Asymmetric audit data reconciliation method based on amount feature bucketing and greedy strategy

Abstract

The application discloses an asymmetric audit data reconciliation method based on amount feature bucketing and a greedy strategy, adaptive semantic reconstruction and context enhancement, and constructs a data preprocessing layer with self-healing capability: an amount fingerprint-based hyperplane dimension reduction, an amount hash bucketing technology is introduced, a high-dimensional Cartesian product search space is compressed to a discrete low-dimensional subspace: multi-dimensional orthogonal constraint pruning, within the feature bucket, the system performs strict "zero tolerance" logical filtering, and 99% of the false matches are removed through three orthogonal dimensions: a confidence evaluation model of three-way feature fusion, for the candidate pairs that pass the pruning, a nonlinear multi-modal scoring matrix is constructed: a maximum coverage priority global greedy locking, a reverse thinking scheduling strategy is adopted. The application aims to reconstruct a deterministic audit evidence chain from chaotic transaction flow in an extremely short time window, and realizes mathematical dimension reduction from exponential complexity to quasi-linear complexity.

Description

Technical Field

[0001] This invention belongs to the field of financial technology, specifically an asymmetric audit data reconciliation method based on amount feature binning and greedy strategy. Background Technology

[0002] In the daily operations of financial auditing, third-party payment verification, and large enterprise financial shared service centers, the reconciliation of bank statements with internal accounting books is a core element in ensuring the safety of funds and the accuracy of financial statements.

[0003] With the diversification of electronic payment methods and the increasing complexity of business operations, the proportion of traditional "1:1 exact matching" scenarios (i.e., a bank transaction corresponds to an accounting voucher, with the amount and date being completely identical) is declining year by year. The more challenging phenomenon of **"asymmetric matching" (i.e., N:M matching)** is becoming increasingly common. For example:

[0004] 1. 1:N scenario: The enterprise records a total remittance, while the bank splits it into multiple transaction fees and principal payments due to cross-bank or channel reasons.

[0005] 2. N:1 scenario: The sales revenue of multiple stores on the same day is combined into a single total amount deposited into the bank.

[0006] 3. N:M scenario: The most complex combination of fund collection, splitting or multi-party transactions, which requires multiple records to be combined on both sides before the transaction can be balanced.

[0007] Currently, existing technologies mainly employ the following approaches to address such complex asymmetric reconciliation problems, but all of them have significant drawbacks:

[0008] 1. Traditional rule-based matching systems:

[0009] Such systems typically have pre-defined matching logic (e.g., "equal amounts and dates within 3 days").

[0010] Limitations: It can only handle simple 1:1 or highly standardized 1:N scenarios. When faced with cross-month transactions, unclear summaries, or complex N:M combination transactions, the rules quickly become invalid, resulting in a large amount of data being missed because the rules cannot be followed, ultimately requiring time-consuming manual verification of each transaction.

[0011] 2. A brute-force search-based combinatorial matching method:

[0012] Try to perform mathematical permutations and combinations on the records in the unmatched pool to calculate whether the sum of the amounts is equal.

[0013] Drawback: The computational complexity explodes exponentially. When the number of unmatched records exceeds several dozen, the computational load will exceed the capacity of a regular server, leading to system crashes or unacceptable runtime, making it unsuitable for engineering practices involving massive amounts of data.

[0014] 3. Prediction methods based on general AI models:

[0015] Use deep learning or machine learning models to predict the matching probability of trading pairs.

[0016] Limitations: The models are typically "black boxes," lacking the absolute explainability required for auditing. Furthermore, training data is extremely expensive, and the models struggle to generalize to the accounting practices of different banks and companies.

[0017] Summary of the main drawbacks of existing technologies:

[0018] In addition to the limitations of the aforementioned approaches themselves, existing technologies also suffer from the following deeper problems in practical applications, making it difficult to meet the high standards required by modern auditing:

[0019] 1. Poor adaptability to "dirty data" and non-standard data (data quality bottleneck):

[0020] Existing technologies typically assume that input data is highly structured and accurate. However, the data sources in actual audit working papers are diverse, often containing "dirty data" such as OCR recognition errors, missing key fields (e.g., empty counterparty name), and disordered date formats (missing years). Traditional algorithms often directly report errors or generate a large number of incorrect matches when encountering such data, lacking adaptive cleaning and semantic completion capabilities.

[0021] 2. A high false positive rate poses serious audit compliance risks:

[0022] In an auditing context, false positives pose a greater risk than no matches because they create erroneous audit evidence. Many existing algorithms that prioritize high matching rates tend to relax constraints (e.g., only considering equal monetary combinations while ignoring compliance of lending directions or semantic relevance of summaries). This results in reconciliation results that appear to balance out monetary amounts but are actually flawed in their business logic, significantly increasing the workload for auditors in subsequent manual review and false positive removal, and severely impacting audit quality.

[0023] 3. Lack of dynamic equilibrium capability with multidimensional constraints:

[0024] Effective reconciliation requires simultaneously satisfying multiple constraints, including "strictly equal amounts," "similar dates," "correct lending / borrowing direction," and "semantic relevance of summaries." Current technologies struggle to balance these dimensions: either sacrificing semantic and date constraints for computational speed (leading to haphazard matching) or imposing too many constraints, resulting in extremely slow computation. There is a lack of an efficient mechanism to maintain computational performance while imposing strict financial constraints.

[0025] 4. Lack of optimal solution selection strategy, prone to getting trapped in local optima:

[0026] When a set of data objectively has multiple possible combinations for matching (e.g., records A+B=E, and C+D=E, but in reality A should match C, and B should match D in another combination), existing technologies often adopt a "first-come, first-served" or random selection strategy. This approach, lacking a global perspective, easily leads to simpler combinations occupying key data first, making it impossible to achieve subsequent more complex and necessary matches, thereby reducing the overall data matching coverage and accuracy. Summary of the Invention

[0027] This invention aims to overcome the long-standing NP-hard (non-deterministic polynomial difficulty) combinatorial optimization problem of "N:M reconciliation of unstructured data" in the field of financial auditing. Addressing the "computing power collapse" and "logical distortion" issues that arise when existing technologies deal with massive amounts of dirty data, this invention provides an asymmetric audit data reconciliation method based on amount feature binning and a greedy strategy. This method is based on global greedy optimization using Discrete Amount Fingerprinting and Orthogonal Constraint Pruning, aiming to reconstruct a deterministic audit evidence chain from chaotic transaction flows within a very short time window, achieving a mathematical dimensionality reduction from exponential complexity to quasi-linear complexity.

[0028] The technical solution of this invention is as follows: an asymmetric audit data reconciliation method based on amount feature binning and a greedy strategy, which constructs a closed-loop intelligent reconciliation calculation engine, comprising the following five tightly coupled steps:

[0029] Step S1: Adaptive semantic reconstruction and context enhancement construct a data preprocessing layer with "self-healing capabilities":

[0030] Step S2: Based on the hyperplane dimensionality reduction of monetary fingerprints, a monetary hashing bucketing technique is introduced to compress the high-dimensional Cartesian product search space into a discrete low-dimensional subspace.

[0031] Step S3: Multidimensional orthogonal constraint pruning. Within the feature bucket, the system performs strict "zero-tolerance" logical filtering, eliminating 99% of false matches through three orthogonal dimensions.

[0032] Step S4: The confidence evaluation model based on three-way feature fusion constructs a non-linear multimodal scoring matrix for pruned candidate pairs:

[0033] Step 5: Global greedy locking with maximum coverage priority, employing a reverse-thinking scheduling strategy.

[0034] Preferably, step S1 is as follows:

[0035] S11 Missing Metadata Inference: To address the common issue of missing account names in bank statements, the engine does not rely on external dictionaries. Instead, it constructs a dynamic mapping graph based on transaction polarity, i.e. the positive or negative sign of the amount, and automatically maps the "borrow / loan" logic to the "receiver / payer" field, achieving a 100% completion rate for key fields.

[0036] S12, Timeline Anchor Point Repair: For time fragments containing only "month-day", a heuristic algorithm combining global time window and local context is adopted; first, the document-level annual fingerprint is locked, and if it fails, backtracking inference is performed based on the file creation time or the current accounting period to ensure that each transaction can be accurately anchored on the correct time axis.

[0037] Preferably, step S2 is as follows:

[0038] S21. Combined Feature Hashing: Defining a Sliding Window For unmatched sets, the system generates full permutations and combinations. Instead of directly comparing the combinations, it calculates the scalar fingerprint of the combination, which is the absolute value of the total amount rounded down to the smallest monetary unit (cents). The sliding window... The value range is usually from 1 to 5, which is used to balance computing power and matching success rate;

[0039] S22. Hash Collision Exploitation: Utilizing the natural aggregation property of hash buckets, all potential combinations with equal "algebraic sums of amounts" are forcibly mapped to the same physical storage bucket. This reduces the time complexity of the search algorithm from... The exponential brute-force exhaustive search was instantly reduced to The polynomial level fundamentally solves the computing power bottleneck.

[0040] Preferably, step S3 is as follows:

[0041] S31. Vector Direction Constraint: This mandates that all non-zero sub-vectors within a combination must be isomorphic in the direction of capital flow, completely eliminating the logical loophole of "positive and negative offsetting to make up numbers".

[0042] S32. Spatiotemporal Neighborhood Constraint: Introducing the principle of temporal locality, it is mandatory that the transaction time of all elements in the portfolio must converge within the same accounting month or a preset ∆ window;

[0043] S33. Semantic consistency constraint: Calculate the semantic entropy of the summary within the combination using a lightweight NLP algorithm, and eliminate random combinations with excessive semantic dispersion.

[0044] Preferably, the formula for the multimodal scoring matrix in step S4 is as follows: .

[0045] Preferably, step S4 is as follows:

[0046] S41, integrates micro-character features, statistical distribution features, and macro-semantic features;

[0047] S42. Introduce a complexity penalty factor to reduce the weight of combinations involving too many fragments. By introducing the complexity penalty factor, the solution space with fewer fragments is converged first, thereby reducing the overall depth of the system's recursive search and improving the utilization of computing resources.

[0048] Preferably, step S5 is as follows:

[0049] S51, Priority Queue Scheduling: The system maintains a dynamic max-heap, whose sorting key is not just the matching score, but the "record coverage";

[0050] S52, Global Optimal Approximation: The algorithm forces priority processing of complex combinations involving the most records;

[0051] S53. Atomic Lock Mechanism: Once the optimal solution at the top of the heap is confirmed, the system immediately locks all the atomic records involved to ensure that they are not reused by subsequent suboptimal solutions, thus guaranteeing the uniqueness and exclusivity of the final reconciliation result.

[0052] By adopting the above technical solution, the present invention has the following beneficial effects compared with the prior art:

[0053] 1. Order-of-magnitude performance leap: Compared with deep learning-based prediction models, this invention does not require GPU training and has zero cold start time. Compared with traditional brute-force search, this solution uses monetary fingerprint binning technology to improve the processing speed by 2-3 orders of magnitude (from several hours to seconds) when processing N:M matching with tens of thousands of data points.

[0054] 2. Audit-Grade Integrity: This invention abandons the probabilistic guessing of black-box models and internalizes financial logic into the algorithm boundary through orthogonal constraint pruning. Each matching result output by the system has strong mathematical consistency in four dimensions: amount, date, direction, and semantics, with a false positive rate approaching zero.

[0055] 3. Strong robustness to extremely "dirty data": Through the built-in context inference engine and multimodal text similarity algorithm, this system can penetrate noise interference such as OCR recognition errors, differences in summary abbreviations, and missing key fields, and still maintain a high recall rate in environments with extremely poor data quality;

[0056] 4. Optimization of global resource allocation: The maximum coverage priority strategy mathematically guarantees that the system always prioritizes solving the matching task with the "highest entropy" (most complex), so that the remaining simple tasks naturally fall into the 1:1 matching interval, maximizing the overall data matching rate (MatchRate). Attached Figure Description

[0057] Figure 1 This is a flowchart illustrating the overall logic of the core algorithm of this invention.

[0058] Figure 2 This is a schematic diagram of the hash bucketing principle based on monetary fingerprints in this invention;

[0059] Figure 3 A comparative diagram of greedy strategies prioritizing maximum coverage. Detailed Implementation

[0060] The embodiments of the present invention are described in detail below: These embodiments are implemented based on the technical solution of the present invention, and provide detailed implementation methods and specific operation processes. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention.

[0061] Example 1

[0062] An asymmetric audit data reconciliation method based on monetary feature binning and a greedy strategy is proposed, which constructs a closed-loop intelligent reconciliation calculation engine. Its core processing logic includes the following five tightly coupled steps:

[0063] Step 1: Adaptive Semantic Reconstruction and Context Enhancement (DataContextReconstruction)

[0064] Unlike traditional simple cleaning, this step constructs a data preprocessing layer with "self-healing capabilities":

[0065] 1.1 Missing Metadata Inference: To address the common issue of missing account names in bank statements, the engine does not rely on external dictionaries. Instead, it constructs a dynamic mapping graph based on transaction polarity—the positive or negative sign of the amount—and automatically maps the "borrow / loan" logic to the "receiver / payer" field, achieving a 100% completion rate for key fields.

