CN103064850B - Excavate the method and system of cheating data - Google Patents
Excavate the method and system of cheating data Download PDFInfo
- Publication number
- CN103064850B CN103064850B CN201110320404.5A CN201110320404A CN103064850B CN 103064850 B CN103064850 B CN 103064850B CN 201110320404 A CN201110320404 A CN 201110320404A CN 103064850 B CN103064850 B CN 103064850B
- Authority
- CN
- China
- Prior art keywords
- cheating
- data
- ugc
- current
- fraction
- 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.)
- Active
Links
Landscapes
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention provides excavating the method and system of cheating data.Wherein, the method includes:Go out feature of practising fraud from current UGC extracting datas;Using the cheating fraction of the current UGC data of the cheating feature calculation for extracting;Judge whether value is, in the number range for representing cheating data of setting, if it is, the determination current UGC data are cheating data, the cheating data to be suppressed to the cheating fraction.
Description
Technical field
The present invention relates to data mining technology, more particularly to excavates the method and system of cheating data.
Background technology
At present, there is the content (UGC of a large number of users participation creation in Web Community:User Generated
Content) such as microblogging, blog etc..As the level of user is different with purpose, the data bulk in Web Community is will result in
It is huge, and there is the very different problem of quality.Such as, there are some users that cheating data, the cheating are issued in Web Community
Data can be such as pornographic data of low quality data etc., alternatively reach increase page light exposure by certain means or promote business
Product or the data of website (include microblogging, blog, have a talk about), this have impact on the development of Web Community.Therefore, excavate Web Community
Middle cheating data are most important.The existing conventional scheme for excavating Web Community's cheating data is given below:
Scheme one:
The program one carries out manual examination and verification by staff by the way of manual examination and verification, determines user in network
Whether the data that community issues belong to cheating data.But, the mode cost of this manual examination and verification is too high, it is difficult to process be in daily
The UGC of geometry level rapid growth, it is impossible to meet demand.
Scheme two:
The program two adopts traditional anti-cheating mode, specially:Based on link analysis, excavate by exchange link or
The data of the page rank that person's purchase link has been obtained, determine that the data excavated are cheating data.As can be seen that the program
Two for scheme one, reduces manual intervention, reduces cost.But, due to the link letter in Web Community between UGC
Manner of breathing contrast is sparse, it is difficult to a linked, diagram is built, so link analysis method single in scheme two is less suitable for, it is impossible to
Cheating data in Web Community are excavated accurately.
Therefore, a kind of method that cheating data in Web Community are accurately excavated on the basis of manual intervention is reduced is current
Technical problem urgently to be resolved hurrily.
The content of the invention
The invention provides the method and system of cheating data is excavated, accurately to excavate on the basis of manual intervention is reduced
Go out.
The technical scheme that the present invention is provided includes:
A kind of method for excavating cheating data, the method include:
Go out feature of practising fraud from current UGC extracting datas;
Using the cheating fraction of the current UGC data of the cheating feature calculation for extracting;
The cheating fraction is judged whether in the number range for representing cheating data of setting, if it is, really
The fixed current UGC data are cheating data, and the cheating data are suppressed.
A kind of system for excavating cheating data, the system include:
Feature deriving means, for going out feature of practising fraud from current UGC extracting datas;
Cheating data mining device, for utilizing the cheating fraction of the current UGC data of cheating feature calculation for extracting;
Cheating data judgment means, for judging the cheating fraction whether in the number for representing cheating data for setting
In the range of value, if it is, determining that the current UGC data are data of practising fraud, the cheating data are suppressed.
As can be seen from the above technical solutions, it is in the present invention, special from the cheating that current UGC extracting datas go out by fusion
Calculating cheating fraction is levied, determines whether current UGC data are cheating data using the cheating fraction, this is avoided artificial participation, more
The link analysis for being not needed upon current UGC data determine cheating data, realize and accurately dig on the basis of manual intervention is reduced
The purpose of data of practising fraud in excavating Web Community, saves cost.
