CN109165950A - A kind of abnormal transaction identification method based on financial time series feature, equipment and readable storage medium storing program for executing - Google Patents
A kind of abnormal transaction identification method based on financial time series feature, equipment and readable storage medium storing program for executing Download PDFInfo
- Publication number
- CN109165950A CN109165950A CN201810909752.8A CN201810909752A CN109165950A CN 109165950 A CN109165950 A CN 109165950A CN 201810909752 A CN201810909752 A CN 201810909752A CN 109165950 A CN109165950 A CN 109165950A
- Authority
- CN
- China
- Prior art keywords
- time series
- data
- financial
- transaction
- account
- 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.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q20/00—Payment architectures, schemes or protocols
- G06Q20/38—Payment protocols; Details thereof
- G06Q20/40—Authorisation, e.g. identification of payer or payee, verification of customer or shop credentials; Review and approval of payers, e.g. check credit lines or negative lists
- G06Q20/401—Transaction verification
- G06Q20/4016—Transaction verification involving fraud or risk level assessment in transaction processing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q20/00—Payment architectures, schemes or protocols
- G06Q20/38—Payment protocols; Details thereof
- G06Q20/40—Authorisation, e.g. identification of payer or payee, verification of customer or shop credentials; Review and approval of payers, e.g. check credit lines or negative lists
- G06Q20/401—Transaction verification
- G06Q20/4018—Transaction verification using the card verification value [CVV] associated with the card
Landscapes
- Business, Economics & Management (AREA)
- Engineering & Computer Science (AREA)
- Accounting & Taxation (AREA)
- Computer Security & Cryptography (AREA)
- Finance (AREA)
- Strategic Management (AREA)
- Physics & Mathematics (AREA)
- General Business, Economics & Management (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Financial Or Insurance-Related Operations Such As Payment And Settlement (AREA)
Abstract
The present invention provides a kind of abnormal transaction identification method based on financial time series feature; equipment and readable storage medium storing program for executing; largely financial transaction flowing water information datas to be detected doubtful abnormal or relevant to certain determining exception accounts can be utilized; by neural network model extracted in self-adaptive financial time series feature, be then based on linear layer and softmax layers of operation in neural network carry out Transaction Account number to be detected whether be multiple level marketing account Classification and Identification.Exception financial transaction recognition methods proposed by the present invention can be based on SoftSeq2Seq-Attention neural network model extracted in self-adaptive financial time series feature, reduce the investment of labor-intensive characteristics engineering to a certain extent.Using compared with single type financial transaction pipelined data and less feature, abnormal financial account detection recognition effect well can be obtained.
Description
Technical field
The present invention relates to financial transaction field more particularly to a kind of abnormal transaction identifications based on financial time series feature
Method, equipment and readable storage medium storing program for executing.
Background technique
Abnormal financial transaction refers to that there are abnormal case or the finance of feature for transaction amount, trading frequency, loco etc.
Transaction.Abnormal financial transaction includes many fields, such as money laundering, credit card fraud, illegal fund collection, multiple level marketing.And pyramid schemes one
It is directly one of financial security field urgent problem to be solved, is in the nature by the offline illegal transfer for realizing finance of development and to gather
Collection upsets social economic order, endangers personal safety, social stability.The analysis of data of financial transaction flow point is to carry out the knowledge of multiple level marketing account
A kind of effective means that other or multiple level marketing organizational structure is excavated.But identification multiple level marketing account is analysed currently based on data of financial transaction flow point
Method, be also largely dependent upon artificial constructed rule system, depend on artificial judgment, spend human and material resources.Such as gold
Melt the mathematical statistics method of time series or uses the conventional methods such as the multiple level marketings network theory such as sociology, marketing analysis, it can not
Effectively handle the large-scale dataset continued to bring out.
Currently, based on machine learning or the abnormal financial transaction of deep learning method identification in anti money washing, credit card fraud
The fields such as detection have been achieved with remarkable progress.Many researchs by the neural network models such as RBF, CNN be applied to credit card fraud or
Anti money washing detection field and significant effect, in addition, the research of the methods of naive Bayesian and support vector machines in anti-money laundering field
In also using very extensive.But Study of recognition of the country in relation to multiple level marketing financial transaction account number is also relatively fewer.Existing multiple level marketing
Tissue identification or research method are built upon on the basis of multiple level marketing hoc network topologies feature, using branch's tree-model or shellfish more
Traditional data analysing methods such as Ye Si estimation.These methods in some scenarios open up by available more accurate multiple level marketing network
Structure or multiple level marketing tissue signature are flutterred, but the multiple level marketing recognition methods based on building network topology characteristic is often possible to need largely
The multiple level marketing related data of complicated type, i.e., various other categorical datas in addition to financial transaction pipelined data, as communicating data,
Note data, wechat chat data etc..On the one hand, Preprocessing early period of these multiple types of data will lead to more complicated
Labor-intensive characteristics Engineering Task.On the other hand, in the situation that data volume lacks or multiple level marketing group organization data is sufficiently complete
Under, it is likely that the analysis result for causing the recognition methods based on multiple level marketing network topology characteristic to obtain is inaccurate, availability is low.And
Currently, there is no one way to adaptively extracting financial time series feature for passing in conjunction with RNN neural network model
The identification of financial account is sold, so that complicated labor-intensive characteristics Engineering Task be alleviated or avoided.
Summary of the invention
In order to overcome the deficiencies in the prior art described above, the present invention provides a kind of exception based on financial time series feature
Transaction identification method, method include:
Step 1 carries out data prediction to the raw financial transaction journal data set of input, extracts raw financial transaction
The cleaning data and crucial item data that pipelined data is concentrated, obtain key item data acquisition system D；
Step 2 constructs financial time series, constructs financial time series data set based on key item data acquisition system D
Dfinput；
Step 3 is based on financial time series data set Dfinput, data are carried out according to determining multiple level marketing card number listing file
Mark；The financial time series data set Train marked is inputted into SoftSeq2Seq-Attention neural network model,
Carry out model training and financial time series feature extraction；
Step 4 carries out detection identification to account, identifies financial transaction flowing water information, and construct financial transaction stream to be detected
Water message data setAccording to step 1 to financial transaction flowing water message data set to be detectedPretreatment operation is carried out, is obtained
To pretreatment operation result key item data acquisition system Dtest, financial time series data set is then constructed by step 2
It willMiddle data are input in trained SoftSeq2Seq-Attention neural network model, extract gold
Melt time series feature vector set
Based on Decoder middle layer financial time series feature vector, by linear linear layer and softmax layers of progress finance
The Classification and Identification of Transaction Account number obtains account ciCorresponding finance time transaction sequence set
Classification results Making by Probability Sets, then according to account detection recognition method carry out account ciThe calculating of final classification result, with
It is the abnormal probability value of multiple level marketing account to corresponding account.
In the present invention, in step 1, extracts the transaction card number in transaction journal data, trade date, transaction amount, plucks
Illustrate, the basic data as feature extraction work；Card number of trading is known as the mark ID of financial account for multiple level marketing account
Not；
By trade date by " YYYYMMDDhhmmss " unified formatting processing, transaction amount combination receipt and payment flag bit is received
Paying mark is that " into " then the amount of money is positive number, and receipt and payment mark is that " out ", then the amount of money was negative；After format normalized, friendship is washed
Easy transaction journal data of the amount of money absolute value less than 50, and transaction card number are empty transaction journal data.
In the present invention, step 2 further include: after above-mentioned processing work, the transaction journal letter based on critical data item
Breath, the method for constructing financial time series are as follows:
1, the transaction card number set C={ c in transaction journal information is counted1,c2,...,cn, wherein n is card number sum；
2, with card number ciFor key assignments key, i.e. financial account identifies ID, ci∈ C, by ciCorresponding whole transaction journal information
Construct list li, liFor ciCorresponding value contents value, li[m]=[ci,timem,moneym,summarym], m is only indicated here
List liIn some element subscript.Obtain key-value pair data set D={ d1,d2,...,dn, if di∈ D, then di=(ci,
li)；
3, to all di∈ D, by its liList carries out ascending sort according to this content of trade date, i.e., by transaction flow
Water list is in chronological sequence sequentially resequenced, by liIt is updated to ranking results list；
4, to all di∈ D, di=(ci,li), utilize its li[m]=[ci,timem,moneym,summarym] content,
Construct initial input vectorWherein, moneymFor former transaction amount value；
tivecmFor timemCorresponding vector indicates；summvecmFor transaction summarization item summarymCorresponding vector indicates.tivecmLife
Be at method, based on exchange hour set whole in D call sklearn HashingVectorizer method it is carried out to
Quantization means, vector dimension take 5 dimensions.summvecmGeneration method be to be called based on transaction summarization set whole in D
The HashingVectorizer method of sklearn carries out vectorization expression to it, and vector dimension takes 10 dimensions.ThenBy 16
Initial characteristics vector is tieed up to constitute；
5, by step 4) method, data set D is obtainedinput, haveInput vectorHere m only indicates listIn some element subscript；
6, it is sampled using sliding window method and generates fixed length financial time series data, while expanding data amount can be played the role of.
Firstly, washing DinputInLength is less than 15Key-value pair data；
Use length for 50, the sliding window that interval steps are 25 is right from the front to the backCarry out data segmentation；If the last one
The data length that sliding window includes then carries out mending 0 operation less than 50 but greater than 15；Otherwise, give up this partial data；
Generation hasBelong to account ciTime series data set；
One time series data content isID is identified by card number ciAnd coding j is collectively constituted；
7, it can be obtained by step 6),J=1,2 ..., k,Here m only indicates listIn some element subscript；IfI=1,2 ..., n, then DfinputFor constructed financial time series data set.
In the present invention, step 3 further include:
The list entries of SoftSeq2Seq-Attention modelFor the financial time sequence of corresponding fixed length
ColumnWhereinM=1,2 ..., Tx.Encoder and
The RNN hidden layer h of the part Decoder<i>And s<i>It is all made of standard GRU (Gated Recurrent Unit) gating cycle unit；
Based on using local attention mechanism, taking attention position is pt=5i+1, wherein i=1,2 ..., 9, window is big
Small Dsize=4 or pt=2i+1, Dsize=2, then attention mechanism intermediate result s<t>Pass through [pt-Dsize,pt+Dsize]
Information and s in window<t-1>It acquires；
s<t>Hidden layer intermediate result vectorAfter splicing, as the financial time series feature extracted
Vector indicates, or indicates using the summation of the vector or average operation result as financial time series feature vector；It will be final
The expression of financial time series feature vector is fed for a linear layer, this linear layer exports length and is equal to account number classification number | C ' |；Line
Property layer result input the softmax layers of calculating for carrying out corresponding class probability again；
It is concentrated in the training data of the neural network model, the output y of the part Decoder is card number ciCorresponding classification knot
Fruit vector, yp=(yp1,yp2)；yp1、yp2Respectively indicate financial time seriesCorresponding account ciBe classified as multiple level marketing or
Normal probability；
The softmax layer calculation method of the part Decoder is as follows:
In SoftSeq2Seq-Attention neural network model, input layerIt is tieed up for 50, wherein often
One-dimensional xiIt is made of 16 dimension initial characteristics vectors；Loss function is the intersection entropy loss letter that financial time series correspond to account number classification
Number, it may be assumed that
Wherein, S is training dataset, and C ' is account number classification number, and s is financial time series data, pc′(s) prediction s is indicated
Corresponding account is the probability of corresponding classification account；Indicate whether c ' class is the correct account number classification of s, value is 0 here
Or 1；When model training, using backpropagation and stochastic gradient descent algorithm undated parameter；
Training dataset is constructed, according to multiple level marketing card number listing file, to financial time series data set DfinputIn finance
Time series initial input characteristic vector dataClassification annotation is carried out, training set Train=is constructed
(x1,y1),(x2,y2),...,(xn,yn)；, (x herei,yi) only indicate training data isomery finance feature With
Corresponding account type marks yi, yiValue represents multiple level marketing account for 0 or 1,1, and 0 represents normal account.If labeled data collection, then
Direct construction training set Train=(x1,y1),(x2,y2),...,(xn,yn)；
It is required that two class data volume specific gravity are preferably between 1:1 to 1:2 in training dataset Train；Further division
Train is training set train and verifying collection test, is divided according to the ratio of 7:3, train specific gravity is 7/10.
In the present invention, in step 4, account detection identification method includes, according to data prediction and format method for normalizing
It arranges data to be tested collection and financial time series construction method constructs financial time series data set to be detected, it will be to be checked
It surveys financial time series data set and inputs SoftSeq2Seq-Attention neural network model, carry out account number classification identification, institute
Obtaining last softmax layers of output result is account ciCorresponding wherein oneTransaction data
Classification results probability；
According to account ciIt is correspondingThe all classification probability of outcome of data in set
Calculate its average valueObtain ciCorresponding final classification probability of outcomeC is determined according to thisiFinal classification.
A kind of equipment for realizing the abnormal transaction identification method based on financial time series feature, comprising:
Memory, for storing computer program and the abnormal transaction identification method based on financial time series feature；
Processor, for executing the computer program and based on the abnormal transaction identification side of financial time series feature
Method, the step of to realize abnormal transaction identification method based on financial time series feature.
A kind of computer readable storage medium with the abnormal transaction identification method based on financial time series feature, institute
It states and is stored with computer program on computer readable storage medium, the computer program is executed by processor to realize based on gold
The step of melting the abnormal transaction identification method of time series feature.
As can be seen from the above technical solutions, the invention has the following advantages that
The abnormal financial transaction recognition methods based on time series feature that invention broadly provides a kind of, method being capable of benefits
With largely financial transaction flowing water information datas to be detected doubtful abnormal or relevant to certain determining exception accounts, pass through nerve net
Network model adaptation extracts financial time series feature, be then based in neural network linear layer and softmax layers of operation into
Row Transaction Account number to be detected whether be multiple level marketing account Classification and Identification.
Exception financial transaction recognition methods proposed by the present invention can be based on SoftSeq2Seq-Attention neural network
Model adaptation extracts financial time series feature, reduces the investment of labor-intensive characteristics engineering to a certain extent.It utilizes
Financial transaction pipelined data and less feature compared with single type can obtain abnormal financial account detection identification effect well
Fruit.This method can provide auxiliary for the multiple level marketing investigation of relevant staff and study and judge information, improve working efficiency, when saving
Between.With the discovery of more multiple level marketing flag datas, disaggregated model can obtain further perfect, detection recognition result accuracy rate
There is increase trend.
Detailed description of the invention
In order to illustrate more clearly of technical solution of the present invention, attached drawing needed in description will be made below simple
Ground introduction, it should be apparent that, drawings in the following description are only some embodiments of the invention, for ordinary skill
For personnel, without creative efforts, it is also possible to obtain other drawings based on these drawings.
Fig. 1 is the abnormal transaction identification method flow diagram based on financial time series feature；
Fig. 2 is module data flow graph；
Fig. 3 is SoftSeq2Seq-Attention model structure schematic diagram；
Fig. 4 is abnormal financial transaction recognition methods flow chart.
Specific embodiment
The present invention provides a kind of abnormal transaction identification method based on financial time series feature, as shown in Figure 1,
S1 carries out data prediction to the raw financial transaction journal data set of input, extracts raw financial transaction journal
Cleaning data and crucial item data in data set, obtain key item data acquisition system D；
S2 constructs financial time series, constructs financial time series data set D based on key item data acquisition system Dfinput；
S3 is based on financial time series data set Dfinput, data mark is carried out according to determining multiple level marketing card number listing file
Note；The financial time series data set Train marked is inputted into SoftSeq2Seq-Attention neural network model, into
Row model training and financial time series feature extraction；
S4 carries out detection identification to account, identifies financial transaction flowing water information, and constructs financial transaction flowing water letter to be detected
Cease data setAccording to step 1 to financial transaction flowing water message data set to be detectedPretreatment operation is carried out, pre- place is obtained
Manage operating result key item data acquisition system Dtest, financial time series data set is then constructed by step 2It willIn
Data are input in trained SoftSeq2Seq-Attention neural network model, extract financial time series feature vector
SetBased on Decoder middle layer financial time series feature vector, by linear linear layer and softmax layers of progress gold
The Classification and Identification for melting Transaction Account number obtains account ciCorresponding finance time transaction sequence set
Classification results Making by Probability Sets, then according to account detection recognition method carry out account ciThe calculating of final classification result, with
It is the abnormal probability value of multiple level marketing account to corresponding account.
In the present invention, a large amount of financial transactions to be detected doubtful abnormal or relevant to the abnormal account of certain determinations can be utilized
Flowing water information data generates customized financial time series data and using customized modification based on Attention mechanism
Seq2Seq neural network model extracted in self-adaptive financial time series feature is known by softmax layers of row multiple level marketing account number classification
Not.Abnormal financial time series feature extraction based on financial transaction pipelined data；Exception based on financial transaction pipelined data
Account detection identification.
This method can be divided into 3 modules, be from left to right data prediction respectively as shown in Fig. 2 module data flow graph
Module, financial time series characteristic extracting module, account detect identification module.
Data preprocessing module: for the raw financial transaction journal data of input, data cleansing and data item lattice are carried out
Formula normalization.It extracts in transaction journal data, such as the several critical datas of time, the amount of money with the suspicious financial transaction degree of correlation greatly
?.It is then based on the key item pipelined data of each transaction card number, constructs financial time series.Financial time series feature extraction
Module: this module financial time series feature by the customed Seq2Seq neural network model based on Attention mechanism from
It adapts to extract building, raw financial time series data collection has been marked by input and has carried out SoftSeq2Seq-Attention mind
Through network model parameter training, gold then can be extracted by the SoftSeq2Seq-Attention neural network model built
Melt the expression of time series feature vector.Account detects identification module: based on the SoftSeq2Seq-Attention nerve built
Network model carries out financial time series to the financial time series data that financial transaction flowing water message data set to be detected generates
Feature extraction and Classification and Identification, i.e., on the basis of extracted in self-adaptive financial time series feature vector using softmax layers into
The identification of row multiple level marketing account number classification.
Input data of the invention is raw financial transaction journal data, generally comprises transaction account in transaction journal data
Number, transaction card number, transaction amount, opponent's account, explanatory memorandum, the multinomial Transaction Information such as loco.And it is used for abnormal financial
When transaction identification, such as the contribution degree of IP address, teller's data item of trading is little, therefore to carry out critical data item information
It extracts.In this method, transaction card number, trade date, the transaction amount, these four passes of explanatory memorandum in transaction journal data are extracted
Key data item, the basic data as further feature extraction work.Wherein, transaction card number is used as the mark ID of financial account
It is identified in multiple level marketing account.
In addition, the format of trade date or transaction amount will be different in the transaction journal data that different channels obtain.
In this method, by trade date by " YYYYMMDDhhmmss " unified formatting processing, example such as " 20160405103419 ", table
34 divide 19 seconds when showing 5 days 10 April in 2016.Transaction amount combination receipt and payment flag bit, receipt and payment mark are that " into " then the amount of money is positive number,
Receipt and payment mark is that " out ", then the amount of money was negative.After format normalized, transaction amount absolute value is also washed less than 50
Transaction journal data, and transaction card number are empty transaction journal data.
After above-mentioned processing work, the transaction journal information based on critical data item constructs the side of financial time series
Method is as follows:
1) the transaction card number set C={ c in transaction journal information is counted1,c2,...,cn, wherein n is that card number is total (no
Do repetition statistics).
2) with card number ciFor key assignments key, i.e. financial account identifies ID, ci∈ C, by ciCorresponding whole transaction journal information
Construct list li, liFor ciCorresponding value contents value, li[m]=[ci,timem,moneym,summarym], m is only indicated here
List liIn some element subscript.Obtain key-value pair data set D={ d1,d2,...,dn, if di∈ D, then di=(ci,
li)。
3) to all di∈ D, by its liList carries out ascending sort according to this content of trade date, i.e., by transaction flow
Water list is in chronological sequence sequentially resequenced, by liIt is updated to ranking results list.
4) to all di∈ D, di=(ci,li), utilize its li[m]=[ci,timem,moneym,summarym] content,
Construct initial input vectorWherein, moneymFor former transaction amount value；
tivecmFor timemCorresponding vector indicates；summvecmFor transaction summarization item summarymCorresponding vector indicates.tivecmLife
Be at method, based on exchange hour set whole in D call sklearn HashingVectorizer method it is carried out to
Quantization means, vector dimension take 5 dimensions.summvecmGeneration method be to be called based on transaction summarization set whole in D
The HashingVectorizer method of sklearn carries out vectorization expression to it, and vector dimension takes 10 dimensions.ThenBy 16
Initial characteristics vector is tieed up to constitute.
5) by step 4) method, data set D is obtainedinput, haveInput vectorHere m only indicates listIn some element subscript.
6) fixed length financial time series data are generated using the sampling of sliding window method, while expanding data amount can be played the role of.
Firstly, washing DinputInLength is less than 15Key-value pair data.Finally, using length for 50, interval steps are
25 sliding window is right from the front to the backCarry out data segmentation.If the data length that the last one sliding window includes is less than 50 but is greater than
15, then it carries out mending 0 operation；Otherwise, give up this partial data.At this point, generating hasBelong to
Account ciTime series data set.At this point, a time series data content isID is identified by card number ciWith
And coding j is collectively constituted.
7) it can be obtained by step 6),J=1,2 ..., k,Here m only indicates listIn some element subscript.IfI=1,2 ..., n, then DfinputFor constructed financial time series data set.
In the present invention, time series data is a kind of special sequence data, is made of a series of values changed over time,
The time interval of these values can be equidistant or unequal spacing.Time series data can be continuous or discrete.
Finance data belongs to the coefficient product of many factors such as society, economy, psychology, therefore, financial time series data it is interior
Hold and trend also contains information or feature abundant.This method uses the Seq2Seq neural network based on Attention mechanism
Model, the feature vector for carrying out extracted in self-adaptive financial time series data indicate.
Seq2Seq (Sequence to Sequence) neural network model belongs to RNN (Recurrent neural
Network one kind).It is made of encoder-decoder frame, working mechanism are as follows: sequence will be inputted using encoder first
Column are encoded to a vector space, and the intermediate vector for obtaining a fixed dimension indicates.Then it is decoded, is obtained using decoder
Required output.The advantages of model be very flexibly, do not limit the neural network type that encoder, decoder are used, and
And this is the process of one end-to-end (end-to-end), and the understanding process of input data and feature vector are generated and closed one
It rises, is not separated by processing.And the shortcomings that this model is that encoder was provided is all the vector of a fixed dimension, there are letters
Breath loss, and the long-term memory scarce capacity to input data of RNN, may learn farther away less than distance in input in this way
The big useful information of contribution degree.Therefore, to solve these problems, attention mechanism, i.e. attention mechanism are introduced,
That is different concern weights are given to different inputs, when generating every dimensional feature value of result feature vector sufficiently to dig
Dig the characteristic information in list entries.
As shown in Fig. 3 SoftSeq2Seq-Attention model structure, this method is customed to have modified standard
The part decoder of Seq2Seq neural network structure joined linear layer and softmax layers of composition and classification model.Additionally plus
Enter attention mechanism, the local attention mechanism used herein for being analogous to local attention, relatively
Global attention mechanism reduces model complexity to a certain extent, simplifies model part parameter.
The list entries of SoftSeq2Seq-Attention modelFor the financial time sequence of corresponding fixed length
ColumnWhereinM=1,2 ..., Tx。Encoder
With the RNN hidden layer h of the part Decoder<i>And s<i>It is all made of standard GRU (Gated Recurrent Unit) gating cycle list
Member.
Since using local attention mechanism, taking attention position is pt=5i+1, wherein i=1,2 ..., 9, window is big
Small Dsize=4 (can also attempt pt=2i+1, Dsize=2), then attention mechanism intermediate result s<t>Pass through [pt-Dsize,
pt+ Dsize] information and s in window<t-1>It acquires.s<t>Hidden layer intermediate result vector After splicing, make
Indicate for the financial time series feature vector that extracts (additionally can using the summation of these vectors or average operation result as
Financial time series feature vector indicates), the expression of final financial time series feature vector is fed for a linear layer, this is linear
Layer output length is equal to account number classification number | C ' |.Linear layer result inputs the softmax layers of calculating for carrying out corresponding class probability again.
It is concentrated in the training data of the neural network model, the output y of the part Decoder is card number ciCorresponding classification knot
Fruit vector, yp=(yp in this method1,yp2)。yp1、yp2Respectively indicate financial time seriesCorresponding account ciClassification
For multiple level marketing or normal probability.The softmax layer calculation method of the part Decoder is as follows:
In SoftSeq2Seq-Attention neural network model shown in Fig. 3, input layerIt is tieed up for 50,
In per one-dimensional xiIt is made of 16 dimension initial characteristics vectors.Loss function is the cross entropy damage that financial time series correspond to account number classification
Lose function, it may be assumed that
Wherein, S is training dataset, and C ' is account number classification number (only ' multiple level marketing ' and ' normal ' two class), and s is the financial time
Sequence data, pc′(s) indicate that the corresponding account of prediction s is the probability of corresponding classification account.Indicate c ' class whether be s just
True account number classification, value is 0 or 1 here.When model training, ginseng is updated using backpropagation and stochastic gradient descent algorithm
Number.
About the building of training dataset, according to multiple level marketing card number listing file, to financial time series data set DfinputIn
Financial time series initial input characteristic vector dataClassification annotation is carried out, training set is constructed
Train=(x1,y1),(x2,y2),...,(xn,yn)；, (x herei,yi) only indicate training data isomery finance featureY is marked with corresponding account typei, yiValue represents multiple level marketing account for 0 or 1,1, and 0 represents normal account.If having marked
Data set is infused, then direct construction training set Train=(x1,y1),(x2,y2),...,(xn,yn)；.It is required that training dataset
In Train, two class data volume specific gravity are preferably between 1:1 to 1:2.Then, further division Train be training set train and
Verifying collection test, is divided according to the ratio of 7:3, and train specific gravity is 7/10.
In the present invention, after the method completes the training and building of SoftSeq2Seq-Attention neural network model,
Account detection identification operation can be carried out to financial transaction flowing water message data set to be detected using this neural network model.
Data prediction and format method for normalizing arrange data to be tested collection in account detection identification, then according to finance
Time series construction method constructs financial time series data set to be detected, then that financial time series data set to be detected is defeated
Enter SoftSeq2Seq-Attention neural network model, carries out account number classification identification, the last softmax layers of output knot of gained
Fruit is account ciCorresponding wherein oneThe classification results probability of transaction data.Then, according to
Account ciIt is correspondingThe all classification probability of outcome of data calculates its average value in setObtain ciCorresponding final classification probability of outcomeC is determined according to thisiFinal classification.
In the present invention, the intermediate result vector of Decoder hidden layer outputAs neural network model
The correspondence financial time series feature vector of output indicates.This characteristic synthetic is extracted the local feature of input vector and one
Determine degree and save long-term memory ability using GRU unit, it is contemplated that the influence between long-term information, available expression energy
Power and the more preferable financial time series feature vector of classifying quality indicate.And Financial time series feature to
Amount can also be in the abnormal transaction identification method of other Feature Engineerings.
In the present invention, steps are as follows for the overall flow of abnormal financial transaction recognition methods:
Step 1 carries out data prediction, does data cleansing and key to the raw financial transaction journal data set of input
Item data extracts, and obtains key item data acquisition system D.
Step 2 carries out financial time series data set D based on D according to financial time series construction stepfinputStructure
It builds.
Step 3 is based on financial time series data set Dfinput, data are carried out according to determining multiple level marketing card number listing file
Mark.The financial time series data set Train marked is inputted into SoftSeq2Seq-Attention neural network model,
Carry out model training and financial time series feature extraction.
Step 4, account detection identification, for financial transaction flowing water message data set to be detectedAccording to step 1 into
Row pretreatment operation obtains pretreatment operation result key item data acquisition system Dtest, financial time sequence is then constructed by step 2
Column data collectionIt willMiddle data are input to trained SoftSeq2Seq-Attention neural network model
In, financial time series feature vector set can be extractedIt is then based on Decoder middle layer financial time series feature
Vector obtains account c by the Classification and Identification of linear linear layer and softmax layers of progress financial transaction account numberiCorresponding gold
Melt time transaction sequence setClassification results Making by Probability Sets, then carry out account ciMost
The calculating of whole classification results, to obtain corresponding account as the abnormal probability value of multiple level marketing account.Abnormal financial transaction of the invention is known
Other method flow, shown in following Fig. 4 exception financial transaction recognition methods flow chart.
The present invention provides a practical case:
Case: certain user inputs raw financial transaction journal labeled data collectionWith raw financial transaction journal number to be detected
According to collection
Step 1, it is assumed thatMiddle multiple level marketing and normal account respectively have 100, and each account has 100 a plurality of transaction journal data,
I.e. multiple level marketing has 10,000 or so transaction journal data with normal respectively.There are 50 accounts to be detected, each account also has about
100 transaction journal data amount to about 5000 transaction journal data.
Step 2, respectively to inputData andData prediction is carried out, data cleansing and crucial item data are done
It extracts, respectively obtains processing result key item data acquisition system D and Dtest。
Step 3 is based on D and DtestFinancial time series building is carried out respectively, obtains building result financial time series number
According to collection DfinputWithNext, entering step four.
Step 4 is based on financial time series data set Dfinput, training set Train is constructed, by financial time series training
Data set inputs SoftSeq2Seq-Attention neural network model, carries out model training, obtains trained
SoftSeq2Seq-Attention neural network model andFinancial time series feature vector set.
Step 5, by data setIn data be input in SoftSeq2Seq-Attention model, generate to
The corresponding financial time series feature vector set of detection data collectionThe corresponding finance of preliminary account can also be obtained simultaneously
The testing result of time transaction sequence set, i.e., multiple level marketing that SoftSeq2Seq-Attention model finally provided be classified as or
Normal probability value, then category calculates the arithmetic average of these probability sets, obtain account final classification probability results to
AmountThus vector can obtain account multiple level marketing exception probability value and identify account result classification.
Step 6 can be obtained by step 5The detection recognition result of corresponding 50 accounts to be detected, can be described as:
[(c1,y1),(c2,y2),...,(c50,y50)], wherein ciFor card number i (account i), y to be detectediFor classification results label (yiIt takes
Value represents multiple level marketing account for 0 or 1,1, and 0 represents normal account).In addition, also can get account ciThe multiple level marketing exception probability of classification
Value.
So far, application case is finished.
The present invention also provides a kind of equipment for realizing the abnormal transaction identification method based on financial time series feature, packets
It includes:
Memory, for storing computer program and the abnormal transaction identification method based on financial time series feature；
Processor, for executing the computer program and based on the abnormal transaction identification side of financial time series feature
Method, the step of to realize abnormal transaction identification method based on financial time series feature.
The present invention also provides a kind of computers with the abnormal transaction identification method based on financial time series feature can
Storage medium is read, is stored with computer program on the computer readable storage medium, the computer program is held by processor
The step of row is to realize the abnormal transaction identification method based on financial time series feature.
The foregoing description of the disclosed embodiments enables those skilled in the art to implement or use the present invention.
Various modifications to these embodiments will be readily apparent to those skilled in the art, as defined herein
General Principle can be realized in other embodiments without departing from the spirit or scope of the present invention.Therefore, of the invention
It is not intended to be limited to the embodiments shown herein, and is to fit to and the principles and novel features disclosed herein phase one
The widest scope of cause.
Claims (7)
1. a kind of abnormal transaction identification method based on financial time series feature, which is characterized in that method includes:
Step 1 carries out data prediction to the raw financial transaction journal data set of input, extracts raw financial transaction journal
Cleaning data and crucial item data in data set, obtain key item data acquisition system D；
Step 2 constructs financial time series, constructs financial time series data set D based on key item data acquisition system Dfinput；
Step 3 is based on financial time series data set Dfinput, data mark is carried out according to determining multiple level marketing card number listing file
Note；The financial time series data set Train marked is inputted into SoftSeq2Seq-Attention neural network model, into
Row model training and financial time series feature extraction；
Step 4 carries out detection identification to account, identifies financial transaction flowing water information, and constructs financial transaction flowing water letter to be detected
Cease data setAccording to step 1 to financial transaction flowing water message data set to be detectedPretreatment operation is carried out, is obtained pre-
Processing operation result key item data acquisition system Dtest, financial time series data set is then constructed by step 2
It willMiddle data are input in trained SoftSeq2Seq-Attention neural network model, extract finance
Time series feature vector set
Based on Decoder middle layer financial time series feature vector, by linear linear layer and softmax layers of progress finance
The Classification and Identification of Transaction Account number obtains account ciCorresponding finance time transaction sequence set
Classification results Making by Probability Sets, then according to account detect identification side
Method carries out account ciThe calculating of final classification result, to obtain corresponding account as the abnormal probability value of multiple level marketing account.