[0066] 1.2 Timeline Anchor Point Repair: For time fragments containing only "month-day", a heuristic algorithm combining a global time window and a local context is employed. First, the document-level annual fingerprint is locked. If this fails, backtracking is performed based on the file creation time or the current accounting period to ensure that each transaction is accurately anchored to the correct timeline.

[0067] Step 2: Hyperplane Dimensionality Reduction Based on Financial Fingerprint

[0068] In order to solve To address the combinatorial explosion problem, this invention introduces a monetary hashing bucketing technique, compressing the high-dimensional Cartesian product search space into a discrete low-dimensional subspace:

[0069] 2.1 Combinatorial Feature Hashing: Defining a Sliding Window The system generates all permutations and combinations for the unmatched set. Instead of directly comparing combinations, the system calculates the scalar fingerprint of the combination—the absolute value of the total amount rounded down to the smallest monetary unit (cent).

[0070] 2.2 Hash Collision Exploitation: Leveraging the inherent aggregation property of hash buckets, all potential combinations with equal "algebraic sums of amounts" (regardless of the number of records they consist of) are forcibly mapped to the same physical storage bucket. This reduces the time complexity of the search algorithm from... The exponential brute-force exhaustive search was instantly reduced to The polynomial level fundamentally solves the computing power bottleneck.

[0071] Step 3: Multi-Dimensional Orthogonal Pruning

[0072] Within the feature bucket, the system performs strict "zero-tolerance" logical filtering, eliminating 99% of false matches through three orthogonal dimensions:

[0073] 3.1 Vector Direction Constraint: It is mandatory that all non-zero sub-vectors (transaction records) in the combination must be isomorphic (all positive or all negative) in the direction of fund flow, completely eliminating the logical loophole of "positive and negative canceling out to make up the numbers".

[0074] 3.2 Spatiotemporal Neighborhood Constraint: The principle of temporal locality is introduced, which requires that the transaction time of all elements in the portfolio must converge within the same accounting month or a preset $\Deltat$ window.

[0075] 3.3 Semantic consistency constraint: The semantic entropy of the summary within the combination is calculated using a lightweight NLP algorithm, and random combinations with excessively high semantic dispersion (i.e., completely irrelevant) are eliminated.

[0076] Step 4: Tri-Fusion Confidence Scoring

[0077] For candidate pairs that have been pruned, a nonlinear multimodal scoring matrix is ​​constructed: .

[0078] Among them, the weighting coefficient , , It is a constant preset based on the accuracy performance of historical reconciliation datasets, and satisfies... + + = $1. In a preferred embodiment, for financial transaction reconciliation scenarios, since the accuracy of the amount has the highest priority, it can be set to... ∈[0.6,0.8], and The remaining weights are allocated based on time span and abstract richness, respectively.

[0079] 4.1 This model innovatively integrates micro-character features (LevenshteinDistance), statistical distribution features (TF-IDF), and macro-semantic features (SentenceEmbedding).

[0080] 4.2 Introducing a complexity penalty factor (ComplexityPenalty) to reduce the weight of combinations involving too many fragments. By introducing the complexity penalty factor, the solution space with fewer fragments is converged first, thereby reducing the overall depth of the system's recursive search and improving the utilization of computing resources.

[0081] Step 5: Max-Coverage Greedy Locking

[0082] This is the core of the decision-making process in this invention, employing a reverse-thinking scheduling strategy:

[0083] 5.1 Priority queue scheduling: The system maintains a dynamic maxheap, whose sorting key is not just the matching score, but the "total coverage".

[0084] 5.2 Global Optimal Approximation: The algorithm prioritizes complex combinations involving the most records (e.g., 4-to-1, 3-to-2). This strategy avoids the local optimum trap of traditional algorithms that "first consume simple, small-value records (e.g., 100 yuan general reimbursement)," leading to subsequent large, complex transactions failing to match due to a lack of "components."

[0085] 5.3 Atomic Lock Mechanism: Once the optimal solution at the top of the heap is confirmed, the system immediately locks all the atomic records involved (AtomicLock) to ensure that they are not reused by subsequent suboptimal solutions, thus guaranteeing the uniqueness and exclusivity of the final reconciliation result.

[0086] like Figure 1 As shown, the intelligent reconciliation calculation engine of this invention operates on a complete closed-loop principle from data input to evidence output, embodying the systematic engineering of transforming unstructured data into a deterministic audit evidence chain. The process first repairs data quality bottlenecks through adaptive semantic reconstruction and context enhancement modules, and then enters the core dimensionality reduction matching stage. It uses monetary fingerprint binning technology to compress the exponential global search space into a controllable local intra-bucket search, and removes most false matches through strict multidimensional orthogonal constraint pruning. On this basis, an innovative three-way feature fusion scoring model is used to accurately evaluate the confidence of the remaining candidates, and finally locks the optimal solution through a global greedy strategy that prioritizes maximum coverage, achieving efficient convergence and automated evidence generation in complex N:M scenarios.

[0087] like Figure 2 As shown, the core mechanism of this invention for solving the N:M combinatorial explosion problem—based on the mathematical dimensionality reduction principle of monetary fingerprints—demonstrates how the system extracts mathematically invariant "monetary fingerprints" by calculating the absolute value of the algebraic sum of discrete transaction combinations down to the cent. This proactively leverages hash collision characteristics to force potential combinations with different dates and digests but mathematically equal total amounts to be mapped to the same physical storage bucket. This mechanism successfully transforms the massive Cartesian product search, which originally required a global scope, into a local, small-scale search within isolated buckets, fundamentally overcoming the computational bottleneck of large-scale asymmetric reconciliation.

[0088] like Figure 3As shown, by comparing the effects of two different scheduling strategies, this invention intuitively illustrates how it overcomes the local optimum trap of traditional algorithms through reverse thinking to approach the global optimum. The comparison in the figure shows that the traditional strategy of prioritizing simple combinations easily leads to the premature occupation of key records, making subsequent complex N:M combinations unmatchable due to the lack of necessary components. In contrast, the "maximum coverage priority" strategy adopted in this invention forcibly prioritizes locking the most complex combinations involving the largest number of records. This global perspective scheduling strategy not only ensures the priority establishment of the most critical and complex audit evidence but also effectively releases resources, significantly improving the overall data's final matching coverage.

[0089] Comparative Example 1

[0090] A deep learning-based intelligent matching method for bank data employs a data-driven deep learning approach to automate the matching of bank details and bank statements. The specific implementation steps are as follows:

[0091] 1. Data Collection and Construction: Collect bank details and account statements transaction records that have been manually verified in historical audit projects and their final matching results to construct a large-scale supervised training dataset.

[0092] 2. AI Feature Engineering: This involves applying artificial intelligence feature engineering techniques to deeply mine raw transaction data. In addition to extracting basic features such as amount, date, summary, and transaction type, it also constructs derived advanced features (such as time interval patterns and summary semantic vectors).

[0093] 3. Model building and learning: Select advanced deep learning model architectures (such as deep neural networks (DNN), convolutional neural networks (CNN) or attention-based Transformer models) and use training datasets to learn and model the representations of the input transaction features.

[0094] 4. Matching Probability Prediction Training: The training model aims to minimize the prediction error and learns how to accurately predict the matching probability value between any two sets of transaction records that belong to the same business.

[0095] 5. End-to-end inference application: Deploy the trained deep learning model to the production environment, receive new unmatched bank data as input, the model directly performs end-to-end inference prediction, outputs the matching probability of each potential combination, and recommends the final matching result accordingly.

[0096] The beneficial effects are described below:

[0097] 1. Powerful adaptive learning capability: Utilizing the superior non-linear fitting ability of deep learning models, it can automatically learn complex patterns from data and adapt to diverse report formats without the need for manual hard-coding of rules.

[0098] 2. Strong ability to handle complex scenarios: As training data accumulates, the model's ability to generalize to highly non-standardized summaries or complex N:M relationships is stronger than that of traditional methods.

[0099] 3. Discover hidden patterns: It can uncover hidden matching patterns and complex feature combinations that are difficult for human experts to identify.

[0100] 4. Achieve end-to-end automation: Supports fully automated processing from raw data input to final matching result output.

[0101] The difference between this comparative example and Example 1 is as follows:

[0102] 1. Data and computing power dependence: This comparison depends on massive labeled data and GPU computing power; Example 1 has a fast cold start and can run on CPU.

[0103] 2. Explanatory power and transparency: The comparative example has a "black box" problem and weak audit explanatory power; Example 1 is based on clear financial logic and has strong explanatory power.

[0104] Comparative Example 2

[0105] A global optimization search method based on genetic algorithm (GA) treats the reconciliation problem as a complex combinatorial optimization problem and uses a mechanism simulating biological evolution to search for an approximate optimal solution. The specific steps are as follows:

[0106] 1. Encoding and Population Initialization: Encode potential transaction matching relationships as "chromosomes" (e.g., use binary strings to represent whether a record belongs to a certain matching combination), and randomly generate the initial population.

[0107] 2. Fitness Function Definition: Construct an objective function to evaluate the quality of each solution. Fitness metrics typically include: the total number of successfully matched records (the higher the better), the absolute value of the monetary difference (the smaller the better), and the date deviation (the smaller the better), etc.

[0108] 3. Evolutionary Iteration: Through operations such as selection (survival of the fittest), crossover (recombination of parental traits), and mutation (introduction of random perturbations), the natural selection process is simulated, enabling the population to evolve generation by generation.

[0109] 4. Convergence and Output: When the preset number of iterations is reached or the fitness no longer improves significantly, the algorithm stops and outputs the optimal solution with the highest fitness in the current population as the final reconciliation solution.

[0110] The beneficial effects are described below:

[0111] 1. Global search capability: It has the ability to escape the trap of local optima and, in theory, can find a matching solution that is close to the global optimum in a huge search space.

[0112] 2. High versatility: It has less dependence on the specific mathematical form of the problem and is suitable for complex optimization scenarios where the objective function is difficult to differentiate or is nonlinear.

[0113] The difference between this comparative example and Example 1 is as follows:

[0114] 1. Result determinism: The comparison example is a probabilistic search algorithm, which is affected by random seeds. The results of each run may be inconsistent, which does not meet the "reproducibility" requirement of auditing; Example 1 is a deterministic algorithm with stable results.

[0115] 2. Computational efficiency: The convergence speed of this comparison is usually slow, and the computation time is extremely long when dealing with data of tens of thousands or more; Example 1 uses hash dimensionality reduction, which has quasi-linear complexity and is extremely fast.

[0116] Comparative Example 3

[0117] A complex relational query method based purely on the database layer attempts to rely entirely on the powerful data processing capabilities of relational databases to achieve matching by writing complex SQL statements or stored procedures. The specific steps are as follows:

[0118] 1. Data loading: Import bank statement and corporate account statement data into a high-performance relational database (such as Oracle or PostgreSQL).

[0119] 2. Complex Join Queries: Write highly complex SQL scripts containing nested joins, recursive CTEs, or window functions. Attempt to find combinations of records with equal amounts using database set operations.

[0120] 3. Stored procedure processing: For loop logic that is difficult to express in SQL, use database stored procedures (PL / SQL, etc.) to perform row-by-row traversal and logical judgment using cursors.

[0121] The beneficial effects are described below:

[0122] 1. No external systems required: Directly utilize existing enterprise database infrastructure without introducing additional Python or AI computing environments.

[0123] 2. Good data consistency: By utilizing the database's own transaction mechanism, consistency and atomicity are guaranteed during data processing.

[0124] The difference between this comparative example and Example 1 is as follows:

[0125] 1. Ability to handle N:M problems: Databases excel at set operations, but are extremely poor at handling N:M problems involving permutations and combinations. Forcing such queries can easily lead to database performance crashes or deadlocks; Example 1 is specifically designed for this type of problem.

[0126] 2. Semantic processing capability: SQL is difficult to efficiently implement complex text similarity calculations (such as TF-IDF or Embedding) and cannot handle dirty OCR data; Example 1 has built-in powerful multimodal text processing capabilities.

[0127] The specific embodiments of the present invention have been described in detail above. It should be understood that the described embodiments are merely one of the preferred embodiments of the present invention and do not constitute a limitation on the scope of protection of the present invention. Those skilled in the art can make various modifications, substitutions, or variations within the scope defined by the claims without departing from the essence of the present invention, and all such modifications, substitutions, or variations should be included within the scope of protection of the present invention.

Claims

1. An asymmetric audit data reconciliation method based on amount feature binning and a greedy strategy applied to an intelligent reconciliation calculation engine, wherein the engine is deployed in a computer device with a processor and memory, characterized in that, Building a closed-loop intelligent reconciliation calculation engine involves the following five tightly coupled steps: Step S1: Adaptive semantic reconstruction and context enhancement construct a data preprocessing layer with "self-healing capabilities": Step S2: Based on the hyperplane dimensionality reduction of monetary fingerprints, a monetary hashing bucketing technique is introduced to compress the high-dimensional Cartesian product search space into a discrete low-dimensional subspace. Step S3: Multidimensional orthogonal constraint pruning. Within the feature bucket, the system performs strict "zero-tolerance" logical filtering, eliminating 99% of false matches through three orthogonal dimensions. Step S4: The confidence evaluation model based on three-way feature fusion constructs a non-linear multimodal scoring matrix for pruned candidate pairs: Step 5: Global greedy locking with maximum coverage priority, employing a reverse-thinking scheduling strategy.

2. The asymmetric audit data reconciliation method based on amount feature binning and greedy strategy according to claim 1, characterized in that, The specific steps of S1 are as follows: S11. Missing Metadata Inference: To address the common issue of missing account names in bank statements, the engine does not rely on external dictionaries. Instead, it constructs a dynamic mapping graph based on transaction polarity, i.e., the positive or negative sign of the amount, and automatically maps the "borrow / loan" logic to the "receiver / payer" field, achieving a 100% completion rate for key fields. S12, Timeline Anchor Point Repair: For time fragments containing only "month-day", a heuristic algorithm combining global time window and local context is adopted; first, the document-level annual fingerprint is locked, and if it fails, backtracking inference is performed based on the file creation time or the current accounting period to ensure that each transaction can be accurately anchored on the correct time axis.

3. The asymmetric audit data reconciliation method based on amount feature binning and greedy strategy according to claim 1, characterized in that, Step S2 is as follows: S21. Combined Feature Hashing: Defining a Sliding Window For unmatched sets, the system generates a full set of permutations and combinations. Instead of directly comparing the combinations, the system calculates the scalar fingerprint of the combination, which is the absolute value of the total amount and rounds it down to the smallest monetary unit (cent). S22. Hash Collision Exploitation: Utilizing the natural aggregation property of hash buckets, all potential combinations with equal "algebraic sums of amounts" are forcibly mapped to the same physical storage bucket. This reduces the time complexity of the search algorithm from... The exponential brute-force exhaustive search was instantly reduced to The polynomial level fundamentally solves the computing power bottleneck.

4. The asymmetric audit data reconciliation method and system based on amount feature binning and greedy strategy according to claim 1, characterized in that, Step S3 is as follows: S31. Vector Direction Constraint: This mandates that all non-zero sub-vectors within a combination must be isomorphic in the direction of capital flow, completely eliminating the logical loophole of "positive and negative offsetting to make up the numbers". S32. Spatiotemporal Neighborhood Constraint: Introducing the principle of temporal locality, it is mandatory that the transaction time of all elements in the portfolio must converge within the same accounting month or a preset $\Deltat$ window; S33. Semantic consistency constraint: Calculate the semantic entropy of the summary within the combination using a lightweight NLP algorithm, and eliminate random combinations with excessive semantic dispersion.

5. The asymmetric audit data reconciliation method based on amount feature binning and greedy strategy according to claim 1, characterized in that, The formula for the multimodal scoring matrix in step S4 is as follows: 。 6. The asymmetric audit data reconciliation method based on amount feature binning and greedy strategy according to claim 1, characterized in that, Step S4 is as follows: S41, integrates micro-character features, statistical distribution features, and macro-semantic features; S42. Introduce a complexity penalty factor to reduce the weight of combinations involving too many fragments. By introducing the complexity penalty factor, the solution space with fewer fragments is converged first, thereby reducing the overall depth of the system's recursive search and improving the utilization of computing resources.

7. The asymmetric audit data reconciliation method based on amount feature binning and greedy strategy according to claim 1, characterized in that, Step S5 is as follows: S51, Priority Queue Scheduling: The system maintains a dynamic max-heap, whose sorting key is not just the matching score, but the "record coverage"; S52, Global Optimal Approximation: The algorithm forces priority to process complex combinations involving the most records; S53. Atomic Lock Mechanism: Once the optimal solution at the top of the heap is confirmed, the system immediately locks all the atomic records involved to ensure that they are not reused by subsequent suboptimal solutions, thus guaranteeing the uniqueness and exclusivity of the final reconciliation result.

Images