Summary of the invention
This specification one or more embodiment describes a kind of method and apparatus, can be based on the incidence number between user
According to, and influence of the other users to user credit to be assessed with user to be assessed with incidence relation, improve credit evaluation
Accuracy.
According in a first aspect, providing a kind of method of credit evaluation, comprising: based on human relation network it is determining with it is to be evaluated
Estimate the first level-one association user that user has incidence relation, wherein the human relation network is in the use for being associated operation
Incidence relation is established between family, and there are between the user of incidence relation by the operation associated corresponding text information record
Relating attribute;According to the text envelope in the relating attribute of the user to be assessed and the first level-one association user
Breath, at least one level-one for obtaining the user to be assessed are associated with vocabulary, and raw based at least one level-one association vocabulary
At the first level-one linked character of the user to be assessed;Based on the human relation network, determine the user's to be assessed
At least one second level association user, wherein at least one described second level association user has with the first level-one association user
Incidence relation;According to the text envelope in the relating attribute of the first level-one association user and each second level association user
Breath, at least one second level for obtaining the user to be assessed are associated with vocabulary, and raw based at least one second level association vocabulary
At the first second level linked character of the user to be assessed;At least it is based on the first level-one linked character and first second level
Linked character assesses the credit rating of the user to be assessed by credit evaluation model trained in advance.
Embodiment according to one aspect, the text information include the operation associated corresponding phrase, sentence information;
And the text information according in the relating attribute of the user to be assessed and the first level-one association user,
At least one level-one association vocabulary for obtaining the user to be assessed includes: to carry out word cutting to the phrase, sentence information to obtain
Initial vocabulary;Each initial vocabulary is matched with the keyword in pre-generated keyword set respectively;It will be described
The initial vocabulary being matched in keyword set is associated with vocabulary as the level-one.
According to the embodiment of another aspect, the text information includes believing from the operation associated corresponding phrase, sentence
The association vocabulary extracted in advance in breath;And the pass according to the user to be assessed and the first level-one association user
The text information in attribute, at least one level-one association vocabulary for obtaining the user to be assessed includes: by the text
This information is associated with vocabulary as at least one level-one.
In one embodiment, the keyword in the keyword set extracts by the following method: obtaining artificial calibration
User's positive sample and user's negative sample;Based on human relation network determine user's positive sample and user's negative sample with
Sample text information in the relating attribute of other users;The key in keyword set is determined according to the sample text information
Word.
In one embodiment, each keyword in the keyword set is also corresponding with statistical indicator, the statistics
Index includes at least one of the following: the number occurred in the text information of user's positive sample, in the text envelope of user's negative sample
The probability of the number, deviation user's positive sample that occur in breath, the probability for being biased to user's negative sample.
According to a kind of possible design, each level-one association vocabulary is corresponding with the pass being matched in the keyword set
The statistical indicator of keyword;And it is described based at least one described level-one association vocabulary generate the first of the user to be assessed
Level-one linked character includes: to obtain the corresponding statistical indicator of each level-one association vocabulary;It is corresponding to each level-one association vocabulary
Statistical indicator carries out the first predetermined process, generates the first level-one linked character, wherein first predetermined process include with
At least one of lower: maximizing minimizes, sums, averaging, seeking weighted sum.
In one embodiment, first second level for generating the user to be assessed based on second level association vocabulary is closed
Connection feature includes: to be associated with vocabulary according to the second level, generates the first level-one association user and each second level association user point
Not corresponding second level-one linked character;Second predetermined process is carried out to each second level-one linked character, processing result is made
For the first second level linked character of the user to be assessed, wherein second predetermined process, which includes at least one of the following:, to be asked most
It is worth greatly, minimize, sum, average, seeks weighted sum.
In one embodiment, the predetermined operation include: transfer accounts, give bonus, plusing good friend.
According to second aspect, a kind of device of credit evaluation is provided, comprising: the first determination unit is configured to interpersonal
The determining first level-one association user with user to be assessed with incidence relation of relational network, wherein the human relation network
Incidence relation is established between the user for being associated operation, and is existed by the operation associated corresponding text information record
Relating attribute between the user of incidence relation;First generation unit is configured to according to the user to be assessed and described first
The text information in the relating attribute of level-one association user obtains at least one level-one conjunctive word of the user to be assessed
It converges, and generates the first level-one linked character of the user to be assessed based at least one described level-one association vocabulary;Second really
Order member, is configured to the human relation network, determines at least one second level association user of the user to be assessed,
In, at least one described second level association user and the first level-one association user have incidence relation;Second generation unit, matches
It is set to according to the text information in the relating attribute of the first level-one association user and each second level association user, obtains
At least one second level of the user to be assessed is associated with vocabulary, and based at least one second level association vocabulary generate it is described to
Assess the first second level linked character of user;Credit evaluation unit, be configured at least based on the first level-one linked character and
The first second level linked character assesses the credit rating of the user to be assessed by credit evaluation model trained in advance.
According to the third aspect, a kind of computer readable storage medium is provided, computer program is stored thereon with, when described
When computer program executes in a computer, enable computer execute first aspect method.
According to fourth aspect, a kind of calculating equipment, including memory and processor are provided, which is characterized in that described to deposit
It is stored with executable code in reservoir, when the processor executes the executable code, the method for realizing first aspect.
The method and apparatus provided by this specification embodiment are primarily based on human relation network determination and use to be assessed
Family has the level-one association user of incidence relation, then according to the text in the relating attribute of user to be assessed and level-one association user
This information obtains at least one level-one association vocabulary of user to be assessed, and generates one based at least one level-one association vocabulary
Grade linked character, then determine user to be assessed, be associated with at least one second level of incidence relation with level-one association user
User, and it is directed to each level-one association user, it is closed according to the second level that its corresponding second level association user generates user to be assessed
Join feature, be then based on level-one linked character and second level linked character, is assessed by credit evaluation model trained in advance to be evaluated
Estimate the credit rating of user.In this way, the associated data between user can be made full use of, have due to considering with user to be assessed
Influence of the other users of incidence relation to user credit to be assessed, so as to improve the accuracy of credit evaluation.
Specific embodiment
With reference to the accompanying drawing, the scheme provided this specification is described.
Fig. 1 is the implement scene schematic diagram of one embodiment that this specification discloses.In the implement scene, multiple users
It can be interacted by network and computing platform.Good friend can be added between user by computing platform, transfers accounts, is rubescent
Packet etc. operation.Computing platform can be operated by these between user, human relation network be established, to indicate between user
Incidence relation.Human relation network can be updated according to the period, such as be updated within 1 day once, can also be often to detect
Operation between user is once updated, and is not construed as limiting in this implement scene to this.
It as described in Figure 2, is the specific example of a human relation network.In the human relation network, circle (node)
Indicate user, line (side) indicates incidence relation, and being wired the user that (side) connects together has direct correlation relationship.Here
Incidence relation, can be addition good friend, the connection relationship that predetermined operation of transferring accounts, give bonus etc. is realized.The category of line (side)
Property in by above-mentioned predetermined operation corresponding text information record, there are the relating attributes between the user of incidence relation.In order to just
Line between explanation, user can also be described as side, and the relating attribute between user can also be described as side attribute.
When needing to carry out credit evaluation to some user to be assessed, can first based on human relation network determine with it is to be evaluated
Estimate the level-one association user that user has incidence relation, user to be assessed corresponds in the side attribute of each level-one association user
Text information obtains one group of level-one and is associated with vocabulary, and a level-one linked character can be generated according to every group of level-one association vocabulary.
It then, can and its use with incidence relation determining based on human relation network to each level-one association user
Second level association user of the family as user to be assessed.For each level-one association user, corresponding to second level association user
Text information in side attribute, available one group of second level are associated with vocabulary.It is closed according to each group second level of a level-one association user
Join vocabulary, a second level linked character for generating user to be assessed can be summarized.In one embodiment, it is closed according to every group of second level
A level-one linked character of corresponding level-one association user can be generated in connection vocabulary, according to all the one of a level-one association user
The second level linked character of user to be assessed can be generated in grade linked character.
Then, each level-one linked character and second level linked character based on user to be assessed pass through letter trained in advance
The credit rating of user to be assessed is assessed with assessment models.In this way, being associated with for user to be assessed and other users can be made full use of
Data extract more features, improve the accuracy assessed user credit.Specifically executing for above-mentioned scene is described below
Journey.
Fig. 3 shows the method flow diagram of the credit evaluation according to one embodiment.The executing subject of this method, which can be, appoints
What has calculating, the system of processing capacity, unit, platform or server, such as computing platform shown in FIG. 1 etc..Such as figure
3 show, method includes the following steps: step 31, has incidence relation with user to be assessed based on human relation network is determining
First level-one association user, wherein human relation network establishes incidence relation between the user for being associated operation, and passes through
There are the relating attributes between the user of incidence relation for operation associated corresponding text information record;Step 32, according to be assessed
Text information in the relating attribute of user and the first level-one association user obtains at least one level-one association of user to be assessed
Vocabulary, and the first level-one linked character that vocabulary generates user to be assessed is associated with based at least one level-one;Step 33, it is based on people
Border relational network determines at least one second level association user of user to be assessed, wherein at least one above-mentioned second level association user
There is incidence relation with the first level-one association user;Step 34, according to the first level-one association user and each second level association user
Relating attribute in text information, obtain user to be assessed at least one second level association vocabulary, and based at least one two
Grade association vocabulary generates the first second level linked character of user to be assessed;Step 35, at least based on the first level-one linked character and
First second level linked character assesses the credit rating of user to be assessed by credit evaluation model trained in advance.
Firstly, in step 31, based on determining first level-one with user to be assessed with incidence relation of human relation network
Association user.It is worth noting that human relation network is used for based on the network structure of the operation associated foundation between user
Record the incidence relation between user and user.There can also be relating attribute between user, which passes through above-mentioned pass
Text information record in connection operation.Level-one association user is that have the user of direct correlation relationship.In practice, between user
It is operation associated for example can include but is not limited to it is at least one of following: plusing good friend transfers accounts, gives bonus, etc..
Human relation network as shown in connection with fig. 2, between user 21 and user 22 have it is operation associated, such as user 21 to
User 22 has carried out transfer operation, then is connected between user 21 and user 22 by line 24, indicates that they have incidence relation.
Line 24 can have the relating attribute between user 21 and user 22 with corresponding record, which passes through the text in operation associated
This information record, such as the message of transferring accounts " dinner cost at noon " when user 21 transfers accounts to user 22.Equally, have operation associated
It can be connected by line 25 between user 22 and user 23, line 25 can be by being associated between user 22 and user 23
Operate the relating attribute between corresponding text information record user 22 and user 23.
It is appreciated that in human relation network shown in Fig. 2: embodiment according to one aspect, line (such as line 24)
Can be it is nondirectional, as long as recognizing specifically, one in user 21 and user 22 has carried out operation associated to another
Mutually there is incidence relation for them;According to the embodiment of another aspect, line (such as line 24) is also possible to directive, tool
For body, user 21 has carried out operation associated (such as transferring accounts) to user 22, then user 22 is that have incidence relation with user 21
User, and user 21 be not necessarily with user 22 have incidence relation user.For ease of description, in this specification embodiment with
Line is directionless to be illustrated.
In human relation network shown in Fig. 2, user 21, user 23 etc. have direct correlation relationship with user 22,
It can be the level-one association user of user 22.User 22 is the level-one association user of user 21, but user 23 and user 21 do not have
Direct correlation relationship, therefore, user 23 and user 21 are not mutually the level-one association users of other side.
Design according to one aspect, above-mentioned text information can be phrase, sentence information.For example, when operation associated
When being plusing good friend, above-mentioned text information can be verifying message when sending addition good friend's request, such as " owing you people of money ", or
Person is the remarks or label for good friend's setting, such as " confirmed gambler ", " braggart ", " honest " etc..When it is operation associated be to transfer accounts
Or when giving bonus, above-mentioned text information can be message sentence " credit card repayment ", " rent next month " etc..In some implementations
In example, above-mentioned text information can also be the keyword extracted from the information such as message, remarks, label, such as: " credit card ",
" refund ", " next month ", " rent ", " confirmed gambler ", " boast ", " king ", " honest " etc..
According to the design of another aspect, above-mentioned text information can also be the conjunctive word extracted from phrase, sentence information
It converges.When above-mentioned text information is association vocabulary, the information such as message, remarks, label in predetermined operation can be first passed through
The segmenting method of such as N-gram, Bigram etc carry out word cutting processing, obtain initial vocabulary.By taking Bigram as an example, for staying
Say " dinner cost at noon ", participle the result is that " noon ", " noon ", " meal ", " dinner cost ".Then, participle can be obtained
Initial vocabulary is matched with the keyword in pre-generated keyword set respectively, by what is be matched in keyword set
Initial vocabulary conduct, for record the relating attribute between user it is operation associated in text information.
Wherein, the keyword in above-mentioned keyword set can extract in the following manner: firstly, selection user's positive sample
With user's negative sample, user's positive sample and user's negative sample can be the user's samples artificially demarcated.For example, for some money
The user of gold management platform (such as Alipay) checks its credit card repayment record or the creditor-debtor entry in the platform, if letter
With good, such as refund on time every time, be then demarcated as user's positive sample, whereas if credit is bad, such as occur refunding it is overdue,
Situations such as debt-credit is not gone back, then be demarcated as user's negative sample.
Then, it for user's positive sample and user's negative sample, respectively by above-mentioned human relation network, obtains and other use
Sample text information in the relating attribute at family, and word cutting processing is carried out, obtained word cutting vocabulary is added crucial as keyword
Set of words.Optionally, obtained word cutting vocabulary can also be filtered, such as removal stop words, function word, by remaining word
It converges and keyword set is added.
In one embodiment, the keyword in above-mentioned keyword set can also respectively correspond statistical indicator, as
Their statistical parameters in user's positive and negative samples.Statistical indicator can include but is not limited at least one of following: user just
The number that occurs in the text information of sample, is biased to user's positive sample at the number occurred in the text information of user's negative sample
Probability, be biased to the probability, etc. of user's negative sample.When user's positive and negative samples quantity is consistent, vocabulary is in user's positive sample
The number occurred in text information and/or the number occurred in the text information of user's negative sample are biased to positive (negative) sample of user
This probability all can serve as their statistical indicator.At this point, vocabulary is biased to user by taking the probability for being biased to user's positive sample as an example
The probability of positive sample=number occurred in the text information of user's positive sample/(goes out in the text information of user's positive sample
The number for existing number+occur in the text information of user's negative sample).It, can be with when user's positive and negative samples quantity is inconsistent
The probability of positive (negative) sample of user is biased to as statistical indicator by vocabulary.It is false still for being biased to the probability of user's positive sample
If user's positive and negative samples quantity ratio is M:N, then vocabulary is biased to probability=N × vocabulary of user's positive sample in user's positive sample
The number occurred in text information/(N × number+M occurred in the text information of user's positive sample × in user's negative sample
The number occurred in text information).
It is readily appreciated that, the number that vocabulary occurs in the text information of user's positive sample is more, is biased to user's positive sample
Probability is bigger, is easier to be biased to user's positive sample, the number occurred in the text information of user's negative sample is more, is biased to use
The probability of family positive sample is smaller, is easier to be biased to user's negative sample.For user's positive and negative samples of equivalent, in the positive sample of user
The number difference occurred in the text information of this and user's negative sample is smaller (being less than preset times), or for arbitrarily distributing number
User's positive and negative samples of amount, are biased to vocabulary of the probability in middle position (near such as 0.5) of positive (negative) sample of user, and vocabulary is inclined
It is all weaker to user's positive sample and the trend for being biased to user's negative sample, it is believed that be neutral words.For neutral vocabulary, from statistics
For, when user's sample size is sufficiently large, it is biased to a range of the Probabilistic Stability of positive (negative) sample of user near 0.5
It is interior, however, if the positive sample of user may be biased in primary divide using 0.5 as the boundary probability for being biased to user's positive and negative samples
This, may also be biased to user's negative sample in another division, influence the stability of data processing.Therefore, in one embodiment
In, it may filter out such neutral words, such as be biased to the probability of positive (negative) sample of user in preset range (such as 0.4-0.6
In) vocabulary improve the stability of data processing to keep the vocabulary in keyword set more stable.
It, can also be according to it in user's positive sample for the keyword in keyword set in a possible design
Text information in the number occurred, the number that occurs in the text information of user's negative sample or be biased to positive (negative) sample of user
This probability distinguishes positive vocabulary and negative sense vocabulary.Wherein, positive vocabulary is that occur in the text information of user's positive sample
Number is on the high side, or is biased to the biggish vocabulary of probability of user's positive sample, and negative sense vocabulary is the text information in user's negative sample
The number of middle appearance is on the high side, or is biased to the lesser vocabulary of probability of user's positive sample.
The incidence relation for being associated operation between user is made full use of by human relation network, not only can clearly be used
Incidence relation between family, can also be according to the relating attribute between operation associated corresponding text information record user.Such as Fig. 2
Shown, user to be assessed may have level-one incidence relation with multiple users, for ease of description, in this specification embodiment
It is described by some level-one association user of user to be assessed.In other words, " described in this specification embodiment
" first " in one level-one association user ", is not offered as sequence or quantity, but refer to one of them, some.
Step 32, according to the text information in the relating attribute of user to be assessed and the first level-one association user, obtain to
At least one level-one for assessing user is associated with vocabulary, and generates the first of user to be assessed based at least one level-one association vocabulary
Level-one linked character.It is appreciated that the case where above-mentioned text information is the association vocabulary extracted from phrase, sentence information
Under, at least one level-one that can directly acquire these association vocabulary as user to be assessed is associated with vocabulary.In above-mentioned text envelope
In the case that breath is phrase, sentence information, phrase, sentence letter can be extracted according to the method above-mentioned for extracting association vocabulary
Association vocabulary in breath, at least one level-one as user to be assessed are associated with vocabulary, it may be assumed that cut to phrase, sentence information
Word obtains initial vocabulary;Each initial vocabulary is matched with the keyword in pre-generated keyword set respectively;It will
The initial vocabulary being matched in keyword set is associated with vocabulary as level-one.Details are not described herein.
It is associated with vocabulary according at least one level-one of user to be assessed, the level-one that can further generate user to be assessed is closed
Join feature.Wherein, level-one linked character can be for embodying the single people between user to be assessed and its level-one association user
The feature of border relationship.
In one embodiment, the vocabulary in keyword set is corresponding with statistical indicator, since each level-one is associated with vocabulary
It is the vocabulary being matched in keyword set, so each level-one association vocabulary also corresponds to the statistics of be matched to keyword
Index.At this point it is possible to according to the corresponding statistical indicator of each level-one of user to be assessed and level-one association user association vocabulary into
The first predetermined process of row generates level-one linked character.First predetermined process for example can be the statistical indicator to each keyword
Summation, maximum value, minimum value, average value, seek weighted sum etc..For weighted sum, it is assumed that user to be assessed and one
It having carried out transferring accounts for 2 times between level-one association user, the message transferred accounts for 2 times is " dinner cost of today ", " the deficient dinner cost of institute's last month " respectively,
Acquired level-one association vocabulary is, for example: " dinner cost ", the corresponding Probability p for being biased to positive sample1=0.7, frequency of occurrence c1=2;
" today ", the corresponding Probability p for being biased to positive sample2=0.7, frequency of occurrence c2=1;" last month ", the corresponding Probability p for being biased to positive sample3
=0.3, frequency of occurrence c3=1.Then calculate the level-one linked character that user to be assessed corresponds to the level-one association user are as follows: (p1×
c1+p2×c2+p3×c3)/(c1+c2+c30.7 × 2+0.7+0.3 of)=()/(2+1+1).
In another embodiment, the keyword in keyword set can be corresponding with positive vocabulary and negative sense vocabulary mark
Label, so each level-one association vocabulary is also corresponding with positive vocabulary and negative sense vocabulary label.At this point, can also count it is above-mentioned at least
Positive vocabulary number in one level-one conjunctive word determines level-one linked character according to positive (negative) specific gravity to vocabulary.Also join
It is illustrated according to example above.It is assumed that " dinner cost " is positive vocabulary 1, frequency of occurrence c1=2;" today " is positive vocabulary 2, out
Occurrence number c2=1;" last month " is negative sense vocabulary 1, frequency of occurrence c3=1.The specific gravity of positive vocabulary may is that positive vocabulary
Number/(positive vocabulary number+negative sense vocabulary number)=2/ (2+1).It can also be using the frequency of occurrence of level-one conjunctive word as weight
It is calculated, the specific gravity of positive vocabulary may is that (positive 1 frequency of occurrence c of vocabulary12 frequency of occurrence c of+positive vocabulary2)/(is positive
1 frequency of occurrence c of vocabulary12 frequency of occurrence c of+positive vocabulary21 frequency of occurrence c of+negative sense vocabulary3)=(2+1)/(2+1+1).
In this way, the level-one linked character that user to be assessed corresponds to each level-one association user can be generated.Optionally, also
These level-one linked characters can be added to the level-one linked character set of user to be assessed.
Step 33, it is based on human relation network, determines at least one second level association user of user to be assessed, wherein on
Stating at least one second level association user and the first level-one association user has incidence relation.It is appreciated that with user's to be assessed
Level-one association user has the user of incidence relation, can be referred to as the second level association user of user to be assessed.By taking Fig. 2 as an example,
User 22 is the level-one association user of user 21, and user 23, user 26 etc. with user 22 there is the user of direct correlation relationship to be
The second level association user of user 21.It optionally, further include user 21 with the associated user of user 22, in the second level for determining user 21
When association user, itself can also be excluded.
In this way, being directed to each corresponding level-one association user of user to be assessed, available user to be assessed is at least
One second level association user.
Step 34, according to the text information in the relating attribute of the first level-one association user and each second level association user,
At least one second level association vocabulary of user to be assessed is obtained, and user to be assessed is generated based at least one second level association vocabulary
The first second level linked character.Wherein, " first " in " the first second level linked character " here is for expressing and the " the 1st
The corresponding relationship of grade association user ", rather than the restriction to serial number or quantity.
It is appreciated that some level-one association user and any one second level association user associated there, in above-mentioned text
In the case that this information is the keyword extracted from phrase, sentence information, these keywords can be directly acquired as to be evaluated
Estimate at least one second level association vocabulary of user.It, can be according to preceding in the case where above-mentioned text information is phrase, sentence information
The method for the extraction keyword stated extracts phrase, the keyword in sentence information, as user to be assessed at least one two
Grade association vocabulary, details are not described herein.
It is associated with vocabulary according at least one second level of user to be assessed, the second level that can further generate user to be assessed is closed
Join feature.Wherein, second level linked character can be used for embodying the whole man of some corresponding level-one association user of user to be assessed
The feature of border relationship.
It is appreciated that user to be assessed is any user in human relation network, and therefore, the level-one with user to be assessed
Similarly, for any one level-one association user of user to be assessed, its level-one linked character (side also can be generated in linked character
Method is with described previously), the spy of the single interpersonal relationships to indicate the level-one association user and some above-mentioned second level association user
Sign.And the second level linked character that user to be assessed corresponds to the level-one association user can be, the level-one of the level-one association user
The comprehensive of linked character embodies.
In this way, according to an implementation, for any one corresponding level-one of user to be assessed (user 21 in such as Fig. 2)
Association user (user 22 in such as Fig. 2) can firstly generate each of the level-one association user (user 22 in such as Fig. 2)
Level-one linked character (the level-one linked character of such as user 22 and user 23, user 22 and the level-one linked character of user 26, user
22 and user 27 level-one linked character), then these level-ones of the level-one association user (user 22 in such as Fig. 2) are associated with
Feature carries out the second predetermined process, to generate the second level linked character of user to be assessed (user 21 in such as Fig. 2).Its
In, which can include but is not limited to: summation, minimum value, average value, seeks weighted sum etc. at maximum value.In the hope of
For weighted sum, each second level association user can be corresponded to, one weight is set, the weight can with the level-one association user with
Operation associated number between corresponding second level association user is positively correlated, be also possible to and with the level-one association user associated two
Quantity (such as N, N are positive integer) negatively correlated (such as 1/N) of grade association user, etc., as long as second level association user can rationally be embodied
Different degree, this specification embodiment is not construed as limiting this.
According to another embodiment, can also for user to be assessed (user 21 in such as Fig. 2) it is corresponding any one
Level-one association user (user 22 in such as Fig. 2) obtains the level-one association user (user 22 in such as Fig. 2) and each second level
The corresponding level-one of text information in the relating attribute of association user (such as user 23, user 26 with user 27) is associated with vocabulary, and
These level-ones association vocabulary is summarized, and their statistical indicator is carried out including but not limited to below at least one pre-
Fixed processing, to generate the second level linked character of user to be assessed (user 21 in such as Fig. 2): summation, maximum value, minimum
Value, average value seek weighted sum etc..Wherein, in the hope of weighted sum, each level-one is associated with the corresponding weight of vocabulary can be with this
Level-one conjunctive word remittance frequency of occurrence is positively correlated.
In this way, can correspond to each level-one association user of user to be assessed, a second level linked character is generated respectively.
Optionally, these second level linked characters can also be added to the second level linked character set of user to be assessed.
Step 35, it is at least based on the first level-one linked character and the first second level linked character, passes through credit trained in advance
Assessment models assess the credit rating of user to be assessed.It is appreciated that " the first level-one linked character " with " the first second level is associated with spy
The feature that sign " is generated both for some corresponding level-one association user of user to be assessed, in fact, being directed to user to be assessed
Each corresponding level-one association user can correspond to and generate a level-one linked character and a second level linked character.
