CN113919862A - Marketing arbitrage black product identification method based on dynamic attention-drawing network - Google Patents

Marketing arbitrage black product identification method based on dynamic attention-drawing network Download PDF

Info

Publication number
CN113919862A
CN113919862A CN202111040219.0A CN202111040219A CN113919862A CN 113919862 A CN113919862 A CN 113919862A CN 202111040219 A CN202111040219 A CN 202111040219A CN 113919862 A CN113919862 A CN 113919862A
Authority
CN
China
Prior art keywords
user
attention
network
nodes
merchant
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202111040219.0A
Other languages
Chinese (zh)
Inventor
傅剑文
陈心童
章建森
韩弘炀
周文彬
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Tianyi Electronic Commerce Co Ltd
Original Assignee
Tianyi Electronic Commerce Co Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Tianyi Electronic Commerce Co Ltd filed Critical Tianyi Electronic Commerce Co Ltd
Priority to CN202111040219.0A priority Critical patent/CN113919862A/en
Priority to JP2023513542A priority patent/JP2023543128A/en
Priority to PCT/CN2021/142967 priority patent/WO2023029324A1/en
Publication of CN113919862A publication Critical patent/CN113919862A/en
Pending legal-status Critical Current

Links

Images

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
    • G06Q30/00—Commerce
    • G06Q30/02—Marketing; Price estimation or determination; Fundraising
    • G06Q30/0207—Discounts or incentives, e.g. coupons or rebates
    • G06Q30/0225—Avoiding frauds
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Business, Economics & Management (AREA)
  • General Engineering & Computer Science (AREA)
  • Accounting & Taxation (AREA)
  • Strategic Management (AREA)
  • Finance (AREA)
  • Development Economics (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Probability & Statistics with Applications (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Game Theory and Decision Science (AREA)
  • Evolutionary Biology (AREA)
  • Economics (AREA)
  • Marketing (AREA)
  • General Business, Economics & Management (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention discloses a marketing arbitrage black product identification method based on a dynamic attention-deficit hyperactivity network. The invention has the following advantages: 1) defining a novel dynamic graph network structure consisting of user nodes, merchant nodes, multidimensional relations and a plurality of moments; 2) a novel multilateral attention unit is designed, so that unification of user nodes, merchant nodes and multidimensional relations can be realized more effectively, and information aggregation of node space layers can be realized better; 3) on the aspect of time dynamics of the graph network, a dynamic mode for transmitting attention information through a plurality of GRU units is innovatively designed, so that evolution information of the graph network on a time axis is effectively transmitted; 4) the multi-edge attention unit and the multi-path GRU units are effectively combined, a novel processing unit is provided, and a dynamic attention diagram network built by the unit can not only finish effective information transmission in a time dimension, but also realize high-efficiency aggregation of multilayer adjacent information in a space dimension.

Description

Marketing arbitrage black product identification method based on dynamic attention-drawing network
Technical Field
The invention relates to the fields of emerging information technologies and artificial intelligence, in particular to a marketing arbitrage black birth identification method based on a dynamic attention-drawing network.
Background
Marketing arbitrage is a novel internet fraud means, arbitrage black products arbitrage money vouchers, discounts, full reductions, standing-returning-standing reductions and other marketing activities which are released by an electronic commerce platform or a payment platform in a large quantity through various illegal modes, vital interests of normal consumers are greatly injured, and the marketing cost input by a platform party does not play the due value of the marketing cost. With the rapid development of the current internet industry, marketing arbitrage and black production tend to be ganged more, and the means become more concealed and complicated.
The traditional solution usually sets some risk strategies in a wind control system by means of expert experience to capture the arbitrage black-producing personnel hit by the risk strategies, but as the black-producing means are more ganged, complicated and concealed, the method is often limited by the fact that the expert cannot more comprehensively identify some characteristic hidden or novel arbitrage black-producing groups.
It is also proposed to judge the arbitrage risk of each node on the graph by calculating risk operators by constructing a graph network and defining some risk operators on the graph network, but such methods usually have the following disadvantages: 1) in essence, this method defines the expert experience policy in a more complex form on the graph network, and is also limited by the operator itself, not a machine learning algorithm. 2) The method usually only utilizes topological information on a graph network space, and does not utilize characteristic information hidden in network nodes. 3) Typically only information in low adjacencies is utilized and the potential information embedded in high adjacencies is not fully utilized.
With the development and progress of science and technology, the problem of better solving the scene through a graph neural network is provided. The graph neural network is an algorithm which is based on historical samples and can be learned by self, the limitation of expert experience can be avoided, meanwhile, the characteristic information of the nodes is introduced, and multi-layer adjacent information aggregation can be carried out, so that potential arbitrage risk nodes can be identified more comprehensively.
However, most of the current methods based on the graph neural network only use static graphs rather than dynamic graphs, that is, only a graph network structure at a certain time is used, and the evolution of the graph network on a time axis is not considered. The static graph method only considers the information on the network space structure of the graph, but ignores the information transmitted on the time axis
Disclosure of Invention
The technical problem to be solved by the invention is to overcome the defects of the prior art, and provide a marketing arbitrage black product identification method based on a dynamic attention-driven graph network, which can greatly improve the comprehensiveness and accuracy of a wind control system in identifying arbitrage black products, so that the scene problem can be better solved.
In order to solve the technical problems, the invention provides the following technical scheme:
the invention provides a marketing arbitrage black product identification method based on a dynamic attention-drawing network, which comprises the following steps of:
a dynamic graph network structure is defined, which is composed of two types of nodes, four types of relation edges and T times:
1) two types of nodes: the graph network comprises user nodes VcAnd merchant node VbTwo kinds of nodes
2) The four relationship edges are:
A) user and user's device relationship edge Ed：
According to the identity of the IP of different user access devices in a certain period of time in the past in the access log, generating a device relation edge between a user node and the user node;
B) payment relationship edge E of user and userp：
Generating a payment relation edge between a user node and a user node according to the account transfer payment behavior existing between the user and the user in a certain period of time in the past in a service system;
C) social relationship edge E between user and users：
Generating a social relationship edge between a user node and a user node according to the fact that a user and the user share marketing activity information and successfully invite another user to participate in the marketing activity in a certain past period of time in a service system;
D) user and merchant transaction relationship edge Et：
Generating a transaction relation edge between a user node and a merchant node according to the transaction payment behavior of a user in the merchant in a certain period of time in the past in a service system;
3) the T moments refer to:
selecting consecutive T time points on a time axis by taking T0 time as an initial time and taking omega as an interval; generating a relationship graph network consisting of the 2 types of nodes and the 4 types of relationship edges at each time point; thus, the evolution of the graph network along with the T moments of time is reflected;
generating T network graphs at T moments in graph database, wherein the T network graphs are respectively
Figure BDA0003248812170000031
Figure BDA0003248812170000032
Figure BDA0003248812170000033
Set V of users at each time pointcRespectively leading the characteristic vectors into a distributed database module from a database module, generating m-dimensional characteristic vectors representing the attributes of each user at each time point, wherein the characteristic fields of the characteristic vectors can be constructed by selecting common natural human attribute characteristics, and statistical characteristics extracted based on expert experience can be introduced, for example: the number of times of accessing a certain activity in a certain period of time in the past, and the like to obtain a better recognition effect; thus, for each time point, a set V of userscAn attribute feature matrix is generated
Figure BDA0003248812170000034
Storing the data in a distributed database module in a form of a table; n in subscript0，n1,.. respectively representing the number of user sets at each time point;
similarly, the commercial tenant set V of each time pointbRespectively importing the data from the database module to the distributed database module, and generating k-dimensional feature vectors representing the attributes of each merchant at each time point; thus, for each time point, a merchant set VbAn attribute feature matrix is generated
Figure BDA0003248812170000041
Storing the data in a distributed database module in a form of a table; subscript l0，l1,.. respectively representing the number of merchant sets at each point in time;
third, the history-based arbitrage user positive and negative samples are respectively a user set V of each time pointcLabeling and storing the label in a distributed database module in a form of a table; after the label is printed, the user set VcBecomes labeled (V)labeled) And no label (V)unlabeled) Two kinds of samples, VlabeledThe samples are divided into known arbitrage user samples V labeled1 and known non-arbitrage user sample VlabeledSetting all merchant nodes on the graph network as label-free samples;
designing a novel graph attention machine mechanism, which is characterized in that 1) independent mapping matrixes are defined for user nodes and merchant nodes respectively, and the two different dimensionality characteristic vectors are mapped into the same dimensionality through the mapping matrixes, so that the unification of the user nodes and the merchant nodes in information aggregation is realized; 2) based on four relation edges defined on the graph, a multilateral attention mechanism is defined, namely, a set of independent attention learning parameters are introduced into each relation edge, so that unification of multiple relation edges during information aggregation is realized;
the specific operation is as follows:
1) firstly, defining 4 sets of independent attention learning parameters according to four relation edges and initializing the learning parameters:
A) user-to-user device relationship Ed: defining a shared linear mapping matrix
Figure BDA0003248812170000042
And attention sharing vector ad∈R2f；
B) User-to-user payment relationship Ep: defining a shared linear mapping matrix
Figure BDA0003248812170000051
And attention sharing vector ap∈R2f；
C) Social relationship edge E between user and users: defining a shared linear mapping matrix
Figure BDA0003248812170000052
And attention sharing vector as∈R2f；
D) User and merchant transaction relationship edge Et: defining independent shared linear mapping matrices for user nodes and merchant nodes
Figure BDA0003248812170000053
And
Figure BDA0003248812170000054
the system comprises a database, a database and a database, wherein the database is used for storing a user characteristic vector space and a merchant characteristic vector space; defining a shared attention vector at∈R2f；
2) And calculating the attention coefficient of each side in the graph according to the following formula:
a) based on EdAttention coefficient of relationship edge:
Figure BDA0003248812170000055
Figure BDA0003248812170000056
b) based on EpAttention coefficient of the relationship:
Figure BDA0003248812170000057
Figure BDA0003248812170000061
c) based on EsAttention coefficient of the relationship:
Figure BDA0003248812170000062
Figure BDA0003248812170000063
d) based on EtAttention coefficient of the relationship:
Figure BDA0003248812170000064
Figure BDA0003248812170000065
in the above expression, the symbol | | | represents the concatenation of vectors, the subscripts i, j respectively represent two adjacent nodes based on a certain relation edge in the graph, hiOr hjAll represent that the feature vector corresponding to the node is taken from the previously generated user feature matrix, and in the calculation EtRelation attention coefficient hiAnd hjOne of the feature vectors is taken from a user feature matrix, and the other feature vector is taken from a merchant feature matrix;
3) and finally, performing information aggregation calculation through the following formula:
Figure BDA0003248812170000066
in the above formula, Nd(i)、Np(i)、Ns(i)、Nt(i) The symbols respectively represent first-order neighbors connected by the node i through an equipment relationship edge, a payment relationship edge, a social relationship edge and a transaction relationship edge;
the series of processes in the fourth step is called as follows: a multilateral attention unit (MultiEdgeGAT);
fifthly, the fourth step is just to complete information aggregation of the graph network at a spatial level at a certain moment, because the transmission of network information at a time dimension must be completed in consideration of the dynamic property of the graph network at the time dimension, the invention innovatively provides a mechanism (called as multi-GRU unit-multi GRU) for transmitting attention parameters at the time dimension by using a multi-GRU mode, and a brand new processing unit (called as dynamic attention unit) is formed by combining a graph attention information aggregation mechanism (multi-edge attention unit-multi-edge GAT) at the spatial level, and the operation is as follows:
1) since the graph attention mechanism will generate 9 parameters to be learned, i.e. 5 shared linear mapping matrices and 4 shared attention vectors, there are 9 independent GRU units responsible for passing parameter information in the time dimension, which are:
Figure BDA0003248812170000071
for conveying a linear mapping matrix, an
Figure BDA0003248812170000072
Used for conveying attention vectors; initializing parameters of the GRU units;
2) the GRU propagation formula here is as follows:
because the internal structure of the GRU units is the same, the GRU units are illustrated for convenience of explanation
Figure BDA0003248812170000073
The formula is illustrated for an example:
Figure BDA0003248812170000074
Figure BDA0003248812170000075
Figure BDA0003248812170000076
Figure BDA0003248812170000077
the above formula completes the attention parameter at the moment t-1
Figure BDA0003248812170000081
By passing
Figure BDA0003248812170000082
Communicating an attention parameter at time t of generation
Figure BDA0003248812170000083
Similarly, the other 8 attention parameters W for generating the t moment can be transmitted from the t-1 moment through other GRUs in the same structurept、Wst、
Figure BDA0003248812170000088
In the above equation |, represents the hadamard product,
Figure BDA0003248812170000084
is composed of
Figure BDA0003248812170000085
The parameters to be learned by the GRU are similar to other GRUs; the 9 GRU units are gathered togetherReferred to as a multiway GRU unit-multiway GRU;
3) combining the multilateral attention unit and the multipath GRU unit to generate a brand new processing unit called dynamic attention unit (dynamic GAT), wherein the combination formula is as follows:
to simplify the formula and make it easier to express, W isd、Wp、Ws、Wtc、WtbThese 5 parameters are combined into W, ad、ap、as、atThese 4 parameters are combined into a;
Wt，at＝MultiGRU(Ht，Wt-1，at-1)
H′t’＝MultiEdgeGAT(Ht，Wt，at)；
sixth, the above step five explains information transfer and information aggregation of one dynamic attention unit on the time axis, where it is natural to perform spatial multi-layer aggregation at each time point, then the expression of the dynamic attention unit can be expressed as:
Figure BDA0003248812170000086
Figure BDA0003248812170000087
the superscript l in the above formula represents the level of spatial aggregation, wherein, adding one layer of aggregation means that the next layer of adjacent information of the node in the graph is pressed to the node, so that the information contained in the multilayer adjacent relation is well utilized, and a dynamically learnable graph network which not only contains information transfer of time dimension, but also contains spatial dimension multilayer aggregation is constructed;
seventhly, outputting the last moment of the last layer of the dynamic attention diagram network
Figure BDA0003248812170000091
Characterization by Softmax, activating and mapping to predict the probability P of node arbitrage risk;
eighthly, calling all the steps as a complete learning iteration; when learning iteration is carried out, required information is synchronized into a server memory from a distributed database module, parameters required to be learned are initialized, and then multiple rounds of complete learning iteration are carried out; for labeled nodes V only after each complete learning iterationlabeledPerforming loss calculation through a cross entropy loss function, performing gradient updating through an Adam optimization algorithm, and finally learning parameters required by the dynamic attention-seeking network through N times of complete learning iteration; completing the learning of the dynamic attention diagram network;
and deploying the learned dynamic attention network into an air control decision module, performing online marketing arbitrage risk probability prediction, setting a probability decision threshold value P as sigma, judging as a marketing arbitrage risk group when the predicted risk probability is greater than sigma, and performing subsequent processing operations such as interception.
Compared with the prior art, the invention has the following beneficial effects:
the patent proposes a method for identifying arbitrage black products based on a dynamic attention-graph network, which is innovative and has the following advantages: 1) defining a novel dynamic graph network structure consisting of user nodes, merchant nodes, multidimensional relations and a plurality of moments; 2) a novel multilateral attention unit is designed, so that unification of user nodes, merchant nodes and multidimensional relations can be realized more effectively, and information aggregation of node space layers is realized better; 3) on the aspect of time dynamics of the graph network, a dynamic mode for transmitting attention information through a plurality of GRU units is innovatively designed, so that evolution information of the graph network on a time axis is effectively transmitted; 4) the multi-edge attention unit and the multi-path GRU units are effectively combined, a novel processing unit of the dynamic attention unit is provided, and the dynamic attention network built by the unit can not only finish effective information transmission of a time dimension, but also realize efficient aggregation of multilayer adjacent information of a space dimension.
The method provided by the patent can greatly improve the comprehensiveness and accuracy of the wind control system in identifying the arbitrage black products, so that the scene problem can be better solved.
Drawings
The accompanying drawings, which are included to provide a further understanding of the invention and are incorporated in and constitute a part of this specification, illustrate embodiments of the invention and together with the description serve to explain the principles of the invention and not to limit the invention. In the drawings:
FIG. 1 is a diagram of an exemplary network composition of the present invention;
FIG. 2 is a diagram of a multiple GRU unit architecture;
FIG. 3 is a diagram of a dynamic attention unit architecture;
FIG. 4 is a diagram of the overall structure of a dynamic attention map network;
fig. 5 is a flow chart of an implementation of the present invention.
Detailed Description
The preferred embodiments of the present invention will be described in conjunction with the accompanying drawings, and it will be understood that they are described herein for the purpose of illustration and explanation and not limitation.
Example 1
The invention provides a marketing arbitrage black birth identification method based on a dynamic attention-drawing network, which comprises the following steps of:
a dynamic graph network structure is defined, which is composed of two types of nodes, four types of relation edges and T times:
1) two types of nodes: the graph network comprises user nodes VcAnd merchant node VbTwo kinds of nodes
2) The four relationship edges are:
A) user and user's device relationship edge Ed：
According to the identity of the IP of different user access devices in a certain period of time in the past in the access log, generating a device relation edge between a user node and the user node;
B) payment relationship edge E of user and userp：
Generating a payment relation edge between a user node and a user node according to the account transfer payment behavior existing between the user and the user in a certain period of time in the past in a service system;
C) social relationship edge E between user and users：
Generating a social relationship edge between a user node and a user node according to the fact that a user and the user share marketing activity information and successfully invite another user to participate in the marketing activity in a certain past period of time in a service system;
D) user and merchant transaction relationship edge Et：
Generating a transaction relation edge between a user node and a merchant node according to the transaction payment behavior of a user in the merchant in a certain period of time in the past in a service system;
3) the T moments refer to:
selecting consecutive T time points on a time axis by taking T0 time as an initial time and taking omega as an interval; generating a relationship graph network consisting of the 2 types of nodes and the 4 types of relationship edges at each time point; thus, the evolution of the graph network along with the T moments of time is reflected; legends are seen in fig. 1;
generating T network graphs at T moments in graph database, wherein the T network graphs are respectively
Figure BDA0003248812170000113
Figure BDA0003248812170000111
Figure BDA0003248812170000112
Set V of users at each time pointcRespectively leading the data from the database module to the distributed database module, generating m-dimensional feature vectors representing the attributes of each user at each time point, wherein the feature domain of the feature vectors can be constructed by selecting common natural human attribute features, and statistical class features extracted based on expert experience can be introducedCharacterization, for example: the number of times of accessing a certain activity in a certain period of time in the past, and the like to obtain a better recognition effect; thus, for each time point, a set V of userscAn attribute feature matrix is generated
Figure BDA0003248812170000121
Storing the data in a distributed database module in a form of a table; n in subscript0，n1,.. respectively representing the number of user sets at each time point;
similarly, the commercial tenant set V of each time pointbRespectively importing the data from the database module to the distributed database module, and generating k-dimensional feature vectors representing the attributes of each merchant at each time point; thus, for each time point, a merchant set VbAn attribute feature matrix is generated
Figure BDA0003248812170000122
Storing the data in a distributed database module in a form of a table; subscript l0，l1,.. respectively representing the number of merchant sets at each point in time;
third, the history-based arbitrage user positive and negative samples are respectively a user set V of each time pointcLabeling and storing the label in a distributed database module in a form of a table; after the label is printed, the user set VcBecomes labeled (V)labeled) And no label (V)unlabeled) Two kinds of samples, VlabeledThe samples are divided into known arbitrage user samples V labeled1 and known non-arbitrage user sample VlabeledSetting all merchant nodes on the graph network as label-free samples;
designing a novel graph attention machine mechanism, which is characterized in that 1) independent mapping matrixes are defined for user nodes and merchant nodes respectively, and the two different dimensionality characteristic vectors are mapped into the same dimensionality through the mapping matrixes, so that the unification of the user nodes and the merchant nodes in information aggregation is realized; 2) based on four relation edges defined on the graph, a multilateral attention mechanism is defined, namely, a set of independent attention learning parameters are introduced into each relation edge, so that unification of multiple relation edges during information aggregation is realized;
the specific operation is as follows:
1) firstly, defining 4 sets of independent attention learning parameters according to four relation edges and initializing the learning parameters:
A) user-to-user device relationship Ed: defining a shared linear mapping matrix
Figure BDA0003248812170000131
And attention sharing vector ad∈R2f；
B) User-to-user payment relationship Ep: defining a shared linear mapping matrix
Figure BDA0003248812170000132
And attention sharing vector ap∈R2f；
C) Social relationship edge E between user and users: defining a shared linear mapping matrix
Figure BDA0003248812170000133
And attention sharing vector at∈R2f；
D) User and merchant transaction relationship edge Et: defining independent shared linear mapping matrices for user nodes and merchant nodes
Figure BDA0003248812170000134
And
Figure BDA0003248812170000135
the method is used for mapping a user characteristic vector space and a merchant characteristic vector space from m dimension and l dimension to a unified f-dimension vector space. Defining a shared attention vector at∈R2f；
2) And calculating the attention coefficient of each side in the graph according to the following formula:
a) based on EdAttention coefficient of relationship edge:
Figure BDA0003248812170000141
Figure BDA0003248812170000142
b) based on EpAttention coefficient of the relationship:
Figure BDA0003248812170000143
Figure BDA0003248812170000144
c) based on EsAttention coefficient of the relationship:
Figure BDA0003248812170000145
Figure BDA0003248812170000146
d) based on EtAttention coefficient of the relationship:
Figure BDA0003248812170000147
Figure BDA0003248812170000148
in the above expression, the symbol | | | represents the concatenation of vectors, the subscripts i, j respectively represent two adjacent nodes based on a certain relation edge in the graph, hiOr hjAll represent that the feature vector corresponding to the node is taken from the previously generated user feature matrix, and in the calculation EtRelation attention coefficient hiAnd hjOne feature vector is taken from the user featuresThe other is taken from the merchant feature matrix;
3) and finally, performing information aggregation calculation through the following formula:
Figure BDA0003248812170000151
in the above formula, Nd(i)、Np(i)、Ns(i)、Nt(i) The symbols respectively represent first-order neighbors connected by the node i through an equipment relationship edge, a payment relationship edge, a social relationship edge and a transaction relationship edge;
the series of processes in the fourth step is called as follows: a multilateral attention unit (MultiEdgeGAT);
fifthly, the fourth step is just to complete information aggregation of the graph network at a spatial level at a certain moment, because the transmission of network information at a time dimension must be completed in consideration of the dynamic property of the graph network at the time dimension, the invention innovatively provides a mechanism (called as multi-GRU unit-multi GRU) for transmitting attention parameters at the time dimension by using a multi-GRU mode, and a brand new processing unit (called as dynamic attention unit) is formed by combining a graph attention information aggregation mechanism (multi-edge attention unit-multi-edge GAT) at the spatial level, and the operation is as follows:
1) since the graph attention mechanism will generate 9 parameters to be learned, i.e. 5 shared linear mapping matrices and 4 shared attention vectors, there are 9 independent GRU units responsible for passing parameter information in the time dimension, which are:
Figure BDA0003248812170000152
for conveying a linear mapping matrix, an
Figure BDA0003248812170000153
For conveying attention vectors. Initializing parameters of the GRU units;
2) the GRU propagation formula here is as follows:
because the internal structure of the GRU units is the same, the GRU units are illustrated for convenience of explanation
Figure BDA0003248812170000154
The formula is illustrated for an example:
Figure BDA0003248812170000161
Figure BDA0003248812170000162
Figure BDA0003248812170000163
Figure BDA0003248812170000164
the above formula completes the attention parameter at the moment t-1
Figure BDA0003248812170000165
By passing
Figure BDA0003248812170000166
Communicating an attention parameter at time t of generation
Figure BDA0003248812170000167
Similarly, the other 8 attention parameters W for generating the t moment can be transmitted from the t-1 moment through other GRUs in the same structurept、Wst、
Figure BDA00032488121700001610
In the above equation |, represents the hadamard product,
Figure BDA0003248812170000168
is composed of
Figure BDA0003248812170000169
The parameters to be learned by the GRU are similar to other GRUs; the 9 GRU units are collected together and are called a multi-path GRU unit-MultiGRU; see the structure shown in FIG. 2;
3) combining the multilateral attention unit and the multipath GRU unit to generate a brand new processing unit called dynamic attention unit (dynamic GAT), wherein the combination formula is as follows:
to simplify the formula and make it easier to express, W isd、Wp、Ws、Wtc、WtbThese 5 parameters are combined into W, ad、ap、as、atThese 4 parameters are combined into a;
Wt，at＝MultiGRU(Ht，Wt-1，at-1)
H′t’＝MultiEdgeGAT(Ht，Wt，at)
see FIG. 3 for a dynamic attention cell structure;
sixth, the above step five explains information transfer and information aggregation of one dynamic attention unit on the time axis, where it is natural to perform spatial multi-layer aggregation at each time point, then the expression of the dynamic attention unit can be expressed as:
Figure BDA0003248812170000173
Figure BDA0003248812170000171
the superscript l in the above formula represents the level of spatial aggregation, wherein, adding one layer of aggregation means that the next layer of adjacent information of the node in the graph is pressed to the node, so that the information contained in the multilayer adjacent relation is well utilized, and a dynamically learnable graph network which not only contains information transfer of time dimension, but also contains spatial dimension multilayer aggregation is constructed; see FIG. 4 for a dynamic attention diagram network overall structure;
seventhly, outputting the last moment of the last layer of the dynamic attention diagram network
Figure BDA0003248812170000172
Representing the probability P of activating and mapping the representation to the predicted node arbitrage risk through Softmax;
and eighthly, calling all the steps as a complete learning iteration. When the learning iteration is performed, the required information is synchronized into the server memory from the distributed database module, the parameters required to be learned are initialized, and then the complete learning iteration of multiple rounds is performed. For labeled nodes V only after each complete learning iterationlabeledAnd finally, learning parameters required by the dynamic attention-seeking network through N times of complete learning iteration by performing loss calculation through a cross entropy loss function and performing gradient updating through an Adam optimization algorithm. Completing the learning of the dynamic attention diagram network;
and deploying the learned dynamic attention network into an air control decision module, performing online marketing arbitrage risk probability prediction, setting a probability decision threshold value P as sigma, judging as a marketing arbitrage risk group when the predicted risk probability is greater than sigma, and performing subsequent processing operations such as interception.
Specifically, examples are as follows:
1. the user node attribute feature vector takes the experience value m of 256
2. The merchant node attribute feature vector has an experience value of l being 128
T time, T value T is 10
3. With a label (V)labeled) And no label (V)unlabeled) Two types of sample ratios:
Figure BDA0003248812170000181
4. spatial aggregation level (L), span: { L |2 ≦ L ≦ 3, L ∈ Z }
5. And (3) taking a decision threshold value sigma during prediction and identification: σ is 0.5.
The invention has the following technical points:
1. a two-class node is defined: the user node and the merchant node are defined by four relation edges: user and user equipment relationship edge EdUser and user payment relationship edge EpUser and user social relationship edge EsUser and merchant transaction relationship edge EtAnd a new dynamic graph network structure consisting of a plurality of moments;
2. a novel multilateral attention unit is designed, so that unification of user nodes, merchant nodes and multidimensional relations can be realized more effectively, and information aggregation of node space layers is realized better;
3. the innovative design of a dynamic mode for transmitting attention information through multiple GRU units enables evolution information of the graph network on a time axis to be effectively transmitted.
4. The multi-edge attention unit and the multi-path GRU units are effectively combined, a novel processing unit of the dynamic attention unit is provided, and the dynamic attention network built by the unit can not only finish effective information transmission of a time dimension, but also realize efficient aggregation of multilayer adjacent information of a space dimension.
Finally, it should be noted that: although the present invention has been described in detail with reference to the foregoing embodiments, it will be apparent to those skilled in the art that changes may be made in the embodiments and/or equivalents thereof without departing from the spirit and scope of the invention. Any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims (1)

1. A marketing arbitrage black yield identification method based on a dynamic attention-drawing network is characterized by comprising the following steps:
a dynamic graph network structure is defined, which is composed of two types of nodes, four types of relation edges and T times:
1) two types of nodes: picture netThe network contains user nodes VcAnd merchant node VbTwo kinds of nodes
2) The four relationship edges are:
A) user and user's device relationship edge Ed：
According to the identity of the IP of different user access devices in a certain period of time in the past in the access log, generating a device relation edge between a user node and the user node;
B) payment relationship edge E of user and userp：
Generating a payment relation edge between a user node and a user node according to the account transfer payment behavior existing between the user and the user in a certain period of time in the past in a service system;
C) social relationship edge E between user and users：
Generating a social relationship edge between a user node and a user node according to the fact that a user and the user share marketing activity information and successfully invite another user to participate in the marketing activity in a certain past period of time in a service system;
D) user and merchant transaction relationship edge Et：
Generating a transaction relation edge between a user node and a merchant node according to the transaction payment behavior of a user in the merchant in a certain period of time in the past in a service system;
3) the T moments refer to:
selecting consecutive T time points on a time axis by taking T0 time as an initial time and taking omega as an interval; generating a relationship graph network consisting of the 2 types of nodes and the 4 types of relationship edges at each time point; thus, the evolution of the graph network along with the T moments of time is reflected;
generating T network graphs at T moments in graph database, wherein the T network graphs are respectively
Figure FDA0003248812160000021
Figure FDA0003248812160000022
Figure FDA0003248812160000023
Set V of users at each time pointcRespectively leading the characteristic vectors into a distributed database module from a database module, generating m-dimensional characteristic vectors representing the attributes of each user at each time point, wherein the characteristic fields of the characteristic vectors can be constructed by selecting common natural human attribute characteristics, and statistical characteristics extracted based on expert experience can be introduced, for example: the number of times of accessing a certain activity in a certain period of time in the past, and the like to obtain a better recognition effect; thus, for each time point, a set V of userscAn attribute feature matrix is generated
Figure FDA0003248812160000024
Storing the data in a distributed database module in a form of a table; n in subscript0，n1,.. respectively representing the number of user sets at each time point;
similarly, the commercial tenant set V of each time pointbRespectively importing the data from the database module to the distributed database module, and generating k-dimensional feature vectors representing the attributes of each merchant at each time point; thus, for each time point, a merchant set VbAn attribute feature matrix is generated
Figure FDA0003248812160000025
Storing the data in a distributed database module in a form of a table; subscript l0L 1.. denotes the number of merchant sets at each time point, respectively;
third, the history-based arbitrage user positive and negative samples are respectively a user set V of each time pointcLabeling and storing the label in a distributed database module in a form of a table; after the label is printed, the user set VcBecomes labeled (V)labeled) And no label (V)unlabeled) Two kinds of samples, VlabeledThe samples are divided into known arbitrage user samplesVlabeled1 and known non-arbitrage user sample VlabeledSetting all merchant nodes on the graph network as label-free samples;
designing a novel graph attention machine mechanism, which is characterized in that 1) independent mapping matrixes are defined for user nodes and merchant nodes respectively, and the two different dimensionality characteristic vectors are mapped into the same dimensionality through the mapping matrixes, so that the unification of the user nodes and the merchant nodes in information aggregation is realized; 2) based on four relation edges defined on the graph, a multilateral attention mechanism is defined, namely, a set of independent attention learning parameters are introduced into each relation edge, so that unification of multiple relation edges during information aggregation is realized;
the specific operation is as follows:
1) firstly, defining 4 sets of independent attention learning parameters according to four relation edges and initializing the learning parameters:
A) user-to-user device relationship Ed: defining a shared linear mapping matrix
Figure FDA0003248812160000031
And attention sharing vector ad∈R2f；
B) User-to-user payment relationship Ep: defining a shared linear mapping matrix
Figure FDA0003248812160000032
And attention sharing vector ap∈R2f；
C) Social relationship edge E between user and users: defining a shared linear mapping matrix
Figure FDA0003248812160000033
And attention sharing vector as∈R2f；
D) User and merchant transaction relationship edge Et: defining independent shared linear mapping matrices for user nodes and merchant nodes
Figure FDA0003248812160000041
And
Figure FDA0003248812160000042
the system comprises a database, a user characteristic vector space and a merchant characteristic vector space, wherein the database is used for mapping the user characteristic vector space and the merchant characteristic vector space from m dimension and l dimension to a unified f-dimension vector space; defining a shared attention vector at∈R2f；
2) And calculating the attention coefficient of each side in the graph according to the following formula:
a) based on EdAttention coefficient of relationship edge:
Figure FDA0003248812160000043
Figure FDA0003248812160000044
b) based on EpAttention coefficient of the relationship:
Figure FDA0003248812160000045
Figure FDA0003248812160000046
c) based on EsAttention coefficient of the relationship:
Figure FDA0003248812160000047
Figure FDA0003248812160000048
d) based on EtAttention coefficient of relationship：
Figure FDA0003248812160000049
Figure FDA0003248812160000051
In the above expression, the symbol | | | represents the concatenation of vectors, the subscripts i, j respectively represent two adjacent nodes based on a certain relation edge in the graph, hiOr hjAll represent that the feature vector corresponding to the node is taken from the previously generated user feature matrix, and in the calculation EtRelation attention coefficient hiAnd hjOne of the feature vectors is taken from a user feature matrix, and the other feature vector is taken from a merchant feature matrix;
3) and finally, performing information aggregation calculation through the following formula:
Figure FDA0003248812160000052
in the above formula, Nd(i)、Np(i)、Ns(i)、Nt(i) The symbols respectively represent first-order neighbors connected by the node i through an equipment relationship edge, a payment relationship edge, a social relationship edge and a transaction relationship edge;
the series of processes in the fourth step is called as follows: a multilateral attention unit (MultiEdgeGAT);
fifthly, the fourth step is just to complete information aggregation of the graph network at a spatial level at a certain moment, because the transmission of network information at a time dimension must be completed in consideration of the dynamic property of the graph network at the time dimension, the invention innovatively provides a mechanism (called as multi-GRU unit-multi GRU) for transmitting attention parameters at the time dimension by using a multi-GRU mode, and a brand new processing unit (called as dynamic attention unit) is formed by combining a graph attention information aggregation mechanism (multi-edge attention unit-multi-edge GAT) at the spatial level, and the operation is as follows:
1) since the graph attention mechanism will generate 9 parameters to be learned, i.e. 5 shared linear mapping matrices and 4 shared attention vectors, there are 9 independent GRU units responsible for passing parameter information in the time dimension, which are:
Figure FDA0003248812160000061
for conveying a linear mapping matrix, an
Figure FDA0003248812160000062
Used for conveying attention vectors; initializing parameters of the GRU units;
2) the GRU propagation formula here is as follows:
because the internal structure of the GRU units is the same, the GRU units are illustrated for convenience of explanation
Figure FDA0003248812160000063
The formula is illustrated for an example:
Figure FDA0003248812160000064
Figure FDA0003248812160000065
Figure FDA0003248812160000066
Figure FDA0003248812160000067
the above formula completes the attention parameter at the moment t-1
Figure FDA0003248812160000068
By passing
Figure FDA0003248812160000069
Communicating an attention parameter at time t of generation
Figure FDA00032488121600000612
Similarly, the other 8 attention parameters W for generating the t moment can be transmitted from the t-1 moment through other GRUs in the same structurept、Wst、
Figure FDA00032488121600000613
In the above equation |, represents the hadamard product,
Figure FDA00032488121600000610
W～d、U～d、B～dis composed of
Figure FDA00032488121600000614
The parameters to be learned by the GRU are similar to other GRUs; the 9 GRU units are collected together and are called a multi-path GRU unit-MultiGRU;
3) combining the multilateral attention unit and the multipath GRU unit to generate a brand new processing unit called dynamic attention unit (dynamic GAT), wherein the combination formula is as follows:
to simplify the formula and make it easier to express, W isd、Wp、Ws、Wtc、WtbThese 5 parameters are combined into W, ad、ap、as、atThese 4 parameters are combined into a;
Wt，at＝MultiGRU(Ht，Wt-1，at-1)
H′t’＝MultiEdgeGAT(Ht，Wt，at)；
sixth, the above step five explains information transfer and information aggregation of one dynamic attention unit on the time axis, where it is natural to perform spatial multi-layer aggregation at each time point, then the expression of the dynamic attention unit can be expressed as:
Figure FDA0003248812160000071
Figure FDA0003248812160000072
the superscript l in the above formula represents the level of spatial aggregation, wherein, adding one layer of aggregation means that the next layer of adjacent information of the node in the graph is pressed to the node, so that the information contained in the multilayer adjacent relation is well utilized, and a dynamically learnable graph network which not only contains information transfer of time dimension, but also contains spatial dimension multilayer aggregation is constructed;
seventhly, outputting the last moment of the last layer of the dynamic attention diagram network
Figure FDA0003248812160000073
Representing the probability P of activating and mapping the representation to the predicted node arbitrage risk through Softmax;
eighthly, calling all the steps as a complete learning iteration; when learning iteration is carried out, required information is synchronized into a server memory from a distributed database module, parameters required to be learned are initialized, and then multiple rounds of complete learning iteration are carried out; for labeled nodes V only after each complete learning iterationlabeledPerforming loss calculation through a cross entropy loss function, performing gradient updating through an Adam optimization algorithm, and finally learning parameters required by the dynamic attention-seeking network through N times of complete learning iteration; completing the learning of the dynamic attention diagram network;
and deploying the learned dynamic attention network into an air control decision module, performing online marketing arbitrage risk probability prediction, setting a probability decision threshold value P as sigma, judging as a marketing arbitrage risk group when the predicted risk probability is greater than sigma, and performing subsequent processing operations such as interception.
CN202111040219.0A 2021-09-06 2021-09-06 Marketing arbitrage black product identification method based on dynamic attention-drawing network Pending CN113919862A (en)

Priority Applications (3)

Application Number Priority Date Filing Date Title
CN202111040219.0A CN113919862A (en) 2021-09-06 2021-09-06 Marketing arbitrage black product identification method based on dynamic attention-drawing network
JP2023513542A JP2023543128A (en) 2021-09-06 2021-12-30 Marketing arbitrage network dark industry identification method based on dynamic attention graph network
PCT/CN2021/142967 WO2023029324A1 (en) 2021-09-06 2021-12-30 Marketing arbitrage underground industry identification method based on dynamic attention graph network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111040219.0A CN113919862A (en) 2021-09-06 2021-09-06 Marketing arbitrage black product identification method based on dynamic attention-drawing network

Publications (1)

Publication Number Publication Date
CN113919862A true CN113919862A (en) 2022-01-11

Family

ID=79234169

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111040219.0A Pending CN113919862A (en) 2021-09-06 2021-09-06 Marketing arbitrage black product identification method based on dynamic attention-drawing network

Country Status (3)

Country Link
JP (1) JP2023543128A (en)
CN (1) CN113919862A (en)
WO (1) WO2023029324A1 (en)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116821451B (en) * 2023-06-19 2024-03-22 苏州浩海医疗科技有限公司 User management system and method based on full life cycle

Family Cites Families (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109934706B (en) * 2017-12-15 2021-10-29 创新先进技术有限公司 Transaction risk control method, device and equipment based on graph structure model
US10824949B2 (en) * 2018-09-27 2020-11-03 Babylon Partners Limited Method and system for extracting information from graphs
CN111260462B (en) * 2020-01-16 2022-05-27 东华大学 Transaction fraud detection method based on heterogeneous relation network attention mechanism
CN112435034A (en) * 2020-12-01 2021-03-02 天翼电子商务有限公司 Marketing arbitrage black product identification method based on multi-network graph aggregation
CN112381547A (en) * 2020-12-02 2021-02-19 天翼电子商务有限公司 Marketing cheating monitoring method based on heterogeneous graph convolution neural network

Also Published As

Publication number Publication date
JP2023543128A (en) 2023-10-13
WO2023029324A1 (en) 2023-03-09

Similar Documents

Publication Publication Date Title
CN112766500B (en) Method and device for training graph neural network
Dong et al. Federated class-incremental learning
US11836615B2 (en) Bayesian nonparametric learning of neural networks
Ru et al. Multi-temporal scene classification and scene change detection with correlation based fusion
Save et al. A novel idea for credit card fraud detection using decision tree
Srivastava et al. New perspectives on gray sheep behavior in E-commerce recommendations
Liu et al. Fedbcd: A communication-efficient collaborative learning framework for distributed features
Tang et al. Toward predicting collective behavior via social dimension extraction
CN108874914A (en) A kind of information recommendation method based on the long-pending and neural collaborative filtering of picture scroll
Qi et al. Model aggregation techniques in federated learning: A comprehensive survey
CN111008335B (en) Information processing method, device, equipment and storage medium
CN110263236A (en) Social network user multi-tag classification method based on dynamic multi-view learning model
CN113283902A (en) Multi-channel block chain fishing node detection method based on graph neural network
CN112541575A (en) Method and device for training graph neural network
Xu et al. Hp-gmn: Graph memory networks for heterophilous graphs
Chiaro et al. FL-Enhance: A federated learning framework for balancing non-IID data with augmented and shared compressed samples
CN113919862A (en) Marketing arbitrage black product identification method based on dynamic attention-drawing network
CN112435034A (en) Marketing arbitrage black product identification method based on multi-network graph aggregation
CN114139066A (en) Collaborative filtering recommendation system based on graph neural network
Sahu et al. Matrix factorization in cross-domain recommendations framework by shared users latent factors
König et al. A dynamic model of network formation with strategic interactions
CN116975686A (en) Method for training student model, behavior prediction method and device
Hajirahimi et al. A novel class of reliability-based parallel hybridization (RPH) models for time series forecasting
CN113297575B (en) Multi-channel graph vertical federal model defense method based on self-encoder
CN115470520A (en) Differential privacy and denoising data protection method under vertical federal framework

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication