CN115660688A - Financial transaction abnormity detection method and cross-region sustainable training method thereof - Google Patents
Financial transaction abnormity detection method and cross-region sustainable training method thereof Download PDFInfo
- Publication number
- CN115660688A CN115660688A CN202211301695.8A CN202211301695A CN115660688A CN 115660688 A CN115660688 A CN 115660688A CN 202211301695 A CN202211301695 A CN 202211301695A CN 115660688 A CN115660688 A CN 115660688A
- Authority
- CN
- China
- Prior art keywords
- transaction
- node
- financial
- meta
- nodes
- 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
- 238000001514 detection method Methods 0.000 title claims abstract description 101
- 238000000034 method Methods 0.000 title claims abstract description 42
- 238000012549 training Methods 0.000 title claims abstract description 35
- 230000007246 mechanism Effects 0.000 claims abstract description 25
- 238000009499 grossing Methods 0.000 claims abstract description 16
- 239000011159 matrix material Substances 0.000 claims description 60
- 239000013598 vector Substances 0.000 claims description 50
- 238000005070 sampling Methods 0.000 claims description 32
- 230000002159 abnormal effect Effects 0.000 claims description 24
- 230000006870 function Effects 0.000 claims description 24
- 239000000126 substance Substances 0.000 claims description 21
- 230000004931 aggregating effect Effects 0.000 claims description 15
- 230000008569 process Effects 0.000 claims description 11
- 238000010276 construction Methods 0.000 claims description 10
- 230000017105 transposition Effects 0.000 claims description 7
- 230000004913 activation Effects 0.000 claims description 6
- 238000012545 processing Methods 0.000 claims description 5
- 230000005856 abnormality Effects 0.000 claims description 4
- 238000013507 mapping Methods 0.000 claims description 3
- 238000012986 modification Methods 0.000 claims description 2
- 230000004048 modification Effects 0.000 claims description 2
- 238000011156 evaluation Methods 0.000 claims 1
- 238000003062 neural network model Methods 0.000 abstract description 4
- 206010000117 Abnormal behaviour Diseases 0.000 abstract 1
- 238000011161 development Methods 0.000 description 5
- 230000002547 anomalous effect Effects 0.000 description 4
- 238000005516 engineering process Methods 0.000 description 4
- 230000006399 behavior Effects 0.000 description 3
- 230000007547 defect Effects 0.000 description 3
- 238000013528 artificial neural network Methods 0.000 description 2
- 230000008859 change Effects 0.000 description 2
- 238000012937 correction Methods 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 230000003993 interaction Effects 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 238000012935 Averaging Methods 0.000 description 1
- 230000002411 adverse Effects 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000008901 benefit Effects 0.000 description 1
- 238000004891 communication Methods 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 239000002360 explosive Substances 0.000 description 1
- PCHJSUWPFVWCPO-UHFFFAOYSA-N gold Chemical compound [Au] PCHJSUWPFVWCPO-UHFFFAOYSA-N 0.000 description 1
- 239000010931 gold Substances 0.000 description 1
- 229910052737 gold Inorganic materials 0.000 description 1
- 238000013508 migration Methods 0.000 description 1
- 230000005012 migration Effects 0.000 description 1
- 238000007781 pre-processing Methods 0.000 description 1
- 230000002265 prevention Effects 0.000 description 1
- 230000000452 restraining effect Effects 0.000 description 1
- 238000012954 risk control Methods 0.000 description 1
- 238000012216 screening Methods 0.000 description 1
- 230000035945 sensitivity Effects 0.000 description 1
- 230000003068 static effect Effects 0.000 description 1
- 230000009466 transformation Effects 0.000 description 1
Images
Landscapes
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
- Financial Or Insurance-Related Operations Such As Payment And Settlement (AREA)
Abstract
The invention relates to the field of financial risk management, and discloses a financial transaction abnormity detection method and a cross-regional sustainable training method thereof, wherein the financial transaction abnormity detection method is characterized in that a heterostructure information graph formed by a plurality of nodes and a plurality of paths is constructed, the problems that the homograph in the prior art is difficult to retain multi-type semantic information and cannot capture dynamic space-time characteristics are solved, high-order semantics such as time information are fully mined, the amount of information can be greatly enriched and obtained, a deep graph neural network model is adopted, the nodes, the paths and the network structure are fused based on an attention mechanism to obtain graph embedded representation, abnormal behaviors are detected based on the graph embedded representation, and the efficiency and the precision of abnormity detection are improved; meanwhile, a cross-region sustainable training method is provided, cross-region sustainable learning of a financial transaction anomaly detection model is achieved through a knowledge playback strategy and a parameter smoothing strategy, cross-region deployment is facilitated, and the method is suitable for financial risk management tasks such as financial transaction fraud detection.
Description
Technical Field
The invention relates to the field of financial risk management, in particular to a financial transaction abnormity detection method and a cross-regional sustainable training method thereof.
Background
With the rapid expansion of electronic commerce and the explosive development of communication technologies, economic activities have developed the demand for cross-regional, cross-time, and low cost. Digital finance as a novel financial state has powerful physics penetrability and low-cost advantage, has avoided the restriction of physics site and then raises the efficiency, brings the convenience. However, with the development of digital technology, digital financial fraud is derived, which affects the progress of legal transaction process and causes economic loss to users.
For digital finance, financial fraud and other abnormal transactions not only can seriously hurt the trust of users on financial science and technology and cause economic and reputation losses to various financial institutions and enterprises, but also bring adverse effects to the innovative development of digital finance and the digital transformation and upgrade of the traditional financial industry, and these threatens to the development of the digital financial industry fatally. Therefore, financial anti-fraud has become a key ring of financial risk prevention capability, and the financial anomaly detection technology has important application value and research significance for the security of the financial industry and even the security of national information.
The main purpose of financial anomaly detection is to detect whether financial transactions are anomalous, including anomalous transaction detection and anomalous behavior detection. Traditional financial anti-fraud solutions rely on one hand on rule-based models that detect financial fraud by building manual features derived from historical transaction data to discover potential anomalous behavior, and by expert experience to set important features. However, rule-based approaches rely heavily on a priori knowledge of humans, resulting in detection bias and are prone to collapse when dealing with more complex rule patterns; on the other hand, it relies on statically structured data, however, existing anti-fraud models are difficult to react quickly and accurately due to the open dynamic nature of the data and the heterogeneous nature of fraudulent data. In addition, most existing model designs also cause the system to lack interpretability and efficiency, and are difficult to efficiently process continuous learning in a real open environment.
In recent years, with the development of graph representation learning, more and more financial researches use a deep graph-based neural network method, high-order implicit information in data is explored by modeling entities such as customers and merchants as nodes and modeling interactions among the entities as edges, and graph representation learning is promoted to be used for revealing implicit modes behind large-scale financial transactions.
Existing depth map neural network based solutions present two significant challenges:
1. the transaction data in the real business scene comprises various types of additional entities besides the customer and merchant entities, such as transaction time and transaction space. However, most of the existing solutions only consider homogeneous networks, and decompose heterogeneous interaction into a plurality of homogeneous connections, which results in loss of high-order semantic information and incapability of learning spatio-temporal dynamic characteristics by a model.
2. The existing solution is only applicable to a narrow range due to the limitation of data acquisition in practical application. When a financial business is extended to a new area, such as a new city or even a new country, it is common to either use a previously trained, i.e. static, model or to develop a completely new model. Because consumption characteristics and behavior characteristics of entities in different areas are different to a greater or lesser extent, a financial model trained through the existing data learning can not be directly used due to mismatching of target characteristics caused by geographical differences; and developing a new model requires a high monetary and time cost.
Disclosure of Invention
The technical problem to be solved by the invention is as follows: the financial transaction abnormity detection method and the cross-regional sustainable training method thereof can improve the efficiency and the precision of financial transaction abnormity detection and can be conveniently deployed in a cross-regional manner.
The technical scheme adopted by the invention for solving the technical problems is as follows:
the financial transaction abnormity detection method comprises the following steps:
a1, constructing a financial heterogeneous information graph according to-be-detected financial transaction data of a target area; the entity nodes of the financial heterogeneous information graph comprise user nodes, merchant nodes, time nodes and transaction nodes, the characteristic vectors of the nodes are obtained according to corresponding knowledge information codes, and the types of meta-paths comprise user-transaction paths, merchant-transaction paths and time-transaction paths;
a2, obtaining the embedded representation of each node through a coding network according to the characteristic vector of each node; according to the embedded representation of the transaction node and the neighbor nodes thereof, classifying and aggregating based on an attention mechanism according to the meta-path types of the transaction node and the neighbor nodes thereof to obtain the embedded representation of each meta-path type of the transaction node; then, according to the meta-path type, aggregating the embedded representation of the meta-path type corresponding to each transaction node to obtain the embedded representation of each meta-path type of the financial heterogeneous information graph;
a3, according to the embedded representation of each meta path type of the financial heterogeneous information graph, fusing the embedded representation of each meta path type based on a self-attention mechanism to obtain graph embedded representation of the financial heterogeneous information graph;
and A4, based on the map embedded representation of the financial heterogeneous information map, adopting a classification label of the all-connection layer output abnormity detection as an abnormity detection result.
Specifically, in step A1, the node feature vector is obtained according to the corresponding knowledge information code, and includes:
the corresponding knowledge information of the user node is the unique identity ID of the transaction user;
the corresponding knowledge information of the merchant node is the unique identity ID of the transaction merchant;
the time node corresponds to the knowledge information and is timestamp information of the transaction;
the transaction node corresponding knowledge information comprises a transaction serial number, a transaction place, a transaction amount and a transaction type.
Further, feature vector coding of the time node comprises: firstly, 24 hours per day are divided into time windows according to time intervals, then coding is carried out according to digital information or onehot mode of the time windows according to the time periods corresponding to the time windows to which the transaction timestamps belong, and feature vector codes of time nodes are obtained.
Further, the feature vector coding of the trading node comprises the following steps: firstly, coding is carried out according to the digital information of a transaction serial number, coding is carried out according to the longitude and latitude of a transaction place, coding is carried out according to the digital information of a transaction amount, the transaction type is coded according to an onehot mode, and then the serial number, the transaction place, the transaction amount and the transaction type obtained by coding are spliced to form a feature vector code of a transaction node.
Specifically, in the step A2, the embedded representation of each node is obtained through a coding network according to the feature vector of each node; according to the embedded representation of the transaction node and the neighbor nodes thereof, classifying and aggregating based on an attention mechanism according to the meta-path types of the transaction node and the neighbor nodes thereof to obtain the embedded representation of each meta-path type of the transaction node; then, according to the meta-path type, aggregating the embedded representation of the meta-path type corresponding to each transaction node to obtain the embedded representation of each meta-path type of the financial heterogeneous information graph, including:
a21, mapping the feature vectors of all nodes in the financial heterogeneous information graph to a feature space through a coding network to obtain the embedded representation of each node in the feature space
Wherein the content of the first and second substances,
for the embedded representation of the ith node in the feature space, W x In order to encode the learnable parameter matrix of the network,
the feature vector of the ith node;
a22, adopting an attention mechanism to obtain attention scores between the transaction nodes and the neighbor nodes thereof
Wherein, att η It is shown that the attention mechanism is,
for the attention score between trading node i and neighbor node j,
the type of the meta-path between the transaction node i and the neighbor node j is determined;
a23, using a SoftMax function to score the attention of each meta-path type according to the meta-path type
Performing standardization to obtain attention weight
Wherein the content of the first and second substances,
the meta-path type in the neighbor node of the transaction node i is the mth type meta-path
A set of neighbor nodes of (1);
a24, according to the meta-path type of the transaction node and the neighbor nodes thereof, based on attention weight
Classifying and aggregating to obtain embedded representation of each meta-path type of transaction node i
Wherein sigma is a sigmoid function;
a25, embedded representation of Each transaction node
Aggregating according to meta-path type to obtain embedded representation of each meta-path type
Further, in step A22, the attention score is calculated using sigmoid function
Wherein the content of the first and second substances,
for the learnable parameter matrix, superscript T represents matrix transpose, and |, is the matrix connection symbol; sigmoid is an activation function.
Specifically, in step A3, the step of fusing the embedded representation of each meta-path type according to the embedded representation of each meta-path type of the financial heterogeneous information graph based on the self-attention mechanism to obtain the graph embedded representation of the financial heterogeneous information graph includes:
a31, obtaining the attention scores among the meta-path types by adopting a self-attention mechanism
Wherein the content of the first and second substances,
representing mth type meta path
The attention score of (a);
representing meta path
An embedded representation of (a); m represents the number of meta path types; att ψ Indicating a mechanism of attention;
a32, using SoftMax function, score attention
Performing standardization to obtain attention weight
A33, attention-based weighting
Fusing the embedded representations of the meta-path types;
wherein Z represents a graph-embedded representation of the financial heterogeneous information graph.
Further, in step A31, the attention score is calculated as follows
Wherein the content of the first and second substances,
the method is characterized in that the method is embedded expression of an mth type element path of a transaction node i, tanh represents a hyperbolic tangent activation function, q, W' and b are learnable parameters, superscript T represents matrix transposition, and N is the number of the transaction nodes.
In order to facilitate the implementation of the financial transaction abnormity detection method in a cross-region manner, the invention also provides a cross-region sustainable training method for the financial transaction abnormity detection method, which comprises the following steps:
b1, according to a continuous learning strategy, constructing initial parameters and a playback sample set of a financial transaction abnormity detection model of a target region;
the continuous learning strategy comprises: a knowledge playback strategy and a parameter smoothing strategy;
the knowledge playback strategy is: sampling a preset number of samples from the samples of the reference area to construct a playback sample set; the parameter smoothing strategy is as follows: extracting a parameter matrix of a financial transaction abnormity detection model of a reference region, taking the parameter matrix as an initial parameter matrix of a target region, and evaluating the parameter matrix to obtain a parameter importance degree score; the reference area is an area which is trained by a financial transaction abnormity detection model, and is a random area initially, and a parameter matrix of the random area is obtained in a random mode;
b2, constructing a financial heterogeneous information graph according to the playback sample set and the sample set of the target area, and obtaining an abnormal detection result of the target area according to the financial transaction abnormal detection method;
and B3, calculating training loss by adopting a cross entropy loss function according to the abnormal detection result and the real label, performing iterative training on the financial transaction abnormal detection model of the target area until a preset iteration turn or model convergence is achieved, obtaining the financial transaction abnormal detection model of the target area, and constraining the training process according to the parameter importance degree score obtained in the step B1 based on a parameter smoothing strategy.
Further, in step B1, sampling a preset number of samples from the samples in the reference region to construct a playback sample set, including:
firstly, randomly sampling transaction nodes from a financial heterogeneous information graph of a reference area, and calculating the mean value of characteristic vectors of the sampling transaction nodes as a playback prototype c l-1 ：
Wherein the content of the first and second substances,
a set of transaction nodes obtained for random sampling,
is composed of
The feature vector of the intermediate transaction node i,
the number of transaction nodes obtained by random sampling;
then, a playback sample set is constructed as follows:
the first method is to obtain the proximity degree of the feature vector of each transaction node and the playback prototype according to sampling, and the proximity degree is preset according to the number from near to far
Selecting a transaction node, wherein the selected transaction node and neighbor nodes thereof form a playback sample set, and the number of the transaction nodes obtained by random sampling is greater than the preset number;
generating Gaussian noise, correcting the feature vector of the transaction node obtained by random sampling based on the generated Gaussian noise, forming a playback sample set based on the transaction node obtained by correction and the neighbor node of the transaction node obtained by random sampling, wherein the number of the transaction nodes obtained by random sampling is a preset number;
and in the third mode, firstly, processing is carried out according to the first mode, and then processing is carried out according to the second mode, so that the construction of the playback sample set is completed.
Further, in step B1, a parameter importance degree score is obtained by evaluating a parameter matrix of the reference region financial transaction anomaly detection model according to the gradient change of the reference region transaction node feature vector under the condition of the parameter matrix.
Specifically, the method for evaluating the parameter matrix by adopting the snow-charging information matrix comprises the following steps:
wherein, theta l-1 Parameter matrix, X, representing a financial transaction anomaly detection model of a reference area l-1 Gold representing reference areaFusing a set of feature vectors of each transaction node in the heterogeneous information graph; g (x; theta) l-1 ) Represents the value calculated at Θ by a loss function l-1 The gradient of the eigenvector of the trading node x under the condition is changed, and T represents matrix transposition.
In particular, the loss function in step B3
Comprises the following steps:
wherein the content of the first and second substances,
in order to achieve a cross-entropy loss,
respectively, regularization constraint items for constraining the updating process according to the parameter importance degree scores;
the cross entropy loss is calculated as follows:
wherein the content of the first and second substances,
a true tag representing the ith transaction node,
A prediction tag representing the ith transaction node, wherein N is the number of the transaction nodes;
the regularization constraint term is calculated as follows:
wherein, theta l A parameter matrix representing a financial transaction anomaly detection model of the target area, S represents an area set of which the training of the financial transaction anomaly detection model is finished before the target area, and theta s A parameter matrix, Θ, representing a financial transaction anomaly detection model for the s-th region of the set of regions l-1 A parameter matrix representing a reference region financial transaction anomaly detection model for the target region;
a matrix representing the importance degree of the financial transaction anomaly detection model parameters of the target area; gamma and lambda are smoothing factors.
The beneficial effects of the invention are:
according to the invention, by constructing the heterostructure information graph, the defects that the previous homogeneity graph is difficult to retain multi-type semantic information and cannot capture dynamic space-time characteristics are overcome, high-order semantics such as time information are fully excavated, the degree of information quantity can be greatly enriched and obtained, a deep graph neural network model is used for fusing nodes, node connection and network structures, learning low-dimensional embedded representation and detecting abnormal results, the abnormal detection efficiency is improved, and the financial abnormality detection precision is also improved.
Meanwhile, the actual requirements of cross-region business expansion in a real financial scene are fully considered, the defect that the current financial abnormity detection system is limited to a single region is overcome based on continuous learning of a depth map neural network model, and cross-region deployment can be conveniently carried out.
The invention is suitable for various financial risk management tasks such as financial transaction fraud detection and the like, and helps financial enterprises to carry out risk control and guarantee financial safety in cross-regional financial business.
Drawings
FIG. 1 is a flow chart of a financial transaction anomaly detection model training in an embodiment of the present invention;
FIG. 2 is a diagram illustrating a process of constructing a financial heterogeneous information map according to an embodiment of the present invention;
FIG. 3 is a flow chart of anomaly detection for financial transactions using a model in an embodiment of the invention.
Detailed Description
The invention aims to provide a financial transaction abnormity detection method and a cross-regional sustainable training method thereof, which can improve the efficiency and the precision of financial transaction abnormity detection and can carry out cross-regional deployment conveniently.
Continuous training, i.e., continuous Learning (CL), has the ability to use the knowledge of one task on another task and to learn the next task without forgetting how to do the previous task. This learning paradigm needs to address: 1. how to use the experience of the previous task to enable the current task to be learned faster and better; 2. when the current task is learned, the task which is learned before can not be forgotten. Currently, the main challenge of continuous learning is disaster forgetting, i.e., knowledge obtained in the past is forgotten when a new task is learned, which will result in the previous task performing less than ever. In order to realize cross-regional sustainable training, the invention adopts a knowledge playback strategy and a parameter smoothing strategy, and solves the problems of catastrophic forgetting and knowledge migration in cross-regional sustainable learning.
Meanwhile, the invention overcomes the defects that the previous homograph is difficult to retain multi-type semantic information and cannot capture dynamic space-time characteristics by constructing a heterostructure information graph, fully excavates high-order semantics such as time information and the like, greatly enriches the degree of acquirable information quantity, fuses nodes, node connections and network structures by using a deep graph neural network model, learns low-dimensional embedded representation, and detects abnormal results, thereby not only improving the cross-region abnormal detection efficiency, but also improving the financial abnormal detection accuracy.
Example (b):
the method comprises two parts of training a financial transaction abnormity detection model and carrying out financial transaction abnormity detection by using the model.
As shown in fig. 1, the process of training the financial transaction anomaly detection model includes the following steps:
s1, constructing initial parameters and a playback sample set of a financial transaction abnormity detection model of a target region:
in the step, aiming at the data sample data of the cross-regional financial transaction, firstly, different transaction regions are divided according to the geography latitude and longitude and the occurrence sequence of the regions is randomly set so as to simulate the situation that the transaction regions are continuously increased in the real transaction scene. In addition, the embodiment also performs a series of preprocessing operations on the data samples, including deleting repeated information, correcting existing errors, and ensuring data consistency and data quality of transactions.
The method is a cross-region financial transaction abnormity detection method capable of continuously learning, therefore, during training, a region-by-region training mode is adopted, and each divided region is trained respectively, so that a financial transaction abnormity detection model corresponding to the region is obtained; and aiming at the financial transaction abnormity detection model of each region, the construction mode of the initial parameters and the playback sample set is adopted to construct the initial parameters and the playback sample set by adopting a sustainable learning strategy, so that the financial transaction abnormity detection model corresponding to the region is suitable for the region which is trained and covered by the model before the region is trained.
Specifically, the continuous learning strategy includes: knowledge playback strategies and parameter smoothing strategies.
The knowledge playback strategy is: sampling a preset number of samples from the samples of the reference area to construct a playback sample set; the parameter smoothing strategy is as follows: and extracting a parameter matrix of the financial transaction abnormity detection model of the reference region, taking the parameter matrix as an initial parameter matrix of the target region, and evaluating the parameter matrix to obtain a parameter importance degree score.
The reference area is an area which is trained by the financial transaction abnormity detection model, and is a random area initially, and a parameter matrix of the random area is obtained in a random mode.
Further, if the target region is the first region, that is, the first region to be model-trained, since there is no learning basis before, that is, there is no model that has already been trained in other regions, another region may be randomly selected as the reference region, and the random initial parameter of the model is used as the initial parameter. If the target area is not the first area, that is, there is a model that has already been trained in other areas, for the target area, the "existing area" in which the model training has already been completed, that is, the area covered by the existing model, may be used as the reference area for the target area learning.
Of course, in order to ensure the maximization of the coverage area of the model after the training of the target area is completed, it is preferable to select a training-completed area before the target area from the existing areas as the reference area.
Furthermore, the knowledge replay strategy can adopt a small empirical knowledge register to randomly replay samples from the selected reference region, so as to assist the training of the anomaly detection model of the current target region.
Specifically, the present embodiment employs a prototype method (protocols), i.e. Feature averaging (MF), to construct a playback sample set for alleviating the training instability problem caused by random sampling, including:
firstly, randomly sampling transaction nodes from a financial heterogeneous information graph of a reference area, and calculating the mean value of characteristic vectors of the sampling transaction nodes as a playback prototype c l-l ：
Wherein, the first and the second end of the pipe are connected with each other,
a set of transaction nodes obtained for random sampling,
is composed of
The feature vector of the intermediate transaction node i,
the number of transaction nodes obtained by random sampling;
and then, generating Gaussian noise, correcting the feature vector of the transaction node obtained by random sampling based on the generated Gaussian noise, forming a playback sample set based on the transaction node obtained by correction and the neighbor node of the transaction node obtained by random sampling, wherein the number of the transaction nodes obtained by random sampling is a preset number. Through robustness experiments, when the playback ratio is 0.1, the model performs optimally, so the preset number of the embodiment is the number of samples when the playback ratio is 0.1.
Of course, in addition to the above, it is also possible to adopt: judging the proximity degree of the characteristic vector of each transaction node and the playback prototype through the Euclidean distance, and according to the preset quantity from near to far
Selecting a transaction node, wherein the selected transaction node and neighbor nodes thereof form a playback sample set, and the number of the transaction nodes obtained by random sampling is greater than the preset number; or, firstly, screening in a Euclidean distance mode, and then correcting in a Gaussian noise mode to complete the construction of a playback sample set.
The parameter smoothing strategy is used for restraining the updating of important parameters through a smoothing parameter method, and can relieve the problem of catastrophic forgetting among cross-region fraud detection tasks. Further, for the parameter matrix of the reference region financial transaction anomaly detection model, the parameter importance degree score can be obtained by evaluating according to the gradient change of the reference region transaction node feature vector under the condition of the parameter matrix.
Specifically, in this embodiment, the snow-charging information matrix is adopted to evaluate the parameter matrix, which includes:
wherein, theta l-1 Parameter matrix, X, representing a financial transaction anomaly detection model of a reference area l-1 A set of feature vectors of each transaction node in a financial heterogeneous information graph representing a reference region; g (x; theta) l-1 ) Represents the value calculated at Θ by a loss function l-1 The gradient of the eigenvector of the trading node x under the condition is changed, and T represents the matrix transposition.
S2, constructing a financial heterogeneous information graph according to the playback sample set and the sample set of the target area:
the method for constructing the financial heterogeneous information graph mainly comprises three parts, namely entity node construction, meta-path construction and node feature construction.
Specifically, as shown in fig. 2, the method includes:
s21, entity node construction and node feature construction:
in the invention, four types of entities are extracted from a financial transaction data sample as entity nodes, namely a user node, a merchant node, a time node and a transaction node, and a node feature vector is obtained according to corresponding knowledge information codes.
Specifically, in an embodiment, the node feature vector is obtained by encoding the corresponding knowledge information, and includes:
the corresponding knowledge information of the user node is the unique identity ID of the transaction user;
the corresponding knowledge information of the merchant node is the unique identity ID of the transaction merchant;
the time node corresponds to the knowledge information and is the time stamp information of the transaction;
the transaction node corresponding knowledge information comprises a transaction serial number, a transaction place, a transaction amount and a transaction type.
The user ID, the merchant ID, the transaction code, and the time identifier are all structured, and therefore, can be directly used as a feature vector code of a corresponding node.
In order to capture the sensitivity of the financial anomaly detection task to the transaction time, the invention adds a time node using timestamp information as an identifier, but in order to facilitate the classification processing, further, the feature vector coding of the time node comprises: firstly, 24 hours per day are divided into time windows according to time intervals, then, coding is carried out according to digital information of the time windows according to the time periods corresponding to the time windows to which the transaction timestamps belong, and the feature vector codes of the time nodes are obtained. Of course, onehot encoding can also be adopted.
In order to further describe the inherent semantics of each transaction node and improve the efficiency of the anomaly detection task, when a financial heterogeneous information graph is constructed, features related to transactions are all used as attribute features of the transaction nodes, and specifically, feature vector coding of the transaction nodes comprises the following steps: firstly, coding is carried out according to the digital information of a transaction serial number, coding is carried out according to the longitude and latitude of a transaction place, coding is carried out according to the digital information of a transaction amount, the transaction type is coded according to an onehot mode, and then the serial number, the transaction place, the transaction amount and the transaction type obtained by coding are spliced to form a feature vector code of a transaction node.
S22, meta path construction:
the meta path is widely applied to semantic exploration of heterogeneous networks and aims to define the type of the corresponding relation of the connecting edges among the nodes. The invention constructs three types of meta paths, namely a user-transaction path, a merchant-transaction path and a time-transaction path, so as to describe heterogeneous associated information derived from given semantics.
Wherein "transaction-user-transaction" describes different transactions conducted by the same user; "transaction-timestamp-transaction" describes different transactions that occur at the same time period; "transaction-merchant-transaction" describes different transactions that occur with the same merchant.
Therefore, the financial heterogeneous information graph constructed by the invention is a heterogeneous network structure with various nodes and relationship types, can fully utilize different types of information related to transactions, constructs a heterogeneous transaction information network through various types of meta-paths, and integrates abundant semantic information.
S3, obtaining the embedded representation of each node through a coding network according to the feature vector of each node; according to the embedded representation of the transaction node and the neighbor nodes thereof, classifying and aggregating based on an attention mechanism according to the meta-path types of the transaction node and the neighbor nodes thereof to obtain the embedded representation of each meta-path type of the transaction node; then, according to the meta-path type, aggregating the embedded representation of the meta-path type corresponding to each transaction node to obtain the embedded representation of each meta-path type of the financial heterogeneous information graph, including:
s31, mapping the feature vectors of all nodes in the financial heterogeneous information graph to a feature space through a coding network to obtain the embedded representation of each node in the feature space
Wherein the content of the first and second substances,
for the embedded representation of the ith node in the feature space, W x In order to encode the learnable parameter matrix of the network,
a feature vector of an ith node;
s32, obtaining the attention score between the transaction node and the neighbor node thereof by adopting an attention mechanism
Wherein, att η It is shown that the attention mechanism is,
for the attention score between trading node i and neighbor node j,
the type of the meta-path between the transaction node i and the neighbor node j is determined;
specifically, in the present embodiment, the attention score is calculated by using sigmoid function
Wherein the content of the first and second substances,
for the learnable parameter matrix, superscript T represents matrix transpose, and |, is the matrix connection symbol; sigmoid is an activation function.
S33, scoring the attention of each meta path type according to the meta path type by using a SoftMax function
Performing standardization to obtain attention weight
Wherein the content of the first and second substances,
the meta-path type in the neighbor node of the transaction node i is the m-th type meta-path
A set of neighbor nodes of (a);
s34, according to the meta-path type of the transaction node and the neighbor nodes thereof, based on attention weight
Classifying and aggregating to obtain embedded representation of each meta-path type of transaction node i
Wherein sigma is a sigmoid function;
s35, embedding representation of each transaction node
Aggregating according to meta-path type to obtain embedded representation of each meta-path type
S4, according to the embedded representation of each meta path type of the financial heterogeneous information graph, based on an attention mechanism, fusing the embedded representation of each meta path type to obtain a graph embedded representation of the financial heterogeneous information graph, wherein the graph embedded representation comprises the following steps:
s41, obtaining attention scores among all element path types by adopting a self-attention mechanism
Wherein, the first and the second end of the pipe are connected with each other,
representing mth type meta path
The attention score of (a);
representing meta path
An embedded representation of (a); m represents the number of meta-path types, M =3 in the present embodiment; att ψ Indicating a mechanism of attention;
specifically, the attention score is calculated as follows
Wherein, the first and the second end of the pipe are connected with each other,
the method is characterized in that the method is embedded expression of an mth type element path of a transaction node i, tanh represents a hyperbolic tangent activation function, q, W' and b are learnable parameters, superscript T represents matrix transposition, and N is the number of the transaction nodes.
S42, scoring attention by using SoftMax function
Performing standardization to obtain attention weight
S43, attention-based weighting
Fusing meta path classesAn embedded representation of a type;
wherein Z represents a graph-embedded representation of the financial heterogeneous information graph.
S5, acquiring an abnormal detection result based on the graph embedding representation:
in the step, based on the obtained graph embedding representation, a classification label of fraud detection is output by adopting a full connection layer as an abnormal detection result; if the output result is "1" indicating "abnormal transaction", and "0" indicating "not abnormal transaction".
S6, updating model parameters:
in the step, according to an abnormal detection result and a real label, a cross entropy loss function is adopted to calculate training loss, iterative training is carried out on a financial transaction abnormal detection model of a target area until a preset iteration round or model convergence is achieved, the financial transaction abnormal detection model of the target area is obtained, and based on a parameter smoothing strategy, a training process is restrained according to the parameter importance degree score obtained in the step B1
In particular, the loss function
Comprises the following steps:
wherein the content of the first and second substances,
in order to achieve a cross-entropy loss,
respectively, regularization constraint items for constraining the updating process according to the parameter importance degree scores;
the cross entropy loss is calculated as follows:
wherein the content of the first and second substances,
a true tag representing the ith transaction node,
A prediction tag representing the ith transaction node, wherein N is the number of the transaction nodes;
the regularization constraint term is calculated as follows:
wherein, theta l A parameter matrix representing a financial transaction anomaly detection model of the target area, S represents an area set of which the training of the financial transaction anomaly detection model is finished before the target area, and theta s A parameter matrix, Θ, representing a financial transaction anomaly detection model for the s-th region of the set of regions l-1 A parameter matrix representing a reference region financial transaction anomaly detection model for the target region;
a matrix representing the degree of importance of the financial transaction anomaly detection model parameters of the target area; gamma and lambda are smoothing factors.
After the anomaly detection model of the target area is obtained through training, the financial transaction data of the corresponding area can be detected by using the model, so that whether the transaction is abnormal transactions such as fraud or the like is judged, and the process is shown in fig. 3 and specifically comprises the following steps:
a1, constructing a financial heterogeneous information graph according to financial transaction data to be detected in a target area;
a2, obtaining the embedded representation of each node through a coding network according to the characteristic vector of each node; according to the embedded representation of the transaction node and the neighbor nodes thereof, classifying and aggregating based on an attention mechanism according to the meta-path types of the transaction node and the neighbor nodes thereof to obtain the embedded representation of each meta-path type of the transaction node; then, according to the meta-path type, aggregating the embedded representation of the meta-path type corresponding to each transaction node to obtain the embedded representation of each meta-path type of the financial heterogeneous information graph;
a3, according to the embedded representation of each meta path type of the financial heterogeneous information graph, fusing the embedded representation of each meta path type based on a self-attention mechanism to obtain graph embedded representation of the financial heterogeneous information graph;
and A4, based on the map embedded representation of the financial heterogeneous information map, adopting a classification label of the all-connection layer output abnormity detection as an abnormity detection result.
Although the present invention has been described herein with reference to the preferred embodiments thereof, which are intended to be illustrative only and not to be limiting of the invention, it will be understood that numerous other modifications and embodiments can be devised by those skilled in the art that will fall within the spirit and scope of the principles of this disclosure.
Claims (13)
1. The financial transaction abnormity detection method is characterized by comprising the following steps:
a1, constructing a financial heterogeneous information graph according to-be-detected financial transaction data of a target area; the entity nodes of the financial heterogeneous information graph comprise user nodes, merchant nodes, time nodes and transaction nodes, the characteristic vectors of the nodes are obtained according to corresponding knowledge information codes, and the types of meta-paths comprise user-transaction paths, merchant-transaction paths and time-transaction paths;
a2, obtaining the embedded representation of each node through a coding network according to the characteristic vector of each node; according to the embedded representation of the transaction node and the neighbor nodes thereof, classifying and aggregating based on an attention mechanism according to the meta-path types of the transaction node and the neighbor nodes thereof to obtain the embedded representation of each meta-path type of the transaction node; then, according to the meta-path type, aggregating the embedded representation of the meta-path type corresponding to each transaction node to obtain the embedded representation of each meta-path type of the financial heterogeneous information graph;
a3, according to the embedded representation of each meta path type of the financial heterogeneous information graph, fusing the embedded representation of each meta path type based on an attention mechanism to obtain graph embedded representation of the financial heterogeneous information graph;
and A4, based on the map embedded representation of the financial heterogeneous information map, adopting a classification label of the all-connection layer output abnormity detection as an abnormity detection result.
2. The financial transaction anomaly detection method according to claim 1,
in step A1, the node feature vector is obtained by encoding the corresponding knowledge information, and includes:
the corresponding knowledge information of the user node is the unique identity ID of the transaction user;
the corresponding knowledge information of the merchant node is the unique identity ID of the transaction merchant;
the time node corresponds to the knowledge information and is timestamp information of the transaction;
the transaction node corresponding knowledge information comprises a transaction serial number, a transaction place, a transaction amount and a transaction type.
3. The financial transaction anomaly detection method according to claim 2,
feature vector encoding of a time node, comprising: firstly, 24 hours per day are divided into time windows according to time intervals, then coding is carried out according to digital information or an onehot mode of the time windows according to the time periods corresponding to the time windows to which the transaction timestamps belong, and the feature vector codes of the time nodes are obtained.
4. The financial transaction anomaly detection method according to claim 2,
feature vector encoding of a transaction node, comprising: firstly, coding is carried out according to the digital information of a transaction serial number, coding is carried out according to the longitude and latitude of a transaction place, coding is carried out according to the digital information of a transaction amount, the transaction type is coded according to an onehot mode, and then the serial number, the transaction place, the transaction amount and the transaction type obtained by coding are spliced to form a feature vector code of a transaction node.
5. The financial transaction abnormality detection method according to any one of claims 1 to 4, wherein in step A2, the embedded representation of each node is obtained through the coding network based on the feature vector of each node; according to the embedded representation of the transaction node and the neighbor nodes thereof, classifying and aggregating based on an attention mechanism according to the meta-path types of the transaction node and the neighbor nodes thereof to obtain the embedded representation of each meta-path type of the transaction node; then, according to the meta-path type, aggregating the embedded representation of the meta-path type corresponding to each transaction node to obtain the embedded representation of each meta-path type of the financial heterogeneous information graph, including:
a21, mapping the feature vectors of all nodes in the financial heterogeneous information graph to a feature space through a coding network to obtain the embedded representation of each node in the feature space
Wherein the content of the first and second substances,
for the embedded representation of the ith node in the feature space, W x In order to encode the learnable parameter matrix of the network,
the feature vector of the ith node;
a22, adopting an attention mechanism to obtain attention scores between the transaction nodes and the neighbor nodes thereof
Wherein, att η A mechanism of attention is shown that is,
for the attention score between trading node i and neighbor node j,
the type of the meta-path between the transaction node i and the neighbor node j is determined;
a23, using a SoftMax function to score the attention of each meta-path type according to the meta-path type
Performing standardization to obtain attention weight
Wherein the content of the first and second substances,
the meta-path type in the neighbor node of the transaction node i is the m-th type meta-path
A set of neighbor nodes of (1);
a24, according to the meta-path type of the transaction node and the neighbor nodes thereof, based on attention weight
Classifying and aggregating to obtain embedded representation of each meta-path type of transaction node i
Wherein sigma is a sigmoid function;
a25, embedded representation of Each transaction node
Aggregating according to meta-path type to obtain embedded representation of each meta-path type
6. The financial transaction abnormality detection method according to claim 5,
in step A22, the attention score is calculated by using sigmoid function
Wherein the content of the first and second substances,
for learnable parameter matrix, superscript T represents matrix transposition, and | is matrix connectionReceiving symbols; sigmoid is an activation function.
7. The financial transaction abnormality detection method according to any one of claims 1 to 4, wherein in the step A3, the graph-embedded representation of the financial heterogeneous information graph is obtained by fusing the embedded representations of the meta path types based on the self-attention mechanism based on the embedded representation of the meta path types of the financial heterogeneous information graph, and the method includes:
a31, obtaining attention scores among the meta-path types by adopting a self-attention mechanism
Wherein the content of the first and second substances,
representing mth type meta path
The attention score of (a);
representing meta path
An embedded representation of (a); m represents the number of meta path types; att ψ Indicating a mechanism of attention;
a32, using SoftMax function, score attention
Performing standardization to obtain attention weight
A33, attention-based weighting
Fusing the embedded representations of the meta-path types;
wherein Z represents a graph-embedded representation of the financial heterogeneous information graph.
8. The financial transaction anomaly detection method according to claim 7,
in step A31, the attention score is calculated as follows
Wherein the content of the first and second substances,
the method is characterized in that the method is embedded expression of an mth type element path of a transaction node i, tanh represents a hyperbolic tangent activation function, q, W' and b are learnable parameters, superscript T represents matrix transposition, and N is the number of the transaction nodes.
9. A cross-regional sustainable training method for the financial transaction anomaly detection method according to any one of claims 1 to 8, comprising the steps of:
b1, according to a continuous learning strategy, constructing initial parameters and a playback sample set of a financial transaction abnormity detection model of a target region;
the continuous learning strategy comprises: a knowledge playback strategy and a parameter smoothing strategy;
the knowledge playback strategy is: sampling a preset number of samples from the samples of the reference area to construct a playback sample set; the parameter smoothing strategy is as follows: extracting a parameter matrix of a financial transaction abnormity detection model of a reference region, taking the parameter matrix as an initial parameter matrix of a target region, and evaluating the parameter matrix to obtain a parameter importance degree score; the reference area is an area which is trained by a financial transaction abnormity detection model, and is a random area initially, and a parameter matrix of the random area is obtained in a random mode;
b2, constructing a financial heterogeneous information graph according to the playback sample set and the sample set of the target area, and obtaining an abnormal detection result of the target area according to the financial transaction abnormal detection method of any one of claims 1 to 8;
and B3, calculating training loss by adopting a cross entropy loss function according to the abnormal detection result and the real label, performing iterative training on the financial transaction abnormal detection model of the target area until a preset iteration turn or model convergence is achieved, obtaining the financial transaction abnormal detection model of the target area, and constraining the training process according to the parameter importance degree score obtained in the step B1 based on a parameter smoothing strategy.
10. The method of cross-regional sustainable training of claim 9,
in step B1, sampling a preset number of samples from the samples in the reference region to construct a playback sample set, including:
firstly, randomly sampling transaction nodes from a financial heterogeneous information graph of a reference area, and calculating the mean value of characteristic vectors of the sampling transaction nodes as a playback prototype c l-1 ：
Wherein the content of the first and second substances,
a set of transaction nodes obtained for random sampling,
is composed of
The feature vector of the intermediate transaction node i,
the number of transaction nodes obtained by random sampling;
then, a playback sample set is constructed as follows:
the first method is to obtain the proximity degree of the feature vector of each transaction node and the playback prototype according to sampling, and the proximity degree is preset according to the number from near to far
Selecting transaction nodes, forming a playback sample set by the selected transaction nodes and neighbor nodes thereof, wherein the number of the transaction nodes obtained by random sampling is larger than the preset number;
generating Gaussian noise, modifying the feature vector of the transaction node obtained by random sampling based on the generated Gaussian noise, forming a playback sample set based on the transaction node obtained by modification and the neighbor node of the transaction node obtained by random sampling, wherein the number of the transaction nodes obtained by random sampling is a preset number;
and in the third mode, firstly, processing is carried out according to the first mode, and then processing is carried out according to the second mode, so that the construction of the playback sample set is completed.
11. The method according to claim 9, wherein in step B1, a parameter importance score is evaluated and obtained for a parameter matrix of the financial transaction anomaly detection model in the reference area according to gradient changes of feature vectors of transaction nodes in the reference area under the condition of the parameter matrix.
12. The cross-regional sustainable training method of claim 11,
adopting a snow cost information matrix to evaluate the parameter matrix, wherein the evaluation comprises the following steps:
wherein, theta l-1 Parameter matrix, X, representing a financial transaction anomaly detection model of a reference area l-1 A set of feature vectors of each transaction node in a financial heterogeneous information graph representing a reference region; g (x; theta) l-1 ) Represents the value calculated at Θ by a loss function l-1 The gradient of the feature vector of the transaction node x under the condition changes,
representing a matrix transposition.
13. The method according to any one of claims 9-12, wherein the loss function in step B3 is a loss function
Comprises the following steps:
wherein the content of the first and second substances,
in order to achieve a cross-entropy loss,
respectively updated according to the score of the importance degree of the parameterRegularization constraint terms for the process constraint;
the cross entropy loss is calculated as follows:
wherein the content of the first and second substances,
a true tag representing the ith transaction node,
A prediction label representing the ith transaction node, wherein N is the number of the transaction nodes;
the regularization constraint term is calculated as follows:
wherein, theta l A parameter matrix representing a financial transaction anomaly detection model of the target area, S represents an area set of which the training of the financial transaction anomaly detection model is finished before the target area, and theta s A parameter matrix, Θ, representing a financial transaction anomaly detection model for the s-th region of the set of regions l-1 A parameter matrix representing a reference region financial transaction anomaly detection model for the target region;
a matrix representing the degree of importance of the financial transaction anomaly detection model parameters of the target area; gamma and lambda are smoothing factors.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211301695.8A CN115660688B (en) | 2022-10-24 | 2022-10-24 | Financial transaction anomaly detection method and cross-regional sustainable training method thereof |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211301695.8A CN115660688B (en) | 2022-10-24 | 2022-10-24 | Financial transaction anomaly detection method and cross-regional sustainable training method thereof |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN115660688A true CN115660688A (en) | 2023-01-31 |
| CN115660688B CN115660688B (en) | 2024-04-30 |
Family
ID=84991891
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211301695.8A Active CN115660688B (en) | 2022-10-24 | 2022-10-24 | Financial transaction anomaly detection method and cross-regional sustainable training method thereof |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115660688B (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| 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 |
| CN117708821A (en) * | 2024-02-06 | 2024-03-15 | 山东省计算中心（国家超级计算济南中心） | Method, system, equipment and medium for detecting Lesu software based on heterogeneous graph embedding |
Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109003089A (en) * | 2018-06-28 | 2018-12-14 | 中国工商银行股份有限公司 | risk identification method and device |
| CN111090780A (en) * | 2019-12-09 | 2020-05-01 | 中国建设银行股份有限公司 | Method and device for determining suspicious transaction information, storage medium and electronic equipment |
| WO2021179838A1 (en) * | 2020-03-10 | 2021-09-16 | 支付宝(杭州)信息技术有限公司 | Prediction method and system based on heterogeneous graph neural network model |
| CN113506179A (en) * | 2021-09-13 | 2021-10-15 | 北京大学深圳研究生院 | Method for detecting abnormal entity in digital currency transaction and storage medium |
| CN113706279A (en) * | 2021-06-02 | 2021-11-26 | 同盾科技有限公司 | Fraud analysis method and device, electronic equipment and storage medium |
| CN114187112A (en) * | 2021-12-15 | 2022-03-15 | 深圳前海微众银行股份有限公司 | Training method of account risk model and determination method of risk user group |
| CN114386727A (en) * | 2020-10-19 | 2022-04-22 | 腾讯科技（深圳）有限公司 | Risk identification method, device, equipment and storage medium |
| CN114565053A (en) * | 2022-03-10 | 2022-05-31 | 天津大学 | Deep heterogeneous map embedding model based on feature fusion |
-
2022
- 2022-10-24 CN CN202211301695.8A patent/CN115660688B/en active Active
Patent Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109003089A (en) * | 2018-06-28 | 2018-12-14 | 中国工商银行股份有限公司 | risk identification method and device |
| CN111090780A (en) * | 2019-12-09 | 2020-05-01 | 中国建设银行股份有限公司 | Method and device for determining suspicious transaction information, storage medium and electronic equipment |
| WO2021179838A1 (en) * | 2020-03-10 | 2021-09-16 | 支付宝(杭州)信息技术有限公司 | Prediction method and system based on heterogeneous graph neural network model |
| CN114386727A (en) * | 2020-10-19 | 2022-04-22 | 腾讯科技（深圳）有限公司 | Risk identification method, device, equipment and storage medium |
| CN113706279A (en) * | 2021-06-02 | 2021-11-26 | 同盾科技有限公司 | Fraud analysis method and device, electronic equipment and storage medium |
| CN113506179A (en) * | 2021-09-13 | 2021-10-15 | 北京大学深圳研究生院 | Method for detecting abnormal entity in digital currency transaction and storage medium |
| CN114187112A (en) * | 2021-12-15 | 2022-03-15 | 深圳前海微众银行股份有限公司 | Training method of account risk model and determination method of risk user group |
| CN114565053A (en) * | 2022-03-10 | 2022-05-31 | 天津大学 | Deep heterogeneous map embedding model based on feature fusion |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| 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 |
| CN117708821A (en) * | 2024-02-06 | 2024-03-15 | 山东省计算中心（国家超级计算济南中心） | Method, system, equipment and medium for detecting Lesu software based on heterogeneous graph embedding |
| CN117708821B (en) * | 2024-02-06 | 2024-04-30 | 山东省计算中心（国家超级计算济南中心） | Method, system, equipment and medium for detecting Lesu software based on heterogeneous graph embedding |
Also Published As
| Publication number | Publication date |
|---|---|
| CN115660688B (en) | 2024-04-30 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Bowden et al. | Input determination for neural network models in water resources applications. Part 1—background and methodology | |
| CN115660688B (en) | Financial transaction anomaly detection method and cross-regional sustainable training method thereof | |
| Remesan et al. | Hydrological data driven modelling | |
| Hamzah et al. | Imputation methods for recovering streamflow observation: A methodological review | |
| Guo et al. | Integration of support vector regression with distributed Gauss-Newton optimization method and its applications to the uncertainty assessment of unconventional assets | |
| Yin et al. | Bayesian set pair analysis and machine learning based ensemble surrogates for optimal multi-aquifer system remediation design | |
| Xu et al. | Deep transfer learning based on transformer for flood forecasting in data-sparse basins | |
| Jiang et al. | Simultaneous identification of contaminant sources and hydraulic conductivity field by combining geostatistics method with self-organizing maps algorithm | |
| Han et al. | Comprehensive analysis for production prediction of hydraulic fractured shale reservoirs using proxy model based on deep neural network | |
| CN113314188B (en) | Graph structure enhanced small sample learning method, system, equipment and storage medium | |
| Godoy et al. | Ensemble random forest filter: An alternative to the ensemble Kalman filter for inverse modeling | |
| CN110889493A (en) | Method and device for adding disturbance aiming at relational network | |
| Liu et al. | DCENet: A dynamic correlation evolve network for short-term traffic prediction | |
| Jhin et al. | Learnable path in neural controlled differential equations | |
| CN117134978A (en) | Vehicle identity verification method and system based on local and global behavior pattern analysis | |
| Chen et al. | A cellular automaton integrating spatial case-based reasoning for predicting local landslide hazards | |
| Chen et al. | CNN-LSTM-attention deep learning model for mapping landslide susceptibility in Kerala, India | |
| Das et al. | A Bayesian sparse generalized linear model with an application to multiscale covariate discovery for observed rainfall extremes over the United States | |
| Thiam et al. | Reservoir interwell connectivity estimation from small datasets using a probabilistic data driven approach and uncertainty quantification | |
| Huang et al. | Short-term traffic flow prediction based on graph convolutional network embedded lstm | |
| Tanaka et al. | Methods for Probabilistic Uncertainty Quantification with Reliable Subsurface Assessment and Robust Decision-Making | |
| Liao et al. | Traj2Traj: A road network constrained spatiotemporal interpolation model for traffic trajectory restoration | |
| Landwehr et al. | Demonstration of the impacts of anti-sedimentation techniques on Japanese reservoir siltation via mass data ANN analysis | |
| Pribić | Stochastic deep learning for compressive-sensing radar | |
| Kor | Decision-Driven Data Analytics for Well Placement Optimization in Field Development Scenario-Powered by 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 |