “if” as used herein can be interpreted as "when” or “when” or "in response to determination”.
  • This manual provides a transaction feature generation scheme, which can combine the original features of transaction data to obtain several new features, and then use the trained transaction feature generation model to predict the new features, and select the predicted results to meet specified conditions Several of the new features are used as newly generated transaction features for subsequent illegal transaction identification, so as to realize the automatic generation of transaction features, which greatly improves the efficiency of transaction feature generation.
  • Fig. 1 is a schematic flowchart of a training method for a transaction feature generation model shown in an exemplary embodiment of this specification.
  • the training method of the transaction feature generation model may include the following steps:
  • Step 102 Obtain a sample data set.
  • the sample data set includes several pieces of sample transaction data with transaction tags, and the transaction tags are used to mark whether the corresponding sample transaction data is an illegal transaction.
  • transaction data that has been identified for illegal transactions can be obtained as sample transaction data, the sample transaction data has a transaction tag, and the transaction tag is used to mark whether the corresponding sample transaction data is an illegal transaction.
  • the illegal transactions may include: cash transactions, money laundering transactions, etc.
  • the transaction label "0" may indicate that the corresponding sample transaction data is not an illegal transaction; the transaction label "1" may indicate that the corresponding sample transaction data is an illegal transaction, etc.
  • sample transaction data can be divided according to application scenarios, and the sample transaction data in the same application scenario are divided into the same sample data set.
  • sample transaction data for transactions through the e-commerce platform can be divided into sample data set 1.
  • sample transaction data for payment through the consumer credit product can be divided into sample data set 2.
  • one sample data set can be acquired, or multiple sample data sets can be acquired, which is not particularly limited in this specification.
  • Step 104 Obtain several original characteristics of the sample transaction data, and determine one or more combinations of the original characteristics.
  • the original feature is the feature of the sample transaction data itself, such as: transaction amount, number of transactions, distance between transaction location and merchant, merchant category, user category, etc.
  • the types of different original features may be the same or different.
  • the type of transaction amount is numeric
  • the number of transactions is also numeric
  • the type of merchant category is a string
  • the type of user is also a string.
  • numeric type can also be subdivided into floating-point type and integer type.
  • the above transaction amount is a floating point type, for example, the transaction amount is 108.75 yuan and so on.
  • the number of transactions mentioned above is an integer type, for example, the number of transactions is 5.
  • floating-point type and integer type can be unified as numeric type without distinction.
  • categories such as floating-point type, integer type, etc. can be used as the original feature categories, and numeric types are not retained.
  • the original feature "the total transaction amount of the user in the last month” and the original feature "the total transaction amount of the user in the same period last year” can be obtained.
  • the original feature "number of transactions yesterday” and the original feature “number of transactions the day before yesterday” can be obtained.
  • the floating-point original feature the total transaction amount in the past 30 days
  • the integer original feature the number of transactions in the past 30 days
  • the character type original feature "user category” and the integer type original feature “number of transactions in the past 30 days” will not be obtained at the same time, and the two cannot be combined.
  • an original feature group several original features acquired each time can be referred to as an original feature group.
  • multiple acquisitions can be performed to obtain multiple original feature groups.
  • the acquisition of the original features in this embodiment usually takes the sample data set as the operation object, and does not acquire across sample data sets.
  • the original features in each original feature group are from the same sample data set.
  • a combination mode that matches its type may be determined for the original features in the original feature group.
  • the combination mode may include four arithmetic operations such as addition, subtraction, multiplication and division, calculation of logarithmic sum, and calculation of square sum.
  • the combination method may also be to sort the original features by size first, and then extract some of the original features according to the sorting result to perform four arithmetic operations, calculate the square sum, and other operations, which are not particularly limited in this specification.
  • the combination can be subtraction, for example: "number of transactions yesterday” divided by "number of transactions the day before yesterday” can be Get the difference in the number of transactions.
  • the combination method can be an operation on the length of the original feature string, such as the four arithmetic operations on the length of the string, etc.
  • the specific operation method can refer to the combination of numeric original features. This will not be repeated one by one.
  • Step 106 Determine the feature vector of the new feature obtained by combining the original features based on each combination mode.
  • a combination method is adopted to combine the original features in the original feature group, and a new feature can be obtained.
  • the original features in the original feature group are combined to obtain several new features.
  • the feature vector of each new feature can be determined.
  • the original features in the sample data set can be combined based on a combination method, and the value of the new feature obtained by the combination can be used as the feature vector of the new feature.
  • the original feature group obtained from a sample data set includes original features of two numerical types, original feature a and original feature b, and the combination method is addition.
  • the values of feature a are [a 1 ,a 2 ,...,a 100 ]
  • the values of original feature b are [b 1 ,b 2 ,...,b 100 ]
  • the feature vector of the new feature a+b obtained can be [a 1 + b 1 , a 2 + b 2 ,..., a 100 + b 100 ].
  • the feature vector of the new feature can also be generated based on the meta information of the original feature and the combination method.
  • the meta information of the original feature may include the mean value, variance, and number of unique data of the original feature in the sample data set.
  • the meta information of each original feature can form a meta information vector.
  • the presentation form of the combination mode may be a 0/1 vector (Onehot vector), or a low-dimensional vector obtained after embedding the 0/1 vector.
  • the 0/1 vector uses a very long vector to represent the combination mode, the dimension of the vector is the total number of combination modes, and each dimension corresponds to a unique combination mode. If a certain combination method is adopted, the element value of the dimension of the combination method is 1; if a certain combination method is not adopted, the element value of the dimension of the combination method is 0.
  • the 0/1 vector of the combination has 80 dimensions, and each dimension corresponds to a unique combination.
  • the element value of the dimension corresponding to this addition is 1, and the element values of other dimensions are all 0, that is, there is 1 element in the 80-dimensional 0/1 vector
  • the element value of is 1, and the remaining 79 elements have element values of 0.
  • the feature vector of the new feature can be generated based on the meta information vector of the original feature a, the meta information vector of the original feature b, and the combined 0/1 vector. Simpler, these three vectors can be spliced together to obtain the feature vector of the new feature.
  • the feature vector of the new feature can be generated based on the meta information of the original feature, the combination method, and the meta feature of the sample data set.
  • the meta-features of the sample data set may include: the ratio of the sample transaction data volume marked as legal transactions to the sample transaction data volume marked as illegal transactions in the sample data set (positive and negative sample ratio), and the number of original features , Numerical original feature quantity, string-based original feature quantity, etc., the meta features of the sample data set can form a meta feature vector.
  • the original feature's meta-information vector, the combined 0/1 vector, and the sample data set's meta feature vector can be spliced together to obtain the feature vector of the new feature.
  • Step 108 Calculate the difference between the new feature and the transaction tag as the feature tag of the new feature.
  • the value of the new feature of each sample transaction data can be calculated according to the combination method, and then the transaction label of the sample transaction data can be calculated As the true value, calculate the mean square error between the value of the new feature of the sample transaction data and the true value, and use the mean square error as the difference between the new feature and the transaction tag.
  • C represents the feature label of the new feature ab
  • N represents the number of sample transaction data in the sample data set
  • fab i represents the new feature of the i-th sample transaction data
  • the value of ab, l i represents the transaction label of the i-th sample transaction data, where the value of i is 1 to N.
  • the difference between the new feature and the transaction label can also be calculated by algorithms such as Euclidean distance, Mahalanobis distance, etc. This specification does not make special restrictions on this.
  • Step 110 Training the transaction feature generation model according to the feature vector of the new feature and the feature label.
  • the feature vector of the new feature can be input into the transaction feature generation model, and the prediction result of the new feature can be output, and then the difference between the prediction result and the feature label of the new feature can be calculated and based on This difference adjusts the parameters of the transaction feature generation model to realize the training of the transaction feature generation model.
  • a supervised machine learning model may be used as the transaction feature generation model, such as a neural network model, etc., which is not particularly limited in this specification.
  • Fig. 2 is a schematic flowchart of a method for generating transaction features shown in an exemplary embodiment of the present specification.
  • the method for generating the transaction feature may include the following steps:
  • Step 202 Obtain a target data set, where the target data set includes several pieces of transaction data.
  • transaction data in the application scenario can be acquired to obtain a target data set.
  • transaction data that has been identified for illegal transactions in this scenario can be obtained to obtain a target data set, and the transaction data in the target data set has a transaction label.
  • the transaction data in this scenario has not been identified as an illegal transaction
  • the transaction data in the scenario without a transaction tag can also be obtained to obtain the target data set, which is not particularly limited in this specification.
  • the illegal transactions may include: cash transactions, money laundering transactions, etc.
  • Step 204 Obtain several original features of the transaction data, and determine one or more combinations of the original features.
  • the original feature is the feature of the transaction data itself in the target data set, for example: transaction amount, number of transactions, distance between transaction location and merchant, merchant category, etc.
  • each group of original feature groups can be obtained from the target data set, and each group of original feature groups includes several original features, and then a combination mode can be determined for the original features in each original feature group.
  • step 104 For the method of obtaining the original features and determining the combination mode, reference may be made to step 104 in the embodiment shown in FIG. 1, which will not be repeated in this specification.
  • Step 206 Determine the feature vector of the new feature obtained by combining the original features based on each combination mode.
  • a combination method is used to combine the original features in the original feature group to obtain a new feature, and various combinations are used for all the original feature groups.
  • the original features in the original feature group are combined to obtain several new features.
  • the feature vector of each new feature can be determined.
  • the original features in the sample data set can be combined based on a combination method, and the value of the new feature obtained by the combination can be used as the feature vector of the new feature.
  • the feature vector of the new feature can also be generated based on the meta information of the original feature and the combination method.
  • the feature vector of the new feature can be generated based on the meta information of the original feature, the combination method, and the meta feature of the data set.
  • the meta features of the data set may include: the number of original features, the number of numeric original features, the number of string original features, the ratio of positive and negative samples (if any), and the like.
  • the method for generating the above-mentioned feature vector may refer to the method for generating the feature vector recorded in step 106 in the embodiment shown in FIG. 1, which will not be repeated here in this specification.
  • Step 208 Input the feature vector into the trained transaction feature generation model, and output the prediction result of the new feature.
  • Step 210 Select several new features whose prediction results meet specified conditions as transaction features generated for the target data set.
  • the feature vector of each new feature can be input into the trained transaction feature generation model, and the prediction result of the new feature can be output, and then each new feature can be sorted in the descending order of the prediction result.
  • Euclidean distance to calculate the difference between the aforementioned new feature and the label as an example, the smaller the prediction result, the smaller the difference between the new feature and the transaction label, and the more reliable the new feature.
  • the top N new features can be selected as the new transaction features generated for the target data set.
  • the value of N can be preset.
  • the transaction data in the scenario to which the target data set belongs can be identified based on the newly generated transaction characteristics.
  • this manual can combine the original features of the transaction data to obtain several new features, and then use the trained transaction feature generation model to predict the new features, and select a number of prediction results that meet the specified conditions
  • the new feature is used as a newly generated transaction feature for subsequent illegal transaction identification, so as to realize the automatic generation of transaction characteristics, which greatly improves the efficiency of transaction feature generation.
  • This specification also provides a feature generation method, which can automatically generate new features for result prediction based on original features, and improve feature generation efficiency.
  • the feature generation method may include the following steps:
  • Step 302 Obtain a target data set, which includes several pieces of data.
  • the target data set is a data set for generating new features.
  • the data in the target data set may be different.
  • the data in the target data set may be transaction data, and each transaction data may include original characteristics such as transaction amount and number of transactions.
  • the data in the target data set may be user data, and each piece of user data may include original characteristics such as user age, user gender, user historical purchase records, user historical browsing records, and so on.
  • the data in the target data set may be text data, and each piece of text data may include original features such as words and word frequencies included in the corresponding text.
  • the data in the target data set may be user data, and each piece of user data may include original characteristics such as user age, user gender, and user behavior trajectory.
  • Step 304 Obtain several original features of the data, and determine one or more combinations of the original features.
  • the method of acquiring and combining the original features can be determined by referring to the method of acquiring and combining the original features in the transaction feature generation task in the embodiment shown in FIG. 2, which will not be repeated in this specification.
  • Step 306 Determine the feature vector of the new feature obtained by combining the original features based on each combination mode.
  • a combination of the original features can be used to obtain a new feature, and various combinations of the original features can be combined.
  • Several new features are available.
  • the feature vector of each new feature can be determined.
  • the method for determining the feature vector can refer to the embodiment shown in FIG. 1 or FIG. 2, which will not be repeated here in this specification.
  • Step 308 Input the feature vector into the trained feature generation model, and output the prediction result of the new feature.
  • Step 310 Select a number of new features whose prediction results meet specified conditions as features generated for the target data set.
  • the feature generation model can be trained in the same manner as the aforementioned feature generation method to obtain a trained feature generation model.
  • the difference between the new feature and the data label can be used as the feature label of the new feature to adjust the parameters of the feature generation model.
  • the specific training process can refer to the training process of the transaction feature generation model in the transaction feature generation scenario shown in FIG. 1, and this description will not repeat them one by one here.
  • the feature vector of each new feature can be input to the trained feature generation model, and the prediction result of the new feature can be output, and several new features whose prediction results meet specified conditions can be selected for the result of subsequent data prediction.
  • the prediction task is to identify users at risk to identify risky users.
  • One or sample data sets with user tags can be used to train the feature generation model to obtain the trained feature generation model.
  • the user tag is used to mark whether the corresponding user is a risk user
  • sample data set 1 in the first transaction scenario and sample data set 2 in the second transaction scenario can be used to train the feature generation model.
  • Each piece of user data in the sample data set 1 and the sample data set 2 All have user tags.
  • the target data set can be obtained from the third transaction scenario, and then the aforementioned trained feature generation model can be used for the scenario Predict new features.
  • the new feature can be used to predict the user data in the third transaction scenario to predict whether the corresponding user is a risk user.
  • this embodiment can combine the original features of the data to obtain several new features, and then use the trained feature generation model to predict the new features, and select several new features whose prediction results meet the specified conditions.
  • the feature is used as a newly generated feature for subsequent data prediction, so as to realize the automatic generation of the feature, which greatly improves the efficiency of feature generation.
  • this specification also provides an embodiment of a device for generating transaction features.
  • the embodiment of the device for generating transaction features in this specification can be applied on a server.
  • the device embodiments can be implemented by software, or by hardware or a combination of software and hardware. Taking software implementation as an example, as a logical device, it is formed by reading the corresponding computer program instructions in the non-volatile memory into the memory through the processor of the server where it is located. From a hardware perspective, as shown in Figure 4, a hardware structure diagram of the server where the device for generating transaction features is located in this manual, except for the processor, memory, network interface, and non-volatile memory shown in Figure 4 In addition, the server where the device is located in the embodiment usually may include other hardware according to the actual function of the server, which will not be repeated here.
  • Fig. 5 is a block diagram of a device for generating transaction features shown in an exemplary embodiment of the present specification.
  • the transaction feature generation device 400 can be applied to the server shown in FIG. 4, and includes: a data set acquisition unit 401, a feature acquisition unit 402, a feature combination unit 403, a feature prediction unit 404, and a feature Generating unit 405.
  • the data set obtaining unit 401 obtains a target data set, and the target data set includes several pieces of transaction data;
  • the feature acquiring unit 402 acquires several original features of the transaction data, and determines one or more combinations of the original features;
  • the feature combining unit 403 determines the feature vector of the new feature obtained by combining the original features based on each combination mode
  • the feature prediction unit 404 inputs the feature vector into the trained transaction feature generation model, and outputs the prediction result of the new feature;
  • the feature generating unit 405 selects several new features whose prediction results meet the specified conditions as the transaction features generated for the target data set.
  • the feature obtaining unit 402 obtains several original features of the same type in the transaction data, and determines a combination mode that matches the type as the combination mode of the original features.
  • the combination mode includes one or more of the following:
  • the combination mode includes one or more of the following:
  • the feature vector is generated based on the meta information of the original feature and the combination manner.
  • the meta information includes one or more of the following:
  • the mean value, variance, and number of unique data of the original feature are the mean value, variance, and number of unique data of the original feature.
  • the feature vector is generated based on the meta information of the original feature, the combination mode, and the meta feature of the data set .
  • the meta feature of the data set includes one or more of the following:
  • this specification also provides an embodiment of the training device for the transaction feature generation model.
  • the embodiment of the device for generating transaction features in this specification can be applied on a server.
  • the device embodiments can be implemented by software, or by hardware or a combination of software and hardware. Taking software implementation as an example, as a logical device, it is formed by reading the corresponding computer program instructions in the non-volatile memory into the memory through the processor of the server where it is located. From a hardware perspective, as shown in Figure 6, a hardware structure diagram of the server where the training device for the transaction feature generation model of this manual is located, except for the processor, memory, network interface, and non-volatile memory shown in Figure 6. In addition to the memory, the server where the device is located in the embodiment generally may include other hardware according to the actual function of the server, which is not repeated here.
  • Fig. 7 is a block diagram of a training device for a transaction feature generation model shown in an exemplary embodiment of this specification.
  • the training device 600 of the transaction feature generation model can be applied to the server shown in FIG. 6, and includes: a sample acquisition unit 601, a feature acquisition unit 602, a feature combination unit 603, a difference calculation unit 604, and Model training unit 605.
  • the sample obtaining unit 601 obtains a sample data set, the sample data set includes several pieces of sample transaction data with transaction tags, and the transaction tags are used to mark whether the corresponding sample transaction data is an illegal transaction;
  • the feature acquiring unit 602 acquires several original features of the sample transaction data, and determines one or more combinations of the original features;
  • the feature combining unit 603 determines the feature vector of the new feature obtained by combining the original features based on each combination mode
  • a difference calculation unit 604 which calculates the difference between the new feature and the transaction tag, as the feature tag of the new feature
  • the model training unit 605 trains the transaction feature generation model according to the feature vector of the new feature and the feature label.
  • the difference calculation unit 604 calculates the difference calculation unit 604:
  • the relevant part can refer to the part of the description of the method embodiment.
  • the device embodiments described above are merely illustrative.
  • the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in One place, or it can be distributed to multiple network units.
  • Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution in this specification. Those of ordinary skill in the art can understand and implement it without creative work.
  • a typical implementation device is a computer.
  • the specific form of the computer can be a personal computer, a laptop computer, a cellular phone, a camera phone, a smart phone, a personal digital assistant, a media player, a navigation device, an email receiving and sending device, and a game control A console, a tablet computer, a wearable device, or a combination of any of these devices.
  • this specification also provides a device for generating transaction features, which includes a processor and a memory for storing machine executable instructions.
  • the processor and the memory are usually connected to each other via an internal bus.
  • the device may also include an external interface to be able to communicate with other devices or components.
  • the processor is prompted to:
  • Target data set including several pieces of transaction data
  • the acquiring several original characteristics of the transaction data includes:
  • the method for determining the combination of the original features includes:
  • the combination mode matching the type is determined as the combination mode of the original feature.
  • the combination mode includes one or more of the following:
  • the combination mode includes one or more of the following:
  • the feature vector is generated based on the meta information of the original feature and the combination manner.
  • the meta information includes one or more of the following:
  • the mean value, variance, and number of unique data of the original feature are the mean value, variance, and number of unique data of the original feature.
  • the feature vector is generated based on the meta information of the original feature, the combination mode, and the meta feature of the data set .
  • the meta feature of the data set includes one or more of the following:
  • this specification also provides a training device for the transaction feature generation model.
  • the device includes a processor and a memory for storing machine executable instructions.
  • the processor and the memory are usually connected to each other via an internal bus.
  • the device may also include an external interface to be able to communicate with other devices or components.
  • the processor is prompted to:
  • sample data set including several pieces of sample transaction data with transaction tags, and the transaction tags are used to mark whether the corresponding sample transaction data is an illegal transaction
  • the acquiring several original characteristics of the sample transaction data includes:
  • the method for determining the combination of the original features includes:
  • the combination mode matching the type is determined as the combination mode of the original feature.
  • the calculating the difference between the new feature and the transaction tag includes:
  • this specification also provides a computer-readable storage medium on which a computer program is stored, and when the program is executed by a processor, the following steps are implemented:
  • Target data set including several pieces of transaction data
  • the acquiring several original characteristics of the transaction data includes:
  • the method for determining the combination of the original features includes:
  • the combination mode matching the type is determined as the combination mode of the original feature.
  • the combination mode includes one or more of the following:
  • the combination mode includes one or more of the following:
  • the feature vector is generated based on the meta information of the original feature and the combination manner.
  • the meta information includes one or more of the following:
  • the mean value, variance, and number of unique data of the original feature are the mean value, variance, and number of unique data of the original feature.
  • the feature vector is generated based on the meta information of the original feature, the combination mode, and the meta feature of the data set .
  • the meta feature of the data set includes one or more of the following:
  • this specification also provides a computer-readable storage medium with a computer program stored on the computer-readable storage medium, which implements the following steps when executed by a processor :
  • sample data set including several pieces of sample transaction data with transaction tags, and the transaction tags are used to mark whether the corresponding sample transaction data is an illegal transaction
  • the acquiring several original characteristics of the sample transaction data includes:
  • the method for determining the combination of the original features includes:
  • the combination mode matching the type is determined as the combination mode of the original feature.
  • the calculating the difference between the new feature and the transaction tag includes:

Landscapes

  • Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Theoretical Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

说明书披露一种交易特征生成模型的训练、交易特征的生成方法和装置。所述交易特征是生成方法可包括:获取目标数据集,所述目标数据集中包括若干条交易数据;获取所述交易数据的若干原始特征,并确定所述原始特征的一种或多种组合方式;确定基于每种组合方式对所述原始特征进行组合所得到的新特征的特征向量;将所述特征向量输入已训练的交易特征生成模型,输出所述新特征的预测结果;选取预测结果满足指定条件的若干新特征,作为为所述目标数据集生成的交易特征。

Description

交易特征生成模型的训练、交易特征的生成方法和装置 技术领域
本说明书涉及机器学习技术领域,尤其涉及一种交易特征生成模型的训练、交易特征的生成方法和装置。
背景技术
在机器学习技术领域,输入机器学习模型的特征往往依赖于人工的经验总结,这需要技术人员具有较多的先验知识,并且还需要消耗大量的时间进行验证,导致特征生成的效率低下。
发明内容
有鉴于此,本说明书提供一种交易特征生成模型的训练、交易特征的生成方法和装置。
具体地,本说明书是通过如下技术方案实现的:
一种交易特征的生成方法,所述交易特征用于对非法交易进行识别,所述方法包括:
获取目标数据集,所述目标数据集中包括若干条交易数据;
获取所述交易数据的若干原始特征,并确定所述原始特征的一种或多种组合方式;
确定基于每种组合方式对所述原始特征进行组合所得到的新特征的特征向量;
将所述特征向量输入已训练的交易特征生成模型,输出所述新特征的预测结果;
选取预测结果满足指定条件的若干新特征,作为为所述目标数据集生成的交易特征。
一种交易特征生成模型的训练方法,包括:
获取样本数据集,所述样本数据集中包括若干条带有交易标签的样本交易数据,所述交易标签用于标注对应的样本交易数据是否为非法交易;
获取所述样本交易数据的若干原始特征,并确定所述原始特征的一种或多种组合方式;
确定基于每种组合方式对所述原始特征进行组合所得到的新特征的特征向量;
计算所述新特征与所述交易标签的差异,作为所述新特征的特征标签;
根据所述新特征的特征向量和所述特征标签对交易特征生成模型进行训练。
一种特征生成方法,所述方法包括:
获取目标数据集,所述目标数据集中包括若干条数据;
获取所述数据的若干原始特征,并确定所述原始特征的一种或多种组合方式;
确定基于每种组合方式对所述原始特征进行组合所得到的新特征的特征向量;
将所述特征向量输入已训练的特征生成模型,输出所述新特征的预测结果;
选取预测结果满足指定条件的若干新特征,作为为所述目标数据集生成的特征。
一种交易特征的生成装置,所述交易特征用于对非法交易进行识别,所述装置包括:
数据集获取单元,获取目标数据集,所述目标数据集中包括若干条交易数据;
特征获取单元,获取所述交易数据的若干原始特征,并确定所述原始特征的一种或多种组合方式;
特征组合单元,确定基于每种组合方式对所述原始特征进行组合所得到的新特征的特征向量;
特征预测单元,将所述特征向量输入已训练的交易特征生成模型,输出所述新特征的预测结果;
特征生成单元,选取预测结果满足指定条件的若干新特征,作为为所述目标数据集生成的交易特征。
一种交易特征生成模型的训练装置,包括:
样本获取单元,获取样本数据集,所述样本数据集中包括若干条带有交易标签的样本交易数据,所述交易标签用于标注对应的样本交易数据是否为非法交易;
特征获取单元,获取所述样本交易数据的若干原始特征,并确定所述原始特征的一种或多种组合方式;
特征组合单元,确定基于每种组合方式对所述原始特征进行组合所得到的新特征 的特征向量;
差异计算单元,计算所述新特征与所述交易标签的差异,作为所述新特征的特征标签;
模型训练单元,根据所述新特征的特征向量和所述特征标签对交易特征生成模型进行训练。
一种交易特征的生成装置,包括:
处理器;
用于存储机器可执行指令的存储器;
其中,通过读取并执行所述存储器存储的与交易特征的生成逻辑对应的机器可执行指令,所述处理器被促使:
获取目标数据集,所述目标数据集中包括若干条交易数据;
获取所述交易数据的若干原始特征,并确定所述原始特征的一种或多种组合方式;
确定基于每种组合方式对所述原始特征进行组合所得到的新特征的特征向量;
将所述特征向量输入已训练的交易特征生成模型,输出所述新特征的预测结果;
选取预测结果满足指定条件的若干新特征,作为为所述目标数据集生成的交易特征。
一种交易特征生成模型的训练装置,包括:
处理器;
用于存储机器可执行指令的存储器;
其中,通过读取并执行所述存储器存储的与交易特征生成模型的训练逻辑对应的机器可执行指令,所述处理器被促使:
获取样本数据集,所述样本数据集中包括若干条带有交易标签的样本交易数据,所述交易标签用于标注对应的样本交易数据是否为非法交易;
获取所述样本交易数据的若干原始特征,并确定所述原始特征的一种或多种组合方式;
确定基于每种组合方式对所述原始特征进行组合所得到的新特征的特征向量;
计算所述新特征与所述交易标签的差异,作为所述新特征的特征标签;
根据所述新特征的特征向量和所述特征标签对交易特征生成模型进行训练。
由以上描述可以看出,本实施例可对数据的原始特征进行组合得到若干新特征,然后采用已训练的特征生成模型对所述新特征进行预测,并可选取预测结果满足指定条件的若干新特征作为新生成的特征,以用于后续的数据预测,从而实现特征的自动生成,极大的提高了特征生成的效率。
附图说明
图1是本说明书一示例性实施例示出的一种交易特征生成模型的训练方法的流程示意图。
图2是本说明书一示例性实施例示出的一种交易特征的生成方法的流程示意图。
图3是本说明书一示例性实施例示出的一种特征生成方法的流程示意图。
图4是本说明书一示例性实施例示出的一种用于交易特征的生成装置的一结构示意图。
图5是本说明书一示例性实施例示出的一种交易特征的生成装置的框图。
图6是本说明书一示例性实施例示出的一种用于交易特征生成模型的训练装置的一结构示意图。
图7是本说明书一示例性实施例示出的一种交易特征生成模型的训练装置的框图。
具体实施方式
这里将详细地对示例性实施例进行说明,其示例表示在附图中。下面的描述涉及附图时,除非另有表示,不同附图中的相同数字表示相同或相似的要素。以下示例性实施例中所描述的实施方式并不代表与本说明书相一致的所有实施方式。相反,它们仅是与如所附权利要求书中所详述的、本说明书的一些方面相一致的装置和方法的例子。
在本说明书使用的术语是仅仅出于描述特定实施例的目的,而非旨在限制本说明书。在本说明书和所附权利要求书中所使用的单数形式的“一种”、“所述”和“该”也旨在包括多数形式,除非上下文清楚地表示其他含义。还应当理解,本文中使用的术语“和/或”是指并包含一个或多个相关联的列出项目的任何或所有可能组合。
应当理解,尽管在本说明书可能采用术语第一、第二、第三等来描述各种信息,但这些信息不应限于这些术语。这些术语仅用来将同一类型的信息彼此区分开。例如,在不脱离本说明书范围的情况下,第一信息也可以被称为第二信息,类似地,第二信息也可以被称为第一信息。取决于语境,如在此所使用的词语“如果”可以被解释成为“在……时”或“当……时”或“响应于确定”。
本说明书提供一种交易特征的生成方案,可对交易数据的原始特征进行组合得到若干新特征,然后采用已训练的交易特征生成模型对所述新特征进行预测,并可选取预测结果满足指定条件的若干新特征作为新生成的交易特征,以用于后续的非法交易识别,从而实现交易特征的自动生成,极大的提高了交易特征生成的效率。
下面分别从交易特征生成模型的训练和交易特征的生成的两个方面来描述本说明书的具体实现过程。
一、交易特征生成模型的训练
图1是本说明书一示例性实施例示出的一种交易特征生成模型的训练方法的流程示意图。
请参考图1,所述交易特征生成模型的训练方法可包括以下步骤:
步骤102,获取样本数据集,所述样本数据集中包括若干条带有交易标签的样本交易数据,所述交易标签用于标注对应的样本交易数据是否为非法交易。
在本实施例中,可获取已进行过非法交易识别的交易数据作为样本交易数据,所述样本交易数据带有交易标签,所述交易标签用于标注对应的样本交易数据是否为非法交易。所述非法交易可包括:套现交易、洗钱交易等。
例如,交易标签“0”可表示对应的样本交易数据不是非法交易;交易标签“1”可表示对应的样本交易数据是非法交易等。
在本实施例中,可按照应用场景对样本交易数据进行划分,将同一应用场景中的样本交易数据划分到同一个样本数据集中。
例如,以某电商平台为例,可将通过该电商平台进行交易的样本交易数据划分到样本数据集1中。
再例如,以某消费信贷产品为例,可将通过该消费信贷产品进行支付的样本交易数据划分到样本数据集2中。
在本实施例中,可获取一个样本数据集,也可获取多个样本数据集,本说明书对此不作特殊限制。
步骤104,获取所述样本交易数据的若干原始特征,并确定所述原始特征的一种或多种组合方式。
在本实施例中,所述原始特征是所述样本交易数据自身的特征,例如:交易金额、交易次数、交易位置与商家距离、商家类别、用户类别等。
不同原始特征的类型可能相同,也可能不同。例如,交易金额的类型是数值型,交易次数也是数值型,商家类别的类型是字符串,用户类型也是字符串等。
其中,数值型还可细分为浮点型和整数型。
上述交易金额是浮点型,例如,交易金额108.75元等。
上述交易次数是整数型,例如,交易次数5次等。
本说明书不对原始特征类型的划分进行限制。
例如,可将浮点型和整数型统一归为数值型不进行区分。
再例如,可以浮点型、整数型等类别作为原始特征的类别,不保留数值型等。
在本实施例中,为便于特征组合,在获取样本数据集中的原始特征时,每次可获取若干类型相同的原始特征,获取的原始特征的数量可以是2,也可以是3等,本说明书对此不作特殊限制。
在一个例子中,可获取类型相同的若干原始特征。
以浮点型原始特征为例,可获取原始特征“上个月用户的交易总金额”以及原始特征“去年同期用户的交易总金额”。
以整数型原始特征为例,可获取原始特征“昨天交易次数”以及原始特征“前天交易次数”等。
在其他例子中,当类型不同的原始特征可以组合时,也可获取类型不同的若干原始特征。
例如,获取浮点型的原始特征“过去30天交易总金额”,以及整数型的原始特征“过去30天的交易次数”等。
值得注意的是,在进行原始特征获取时,如果两个原始特征无法组合,则不会同 时获取这两个原始特征。
例如,不会同时获取字符型原始特征“用户类别”和整数型原始特征“过去30天的交易次数”,二者无法组合。
在本实施例中,为便于描述,可将每次获取到的若干原始特征称为一个原始特征组,在本步骤中,可进行多次获取,得到多个原始特征组。
值得注意的是,本实施例中原始特征的获取通常以样本数据集为操作对象,不会跨样本数据集获取。换言之,每个原始特征组中的原始特征均来自同一个样本数据集。
在本实施例中,针对获取到的每个原始特征组,可为所述原始特征组中的原始特征确定匹配其类型的组合方式。
以数值型原始特征为例,所述组合方式可包括:加减乘除等四则运算、计算对数和、计算平方和等运算方式。所述组合方式还可以是先对原始特征进行大小排序,然后根据排序结果抽取部分原始特征进行四则运算、计算平方和等运算,本说明书对此不作特殊限制。
举例来说,假设某原始特征组中的原始特征为“昨天交易次数”和“前天交易次数”,则组合方式可以为减法,例如:采用“昨天交易次数”除以“前天交易次数”,可得到交易次数的差值。
以字符串类型的原始特征为例,所述组合方式可以是针对原始特征字符串长度的运算,例如字符串长度的四则运算等,具体运算方式可参考数值型原始特征的组合方式,本说明书在此不再一一赘述。
步骤106,确定基于每种组合方式对所述原始特征进行组合所得到的新特征的特征向量。
在本实施例中,针对前述步骤104获取到每个原始特征组,采用一种组合方式对所述原始特征组中的原始特征进行组合,就可得到一个新特征,采用各种组合方式对各原始特征组中的原始特征进行组合,可得到若干新特征。
在本步骤中,可确定每个新特征的特征向量。
在一个例子中,可基于组合方式对样本数据集中的原始特征进行组合,并可将组合得到的新特征的取值作为所述新特征的特征向量。
举例来说,假设从某样本数据集中获取到的原始特征组中包括原始特征a和原始特 征b两个数值类型的原始特征,组合方式是加法,该样本数据集中一共有100条交易数据,原始特征a的取值分别为[a 1,a 2,…,a 100],原始特征b的取值分别为[b 1,b 2,…,b 100],则采用加法对原始特征a和原始特征b进行组合后,得到的新特征a+b的特征向量可以为[a 1+b 1,a 2+b 2,…,a 100+b 100]。
在另一个例子中,为降低计算量,也可以基于原始特征的元信息和组合方式生成新特征的特征向量。
其中,所述原始特征的元信息可包括样本数据集中该原始特征的均值、方差、唯一数据个数等。每个原始特征的元信息可组成一个元信息向量。
所述组合方式的展现形式可以是一个0/1向量(Onehot向量),也可以是对0/1向量进行嵌入处理之后得到的低维向量。所述0/1向量是采用一个很长的向量来表示组合方式,该向量的维度是组合方式的总数量,每个维度对应唯一一种组合方式。若采用某种组合方式,则该组合方式所在维度的元素值是1;若未采用某种组合方式,则该组合方式所在维度的元素值是0。
假设,一共有80种组合方式,则组合方式的0/1向量有80维,每个维度对应唯一一种组合方式。仍以前述加法为例,该80维的0/1向量中,这加法对应维度的元素值是1,其他维度的元素值都是0,即该80维的0/1向量中有1个元素的元素值是1,其他剩余79个元素的元素值都是0。
仍以前述原始特征a和原始特征b为例,可基于原始特征a的元信息向量、原始特征b的元信息向量和组合方式的0/1向量生成新特征的特征向量。较为简单的,可将这三个向量拼接到一起得到新特征的特征向量。
在其他例子中,当样本数据集有多个,即样本数据集的数量大于1时,可基于原始特征的元信息、组合方式和样本数据集的元特征生成新特征的特征向量。
其中,所述样本数据集的元特征可包括:该样本数据集中被标注为合法交易的样本交易数据量与被标注为非法交易的样本交易数据量的比值(正负样本比例)、原始特征数量、数值型原始特征数量、字符串型原始特征数量等,所述样本数据集的元特征可组成一个元特征向量。
在本例中,可将原始特征的元信息向量、组合方式的0/1向量和样本数据集的元特征向量拼接到一起,得到新特征的特征向量。
步骤108,计算所述新特征与所述交易标签的差异,作为所述新特征的特征标签。
在本实施例中,组合得到的新特征与各样本交易数据交易标签之间的差异越小,可认为得到的新特征越可靠,由此本例中采用新特征和交易标签的差异作为新特征的特征标签。
在本实施例中,针对所述样本数据集中的各条样本交易数据,可根据所述组合方式,计算每条样本交易数据的新特征的取值,然后可将所述样本交易数据的交易标签作为真实值,计算所述样本交易数据的新特征的取值与所述真实值的均方误差,并将所述均方误差作为所述新特征与所述交易标签的差异。
在本实施例中,可采用如下公式计算新特征的特征标签:
Figure PCTCN2020070847-appb-000001
仍以对原始特征a和原始特征b进行组合为例,上述公式中C表示新特征ab的特征标签,N表示样本数据集中样本交易数据的数量,fab i表示第i条样本交易数据的新特征ab的取值,l i表示第i条样本交易数据的交易标签,其中i的取值是1到N。
在其他例子中,也可通过欧氏距离、马氏距离等算法来计算新特征与交易标签的差异,本说明书对此不作特殊限制。
步骤110,根据所述新特征的特征向量和所述特征标签对交易特征生成模型进行训练。
基于前述步骤106和108,可将新特征的特征向量输入交易特征生成模型,输出所述新特征的预测结果,然后可计算该预测结果和所述新特征的特征标签之间的差异,并基于该差异调整交易特征生成模型的参数,以实现对交易特征生成模型的训练。
在本实施例中,可采用有监督的机器学习模型作为所述交易特征生成模型,例如神经网络模型等,本说明书对此不作特殊限制。
二、交易特征的生成
图2是本说明书一示例性实施例示出的一种交易特征的生成方法的流程示意图。
请参考图2,所述交易特征的生成方法可包括以下步骤:
步骤202,获取目标数据集,所述目标数据集中包括若干条交易数据。
在本实施例中,针对需要进行非法交易识别的某个应用场景,可获取该应用场景下的交易数据,得到目标数据集。
例如,可获取该场景下已进行过非法交易识别的交易数据,得到目标数据集,该目标数据集中的交易数据带有交易标签。
再例如,当该场景下的交易数据未进行过非法交易识别时,也可以获取该场景下不带有交易标签的交易数据,得到目标数据集,本说明书对此不作特殊限制。
所述非法交易可包括:套现交易、洗钱交易等。
步骤204,获取所述交易数据的若干原始特征,并确定所述原始特征的一种或多种组合方式。
在本实施例中,所述原始特征是所述目标数据集中交易数据自身的特征,例如:交易金额、交易次数、交易位置与商家距离、商家类别等。
在本实施例中,可从所述目标数据集中获取若干组原始特征组,每组原始特征组中包括若干个原始特征,然后可分别为每个原始特征组中的原始特征确定组合方式。
上述原始特征的获取、组合方式的确定方法可参考前述图1所示实施例中的步骤104,本说明书在此不再一一赘述。
步骤206,确定基于每种组合方式对所述原始特征进行组合所得到的新特征的特征向量。
在本实施例中,针对前述步骤204获取到每个原始特征组,采用一种组合方式对所述原始特征组中的原始特征进行组合,就可得到一个新特征,采用各种组合方式对所有原始特征组中的原始特征进行组合,可得到若干新特征。
在本步骤中,可确定每个新特征的特征向量。
在一个例子中,可基于组合方式对样本数据集中的原始特征进行组合,并可将组合得到的新特征的取值作为所述新特征的特征向量。
在另一个例子中,也可以基于原始特征的元信息和组合方式生成新特征的特征向量。
在其他例子中,当用于训练交易特征生成模型的样本数据集的数量是多个时,可基于原始特征的元信息、组合方式和数据集的元特征生成新特征的特征向量。
所述数据集的元特征可包括:原始特征数量、数值型原始特征数量、字符串型原始特征数量、正负样本比例(如有)等。
上述特征向量的生成方式可参考前述图1所示实施例中步骤106中记载的特征向量生成方式,本说明书在此不再一一赘述。
步骤208,将所述特征向量输入已训练的交易特征生成模型,输出所述新特征的预测结果。
步骤210,选取预测结果满足指定条件的若干新特征,作为为所述目标数据集生成的交易特征。
在本实施例中,可将各新特征的特征向量输入已训练的交易特征生成模型,输出所述新特征的预测结果,然后可按照预测结果从小到大的顺序对各新特征进行排序。以采用欧式距离计算前述新特征与标签之间的差异为例,预测结果越小,说明新特征和交易标签的差异越小,新特征越可靠。
在本实施例中,在排序后,可选取排列在前N位的新特征作为为所述目标数据集生成的新的交易特征。其中,N的取值可预先设置。
在本实施例中,在为所述目标数据集确定新生成的交易特征后,可基于新生成的交易特征对所述目标数据集所属场景中的交易数据进行非法交易识别。
由以上描述可以看出,本说明书可对交易数据的原始特征进行组合得到若干新特征,然后采用已训练的交易特征生成模型对所述新特征进行预测,并可选取预测结果满足指定条件的若干新特征作为新生成的交易特征,以用于后续的非法交易识别,从而实现交易特征的自动生成,极大的提高了交易特征生成的效率。
本说明书还提供一种特征生成方法,可基于原始特征自动生成用于结果预测的新特征,提高特征生成效率。
请参考图3,所述特征生成方法可包括以下步骤:
步骤302,获取目标数据集,所述数据集中包括若干条数据。
在本实施例中,所述目标数据集是要进行新特征生成的数据集。在不同的预测任务中,所述目标数据集中的数据可能不同。
例如,在非法交易识别的任务中,所述目标数据集中的数据可以为交易数据,每条交易数据可包括交易金额、交易次数等原始特征。
再例如,在商品推荐的任务中,所述目标数据集中的数据可以为用户数据,每条用户数据可包括用户年龄、用户性别、用户历史购买记录、用户历史浏览记录等原始特 征。
又例如,在文本分类的任务中,所述目标数据集中的数据可以为文本数据,每条文本数据可包括对应文本所包括的词语、词频等原始特征。
又例如,在风险用户识别的任务中,所述目标数据集中的数据可以为用户数据,每条用户数据可包括用户年龄、用户性别、用户行为轨迹等原始特征。
步骤304,获取所述数据的若干原始特征,并确定所述原始特征的一种或多种组合方式。
在本实施例中,可从所述数据的各原始特征中获取若干类型相同的原始特征,并确定这些原始特征的组合方式。
当然,在其他例子中,也可从所述数据的各原始特征中获取若干类型并不完全相同的原始特征,并确定这些原始特征的组合方式。
其中,原始特征的获取和组合方式的确定可参考前述图2所示实施例中交易特征生成任务中原始特征的获取和组合方式的确定,本说明书在此不再一一赘述。
步骤306,确定基于每种组合方式对所述原始特征进行组合所得到的新特征的特征向量。
在本实施例中,针对前述步骤304获取到每组原始特征,采用一种组合方式对该组原始特征进行组合,就可得到一个新特征,采用各种组合方式对各组原始特征进行组合,可得到若干新特征。
在本步骤中,可确定每个新特征的特征向量。
其中,所述特征向量的确定方法可参考前述图1或图2所示的实施例,本说明书在此不再一一赘述。
步骤308,将所述特征向量输入已训练的特征生成模型,输出所述新特征的预测结果。
步骤310,选取预测结果满足指定条件的若干新特征,作为为所述目标数据集生成的特征。
在本实施例中,可采用与前述特征生成方法相同的方式对特征生成模型进行训练,得到已训练的特征生成模型。
在训练的过程中,可采用新特征与数据标签的差异作为新特征的特征标签,以对特征生成模型的参数进行调整。
具体的训练过程可参考前述图1所示的交易特征生成场景中交易特征生成模型的训练过程,本说明书在此不再一一赘述。
在本实施例中,可将各新特征的特征向量输入已训练的特征生成模型,输出所述新特征的预测结果,并可选取预测结果满足指定条件的若干新特征,用于后续数据的结果预测。
举例来说,假设预测任务是对用户进行风险识别,以识别出风险用户。
可先采用具有用户标签的一个或样本数据集来训练特征生成模型,得到已训练的特征生成模型。所述用户标签用于标记对应用户是否为风险用户
例如,可采用第一交易场景下的样本数据集1和第二交易场景下的样本数据集2来训练特征生成模型,所述样本数据集1和所述样本数据集2中的每条用户数据均带有用户标签。
在得到所述已训练的特征生成模型之后,针对要进行风险用户识别的第三交易场景,可从该第三交易场景中获取目标数据集,然后采用前述已训练的特征生成模型来为该场景预测新特征。在预测得到新特征之后,可采用新特征对该第三交易场景中的用户数据进行预测,以预测对应用户是否为风险用户。
由以上描述可以看出,本实施例可对数据的原始特征进行组合得到若干新特征,然后采用已训练的特征生成模型对所述新特征进行预测,并可选取预测结果满足指定条件的若干新特征作为新生成的特征,以用于后续的数据预测,从而实现特征的自动生成,极大的提高了特征生成的效率。
与前述交易特征的生成方法的实施例相对应,本说明书还提供了交易特征的生成装置的实施例。
本说明书交易特征的生成装置的实施例可以应用在服务器上。装置实施例可以通过软件实现,也可以通过硬件或者软硬件结合的方式实现。以软件实现为例,作为一个逻辑意义上的装置,是通过其所在服务器的处理器将非易失性存储器中对应的计算机程序指令读取到内存中运行形成的。从硬件层面而言,如图4所示,为本说明书交易特征的生成装置所在服务器的一种硬件结构图,除了图4所示的处理器、内存、网络接口、以及非易失性存储器之外,实施例中装置所在的服务器通常根据该服务器的实际功能, 还可以包括其他硬件,对此不再赘述。
图5是本说明书一示例性实施例示出的一种交易特征的生成装置的框图。
请参考图5,所述交易特征的生成装置400可以应用在前述图4所示的服务器中,包括有:数据集获取单元401、特征获取单元402、特征组合单元403、特征预测单元404以及特征生成单元405。
其中,数据集获取单元401,获取目标数据集,所述目标数据集中包括若干条交易数据;
特征获取单元402,获取所述交易数据的若干原始特征,并确定所述原始特征的一种或多种组合方式;
特征组合单元403,确定基于每种组合方式对所述原始特征进行组合所得到的新特征的特征向量;
特征预测单元404,将所述特征向量输入已训练的交易特征生成模型,输出所述新特征的预测结果;
特征生成单元405,选取预测结果满足指定条件的若干新特征,作为为所述目标数据集生成的交易特征。
可选的,所述特征获取单元402,获取所述交易数据中类型相同的若干原始特征,并确定与所述类型匹配的组合方式,作为所述原始特征的组合方式。
可选的,当所述类型是数值型时,所述组合方式包括以下一种或多种:
四则运算、计算对数和、计算平方和。
可选的,当所述类型是字符串时,所述组合方式包括以下一种或多种:
字符串长度的四则运算、字符串长度的对数和、字符串长度的平方和。
可选的,所述特征向量基于所述原始特征的元信息和所述组合方式生成。
可选的,所述元信息包括以下一种或多种:
所述原始特征的均值、方差、唯一数据个数。
可选的,当对所述交易特征生成模型进行训练的样本数据集的数量大于1时,所述特征向量基于所述原始特征的元信息、所述组合方式和所述数据集的元特征生成。
可选的,所述数据集的元特征包括以下一种或多种:
原始特征数量、数值型原始特征数量、正负样本比例。
与前述交易特征生成模型的训练方法的实施例相对应,本说明书还提供了交易特征生成模型的训练装置的实施例。
本说明书交易特征的生成装置的实施例可以应用在服务器上。装置实施例可以通过软件实现,也可以通过硬件或者软硬件结合的方式实现。以软件实现为例,作为一个逻辑意义上的装置,是通过其所在服务器的处理器将非易失性存储器中对应的计算机程序指令读取到内存中运行形成的。从硬件层面而言,如图6所示,为本说明书交易特征生成模型的训练装置所在服务器的一种硬件结构图,除了图6所示的处理器、内存、网络接口、以及非易失性存储器之外,实施例中装置所在的服务器通常根据该服务器的实际功能,还可以包括其他硬件,对此不再赘述。
图7是本说明书一示例性实施例示出的一种交易特征生成模型的训练装置的框图。
请参考图7,所述交易特征生成模型的训练装置600可以应用在前述图6所示的服务器中,包括有:样本获取单元601、特征获取单元602、特征组合单元603、差异计算单元604以及模型训练单元605。
其中,样本获取单元601,获取样本数据集,所述样本数据集中包括若干条带有交易标签的样本交易数据,所述交易标签用于标注对应的样本交易数据是否为非法交易;
特征获取单元602,获取所述样本交易数据的若干原始特征,并确定所述原始特征的一种或多种组合方式;
特征组合单元603,确定基于每种组合方式对所述原始特征进行组合所得到的新特征的特征向量;
差异计算单元604,计算所述新特征与所述交易标签的差异,作为所述新特征的特征标签;
模型训练单元605,根据所述新特征的特征向量和所述特征标签对交易特征生成模型进行训练。
可选的,所述差异计算单元604:
针对所述样本数据集中的各条样本交易数据,根据所述组合方式,计算所述样本交易数据的新特征的取值;
将所述样本交易数据的交易标签作为真实值,计算所述样本交易数据的新特征的 取值与所述真实值的均方误差,并将所述均方误差作为所述新特征与所述交易标签的差异。
上述装置中各个单元的功能和作用的实现过程具体详见上述方法中对应步骤的实现过程,在此不再赘述。
对于装置实施例而言,由于其基本对应于方法实施例,所以相关之处参见方法实施例的部分说明即可。以上所描述的装置实施例仅仅是示意性的,其中所述作为分离部件说明的单元可以是或者也可以不是物理上分开的,作为单元显示的部件可以是或者也可以不是物理单元,即可以位于一个地方,或者也可以分布到多个网络单元上。可以根据实际的需要选择其中的部分或者全部模块来实现本说明书方案的目的。本领域普通技术人员在不付出创造性劳动的情况下,即可以理解并实施。
上述实施例阐明的系统、装置、模块或单元,具体可以由计算机芯片或实体实现,或者由具有某种功能的产品来实现。一种典型的实现设备为计算机,计算机的具体形式可以是个人计算机、膝上型计算机、蜂窝电话、相机电话、智能电话、个人数字助理、媒体播放器、导航设备、电子邮件收发设备、游戏控制台、平板计算机、可穿戴设备或者这些设备中的任意几种设备的组合。
与前述交易特征的生成方法的实施例相对应,本说明书还提供一种交易特征的生成装置,该装置包括:处理器以及用于存储机器可执行指令的存储器。其中,处理器和存储器通常借由内部总线相互连接。在其他可能的实现方式中,所述设备还可能包括外部接口,以能够与其他设备或者部件进行通信。
在本实施例中,通过读取并执行所述存储器存储的与交易特征的生成逻辑对应的机器可执行指令,所述处理器被促使:
获取目标数据集,所述目标数据集中包括若干条交易数据;
获取所述交易数据的若干原始特征,并确定所述原始特征的一种或多种组合方式;
确定基于每种组合方式对所述原始特征进行组合所得到的新特征的特征向量;
将所述特征向量输入已训练的交易特征生成模型,输出所述新特征的预测结果;
选取预测结果满足指定条件的若干新特征,作为为所述目标数据集生成的交易特征。
可选的,所述获取所述交易数据的若干原始特征,包括:
获取所述交易数据中类型相同的若干原始特征;
所述确定所述原始特征的组合方式,包括:
确定与所述类型匹配的组合方式,作为所述原始特征的组合方式。
可选的,当所述类型是数值型时,所述组合方式包括以下一种或多种:
四则运算、计算对数和、计算平方和。
可选的,当所述类型是字符串时,所述组合方式包括以下一种或多种:
字符串长度的四则运算、字符串长度的对数和、字符串长度的平方和。
可选的,所述特征向量基于所述原始特征的元信息和所述组合方式生成。
可选的,所述元信息包括以下一种或多种:
所述原始特征的均值、方差、唯一数据个数。
可选的,当对所述交易特征生成模型进行训练的样本数据集的数量大于1时,所述特征向量基于所述原始特征的元信息、所述组合方式和所述数据集的元特征生成。
可选的,所述数据集的元特征包括以下一种或多种:
原始特征数量、数值型原始特征数量、正负样本比例。
与前述交易特征生成模型的训练方法的实施例相对应,本说明书还提供一种交易特征生成模型的训练装置,该装置包括:处理器以及用于存储机器可执行指令的存储器。其中,处理器和存储器通常借由内部总线相互连接。在其他可能的实现方式中,所述设备还可能包括外部接口,以能够与其他设备或者部件进行通信。
在本实施例中,通过读取并执行所述存储器存储的与交易特征生成模型的训练逻辑对应的机器可执行指令,所述处理器被促使:
获取样本数据集,所述样本数据集中包括若干条带有交易标签的样本交易数据,所述交易标签用于标注对应的样本交易数据是否为非法交易;
获取所述样本交易数据的若干原始特征,并确定所述原始特征的一种或多种组合方式;
确定基于每种组合方式对所述原始特征进行组合所得到的新特征的特征向量;
计算所述新特征与所述交易标签的差异,作为所述新特征的特征标签;
根据所述新特征的特征向量和所述特征标签对交易特征生成模型进行训练。
可选的,所述获取所述样本交易数据的若干原始特征,包括:
获取所述样本交易数据中类型相同的若干原始特征;
所述确定所述原始特征的组合方式,包括:
确定与所述类型匹配的组合方式,作为所述原始特征的组合方式。
可选的,所述计算所述新特征与所述交易标签的差异,包括:
针对所述样本数据集中的各条样本交易数据,根据所述组合方式,计算所述样本交易数据的新特征的取值;
将所述样本交易数据的交易标签作为真实值,计算所述样本交易数据的新特征的取值与所述真实值的均方误差,并将所述均方误差作为所述新特征与所述交易标签的差异。
与前述交易特征的生成方法的实施例相对应,本说明书还提供一种计算机可读存储介质,所述计算机可读存储介质上存储有计算机程序,该程序被处理器执行时实现以下步骤:
获取目标数据集,所述目标数据集中包括若干条交易数据;
获取所述交易数据的若干原始特征,并确定所述原始特征的一种或多种组合方式;
确定基于每种组合方式对所述原始特征进行组合所得到的新特征的特征向量;
将所述特征向量输入已训练的交易特征生成模型,输出所述新特征的预测结果;
选取预测结果满足指定条件的若干新特征,作为为所述目标数据集生成的交易特征。
可选的,所述获取所述交易数据的若干原始特征,包括:
获取所述交易数据中类型相同的若干原始特征;
所述确定所述原始特征的组合方式,包括:
确定与所述类型匹配的组合方式,作为所述原始特征的组合方式。
可选的,当所述类型是数值型时,所述组合方式包括以下一种或多种:
四则运算、计算对数和、计算平方和。
可选的,当所述类型是字符串时,所述组合方式包括以下一种或多种:
字符串长度的四则运算、字符串长度的对数和、字符串长度的平方和。
可选的,所述特征向量基于所述原始特征的元信息和所述组合方式生成。
可选的,所述元信息包括以下一种或多种:
所述原始特征的均值、方差、唯一数据个数。
可选的,当对所述交易特征生成模型进行训练的样本数据集的数量大于1时,所述特征向量基于所述原始特征的元信息、所述组合方式和所述数据集的元特征生成。
可选的,所述数据集的元特征包括以下一种或多种:
原始特征数量、数值型原始特征数量、正负样本比例。
与前述交易特征生成模型的训练方法的实施例相对应,本说明书还提供一种计算机可读存储介质,所述计算机可读存储介质上存储有计算机程序,该程序被处理器执行时实现以下步骤:
获取样本数据集,所述样本数据集中包括若干条带有交易标签的样本交易数据,所述交易标签用于标注对应的样本交易数据是否为非法交易;
获取所述样本交易数据的若干原始特征,并确定所述原始特征的一种或多种组合方式;
确定基于每种组合方式对所述原始特征进行组合所得到的新特征的特征向量;
计算所述新特征与所述交易标签的差异,作为所述新特征的特征标签;
根据所述新特征的特征向量和所述特征标签对交易特征生成模型进行训练。
可选的,所述获取所述样本交易数据的若干原始特征,包括:
获取所述样本交易数据中类型相同的若干原始特征;
所述确定所述原始特征的组合方式,包括:
确定与所述类型匹配的组合方式,作为所述原始特征的组合方式。
可选的,所述计算所述新特征与所述交易标签的差异,包括:
针对所述样本数据集中的各条样本交易数据,根据所述组合方式,计算所述样本交易数据的新特征的取值;
将所述样本交易数据的交易标签作为真实值,计算所述样本交易数据的新特征的取值与所述真实值的均方误差,并将所述均方误差作为所述新特征与所述交易标签的差异。
上述对本说明书特定实施例进行了描述。其它实施例在所附权利要求书的范围内。在一些情况下,在权利要求书中记载的动作或步骤可以按照不同于实施例中的顺序来执行并且仍然可以实现期望的结果。另外,在附图中描绘的过程不一定要求示出的特定顺序或者连续顺序才能实现期望的结果。在某些实施方式中,多任务处理和并行处理也是可以的或者可能是有利的。
以上所述仅为本说明书的较佳实施例而已,并不用以限制本说明书,凡在本说明书的精神和原则之内,所做的任何修改、等同替换、改进等,均应包含在本说明书保护的范围之内。

Claims (19)

  1. 一种交易特征的生成方法,所述交易特征用于对非法交易进行识别,所述方法包括:
    获取目标数据集,所述目标数据集中包括若干条交易数据;
    获取所述交易数据的若干原始特征,并确定所述原始特征的一种或多种组合方式;
    确定基于每种组合方式对所述原始特征进行组合所得到的新特征的特征向量;
    将所述特征向量输入已训练的交易特征生成模型,输出所述新特征的预测结果;
    选取预测结果满足指定条件的若干新特征,作为为所述目标数据集生成的交易特征。
  2. 根据权利要求1所述的方法,所述获取所述交易数据的若干原始特征,包括:
    获取所述交易数据中类型相同的若干原始特征;
    所述确定所述原始特征的组合方式,包括:
    确定与所述类型匹配的组合方式,作为所述原始特征的组合方式。
  3. 根据权利要求2所述的方法,
    当所述类型是数值型时,所述组合方式包括以下一种或多种:
    四则运算、计算对数和、计算平方和。
  4. 根据权利要求2所述的方法,
    当所述类型是字符串时,所述组合方式包括以下一种或多种:
    字符串长度的四则运算、字符串长度的对数和、字符串长度的平方和。
  5. 根据权利要求1所述的方法,
    所述特征向量基于所述原始特征的元信息和所述组合方式生成。
  6. 根据权利要求5所述的方法,所述元信息包括以下一种或多种:
    所述原始特征的均值、方差、唯一数据个数。
  7. 根据权利要求1所述的方法,
    当对所述交易特征生成模型进行训练的样本数据集的数量大于1时,所述特征向量基于所述原始特征的元信息、所述组合方式和所述数据集的元特征生成。
  8. 根据权利要求7所述的方法,所述数据集的元特征包括以下一种或多种:
    原始特征数量、数值型原始特征数量、正负样本比例。
  9. 一种交易特征生成模型的训练方法,包括:
    获取样本数据集,所述样本数据集中包括若干条带有交易标签的样本交易数据,所述交易标签用于标注对应的样本交易数据是否为非法交易;
    获取所述样本交易数据的若干原始特征,并确定所述原始特征的一种或多种组合方 式;
    确定基于每种组合方式对所述原始特征进行组合所得到的新特征的特征向量;
    计算所述新特征与所述交易标签的差异,作为所述新特征的特征标签;
    根据所述新特征的特征向量和所述特征标签对交易特征生成模型进行训练。
  10. 根据权利要求9所述的方法,所述获取所述样本交易数据的若干原始特征,包括:
    获取所述样本交易数据中类型相同的若干原始特征;
    所述确定所述原始特征的组合方式,包括:
    确定与所述类型匹配的组合方式,作为所述原始特征的组合方式。
  11. 根据权利要求9所述的方法,所述计算所述新特征与所述交易标签的差异,包括:
    针对所述样本数据集中的各条样本交易数据,根据所述组合方式,计算所述样本交易数据的新特征的取值;
    将所述样本交易数据的交易标签作为真实值,计算所述样本交易数据的新特征的取值与所述真实值的均方误差,并将所述均方误差作为所述新特征与所述交易标签的差异。
  12. 一种特征生成方法,所述方法包括:
    获取目标数据集,所述目标数据集中包括若干条数据;
    获取所述数据的若干原始特征,并确定所述原始特征的一种或多种组合方式;
    确定基于每种组合方式对所述原始特征进行组合所得到的新特征的特征向量;
    将所述特征向量输入已训练的特征生成模型,输出所述新特征的预测结果;
    选取预测结果满足指定条件的若干新特征,作为为所述目标数据集生成的特征。
  13. 根据权利要求12所述的方法,所述获取所述数据的若干原始特征,包括:
    获取所述数据中类型相同的若干原始特征;
    所述确定所述原始特征的组合方式,包括:
    确定与所述类型匹配的组合方式,作为所述原始特征的组合方式。
  14. 一种交易特征的生成装置,所述交易特征用于对非法交易进行识别,所述装置包括:
    数据集获取单元,获取目标数据集,所述目标数据集中包括若干条交易数据;
    特征获取单元,获取所述交易数据的若干原始特征,并确定所述原始特征的一种或多种组合方式;
    特征组合单元,确定基于每种组合方式对所述原始特征进行组合所得到的新特征的 特征向量;
    特征预测单元,将所述特征向量输入已训练的交易特征生成模型,输出所述新特征的预测结果;
    特征生成单元,选取预测结果满足指定条件的若干新特征,作为为所述目标数据集生成的交易特征。
  15. 根据权利要求14所述的装置,
    所述特征获取单元,获取所述交易数据中类型相同的若干原始特征,并确定与所述类型匹配的组合方式,作为所述原始特征的组合方式。
  16. 一种交易特征生成模型的训练装置,包括:
    样本获取单元,获取样本数据集,所述样本数据集中包括若干条带有交易标签的样本交易数据,所述交易标签用于标注对应的样本交易数据是否为非法交易;
    特征获取单元,获取所述样本交易数据的若干原始特征,并确定所述原始特征的一种或多种组合方式;
    特征组合单元,确定基于每种组合方式对所述原始特征进行组合所得到的新特征的特征向量;
    差异计算单元,计算所述新特征与所述交易标签的差异,作为所述新特征的特征标签;
    模型训练单元,根据所述新特征的特征向量和所述特征标签对交易特征生成模型进行训练。
  17. 根据权利要求16所述的装置,所述差异计算单元:
    针对所述样本数据集中的各条样本交易数据,根据所述组合方式,计算所述样本交易数据的新特征的取值;
    将所述样本交易数据的交易标签作为真实值,计算所述样本交易数据的新特征的取值与所述真实值的均方误差,并将所述均方误差作为所述新特征与所述交易标签的差异。
  18. 一种交易特征的生成装置,包括:
    处理器;
    用于存储机器可执行指令的存储器;
    其中,通过读取并执行所述存储器存储的与交易特征的生成逻辑对应的机器可执行指令,所述处理器被促使:
    获取目标数据集,所述目标数据集中包括若干条交易数据;
    获取所述交易数据的若干原始特征,并确定所述原始特征的一种或多种组合方式;
    确定基于每种组合方式对所述原始特征进行组合所得到的新特征的特征向量;
    将所述特征向量输入已训练的交易特征生成模型,输出所述新特征的预测结果;
    选取预测结果满足指定条件的若干新特征,作为为所述目标数据集生成的交易特征。
  19. 一种交易特征生成模型的训练装置,包括:
    处理器;
    用于存储机器可执行指令的存储器;
    其中,通过读取并执行所述存储器存储的与交易特征生成模型的训练逻辑对应的机器可执行指令,所述处理器被促使:
    获取样本数据集,所述样本数据集中包括若干条带有交易标签的样本交易数据,所述交易标签用于标注对应的样本交易数据是否为非法交易;
    获取所述样本交易数据的若干原始特征,并确定所述原始特征的一种或多种组合方式;
    确定基于每种组合方式对所述原始特征进行组合所得到的新特征的特征向量;
    计算所述新特征与所述交易标签的差异,作为所述新特征的特征标签;
    根据所述新特征的特征向量和所述特征标签对交易特征生成模型进行训练。
PCT/CN2020/070847 2019-05-29 2020-01-08 交易特征生成模型的训练、交易特征的生成方法和装置 WO2020238229A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US16/803,150 US11210673B2 (en) 2019-05-29 2020-02-27 Transaction feature generation

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201910457803.2 2019-05-29
CN201910457803.2A CN110263821B (zh) 2019-05-29 2019-05-29 交易特征生成模型的训练、交易特征的生成方法和装置

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US16/803,150 Continuation US11210673B2 (en) 2019-05-29 2020-02-27 Transaction feature generation

Publications (1)

Publication Number Publication Date
WO2020238229A1 true WO2020238229A1 (zh) 2020-12-03

Family

ID=67915812

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/070847 WO2020238229A1 (zh) 2019-05-29 2020-01-08 交易特征生成模型的训练、交易特征的生成方法和装置

Country Status (2)

Country Link
CN (1) CN110263821B (zh)
WO (1) WO2020238229A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112861123A (zh) * 2021-01-26 2021-05-28 中山大学 一种比特币恶意地址识别方法及装置

Families Citing this family (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US11210673B2 (en) 2019-05-29 2021-12-28 Advanced New Technologies Co., Ltd. Transaction feature generation
CN110263821B (zh) * 2019-05-29 2020-08-14 阿里巴巴集团控股有限公司 交易特征生成模型的训练、交易特征的生成方法和装置
CN110717182A (zh) * 2019-10-14 2020-01-21 杭州安恒信息技术股份有限公司 一种网页木马检测方法、装置、设备及可读存储介质
CN110782349A (zh) * 2019-10-25 2020-02-11 支付宝(杭州)信息技术有限公司 一种模型训练方法和系统
CN110852809A (zh) * 2019-11-15 2020-02-28 腾讯科技(深圳)有限公司 一种数据处理方法、装置、设备及介质
CN111553759A (zh) * 2020-03-25 2020-08-18 平安科技(深圳)有限公司 一种产品信息推送方法、装置、设备及存储介质
CN111861704A (zh) * 2020-07-10 2020-10-30 深圳无域科技技术有限公司 风控特征生成方法及系统
CN114943607A (zh) * 2022-06-02 2022-08-26 支付宝(杭州)信息技术有限公司 特征发现方法、属性预测方法和装置

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US9336494B1 (en) * 2012-08-20 2016-05-10 Context Relevant, Inc. Re-training a machine learning model
CN106709508A (zh) * 2016-11-29 2017-05-24 江南大学 一种利用特征信息的加权典型相关分析方法
CN109819282A (zh) * 2017-11-22 2019-05-28 腾讯科技(深圳)有限公司 一种视频用户类别识别方法、装置和介质
CN110263821A (zh) * 2019-05-29 2019-09-20 阿里巴巴集团控股有限公司 交易特征生成模型的训练、交易特征的生成方法和装置

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US9336494B1 (en) * 2012-08-20 2016-05-10 Context Relevant, Inc. Re-training a machine learning model
CN106709508A (zh) * 2016-11-29 2017-05-24 江南大学 一种利用特征信息的加权典型相关分析方法
CN109819282A (zh) * 2017-11-22 2019-05-28 腾讯科技(深圳)有限公司 一种视频用户类别识别方法、装置和介质
CN110263821A (zh) * 2019-05-29 2019-09-20 阿里巴巴集团控股有限公司 交易特征生成模型的训练、交易特征的生成方法和装置

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112861123A (zh) * 2021-01-26 2021-05-28 中山大学 一种比特币恶意地址识别方法及装置
CN112861123B (zh) * 2021-01-26 2022-02-08 中山大学 一种比特币恶意地址识别方法及装置

Also Published As

Publication number Publication date
CN110263821A (zh) 2019-09-20
CN110263821B (zh) 2020-08-14

Similar Documents

Publication Publication Date Title
WO2020238229A1 (zh) 交易特征生成模型的训练、交易特征的生成方法和装置
CN111177569B (zh) 基于人工智能的推荐处理方法、装置及设备
WO2019223379A1 (zh) 一种产品推荐方法和装置
US8589317B2 (en) Human-assisted training of automated classifiers
TW201946013A (zh) 基於lstm模型的信用風險預測方法及裝置
TW201905773A (zh) 車險風險預測方法、裝置及伺服器
US11210673B2 (en) Transaction feature generation
CN112529663B (zh) 商品推荐方法、装置、终端设备及存储介质
CN111783039B (zh) 风险确定方法、装置、计算机系统和存储介质
CN108734587A (zh) 金融产品的推荐方法及终端设备
CN110647683B (zh) 一种信息推荐方法、装置
US20120166307A1 (en) Determination of permissibility associated with e-commerce transactions
CN107679856A (zh) 基于交易的业务控制方法和装置
CN108804564A (zh) 金融产品的组合推荐方法及终端设备
CN109191133A (zh) 支付渠道的选择方法及终端设备
CN108228684A (zh) 聚类模型的训练方法、装置、电子设备和计算机存储介质
US20150142511A1 (en) Recommending and pricing datasets
WO2020177477A1 (zh) 一种信用服务推荐方法、装置及设备
WO2020155831A1 (zh) 数据标签生成、模型训练、事件识别方法和装置
CN111754287A (zh) 物品筛选方法、装置、设备和存储介质
CN109885745A (zh) 一种用户画像方法、装置、可读存储介质及终端设备
CN109886299A (zh) 一种用户画像方法、装置、可读存储介质及终端设备
CN112446777A (zh) 一种信用评估方法、装置、设备及存储介质
CN110008352A (zh) 实体发现方法及装置
CN114529399A (zh) 用户数据处理方法、装置、计算机设备和存储介质

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: 20813791

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20813791

Country of ref document: EP

Kind code of ref document: A1