An interpretable method for detecting false invoicing based on mesoscience
Through the complex system modeling and symbolic control mechanism of mediating science, the invoice fictitious data system is divided into the mediating area, and an interpretable invoice fictitious detection model is built, which solves the problem of uninterpretation of existing methods and realizes credible and accurate invoice fictitious detection.
Patent Information
- Application Number
- CN202211311760.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-10-25
- Publication Date
- 2025-07-25
- Estimated Expiration
- 2042-10-25
AI Technical Summary
The existing invoice false invoice detection methods lack interpretability, which leads to many doubts about the test results and is difficult to verify, which affects efficiency and credibility.
Based on the complex system analysis method of mediation science, the invoice fictitious data system is divided into several mediation areas, and a symbolic control mechanism embedded in specific tax knowledge is built. Data inference is carried out through self-attention mechanism and symbolic differentiable control mechanism to generate an interpretable link of evidence.
It provides credible and accurate invoice false detection results, improves the interpretability of the detection and the reliability of the evidence link, and solves the problem of uninterpretation of the existing methods.
Smart Images

Figure CN115496571B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical fields of artificial intelligence and tax control technology, and particularly relates to an interpretable invoice fraud detection method based on mesoscience. Background Art
[0002] Invoice fraud refers to a fraudulent act of issuing invoices dishonestly. Taxpaying units and individuals, in order to achieve the purpose of tax evasion, resort to fraud in commodity names, quantity names, commodity unit prices, and amounts, and even fabricate transaction matters and issue invoices fraudulently by taking advantage of familiar relationships. Invoice fraud has always been a cancer and chronic problem in tax governance, causing an average annual loss of about 1.1 trillion yuan in huge tax revenues. Therefore, invoice fraud detection is crucial for tax risk management. Traditionally, since the invoices themselves are real, it is impossible to detect invoice fraud only by verifying the authenticity of invoices. One can only rely on the experience of tax inspectors to analyze and judge whether the invoices are involved in fraud and whether the enterprises have abnormal behaviors related to invoice fraud based on limited invoice information. However, with the continuous growth of the scale of invoice data, it is no longer feasible to rely solely on manual inspection of invoice fraud, and new technical means are needed to quickly, accurately, and reasonably determine whether an invoice is involved in fraud. In recent years, deep learning technology has made amazing progress in various fields. In particular, in the tax control field, some researchers have also used deep learning technology to detect whether enterprises or individuals evade taxes or whether there is invoice fraud.
[0003] The following patents and literatures provide solutions aimed at automatically identifying invoice fraud through machine learning or deep learning technologies:
[0004] Literature 1: An Invoice Fraud Identification Method and System Based on Positive Examples and Unlabeled Learning (CN201910636175.4);
[0005] Literature 2: An Enterprise Invoice Fraud Detection Method Based on a Graph Attention Network with Enhanced Dynamic Edge Features (CN202010507242.5);
[0006] Literature 3: An Invoice Fraud Identification Method Based on Dynamic Network Representation (CN201911066791.7);
[0007] Document 1 proposed an invoice recognition method and system based on positive examples and unlabeled learning. The system is divided into an enterprise feature extraction module, a cyclic multi-spy negative example marking module, a construction module based on k-nearest neighbor regression co-training, and an invoice fraud prediction module. First, the basic information of taxpayers is processed and encoded for features; secondly, the basic features and network features are combined as the feature space, and a binary classifier is trained based on the proposed cyclic multi-spy negative example marking method. The final negative sample set is obtained by taking the intersection of all preliminary negative sample sets through the binary classifier; then, the reliable negative example samples and positive samples mined are used as the training set to construct an invoice fraud prediction model based on the k-nearest neighbor regression co-training algorithm; finally, the features of the unlabeled enterprise samples are input into the trained invoice fraud prediction model to determine whether the enterprise has invoice fraud behavior.
[0008] Document 2 proposed an enterprise invoice fraud detection method based on a graph attention network with dynamic edge enhancement. First, a dynamic enterprise transaction network is constructed based on tax-related data, and node features and edge features are extracted; secondly, the graph attention network is used to extract the transaction network features for each time period; thirdly, the features extracted in the previous step are used with a long short-term memory network to obtain time series features; then, a deep neural network is used to construct an invoice fraud model, and the parameters of the model are adjusted by training the invoice fraud detection model; finally, the trained model is used to detect invoice fraud for target enterprises.
[0009] Document 3 proposed an invoice fraud recognition method based on dynamic network representation. First, enterprises are used as nodes and transaction records as edges to organize enterprise information into a static network; secondly, a representation of the enterprise transaction network is established with each day as a time node, a time window of length 30 is established, and a dynamic network representation result is obtained by moving the time window; thirdly, by referring to the distributed optimization algorithm, the objective function of the representation is decomposed into independent sub-functions to improve the learning efficiency of the model; finally, a classifier is constructed based on LightBGM to identify enterprises with invoice fraud.
[0010] The above methods based on machine learning or deep learning can solve the problem that traditional manual inspection of invoice fraud is time-consuming and laborious, but there are still some problems. The method in Document 1 is based on the enterprise static relationship network and cannot dynamically analyze the invoice fraud risk of taxpaying enterprises by combining historical information. The dynamic edge-enhanced graph attention network model mentioned in Document 2 and the dynamic network representation obtained by moving the time series window in Document 3 both take into account the transaction network characteristics in different time periods, which can be regarded as an improvement to the method in Document 3. However, there is an inherent defect in the methods for invoice fraud detection mentioned in these documents: whether it is the graph attention network mentioned in Document 2, the multi-layer perceptron constructed in Document 3, or the classifiers trained in the feature space in Documents 1 and 3, these algorithms lack the interpretability of the detection process and results. This will lead to many doubts about the invoice fraud detection results and great difficulty in verification, and a large amount of manpower and material resources are still needed for inspection and verification, thus affecting the efficiency and coverage of invoice fraud detection, as well as the credibility, public trust and execution of the results. The reason is that the above invoice fraud detection methods have the following characteristics: 1. Indecomposability: When using a deep learning model to model the invoice fraud problem, the focus is mainly on the input and output ends of the "black box". Although the deep model has a so-called hierarchical structure, there is no clear and distinct functional definition for these layers, and the association between layers cannot be explained. 2. Non-interference: During the training process of the model, the model parameters are generally obtained through the random gradient descent method with multiple iterations, and the process is difficult to intervene. In the model inference process, there are a large number of calculations involved, and it is difficult for the human brain to track and understand the meaning of these operations, so it is impossible to interfere with the model inference process. Therefore, it is urgent to solve the "interpretability" problem of the invoice fraud detection algorithm and provide a credible and accurate "evidence chain". Summary of the Invention
[0011] The purpose of the present invention is to provide an interpretable invoice fraud detection method based on mesoscience. This method is based on the complex system analysis method of mesoscience theory, divides the multi-level and multi-scale invoice fraud data system into several meso-regions; for the functions and characteristics of each meso-region, constructs a symbolic control mechanism embedded with specific tax knowledge to achieve knowledge-driven data calculation and reasoning. Creatively combines the complex system modeling theory of mesoscience and the understandable and editable characteristics of the symbolic system, and proposes a symbolic hierarchical and hierarchical interpretable invoice fraud detection model SHiL, aiming to provide a credible and accurate invoice fraud "evidence chain" and solve the interpretability problem of the existing invoice fraud detection methods.
[0012] The present invention is implemented by adopting the following technical solutions:
[0013] An interpretable invoice fraud detection method based on mesoscience, comprising the following steps:
[0014] S1, Tax transaction dynamic graph embedding based on self-attention mechanism; First, transform tax transaction data into a tax dynamic network through steps such as feature processing and feature embedding. The tax dynamic network preserves transaction information and other structural information between enterprises at different time slices;
[0015] S2, Determination of the meso-region in the invoice fraud data system based on mesoscience; According to the mesoscience complex data system modeling theory, in the tax dynamic transaction network, "invoice fraud" is the dominant factor in invoice fraud behavior. Therefore, based on the dominant factor "invoice flow direction", divide the data system of the tax network into different meso-regions;
[0016] S3, Construct a symbolic differentiable programming control mechanism for the divided meso-regions; On the basis of the divided meso-regions, construct an artificially participable control mechanism through two steps: symbolic representation and construction of the control mechanism;
[0017] S4, By coupling the calculation and reasoning of the control mechanism, detect whether there is invoice fraud behavior in the enterprise, and at the same time generate a cross-regional reasoning path to provide an interpretable evidence chain for the reasoning result. Optimally select the reasoning paths formed by coupling the reasoning steps of each meso-region based on the multi-objective evolutionary algorithm based on Chebyshev aggregation, and finally select an optimal solution to provide evidence for the reasoning result.
[0018] A further improvement of the present invention is that the specific implementation method of step S1 is:
[0019] First, transform tax transaction data into a tax dynamic network for downstream data reasoning. The specific implementation steps are as follows:
[0020] Step1: Feature preprocessing; For text features, use the BERT training language model to represent them as word vectors; For categorical attribute values, adopt the One-Hot encoding method to represent them as vectors; For numerical attribute values, adopt the standardization method for processing; Finally, splice the above features to obtain all the features of the enterprise. After processing the above features for each enterprise, obtain the feature matrix X of the enterprise;
[0021] Step2: Define a tax dynamic graph G = {G 1 , G 2 ,..., G T ,}, which contains static graphs at T moments. G t = {V, E t , X} is the static graph at the t-th moment; All static graphs share a node set, that is, the set composed of all enterprises, but each static graph has its own edge set E t , X is the feature matrix constructed in step Step1. Since the edge sets are different, each static graph has its own adjacency matrix A tTo represent the transaction relationships between enterprises, such as indicating that at the t-th moment, there is no transaction between the i-th enterprise and the j-th enterprise;
[0022] Step3: Dynamic graph embedding based on the self-attention mechanism; for the static graph G at any moment T ={V, E t , X}, for the node v ∈ V, its neighbor node set is denoted as Calculate the attention value of v to each neighbor u: The calculation formula is: After calculating the attention of node v to all its neighbors, normalize the attention: Weightedly aggregate the vectors of node v's neighbors according to the attention to obtain the output vector z of node v v , and the calculation formula is: At this time, this aggregation preserves the local structural information of node v; to obtain the dynamic embedding, calculate the new representation of node v at each moment based on the self-attention mechanism, and this new representation should preserve the dynamic information of the node changing over time, denoted as The specific calculation process is as follows: 1) Calculate the attention values for all moments at any moment. First, map the features of node v at all moments to the Query space: X v W q , the Key space: X v W k ; the attention value between moment i and moment j is: where is the scaling factor and is the dimension number of the Query space; to ensure that at any moment, only the moments before it can be attended to and not the moments after it, add a mask value M ij to the attention value, and the calculation formula is: The masked attention value is Perform normalization on it 2) Using the normalized attention as weights, aggregate the features of all moments to obtain the output features at each moment; first, weight the features of node v at all moments according to the attention to obtain the output at each moment: Z v = β v (X v );
[0023] The representation of the node learned in this way preserves the dynamic structural information of the graph.
[0024] A further improvement of the present invention lies in that the specific implementation method of step S2 is:
[0025] Based on the mesoscience complex system modeling theory, divide the invoice fraud data system into several meso-regions, and the specific implementation steps are as follows:
[0026] Step 1: Divide the meso-regions. During the evolution of the dynamic tax network in step S1, the dominant factor in invoice fraud is "invoice flow direction". Based on the mesoscience complex system modeling theory, the following levels and meso-regions are divided in the invoice fraud data system; invoice fraud has three levels, namely "data item - invoice", "invoice - enterprise", and "enterprise - gang"; among them, "data item", "invoice", "enterprise", and "gang" are regarded as boundary scales.
[0027] The invoice fraud data system is divided into three meso-regions, namely the meso-region M1 between the boundary scale data item E1 and the boundary scale invoice E2; the meso-region M2 between the boundary scale invoice E2 and the boundary scale fraudulent enterprise E3; the meso-region M3 between the boundary scale enterprise E3 and the boundary scale fraudulent gang E4; where the data item is the tax transaction network defined in step S1, and each meso-region forms a coupling through parameter and data transfer; the invoice fraud detection model has 3 layers, expressed as Each layer can be regarded as a triple Layer i =(E i , M i , E i+1 ), where E i and E i+1 represent adjacent boundary scales, corresponding to the input and output of the meso-region M i .
[0028] A further improvement of the present invention is that there is a "data constraint relationship" in the meso-region of the "data item - invoice" level, which is affected by the competition and coordination of control mechanisms such as product name control, amount control, and tax rate control.
[0029] A further improvement of the present invention is that there is a "supply and marketing relationship" in the meso-region of the "invoice - enterprise" level, which is affected by the competition and coordination of control mechanisms such as the relationship between purchase, sale, and inventory and the relationship between raw material production and investment.
[0030] A further improvement of the present invention is that there is an "interest relationship" in the meso-region of the "enterprise - gang" level, which is affected by the competition and coordination of control mechanisms such as the holding relationship and the transaction relationship.
[0031] A further improvement of the present invention is that the specific implementation method of step S3 is:
[0032] Based on the meso-regions divided in step S2, in order to achieve the divide-and-conquer of the model and the participation of humans, through the external tax domain knowledge K i in the meso-region, construct the control mechanism set R i in the meso-region M i , satisfying R i (Ei , K i ) → E i+1 , and apply it to the invoice fraud data system with multi-level and multi-scale coupling relationships. The specific implementation steps are as follows:
[0033] Step1. Symbolically represent the control mechanism
[0034] The control mechanism R for invoice fraud detection i includes tax law knowledge, inference rules, and calculation units; thus, R i is composed of the combination of Rules and CUs, satisfying where Rules represents the set of logical rules, CUs represents the set of calculation units, and × represents the Cartesian product;
[0035] The symbolic units and calculation units that make up the logical rules interact with each other, and then form the control mechanism R in the meso-region i ;
[0036] Step2. Construct a differentiable programming control mechanism
[0037] For the control mechanism required for invoice fraud detection, fine-grained atomic formulas are its main components; based on a bottom-up differentiable programming model to learn the control mechanism required in the meso-region, including atomic formula extraction, control mechanism learning, and discovery of precedence relationships. The specific steps are as follows:
[0038] Step201. First, for the existing tax domain knowledge such as relevant tax laws, apply the UIE information extraction algorithm to extract the corresponding entities and represent them with atomic formulas; then use the pre-trained model BERT to embed semantics, encode the tax knowledge in text, and obtain the attention weight vector T = {t1, t2,..., t N} through constructing a word-level attention mechanism, where n represents the length of the knowledge. For vectors with high weights, judge the relationship type between them, and thus extract the atomic formula <Atom> := (Pr, Sub, Obj) from the facts in the tax knowledge, where Pr represents the logical predicate, and Sub and Obj represent the subject and object of the atomic formula;
[0039] Step202. The relationships between the atomic formulas extracted in Step201 are represented by the matrix M, and the attention mechanism is used to analyze the weights of each atomic formula:
[0040]
[0041] where n is the number of relationships between the atomic formulas extracted in Step201, and the matrix element M ijIt represents the confidence that atomic formulas i and j exist in the rule, and then models the relationship based on the LSTM network model to extract the set of atomic formulas and their order;
[0042] Step203. Since the inference process of the intermediate region requires multiple control mechanisms to be jointly completed in a set order, sort and combine the control mechanisms obtained in Step202; First, determine the required topics and related principle I in the intermediate region; Second, calculate the attention degree of different control mechanisms for completing this inference through the attention mechanism α(·), so as to obtain the order relationship according to the attention degree; The formula is as follows:
[0043]
[0044]
[0045] Among them, represents the cosine similarity between the j-th control mechanism related to principle I and principle I, N represents the number of control mechanisms learned in Step202, TopN(·) represents descending order according to the attention degree, and finally the entire inference process is completed;
[0046] Through Step2, the order of multiple logical relationships in the intermediate region inference process is mined.
[0047] A further improvement of the present invention is that in Step1 of S3, the logical rule is expressed in the form of <Rules>::=(ME, Thm, Atoms, Opers)∈Rules, where ME represents the tax intermediate region divided in Step S2; Thm represents the topic of the intermediate region, Atoms is the set of atomic formulas that make up the logical rule; Opers represents the set of operation symbols that connect the atomic formulas to form the form of a first-order logical rule.
[0048] A further improvement of the present invention is that in Step1 of S3, the calculation unit is expressed as <CUs>::=(ME, Thm, Ic, Am)∈CUs, Ic is the input data type, and Am is the neural network model for identifying the taxpayer identity.
[0049] A further improvement of the present invention is that the specific implementation method of Step S4 is:
[0050] The detection process of invoice fraud detects whether an enterprise has invoice fraud behavior by coupling the calculation and inference of each control mechanism in different intermediate regions, and the inference path across intermediate regions provides an interpretable evidence chain for the inference result; The specific implementation steps are as follows:
[0051] Step1. Couple the reasoning process of the cross-media region. A complete process of detecting false invoice issuing is inferred by the mechanisms in the 3 media regions divided in Step S2. among them Let S i represent an actual operation of mechanism R i , understood as the reasoning step from E i to E i+1 . Thus, an inference path from the input data set I to the conclusion O is formed.
[0052] Step2. Generate the optimal inference path. During the reasoning process of the control mechanism, there are multiple inference paths that meet the final conclusion. The problem of generating the optimal inference path is transformed into a multi-step multi-objective combinatorial optimization problem. The specific steps are as follows:
[0053] Step201: Define the optimization objectives. To find the optimal inference path, it is necessary to define the optimization objectives. Here, two optimization objectives are considered. One is the accuracy of the inference path, and the other is the interpretability of the inference path.
[0054] (1) Inference path accuracy accuracy: Let the predicted inference path be P = (S1 → S2 →,..., → S i ,..., → S n ), and the true inference path be P' = (S'1 → S'2 →,..., → S' i ,..., → S' n ). Vectorize the inference path and use the cosine similarity to determine the accuracy between the predicted path P and the true path P'. The specific calculation formula is: The larger the accuracy value, the more accurate it is.
[0055] (2) Definition of interpretability:
[0056] a) Tax knowledge coupling degree σ: During the reasoning process, embed external tax knowledge into the reasoning control medium for reasoning. Use σ to measure the coupling degree between the reasoning steps in S3 and the external tax knowledge in the media region. If the reasoning step is expressed as and the external tax knowledge is expressed as then represents the optional calculation method. The higher the value of σ, the higher the degree of understandability.
[0057] b) Symbolic interpretability e i : Represents the understandability of the rules contained in the media region. Develop a symbolic interpretability dictionary by assigning values by domain experts to quantify the logical rules contained in the symbols, that is, e i = {(s i, v i )|(s i , v i ) ∈ dict(s, v)}, where s represents a symbol and v represents the artificial assignment of the interpretability of the symbol;
[0058] c) Inference path coupling degree r ij : During the inference process, the inference path will be transmitted between different medium regions. The inference path coupling degree is represented based on the inference confidence propagation algorithm of the medium region; the algorithm description is: the higher layer layer of the model k Score R k = ∑ j∈N R j←k , if there exist symbols i and j, and the medium regions they are in are continuous, then represents the confidence propagated from symbol j to symbol i and between medium regions m and n, where n = m + 1, and the propagation rule is: where α - β = 1, β ≥ 0, and j represents one of the symbols in the previous layer of the layer where symbol k is located;
[0059] The tax knowledge coupling degree, compliance interpretability, and inference path coupling degree defined by a), b), and c) respectively define the interpretability metric of the SHiL invoice fraud detection model
[0060] Step202: Generate the optimal inference path. Under the condition of satisfying the precision constraint, maximize the interpretability of the inference path. Since the interpretability objective can be decomposed into multiple sub - objectives, the optimal inference path generation problem is transformed into a multi - step multi - objective combinatorial optimization problem. The optimization objective is:
[0061] maxF(x)=(f1(x),..., f m (x)) T
[0062] subject to x ∈ Ω
[0063] where Ω represents the set of all feasible solutions, and a multi - objective evolutionary algorithm based on Chebyshev aggregation is used to generate the optimal inference path; the specific algorithm steps are: First, formalize the original problem as:
[0064] Input:
[0065] m weight vectors λ1, λ2,..., λ uniformly distributed m corresponding to the m optimization objectives f1(x),..., f m (x) respectively;
[0066] N: The number of sub - problems considered;
[0067] Output: Solve EP
[0068] Step1) Initialization:
[0069] Step1.1): Set EP = Φ;
[0070] Step1.2): Calculate the Euclidean distance between two weight vectors, find the nearest T weight vectors for each weight vector. For each i = 1,..., N, let B(i) = {i1,..., i T}, λ i1 ,..., λ iT be the nearest T weight vectors of λ i ;
[0071] Step1.3): Uniformly and randomly sample in the feasible solution space Ω to generate the initial population x 1 ,..., x N ;
[0072] Let FV i = F(x i );
[0073] Step1.4): Initialize z = (z1, z2,..., z m ) T ;
[0074] Step2) Update:
[0075] For i = 1,..., N, do;
[0076] Step2.1) Copy: Randomly select two indices k, l from B(i), and use genetic operators to generate a new solution y from x k and x l ;
[0077] Step2.2) Improvement: Apply repair and improvement heuristics based on the test problem to y to generate y′;
[0078] Step2.3) Update z: If z j < f j (y′), then |z j = f j (y′), j = 1, 2,..., m;
[0079] Step2.4) Update neighboring solutions: If g te (y′|λ j , z) ≤ g te (x j |λ j , z), j ∈ B(i), if x l= y′, FV j = F(y′);
[0080] Step2.5) Update EP: Remove all vectors dominated by F(y′) from EP;
[0081] Step3) Termination condition: When the termination condition is reached, stop and output EP, otherwise go back to Step2;
[0082] Through the above algorithm, finally generate the optimal inference path of invoice fraud in the solution EP, and this path serves as the evidence chain for invoice fraud.