2. the abnormal transaction identification method according to claim 1 based on financial time series feature, which is characterized in that
In step 1, transaction card number, the trade date, transaction amount, explanatory memorandum in transaction journal data are extracted, as spy
Sign extracts the basic data of work；Card number of trading is identified as the mark ID of financial account for multiple level marketing account；
By trade date by " YYYYMMDDhhmmss " unified formatting processing, transaction amount combination receipt and payment flag bit, receipt and payment mark
Will is that " into " then the amount of money is positive number, and receipt and payment mark is that " out ", then the amount of money was negative；After format normalized, trade gold is washed
Transaction journal data of the volume absolute value less than 50, and transaction card number are empty transaction journal data.
3. the abnormal transaction identification method according to claim 2 based on financial time series feature, which is characterized in that
Step 2 further include: after above-mentioned processing work, the transaction journal information based on critical data item constructs the financial time
The method of sequence is as follows:
1, the transaction card number set C={ c in transaction journal information is counted1,c2,...,cn, wherein n is card number sum；
2, with card number ciFor key assignments key, i.e. financial account identifies ID, ci∈ C, by ciCorresponding whole transaction journal information architecture
List li, liFor ciCorresponding value contents value, li[m]=[ci,timem,moneym,summarym], m only indicates list here
liIn some element subscript；Obtain key-value pair data set D={ d1,d2,...,dn, if di∈ D, then di=(ci,li)；
3, to all di∈ D, by its liList carries out ascending sort according to this content of trade date, i.e., arranges transaction journal
Table is in chronological sequence sequentially resequenced, by liIt is updated to ranking results list；
4, to all di∈ D, di=(ci,li), utilize its li[m]=[ci,timem,moneym,summarym] content, building
Initial input vectorWherein, moneymFor former transaction amount
Value；tivecmFor timemCorresponding vector indicates；summvecmFor transaction summarization item summarymCorresponding vector indicates；tivecm
Generation method be, based on exchange hour set whole in D call the HashingVectorizer method of sklearn to its into
Row vectorization indicates that vector dimension takes 5 dimensions；summvecmGeneration method be to be called based on transaction summarization set whole in D
The HashingVectorizer method of sklearn carries out vectorization expression to it, and vector dimension takes 10 dimensions；ThenBy 16
Initial characteristics vector is tieed up to constitute；
5, by step 4) method, data set D is obtainedinput, haveInput vectorHere m only indicates listIn some element subscript；
6, it is sampled using sliding window method and generates fixed length financial time series data, while expanding data amount can be played the role of；Cleaning
Fall DinputInLength is less than 15Key-value pair data；
Use length for 50, the sliding window that interval steps are 25 is right from the front to the backCarry out data segmentation；If the last one sliding window packet
The data length contained then carries out mending 0 operation less than 50 but greater than 15；Otherwise, give up this partial data；
Generation hasBelong to account ciTime series data set；
One time series data content isID is identified by card number ciAnd coding j is collectively constituted；
7, it can be obtained by step 6),
Here m only indicates listIn some element subscript；
IfThen DfinputFor constructed financial time series data set.
4. the abnormal transaction identification method according to claim 1 based on financial time series feature, which is characterized in that
Step 3 further include:
The list entries of SoftSeq2Seq-Attention modelFor corresponding fixed length financial time seriesWhereinEncoder and
The RNN hidden layer h of the part Decoder<i>And s<i>It is all made of standard GRU (Gated Recurrent Unit) gating cycle unit；
Based on using local attention mechanism, taking attention position is pt=5i+1, wherein i=1,2 ..., 9, window size
Dsize=4 or pt=2i+1, Dsize=2, then attention mechanism intermediate result s<t>Pass through [pt-Dsize,pt+ Dsize] window
Information and s in mouthful<t-1>It acquires；
s<t>Hidden layer intermediate result vectorAfter splicing, as the financial time series feature vector table extracted
Show, or is indicated using the summation of the vector or average operation result as financial time series feature vector；When will be final financial
Between the expression of sequence signature vector be fed for a linear layer, this linear layer exports length and is equal to account number classification number | C ' |；Linear layer knot
Fruit inputs the softmax layers of calculating for carrying out corresponding class probability again；
It is concentrated in the training data of the neural network model, the output y of the part Decoder is card number ciCorresponding classification results to
Amount, yp=(yp1,yp2)；yp1、yp2Respectively indicate financial time seriesCorresponding account ciIt is classified as multiple level marketing or normal
Probability；
The softmax layer calculation method of the part Decoder is as follows:
In SoftSeq2Seq-Attention neural network model, input layerIt is tieed up for 50, wherein per one-dimensional xi
It is made of 16 dimension initial characteristics vectors；Loss function is the cross entropy loss function that financial time series correspond to account number classification, it may be assumed that
Wherein, S is training dataset, and C ' is account number classification number, and s is financial time series data, pc′(s) indicate that prediction s is corresponding
Account be corresponding classification account probability；Indicate whether c ' class is the correct account number classification of s, value is 0 or 1 here；
When model training, using backpropagation and stochastic gradient descent algorithm undated parameter；
Training dataset is constructed, according to multiple level marketing card number listing file, to financial time series data set DfinputIn the financial time
Sequence initial input characteristic vector dataClassification annotation is carried out, training set Train=(x is constructed1,
y1),(x2,y2),...,(xn,yn)；, (x herei,yi) only indicate training data isomery finance featureWith it is right
Account type is answered to mark yi, yiValue represents multiple level marketing account for 0 or 1,1, and 0 represents normal account；If labeled data collection, then directly
Meet building training set Train=(x1,y1),(x2,y2),...,(xn,yn)；
It is required that two class data volume specific gravity are preferably between 1:1 to 1:2 in training dataset Train；Further division Train is
Training set train and verifying collection test, is divided, train specific gravity is 7/10 according to the ratio of 7:3.
5. the abnormal transaction identification method according to claim 1 based on financial time series feature, which is characterized in that
In step 4, account detection identification method includes arranging number to be detected according to data prediction and format method for normalizing
Financial time series data set to be detected is constructed according to collection and financial time series construction method, by financial time sequence to be detected
Column data collection inputs SoftSeq2Seq-Attention neural network model, carries out account number classification identification, and gained is last
Softmax layers of output result are account ciCorresponding wherein oneThe classification knot of transaction data
Fruit probability；
According to account ciIt is correspondingThe all classification probability of outcome of data in set
Calculate its average valueObtain ciCorresponding final classification probability of outcomeC is determined according to thisiFinal classification.
6. a kind of equipment for realizing the abnormal transaction identification method based on financial time series feature characterized by comprising
Memory, for storing computer program and the abnormal transaction identification method based on financial time series feature；
Processor, for executing the computer program and the abnormal transaction identification method based on financial time series feature, with
The step of realizing the abnormal transaction identification method as described in claim 1 to 5 any one based on financial time series feature.
7. a kind of computer readable storage medium with the abnormal transaction identification method based on financial time series feature, special
Sign is, is stored with computer program on the computer readable storage medium, the computer program be executed by processor with
The step of realizing the abnormal transaction identification method as described in claim 1 to 5 any one based on financial time series feature.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810909752.8A CN109165950B (en) | 2018-08-10 | 2018-08-10 | Financial time series characteristic-based abnormal transaction identification method, device and readable storage medium |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810909752.8A CN109165950B (en) | 2018-08-10 | 2018-08-10 | Financial time series characteristic-based abnormal transaction identification method, device and readable storage medium |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109165950A true CN109165950A (en) | 2019-01-08 |
| CN109165950B CN109165950B (en) | 2023-02-03 |
Family
ID=64895519
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810909752.8A Active CN109165950B (en) | 2018-08-10 | 2018-08-10 | Financial time series characteristic-based abnormal transaction identification method, device and readable storage medium |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109165950B (en) |
Cited By (18)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110059692A (en) * | 2019-04-16 | 2019-07-26 | 厦门商集网络科技有限责任公司 | A kind of method and terminal identifying the affiliated industry of enterprise |
| CN110084610A (en) * | 2019-04-23 | 2019-08-02 | 东华大学 | A kind of network trading fraud detection system based on twin neural network |
| CN110175850A (en) * | 2019-05-13 | 2019-08-27 | ***股份有限公司 | A kind of processing method and processing device of Transaction Information |
| CN110263827A (en) * | 2019-05-31 | 2019-09-20 | 中国工商银行股份有限公司 | Abnormal transaction detection method and device based on transaction rule identification |
| CN110276620A (en) * | 2019-06-28 | 2019-09-24 | 深圳前海微众银行股份有限公司 | It is a kind of to determine the method and device traded extremely |
| CN110400082A (en) * | 2019-07-29 | 2019-11-01 | 中国工商银行股份有限公司 | The recognition methods of abnormal transaction enterprise and device |
| CN110458581A (en) * | 2019-07-11 | 2019-11-15 | 阿里巴巴集团控股有限公司 | Merchant business turnover abnormal recognition methods and device |
| CN110751557A (en) * | 2019-10-10 | 2020-02-04 | 中国建设银行股份有限公司 | Abnormal fund transaction behavior analysis method and system based on sequence model |
| CN111062416A (en) * | 2019-11-14 | 2020-04-24 | 支付宝(杭州)信息技术有限公司 | User clustering and feature learning method, device and computer readable medium |
| CN111797177A (en) * | 2020-07-06 | 2020-10-20 | 哈尔滨工业大学（威海） | Financial time sequence classification method for abnormal financial account detection and application |
| CN111831825A (en) * | 2020-07-23 | 2020-10-27 | 咪咕文化科技有限公司 | Account detection method, account detection device, network equipment and storage medium |
| CN111861756A (en) * | 2020-08-05 | 2020-10-30 | 哈尔滨工业大学（威海） | Group partner detection method based on financial transaction network and implementation device thereof |
| CN111915437A (en) * | 2020-06-30 | 2020-11-10 | 深圳前海微众银行股份有限公司 | RNN-based anti-money laundering model training method, device, equipment and medium |
| CN113516228A (en) * | 2021-07-08 | 2021-10-19 | 哈尔滨理工大学 | Network anomaly detection method based on deep neural network |
| CN113609236A (en) * | 2021-08-05 | 2021-11-05 | 中国联合网络通信集团有限公司 | Data processing method, device and equipment |
| CN116383708A (en) * | 2023-05-25 | 2023-07-04 | 北京芯盾时代科技有限公司 | Transaction account identification method and device |
| CN117455497A (en) * | 2023-11-12 | 2024-01-26 | 北京营加品牌管理有限公司 | Transaction risk detection method and device |
| CN117522416A (en) * | 2023-12-28 | 2024-02-06 | 北京芯盾时代科技有限公司 | Transaction account identification method and device |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105357167A (en) * | 2014-08-19 | 2016-02-24 | 阿里巴巴集团控股有限公司 | Service processing method and device |
| CN106372938A (en) * | 2015-07-21 | 2017-02-01 | 华为技术有限公司 | Abnormal account identification method and system |
| CN106886518A (en) * | 2015-12-15 | 2017-06-23 | 国家计算机网络与信息安全管理中心 | A kind of method of microblog account classification |
| CN107066616A (en) * | 2017-05-09 | 2017-08-18 | 北京京东金融科技控股有限公司 | Method, device and electronic equipment for account processing |
| US20180165288A1 (en) * | 2016-12-14 | 2018-06-14 | Microsoft Technology Licensing, Llc | Dynamic Tensor Attention for Information Retrieval Scoring |
| CN108376131A (en) * | 2018-03-14 | 2018-08-07 | 中山大学 | Keyword abstraction method based on seq2seq deep neural network models |
-
2018
- 2018-08-10 CN CN201810909752.8A patent/CN109165950B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105357167A (en) * | 2014-08-19 | 2016-02-24 | 阿里巴巴集团控股有限公司 | Service processing method and device |
| CN106372938A (en) * | 2015-07-21 | 2017-02-01 | 华为技术有限公司 | Abnormal account identification method and system |
| CN106886518A (en) * | 2015-12-15 | 2017-06-23 | 国家计算机网络与信息安全管理中心 | A kind of method of microblog account classification |
| US20180165288A1 (en) * | 2016-12-14 | 2018-06-14 | Microsoft Technology Licensing, Llc | Dynamic Tensor Attention for Information Retrieval Scoring |
| CN107066616A (en) * | 2017-05-09 | 2017-08-18 | 北京京东金融科技控股有限公司 | Method, device and electronic equipment for account processing |
| CN108376131A (en) * | 2018-03-14 | 2018-08-07 | 中山大学 | Keyword abstraction method based on seq2seq deep neural network models |
Non-Patent Citations (2)
| Title |
|---|
| FANG LV等: "Detecting Pyramid Scheme Accounts with Time Series Financial Transactions", 《2018 IEEE DSC》 * |
| 秦学志等: "基于大数据样本的银行异常账户监测方法", 《***管理学报》 * |
Cited By (28)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110059692A (en) * | 2019-04-16 | 2019-07-26 | 厦门商集网络科技有限责任公司 | A kind of method and terminal identifying the affiliated industry of enterprise |
| CN110084610B (en) * | 2019-04-23 | 2023-06-23 | 东华大学 | Network transaction fraud detection system based on twin neural network |
| CN110084610A (en) * | 2019-04-23 | 2019-08-02 | 东华大学 | A kind of network trading fraud detection system based on twin neural network |
| CN110175850A (en) * | 2019-05-13 | 2019-08-27 | ***股份有限公司 | A kind of processing method and processing device of Transaction Information |
| CN110263827A (en) * | 2019-05-31 | 2019-09-20 | 中国工商银行股份有限公司 | Abnormal transaction detection method and device based on transaction rule identification |
| CN110263827B (en) * | 2019-05-31 | 2021-08-20 | 中国工商银行股份有限公司 | Abnormal transaction detection method and device based on transaction rule identification |
| CN110276620A (en) * | 2019-06-28 | 2019-09-24 | 深圳前海微众银行股份有限公司 | It is a kind of to determine the method and device traded extremely |
| CN110458581A (en) * | 2019-07-11 | 2019-11-15 | 阿里巴巴集团控股有限公司 | Merchant business turnover abnormal recognition methods and device |
| CN110458581B (en) * | 2019-07-11 | 2024-01-16 | 创新先进技术有限公司 | Method and device for identifying business turnover abnormality of commercial tenant |
| CN110400082A (en) * | 2019-07-29 | 2019-11-01 | 中国工商银行股份有限公司 | The recognition methods of abnormal transaction enterprise and device |
| CN110400082B (en) * | 2019-07-29 | 2022-12-13 | 中国工商银行股份有限公司 | Abnormal transaction enterprise identification method and device |
| CN110751557A (en) * | 2019-10-10 | 2020-02-04 | 中国建设银行股份有限公司 | Abnormal fund transaction behavior analysis method and system based on sequence model |
| CN111062416A (en) * | 2019-11-14 | 2020-04-24 | 支付宝(杭州)信息技术有限公司 | User clustering and feature learning method, device and computer readable medium |
| CN111915437B (en) * | 2020-06-30 | 2024-06-07 | 深圳前海微众银行股份有限公司 | Training method, device, equipment and medium of money backwashing model based on RNN |
| CN111915437A (en) * | 2020-06-30 | 2020-11-10 | 深圳前海微众银行股份有限公司 | RNN-based anti-money laundering model training method, device, equipment and medium |
| CN111797177A (en) * | 2020-07-06 | 2020-10-20 | 哈尔滨工业大学（威海） | Financial time sequence classification method for abnormal financial account detection and application |
| CN111831825A (en) * | 2020-07-23 | 2020-10-27 | 咪咕文化科技有限公司 | Account detection method, account detection device, network equipment and storage medium |
| CN111831825B (en) * | 2020-07-23 | 2024-03-15 | 咪咕文化科技有限公司 | Account detection method, device, network equipment and storage medium |
| CN111861756A (en) * | 2020-08-05 | 2020-10-30 | 哈尔滨工业大学（威海） | Group partner detection method based on financial transaction network and implementation device thereof |
| CN111861756B (en) * | 2020-08-05 | 2024-05-03 | 哈尔滨工业大学（威海） | Group partner detection method based on financial transaction network and realization device thereof |
| CN113516228A (en) * | 2021-07-08 | 2021-10-19 | 哈尔滨理工大学 | Network anomaly detection method based on deep neural network |
| CN113516228B (en) * | 2021-07-08 | 2023-04-18 | 哈尔滨理工大学 | Network anomaly detection method based on deep neural network |
| CN113609236B (en) * | 2021-08-05 | 2023-07-18 | 中国联合网络通信集团有限公司 | Data processing method, device and equipment |
| CN113609236A (en) * | 2021-08-05 | 2021-11-05 | 中国联合网络通信集团有限公司 | Data processing method, device and equipment |
| CN116383708B (en) * | 2023-05-25 | 2023-08-29 | 北京芯盾时代科技有限公司 | Transaction account identification method and device |
| CN116383708A (en) * | 2023-05-25 | 2023-07-04 | 北京芯盾时代科技有限公司 | Transaction account identification method and device |
| CN117455497A (en) * | 2023-11-12 | 2024-01-26 | 北京营加品牌管理有限公司 | Transaction risk detection method and device |
| CN117522416A (en) * | 2023-12-28 | 2024-02-06 | 北京芯盾时代科技有限公司 | Transaction account identification method and device |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109165950B (en) | 2023-02-03 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109165950A (en) | A kind of abnormal transaction identification method based on financial time series feature, equipment and readable storage medium storing program for executing | |
| CN108960833A (en) | A kind of abnormal transaction identification method based on isomery finance feature, equipment and storage medium | |
| Shrivastava et al. | Failure prediction of Indian Banks using SMOTE, Lasso regression, bagging and boosting | |
| CN111882446B (en) | Abnormal account detection method based on graph convolution network | |
| CN110110335A (en) | A kind of name entity recognition method based on Overlay model | |
| Ba | Improving detection of credit card fraudulent transactions using generative adversarial networks | |
| Kirkos et al. | Identifying qualified auditors' opinions: a data mining approach | |
| CN111292195A (en) | Risk account identification method and device | |
| Li et al. | Transaction fraud detection using gru-centered sandwich-structured model | |
| CN104850868A (en) | Customer segmentation method based on k-means and neural network cluster | |
| CN109726918A (en) | The personal credit for fighting network and semi-supervised learning based on production determines method | |
| Latha et al. | Stock movement prediction using KNN machine learning algorithm | |
| CN110084609A (en) | A kind of transaction swindling behavior depth detection method based on representative learning | |
| Islam et al. | Cost reduce: Credit card fraud identification using machine learning | |
| Chu et al. | Co-training based on semi-supervised ensemble classification approach for multi-label data stream | |
| CN111047428A (en) | Bank high-risk fraud client identification method based on small amount of fraud samples | |
| Negi et al. | Cryptocurrency Price Analysis using Deep Learning | |
| Ketepalli et al. | Anomaly Detection in Credit Card Transaction using Deep Learning Techniques | |
| CN112966728A (en) | Transaction monitoring method and device | |
| CN112784116A (en) | Method for identifying user industry identity in block chain | |
| Yu et al. | Neural network based transaction classification system for chinese transaction behavior analysis | |
| Geetha et al. | A hybrid deep learning and modified butterfly optimization based feature selection for transaction credit card fraud detection | |
| CN111639792A (en) | Method for intelligently adding bank ATM (automatic teller machine) money based on artificial intelligence | |
| CN116245645A (en) | Financial crime partner detection method based on graph neural network | |
| Shaik et al. | Customer loan eligibility prediction using machine 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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |