CN114757581A - Financial transaction risk assessment method and device, electronic equipment and computer readable medium - Google Patents
Financial transaction risk assessment method and device, electronic equipment and computer readable medium Download PDFInfo
- Publication number
- CN114757581A CN114757581A CN202210537782.7A CN202210537782A CN114757581A CN 114757581 A CN114757581 A CN 114757581A CN 202210537782 A CN202210537782 A CN 202210537782A CN 114757581 A CN114757581 A CN 114757581A
- Authority
- CN
- China
- Prior art keywords
- path
- vector
- features
- attention weight
- decision
- 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
Links
- 238000000034 method Methods 0.000 title claims abstract description 48
- 238000012502 risk assessment Methods 0.000 title claims abstract description 29
- 239000013598 vector Substances 0.000 claims abstract description 115
- 230000004927 fusion Effects 0.000 claims abstract description 74
- 230000004931 aggregating effect Effects 0.000 claims abstract description 11
- 238000003066 decision tree Methods 0.000 claims description 27
- 239000011159 matrix material Substances 0.000 claims description 25
- 238000013507 mapping Methods 0.000 claims description 8
- 238000000605 extraction Methods 0.000 claims description 7
- 238000012549 training Methods 0.000 claims description 7
- 238000004590 computer program Methods 0.000 claims description 3
- 238000003860 storage Methods 0.000 claims description 3
- 230000002776 aggregation Effects 0.000 claims description 2
- 238000004220 aggregation Methods 0.000 claims description 2
- 230000008569 process Effects 0.000 abstract description 5
- 239000002131 composite material Substances 0.000 description 8
- 230000006870 function Effects 0.000 description 8
- 238000010586 diagram Methods 0.000 description 6
- 238000005457 optimization Methods 0.000 description 6
- 230000010354 integration Effects 0.000 description 4
- 230000007246 mechanism Effects 0.000 description 4
- 238000010801 machine learning Methods 0.000 description 3
- 238000010606 normalization Methods 0.000 description 3
- 238000007637 random forest analysis Methods 0.000 description 3
- 230000008901 benefit Effects 0.000 description 2
- 238000004422 calculation algorithm Methods 0.000 description 2
- 230000004913 activation Effects 0.000 description 1
- 239000000654 additive Substances 0.000 description 1
- 230000000996 additive effect Effects 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 238000013528 artificial neural network Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000002860 competitive effect Effects 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 238000012417 linear regression Methods 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 238000003062 neural network model Methods 0.000 description 1
- 238000005192 partition Methods 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 230000001105 regulatory effect Effects 0.000 description 1
- 230000011218 segmentation Effects 0.000 description 1
- 230000001960 triggered effect Effects 0.000 description 1
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
- G06Q10/00—Administration; Management
- G06Q10/06—Resources, workflows, human or project management; Enterprise or organisation planning; Enterprise or organisation modelling
- G06Q10/063—Operations research, analysis or management
- G06Q10/0639—Performance analysis of employees; Performance analysis of enterprise or organisation operations
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
- G06F18/253—Fusion techniques of extracted features
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N5/00—Computing arrangements using knowledge-based models
- G06N5/01—Dynamic search techniques; Heuristics; Dynamic trees; Branch-and-bound
-
- 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
- G06Q10/00—Administration; Management
- G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
-
- 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
- G06Q40/00—Finance; Insurance; Tax strategies; Processing of corporate or income taxes
- G06Q40/04—Trading; Exchange, e.g. stocks, commodities, derivatives or currency exchange
Landscapes
- Engineering & Computer Science (AREA)
- Business, Economics & Management (AREA)
- Theoretical Computer Science (AREA)
- Human Resources & Organizations (AREA)
- Physics & Mathematics (AREA)
- Economics (AREA)
- Strategic Management (AREA)
- General Physics & Mathematics (AREA)
- Development Economics (AREA)
- Entrepreneurship & Innovation (AREA)
- Marketing (AREA)
- General Business, Economics & Management (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Game Theory and Decision Science (AREA)
- Finance (AREA)
- Accounting & Taxation (AREA)
- Tourism & Hospitality (AREA)
- General Engineering & Computer Science (AREA)
- Educational Administration (AREA)
- Quality & Reliability (AREA)
- Operations Research (AREA)
- Artificial Intelligence (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Computing Systems (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Computational Linguistics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Software Systems (AREA)
- Mathematical Physics (AREA)
- Technology Law (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a financial transaction risk assessment method, which comprises the following steps: extracting cross features from the original feature vectors of the transaction according to a pre-trained integrated tree model; fusing embedded vectors corresponding to all nodes on the decision path to obtain a path fusion vector; setting attention weight, and aggregating the path fusion vector by using the attention weight to obtain expression characteristics of each transaction; and obtaining a prediction result according to the expression characteristics. The invention also discloses a financial transaction risk assessment device, electronic equipment and a computer readable medium. The prediction process of the invention is transparent, has strong interpretability and high prediction accuracy.
Description
Technical Field
The invention belongs to the technical field of machine learning, and particularly relates to a financial transaction risk assessment method and device, electronic equipment and a computer readable medium.
Background
In recent years, tree-based integrated models have shown competitive performance in financial transaction risk assessment, particularly in prediction accuracy and model generalization capability. However, in contrast to decision trees, the integrated tree model is considered opaque and does not provide enough interpretable information for decisions. Although the accuracy of the risk assessment model is an important evaluation criterion, understandability and portability are also not negligible. According to bank regulatory provisions, financial institutions in some countries are obligated to provide understandable reasons when a credit application is denied. It follows that an opaque model cannot be trusted by decision makers in the financial field and is subject to legal restrictions.
Financial transaction risks, on the other hand, are often triggered by a variety of factors, and a widely used solution in the industry is to craft composite features and then enter them into an interpretable method that can learn the importance of each composite feature. However, this method has problems of difficulty in expansion, high labor costs, and the like.
Disclosure of Invention
Aiming at the defects in the prior art, the invention provides a financial transaction risk assessment method, a device, electronic equipment and a computer readable medium, wherein the method is transparent in prediction process, strong in interpretability and high in prediction accuracy.
In a first aspect, a method for assessing risk of financial transactions includes the steps of:
extracting cross features from the original feature vectors of the transaction according to a pre-trained integrated tree model;
fusing embedded vectors corresponding to all nodes on the decision path to obtain a path fusion vector;
setting attention weight, and aggregating the path fusion vector by using the attention weight to obtain expression characteristics of each transaction;
and obtaining a prediction result according to the expression characteristics.
As an optimization, the method for extracting cross features from the original feature vectors of the transactions according to the pre-trained ensemble tree model comprises the following steps:
Pre-training an ensemble tree model;
inputting the original feature vector of the transaction into the integrated tree model to obtain local cross features;
and forming cross features according to all local cross features.
As optimization, the method for fusing the embedded vectors corresponding to all nodes on the decision path to obtain a path fusion vector includes the following steps:
mapping each node in the cross feature into an embedded vector;
forming an embedded matrix according to the embedded vectors corresponding to all nodes on the decision path and fusing the embedded matrix into a path fusion vector;
and constructing a path fusion matrix through the path fusion vector and the cross features.
As optimization, the setting of attention weight and the aggregation of the path fusion vector by using the attention weight to obtain the expression feature of each transaction includes the following steps:
constructing an attention weight function based on the original feature vector and the path fusion vector;
normalizing the attention weight;
and aggregating the path fusion vectors according to the attention weight to obtain the expression characteristics of each transaction.
As an optimization, the obtaining of the prediction result according to the expression features comprises:
the expression features are mapped to the prediction by applying a layer of linearity.
As optimization, the embedding matrix obtains a path fusion vector through an addition method.
As optimization, the expression features are obtained by means of weighted sum of attention weight and path fusion vector; and/or
The integrated tree model adopts a gradient lifting decision tree model.
In a second aspect, a financial transaction risk assessment apparatus includes:
the cross feature extraction module is used for extracting cross features from the original feature vectors of the transaction according to the pre-trained integrated tree model;
the decision path fusion module is used for fusing the embedded vectors corresponding to all the nodes on the decision path to obtain a path fusion vector;
the attention fusion module is used for setting attention weight and aggregating the path fusion vector by using the attention weight to obtain expression characteristics of each transaction;
and the prediction module is used for obtaining a prediction result according to the expression characteristics.
In a third aspect, an electronic device includes:
one or more processors;
a storage device having one or more programs stored thereon;
when executed by the one or more processors, cause the one or more processors to implement a method of risk assessment of financial transactions as any one of the preceding.
In a fourth aspect, a computer readable medium has a computer program stored thereon, wherein the program when executed by a processor implements a method for risk assessment of financial transactions as any one of the previous methods.
Compared with the prior art, the invention has the following beneficial effects:
1. original features are converted into cross features through the integration tree model to be used by a downstream prediction task, the extraction capability of the integration tree model on important features is reserved, and a plurality of composite features (cross features) are automatically constructed;
2. each node of the decision tree is mapped with an embedded vector, and the embedded vectors of all the nodes on a decision path are fused, so that the fused vector can keep the structural information of the decision tree on one hand, and the fused features and the local cross features extracted by the integrated tree have equivalent expression capability as far as possible; on the other hand, high-level semantic information of local cross features can be captured, so that the proposed model can obtain better performance than an integrated tree;
3. local cross features, which are more concerned in the prediction process, of each sample are learned through an attention network, and the intrinsic interpretability of the model is given, namely, the prediction result is mainly determined based on which composite features.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the embodiments or the prior art descriptions will be briefly described below, it is obvious that the drawings in the following description are only some embodiments of the present invention, and other drawings can be obtained by those skilled in the art without creative efforts.
FIG. 1 is a schematic flow chart of a method for assessing risk of financial transactions according to the present invention;
FIG. 2 is a schematic diagram of an overall framework of a financial transaction risk assessment method according to the present invention;
FIG. 3 is a diagram illustrating an example of a gradient-boosting decision tree model of a financial transaction risk assessment method according to the present invention;
FIG. 4 is a flowchart illustrating a step S1 of the method for assessing risk of financial transaction according to the present invention;
FIG. 5 is a flowchart illustrating a step S2 of the method for assessing risk of financial transaction according to the present invention;
FIG. 6 is a flowchart illustrating a step S3 of the method for assessing risk of financial transaction according to the present invention;
FIG. 7 is a schematic view of an attention network of a financial transaction risk assessment method according to the present invention;
FIG. 8 is a schematic diagram of a financial transaction risk assessment apparatus according to the present invention;
FIG. 9 is a schematic diagram of a cross feature extraction module of the financial transaction risk assessment apparatus according to the present invention;
FIG. 10 is a schematic diagram illustrating a structure of a decision path fusion module of the financial transaction risk assessment apparatus according to the present invention;
fig. 11 is a schematic structural diagram of an attention fusion module of the financial transaction risk assessment apparatus according to the present invention.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be obtained by a person skilled in the art without inventive step based on the embodiments of the present invention, are within the scope of protection of the present invention.
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and do not limit the invention.
In a first aspect, a method for assessing risk of financial transaction, as shown in fig. 1 and fig. 2, includes the following steps:
S1: cross features are extracted from the raw feature vectors of the transaction according to a pre-trained ensemble tree model.
In this step, the Ensemble tree model (DT-based Ensemble) is an Ensemble model using a decision tree as a base classifier, and is mainly divided into two types, i.e., serial Ensemble and parallel Ensemble, in which the former is represented by a gradient-enhanced decision tree (GBDT) and the latter is represented by a Random Forest (RF), which are algorithms commonly used in the field of machine learning. The integrated tree model in the step can adopt models such as a random forest, XGBoost, LightGBM and the like besides a gradient lifting decision tree model.
The cross feature (cross feature) refers to an integral cross feature formed by splicing local cross features of each decision tree in the integrated tree model; for an integrated tree model, the original features correspond to one cross feature, but one cross feature may correspond to multiple original features.
The local cross feature (localcrosfeature) refers to a composite feature formed by combining partial feature intervals in a decision path. In the aspect of semantics, the local cross feature, the leaf node and the decision path are equivalent.
In this embodiment, a gradient boosting decision tree model is taken as an example for explanation. The gradient boost decision tree model comprises a plurality of decision trees, wherein each leaf node in each decision tree corresponds to one leaf node A decision path is formed, and each internal node on the path uses two decision edges to segment a certain feature in the original features. Unlike neural network-based feature expression learning, the tree-based model does not learn embedded features to predict; instead, they predict by learning decision rules from the data. The advantage of this approach is its effectiveness and interpretability. Specifically, assume that there is a decision tree T ═ V, E, where V is the set of nodes and E is the set of edges. V can be divided into three subsets, each being a root node V of the treeRInternal node VIAnd leaf node VLI.e. V ═ VR}∪VI∪VL. Each internal node vi∈VIFeature x segmentation using two decision edgesi. For numerical features (e.g., revenue), the node selects a threshold aje.R and partition the feature into xi<aj[ and [ x ]i≥aj](ii) a For binary features (class features are converted to binary features by one-hot coding), the node makes a decision based on whether the feature is equal to a value, i.e., a decision edge such as [ x ]i=aj]And [ x ]i≠aj]。
Connection vRAnd any leaf node vl∈VLThe path of (a) represents a decision rule or decision path and can also be regarded as a local intersection feature. Local intersection features combine multiple feature ranges together, as shown by leaf node v in FIG. 3 7Represents [ x ]0≥a0]∧[x2≥a0]∧[x4≥a3]. Given a feature vector x, the decision tree model determines on which leaf node x will fall, and can be thought of as mapping the original feature vector onto the leaf nodes according to the tree structure. Under this prediction mechanism, the decision paths corresponding to leaf nodes can be used as the most prominent cross features in tree model prediction, and therefore, the model based on the decision tree is self-interpretable in nature. The mode of generating the cross features through the tree model avoids labor-intensive feature engineering, and effectively solves the problems of difficult expansion and lack of universality in the composite characteristics of manual manufacturing.
S2: and fusing the embedded vectors corresponding to all nodes on the decision path to obtain a path fusion vector.
In this step, for the cross feature obtained in step S1, each node therein is mapped to a learnable dense embedded vector. Considering that the decision paths of these nodes partially overlap, the inventors believe that the local intersection features represented by neighboring leaf nodes should have similar embedded vectors. As shown in fig. 2, since each leaf node corresponds to a decision path, the local intersection feature can be mapped to an embedding matrix composed of embedding vectors corresponding to all nodes on the path; and as the lengths of all decision paths are not the same and the difference brings trouble to the training of subsequent models, the path embedding matrix is fused into an embedded vector in the step, which is called as a path fusion vector.
S3: and setting attention weights, and aggregating the path fusion vectors by using the attention weights to obtain the expression features r of each transaction.
In this step, the attention weight may adopt a function based on the original feature vector and the path fusion vector, and the attention weight reflects which local cross feature is more concerned when predicting the feature vector x.
S4: and obtaining a prediction result according to the expression characteristic r.
In this step, after the expression feature r is obtained, the expression feature r may be mapped into a final prediction by applying a layer of linear layer, such as:
wherein a ∈ RdIs the weight of the final linear regression layer, and b is the offset. It can be seen that this embodiment is a shallow additive model. The contribution of each component can be easily evaluated for the purpose of interpreting the prediction.
Taking binary classification task as an example, the corresponding loss function is a cross entropy loss function:
in this embodiment, composite features are automatically constructed by employing a tree integration model, high-level semantic information of the composite features is captured by learnable embedded vectors, and the inherent interpretability of the model is provided using an attention network. In an accurate aspect, the invention can have better performance than the current and traditional machine learning algorithms for credit scoring; in interpretable aspects, the present invention is transparent in the process of generating the prediction results, enabling the positioning of the key decision rules upon which the prediction is based.
As a preferred embodiment, as shown in fig. 2 and 4, the step S1 includes the following steps:
s101: and pre-training the ensemble tree model.
S102: and inputting the original feature vector of the transaction into the integrated tree model to obtain local cross features.
In this step, the integration tree model takes a gradient lifting decision tree model as an example, and the gradient lifting decision tree is taken as a set of decision trees, some of which areWherein the t-th decision tree maps an original feature vector x to a leaf node Qt(x) Each leaf node corresponds to a prediction weight. Regarding each activated leaf node as a local cross feature, using a unique heat vectorIs represented by wherein LtRepresenting the number of leaves of the t-th decision tree.
S103: and forming the cross feature according to all the local cross features.
In this step, if N is usedL=∑τLtRepresenting the total number of leaf nodes in the forest, wherein tau is the number of decision trees; then each cross feature can use one multiple heat vectorRepresents:where q is a sparse vector with 1's elements representing activated leaf nodes of each decision tree and 0's elements representing all unactivated leaf nodes in the forest. Unlike the traditional gradient boosting decision tree model which sums the weights of all activated leaf nodes (i.e., local cross features) as the final prediction, the selection in this step is to keep all activated leaf nodes as the overall cross features.
As a preferred embodiment, as shown in fig. 2 and 5, the step S2 includes the following steps:
s201: and mapping each node in the cross features into an embedded vector.
In this step, assuming that the forest has N nodes except the root node of the tree, each node is divided into N nodesAre all mapped into a learnable embedded vector ej∈Rd. WhereinRefers to a set of nodes of each decision tree from which root nodes are removed.
S202: and forming an embedded matrix according to the embedded vectors corresponding to all the nodes on the decision path and fusing the embedded matrix into a path fusion vector.
In this step, since each leaf node corresponds to one decision path, the local intersection feature f can be usediMapping an embedding matrix P consisting of corresponding embedding vectors of all nodes on a pathi∈R|P(i)|×d:
Pi=[e1,…,ej],ej∈P(i);
Wherein P (i) represents a leaf node viAll nodes on the decision path, except the root node, including the leaf nodes themselves, correspond to a set of embedded vectors.
Since not all decision path lengths | p (i) | are the same, and this difference can cause trouble in training of subsequent models, the path embedding matrix is merged into one embedding vector, which is called a path merging vector. In the step, a simple addition method is adopted to obtain a path fusion vector p i∈Rd:
pi=∑ej,ej∈P(i);
By the method, the path fusion vectors corresponding to the adjacent leaf nodes all contain node information of the overlapped part on the decision path, so that the structural information of the model is introduced into the path fusion vectors.
S203: and constructing a path fusion matrix through the path fusion vector and the cross features.
In the step, a path fusion matrix E is constructed through the path fusion vector and the cross feature qp∈Rτ×d:
For a sample, the corresponding path fusion matrix is composed of the path fusion vectors corresponding to the decision paths activated by the decision trees.
As a preferred embodiment, as shown in fig. 2 and fig. 6, the step S3 includes the following steps:
s301: constructing an attention weight function based on the original feature vector and the path fusion vector.
In this step, in order to solve the generalization and scalability problems, the attention weight is considered to be modeled as a fusion vector p based on the original feature vector x and the pathtRather than learning freely from the data. In particular, a multi-layered perceptron may be employed to parameterize attention weights, referred to as an attention network, as shown in fig. 7, which is defined as:
wherein p istThe path fusion vector is output by the t-th tree in the forest; Respectively representing a weight matrix and a bias vector of the hidden layer; d is a radical ofhIs the dimension of the implicit vector; d is a radical ofxIs the dimension of the original feature vector;mapping the hidden layer output to attention weight; σ denotes the ReLU activation function.
S302: the attention weight is normalized.
In this step, the normalization process may use a softmax function.
S303: and aggregating the path fusion vectors according to the attention weight to obtain the expression characteristic r of each transaction.
In this step, the inventor considers that if only each path fusion vector in the path fusion matrix is multiplied by the corresponding weight, the attention weight will not be valid correctly, because the meaning of all 0 vectors may be different for different path fusion vectors. Therefore, in this step, the attention weight is used to aggregate the path fusion vectors to obtain the expression features r (x, E) of each transactionP)∈Rd:
The path fusion vector with the smaller attention weight only has a smaller influence on the finally generated expression feature, so that the expression feature obtained by the weighted sum mode can keep the influence of the attention weight.
In this embodiment, although the way of mapping cross features into embedded matrices can learn the high-level semantic information and potential relevance of cross features, for different original feature vectors, if the integrated tree model assigns them the same cross features, they have the same path fusion matrix, which limits modeling fidelity. This is because in real applications, transactions with the same cross-signature may also have different risks.
In theory, in this step, the path fusion vectors may be directly summed to obtain the expression features without using the attention mechanism. However, in order to improve the expression ability of the expression features as much as possible, the inventor considers that it is important to assign different scores to the local cross features for different original feature vectors, so the preferred embodiment adopts the weight of the personalized local cross features instead of using a simple summation mechanism. The invention captures the different importance of the local cross features in predicting different samples by assigning an attention weight to the embedded features of each local cross feature, and knows how each local cross feature contributes to the prediction of the current sample.
Although the shallow neural network model without the fully-connected hidden layer is adopted in the embodiment, the embedded and attention mechanism has strong expression capability and ensures the effectiveness of the model.
In a second aspect, a financial transaction risk assessment apparatus, as shown in fig. 8, includes:
the cross feature extraction module is used for extracting cross features from the original feature vectors of the transaction according to the pre-trained integrated tree model;
The decision path fusion module is used for fusing the embedded vectors corresponding to all nodes on the decision path to obtain a path fusion vector;
the attention fusion module is used for setting attention weight and aggregating the path fusion vector by using the attention weight to obtain expression characteristics of each transaction;
and the prediction module is used for obtaining a prediction result according to the expression characteristics.
In this embodiment, the structure of the financial transaction risk assessment apparatus corresponds to S1 to S4 of the financial transaction risk assessment method of the first aspect.
As a preferred embodiment, the cross feature extraction module, as shown in fig. 9, includes:
the pre-training module is used for pre-training the integrated tree model;
the local cross feature acquisition module is used for inputting the original feature vector of the transaction into the integrated tree model to obtain local cross features;
and the cross feature acquisition module is used for forming cross features according to all the local cross features.
The structure of the cross feature extraction module in this embodiment corresponds to S101 to S103 of the financial transaction risk assessment method of the first aspect.
As a preferred embodiment, the decision path fusion module, as shown in fig. 10, includes:
An embedded vector module for mapping each node in the cross feature into an embedded vector;
the embedded matrix fusion module is used for forming an embedded matrix according to the embedded vectors corresponding to all the nodes on the decision path and fusing the embedded matrix into a path fusion vector;
and the path fusion matrix constructing module is used for constructing a path fusion matrix through the path fusion vector and the cross feature.
The structure of the decision path fusion module in this embodiment corresponds to S201 to S203 of the financial transaction risk assessment method of the first aspect.
As a preferred embodiment, the attention fusion module, as shown in fig. 11, includes:
an attention weight setting module for constructing an attention weight function based on the original feature vector and the path fusion vector;
the normalization module is used for carrying out normalization processing on the attention weight;
and the expression feature acquisition module is used for aggregating the path fusion vector according to the attention weight so as to obtain the expression feature of each transaction.
The structure of the attention fusion module in this embodiment corresponds to S301 to S303 of the financial transaction risk assessment method of the first aspect.
In a third aspect, an electronic device comprises:
One or more processors;
a storage device having one or more programs stored thereon;
when executed by the one or more processors, cause the one or more processors to implement a method of risk assessment of a financial transaction as in any one of the first aspects.
In a fourth aspect, a computer readable medium has a computer program stored thereon, wherein the program when executed by a processor implements a method for risk assessment of a financial transaction as in any one of the first aspects.
All possible combinations of the technical features in the above embodiments may not be described for the sake of brevity, but should be considered as being within the scope of the present disclosure as long as there is no contradiction between the combinations of the technical features.
Claims (10)
1. A financial transaction risk assessment method is characterized by comprising the following steps:
extracting cross features from the original feature vectors of the transaction according to a pre-trained integrated tree model;
fusing embedded vectors corresponding to all nodes on the decision path to obtain a path fusion vector;
setting attention weight, and aggregating the path fusion vector by using the attention weight to obtain expression characteristics of each transaction;
And obtaining a prediction result according to the expression characteristics.
2. The method as claimed in claim 1, wherein the step of extracting cross features from the original feature vectors of the transaction according to the pre-trained ensemble tree model comprises the steps of:
pre-training an ensemble tree model;
inputting the original feature vector of the transaction into the integrated tree model to obtain local cross features;
and forming cross features according to all local cross features.
3. The method for assessing risk of financial transaction according to claim 1, wherein said fusing the embedded vectors corresponding to all nodes on the decision path to obtain a path fusion vector comprises the following steps:
mapping each node in the cross feature into an embedded vector;
forming an embedded matrix according to the embedded vectors corresponding to all the nodes on the decision path and fusing the embedded matrix into a path fusion vector;
and constructing a path fusion matrix through the path fusion vector and the cross features.
4. The financial transaction risk assessment method according to claim 1, wherein the setting of attention weight and the aggregation of the path fusion vector by using the attention weight to obtain the expression feature of each transaction comprises the following steps:
Constructing an attention weight function based on the original feature vector and the path fusion vector;
normalizing the attention weight;
and aggregating the path fusion vectors according to the attention weight to obtain the expression characteristics of each transaction.
5. The method as claimed in claim 1, wherein the obtaining of the predicted result according to the expression feature comprises:
the expression features are mapped to the predicted result by applying a layer of linearity.
6. The method of claim 3, wherein the method further comprises:
and the embedded matrix obtains a path fusion vector through an addition method.
7. The method of claim 1, wherein the method further comprises:
the expression features are obtained by means of weighted sum of attention weight and path fusion vector; and/or
The integrated tree model adopts a gradient lifting decision tree model.
8. A financial transaction risk assessment apparatus, comprising:
the cross feature extraction module is used for extracting cross features from the original feature vectors of the transactions according to the pre-trained integrated tree model;
the decision path fusion module is used for fusing the embedded vectors corresponding to all the nodes on the decision path to obtain a path fusion vector;
The attention fusion module is used for setting attention weight and aggregating the path fusion vector by using the attention weight to obtain expression characteristics of each transaction;
and the prediction module is used for obtaining a prediction result according to the expression characteristics.
9. An electronic device, comprising:
one or more processors;
a storage device having one or more programs stored thereon;
the one or more programs, when executed by the one or more processors, cause the one or more processors to implement the method recited in any of claims 1-7.
10. A computer-readable medium, on which a computer program is stored which, when being executed by a processor, carries out the method according to any one of claims 1-7.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210537782.7A CN114757581A (en) | 2022-05-18 | 2022-05-18 | Financial transaction risk assessment method and device, electronic equipment and computer readable medium |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210537782.7A CN114757581A (en) | 2022-05-18 | 2022-05-18 | Financial transaction risk assessment method and device, electronic equipment and computer readable medium |
Publications (1)
Publication Number | Publication Date |
---|---|
CN114757581A true CN114757581A (en) | 2022-07-15 |
Family
ID=82335983
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202210537782.7A Pending CN114757581A (en) | 2022-05-18 | 2022-05-18 | Financial transaction risk assessment method and device, electronic equipment and computer readable medium |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN114757581A (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN115641201A (en) * | 2022-09-27 | 2023-01-24 | 厦门国际银行股份有限公司 | Data anomaly detection method, system, terminal device and storage medium |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112270547A (en) * | 2020-10-27 | 2021-01-26 | 上海淇馥信息技术有限公司 | Financial risk assessment method and device based on feature construction and electronic equipment |
CN114298417A (en) * | 2021-12-29 | 2022-04-08 | ***股份有限公司 | Anti-fraud risk assessment method, anti-fraud risk training method, anti-fraud risk assessment device, anti-fraud risk training device and readable storage medium |
-
2022
- 2022-05-18 CN CN202210537782.7A patent/CN114757581A/en active Pending
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112270547A (en) * | 2020-10-27 | 2021-01-26 | 上海淇馥信息技术有限公司 | Financial risk assessment method and device based on feature construction and electronic equipment |
CN114298417A (en) * | 2021-12-29 | 2022-04-08 | ***股份有限公司 | Anti-fraud risk assessment method, anti-fraud risk training method, anti-fraud risk assessment device, anti-fraud risk training device and readable storage medium |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN115641201A (en) * | 2022-09-27 | 2023-01-24 | 厦门国际银行股份有限公司 | Data anomaly detection method, system, terminal device and storage medium |
CN115641201B (en) * | 2022-09-27 | 2023-11-07 | 厦门国际银行股份有限公司 | Data anomaly detection method, system, terminal equipment and storage medium |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN112529168B (en) | GCN-based attribute multilayer network representation learning method | |
Zhang et al. | A high-order possibilistic $ C $-means algorithm for clustering incomplete multimedia data | |
Lu et al. | GLR: A graph-based latent representation model for successive POI recommendation | |
CN113220886A (en) | Text classification method, text classification model training method and related equipment | |
CN111931505A (en) | Cross-language entity alignment method based on subgraph embedding | |
CN113449204B (en) | Social event classification method and device based on local aggregation graph attention network | |
Hao et al. | An unsupervised detection method for shilling attacks based on deep learning and community detection | |
CN112257959A (en) | User risk prediction method and device, electronic equipment and storage medium | |
CN112287020A (en) | Big data mining method based on graph analysis | |
CN113887213A (en) | Event detection method and device based on multilayer graph attention network | |
CN115858919A (en) | Learning resource recommendation method and system based on project field knowledge and user comments | |
CN111985207A (en) | Method and device for acquiring access control policy and electronic equipment | |
CN114757581A (en) | Financial transaction risk assessment method and device, electronic equipment and computer readable medium | |
Jimenez et al. | An empirical study on identifying sentences with salient factual statements | |
CN117009613A (en) | Picture data classification method, system, device and medium | |
CN116522232A (en) | Document classification method, device, equipment and storage medium | |
CN108304568B (en) | Real estate public expectation big data processing method and system | |
Bu et al. | A High‐Order CFS Algorithm for Clustering Big Data | |
CN114491296A (en) | Proposal affiliate recommendation method, system, computer device and readable storage medium | |
CN117217929A (en) | Registered object risk identification method, device, computer equipment and storage medium | |
CN114579761A (en) | Information security knowledge entity relation connection prediction method, system and medium | |
Patel | Detection of Maliciously Authored News Articles | |
CN114676755A (en) | Non-supervision domain self-adaptive classification method based on graph convolution network | |
Shukla et al. | Sentiment analysis of international relations with artificial intelligence | |
Li et al. | SLBCNN: a improved deep learning model for few-shot charge prediction |
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 | ||
RJ01 | Rejection of invention patent application after publication | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20220715 |