[0083] The present invention has at least the following beneficial technical effects:
[0084] An interpretable invoice fraud detection method based on mesoscience provided by the present invention creatively combines the complex system modeling theory of mesoscience and the understandable and editable characteristics of symbolic systems, and constructs an interpretable invoice fraud detection model SHiL of "hierarchical controllable, manually participable, and result traceable". Compared with the existing mainstream invoice fraud detection methods using deep learning technology, the method provided by the present invention starts from the complex system modeling of mesoscience, divides the invoice fraud data system into several meso-regions with clear functions, and fundamentally breaks through the interpretability bottleneck of the existing black box models. At the same time, a symbolic control mechanism integrating specific tax domain knowledge is constructed for each meso-region to realize knowledge-driven data reasoning. Finally, the generation of the evidence chain can also be achieved through the traceability of the inference results. The invoice fraud detection method provided by the present invention will improve the interpretability of invoice fraud detection and provide a reliable and accurate "evidence chain". In addition, the interpretable method proposed by the present invention can also be applied to fields with high interpretability requirements such as unmanned driving and medical health. Description of the Drawings
[0085] Figure 1 It is the overall flowchart of the interpretable invoice fraud detection method based on mesoscience of the present invention.
[0086] Figure 2 It is the schematic diagram of the meso-regions of invoice fraud divided by the present invention.
[0087] Figure 3 It is the schematic diagram of the process of constructing the control mechanism in the invoice-enterprise meso-region of the present invention.
[0088] Figure 4 It is the schematic diagram of the process of detecting invoice fraud and generating the inference evidence chain. Detailed Embodiments
[0089] To more clearly illustrate the technical solution of the present invention, the present invention selects the invoice statistical information of 7 enterprises in SH City as a specific embodiment, and will describe in detail a method for detecting false invoice opening based on mesoscience and interpretability of the present invention in combination with specific embodiments and drawings hereinafter.
[0090] As Figure 1 shown, the method for detecting false invoice opening based on mesoscience and interpretability provided by the present invention is divided into three stages: determining the meso-region in the false invoice opening data system, constructing a symbolic differentiable programming control mechanism for the divided meso-regions, and coupling the calculation and reasoning of the control mechanism to generate an inference path across regions.
[0091] The method for detecting false invoice opening based on mesoscience and interpretability provided by the present invention includes the following steps:
[0092] S1, Tax transaction dynamic graph embedding based on self-attention mechanism.
[0093] First, convert tax transaction data into a semantically equivalent tax dynamic network for downstream data inference. Then, based on the mesoscience complex system modeling theory, divide the false invoice opening data system into several meso-regions. The specific implementation steps are as follows:
[0094] Convert tax data into a semantically equivalent tax dynamic network.
[0095] Tax data is generally composed of enterprise-to-enterprise transaction information at different times, so it can be modeled as a network structure. The present invention proposes a dynamic graph embedding method based on self-attention mechanism to obtain a tax dynamic network.
[0096] Step1: Feature preprocessing. For a certain enterprise, the attribute value of the general taxpayer flag field {Sfybnsrbz} is "yes", the industry name field {Xymc} is "sales industry", and the attribute value of the unit price field {Dj} is "200". For the text type {Xymc} attribute, use the BERT pre-trained language model to represent it as a word vector; for the categorical {Sfybnsrbz} attribute value, adopt the One-Hot encoding method to represent it as a vector; for the numerical type {Dj} attribute value, use the standardization method for processing. Finally, splice the above features to obtain all the features of the enterprise, and after processing the above features for each enterprise, the feature matrix X of the enterprise can be obtained.
[0097] Step2: Define a tax dynamic graph G = {G 1 , G 2 ,..., G T ,}, which includes static graphs at T moments, G t = {V, E t, X} is the static graph at the t-th moment; all static graphs share a node set, which is the set composed of all enterprises, but each static graph has its own edge set E t , because the transaction relationships at each moment are different. For example, at time t-1, there are transaction relationships between enterprise 1 and enterprise 2, enterprise 1 and enterprise 7, enterprise 1 and enterprise 5, enterprise 2 and enterprise 6, enterprise 2 and enterprise 7, enterprise 3 and enterprise 6, and enterprise 4 and enterprise 6. These transaction relationships together constitute the edge set E of the tax network at time t-1 t-1 , as time goes by, the transactions between enterprises will change; X is the feature matrix constructed in Step1. Due to different edge sets, each static graph has its own adjacency matrix A t used to represent the transaction relationships between enterprises, such as indicating that at the t-th moment, there is no transaction between the i-th enterprise and the j-th enterprise
[0098] Step3: Dynamic graph embedding based on the self-attention mechanism. For the static graph G T = {V, E t , X} at any moment, for the node v ∈ V, its neighbor node set is denoted as Calculate the attention value of v to each neighbor u: The calculation formula is: After calculating the attention of node v to all its neighbors, normalize the attention: Weightedly aggregate the vectors of the neighbors of node v according to the attention to obtain the output vector z of node v v , and the calculation formula is: At this time, this aggregation retains the local structure information of node v. To obtain the dynamic embedding, calculate the new representation of node v at each moment based on the self-attention mechanism. This new representation should preserve the dynamic information of the node changing over time, denoted as The specific calculation process is as follows: 1) Calculate the attention values of all moments for any moment. First, map the features of node v at all moments to the Query space: X v W q and the Key space: X v W k ; The attention value between time i and time j is: where is the scaling factor and is the dimension number of the Query space; To ensure that at any moment, only the moments before it can be attended to and the moments after it cannot be attended to, add a mask value M to the attention value ij , and the calculation formula is: The masked attention value is Perform normalization processing on it 2) Aggregate the features at all times with the normalized attention as the weight to obtain the output features at each time. First, weight the features at all times of node v according to the attention to obtain the output at each time: Z v = β v (X v ). The representation of the node learned in this way can preserve the dynamic structure information of the graph.
[0099] S2. Determination of the meso-region in the invoice fraud data system based on mesoscience.
[0100] Step 1: Divide the meso-region. In the evolution process of the dynamic tax network in step S1, the dominant factor in invoice fraud is "invoice flow direction". Based on the mesoscience complex system modeling theory, the following levels and meso-regions can be divided in the invoice fraud data system. Invoice fraud has three levels, namely "data item - invoice", "invoice - enterprise", and "enterprise - gang". Among them, "data item", "invoice", "enterprise", and "gang" can be regarded as boundary scales.
[0101] In the meso-region of the "data item - invoice" level, there is a "data constraint relationship", which is affected by the competition and coordination of control mechanisms such as product name control, amount control, and tax rate control.
[0102] In the meso-region of the "invoice - enterprise" level, there is a "supply and marketing relationship", which is affected by the competition and coordination of control mechanisms such as the relationship between purchase, sale, and inventory and the relationship between raw material production and investment.
[0103] In the meso-region of the "enterprise - gang" level, there is an "interest relationship", which is affected by the competition and coordination of control mechanisms such as the holding relationship and the transaction relationship.
[0104] As Figure 2 shown, the invoice fraud data system of the present invention can be divided into three meso-regions, namely the meso-region M1 between the boundary scale data item E1 and the boundary scale invoice E2; the meso-region M2 between the boundary scale invoice E2 and the boundary scale fraud enterprise E3; the meso-region M3 between the boundary scale enterprise E3 and the boundary scale fraud gang E4. Among them, the data item is the tax transaction network defined in step S1, and each meso-region forms a coupling through parameter and data transmission. Then, the invoice fraud detection model proposed by the present invention has 3 layers, denoted as Each layer can be regarded as a triple Layer i =(E i , M i , E i+1 ), where E i and E i+1 represent adjacent boundary scales, corresponding to the meso-region M iInput and output. For example, taking Layer1=(E1, M1, E2) as an example, in this layer, E1 is a data item, E2 is an invoice, and M1 represents the intermediate area between E1 and E2.
[0105] S3. Construct a symbolic differentiable programming control mechanism for the identified intermediate area.
[0106] Based on the intermediate area divided in step S2, in order to achieve the divide-and-conquer of the model and human participation, through the external tax domain knowledge K in the intermediate area i Construct the intermediate area M i The set of control mechanisms R in i satisfies R i (E i , K i )→E i+1 , and apply it to the invoice fraud data system with multi-level and multi-scale coupling relationships. The specific implementation steps are as follows:
[0107] Step1. Symbolically represent the control mechanism.
[0108] The control mechanism R for invoice fraud detection i Includes external tax knowledge, inference rules, and computing units. Therefore, R i Is composed of the combination of Rules and CUs, satisfying Among them, Rules represents the set of logical rules, CUs represents the set of computing units, and × represents the Cartesian product.
[0109] Specifically, the logical rule is represented in the form of <Rules>::=(ME, Thm, Atoms, Opers)∈Rules, where ME represents the tax intermediate area divided in step S2, such as invoice-enterprise, enterprise-gang; Thm represents the theme of the intermediate area, such as special invoice, tax, ordinary invoice, etc., Atoms is the set of atomic formulas that make up the logical rule; Opers represents the set of operation symbols that connect the atomic formulas to form the form of a first-order logical rule.
[0110] Specifically, the computing unit can be represented as <CUs>::=(ME, Thm, Ic, Am)∈CUs, Ic is the input data type, and Am is the neural network model for identifying the taxpayer identity.
[0111] Such as Figure 3As shown, there is external tax knowledge K2 in the M2 mediation area. According to the "Regulations on the Issuance of Special VAT Invoices": "General taxpayers selling goods or providing taxable services shall issue invoices to the purchasers", and "General taxpayers in commercial enterprises may not issue special invoices for the retail sale of cigarettes, alcohol, food, clothing, shoes and hats, cosmetics and other consumer goods". Based on the above control mechanism representation rules, the control mechanism can be represented as shown in the following table:
[0112]
[0113] Step2. Construct a differentiable programming control mechanism.
[0114] Step201: In the mediation area divided by S2, first extract the atomic formulas contained in the external tax knowledge.
[0115] Such as Figure 3 As shown, in this embodiment, there is external tax knowledge K2 in the M2 mediation area: "General taxpayers selling goods or providing taxable services shall issue invoices to the purchasers". First, through the UIE information extraction algorithm (Universal Information Extraction), four entities, namely "general taxpayer", "selling goods", "providing taxable services", and "issuing invoices", can be extracted and represented as R, P, Q, and T respectively. Applying the pre-trained model BERT to embed the semantics of tax domain knowledge and establishing a word-level attention mechanism, the attention weight vector T = {t1, t2,..., t N}, where n represents the length of the knowledge. For vectors with higher weights, judge the relationship type, and thus the atomic formula <Atom> := (Pr, Sub, Obj) can be extracted from the facts in the tax knowledge, where Pr represents the logical predicate, and Sub and Obj represent the subject and object of the atomic formula. In this embodiment, the atomic formula that can be extracted according to the tax law above is: (R ∧ (P ∨ Q)) → T. Further, each entity type is represented by a pre-set predicate. In this embodiment, "general taxpayer" is represented by the taxpayer predicate, "selling goods" is represented by the sellGoods predicate, "providing taxable services" is represented by the ptServices predicate, and "issuing invoices" is represented by the speInvoice predicate. Then (R ∧ (P ∨ Q)) → T can be further represented as taxpayer(X) ∧ (sellGoods(X, Z) ∨ ptServices(X, Z)) → speInvoice(Z, X).
[0116] Step202. Represent the relationship between the atomic formulas extracted in Step201 with a matrix M, and use the attention mechanism to perform weight analysis on each atomic formula:
[0117]
[0118] Among them, n is the number of relationships between the atomic formulas extracted in Step 201, and the matrix element M ij represents the confidence that atomic formulas i and j exist in the rule. Then, based on the LSTM network model, the relationships are modeled to extract the set of atomic formulas and their order, and further a control mechanism is learned, that is, each control mechanism is composed of several atomic formulas. As Figure 3 shown, the control mechanism includes atomic formula 1 and atomic formula 2, and the control mechanism includes atomic formula 1, atomic formula 3, and atomic formula n.
[0119] Step 203. Sort the control mechanisms in Step 202. First, determine the theme and related principle I in the mediation area; secondly, calculate the attention degree of different control mechanisms for completing this reasoning through the attention mechanism α(·), so as to obtain the order relationship according to the attention degree. The formula is as follows:
[0120]
[0121]
[0122] Among them, represents the cosine similarity between the j-th control mechanism related to principle I and principle I, N represents the number of control mechanisms learned in Step 202, and TopN(·) represents descending order according to the attention degree. Finally, the entire reasoning process can be completed. As Figure 3 shown, in this embodiment, the finally formed order of the control mechanisms is (Special invoice issuance) → (Special invoice non-issuance) → (Invoice amount).
[0123] S4. Through the calculation and reasoning of the coupled control mechanism, detect whether the enterprise has the behavior of issuing false invoices, and at the same time generate an inference path across regions to provide an interpretable evidence chain for the inference result. The specific implementation steps are as follows:
[0124] Step 1: Couple the inference processes across mediation areas. A complete process of detecting false invoice issuance needs to be inferred by the mechanisms in the 3 mediation areas divided in Step S2 . Let S i represent an actual operation of mechanism R i , which can be understood as from E i to E i+1The reasoning steps, and the reasoning path from the input data set I to the conclusion O formed thereby As Figure 4 shown, in Figure 4 , the reasoning path of the solid black line is P1, and this reasoning path requires different operations S between i to complete together. Specifically, to generate the reasoning path P1, it is necessary to control the mechanism between the (S1, S2, S3, S4, S5, S6) reasoning steps to complete.
[0125] Step2: Optimal reasoning path generation. During the reasoning process of the control mechanism, there may be multiple reasoning paths that meet the final conclusion. The present invention needs to find the "optimal" reasoning path. Considering the accuracy of the invoice detection model, the present invention pursues the maximization of the interpretability of the reasoning path. Therefore, the present invention can transform the problem of generating the optimal reasoning path into a multi-step multi-objective combinatorial optimization problem. The specific steps are as follows:
[0126] Step201: Define the optimization objective. In order to find the optimal reasoning path, the present invention needs to define the optimization objective. Here, mainly two optimization objectives are considered. One is the accuracy of the reasoning path; the other is the interpretability of the reasoning path.
[0127] (1) Reasoning path accuracy accuracy: Let the predicted reasoning path be P = (S1 → S2 →,..., → S i ,..., → S n ), and the true reasoning path be P' = (S'1 → S'2 →,..., → S' i ,..., → S' n ). Vectorize the reasoning path, and use the cosine similarity to determine the accuracy between the predicted path P and the true path P'. The specific calculation formula is: The larger the accuracy value, the more accurate it is. Specifically, in Figure 4 there are 3 different operations in each of the three control mechanisms , then there may be 27 different reasoning paths in the end. At this time, the one-hot vectors of 3 dimensions of 3 will be concatenated to form the final reasoning path vector Let the vector corresponding to the predicted reasoning path be The vector of the true path is Then the final reasoning path accuracy
[0128] (2) Definition of interpretability:
[0129] a) Tax knowledge coupling degree σ: During the reasoning process, it is necessary to embed external tax knowledge into the reasoning control medium for reasoning. In the present invention, σ is used to measure the coupling degree between the reasoning steps in S3 and the external tax knowledge in the medium region. If the reasoning step is represented as and the external tax knowledge is represented as then represents an optional calculation method. The higher the value of σ, the higher the degree of comprehensibility.
[0130] b) Symbol interpretability e i : It represents the comprehensibility of the rules contained in the medium region. By means of domain expert assignment, a symbol interpretability dictionary is formulated to quantify the logical rules contained in the symbols, that is, e i = {(s i , v i ) | (s i , v i ) ∈ dict(s, v)}, where s represents the symbol and v represents the manual assignment of the interpretability of the symbol.
[0131] c) Reasoning path coupling degree r ij : During the reasoning process, the reasoning path will be transmitted between different medium regions. The present invention proposes a reasoning confidence propagation algorithm based on the medium region to represent the reasoning path coupling degree. The algorithm can be expressed as: The score R k of the higher layer layer k of the model = ∑ j∈N R j←k . If there are symbols i and j, and the medium regions where they are located are continuous, then represents the confidence propagated from symbol j to symbol i and between medium regions m and n, where n = m + 1, and the propagation rule is: where α - β = 1, β ≥ 0, and j represents the symbol among the symbols in the previous layer of the layer where symbol k is located
[0132] The tax knowledge coupling degree, symbol interpretability, and reasoning path coupling degree defined by a), b), and c) respectively can define the interpretability metric of the SHiL invoice fraud detection model
[0133] Step202: Generate the optimal reasoning path. Under the condition of meeting the accuracy constraint, maximize the interpretability of the reasoning path. Since the interpretability objective can be decomposed into multiple sub - objectives, the optimal reasoning path generation problem is transformed into a multi - step multi - objective combinatorial optimization problem. First, according to the optimization objective defined in Step201, the optimization objective of the invoice fraud detection system is:
[0134] maxF(x) = (f1(x), f2(x)) T
[0135] subject to x ∈ Ω
[0136] where f1(x) and f2(x) correspond to accuracy(x) and ε(x) respectively, that is, to maximize the inference accuracy and interpretability. Where F(x) is the total optimization objective, accuracy and ε are the inference path accuracy and inference path interpretability defined in Step 201 respectively, and Ω represents the set of all possible inference paths. The present invention uses a multi-objective evolutionary algorithm based on Chebyshev aggregation to generate the optimal inference path. The specific algorithm steps are as follows:
[0137] Input:
[0138] Two weight vectors λ1 and λ2 with a uniform distribution correspond to the two optimization objectives of accuracy(Ω) and ε(Ω) respectively.
[0139] N: The number of sub-problems considered.
[0140] Output: Solution EP
[0141] Step 1) Initialization:
[0142] Step 1.1): Set EP = Φ.
[0143] Step 1.2): Calculate the Euclidean distance between the two weight vectors, and find the T weight vectors closest to each weight vector. For each i = 1,..., N, let B(i) = {i1,..., i T}, λ i1 ,..., λ iT are the T weight vectors closest to λ i .
[0144] Step 1.3): Uniformly and randomly collect and generate the initial population x 1 ,..., x N in the feasible solution space Ω.
[0145] Let FV i = F(x i )
[0146] Step 1.4): Initialize z = (z1, z2,..., z m ) T .
[0147] Step 2) Update:
[0148] For i = 1,..., N, do
[0149] Step2.1) Copy: Randomly select two serial numbers k and l from B(i), and use genetic operators to generate a new solution y from x k and x l .
[0150] Step2.2) Improvement: Apply repair and improvement heuristics based on the test problem to y to generate y'.
[0151] Step2.3) Update z: If z j < f j (y'), then |z j = f j (y'), j = 1, 2,..., m.
[0152] Step2.4) Update adjacent solutions: If g te (y'|λ j , z) ≤ g te (x j |λ j , z), j ∈ B(i), if x l = y', FV j = F(y').
[0153] Step2.5) Update EP: Remove all vectors dominated by F(y') from EP.
[0154] Step3) Termination condition: When the termination condition is reached, stop and output EP, otherwise go to Step2.
[0155] After the above algorithm, finally, the optimal inference path of invoice fraud can be generated in the solution EP, and this path can be used as an evidence chain for invoice fraud.
[0156] As Figure 4 shown, in this embodiment, there are 2 inferable paths, namely the inference path P1 shown by the solid line and the inference path P2 shown by the dashed line. By running the above algorithm, the optimization objectives of accuracy(Ω) and ε(Ω) of the inference path P1 are dominant compared to the inference path P2. Therefore, the system selects the inference path P1 as the final output path, that is, the inference process of invoice fraud detection.
[0157] Although the present invention has been described in detail above with general descriptions and specific embodiments, based on the present invention, some modifications or improvements can be made, which are obvious to those skilled in the art. Therefore, these modifications or improvements made without departing from the spirit of the present invention all fall within the scope of protection required by the present invention.
Claims
1. An interpretable invoice fraud detection method based on mesoscience, characterized in that, It includes the following steps: S1. Tax transaction dynamic graph embedding based on the self-attention mechanism: First, the tax transaction data is transformed into a tax dynamic network through feature processing and feature embedding steps. The tax dynamic network preserves the transaction information and other structural information between enterprises at different time slices. S2. Determination of the meso-region of the invoice fraud data system based on mesoscience: According to the mesoscience complex data system modeling theory, in the tax dynamic transaction network, "invoice fraud" is the dominant factor of invoice fraud behavior. Therefore, based on the dominant factor "invoice flow direction", the data system of the tax network is divided into different meso-regions. S3. Construct a symbolic differentiable programming control mechanism for the divided meso-regions: Based on the divided meso-regions, an artificial participatory control mechanism is constructed through two steps: symbolic representation of the control mechanism and construction. The specific implementation method is as follows: Based on the mediation regions divided in step S2, in order to achieve the divide-and-conquer of the model and human participation, through the external tax domain knowledge K in the mediation regions i Construct the mediation region M i The set of control mechanisms R in i , satisfying R i (E i , K i ) → E i+1 , and apply it to the invoice fraud data system with multi-level and multi-scale coupling relationships. The specific implementation steps are as follows: Step1. Symbolically represent the control mechanism Control mechanism R for detecting false invoicing i It includes tax law knowledge, inference rules and calculation units; thus R i is composed of a combination of Rules and CUs, satisfying where Rules represents a set of logical rules, CUs represents a set of calculation units, and × represents the Cartesian product; The symbolic units and computational units that make up the logical rules interact with each other, thereby forming the control mechanism R in the meso-region i ; Step2. Construct a differentiable programming control mechanism For the control mechanism required for invoice fraud detection, fine-grained atomic formulas are its main components. Based on a bottom-up differentiable programming model, learn the control mechanism required in the meso-region, including atomic formula extraction, control mechanism learning, and discovery of the order relationship. The specific steps are as follows: Step201. First, for the existing tax laws related to tax domain knowledge, apply the UIE information extraction algorithm to extract the corresponding entities and represent them with atomic formulas. Then, use the pre-trained model BERT to embed semantics, encode the tax knowledge in text, and obtain the attention weight vector T = {t1, t2,..., t N} by constructing a word-level attention mechanism, where n represents the length of the knowledge. For vectors with high weights, judge the type of relationship between them, so as to extract atomic formulas from the facts in the tax knowledge <atom>∷=(Pr, Sub, Obj), where Pr represents the logical predicate, and Sub and Obj represent the subject and object of the atomic formula. < / atom> Step202. The relationship between the atomic formulas extracted in Step201 is represented by the matrix M, and the attention mechanism is used to analyze the weight of each atomic formula: Among them, n is the number of relationships between the atomic formulas extracted in Step 201, and the matrix element M ij represents the confidence that atomic formulas i and j exist in the rule, and then the relationships are modeled based on the LSTM network model to extract the atomic formula set and its order; Step203. Since the reasoning process of the meso-region requires multiple control mechanisms to be completed jointly in a set order, sort and combine the control mechanisms obtained in Step202. First, determine the required topics and related principles I in the meso-region. Secondly, calculate the attention degree of different control mechanisms for completing this reasoning through the attention mechanism α(·), so as to obtain the order relationship according to the attention degree. The formula is as follows: Among them, represents the cosine similarity between the j-th control mechanism related to Principle I and Principle I. N represents the number of control mechanisms learned in Step 202. TopN(·) represents descending order according to the degree of attention, and finally the entire reasoning process is completed; Through Step2, the order of multiple logical relationships in the meso-region reasoning process is mined. S4. By coupling the calculation and reasoning of the control mechanism, detect whether there is invoice fraud behavior in the enterprise, and at the same time generate a cross-region reasoning path to provide an interpretable evidence chain for the reasoning result. Optimally select the reasoning path formed by coupling the reasoning steps of each meso-region based on the Chebyshev aggregation multi-objective evolutionary algorithm, and finally select an optimal solution to provide evidence for the reasoning result.
2. The interpretable invoice fraud detection method based on mesoscience according to claim 1, wherein The specific implementation method of Step S1 is as follows: First, transform the tax transaction data into a tax dynamic network for downstream data reasoning. The specific implementation steps are as follows: Step1: Feature preprocessing; for text features, use the BERT training language model to represent them as word vectors; for categorical attribute values, adopt the One-Hot encoding method to represent them as vectors; for numerical attribute values, adopt the standardization method for processing; finally, splice the above features to obtain all the features of the enterprise. After performing the above feature processing on each enterprise, obtain the feature matrix X of the enterprise. Step2: Define a tax dynamic graph $G = \{G 1 , G 2 , \ldots, G T \}$, which contains static graphs at $T$ moments. $G t = \{V, E t , X\}$ is the static graph at the $t$-th moment. All static graphs share a node set, which is the set composed of all enterprises. However, each static graph has its own edge set $E t $, and $X$ is the feature matrix constructed in Step1. Since the edge sets are different, each static graph has its own adjacency matrix $A t $ to represent the transaction relationship between enterprises. indicates that there is no transaction between the $i$-th enterprise and the $j$-th enterprise at the $t$-th moment. Step3: Dynamic graph embedding based on self-attention mechanism; for the static graph G at any moment T ={V, E t , X}, for the node v ∈ V, its neighbor node set is denoted as Calculate the attention value of v to each neighbor u: The calculation formula is: After calculating the attention of node v to all its neighbors, normalize the attention: Weightedly aggregate the vectors of the neighbors of node v according to the attention to obtain the output vector z of node v v , and the calculation formula is: At this time, this aggregation retains the local structure information of node v; to obtain the dynamic embedding, calculate the new representation of node v at each moment based on the self-attention mechanism, and this new representation should preserve the dynamic information of the node changing over time, denoted as The specific calculation process is as follows: 1) Calculate the attention values for all moments at any moment. First, map the features of all moments of node v to the Query space respectively: X v W q , Key space: X v W k ; The attention value between moment i and moment j is: where is the scaling factor, and is the dimension number of the Query space; to achieve that at any moment, only the moments before it can be attended to, and the moments after it cannot be attended to, add a mask value M ij to the attention value, and the calculation formula is: The masked attention value is Perform normalization on it 2) Using the normalized attention as the weight, aggregate the features of all moments to obtain the output features at each moment; First, weight the features of all moments of node v according to the attention to obtain the output at each moment: Z v = β v (X v ); The representation of nodes learned in this way preserves the dynamic structure information of the graph.
3. An interpretable invoice fraud detection method based on mesoscience according to claim 2, characterized in that, The specific implementation method of step S2 is as follows: Based on the mesoscience complex system modeling theory, the invoice fraud data system is divided into several meso-regions. The specific implementation steps are as follows: Step1: Divide the meso-regions. In the evolution process of the dynamic tax network in step S1, the dominant factor in invoice fraud is "invoice flow direction". Based on the mesoscience complex system modeling theory, the following levels and meso-regions are divided in the invoice fraud data system; invoice fraud has three levels, namely "data item - invoice", "invoice - enterprise", and "enterprise - gang"; among them, "data item", "invoice", "enterprise", and "gang" are regarded as boundary scales. The invoice fraud data system is divided into three intermediate regions, namely the intermediate region M1 between the boundary scale data item E1 and the boundary scale invoice E2; the intermediate region M2 between the boundary scale invoice E2 and the boundary scale fraudulent enterprise E3; the intermediate region M3 between the boundary scale enterprise E3 and the boundary scale fraud gang E4; where the data item is the tax transaction network defined in step S1, and each intermediate region forms a coupling through parameter and data transfer; the invoice fraud detection model has 3 layers, denoted as Each layer can be regarded as a triple Layer i =(E i ,M i ,E i+1 ), where E i and E i+1 represent adjacent boundary scales, corresponding to the input and output of the intermediate region M i .
4. The interpretable invoice fraud detection method based on mesoscience according to claim 3, characterized in that There is a "data constraint relationship" in the meso-region at the "data item - invoice" level, which is affected by the competition and coordination of control mechanisms such as product name control, amount control, and tax rate control.
5. The interpretable invoice fraud detection method based on mesoscience according to claim 3, wherein There is a "supply and marketing relationship" in the meso-region at the "invoice - enterprise" level, which is affected by the competition and coordination of control mechanisms such as the relationship between purchase, sale, and inventory and the relationship between raw material production and operation.
6. The interpretable invoice fraud detection method based on mesoscience according to claim 3, characterized in that, There is an "interest relationship" in the meso-region at the "enterprise - gang" level, which is affected by the competition and coordination of control mechanisms such as shareholding relationship and transaction relationship.
7. The method for detecting false invoicing based on mesoscience and having interpretability according to claim 1, wherein In Step 1 of Step S3, the logical rule is expressed as <rules>In the form of ∷=(ME, Thm, Atoms, Opers) ∈ Rules, where ME represents the tax meso-region divided in step S2; Thm represents the theme of the meso-region, Atoms is the set of atomic formulas constituting the logical rule; Opers represents the set of operation symbols connecting the atomic formulas to form the form of a first-order logical rule. < / rules> 8. The interpretable invoice fraud detection method based on mesoscience according to claim 1, wherein In Step 1 of Step S3, the calculation unit is represented as <cus>In the form of ∷=(ME, Thm, Ic, Am) ∈ CUs, Ic is the input data type, and Am is the neural network model for identifying the taxpayer identity. < / cus> 9. The interpretable invoice fraud detection method based on mesoscience according to claim 1, characterized in that, The specific implementation method of step S4 is as follows: The detection process of invoice fraud detects whether there is invoice fraud behavior in an enterprise by coupling the calculations and inferences of various control mechanisms in different meso-regions. The cross-meso-region inference path provides an interpretable evidence chain for the inference result; the specific implementation steps are as follows: Step1. Couple the reasoning processes in the cross-media regions. A complete process of detecting invoice fraud is inferred by the mechanisms in the 3 media regions divided in Step S2 among them Let S i represent an actual operation of mechanism R i which is understood as the reasoning step from E i to E i+1 and thus the reasoning path from the input data set I to the conclusion O is formed Step2. Generate the optimal inference path. In the inference process of the control mechanism, there are multiple inference paths that conform to the final conclusion. The problem of generating the optimal inference path is transformed into a multi-step multi-objective combinatorial optimization problem. The specific steps are as follows: Step201: Define the optimization objectives. To find the optimal inference path, it is necessary to define the optimization objectives; here, two optimization objectives are considered. One is the accuracy of the inference path, and the other is the interpretability of the inference path. (1) Inference path accuracy: Let the predicted inference path be P = (S1 → S2 →,…, → S i ,…, → S n ), and the true inference path be P′ = (S′1 → S′2 →,…, → S′ i ,…, → S′ n ); Vectorize the inference path and use the cosine similarity to determine the accuracy between the predicted path P and the true path P′. The specific calculation formula is: The larger the accuracy value, the more accurate it indicates. (2) Definition of interpretability: a) Tax knowledge coupling degree σ: During the reasoning process, external tax knowledge is embedded in the reasoning control medium for reasoning; σ is used to measure the coupling degree between the reasoning steps in S3 and the external tax knowledge in the medium region. If the reasoning step is expressed as The external tax knowledge is expressed as Then Indicates an optional calculation method. The higher the value of σ, the higher the degree of comprehensibility. b) Symbol interpretability e i : It represents the comprehensibility of the rules contained in the intermediate region. By means of assignment by domain experts, a symbol interpretability dictionary is formulated to quantify the logical rules contained in the symbols, that is, e i = {(s i , v i ) | (s i , v i ) ∈ dict(s, v)}, where s represents the symbol and v represents the manual assignment of the interpretability of the symbol; c) Inference path coupling degree r ij : During the inference process, the inference path is transmitted between different medium regions. The inference path coupling degree is represented based on the inference confidence propagation algorithm for medium regions. The algorithm is described as follows: for a higher layer layer of the model k Score R k = ∑ j∈ N R j←k , if there exist symbols i and j, and the medium regions they are in are continuous, then represents the confidence propagated from symbol j to symbol i and between medium regions m and n, where n = m + 1, and the propagation rule is: where α - β = 1, β ≥ 0, and j represents one of the symbols in the layer before the layer where symbol k is located; The tax knowledge coupling degree, compliance interpretability, and inference path coupling degree defined by a), b), and c) respectively define the interpretability metric of the SHiL invoice fraud detection model Step202: Generate the optimal inference path. Under the condition of satisfying the accuracy constraint, maximize the interpretability of the inference path. Since the interpretability objective can be decomposed into multiple sub-objectives, the problem of generating the optimal inference path is transformed into a multi-step multi-objective combinatorial optimization problem. The optimization objective is: maxF(x) = (f1(x), …, f e (x)) T subject to x ∈ Ω where Ω represents the set of all feasible solutions, and a multi-objective evolutionary algorithm based on Chebyshev aggregation is used to generate the optimal inference path; the specific algorithm steps are as follows: First, the original problem is formally represented as: Input: m weight vectors λ1, λ2, …, λ that are evenly distributed m respectively correspond to the m optimization objectives of f1(x), …, f m (x); N: The number of sub-problems considered; Output: Solution EP Step 1) Initialization: Step 1.1): Set EP = Φ; Step 1.2): Calculate the Euclidean distance between two weight vectors, and find the T weight vectors closest to each weight vector. For each i = 1, …, N, let B(i) = {i1, …, i T}, λ i1 , …, λ iT be the T weight vectors closest to λ i ; Step 1.3): Uniformly and randomly sample in the feasible solution space Ω to generate the initial population x 1 ,…,x N ; Let FV i = F(x i ); Step 1.4): Initialize z = (z1, z2, …, z m ) T ; Step 2) Update: For i = 1, …, N, do; Step 2.1) Copy: Randomly select two serial numbers k and l from B(i), and use the genetic operator to generate a new solution y from x k and x l ; Step 2.2) Improvement: Apply repair and improvement heuristics based on the test problem to y to generate y ′ ; Step 2.3) Update z: If z j <f j (y ′ ), then |z j = f j (y ′ ), j = 1, 2, …, m; Step 2.4) Update adjacent solutions: If g te (y ′ ∣λ j , z) ≤ g te (x j ∣λ j , z), j ∈ B(i), if x l = y′, FV j = F(y ′ ); Step 2.5) Update EP: Remove all vectors dominated by F(y ′ ) from EP; Step 3) Termination condition: If the termination condition is reached, stop and output EP, otherwise go to Step 2; After the above algorithm, finally generate the optimal inference path of invoice fraud in the solution EP, and this path serves as the evidence chain for invoice fraud.
Citation Information
Patent Citations
A method and system for identifying fraudulent invoices based on positive examples and unlabeled learning
CN110532542B
A method for identifying fraudulent invoices based on dynamic network representation
CN110852856B
Detection method of corporate invoice fraud based on graph attention network with dynamic edge features
CN111724241B