Description of the drawings
Fig. 1 is basic flow sheet provided in an embodiment of the present invention;
Fig. 2 determines schematic diagram for machine learning model provided in an embodiment of the present invention;
Fig. 3 is machine learning model effect detection schematic diagram provided in an embodiment of the present invention;
The Data Mining System Structure figure that Fig. 4 is provided for the present invention.
Specific embodiment
In order that the object, technical solutions and advantages of the present invention are clearer, below in conjunction with the accompanying drawings with specific embodiment pair
The present invention is described in detail.
The method that the present invention is provided includes the flow process shown in Fig. 1:
Referring to Fig. 1, Fig. 1 is basic flow sheet provided in an embodiment of the present invention.As shown in figure 1, the flow process may include it is following
Step:
Step 101, goes out feature of practising fraud from current UGC extracting datas.
In this step 101, current UGC data can be snapshots of web pages, alternatively web page source file content.Wherein, the net
Page snapshot can be the content outside removing html labels in webpage, and web page source file content can be webpage source code.
When current UGC data are snapshots of web pages, the cheating feature that step 101 is extracted at least is piled up including duplicate contents
Degree and/or contact method quantity and/or advertisement score and/or low quality vocabulary score.
Wherein, duplicate contents are piled up degree and are calculated by following formula:
Here, dupscore piles up degree for duplicate contents, D be in current UGC data duplicate contents pile up window
Number, N are the window sums in current UGC data.Based on this, the present invention can pass through the suitable threshold value of setting can be to a certain degree
It is upper to solve the problems, such as to pile up degree and caused cheating due to duplicate contents.
As for contact method quantity, in the present invention, contact method mainly includes:Instant messaging account, cell-phone number, phone
Number, mailbox etc., which can be identified by pattern match.
As for advertisement score and/or low quality vocabulary score, which can adopt same algorithm to calculate.By taking advertisement as an example, then
Advertisement score can be calculated by following formula:
Wherein, docdirtyscore (d) is expressed as advertisement score, and DocLen is current UGC data, and K is current UGC numbers
According to the advertisement total quantity for including, dirtyleniFor the length of i-th advertising words, 1≤i≤K.
And when current UGC data are web page source file content, the cheating feature that step 101 is extracted at least includes the page
Number of links.
Wherein, page link, its objective is to take cribber's in users from networks community by the link on picture
The page is promoted, why step 101 extracts page link, it is therefore an objective to:On the one hand by blacklist to significantly cheating link
Suppress;On the other hand matched with the main domain of picture address by the main domain of hyperlink target page and suppressed.
Step 102, using the cheating fraction of the current UGC data of the cheating feature calculation for extracting.
Step 103, judges the cheating fraction whether in the number range for representing cheating data of setting, if
It is, it is determined that the current UGC data are cheating data, and the cheating data are suppressed.
So far, complete the flow process shown in Fig. 1.
It should be noted that the step 102 shown in Fig. 1, not manual examination and verification cheating data, but by predetermined work
The cheating feature calculation that disadvantage data mining device fusion steps 101 are extracted, this avoid artificial the inaccurate of calculating and ask
Topic, saves cost.
Wherein, data mining device of practising fraud can be machine learning model, alternatively algorithmic formula, and the present invention is not concrete to be limited
It is fixed.So that data mining device of practising fraud is as machine learning model as an example, describes below and how to determine the machine learning model, its other party
The principles such as formula such as algorithmic formula are similar to.
Referring to Fig. 2, Fig. 2 determines schematic diagram for machine learning model provided in an embodiment of the present invention.Which mainly includes following
Four parts:Data preparation, feature extraction, machine learning and effect detection.It is described one by one below:
1, data prepare:
Due in Web Community practise fraud data and it is non-cheating data ratio it is less, in order to reach the relatively flat of positive example counter-example
Weighing apparatus, obtains a UGC data acquisition system substantially practised fraud by coarse sizing first, then randomly chooses other UGC data, finally will
Two parts data fusion forms training dataset together.Meanwhile, also some UGC data are chosen from Web Community, constitute and survey
Examination data set.
