An AI large model-based cross-border e-commerce multi-platform product selection and market analysis method and system

Abstract

The present application relates to the technical field of computer data processing and artificial intelligence application, and particularly relates to a cross-border e-commerce multi-platform product selection and market analysis method and system based on an AI large model. Multi-platform data in a sliding time window are acquired and evidence units are generated, evidence dependency relationships are determined according to source blood relations and confidence reduction is performed; a time sequence evidence conflict graph is constructed, overall confidence of an evidence package and a lower bound of commodity value are calculated; AI output is decomposed into atomic assertions and an assertion-evidence bipartite graph is established; commodity state determination is performed according to overall confidence of the evidence package, the lower bound of commodity value, assertion coverage and conflict intensity; after settlement maturity, deviation attribution is performed, and local state back calculation is performed on affected commodities. The present application provides a cross-border e-commerce multi-platform product selection and market analysis method and system based on an AI large model, which is used for reducing common source evidence repeated weighting, no-evidence assertion participating in state determination and commodity state repeated switching.

Description

Technical Field

[0001] This invention relates to the fields of computer data processing and artificial intelligence application technology, and in particular to a method and system for cross-border e-commerce multi-platform product selection and market analysis based on a large AI model. Background Technology

[0002] Product selection and market analysis in cross-border e-commerce require the aggregation and processing of data from cross-border e-commerce platforms, advertising platforms, logistics service systems, payment settlement systems, and market information sources. This data includes product prices, sales volume, page views, search popularity, review content, advertising costs, platform commissions, logistics costs, taxes, exchange rates, refund records, and settlement records. Existing processing methods typically map fields and normalize indicators from data from different sources, then generate product ratings, market analysis results, and product selection suggestions based on preset weights, statistical models, or artificial intelligence models.

[0003] Data output from multiple data interfaces may originate from the same platform page, the same data provider, the same market report, or the same upstream database. Data records generated by different collection interfaces may still share a common upstream source. When these data records are aggregated based on mutually independent evidence, information corresponding to a common upstream source will be weighted repeatedly. For unstructured data such as comments, reports, platform announcements, and market rules, the analysis output by the AI ​​model may also reference outdated versions of data or form factual assertions that cannot be correlated with the current data record.

[0004] Platform commissions, advertising attribution fees, logistics surcharges, exchange rate gains and losses, refund losses, warehousing costs, and taxes have different confirmation times. The data used to determine the status of goods may include both confirmed and immature data. When using immature cost estimates to directly calculate the value of goods, the cost confirmation results may lead to different status determinations for the same goods within adjacent time windows. When using a single status threshold, short-term market fluctuations, interface delays, version switching, or duplicate source data may also cause the status of goods to switch repeatedly within adjacent time windows. The above statuses can be recorded using the weighted ratio of duplicate evidence, key assertion coverage, the proportion of assertions without evidence, conflicting edge weights, the number of status switches per unit time, prediction settlement deviation, and the proportion of back-calculated goods.

[0005] Existing processing methods handle evidence source dependence, evidence version, evidence validity period, AI output evidence support relationship, and settlement maturity separately, resulting in a lack of unified product status determination relationships. Furthermore, when local cost data, settlement results, or evidence versions change, all products may need to be recalculated. This invention performs dependency reduction on evidence from common upstream sources, constructs a temporal evidence conflict graph based on evidence differences, breaks down AI output into atomic assertions and establishes assertion-evidence support relationships, determines product status based on multiple decision quantities, and identifies the affected product set along graph associations after settlement maturity. Summary of the Invention

[0006] To achieve the above-mentioned objectives, this invention provides a method and system for product selection and market analysis across multiple platforms in cross-border e-commerce based on a large AI model. This system is used to process common source evidence, data version differences, fee settlement time differences, and situations where AI output lacks supporting evidence in multi-platform data. It is also used to reduce repeated switching of product status within adjacent time windows and back-calculation of irrelevant products after settlement data updates.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a method for product selection and market analysis across multiple platforms in cross-border e-commerce based on an AI large-scale model, comprising the following steps:

[0008] Acquire multi-platform product, fee, market, and settlement data within a sliding time window, and generate evidence units with source lineage, validity period, and version identifiers; reduce the confidence of duplicate evidence based on common upstream sources, construct a time-series evidence conflict graph based on evidence differences, and calculate the overall confidence of the evidence package and the lower bound of product value; decompose the AI ​​output into atomic assertions, establish an assertion-evidence bipartite graph, and determine assertion coverage, the proportion of assertions without evidence, and conflict intensity; determine a processable state when each judgment quantity meets the processable condition, a supplementary data state when the evidence completeness is insufficient, a manual review state when the conflict intensity meets the review condition, and a restricted state when the lower bound of product value or the proportion of assertions without evidence meets the restriction condition; perform deviation attribution after settlement is mature, identify affected products, and perform local state backcalculation.

[0009] To further realize the present invention, the following technical solutions may be preferred:

[0010] Preferably, the generation of evidence units with source lineage, validity period, and version identifier includes:

[0011] For structured data, we perform commodity identification mapping, currency conversion, time zone alignment, and fee standardization. For unstructured data, we perform semantic parsing and encapsulate the processing results into evidence units that include evidence object, target market, indicator type, evidence value, original source identifier, derived source identifier, collection link identifier, effective time window, version identifier, settlement maturity, and original data reference location. When the commodity identification mapping result does not meet the mapping confirmation conditions, the corresponding data is written into the mapping set to be confirmed, and the corresponding evidence unit is not merged into the existing commodity evidence package.

