Example sample selection system based on Shapley value

Abstract

The invention discloses an example sample selection system based on a Shapley value, and aims to solve the problems that a current example selection method is long in calculation time and a finally selected example is not high in matching degree with the actual demand of a user. The system comprises a CSV model construction module, a candidate example set acquisition module and an example sample acquisition module. And the CSV model construction module is used for constructing a CSV model based on the Shapley value. The candidate example set acquisition module is used for acquiring a historical candidate example set according to a financial big data arrangement task and a big language model, setting the number of example samples required by a user, and storing the acquired candidate example set and the set number of example samples in the cloud. And the example sample obtaining module is used for obtaining a set number of example samples from a candidate example set by utilizing a CSV model based on the large language model at the cloud. The invention belongs to the technical field of example sample selection.

Description

An example sample selection system based on Shapley values Technical Field

[0001] This invention relates to the field of example sample selection, and in particular to an example sample selection system based on Shapley values. Background Technology

[0002] In recent years, Large Language Models (LLM) have demonstrated exceptional semantic understanding capabilities in tasks unique to the financial sector, such as parsing long-text contracts, extracting unstructured public opinion, and generating multilingual transaction reports. This has provided a new approach to addressing common issues in financial data processing, such as semantic ambiguity in cross-table relationships and alignment discrepancies in time-series data. For example, in customer credit risk assessment, using LLM to learn from unstructured data such as corporate financial statements, guarantee agreement clauses, and industry research reports with a small sample size can quickly build customized risk assessment models suitable for SMEs. However, in practical applications, selecting the few most valuable examples (such as key feature combinations of typical default cases and multidimensional data labels of high-quality customers) from a database of millions of historical risk control cases to minimize the computational cost of contextual learning and maximize the prediction accuracy of the large language model has become a key technical bottleneck restricting the implementation of LLM in financial scenarios. This challenge is known in industry practice as the "demonstration engineering" problem. Current mainstream example selection methods in the industry face the following significant pain points:

[0003] The example selection method based on Retrieval Enhancement Generation (RAG) only matches similar payment scenarios in historical transactions using cosine similarity when processing transaction log data. It ignores the differentiated impact of different transaction types (such as cross-border trade finance and domestic supply chain loans) on subsequent risk assessment models, resulting in a low degree of match between the obtained examples and the actual needs of users. Furthermore, the geographically dispersed deployment of transaction data and risk control systems by some financial institutions leads to inconsistent IO protocols or limited IO controller bandwidth when extracting transaction type features across regions, causing IO bandwidth congestion and increasing IO latency. While the example selection method based on topic models can extract potential topics such as service efficiency and product design when analyzing customer complaint texts, its inference and attribution calculations are performed on the GPU, while the associated data of a single example is mostly stored in memory. The limited bandwidth of the PCIe bus between GPU memory and host memory makes data transfer time far exceed the computation itself, further reducing quantification efficiency. It also cannot quantify the actual contribution of a specific example to the LLM's identification of the sentiment tendency of complaint texts, resulting in a low degree of match between the obtained examples and the actual needs of users. Evaluation methods based on influence functions need to be repeatedly tested on a validation set of tens of millions of credit data. The validation of a single example of a credit scoring model may incur several hours of GPU computing costs, making GPU computing slow and time-consuming. Furthermore, it is easily affected by abnormal transaction data, leading to unstable selection results.

[0004] In summary, the current example selection method is not only computationally time-consuming, but the final selected example also does not match the user's actual needs well. Summary of the Invention

[0005] In order to solve the problems that current example selection methods are not only computationally time-consuming, but also that the selected examples do not match the user's actual needs well, this invention proposes an example sample selection system based on Shapley values.

[0006] The technical solution adopted in this invention is:

[0007] It includes:

[0008] The CSV model building module is used to build CSV models based on Shapley values.

[0009] The candidate example set acquisition module is used to obtain historical candidate example sets based on financial big data processing tasks and large language models, set the number of example samples required by the user, and then store the acquired candidate example set and the set number of example samples in the cloud.

[0010] The example sample acquisition module is used to obtain a set number of example samples from the candidate example set in the cloud based on the large language model and using a CSV model.

[0011] The beneficial effects of this invention are as follows:

[0012] This invention innovatively introduces the Shapley value from cooperative game theory into financial data processing tasks, constructing a constrained Shapley value (CSV) model. Utilizing the CSV model and LLM, and within constraints on the number of examples required by regulatory compliance requirements (e.g., a maximum of 10 historical cases for reference in a single loan approval scenario) and computational resource constraints (e.g., memory limitations of edge computing devices), calculations are performed based on a candidate example set. This significantly reduces computational load, maximizing the improvement of core indicators such as customer credit scoring accuracy and transaction fraud identification recall rate. It also improves the matching degree between selected examples and actual user needs, reduces computation time, and provides key technical support for financial institutions to unlock the value of data elements during digital transformation.

[0013] In the customer segmentation scenario of retail banking, the CSV model can quantify the marginal contribution of each historical customer data example to the LLM-generated customer value score. For example, it can accurately identify the joint impact of key data subsets such as the asset allocation portfolio of high-net-worth customers and the repayment ability characteristics of consumer credit customers. Addressing the data compliance constraints (such as GDPR restrictions on the use of sensitive customer information) and computational efficiency requirements specific to the financial sector, this invention improves the traditional Shapley value approximation algorithm, reducing the example selection complexity from exponential to polynomial level while maintaining accuracy. This effectively solves the industry problem of excessively high example selection costs in scenarios with millions of transaction data points.

[0014] This invention is the first to introduce Shapley values ​​from financial game theory into large language models for learning. By limiting the subset size, it addresses the high computational complexity of traditional methods and accurately evaluates the marginal contribution of each example sample to model performance. This invention can be applied to any LLM task requiring context learning (such as NLG and QA) and supports alternative algorithms (such as greedy strategies), but the CSV method has significant advantages in both theory and practice. Attached Figure Description

[0015] Figure 1 is a flowchart of the present invention;

[0016] Figure 2 is a schematic diagram of the MCSV module;

[0017] Figure 3 is a schematic diagram of the ACSV module;

[0018] Figure 4 is a schematic diagram of the BCSV module; Detailed Implementation

[0019] Specific Implementation Method 1: This implementation method is illustrated with reference to Figures 1-4. The example sample selection system based on Shapley values ​​described in this implementation method includes a CSV model construction module, a candidate example set acquisition module, and an example sample acquisition module.

[0020] This invention models the task of financial big data processing (DW) as a process with a given number of examples. and cost budget Under the constraints, the optimal choice is selected based on the constrained Shapley value and the large language model. An example to maximize the utility function of an LLM. The problem is that the financial big data processing task originates from both the client and server sides, and the required data or information is stored in the cloud.

[0021] The financial big data processing tasks include the following:

[0022] 1. Entity Matching (EM): Consider a pair of tables with a unified schema. and table Each row in the table represents an entity, and each column represents the entity's attributes. The table contains a total of List , The first entity represents the third Entity matching is a binary classification task used to determine a pair of entities. Whether they refer to the same real-world entity, for example, whether "iPhone 13, Apple, 799" and "iPhone 13 Pro, Apple, 999" refer to the same product.

[0023] 2. Schema Mapping (SM): Consider a pair of contained... Column table and Their patterns (i.e., columns) are respectively and Pattern mapping is a binary classification task used to determine two attributes. and Does it refer to the same real-world attribute, for example, attribute? It is "Product_Name", attribute It's Product_Price. Do they refer to the same attribute?

[0024] 3. Error Detection (ED): Consider a system containing... List( ) table Error detection is a binary classification task used to determine the correct tuple in a table. Attributes Check if the value contains errors, such as values ​​that deviate from the true value. For example, check if "Laptop, Dell, 1200, Discount: 120%" contains any errors.

[0025] 4. Missing Value Imputation (MVI): Consider a condition containing... List( ) table Missing value imputation aims to fill in missing values ​​for attributes in a table. Determine the most likely value, for example, infer the brand of "Smartwatch, [MISSING], 199, Color: Black" from the context.

[0026] 5. Financial Big Data Normalization (DN): Consider a set of financial big data... List( ) table Financial big data standardization aims to standardize each tuple in the table according to the requirements of downstream financial big data analysis tasks. Attributes Convert the value to standard form For example, "tableA" can be normalized by removing redundant symbols.

[0027] 6. Data Deduplication (DD) in Financial Big Data: Consider a data structure with... List( ) table Each row in the table corresponds to one entity. The first entity represents the third Each attribute. Financial big data deduplication is a binary classification task used to identify a pair of entities in a table. Whether they refer to the same real-world entity, for example, to identify whether “Name: John Doe, Number: 123 - 456 - 7890” and “Name: John Doe, Number: 123- 456 - 7890” are duplicate records.

[0028] The CSV model building module is used to build CSV models based on constrained Shapley Values ​​(CSV). The CSV model includes the MCSV module, ACSV module, and BCSV module.

[0029] The MCSV module is:

[0030] (1)

[0031] in, Historical candidate sample samples were selected from historical databases based on the task of organizing financial big data. The quantity and historical database are stored in the cloud, facilitating large-scale storage and a unified format. Preprocessed The candidate example set is obtained by combining the candidate example samples. , For the candidate example set The Middle ( ) candidate example samples CSV estimate, It is a non-zero constant; to simplify calculations, it is usually assumed to be... , For each round from the candidate example set A subset consisting of the number of candidate example samples in the middle sampling. Customizable quantity for The absolute value, For a given utility function, The utility function for each candidate example sample added to the subset. For subset The utility value, The number of sample samples ultimately required by the user (for customization) ), Indicates multiplication. This indicates marginal contribution or activation contribution. CSV limits the calculation of Shapley values. Less than This makes the calculation results more consistent with the selections made in the LLM prompts. The actual needs of the example sample.

[0032] The ACSV module is:

[0033] (2)

[0034] (3)

[0035] in, To activate contributions, It is a cross-field association to strengthen the weight, which is a given value.

[0036] This invention defines the CSV model as having the following important properties, making it a utility function for evaluating candidate example samples for LLM. Effective value function of contribution:

[0037] 1) Symmetry: For each subset Any two candidate examples with equal marginal contributions receive the same reward, i.e. ,like ,but ,in, For the first One candidate example sample, For the first One candidate example sample, For the first CSV estimates of candidate example samples, For the first CSV estimates of candidate example samples.

[0038] 2) Additivity: The utility function of all candidate example samples. The sum of their contributions equals The function value over all candidate example samples, i.e. .

[0039] 3) Balance: For any candidate example sample Two subsets involved and The rewards they received in the two subsets were respectively and Then it is in the subset The reward distribution is as follows .

[0040] 4) Zero element: a pair containing at most Candidate example samples whose reward contribution is zero for any subset of an element have a reward of 0, i.e. ,like ,but .

[0041] Before implementing each module, this invention first obtains a candidate example set by using a candidate example set. The specific process is as follows:

[0042] First, a financial big data processing task is obtained from the client and input into the large language model. The large language model retrieves a certain number of candidate example samples from the historical database. Each candidate example sample has attribute or category labels. Then, all candidate example samples are preprocessed to ensure that the candidate example sample format is uniform. Finally, the preprocessed candidate example samples are combined into a candidate example set. Simultaneously, the user defines the desired number of sample instances. Finally, the acquired candidate example set and the set number of sample instances are sent to cloud storage.

[0043] The example sample acquisition module is used to obtain a set number of example samples from the candidate example set in the cloud based on the large language model and computing resource constraints (such as the memory limit of edge computing devices) using a CSV model.

[0044] (1) When the total number of labels in the candidate example set is less than or equal to 1000, the MCSV module is selected. The MCSV module is an approximation algorithm based on marginal contribution, abbreviated as MCSV, which is adapted from the classic Shapley value marginal contribution approximation technique. The time complexity of the MCSV module is O(n log n). Compared to classic algorithms, limiting the permutation size reduces runtime costs. Based on the aforementioned large language model, the MCSV module is used to obtain a set number of example samples from the candidate example set. The specific process is as follows:

[0045] First, set the cost budget and the number of candidate example samples within the subset for each iteration of the large language model, and provide a formula for calculating the utility function. Second, in each iteration, based on the set number of candidate example samples within the subset, [the system] adjusts the candidate example set accordingly. Random sampling is performed, and the sampled candidate examples are randomly sorted to obtain a subset. Calculate subsets based on the utility function formula. The utility value when it is empty, then according to subset The candidate example samples are sorted and added to the subset one by one. The marginal contribution of each candidate example is calculated by comparing the utility value of the set after each addition with the utility value before the addition. The calculation results, and The value is saved for the next iteration. During the iteration process, the cumulative cost is calculated in real time, and the iteration stops when the cost budget is met. Next, based on the marginal contribution of each candidate example sample in each iteration and the number of times the candidate example sample is sampled, the CSV estimate of each candidate example sample is calculated using formula (1), and finally the top candidates with the highest CSV estimates are returned. Example samples.

[0046] (2) The ACSV module is selected when the total number of annotations in the candidate example set is greater than 1000 and the number of candidate example samples is less than 20. The ACSV module is an approximation algorithm based on activation contribution, abbreviated as ACSV. Based on the large language model, the ACSV module is used to obtain a set number of example samples from the candidate example set. The specific process is as follows:

[0047] This invention first proposes the concept of Activated Contribution (AC) and defines Activated Contribution. ,in, This involves strengthening the weights across fields, setting a given value, and defining a cost budget and the number of candidate example samples within the subset for each iteration of the large language model, as well as defining a utility function formula. Secondly, in each iteration, the candidate example set is adjusted based on the set number of candidate example samples within the subset. A subset is sampled uniformly and randomly from the middle. For subsets Each candidate example sample Each candidate example sample is calculated based on the utility function formula and the definition of activation contribution. The activation contribution in the current iteration. The cumulative cost is calculated in real time during the iteration, and the iteration stops when the cost budget is met. Next, based on the activation contribution of each candidate example sample in each iteration and the total number of times the candidate example sample is sampled, the CSV estimate of each candidate example sample is calculated using formula (2). Finally, the top candidates with the highest CSV estimates are returned. A number of example samples. In this way, the AC algorithm can share... The calculation avoids the need for computation in the MCSV module. Repeated computation. By uniformly sampling subsets This ensures the unbiasedness of CSV estimation. Furthermore, to further optimize the sampling strategy and algorithm performance, sample allocation methods based on bias minimization and regret minimization are proposed. Bias minimization-based sample allocation minimizes the deviation between the estimated CSV value and the true value by adjusting the sampling probability. Regret minimization-based sample allocation models the CSV selection problem as a multi-arm recognition problem, reducing the probability of selecting incorrect examples by optimizing the sampling strategy. Theoretically, the time complexity of the ACSV module is O(log n). However, in practical applications, due to its more efficient calculation method, the ACSV module outperforms the MCSV module. The ACSV module is used for a small number of precise candidate example samples, that is, a set of candidate examples with more than 1000 labels and a small number of candidate examples. It can solve most example selection problems.

[0048] (3) The BCSV module is used when the total number of annotations in the candidate example set is too small and the number of candidate example samples is large. The BCSV module is used for batch processing scenarios with a coarse number of candidate example samples. Based on the large language model, the BCSV module is used to obtain a set number of example samples from the candidate example set. The specific process is as follows:

[0049] In batch processing scenarios, the BCSV module first generates batches of test records based on the candidate example set using a text embedding-based batch processing method (such as the method used in BatchER). Secondly, it initializes the large language model by reusing the distribution patterns of CSV estimates from historical batches using a meta-learning method, resulting in an initialized large language model. All test records from a single batch are then input into the initialized large language model for one iteration. The process is as follows: for each batch of test records, the Sentence-BERT method is used to extract feature vectors from all test records and all candidate example samples in that batch. The L2 distance between the feature vectors of each test record and each candidate example sample is calculated, and the closest feature vector is selected. ( ( ) candidate example samples. Then, the ACSV module is used to extract from this... The top candidates were selected from a pool of candidate examples. The BCSV module generates a CSV estimate for each test record using sample samples. During iteration, the cumulative cost is calculated in real time. Iteration stops when the cost budget is met, and the candidate example with the highest final CSV estimate is selected based on the number of sample samples required by the user. This invention combines the CSV estimates of historical batches with the meta-features of the current batch to calculate the expected CSV value of each candidate example in the current batch. It models the sampling process as a meta-learning multi-arm recognition problem, further optimizing example selection and thus more accurately selecting examples suitable for the current batch, improving the efficiency of large-scale financial big data processing. The offline time complexity of the BCSV module is O(n log n). It can operate within an error range comparable to the ACSV module.

[0050] Example 1

[0051] This invention models income verification for credit card applications (a bank risk control scenario), specifically including:

[0052] 1) Setting constraints includes:

[0053] Define activation contribution .

[0054] Number of examples .

[0055] Cost Budget USD (OpenAI API fee rate).

[0056] A subset of each iteration , representing a uniform and random sampling of 8 candidate examples.

[0057] Define utility function ,in, Precision is a metric used to evaluate model performance in binary and multi-class classification tasks. Its core function is to balance precision and recall, and it is particularly suitable for scenarios with imbalanced sample distribution. Privacy protection is calculated by PHI leak detection based on the bank customers and risk control scenarios requiring protection.

[0058] 2) Calculation process of CSV estimate:

[0059] 1200 candidate examples with income inconsistency and occupational anomaly annotations were selected from the LendingClub dataset. Based on the number of candidate examples and the number of annotations, the ACSV module was used to calculate the CSV estimate for each candidate example. The specific process is as follows:

[0060] Due to the large number of candidate examples, the large language model iteratively computes the CSV estimate for each candidate example, as is set up in the following embodiments. Based on an API cost budget of $50 and a call cost of $1.2368 per round, the maximum number of iteration rounds is determined to be 38.

[0061] In each iteration, the subset is input into GPT-3.5 for income verification inference. The number of actual anomalies in the candidate examples of income anomalies predicted by GPT-3.5 is counted, and the precision is calculated. Then, the number of anomalies successfully identified by GPT-3.5 in the candidate examples of actual income anomalies is counted, and the recall is calculated. The utility function of the subset is calculated based on the precision and recall to obtain the utility value. Then, given a cross-field association reinforcement weight of 1.2 for each candidate example, the utility value after adding each candidate example to the subset is calculated in turn. The utility value calculated each time is multiplied by the cross-field association reinforcement weight to obtain the activation contribution of the candidate example. The activation contribution of each candidate example sample is obtained. The cumulative cost is counted in real time during the iteration process. The iteration is stopped after 38 iterations. Based on the activation contribution of each candidate example in each iteration and the total number of times the candidate example is sampled, the CSV estimate of each candidate example is calculated using formula (1). The 10 candidate examples with the highest CSV estimates are directly selected.

[0062] 3) After practical verification, the following optimization effects were obtained:

[0063] 1. Performance Improvement:

[0064] The revenue validation F1 score of the GPT-3.5 model changed from 0.75 to 0.87 (+16%).

[0065] Privacy breach rate reduced to 0.2% (in compliance with GDPR);

[0066] 2. Cost control:

[0067] The traditional method costs $210, while ACSV only costs $47 (saving 77.6%).

[0068] 3. Experimental conditions:

[0069] Hardware: NVIDIA RTX 4090;

[0070] Time taken: 22 minutes (traditional Shapley takes 18 hours);

[0071] 4. Socioeconomic value:

[0072] The simulated bank processes 3 million applications annually.

[0073] Manual review was reduced by 53%, saving 3.6 million yuan in risk control costs;

[0074] Fraud losses were reduced by 28% (approximately 120 million yuan in losses were avoided).

[0075] Example 2:

[0076] This invention models the verification of invoices in the automotive supply chain (industrial finance scenario), specifically including:

[0077] 1) Setting constraints includes:

[0078] Number of examples (Edge device memory limitations);

[0079] Cost Budget US dollars (cost of edge API calls);

[0080] 2) Calculation process of CSV estimate:

[0081] 30,000 automotive supply chain invoices were generated using AUTOINVOICE as a candidate example set. A text-embedding-based batch processing method was used to divide the 30,000 invoices into 500 batches of 60 invoices each. The BCSV module was selected to calculate the CSV estimate for each invoice. The specific process is as follows:

[0082] The large language model is initialized by reusing the distribution patterns of CSV estimates from historical batches using meta-learning methods. Then, the Sentence-BERT tool is used to extract text feature vectors from each batch of invoices and all candidate examples. The L2 distance between each invoice and the feature vectors of all candidate examples is calculated, and the 20 candidate examples with the closest L2 distance are selected. The ACSV module is then called to perform a second calculation on these 20 candidate examples, randomly sampling subsets round by round, calculating the utility value of the subsets, and evaluating the change in utility of the subsets after the addition of a single example, to obtain the CSV estimate of each candidate example. From this, the top 6 optimal candidate examples are selected. The entire calculation process strictly adheres to the 16GB memory limit of the edge device.

[0083] 3) After practical verification, the following optimization effects were obtained:

[0084] 1. Performance Breakthrough:

[0085] The accuracy rate of three-single matching is 89% (compared to 68% for traditional methods);

[0086] Processing speed: 9.2 copies / second (83% above real-time requirement);

[0087] 2. Resource saving:

[0088] Memory usage is only 14.3GB (<16GB limit);

[0089] The calculation took 4.1 minutes (<5-minute threshold);

[0090] 3. Experimental conditions:

[0091] Equipment: Raspberry Pi 4B + Google Coral TPU;

[0092] Candidate example dataset: AUTOINVOICE synthetic data (10,000 images);

[0093] 4. Socioeconomic value:

[0094] Cost savings for mid-sized automakers:

[0095] Reduced accounting staff by 4 million yuan;

[0096] Shorten the payment cycle by 31 days (improving supplier cash flow).

[0097] Example 3:

[0098] This invention models medical insurance fraud detection (insurance technology scenario), specifically including:

[0099] 1) Setting constraints includes:

[0100] Number of examples (Regulatory requirements cover 7 types of fraud patterns).

[0101] Cost Budget US dollars (PHI shielding cost).

[0102] A subset of each iteration , indicating that 10 candidate examples are randomly collected and randomly sorted.

[0103] Define utility function .

[0104] 2) Calculation process of CSV estimate:

[0105] Eighty hundred fraud candidate examples were selected from the CMS dataset, and the CSV estimate for each fraud candidate example was calculated using the MCSV module. The process is as follows:

[0106] In each iteration, a subset is first randomly generated from the 800 fraud candidate examples. The candidate examples in this subset are then selected from the candidate example set. Random sampling and random sorting. The marginal contribution of each candidate example is obtained in turn according to the utility function formula and the sorting of candidate example samples in the subset. The cumulative cost of PHI shielding is calculated in real time during the calculation process. When the cost reaches $32, the iteration stops. According to the marginal contribution of each candidate example sample in each iteration and the number of times each candidate example sample is sampled, the CSV estimate of each candidate example is calculated using formula (1). The 7 candidate examples with the highest CSV estimates are used as the final output.

[0107] 3) After practical verification, the following optimization effects were obtained:

[0108] 1. Detection efficiency:

[0109] The AUC of the large language model is 0.89 (benchmark 0.71).

[0110] The false alarm rate was 4.3% (industry standard <5%).

[0111] 2. Key findings:

[0112] The final output of 7 candidate examples includes features such as "a sudden increase of 300% in daily prescriptions".

[0113] Cross-institutional data associations accounted for 62% of the CSV value.

[0114] 3. Laboratory reproduction:

[0115] Framework: PyTorch 2.1 + HuggingFace Transformers

[0116] Privacy protection: PHI desensitization was performed using Microsoft Presidio.

[0117] 4. Socioeconomic value:

[0118] Simulation of a million-level claims scale:

[0119] Annual fraud prevention losses amounted to 28 million yuan.

[0120] The investigation period has been shortened from 21 days to 5 days, significantly improving service efficiency.

[0121] Example 4:

[0122] This invention models green credit ESG verification (sustainable finance), specifically including:

[0123] 1) Setting constraints includes:

[0124] Define activation contribution .

[0125] Number of examples (Core indicators of EU classification).

[0126] Cost Budget US dollars (cost of accessing satellite data).

[0127] Cross-modal CSV fusion:

[0128] Multi-source candidate set:

[0129] Satellite imagery (Sentinel-2 public data);

[0130] IoT sensor readings (OSIsoft dataset);

[0131] Enterprise ESG Reports (TCFD Database).

[0132] ACSV weighted:

[0133] text ,image ,sensor .

[0134] Meta-feature guidance: Reuse historical carbon verification batch mode.

[0135] 2) Calculation process of CSV estimate:

[0136] A cross-modal ACSV fusion calculation was employed. First, Sentinel-2 satellite imagery, OSIsoft IoT sensor readings, and TCFD enterprise ESG reports were integrated to form a candidate example set. Differentiated weights were assigned to different modalities of data: text-based ESG reports had a weight of 1.0, satellite imagery had a weight of 1.5, and sensor readings had a weight of 0.8. A meta-feature-guided strategy was introduced during the calculation, reusing the CSV value distribution of historical carbon audit batches to optimize the sampling direction, and using the ACSV module to obtain CSV estimates. Simultaneously, Microsoft Presidio was used to anonymize trade-sensitive fields during the calculation process, ultimately selecting four optimal examples.

[0137] 3) After practical verification, the following optimization effects were obtained:

[0138] 1. Verification capability:

[0139] "Greenwashing" identification accuracy rate is 88% (baseline 65%).

[0140] Processing speed: 1400 records / minute;

[0141] 2. Compliance:

[0142] The confidentiality of trade information is 100% masked.

[0143] The sensitivity of detecting discrepancies in carbon emission data has been improved by 3.2 times;

[0144] 3. Laboratory reproduction:

[0145] Tools: Google Earth Engine + Scikit-learn;

[0146] Hardware: Xeon E5-2690v4 + Quadro RTX 6000.

[0147] 4. Socioeconomic value:

[0148] 20 billion yuan in green loan scenarios:

[0149] Annual savings of 9 million yuan in audit costs;

[0150] Reduced exposure to "greenwashing" risk by $450 million.

[0151] Example 5:

[0152] This invention models cross-border payment risk control (in a fintech scenario), specifically including:

[0153] 1) Setting constraints includes:

[0154] Number of examples (PSD2 real-time risk control requirements);

[0155] Cost Budget US dollars (maximum cost per transaction);

[0156] BCSV Extreme Optimization:

[0157] Micro-batch processing: Aggregates 100 transactions every 0.1 seconds;

[0158] dynamic Adjustment:

[0159] Normal traffic ;

[0160] Fluctuation period (Enhance anomaly detection).

[0161] Regret-minimizing sampling: reducing the false screening rate in high-risk areas;

[0162] 2) Calculation process of CSV estimate:

[0163] This case study utilizes the BCSV module for real-time calculations. It first aggregates 100 transactions into micro-batches within a 0.1-second time window. The candidate example selection threshold is dynamically adjusted based on transaction volume: 8 candidate examples are selected during normal traffic, increasing to 12 during volatile periods to enhance anomaly detection. A regret-minimizing sampling strategy is employed, adjusting the sampling probability based on the contribution data of historical high-risk transaction areas to reduce the false positive rate of high-risk transactions. For each micro-batch, the similarity between the features of each transaction example and historical examples is calculated. Combining similarity and marginal contribution, a CSV estimate is calculated, selecting 6 optimal examples. Throughout the process, a strict cost cap of $0.5 per transaction is maintained, and transaction processing latency is monitored in real-time. Ultimately, these 6 examples demonstrate a 95% fraud transaction detection rate with an average processing latency of 37ms, meeting the PSD2 real-time risk control requirements.

[0164] 3) After practical verification, the following optimization effects were obtained:

[0165] 1. Risk control efficiency:

[0166] Fraudulent transaction detection rate: 95% (benchmark: 79%);

[0167] The false interception rate has been reduced to 1.8% (international standard <2%).

[0168] 2. Delay control:

[0169] Average processing latency is 37ms (<50ms requirement);

[0170] Peak throughput: 2400 transactions / second;

[0171] 3. Laboratory reproduction:

[0172] Dataset: SWIFT public messages + synthetic cryptocurrency stream;

[0173] Stress testing: Locust simulates millions of concurrent users;

[0174] 4. Socioeconomic value:

[0175] The daily average number of payment transactions is 120 million.

[0176] Annual reduction of erroneous liquidation losses by $12 million;

[0177] Reduced the risk of compliance penalties by 83%.

[0178] The above examples of the present invention are merely illustrative of the computational model and process of the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is impossible to exhaustively list all possible implementations here. Any obvious variations or modifications derived from the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A sample selection system based on Shapley values, characterized in that: It includes: The CSV model building module is used to build CSV models based on Shapley values. The candidate example set acquisition module is used to acquire historical candidate example sets based on financial big data processing tasks and large language models, set the number of example samples required by the user, and then store the acquired candidate example set and the set number of example samples in the cloud; the example sample acquisition module is used to obtain the set number of example samples from the candidate example set in the cloud based on the large language model and using a CSV model.

2. The sample selection system based on Shapley values ​​according to claim 1, characterized in that: The CSV model includes the following attributes: (1) Symmetry: ,like ,but ,in, For the first One candidate example sample, For the first One candidate example sample, For the first CSV estimates of candidate example samples, For the first (2) Additivity: CSV estimates of candidate example samples; ,in, For the candidate example set The utility function; (3) Balance: the rewards obtained by any candidate sample in the two subsets are respectively and Then in the subset The reward distribution is as follows (4) Zero element: ,like ,but 。 3. The sample selection system based on Shapley values ​​according to claim 2, characterized in that: The CSV model includes an MCSV module, an ACSV module, and a BCSV module.

4. The sample selection system based on Shapley values ​​according to claim 3, characterized in that: The MCSV module is: (1) Among them, ， For obtaining historical candidate example samples Quantity, For preprocessed The candidate example set is obtained by combining candidate example samples. ， For the candidate example set The Middle Candidate example samples CSV estimate, ， For each round from the candidate example set A subset consisting of multiple candidate example samples from the sampling process. for The absolute value, The number of example samples required by the user. For a given utility function, The utility function for each candidate example sample added to the subset. For subset The utility value, Indicates multiplication. Indicates marginal contribution. This indicates the number of times the candidate example sample was sampled.

5. The sample selection system based on Shapley values ​​according to claim 4, characterized in that: The ACSV module is: （2） (3) Among them, To activate contributions, It is a cross-field association to strengthen the weight, which is a given value.

6. The sample selection system based on Shapley values ​​according to claim 5, characterized in that: The specific process of the candidate example set acquisition module is as follows: the financial big data processing task is obtained from the client, the financial big data processing task is input into the big language model, the big language model retrieves a certain number of historical candidate example samples from the historical database stored in the cloud, each historical candidate example sample has attribute labeling or category labeling, all historical candidate example samples are preprocessed to obtain the candidate example set, at the same time, the user defines the required number of example samples, and sends the obtained candidate example set and the set number of example samples to the cloud for temporary storage.

7. The sample selection system based on Shapley values ​​according to claim 6, characterized in that: The financial big data processing tasks include entity matching, pattern mapping, error detection, missing value imputation, financial big data standardization, and financial big data deduplication.

8. The sample selection system based on Shapley values ​​according to claim 7, characterized in that: The specific process of the example sample acquisition module is as follows: 1) When the total number of annotations in the candidate example set is less than or equal to 1000, the MCSV module is selected. Based on the large language model, the MCSV module is used to obtain a set number of example samples from the candidate example set. The specific process is as follows: First, set the cost budget and the number of candidate example samples in the subset in each iteration of the large language model, and define the utility function formula; Second, in each iteration, randomly sample from the candidate example set according to the set number of candidate example samples in the subset, and randomly sort the sampled candidate example samples to obtain a subset. According to the utility function formula and the sorting of candidate example samples in the subset, the marginal contribution of each candidate example sample is obtained in turn. During the iteration process, the cumulative cost is calculated in real time. When the cost budget is met, the iteration stops. Based on the marginal contribution of each candidate example sample in each iteration and the number of times the candidate example sample is sampled, the CSV estimate of each candidate example sample is calculated using formula (1). The candidate example with the highest CSV estimate is selected according to the number of example samples required by the user. 2) When the total number of annotations in the candidate example set is greater than 1000 and the number of candidate example samples is less than 20, the ACSV module is selected. Based on the large language model, the ACSV module is used to obtain a set number of example samples from the candidate example set. The specific process is as follows: First, define the utility function formula, define the activation contribution formula according to the utility function formula, and set the cost budget and the number of candidate example samples in the subset in each iteration of the large language model. Secondly, in each iteration, a subset is obtained by uniformly and randomly sampling from the candidate example set according to the number of candidate example samples in the subset. The activation contribution of each candidate example sample is calculated according to the utility function formula and the activation contribution formula. During the iteration process, the cumulative cost is calculated in real time. When the cost budget is met, the iteration stops. Based on the activation contribution of each candidate example sample in each iteration and the total number of times the candidate example sample is sampled, the CSV estimate of each candidate example sample is calculated using formula (2). The candidate example with the highest CSV estimate is selected according to the number of example samples required by the user. 3) When batch processing of candidate example samples in the candidate example set is performed, the BCSV module is selected. Based on the large language model, the BCSV module is used to obtain a set number of example samples from the candidate example set. The specific process is as follows: First, set the cost budget and generate a batch of test records using the batch processing method based on text embedding according to the candidate example set. Second, use the meta-learning method to reuse the distribution law of the CSV estimate of the historical batch to initialize the large language model and obtain the initialized large language model. Input all test records of a single batch into the initialized large language model for one iteration. The processing process is as follows: extract the feature vectors of all test records and all candidate example samples in the batch, calculate the L2 distance between each test record and each candidate example sample feature vector, and select the front sample with the closest L2 distance to each test record. One candidate example sample, Then call the ACSV algorithm to process the previous... The CSV estimate for each test record is calculated using candidate example samples. The cumulative cost is calculated in real time during the iteration process. The iteration stops when the cost budget is met, and the candidate example with the highest CSV estimate is selected based on the number of example samples required by the user.

9. A sample selection system based on Shapley values ​​according to claim 8, characterized in that: The ACSV algorithm employs a bias minimization method and a regret minimization method in each round of iterative sampling.

10. A sample selection system based on Shapley values ​​according to claim 9, characterized in that: The method used in the BCSV module to extract the feature vectors of all test records and all candidate example samples in the batch is the Sentence-BERT method.

Images