The every UGC data (referred to as training UGC data) concentrated for training data, according to cheating data set in advance
Definition, by staff to each training UGC data mark cheating fraction;
Also, the every UGC data (referred to as test UGC data) concentrated for test data, also according to set in advance
The definition of cheating data, by staff to each test UGC data mark cheating fraction.
In the present invention, it is the foundation for ensureing machine learning model, has prepared the training dataset comprising 3430 UGC data
With the test data set comprising 3113 UGC data.
2, feature extraction:
In the present invention, the evaluation problem of cheating data is solved for ease of using machine learning model method, and needs will be every
One training UGC data are expressed as the combination of feature one by one.Based on this, in the present invention, go out from each training UGC extracting datas
Cheating feature.
Wherein, the cheating feature for extracting from training UGC data depends on the particular type of training UGC data, such as,
As described above, if training UGC data are snapshots of web pages, the cheating feature for extracting at least piles up journey including duplicate contents
Degree and/or contact method quantity and/or advertisement score and/or low quality vocabulary score.If training UGC data are web page source
File content, then the cheating feature for extracting at least include page link quantity.As duplicate contents pile up degree, contact method
Quantity, advertisement or pornographic vocabulary score and page link quantity etc. are all apparent cheating features, therefore, using the work
Disadvantage feature can be good at identification cheating data.
3, machine learning:
In above-mentioned data prepare, each training UGC data mark cheating fraction is given.Preferably, the cheating fraction can be
The codomain of one [0,1], describes the possibility of webpage cheating with this.Wherein, the cheating fraction of UGC data is trained to be noted as
1, then it represents that training UGC data are cheating data, train the cheating fraction of UGC data to be noted as 0, then it represents that the training
UGC data are non-cheating data.
In features described above is extracted, each training UGC data are extracted cheating feature.In the machine learning, can be by
This feature changes into the vector form that machine can be recognized, obtains characteristic vector of practising fraud.Thus, can be by each training UGC numbers
According to the vector for being expressed as N*1, wherein N is the quantity of the cheating feature extracted from training UGC data.And whole training dataset
The vector of a M*N dimension can be just expressed as, wherein N is the quantity of the cheating feature from training UGC data pick-ups, and M is training
The quantity of UGC data in data set, this will successfully excavate cheating data problem and is converted into regression problem.So, it is possible to use
The cheating fraction and the cheating characteristic vector of training UGC data marked in advance by each training UGC data sets up machine learning
Model, when implementing, can be realized using machine learning method, but suitable for this machine learning method including but not limited to certainly
Plan tree, SVMs (SVM), artificial neural network (ANN), gradient are incremented by decision tree (GBDT).
4, effect detection:
The effect detection can be described by the flow process shown in Fig. 3:
Referring to the machine learning model effect detection flow chart that Fig. 3, Fig. 3 are provided for the present invention.As shown in figure 3, the flow process
May include following steps:
Step 301, using the machine learning model as current machine learning model, for each test UGC data, from
Cheating feature is extracted in test UGC data and is expressed as vector form, obtain characteristic vector.
This feature vector is substituted into the cheating fraction that current machine learning model calculates test UGC data by step 302.
Step 303, the cheating for being marked using the cheating fraction and test UGC data of each test UGC data in advance
Fraction calculates the loss value of the test data set.
This step 303 when implementing can be:For each test UGC data, the cheating point of test UGC data is calculated
Number and test UGC data error in advance between the cheating fraction of mark;The error calculated using test UGC data, and adopt
With mean square error (MSE:Mean squared error) and coefficient correlation (SCC:squared correlation
) etc. coefficient method is calculating the loss value of the test data set.Certainly, this step 304 is possible with each test
The cheating fraction marked in advance by the cheating fraction of UGC data and test UGC data, calculates the test using other modes
The loss value of data set, the present invention is not concrete to be limited.
Step 304, judges whether the loss value meets the current demand for excavating cheating data, works as if it is, determining
Front cheating data mining device is used for the cheating fraction for calculating UGC data in the later stage, if it is not, then the optimization cheating data mining
Device, using the cheating data mining device after optimization as current cheating data mining device, return to step 302.
Optimization in this step 304 can be:Whether the cheating feature of analyzing test data integrated test UGC data omits,
Or whether the cheating fraction that marked in advance of analyzing and training data concentration training UGC data is correct etc., the present invention is not concrete
Limit.
So far, complete the flow process shown in Fig. 3.Afterwards, when whether judge UGC data is cheating data, so that it may utilize Fig. 3
The good machine learning model of the effect that detects is judged, and without the need for manually participate in, this accurately can be excavated in Web Community
Cheating data, it is to avoid the artificial subjectivity for participating in.
It should be noted that in the present invention, can not also limit UGC data, user class can be arrived more greatly with granularity
(user-level), exactly find out cheating user to be suppressed.When implementing can be:For a certain user, the user is found out
The UGC data issued within a period of time, for every UGC data, perform according to the flow process shown in above-mentioned Fig. 1, as the user
When there are setting quantity UGC data to be judged as cheating data in the UGC data issued within the time period, determine that the user is
Cheating user, suppresses to cheating user.
Above the method that the present invention is provided is described, below the system that the present invention is provided is described:
The system construction drawing of the excavation cheating data provided for the present invention referring to Fig. 4, Fig. 4.As described in Figure 4, the system bag
Include:
Feature deriving means, for going out feature of practising fraud from current UGC extracting datas;
Cheating data mining device, for utilizing the cheating fraction of the current UGC data of cheating feature calculation for extracting;
Cheating data judgment means, for judging the cheating fraction whether in the number for representing cheating data for setting
In the range of value, if it is, determining that the current UGC data are data of practising fraud, the cheating data are suppressed.
Wherein, when the current UGC data are snapshots of web pages, the cheating feature at least piles up degree including duplicate contents
And/or contact method quantity and/or advertisement score and/or low quality vocabulary score.The current UGC data are webpage source file
During content, the cheating feature at least includes page link quantity.
As shown in figure 4, the cheating data mining device may include:
Unit is set up, for utilizing the training UGC data that the training data collected is concentrated to set up cheating data mining dress
Put, using the cheating data mining device as current cheating data mining device;
Score calculating unit, concentrates for calculating the test data collected using the current cheating data mining device
The cheating fraction of each test UGC data;
Loss calculation unit, for being marked using the cheating fraction of each test UGC data and test UGC data in advance
The cheating fraction of note calculates the loss value of the test data set;
Judging unit, judges whether the loss value meets the current demand for excavating cheating data, works as if it is, determining
Front cheating data mining device is used for the cheating fraction for calculating UGC data in the later stage, if it is not, then the optimization cheating data mining
Device, using the cheating data mining device after optimization as current cheating data mining device, triggers the score calculating unit
Calculate cheating fraction.
Preferably, described foundation when unit is implemented may include:
Conversion subunit, each training UGC extracting datas for concentrating from the training data go out feature of practising fraud, and
By the cheating Feature Conversion for extracting into the form of vector, characteristic vector of practising fraud is obtained;
Subelement is set up, for the cheating fraction and the training UGC data that are marked using each training UGC data in advance
Cheating characteristic vector set up cheating data mining device.
So far, the System describe of present invention offer is provided.
As can be seen from the above technical solutions, it is in the present invention, special from the cheating that current UGC extracting datas go out by fusion
Calculating cheating fraction is levied, determines whether current UGC data are cheating data using the cheating fraction, this is avoided artificial participation, more
The link analysis for being not needed upon current UGC data determine cheating data, realize and accurately dig on the basis of manual intervention is reduced
The purpose of data of practising fraud in excavating Web Community, saves cost.
Presently preferred embodiments of the present invention is the foregoing is only, not to limit the present invention, all essences in the present invention
Within god and principle, any modification, equivalent substitution and improvements done etc. are should be included within the scope of protection of the invention.
Claims (8)
1. a kind of method for excavating cheating data, the method are applied to exist the Web Community of content UGC that user participates in creating
In, it is characterised in that the method includes:
Go out feature of practising fraud from current UGC extracting datas;
Using the cheating fraction of the current UGC data of the cheating feature calculation for extracting;
The cheating fraction is judged whether in the number range for representing cheating data of setting, if it is, determining institute
It is cheating data to state current UGC data, and the cheating data are suppressed;
The cheating fraction of the current UGC data is special using the cheating for extracting by predetermined cheating data mining device
Levy what is calculated, the cheating data mining device is determined by following steps:
A, sets up cheating data mining device using the training UGC data that the training data collected is concentrated, by the cheating number
According to excavating gear as current cheating data mining device;
B, calculates the work that the test data collected concentrates each test UGC data using the current cheating data mining device
Disadvantage fraction;
C, the cheating fraction for being marked using the cheating fraction and test UGC data of each test UGC data in advance calculate institute
State the loss value of test data set;
D, judges whether the loss value meets the current demand for excavating cheating data, if it is, determining current cheating data
Excavating gear is used for the cheating fraction for calculating UGC data in the later stage, if it is not, then the optimization cheating data mining device, will be excellent
Cheating data mining device after change is used as current cheating data mining device, return to step B.
2. method according to claim 1, it is characterised in that the current UGC data are snapshots of web pages;
The cheating feature at least piles up degree and/or contact method quantity and/or advertisement score and/or low including duplicate contents
Quality vocabulary score.
3. method according to claim 1, it is characterised in that the current UGC data are web page source file content;
The cheating feature at least includes page link quantity.
4. method according to claim 1, it is characterised in that the foundation in step A includes:
The each training UGC extracting datas concentrated from the training data go out feature of practising fraud, and by the cheating feature for extracting
The form of vector is converted into, characteristic vector of practising fraud is obtained;
The cheating fraction for being marked using each training UGC data in advance and the cheating characteristic vector of training UGC data are set up
Cheating data mining device.
5. a kind of system for excavating cheating data, the system are applied to exist the Web Community of content UGC that user participates in creating
In, it is characterised in that the system includes:
Feature deriving means, for going out feature of practising fraud from current UGC extracting datas;
Cheating data mining device, for utilizing the cheating fraction of the current UGC data of cheating feature calculation for extracting;
Cheating data judgment means, for judging the cheating fraction whether in the numerical value model for representing cheating data for setting
In enclosing, if it is, determining that the current UGC data are data of practising fraud, the cheating data are suppressed;
The cheating data mining device includes:
Unit is set up, for utilizing the training UGC data that the training data collected is concentrated to set up cheating data mining device, will
The cheating data mining device is used as current cheating data mining device;
Score calculating unit, concentrates each for calculating the test data collected using the current cheating data mining device
The cheating fraction of test UGC data;
Loss calculation unit, for what is marked using the cheating fraction and test UGC data of each test UGC data in advance
Cheating fraction calculates the loss value of the test data set;
Judging unit, judges whether the loss value meets the current demand for excavating cheating data, if it is, determining current work
Disadvantage data mining device is used for the cheating fraction for calculating UGC data in the later stage, if it is not, then the optimization cheating data mining dress
Put, using the cheating data mining device after optimization as current cheating data mining device, trigger the score calculating unit meter
Can be regarded as disadvantage fraction.
6. system according to claim 5, it is characterised in that the current UGC data are snapshots of web pages;
The cheating feature at least piles up degree and/or contact method quantity and/or advertisement score and/or low including duplicate contents
Quality vocabulary score.
7. system according to claim 5, it is characterised in that the current UGC data are web page source file content;
The cheating feature at least includes page link quantity.
8. system according to claim 5, it is characterised in that the unit of setting up includes:
Conversion subunit, each training UGC extracting datas for concentrating from the training data go out feature of practising fraud, and will carry
Form of the cheating Feature Conversion of taking-up into vector, obtains characteristic vector of practising fraud;
Subelement is set up, the work of practise fraud fraction and training UGC data for being marked using each training UGC data in advance
Disadvantage characteristic vector sets up cheating data mining device.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201110320404.5A CN103064850B (en) | 2011-10-20 | 2011-10-20 | Excavate the method and system of cheating data |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201110320404.5A CN103064850B (en) | 2011-10-20 | 2011-10-20 | Excavate the method and system of cheating data |
Publications (2)
Publication Number | Publication Date |
---|---|
CN103064850A CN103064850A (en) | 2013-04-24 |
CN103064850B true CN103064850B (en) | 2017-04-05 |
Family
ID=48107480
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201110320404.5A Active CN103064850B (en) | 2011-10-20 | 2011-10-20 | Excavate the method and system of cheating data |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN103064850B (en) |
Families Citing this family (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106326498A (en) * | 2016-10-13 | 2017-01-11 | 合网络技术(北京)有限公司 | Cheat video identification method and device |
CN108512682B (en) * | 2017-02-28 | 2021-02-26 | 腾讯科技(深圳)有限公司 | Method and device for determining false terminal identification |
CN109241523B (en) * | 2018-08-10 | 2020-12-11 | 北京百度网讯科技有限公司 | Method, device and equipment for identifying variant cheating fields |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101981590A (en) * | 2008-03-31 | 2011-02-23 | 雅虎公司 | Access to trusted user-generated content using social networks |
CN102098332A (en) * | 2010-12-30 | 2011-06-15 | 北京新媒传信科技有限公司 | Method and device for examining and verifying contents |
Family Cites Families (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20090319629A1 (en) * | 2008-06-23 | 2009-12-24 | De Guerre James Allan | Systems and methods for re-evaluatng data |
-
2011
- 2011-10-20 CN CN201110320404.5A patent/CN103064850B/en active Active
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101981590A (en) * | 2008-03-31 | 2011-02-23 | 雅虎公司 | Access to trusted user-generated content using social networks |
CN102098332A (en) * | 2010-12-30 | 2011-06-15 | 北京新媒传信科技有限公司 | Method and device for examining and verifying contents |
Also Published As
Publication number | Publication date |
---|---|
CN103064850A (en) | 2013-04-24 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN104239485B (en) | A kind of dark chain detection method in internet based on statistical machine learning | |
CN107038480A (en) | A kind of text sentiment classification method based on convolutional neural networks | |
CN105447206B (en) | New comment object identifying method and system based on word2vec algorithms | |
CN107247947A (en) | Face character recognition methods and device | |
CN103617435B (en) | Image sorting method and system for active learning | |
CN110263166A (en) | Public sentiment file classification method based on deep learning | |
CN107563995A (en) | A kind of confrontation network method of more arbiter error-duration models | |
CN105654144A (en) | Social network body constructing method based on machine learning | |
CN104268134B (en) | Subjective and objective classifier building method and system | |
CN109005145A (en) | A kind of malice URL detection system and its method extracted based on automated characterization | |
CN102789449B (en) | The method and apparatus that comment text is evaluated | |
CN103064850B (en) | Excavate the method and system of cheating data | |
CN107438083B (en) | Detection method for phishing site and its detection system under a kind of Android environment | |
CN107729908A (en) | A kind of method for building up, the apparatus and system of machine learning classification model | |
CN108664474A (en) | A kind of resume analytic method based on deep learning | |
CN110135282A (en) | A kind of examinee based on depth convolutional neural networks model later plagiarizes cheat detection method | |
CN106778878A (en) | A kind of character relation sorting technique and device | |
CN110175851A (en) | A kind of cheating detection method and device | |
CN104156447A (en) | Intelligent social platform advertisement early warning and handling method | |
CN106021290A (en) | Method for social network association excavation based on multi-scale geographic information | |
CN104063720A (en) | Method for detecting images of prohibited commodities of e-commerce websites based on deep Boltzmann machine | |
CN106681989A (en) | Method for predicting microblog forwarding probability | |
CN103714093B (en) | A kind of method for digging and device of the website emphasis page | |
CN110084609A (en) | A kind of transaction swindling behavior depth detection method based on representative learning | |
CN108491719A (en) | A kind of Android malware detection methods improving NB Algorithm |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
C10 | Entry into substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |