A Cross-Media Region Interpretable Inference Path Generation Method for Invoice Fraud Detection

Abstract

The present invention discloses a method for generating an interpretable inference path across media regions for invoice fraud detection, including the steps: First, aiming at the interpretability of deep learning, a method for measuring interpretability combining subjective and objective factors is proposed, which includes four indicators. On this basis, an inference path selection and sorting algorithm based on Pareto efficiency is proposed, which converts the problem of arranging and selecting fragmented paths into a multi-objective optimization problem. By adjusting the boundaries of different objectives, the optimal inference path with multi-objective balance is selected to form an evidence chain, realizing the interpretability of the model.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence and tax control technology, and specifically relates to a method for generating a cross-region explainable reasoning path for detecting false invoices. Background Art

[0002] False invoicing refers to the practice of issuing invoices that are not true to the facts. To evade taxes, taxpayers and individuals resort to fraudulent methods, such as falsifying product names, quantities, unit prices, and amounts. They even exploit familiar relationships to fabricate transactions and issue false invoices. False invoices have long been a persistent problem in tax governance, resulting in a massive tax loss of approximately 1.1 trillion yuan annually. Therefore, detecting false invoices is crucial for tax risk management. Furthermore, tax data is complex and high-dimensional, covering 19 tax categories and 89 industries, with approximately 60 billion invoices issued annually. In recent years, deep learning technology has been successfully applied in tax control applications, such as tax evasion detection, industry classification, and detection of abnormal corporate behavior. While early models have achieved promising results, end-to-end intelligent learning models have yet to achieve interpretability of the detection process and results. This has led to numerous challenges and difficulty verifying the results. Audits and verifications still require significant resources and manpower, raising questions about the credibility and fairness of the algorithms, and even creating legal issues. This has become a bottleneck for the large-scale application of deep learning in tax control. Therefore, at this stage, it is proposed to integrate relevant theories of mesoscience into the model. Mesoscience is a complex system analysis method that can discover mesoregions in data systems with multi-scale spatiotemporal dynamic structures, thereby rationally decomposing complex systems and facilitating the establishment of interpretable learning models. In tax control scenarios, domain knowledge is combined to identify the spatiotemporal multi-layers of complex data systems and the mesoregions at different levels. For example, the detection of fraudulent invoices requires the exploration of mesoregions at the "invoice-enterprise" and "enterprise-gang" levels. This method organically combines the complex system modeling theory of mesoscience with the understandable and easy-to-edit characteristics of symbolic systems, where the symbols represent knowledge related to physics or sociology and take the form of atomic formulas and operators between atomic formulas. The problem of constructing symbolic control mechanisms within mesoregions is solved by divide and conquer. Fragmented reasoning paths are generated, but currently there is no method to effectively piece together the reasoning paths for fraudulent invoices within mesoregions into a complete, interpretable, accurate, and timely reasoning path across mesoregions.

[0003] To address this problem, the following literature provides references for generating explainable paths to address false invoices and proposes corresponding solutions:

[0004] Reference 1: A Method for Explainable Sentiment Tracing (202110743688.2)

[0005] Document 2: A Method for Identifying the Risk of Circular Invoice Issuance Based on a Knowledge Graph (202010715798.3);

[0006] Document 3: A Method for Identifying False Invoice Issuance Based on Dynamic Network Representation (CN201911066791.7);

[0007] Document 4: A Multi-objective Optimization Method for Vehicle Safety Based on NSGA-II (202110814834.6);

[0008] Document 1 first proposed the concept of emotion traceability. Emotion traceability refers to finding a complete reasoning link in the process of emotion analysis. In terms of actual operation, this patent uses the method of manual annotation to generate a data set. During the training process, each path is vectorized and input into a linear layer and softmax normalization to obtain the score of each emotion traceability path, and the path with the highest score is selected as the final interpretable emotion traceability path.

[0009] Document 2 proposed a method for identifying the risk of circular invoice issuance based on a knowledge graph. This method for identifying the risk of circular invoice issuance based on a knowledge graph starts from the circular invoice transaction relationship, builds the backbone network of taxpayers' transaction relationships based on knowledge graph technology, uses graph algorithms to search for abnormal transaction relationships, reveals the risk of false invoice issuance, and on the basis of identifying circular invoice taxpayers, combines the attribute information of taxpayers to discover entity association relationships and identify false invoice rings.