The credit evaluation model that above-mentioned level-one linked character and the input of second level linked character can be trained in advance, to obtain
The output of credit evaluation model is as a result, determine the credit rating of user to be assessed according to the output result.As shown in figure 4, at least will
Level-one linked character 41 and second level linked character 42 input credit evaluation model, obtain output result.If there are also other features,
Such as user identity feature, probability characteristics of honouring an agreement, can be by above-mentioned level-one linked character and second level linked character and other features
Credit evaluation model trained in advance is inputted, together with the credit rating of determination user to be assessed.
It is appreciated that in the case where user to be assessed only corresponds to a level-one association user and other no features, it can
To be based on above-mentioned first level-one linked character and the first second level linked character, by credit evaluation model assessment trained in advance to
Assess the credit rating of user.
Wherein, credit evaluation model can be such as scorecard model, Random Forest model of training in advance, gradient is promoted
The classification of decision tree (Gradient Boosting Decision Tree, GBTD) model etc or scoring model, herein no longer
It repeats.The output result of credit evaluation model can be classification results, such as the good user of credit, the bad user of credit etc.;?
Marking be can be as a result, such as any score in 1-1000;Etc., this specification embodiment is not construed as limiting this.This is defeated
Result can also be further processed work after (such as normalized) directly as the credit rating assessment result of user out
For the credit rating assessment result of user.
It, can be according to user's to be assessed in the case where above-mentioned credit rating result uses credit field in practical application
Credit rating sets debt-credit condition, for example, credit score is lower than 600, can not carry out debt-credit activity, etc..
Above procedure is looked back, during carrying out credit evaluation to user, based on the associated data between user, sufficiently benefit
Text information when with the incidence relation between user and incidence relation occurs.Specifically, based on user to be assessed and directly
Associated level-one association user occurs corresponding text information when incidence relation and generates level-one linked character, is associated with and is used according to level-one
The family second level that corresponding text information generates user to be assessed when incidence relation occurs with the second level association user of direct correlation is closed
Join feature.Using the level-one linked character and second level linked character as a part of the feature of credit evaluation model, due to considering
There is influence of the other users of incidence relation to user credit to be assessed with user to be assessed, comment so as to improve credit
The accuracy estimated.
According to the embodiment of another aspect, a kind of device of credit evaluation is also provided.Fig. 5 is shown according to one embodiment
The schematic block diagram of device for credit evaluation.As shown in figure 5, the device 500 for credit evaluation includes: first determining single
Member 51 is configured to the determining first level-one association user with user to be assessed with incidence relation of human relation network,
In, human relation network establishes incidence relation between the user for being associated operation, and passes through operation associated corresponding text
There are the relating attributes between the user of incidence relation for information record;First generation unit 52, is configured to according to user to be assessed
With the text information in the relating attribute of the first level-one association user, at least one level-one conjunctive word of user to be assessed is obtained
It converges, and generates the first level-one linked character of user to be assessed based at least one level-one association vocabulary;Second determination unit 53,
It is configured to human relation network, determines at least one second level association user of user to be assessed, wherein second level association user
There is incidence relation with the first level-one association user;Second generation unit 54 is configured to according to the first level-one association user and each
Text information in the relating attribute of a second level association user obtains at least one second level association vocabulary of user to be assessed, and
The first second level linked character of user to be assessed is generated based at least one second level association vocabulary;Credit evaluation unit 55, matches
It is set at least based on the first level-one linked character and the first second level linked character, is assessed by credit evaluation model trained in advance
The credit rating of user to be assessed.
In the present embodiment, the first determination unit 51 can have pass with user to be assessed based on human relation network is determining
The level-one association user of connection relationship.Wherein, human relation network is the network structure based on the operation associated foundation between user,
For recording the incidence relation between user and user.There can also be relating attribute between user, which passes through upper
State operation associated corresponding text information record.In practice, operation associated between user for example can include but is not limited to
At least one of lower: plusing good friend transfers accounts, gives bonus, etc..
According to one embodiment, the text information recorded in above-mentioned relating attribute is when being associated operation between user
Phrase, sentence information.First generation unit 52 can extract association vocabulary as at least one from phrase, sentence information
Level-one is associated with vocabulary.Specifically, the first generation unit 52 can be with: carrying out word cutting to phrase, sentence information and obtains initial vocabulary;It will
Each initial vocabulary is matched with the keyword in pre-generated keyword set respectively;It will be matched in keyword set
The initial vocabulary arrived is associated with vocabulary as level-one.
According to another embodiment, the text information recorded in above-mentioned relating attribute is from above-mentioned phrase, sentence information
The association vocabulary of middle extraction, the first generation unit 52 can directly acquire recorded association vocabulary and close as at least one level-one
Join vocabulary.
According to a possible design, device 500 can also include keyword determination unit (not shown), be configured to pass through
Following methods extract the keyword in keyword set: obtaining the user's positive sample artificially demarcated and user's negative sample;Based on people
Border relational network determines the sample text information in user's positive sample and user's negative sample and the relating attribute of other users;According to
Sample text information determines the keyword in keyword set.
In one embodiment, each keyword in keyword set can also be corresponding with statistical indicator.Statistical indicator
It may include at least one of following: the number that occurs in the text information of user's positive sample, in the text envelope of user's negative sample
The probability of the number, deviation user's positive sample that occur in breath, the probability for being biased to user's negative sample.
According to a kind of embodiment, each level-one association vocabulary is corresponding with the keyword being matched in keyword set
Statistical indicator;And for some level-one association user, the first generation unit 52 is also configured as: being obtained each level-one and is closed
Join the corresponding statistical indicator of vocabulary;First predetermined process is carried out to the corresponding statistical indicator of each level-one association vocabulary, generates one
Grade linked character, wherein the first predetermined process include at least one of the following: maximizing, minimize, sum, averaging,
Seek weighted sum.
In the present embodiment, the second determination unit 53 can be based on human relation network, for 51 institute of the first determination unit
Determining each level-one association user determines at least one second level association user of user to be assessed, wherein second level association user
There is incidence relation with the level-one association user.
Second generation unit 54 can be directed to each level-one association user, according to second level determined by the second determination unit 53
Association user is obtained at least one second level association vocabulary of user to be assessed, and is generated based at least one second level association vocabulary
The second level linked character of user to be assessed.
In further carrying out example, the second generation unit 54 may include:
First generation module is configured to be associated with vocabulary for each level-one association user according to second level and generate level-one association
User and the corresponding second level-one linked character of each second level association user;
Second generation module is configured to each second level-one linked character corresponding to each level-one association user and carries out the
Two predetermined process, using processing result as the second level linked character of user to be assessed, wherein the second predetermined process include with
At least one of lower: maximizing minimizes, sums, averaging, seeking weighted sum.
Credit evaluation unit 55 can be generated at least based on the level-one linked character and second that the first generation unit 52 generates
The second level linked character that unit 54 generates assesses the credit rating of user to be assessed by credit evaluation model trained in advance.
It is worth noting that device 500 shown in fig. 5 is that device corresponding with the embodiment of the method shown in Fig. 3 is implemented
, the corresponding description in the embodiment of the method shown in Fig. 3 is equally applicable to device 500, and details are not described herein.
By apparatus above, using the level-one linked character and second level linked character as the one of the feature of credit evaluation model
Part has influence of the other users of incidence relation to user credit to be assessed with user to be assessed due to considering, thus
The accuracy of credit evaluation can be improved.
According to the embodiment of another aspect, a kind of computer readable storage medium is also provided, is stored thereon with computer journey
Sequence enables computer execute method described in conjunction with Figure 3 when the computer program executes in a computer.
According to the embodiment of another further aspect, a kind of calculating equipment, including memory and processor, the memory are also provided
In be stored with executable code, when the processor executes the executable code, realize the method in conjunction with described in Fig. 3.
Those skilled in the art are it will be appreciated that in said one or multiple examples, function described in the invention
It can be realized with hardware, software, firmware or their any combination.It when implemented in software, can be by these functions
Storage in computer-readable medium or as on computer-readable medium one or more instructions or code transmitted.
Above-described specific embodiment has carried out further the purpose of the present invention, technical scheme and beneficial effects
It is described in detail, it should be understood that being not intended to limit the present invention the foregoing is merely a specific embodiment of the invention
Protection scope, all any modification, equivalent substitution, improvement and etc. on the basis of technical solution of the present invention, done should all
Including within protection scope of the present invention.