CN109919436A - A kind of promise breaking user's probability forecasting method based on sparse features insertion - Google Patents
A kind of promise breaking user's probability forecasting method based on sparse features insertion Download PDFInfo
- Publication number
- CN109919436A CN109919436A CN201910084188.5A CN201910084188A CN109919436A CN 109919436 A CN109919436 A CN 109919436A CN 201910084188 A CN201910084188 A CN 201910084188A CN 109919436 A CN109919436 A CN 109919436A
- Authority
- CN
- China
- Prior art keywords
- user
- data
- class
- feature
- variable
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Landscapes
- Financial Or Insurance-Related Operations Such As Payment And Settlement (AREA)
Abstract
A kind of promise breaking user's probability forecasting method based on sparse features insertion of the present invention: being converted to characteristics of variables for the initial data of user first, then (is similar to one-hot to handle) into a sparse matrix by the multi-class variable mappings in characteristics of variables；On this basis, which is mapped to by probability by basic decision-tree model, then increased in model using the probability as feature, to carry out promise breaking user in predicting.A kind of promise breaking user's probability forecasting method based on sparse features insertion of the present invention, compared with prior art, the advantage is that: effectively increasing the processing capacity of classification coding, the dimension for effectively reducing feature space in the process of subsequent machine learning simultaneously, is conducive to the study and processing of machine learning model.
Description
Technical field
A kind of promise breaking user's probability forecasting method based on sparse features insertion of the present invention is related to user's letter of financial field
With risk assessment technology, and in particular to a kind of promise breaking user probabilistic forecasting side towards property management company's consumer finance field
Method.
Background technique
In recent years, the internet financing corporation of business continues to bring out based on P2P debt-credit, consumer finance etc., in tradition
The field that financial industry can not be related to establishes a huge new industry.But various negative press also come one after another simultaneously,
To the future cloud of these internet financial companies.Wherein, air control ability is always these emerging technologies finance
The gate of vitality of company only possesses good air control technology, is likely to develop in a healthy way in this strand of tide.Traditional air control means rely on
In Central Bank's credit investigation system, there is also certain defects and problem.From the point of view of objective group's angle, P2P loan platform, consumer finance company face
To object be not mostly the target visitor group of Bank Retail Business, thus also lack the credit of this kind of client in Central Bank's credit investigation system
Information；In addition, Central Bank's credit investigation system also relies on the historical data of banking system upload, and this kind of data deficiency timeliness is right
Also just show slightly insufficient in the assessment of the current credit risk of client.
With the fast development and application of the technologies such as current machine study, deep learning and theory, it can be from multidimensional degree
It is set out according to (shopping, communication, trip etc.) to reflect user behavior, user's portrait is portrayed, thus the credit wind of auxiliary judgment user
Danger.Nowadays, domestic well-known Internet company all relies on its data accumulation abundant, excavates the hiding abundant letter in its behind
Breath, so as to complete the credit scoring to user.Such as ant gold clothes rely on user in the shopping number of the platforms such as Taobao, day cat
According to release sesame credit score, similar third party also releases personal credit point；Furthermore China Mobile, China Unicom etc. are based on using
Family communication behavior data carry out credit scoring to user.This credit estimation method based on various dimensions information is also internet
Financing corporation's air control provides a kind of new approaches.(include in the data of third-party platform by crawl user under user's authorization
Shopping, communication, trip etc.), in conjunction with the historical data of the accumulation of itself, complete air control modeling.But these methods are to sparse
The treatment effect of data is all less desirable.
Based on problem above, the present invention proposes a kind of promise breaking user's probability forecasting method based on sparse features insertion, from
The data informations such as certification, shopping of the user on third-party platform are started with, and are first passed around data cleansing and are pre-processed to data,
Then in handling multi-class multivariable process, the sparse features embedding grammar based on proposition, by by the multiclass of no continuous information
Other variable switchs to sparse matrix, and learns the information of sparse matrix by machine learning model, and sparse matrix is finally passed through machine
Device learning model is mapped to dense probabilistic information, to carry out promise breaking user in predicting.
Summary of the invention
It is an object of the invention to propose a kind of promise breaking user's probability forecasting method of sparse features insertion, by machine learning
In technical application to credit estimation method, with promoting technology business development, tested by the information analysis, filtering, intersection of various dimensions
A comprehensive user data portrait is demonstrate,proved and is aggregated to form, auxiliary activities personnel audit the credit risk for judging user, greatly
The efficiency and accuracy of audit are improved, thus the manual skill for substituting a large amount of cost of labor and falling behind.
To achieve the goals above, a kind of promise breaking user's probability forecasting method based on sparse features insertion of the present invention, is adopted
With following technical solution:
The initial data of user is converted to characteristics of variables first by the present invention, then by the multi-class variable in characteristics of variables
It is mapped in a sparse matrix and (is similar to one-hot to handle)；On this basis, by basic decision-tree model that this is sparse
Matrix is mapped to probability, then increases in model using the probability as feature, to carry out promise breaking user in predicting.The present invention mainly wraps
The data cleansing to initial data, Feature Engineering two parts based on machine learning are included, specific as follows:
One, data cleansing
The part converts different dimensional in the initial data (including shopping, communication, trip etc.) of third-party platform for user
Characteristics of variables under degree, in addition to traditional based on nearest consumption time, the nearest consumer finance, the variables such as nearest consuming frequency
Except, data are integrally divided into the dimensions such as behavioral data, consumption data, essential information data according to the data of offer, are passed through
These data are analyzed, converts, ultimately forms the feature of various dimensions multivariable.
There may be store lack of standardization, field disunity, Chinese and English mixing, shortage of data, multi-class for initial data simultaneously
The problems such as variable, converts regular data for initial data using data cleansing, specific technical solution is such as these problems
Under:
1.1 field processing lack of standardization
There may be part messy code data and data lack of standardization for initial data.For messy code data, using delete processing；It is right
In storing nonstandard data, the data of its nonstandardized technique are converted to unified canonical form.
The processing of 1.2 missing datas
There may be a large amount of absent fields in initial data, and for different deletion conditions, there are different data cleansing sides
Formula.Such as the field little for missing ratio, it can be using the methods of filling mean, mode, median " polishing " field；
And for the field that data item largely lacks, each statistics available user's absent field number is as a feature, so not only
The information of missing is remained to a certain extent, it is ensured that the stability of data distribution.
The processing of 1.3 class variables
Client gender (male, female) in addition, there is also a large amount of class variables in initial data, such as in user message table,
Membership grade (common, bronze medal, silver medal, gold medal, diamond member), bank card types (deposit card, credit card) etc., can pass through mapping
Mode class variable is converted.The present invention uses distinct methods for different classifications, for different values without specific excellent
The bad classification divided uses one-hot coding method, such as client gender；For different values, there are the classifications of notable difference to adopt
With numeric type coding method, such as membership grade.Wherein One-Hot is encoded, and also known as an efficient coding, method are using N
Bit status register encodes N number of state, and each state has register-bit independent, and when any
It waits, wherein only one effective.Numeric type coding is then the classification that appearance is directly corresponded to a number.Such as color
In each class variable such as " red ", " green " and " yellow ", be respectively after being encoded using one-hot [1,0,0],
Three vectors in [0,1,0], [0,0,1], and what use numeric type encoded is 1,2,3 three numbers.
Two, based on the Feature Engineering of machine learning
Original characteristics of variables is converted to the training data of model by the part by Feature Engineering processing.Base of the invention
It is divided into two parts in the Feature Engineering of machine learning, is the multi-class of traditional characteristic engineering and the decision Tree algorithms that are proposed respectively
Variable processing method.Primitive character can generate new feature after the processing of traditional characteristic engineering, but part new feature is sparse
Feature, cannot be directly as the training data of model, therefore proposes the multi-class variable side based on machine learning through the invention
Sparse features are converted to one-dimensional characteristic by method, so as to the training data directly as model.Specific technical solution is as follows:
2.1 traditional characteristic engineerings
Variable in primitive character is carried out feature according to time class, amount of money class, address class, telephone number class respectively to mention
It takes, the derivative work of variable.Detailed process is as follows:
2.1.1 time class field:
It include a large amount of time class field in primitive character, these time class fields can reflect user in specific time
Activity periods situation.Therefore reasonable feature is carried out based on time variable to derive, user's row can be portrayed from time dimension
For thus auxiliary judgment user credit superiority and inferiority.The time class field of user is divided according to certain time interval, is counted
The behavior of user in the time interval, to assist portraying user's portrait.Such as by the user authentication time respectively according to year, month, day,
Zhou Jinhang statistics divides, and calculates its authenticated time and loan application time difference, in general the more early user of authenticated time
It is more credible.
2.1.2 amount of money class field
Amount of money class field is that most can directly react the information of user's economic capability in modeling.Amount of money class field directly reflects
The level of consumption of user, such as the shopping amount of money of user, in general its amount of money of doing shopping are bigger, loan user's economic capability
Also stronger, Default Probability is also lower.
2.1.3 address class field
Address information equally can help auditor from another angle to describe user.Address category information is drawn
Point, the address class user portrait of the available user.Such as its inhabitation address of common user can be more stable, shopping
Shipping address it is also relatively stable, be distributed in job site and place of abode.And for the part unemployed, it is unstable
Inhabitation address, or frequently more change jobs, its replacement frequency is also greater than other users from the point of view of shipping address.
2.1.4 telephone number class field
Telephone number information also reflects the partial information of user.Telephone number has different operators, and phone number
Number, the frequency of use etc. of code can all influence the credit default risk of a user.For example telephone number frequency of use is too low, says
Bright user's potentially unstable, it is easier to generate violations.
The 2.2 multi-class variable processing methods based on decision Tree algorithms
Above-mentioned tradition class variable mostly uses the form of direct digital encoding to be put into machine learning model and is handled, still
The problems such as this method introduces noise there are variable, by by variable mappings to higher dimensional space, by the way of register occupy-place into
The processing of row variable can be by the sequence noise remove of variable, but the sparse matrix of higher-dimension has message difficulty, therefore ties
Reason of both closing, can be converted to new feature for initial data after features described above engineering.For these multi-class changes
The sparse features of amount, the present invention propose a kind of method for handling various dimensions variable, the i.e. multi-class variable based on decision Tree algorithms
Sparse features are converted to one-dimensional characteristic by processing method, and detailed process is as follows: multi-class variable ω being carried out one-hot first
Coding, obtains a sparse matrix H, and dimension is equal to the different classes of quantity of ω in former data.Then by sparse matrix H and
It after label information Y, that is, user promise breaking information combines, is fitted using decision-tree model F, the probability value P of model output is made
It is put into following model F ' for new feature, is mapped sparse matrix H for one-dimensional variable P by this step, and as
Feature increases in model.This multi-class information is not only remained in this way, but also reduces the variable dimension again.This method is logical
It crosses machine learning model sparse matrix is reconstructed, reduces final feature while by model by sparse matrix dimensionality reduction
Dimension further reduced the complexity for the mapping that entire machine learning model learns, and effectively reduce the risk of over-fitting.
One-Hot coding:
One-Hot coding is also known as an efficient coding, and method is to be carried out using N bit status register to N number of state
Coding, each register have independent register-bit, while when any, and only one effectively.
For machine learning task, input feature vector variable includes continuous variable, classification type variable, One-Hot volume
Code is mainly for classification type variable.By the way that classification type variable mappings to data variable are handled convenient for machine learning algorithm.
By taking occupation as an example: teacher, student, engineer
Simple classification coding: 1,2,3
One-Hot coding: [001], [010], [100]
Because original classification code book body does not have the attribute of continuous variable, but direct coding make algorithm to the variable with
Continuous variable is handled, and the increasing informatio being originally not present is introduced.One-Hot variable is changed to then for the noise of original introducing
Information be changed into whether information, significantly reduce noise, but simultaneously characteristic dimension promoted so that learning difficulty promoted.
Decision Tree algorithms:
Decision-tree model uses CART generating algorithm, and mode input is training dataset and stops the condition calculating, exports
It is CART decision tree.Algorithm is according to training dataset, since root node, recursively performs the following operation to each node, structure
Build binary decision tree:
(1) training dataset of node is set as D, calculates existing feature to the gini index of the training dataset.Assuming that there is K
A class, the probability that sample point belongs to kth class is pk, then the gini index of probability distribution is defined as
For given training data set D, gini index is
(2) at this point, to each feature A, each cut-off a that it may be taken, the test according to sample point to A=a
D is divided into D for "Yes" or "No"1And D2Two parts calculate gini index when A=a using following formula.
(3) in all possible feature A and the possible cut-off a of all of which, the smallest spy of gini index is selected
Sign and its corresponding cut-off are as optimal characteristics and optimal cut-off.According to optimal characteristics and optimal cut-off, from existing node
Two child nodes are generated, training dataset is assigned in two child nodes according to feature.
(4) to two child node recursive call above-mentioned steps (1), (2), (3), until meeting stop condition.
(5) CART decision tree is generated.
The condition that decision Tree algorithms stop is that the number of samples in node is less than predetermined threshold, which is
0.5 or sample set gini index be less than predetermined threshold (sample substantially belongs to same class), or do not have more features.
A kind of promise breaking user's probability forecasting method based on sparse features insertion of the present invention, it is compared with prior art, excellent
Point is: effectively increasing the processing capacity of classification coding, while effectively reducing feature in the process of subsequent machine learning
The dimension in space is conducive to the study and processing of machine learning model.
Detailed description of the invention
The present invention is based on user's Default Probability prediction technique schematic diagrames that sparse features are embedded in by Fig. 1.
Multi-class variable processing method schematic diagram of the Fig. 2 based on machine learning.
Specific embodiment
In the following with reference to the drawings and specific embodiments, the following further describes the technical solution of the present invention.
The initial data such as certification, shopping based on user on third-party platform are verified.
As shown in Figure 1, a kind of promise breaking user's probability forecasting method based on sparse features insertion, steps are as follows:
One, data cleansing
Data used are behavior and essential information of the desensitization of loan user etc., and user tag table is the promise breaking of user's loan
Situation, 1 indicates promise breaking user, and 0 indicates normal users.This data set includes 120929 loan user data, time window altogether
Mouth is from March, 2016 in April, 2017.It wherein breaks a contract user 3388, normal users 117541, positive negative sample ratio about exists
1:34.7, sample rate of violation are 2.8%.
Since initial data has the problems such as storage lack of standardization, shortage of data, excessive class variable, logarithm is first had to
According to progress cleaning.For storing nonstandard problem, using the method that nonstandardized technique data are converted to normal data, than
User's date of birth " after 90s " is such as converted into canonical form " 1990-01-01 "；Chinese and English is mixed into field, such as " Chinese work
Quotient bank ", " icbc " etc. uniformly replace with Chinese Fields.Aiming at the problem that shortage of data, for there are data item largely to lack
Field, then count each user missing field number as a feature, the information of missing can be retained to a certain extent.
For the problem that class variable is excessive, different coding methods is used according to different classifications.For gender, bank card types etc.
Without point of specific superiority and inferiority between different values, thus one-hot coding method can be used in field, and for membership grade
There are the classifications of notable difference between its different brackets, then are mapped by 0, the 1,2,3,4 of enumeration type, can be retained in this way
Different information also can be reduced feature quantity (if encoding using one-hot, can derive one 5 dimension matrix).
Two, based on the Feature Engineering of machine learning
The primitive character that the first step is obtained by data cleansing is converted to model by the processing of Feature Engineering by the part
Training data.It is first new feature by traditional Feature Engineering method migration by primitive character, it then will be in new feature
The multi-class variable method that part sparse features propose through the invention is converted to one-dimensional variable, which can be directly as model
Training data.Specific embodiment is as follows:
2.1 Feature Engineering
Variable in primitive character is carried out feature according to time class, amount of money class, address class, telephone number class respectively to mention
It takes, the derivative work of variable.Time class field includes authenticated time, loan time, shopping-time etc., such as by the purchase order time
According in every day the morning (6:00-12:00), afternoon (12:00-18:00), at night (18:00-24:00) and morning (00:
00-6:00) divide, and count quantity on order within this four periods and ratio and working day, weekend quantity on order and
Ratio is distributed here to assist portraying user's Shopping Behaviors by its shopping-time.Such as some user its on weekdays or
Carry out that shopping ratio is higher, then its time is more idle person's morning, it is more likely that the unemployed, thus disobeying after user loan
About probability is also higher.
Amount of money class field mainly include each shopping amount of money of user and user on third-party platform informal voucher amount, make
With amount, from these features, it can be seen that user's current economic ability, such as its third party's informal voucher of certain user use amount ratio
Example is higher, then its income possibly can not support its normal overhead, also more may cause promise breaking.
Address class field mainly includes the number of user's difference shipping address, whether replaces address feature, different consignees
Number and different provinces promise breaking ratio.By counting different provinces promise breaking ratios, promise breaking ratio highest 5 can be filtered out
Province is as statistical nature, such as A saves rate of violation highest, then increases whether the character representation user is that A is saved.Here it only unites
First 5 are counted, also for avoiding this method from generating excessive sparse matrix, verifies to obtain preceding 5 by experience and following model
It is a proper.
Telephone number class field mainly includes operator's informaiton, different reserved call numbers, shipping address in order information
Different phone number numbers and fixed-line telephone number in information.Such as fixed-line telephone is filled in, fixed number is filled in
The client of code, than the user for not filling in the field, inhabitation address is more stable, more difficult generation violations.
After features described above engineering, the derivative variable of 54 dimensions is obtained to carry out subsequent modeling.
The 2.2 class variable processing methods based on sparse features insertion
But after features described above engineering, there are multi-class variables, such as bank sort in the variable of generation.Original number
According to the middle banking style in the presence of more than 80 kinds, can derive to obtain the huge of one 100 dimension or more after directly encoding using one-hot
Sparse matrix (remaining feature is also no more than 80 dimensions), it is dry that this sparse matrix can also carry out very big noise to following model training band
It disturbs.And there is no superiority and inferiority difference as similar membership grade between different bank sorts, can not according to numeric type method into
Row mapping.Therefore multi-class variable processing method proposed by the present invention is used, multi-class variable is converted into one-dimensional variable, it is convenient
Subsequent modeling.As shown in Figure 2.
Original bank sort is subjected to one-hot coding first, obtains a sparse matrix H, dimension is equal to former data
Middle bank sort quantity.Then it after sparse matrix H and label information Y being combined, is fitted using decision-tree model F, it will
The probability value P of model output is put into following model F ' as new feature, is mapped to sparse matrix H after this step
One-dimensional variable P.
By above-mentioned steps, by numerical variable of the original bank sort variable mappings between [0,1].But simultaneously upper
During stating, it is used for a label information Y, thus model can be easy to cause excessively quasi- in subsequent modeling using this feature
It closes.So using decision-tree model, and the depth for controlling tree is no more than 4 layers, to a certain extent in subsequent processes
Alleviate over-fitting.
After above-mentioned multi-class variable method, the one-dimensional variable of bank sort is obtained.In addition traditional characteristic engineering obtains
54 dimension variables, the derivative variable of 55 dimensions, which is always obtained, can carry out subsequent modeling.
The 2.3 model training processing methods based on LightGBM
The sparse variable insertion feature that foundation characteristic derived from step 2.1 and step 2.2 increase newly is merged into eigenmatrix,
It is trained using decision-tree model is promoted based on improved gradient, decision tree is obtained by above-mentioned characteristic processing method and is calculated
The feature space of calligraphy learning carries out parameter optimization and model iteration in this feature space, finally obtains the study for prediction
Device.
The above, preferable implementation sample only of the invention, not does any restrictions to technical scope of the invention, therefore
According to the technical essence of the invention to the above any subtle modifications, equivalent variations and modifications implementing sample and being done, still
Belong in the range of technical solution of the present invention.
Claims (4)
1. a kind of promise breaking user's probability forecasting method based on sparse features insertion, it is characterised in that: the method steps are as follows:
Step 1: data cleansing
The characteristics of variables under different dimensions is converted in the initial data of third-party platform by user, in addition to traditional based on nearest
Consumption time, the nearest consumer finance except the variable of nearest consuming frequency, integrally divide data according to the data of offer
Dimension for behavioral data, consumption data, essential information data is ultimately formed more by being analyzed these data, being converted
The feature of dimension multivariable；
Lack of standardization, field disunity, Chinese and English mixing, shortage of data, multi-class change for the existing storage of initial data simultaneously
The problem of amount, converts regular data for initial data using data cleansing；
Step 2: the Feature Engineering based on machine learning
The Feature Engineering based on machine learning is divided into two parts, is traditional characteristic engineering and based on machine learning respectively
Multi-class variable processing method；Primitive character can generate new feature after the processing of traditional characteristic engineering, but part new feature
It is sparse features, cannot be directly as the training data of model, therefore pass through the multi-class variable method based on machine learning, general
Sparse features are converted to one-dimensional characteristic, so as to the training data directly as model；Wherein:
2.1 traditional characteristic engineerings
Variable in primitive character is subjected to feature extraction according to time class, amount of money class, address class, telephone number class respectively, is become
Measure derivative work；
The 2.2 multi-class variable processing methods based on decision Tree algorithms
For the sparse features of multi-class variable, a kind of i.e. multi-class variable processing method based on decision Tree algorithms is proposed, it will
Sparse features are converted to one-dimensional characteristic, and detailed process is as follows: multi-class variable ω being carried out one-hot coding first, obtains one
A sparse matrix H, dimension are equal to the different classes of quantity of ω in former data；Then it is by sparse matrix H and label information Y
It after the promise breaking information of user combines, is fitted using decision-tree model F, the probability value P that model exports is put as new feature
Enter in following model F ', is mapped sparse matrix H for one-dimensional variable P by this step, and increase to as feature
In model；
Decision Tree algorithms:
Decision-tree model uses CART generating algorithm, and mode input is training dataset and stops the condition calculating, and output is
CART decision tree；Algorithm is according to training dataset, since root node, recursively performs the following operation to each node, building
Binary decision tree:
(1) training dataset of node is set as D, calculates existing feature to the gini index of the training dataset；Assuming that there is K
Class, the probability that sample point belongs to kth class is pk, then the gini index of probability distribution is defined as
For given training data set D, gini index is
(2) at this point, to each feature A, each cut-off a that may be taken to it is to the test of A=a according to sample point
D is divided into D by "Yes" or "No"1And D2Two parts calculate gini index when A=a using following formula；
(3) in all possible feature A and the possible cut-off a of all of which, select the smallest feature of gini index and
Its corresponding cut-off is as optimal characteristics and optimal cut-off；According to optimal characteristics and optimal cut-off, generated from existing node
Training dataset is assigned in two child nodes by two child nodes according to feature；
(4) to two the above-mentioned sub-steps of child node recursive call (1), (2), (3), until meeting stop condition；
(5) CART decision tree is generated；
The condition that decision Tree algorithms stop is that the number of samples in node is less than predetermined threshold or the gini index of sample set is less than
Predetermined threshold, or there is no more features.
2. a kind of promise breaking user's probability forecasting method based on sparse features insertion according to claim 1, feature exist
In: the data cleansing, the specific method is as follows:
1.1 field processing lack of standardization
There are part messy code data and data lack of standardization for initial data；For messy code data, using delete processing；For storing not
The data of its nonstandardized technique are converted to unified canonical form by the data of specification；
The processing of 1.2 missing datas
There are a large amount of absent fields in initial data, and for different deletion conditions, there are different data cleansing modes: including
Using the methods of filling mean, mode, median " polishing " field；Or each user's absent field number is counted as one
A feature not only remains the information of missing in this way, it is ensured that the stability of data distribution；
The processing of 1.3 class variables
In addition, there is also class variables in initial data, class variable is converted by way of mapping；Specific to different
Classification uses distinct methods, does not have the classification of specific superiority and inferiority divided using one-hot coding method different values；For not
With value, there are the classifications of notable difference to use numeric type coding method.
3. a kind of promise breaking user's probability forecasting method based on sparse features insertion according to claim 1, feature exist
In: detailed process is as follows for the traditional characteristic engineering:
2.1.1 time class field:
It include time class field in primitive character, these time class fields can reflect activity periods of the user in specific time
Situation；The time class field of user is divided according to time interval, the behavior of user in the time interval is counted, to assist
Portray user's portrait；
2.1.2 amount of money class field
Amount of money class field is that most can directly react the information of user's economic capability in modeling；Amount of money class field directly reflects use
The level of consumption at family specifically includes the shopping amount of money of user, and the shopping amount of money is bigger, and loan user's economic capability is also stronger,
Its Default Probability is also lower；
2.1.3 address class field
Address information equally can help auditor from another angle to describe user；Address category information is divided, it can
To obtain the address class user portrait of the user；One its inhabitation address of common user can be more stable, the place of acceptance of shopping
Location is also relatively stable, is distributed in job site and place of abode；And for the part unemployed, unstable residence
Location, or frequently more change jobs, its replacement frequency is also greater than other users from the point of view of shipping address；
2.1.4 telephone number class field
Telephone number information also reflects the partial information of user；Telephone number has a different operators, and telephone number
Number, frequency of use etc. can all influence the credit default risk of a user；Telephone number frequency of use is too low, illustrates the user
Potentially unstable, it is easier to generate violations.
4. a kind of promise breaking user's probability forecasting method based on sparse features insertion according to claim 1, feature exist
In described One-Hot coding: an also known as efficient coding, method be using N bit status register come to N number of state into
Row coding, each register have independent register-bit, while when any, and only one effectively；
For machine learning task, input feature vector variable includes continuous variable, classification type variable, One-Hot coding master
To be directed to classification type variable；By the way that classification type variable mappings to data variable are handled convenient for machine learning algorithm.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910084188.5A CN109919436A (en) | 2019-01-29 | 2019-01-29 | A kind of promise breaking user's probability forecasting method based on sparse features insertion |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910084188.5A CN109919436A (en) | 2019-01-29 | 2019-01-29 | A kind of promise breaking user's probability forecasting method based on sparse features insertion |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109919436A true CN109919436A (en) | 2019-06-21 |
Family
ID=66961068
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910084188.5A Pending CN109919436A (en) | 2019-01-29 | 2019-01-29 | A kind of promise breaking user's probability forecasting method based on sparse features insertion |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109919436A (en) |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111552723A (en) * | 2020-05-07 | 2020-08-18 | 河北雄安舜耕数据科技有限公司 | System and method for client portrait in data management |
| CN111967596A (en) * | 2020-08-18 | 2020-11-20 | 北京睿知图远科技有限公司 | Feature automatic intersection method based on deep learning in wind control scene |
| CN112633937A (en) * | 2020-12-30 | 2021-04-09 | 上海数鸣人工智能科技有限公司 | Marketing prediction method based on dimension reduction of depth automatic encoder and gradient lifting decision tree |
| CN112700324A (en) * | 2021-01-08 | 2021-04-23 | 北京工业大学 | User loan default prediction method based on combination of Catboost and restricted Boltzmann machine |
| CN113570392A (en) * | 2020-04-29 | 2021-10-29 | 中移动信息技术有限公司 | User grouping method and device, electronic equipment and computer storage medium |
| CN113706195A (en) * | 2021-08-26 | 2021-11-26 | 东北大学秦皇岛分校 | Online consumption behavior prediction method and system based on two-stage combination |
| CN115934809A (en) * | 2023-03-08 | 2023-04-07 | 北京嘀嘀无限科技发展有限公司 | Data processing method and device and electronic equipment |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105045931A (en) * | 2015-09-02 | 2015-11-11 | 南京邮电大学 | Video recommendation method and system based on Web mining |
| CN108564286A (en) * | 2018-04-19 | 2018-09-21 | 天合泽泰（厦门）征信服务有限公司 | A kind of artificial intelligence finance air control credit assessment method and system based on big data reference |
| CN109034658A (en) * | 2018-08-22 | 2018-12-18 | 重庆邮电大学 | A kind of promise breaking consumer's risk prediction technique based on big data finance |
| CN109255506A (en) * | 2018-11-22 | 2019-01-22 | 重庆邮电大学 | A kind of internet finance user's overdue loan prediction technique based on big data |
-
2019
- 2019-01-29 CN CN201910084188.5A patent/CN109919436A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105045931A (en) * | 2015-09-02 | 2015-11-11 | 南京邮电大学 | Video recommendation method and system based on Web mining |
| CN108564286A (en) * | 2018-04-19 | 2018-09-21 | 天合泽泰（厦门）征信服务有限公司 | A kind of artificial intelligence finance air control credit assessment method and system based on big data reference |
| CN109034658A (en) * | 2018-08-22 | 2018-12-18 | 重庆邮电大学 | A kind of promise breaking consumer's risk prediction technique based on big data finance |
| CN109255506A (en) * | 2018-11-22 | 2019-01-22 | 重庆邮电大学 | A kind of internet finance user's overdue loan prediction technique based on big data |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113570392A (en) * | 2020-04-29 | 2021-10-29 | 中移动信息技术有限公司 | User grouping method and device, electronic equipment and computer storage medium |
| CN113570392B (en) * | 2020-04-29 | 2024-04-09 | 中移动信息技术有限公司 | User grouping method, device, electronic equipment and computer storage medium |
| CN111552723A (en) * | 2020-05-07 | 2020-08-18 | 河北雄安舜耕数据科技有限公司 | System and method for client portrait in data management |
| CN111967596A (en) * | 2020-08-18 | 2020-11-20 | 北京睿知图远科技有限公司 | Feature automatic intersection method based on deep learning in wind control scene |
| CN112633937A (en) * | 2020-12-30 | 2021-04-09 | 上海数鸣人工智能科技有限公司 | Marketing prediction method based on dimension reduction of depth automatic encoder and gradient lifting decision tree |
| CN112633937B (en) * | 2020-12-30 | 2023-10-20 | 上海数鸣人工智能科技有限公司 | Marketing prediction method based on dimension reduction and GBDT (Global positioning System) of depth automatic encoder |
| CN112700324A (en) * | 2021-01-08 | 2021-04-23 | 北京工业大学 | User loan default prediction method based on combination of Catboost and restricted Boltzmann machine |
| CN113706195A (en) * | 2021-08-26 | 2021-11-26 | 东北大学秦皇岛分校 | Online consumption behavior prediction method and system based on two-stage combination |
| CN113706195B (en) * | 2021-08-26 | 2023-10-31 | 东北大学秦皇岛分校 | Online consumption behavior prediction method and system based on two-stage combination |
| CN115934809A (en) * | 2023-03-08 | 2023-04-07 | 北京嘀嘀无限科技发展有限公司 | Data processing method and device and electronic equipment |
| CN115934809B (en) * | 2023-03-08 | 2023-07-18 | 北京嘀嘀无限科技发展有限公司 | Data processing method and device and electronic equipment |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109919436A (en) | A kind of promise breaking user's probability forecasting method based on sparse features insertion | |
| CN109918511B (en) | BFS and LPA based knowledge graph anti-fraud feature extraction method | |
| CN108510280B (en) | Financial fraud behavior prediction method based on mobile equipment behavior data | |
| CN102591854B (en) | For advertisement filtering system and the filter method thereof of text feature | |
| CN110555455A (en) | Online transaction fraud detection method based on entity relationship | |
| CN109918452A (en) | A kind of method, apparatus of data processing, computer storage medium and terminal | |
| CN106897930A (en) | A kind of method and device of credit evaluation | |
| CN106372072A (en) | Location-based recognition method for user relations in mobile social network | |
| CN109034194A (en) | Transaction swindling behavior depth detection method based on feature differentiation | |
| CN106022708A (en) | Method for predicting employee resignation | |
| Lu et al. | Telecom fraud identification based on ADASYN and random forest | |
| CN107784517A (en) | A kind of customer information screening method and terminal device | |
| CN108647800A (en) | A kind of online social network user missing attribute forecast method based on node insertion | |
| CN107818376A (en) | Customer loss Forecasting Methodology and device | |
| CN106056137B (en) | A kind of business recommended method of telecommunications group based on data mining multi-classification algorithm | |
| CN113139876A (en) | Risk model training method and device, computer equipment and readable storage medium | |
| CN110059126A (en) | Complicated related network analysis method and system based on LKJ exception Value Data | |
| CN112365339A (en) | Method for judging commercial value credit loan amount of small and medium-sized enterprises | |
| CN115391151B (en) | Method for intelligently discovering alarm tag based on object relationship | |
| CN107222319A (en) | A kind of traffic operation analysis method and device | |
| CN109858901A (en) | Mobile subscriber based on mobile phone signaling big data pays recognition methods | |
| CN115731030A (en) | Method, device and storage medium for mining bank consumption loan customer requirements | |
| CN115587828A (en) | Interpretable method of telecommunication fraud scene based on Shap value | |
| Guo et al. | Statistical decision research of long-term deposit subscription in banks based on decision tree | |
| CN112270548B (en) | Credit card fraud detection method based on deep learning |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| WD01 | Invention patent application deemed withdrawn after publication | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20190621 |