[0010] Document 3 is a method for identifying false invoice issuance based on dynamic network representation. First, taking enterprises as nodes and transaction records as edges, organize enterprise information into a static network. Secondly, taking each day as a time node, establish the representation of the enterprise transaction network, establish a time window of about 30, and obtain the dynamic network representation result by moving the time window. Thirdly, draw on the distributed optimization algorithm, decompose the objective function of the representation into independent sub-functions to improve the learning efficiency of the model. Finally, based on LightBGM, construct a classifier to identify enterprises with false invoice issuance.

[0011] Document 4 proposed a multi-objective optimization method for vehicle safety based on NSGA-II. First, determine the range and default values of vehicle physical parameters. Establish a multi-objective optimization model, including: establishing a model for selecting vehicle physical parameters; performing mean value processing according to the value ranges of different parameters to establish a model for the difference in vehicle physical parameter configuration; establishing a model for vehicle obstacle avoidance effect according to the obstacle avoidance-related data returned by configuring vehicle physical parameters into an automotive simulation simulator; using the NSGA-II algorithm to solve the multi-objective optimization model. By setting the multi-objective optimization model, reduce the search space and obtain the Pareto optimal solution set in a relatively short time.

[0012] Although the above traditional methods can solve specific interpretability problems and optimization problems, it is difficult to directly extend them to the inference path generation of the complex dynamic system in the scenario of invoice fraud. The reason is that the processing performance and universality of the existing interpretation methods are weak. The research objects have preconditions of small scale, clear boundaries, and single goals, which are very different from the complex data systems required by this project and are difficult to apply. At the same time, the existing interpretability measurement methods have subjective bias problems. Moreover, the complex data system in the scenario of invoice fraud has the characteristics of massive and high-dimensional data. There are more than 50 million taxpaying enterprises in China currently, involving 19 tax categories and 89 industries, and about 60 billion invoices are issued every year.

[0013] In Document 1, only the method of manual annotation is used to generate the training set to train the sentiment traceability path scoring model to judge which path has the best interpretability. Inevitably, there will be subjective bias problems when only using the method of manual annotation. Interpretability is the intersection between objective interpretation and subjective understanding ability. At present, the core indicators of interpretability assessment include: plausibility and faithfulness. Document 2 mainly focuses on the risk modeling of cyclic invoicing, with a small application scope and cannot handle complex tax scenarios. At the same time, using the depth-first algorithm to find the closed loop in the directed graph will lead to a slow search speed, and the problem is more obvious when the search space is larger; although Document 3 can consider the transaction network characteristics in different time periods to improve the detection rate, it has the same defect of weak result interpretability as Document 2. This will lead to many doubts about the invoice fraud detection results and great difficulty in verification, and still requires a large amount of manpower and material resources 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 NSGA-II adopted in Document 4 has good effects on low-dimensional multi-objective optimization problems. However, for high-dimensional multi-objective optimization problems, the first problem it faces is the too small selection pressure caused by its Pareto dominance relationship. Secondly, the crowding distance is not applicable in high-dimensional spaces, and the computational complexity is relatively high; the algorithms used in Document 4 all belong to heuristic search, but heuristic search cannot guarantee Pareto efficiency. It can only ensure that the obtained solution is not dominated by the other party (but can still be dominated by the Pareto efficient solution). Summary of the Invention

[0014] The present invention aims to provide a method for generating an interpretable inference path across meso-regions for invoice fraud detection. First, based on the complex system analysis method of mesoscience theory, on the basis of dividing the tax data system into several meso-regions, an interpretable measurement scheme for tax data between and within meso-regions is constructed; then, an interpretable loss function for invoice fraud is designed, and then an inference path selection and sorting algorithm based on Pareto efficiency is used to generate an inference path, converting the problem of arranging and selecting fragmented paths into a multi-objective optimization problem, and by adjusting the boundaries of different objectives, the optimal inference path with multi-objective balance is selected to form an evidence chain, realizing the interpretability of the model.

[0015] The present invention is implemented by the following technical solutions:

[0016] A method for generating an interpretable inference path across meso-regions for invoice fraud detection, comprising the following steps:

[0017] First, on the basis of obtaining a candidate set, that is, path nodes, in the meso-region, an interpretable measurement method is designed according to the characteristics of the path nodes in the meso-region; then, according to the interpretable measurement method, an interpretable loss function is designed, and on this basis, an inference path selection and learning algorithm for maximizing the interpretability of the path under the conditions of meeting the accuracy constraint and the timeliness constraint is realized; then, all possible inference paths are enumerated through the above-mentioned interpretable inference path selection and learning algorithm to find the inference path that meets the Pareto optimality; finally, the optimal inference path is selected from the inference paths that meet the Pareto optimality.

[0018] A method for generating an interpretable inference path across meso-regions for invoice fraud detection, comprising the following steps:

[0019] 1) On the basis of obtaining a candidate set of tax data, that is, path nodes, in the meso-region, an interpretable measurement method is designed according to the characteristics of the path nodes in the meso-region;

[0020] 2) According to tax knowledge and the interpretable measurement method, an interpretable loss function is designed, and on this basis, an inference path selection and learning algorithm for maximizing the interpretability of the path under the conditions of meeting the accuracy constraint and the timeliness constraint is realized;

[0021] 3) All possible inference paths are enumerated through the above-mentioned interpretable inference path selection and learning algorithm to find the inference path that meets the Pareto optimality;

[0022] 4) The optimal inference path is selected from the inference paths that meet the Pareto optimality.

[0023] The further improvement of the present invention lies in that the specific implementation method of step 1) is as follows:

[0024] Based on the mesoscience theory, the tax data system with multi-level and multi-scale dynamic spatio-temporal correlation is divided into several meso-regions, forming a hierarchical structure; aiming at the characteristics of tax data in the meso-region and for the convenience of forming an evidence chain, the tax data in the meso-region is now represented in the form of first-order predicate logic;

[0025] step101. Logic replacement index σ:

[0026] Measure the coupling degree between the inference steps of the model and external tax knowledge. To measure the coupling degree between the inference steps and external knowledge, a coupling degree evaluation algorithm based on the BLEU algorithm is used. The specific formula is as follows:

[0027]

[0028] Among them, the inference step is expressed as The external tax knowledge is expressed as Indicates the number of occurrences of the k-th n-gram in And Indicates the number of occurrences of the k-th n-gram in ;

[0029] step102. Symbol interpretability e i

[0030] Indicates the comprehensibility of the rules contained in the meso-region. By means of manual 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;

[0031] step103. Correlation r between symbols ij

[0032] The algorithm is 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 meso-regions they are in are continuous, then represents the confidence level that propagates from symbol j to symbol i and is between meso-regions m and n, where n = m + 1, and the propagation rule is:

[0033]

[0034] Among them, α - β = 1, β ≥ 0, and j represents one of the symbols in the layer before the layer where symbol k is located;​

[0035] step104. Inference path length d

[0036] During the inference process, there may be multiple inference paths from the input to the output. The inference path length d can be defined as the sum of the number of paths between symbols in the inference behavior, that is: d = ∑w ij .

[0037] A further improvement of the present invention is that the higher the value of σ, the higher the degree of understandability.

[0038] A further improvement of the present invention is that the smaller the path length, the fewer the rules involved, and the higher the interpretability.

[0039] A further improvement of the present invention is that the specific implementation method of step 2) is as follows:

[0040] Step201. Construct an interpretability metric loss function

[0041] According to the four metric indicators defined in step1, the interpretability metric of the model can be defined as:

[0042]

[0043] where ε is a logical value that humans can understand, used to represent the interpretability of the model. The loss function of interpretability is defined in the following form:

[0044]

[0045] Step202. Construct an accuracy loss function

[0046] The output form of the model is a sequence form of "symbol" → "symbol". The accuracy loss function is constructed based on the Levenshtein distance. Assume that the inference path output by the model is W = {s1, s2, s3, …, s m}, and the true label path is where s1 and both represent symbols. The distance between the inference path output by the model and the true label path can be expressed as:

[0047]

[0048] Then the accuracy loss function is defined as:

[0049]

[0050] Step203. Construct a timeliness loss function

[0051] In the actual forensics process, the validity of general evidence gradually decreases over time, thus giving rise to the concept of "timeliness". To fully ensure timeliness while also guaranteeing interpretability, each symbol s has the following attribute values: occurrence time ot, forensics time t, and validity period l;

[0052] The occurrence time ot represents the time when the real meaning represented by symbol s appears; the forensics time t represents the time required to obtain or verify the real meaning represented by symbol s in reality; the validity period l represents the time from the appearance time of the real meaning represented by symbol s to the time when it becomes impossible to obtain evidence or the difficulty of obtaining evidence surges; the specific values of the attribute values t and l are determined by expert evaluation;

[0053] Based on the above attribute values, the time difference between the validity period and the time to complete evidence collection is defined as the time gain. Assuming that the inference path, i.e., the evidence chain, generated now is W = {s1, s2, s3, …, s m}, then the time gain of symbol s i is:

[0054]

[0055] The gain at the path level is obtained here as:

[0056]

[0057] Similarly, assuming the true label path is then the gain at the path level of the true label path is:

[0058]

[0059] Then the timeliness loss function is defined as:

[0060]

[0061] Step204. Let be the loss function of the above interpretability metric, precision, and timeliness objectives, then the total loss function is:

[0062]

[0063] where ω i is the weight of the i-th loss function and satisfies ω i > 0, i ∈ {1, 2, 3}. At the same time, a constraint μ i is introduced as the lower limit of the weight of the i-th loss function, and ω i ≥ μ i ,

[0064] A further improvement of the present invention is that the higher the ε, the stronger the interpretability of the model.

[0065] A further improvement of the present invention is that the specific implementation method of step 3) is as follows:

[0066] Step301. Under the KKT criterion, transform the total loss function into:

[0067]

[0068] After the transformation, the direction of the Pareto efficient solution is the direction in which the gradient descent is zero. Under this condition, the algorithm is divided into two parts: first train the framework parameter θ once using evenly divided scalar weights;

[0069] Step302. Use the weight training algorithm to automatically update the weight ω i , loop and execute the above steps, alternately update the framework parameter θ and the weight ω i ; where the update of the weight ω i involves a quadratic programming problem. Under quadratic programming, the Pareto efficiency condition changes to:

[0070]

[0071] which involves equality constraints and inequality constraints. First, only consider the loose equality constraint problem;

[0072] Step303. Introduce a projection process to generate an effective solution in the set of feasible solutions under all constraints, transform the problem into a non - negative least - squares problem, and solve it using an effective method

[0073]

[0074] Step304. Finally, after alternately updating the weights ω of each objective i , make the three optimization objectives of interpretability, accuracy, and timeliness reach equilibrium. At this time, set a series of different lower limits μ of the objective weights i to obtain a solution set that conforms to the Pareto efficiency.

[0075] A further improvement of the present invention is that the specific implementation method of step 4) is as follows:

[0076] Step401. Select a relatively fair solution as the optimal solution from this solution set through the Technique for Order Preference by Similarity to an Ideal Solution (TOPSIS) based on the entropy weight method as the final inference path; the so-called TOPSIS based on the entropy weight method is to first calculate the objective weights of the indicators by the entropy weight method, and then use the TOPSIS method to evaluate each evaluation object. Based on the TOPSIS, the weights of different indicators are calculated by the entropy weight method. Among them, the timeliness loss function, the accuracy loss function, and the interpretability loss function are all cost-type indicators.

[0077] The present invention has at least the following beneficial technical effects:

[0078] The method for generating an interpretable inference path across different granularity regions for invoice fraud detection provided by the present invention realizes the construction of an interpretable measurement scheme for tax data between and within different granularity regions on the basis of dividing the tax data system into several granularity regions, designs an interpretability loss function for the evidence chain of invoice fraud, overcomes the subjective bias problem existing in the existing interpretable measurement methods, and fundamentally breaks through the interpretability bottleneck of the existing black box model; and converts the problem of arranging and selecting fragmented paths into a multi-objective optimization problem. By adjusting the boundaries of different objectives, an optimal inference path with multi-objective balance is selected to form an evidence chain of invoice fraud and realize the interpretability of the model. The method for generating an interpretable inference path across different granularity regions for invoice fraud detection provided by the present invention will improve the interpretability of invoice fraud detection and provide a reliable and accurate "evidence chain". Description of the Drawings

[0079] Figure 1 It is the overall flowchart for generating an inference path across different granularity regions.

[0080] Figure 2 It is the schematic diagram of the propagation of symbol correlation in different granularity regions of the present invention.

[0081] Figure 3 It is the schematic diagram of the initial input meta-path.

[0082] Figure 4 It is to generate an inference path that satisfies the Pareto optimal solution.

[0083] Figure 5 It is the schematic diagram of the process for detecting invoice fraud and generating an inference evidence chain.

[0084] Figure 6 It is the flowchart of the interpretable inference path. Detailed Embodiments

[0085] In the following, taking the control mechanism of enterprise data items registered in a certain region's national tax from 2015 to 2017 as an example, the specific implementation manner will be described in combination with the accompanying drawings. The enterprise data item control mechanism is used as an input item to detect whether there is any behavior of false invoicing by enterprises, and at the same time generate an inference path across regions to provide an interpretable evidence chain for the inference result.

[0086] step1. Define an interpretability measurement method. In order to overcome the subjective bias problem existing in the existing interpretability measurement methods, the present invention proposes an interpretability measurement method combining subjectivity and objectivity. It mainly includes four measurement indicators:

[0087] step101. Logical replacement index σ: It measures the coupling degree between the model inference steps and external tax knowledge. In order to measure the coupling degree between the inference steps and external knowledge, the present invention proposes a coupling degree evaluation algorithm based on the BLEU algorithm. The specific formula is as follows:

[0088]

[0089] where the inference step is represented as the external tax knowledge is represented as represents the number of times the k-th n-gram appears in , represents the number of times the k-th n-gram appears in ; the higher the value of σ, the higher the degree of understandability.

[0090] Suppose there is an existing inference path: "Inventory anomaly" → "Closed-loop shareholding" → "Profit transfer". For convenience, it is described in words, and the same below. There are multiple existing external tax knowledge, such as: "Inventory anomaly" → "Closed-loop shareholding" → "One address with multiple licenses" → "Profit transfer". Let the inference path be operated with all tax knowledge to obtain an evaluation value, and select the highest value as the final logical replacement index. The logical replacement index value calculated from the inference path and external tax knowledge in the above text is 0.5. Suppose it is the highest logical replacement index, so 0.5 is selected as the logical replacement index of this inference path. step102. Symbolic interpretability e i : It represents the understandability of the rules contained within the region. A symbolic interpretability dictionary is formulated by means of manual assignment 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.

[0091] Now assume that there is a path with symbols interlockCtrl(X,Y), pnCtrl(X), amountCtrl(X), and purchaseCtrl(X). After relevant evaluation by experts, the following symbol interpretability evaluation table can be obtained, namely dict(s,v).

[0092]

[0093] step103. Correlation r between symbols ij The algorithm can be expressed as: the higher layer layer of the model k Score R k = ∑ 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:

[0094]

[0095] 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.

[0096] For example Figure 1 As shown, at the current stage, multiple internal paths, i.e., meta-paths, have been identified within the medium region. When selecting a meta-path in each medium region to connect and form a complete inference path, the correlation r between symbols is required ij to measure the path between medium regions. There are 4 nodes with an out-degree of 0 in medium region M1 that can be connected to 4 nodes with an in-degree of 0. Assume that a path has been formed between medium region M1 and medium region M2. Then, according to the above calculation formula, the correlation r between the symbols in medium region M1 and medium region M2 is obtained ij .

[0097] step104. Inference path length d: During the inference process, there may be multiple inference paths from the input to the output. The inference path length d can be defined as the sum of the number of symbol paths passed by the inference behavior, i.e.: d = ∑w ij . The smaller the path length, the fewer the rules involved, and the higher the interpretability.

[0098] Assume that the final inference path has now been formed, that is, the paths between medium regions have been connected, and assume the path is "Product name does not match" → "Instant invoice" → "Purchase of gold" → "Sale of mobile phones" → "Inventory anomaly" → "Closed-loop shareholding" → "Multiple business licenses at one address" → "Profit transfer", then the path length is 7.

[0099] Step2. Generate the optimal inference path.

[0100] The main goal of this part is to maximize the interpretability of the path under the conditions of meeting the precision constraint and timeliness constraint. Since the interpretability goal can be decomposed into multiple sub-goals, the optimal inference path generation is transformed into a multi-step multi-objective combinatorial optimization problem; and an algorithm is constructed under the KKT conditions. On the premise of theoretical guarantee, scalarized weights are produced to transform the multi-objective problem into a single-objective problem.

[0101] Step201. Construct an interpretability metric loss function. According to the four metric indicators defined in step1, the interpretability metric of the model can be defined as:

[0102]

[0103] where ε is a logical value that humans can understand, used to represent the interpretability of the model. The higher ε is, the stronger the interpretability of the model. Therefore, the loss function of interpretability can be defined in the following form:

[0104]

[0105] Step202. Construct a precision loss function

[0106] The output form of the model of the present invention is a sequence form of "symbol" → "symbol", such as ecStartup(X) → freeTax(X). The precision loss function is constructed based on the Levenshtein distance. Assume that the inference path output by the model is W = {s1, s2, s3, …, s m}, and the true label path is where s1 and both represent symbols. The distance between the inference path output by the model and the true label path can be expressed as:

[0107]

[0108] Then the precision loss function is defined as:

[0109]

[0110] Step203. Construct a timeliness loss function

[0111] In the actual evidence collection process, the validity of general evidence will gradually decrease over time, so the concept of "timeliness" emerges. In order to fully ensure timeliness while ensuring interpretability, the symbols s in the present invention all have attribute values: occurrence time ot, evidence collection time t, and validity period l.

[0112] The occurrence time ot represents the time when the real meaning represented by the symbol s appears. The evidence collection time t represents the time required to collect or verify the real meaning represented by the symbol s in reality. The validity period l represents the time from the appearance time of the real meaning represented by the symbol s to the time when it is impossible to collect evidence or the difficulty of evidence collection surges. The specific values of the attribute values t and l are determined by expert evaluation.

[0113] According to the above attribute values, the time difference between the validity period and the time to complete evidence collection is defined as the time benefit. Assuming that the inference path, i.e., the evidence chain, has been generated as W = {s1, s2, s3, …, s m}, then the time benefit of the symbol s i is:

[0114]

[0115] The benefit at the path level can be obtained here as:

[0116]

[0117] Similarly, assuming the true label path is then the benefit at the path level of the true label path is:

[0118]

[0119] Then the timeliness loss function is defined as:

[0120]

[0121] Step3. Let be the loss function of the above interpretability metric, precision, and timeliness objectives. Then the total loss function is:

[0122]

[0123] where ω i is the weight of the i-th loss function and satisfies ω i > 0, i ∈ {1, 2, 3}. At the same time, a constraint μ i is introduced as the lower bound of the weight of the i-th loss function, and ω i ≥ μ i ,

[0124] Step301. Construct an algorithm under the KKT conditions. On the premise of theoretical guarantee, generate scalarized weights to transform the multi-objective problem into a single-objective problem, which can ensure Pareto efficiency.

[0125] Under the KKT criterion, the total loss function is transformed into:

[0126]

[0127] After the transformation, the direction of the Pareto efficient solution is obtained, that is, the direction where the gradient descends to zero. Under this condition, the algorithm is mainly divided into two parts: first, the framework parameters θ are trained once using the equally distributed scalar weights;

[0128] Step 302. Automatically update weight ω using weight training algorithm i , loop through the above steps, and update the framework parameters θ and weights ω alternately i . The weight ω i The update involves the quadratic programming problem, under which the Pareto efficiency condition changes to:

[0129]

[0130] This involves equality constraints and non-equality constraints. First, we only consider the problem of loose equality constraints.

[0131] Step 303. Introduce a projection process to generate a valid solution from the set of feasible solutions under all constraints, transforming the problem into a non-negative least squares problem that can be solved using an effective method.

[0132]

[0133] Step 304. Finally, the weights of each target are updated alternately i , so that the three optimization goals of interpretability, accuracy and timeliness are balanced. At this time, a series of different target weight lower limits μ can be set i A solution set that meets Pareto efficiency can be obtained

[0134] Step 4. Solution selection. When a single solution is expected, the approach to the ideal solution sorting method based on the entropy weight method can be used to select the maximum value of the objective loss function and select a relatively fair solution as the optimal solution in the solution set as the final reasoning path. Figure 4 As shown in this embodiment, two paths that meet the Pareto optimal solution are generated through the multi-objective optimization selection algorithm, which are

[0135] (1) “Fake exchange” → “Instant invoice” → “Business acquisition” → “Request for invoice” → “Separation of invoice and goods” → “Massive false invoicing”

[0136] (2) “Product name discrepancy” → “Instant invoice” → “Purchase of gold” → “Selling of mobile phones” → “Inventory anomaly” → “Closed-loop shareholding” → “Multiple licenses at one address” → “Profit transfer”

[0137] By using the Technique for Order Preference by Similarity to an Ideal Solution (TOPSIS) based on the entropy weight method, path 2 is selected as the optimal solution, which is the inference path generated by the model. An interpretable evidence chain is generated, which is "Discrepancy in Product Name" → "Instant Invoice" in the M1 medium region, then "Purchased Gold" → "Sold Mobile Phones" → "Abnormal Inventory" in the M2 medium region, and finally "Closed-loop Shareholding" → "Multiple Business Licenses at One Address" → "Profit Transfer" in the M3 medium region.

[0138] Although the present invention has been described in detail with general descriptions and specific embodiments above, 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. A method for generating an interpretable inference path across media regions for invoice fraud detection, characterized in that It includes the following steps: 1) Based on obtaining the candidate set of tax data, i.e., path nodes, in the meso-region, a method for interpretable measurement is designed according to the characteristics of path nodes in the meso-region. The specific implementation method is as follows: Based on the theory of mesoscience, the tax data system with multi-level and multi-scale dynamic spatio-temporal correlation is divided into several meso-regions, forming a hierarchical structure. In view of the characteristics of tax data in the meso-region and for the convenience of forming an evidence chain, the tax data in the meso-region is represented in the form of first-order predicate logic. step101. Logical replacement index σ To measure the coupling degree between the inference steps of the model and external tax knowledge, a coupling degree evaluation algorithm based on the BLEU algorithm is used. The specific formula is as follows: where the reasoning step is represented as denoted as the i-th reasoning step, and the external tax knowledge is represented as k is the n-gram subscript, representing the k-th n-gram; denotes the number of times the k-th n-gram appears in and denotes the number of times the k-th n-gram appears in ; step102. Symbol interpretability e i Indicates the comprehensibility of the rules contained in the intermediate region. By means of manual assignment, an interpretability dictionary of symbols 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, v represents the manual assignment of the interpretability of the symbol, and dict is the set of all symbols and their corresponding manual assignments; step103. Correlation r between symbols ij The algorithm is expressed as: the higher layer of the model, layer k Score R k = ∑ j∈N R j←k , where N is the total number of mediation regions. If there exist symbols i and j and the mediation regions they are in are continuous, then represents the confidence propagated from symbol j to symbol i and between mediation 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. step104. Inference path length d During the inference process, there may be multiple inference paths from the input to the output. The length d of the inference path can be defined as the sum of the number of paths between symbols passed by the inference behavior, that is: d = ∑w ij , where w ij represents the distance between the i-th symbol and the j-th symbol; 2) According to tax knowledge and the interpretable measurement method, an interpretable loss function is designed, and on this basis, an inference path selection learning algorithm that maximizes the interpretability of the path under the conditions of meeting the accuracy constraint and the timeliness constraint is implemented. 3) Enumerate all possible inference paths through the interpretable inference path selection learning algorithm in the previous step to find the inference path that meets the Pareto optimality. 4) Select the optimal inference path from the inference paths that meet the Pareto optimality.

2. The cross-media region interpretable inference path generation method for invoice fraud detection according to claim 1, wherein The higher the value of σ, the higher the degree of understandability.

3. A method for generating an interpretable inference path across media regions for invoice fraud detection according to claim 1, characterized in that The smaller the path length, the fewer rules are involved, and the higher the interpretability.

4. A method for generating an interpretable inference path across media regions for invoice fraud detection according to claim 1, characterized in that The specific implementation method of step 2) is as follows: Step201. Construct an interpretive measurement loss function According to the four measurement indexes defined in step101 - 104, the interpretability measurement of the model is defined as: where ε is a logical value that humans can understand, used to represent the interpretability of the model. The loss function of interpretability is defined in the following form: θ is a framework parameter; Step202. Construct an accuracy loss function The output form of the model is a sequence of "symbol" → "symbol". The loss function of accuracy is constructed based on the Levenshtein distance. Assume that the inference path output by the model is W = {s1, s2, s3, …, s m}, and the true label path is where s1 and both represent symbols. The distance between the inference path output by the model and the true label path can be expressed as: Then the accuracy loss function is defined as: Step203. Construct a timeliness loss function The symbol s has attribute values: occurrence time ot, evidence collection time t, and validity period l; The occurrence time ot represents the time when the real meaning represented by symbol s appears. The evidence collection start time ct represents the start time of real evidence collection of symbol s. The evidence collection time t represents the time required for the real meaning represented by symbol s to be evidenced or verified in reality. The validity period 1 represents the time from the appearance time of the real meaning represented by symbol s to the time when evidence collection is impossible or the difficulty of evidence collection surges. The specific values of the attribute values t and 1 are determined by expert evaluation. Define the time benefit as the difference between the expiration date and the time to complete evidence collection according to the above attribute values. Assume that the inference path, i.e., the evidence chain, has been generated as W = {s1, s2, s3, …, s m}, then the time benefit of symbol s i is: where t k represents the k-th evidence collection time; the path-level benefit obtained here is: Similarly, assuming the true label path is The path-level gain of the true label path is: Then the timeliness loss function is defined as: Step204. Set as the loss function for the above interpretability metric, precision, and timeliness objectives, then the total loss function is: where ω i is the weight of the i-th loss function and satisfies ω i > 0, i ∈ {1, 2, 3}. At the same time, a constraint μ i is introduced as the lower bound of the weight of the i-th loss function, and ω i ≥ μ i , 5. A method for generating an interpretable inference path across different media regions for invoice fraud detection according to claim 4, characterized in that The higher the ε, the stronger the interpretability of the model.

6. A method for generating an interpretable inference path across media regions for invoice fraud detection according to claim 4, characterized in that, The specific implementation method of step 3) is as follows: Step301. Under the KKT criterion, transform the total loss function into: Among them denotes the partial derivative with respect to θ, K represents the total number of loss functions. After transformation, the direction of the Pareto efficient solution is the direction where the gradient descent is zero. Under this condition, the algorithm is divided into two parts: first, train the framework parameter θ once using evenly divided scalar weights; Step302. Automatically update the weight ω using the weight training algorithm i , loop to execute Step301, and alternately update the framework parameter θ and the weight ω i ; where the update of the weight ω i involves a quadratic programming problem. Under quadratic programming, the Pareto efficiency condition changes to: Among them represents the weight during the training process; it involves equality constraints and inequality constraints. First, only consider the problem of loose equality constraints; Step303. Introduce a projection process to generate a valid solution in the set of feasible solutions under all constraints, transform the problem into a non-negative least squares problem, and use an effective method to solve it wherein is the true weight; Step 304. Finally, alternately update the weights ω of each objective i , so that the three optimization objectives of interpretability, accuracy, and timeliness reach an equilibrium. At this time, set a series of different lower bounds μ of the objective weights i to obtain a solution set that conforms to Pareto efficiency.

7. A method for generating an interpretable inference path across media regions for invoice fraud detection according to claim 6, characterized in that The specific implementation method of step 4) is as follows: Step401. Select a relatively fair solution as the optimal solution from this solution set through the Technique for Order Preference by Similarity to an Ideal Solution (TOPSIS) method based on the entropy weight method as the final inference path. The so-called TOPSIS method based on the entropy weight method first calculates the objective weights of the indicators by the entropy weight method, and then uses the TOPSIS method to evaluate each evaluation object. Based on the TOPSIS method, the weights of different indicators are calculated by the entropy weight method. Among them, the timeliness loss function, the accuracy loss function, and the interpretability loss function are all cost-type indicators.

Images