[0012] Preferably, the confidence level for reducing duplicate evidence based on a common upstream source includes:

[0013] The source dependency coefficient between evidence units is determined based on the original source identifier, derived source identifier, acquisition link identifier, content fingerprint, and effective time window. When the source dependency coefficient reaches the dependency judgment condition, an evidence dependency edge is established between the corresponding evidence units, and confidence reduction is performed on the derived evidence within the common upstream source group. When the source dependency coefficient does not reach the dependency judgment condition, the corresponding evidence unit is treated as independent source evidence in the calculation.

[0014] Preferably, the step of constructing a temporal evidence conflict diagram based on evidence differences and calculating the overall confidence level of the evidence package and the lower bound of the commodity value includes:

[0015] Establish conflict edges between evidence units that correspond to the same product and target market but have numerical deviations, opposite trend directions, misaligned time windows, inconsistent versions, inconsistent settlement statuses, or inconsistent semantic results; when the weight of a conflict edge reaches the conflict determination condition, the corresponding conflict edge is included in the reduction item of the overall confidence of the evidence package; determine the lower bound of the product value based on the confirmed revenue, unvested revenue range, confirmed expenses, unvested expense range, and loss range.

[0016] Preferably, the step of decomposing the AI ​​output into atomic assertions and establishing an assertion-evidence bipartite graph includes:

[0017] Atomic assertions, including product subject, target market, attribute type, value or direction, applicable time window, establishment conditions, and referenced evidence identifier, are extracted from the AI ​​output. When the atomic assertion and the evidence unit meet the matching conditions in terms of evidence object, indicator type, target market, effective time window, and version identifier, a supporting edge or an opposing edge is established. When a key atomic assertion does not have a valid supporting edge, the corresponding key atomic assertion is marked as an assertion without evidence.

[0018] Preferably, the product status determination includes:

[0019] When the overall confidence level of the evidence package, the lower bound of the commodity value, and the assertion coverage rate reach the corresponding processing thresholds, and the proportion of assertions without evidence and the intensity of conflict are lower than the corresponding restriction thresholds, the status is determined to be processable. When necessary evidence is missing and the restriction conditions and manual review conditions are not met, the status is determined to be supplementary data. When the intensity of conflict reaches the review threshold or the key atomic assertion has both supporting and opposing edges that meet the conditions, the status is determined to be manually reviewed. When the lower bound of the commodity value is lower than the value restriction threshold or the proportion of assertions without evidence reaches the assertion restriction threshold, the status is determined to be restricted. When there are multiple determination results, the commodity status is determined in the order of restricted status, manual review status, supplementary data status, and processable status.

[0020] Preferably, the method further includes a state hysteresis control step, the state hysteresis control step comprising:

[0021] An entry restriction threshold and an exit restriction threshold are set respectively, and the entry restriction threshold and the exit restriction threshold form a hysteresis interval; when a product meets the entry restriction conditions and continuously reaches the preset number of confirmation windows, it enters the restriction state; when a product in the restriction state meets the exit restriction conditions, it enters the cooling observation state; when the exit restriction conditions are continuously met within multiple consecutive sliding time windows, and there is independent source recovery evidence and no unprocessed high-weight conflict edges, it exits the cooling observation state.

[0022] Preferably, the product status determination further includes:

[0023] When two or more platforms simultaneously output recovery evidence in the same direction, it is determined whether the recovery evidence has a common upstream source based on the evidence dependency edge. If the recovery evidence has a common upstream source, the recovery evidence is used as evidence from one source in the recovery determination. If the recovery evidence does not have a common upstream source, and the recovery evidence meets the recovery conditions within a continuous sliding time window, the corresponding product is switched from the cooling observation state to the processable state or the supplementary data state.

[0024] Preferably, the step of determining the affected goods and performing local state back calculation includes:

[0025] When settlement deviations, evidence versions, or fee rules change, the changed evidence node or parameter node is identified as the starting node, and propagation proceeds along evidence dependency edges, conflict edges, assertion support edges, and product association edges. Propagation stops when it reaches a preset termination condition, and the product nodes traversed by the propagation are identified as the affected product set. If the affected product set does not meet the full recalculation condition, the product status is recalculated only for the affected product set. When the affected product set meets the full recalculation condition, a full recalculation is performed.

[0026] A cross-border e-commerce multi-platform product selection and market analysis system based on AI large-scale models, comprising:

[0027] The evidence processing module is used to acquire data from multiple platforms, generate evidence units, and reduce the confidence level of duplicate evidence based on common upstream sources.

[0028] The graph processing module is used to construct a temporal evidence conflict graph and calculate the overall confidence of the evidence package and the lower bound of the commodity value.

[0029] The output verification module is used to decompose the AI ​​output into atomic assertions, establish an assertion-evidence bipartite graph, and determine assertion coverage, the proportion of assertions without evidence, and conflict intensity.

[0030] The status determination module is used to determine the product as a processable state, a supplementary data state, a manual review state, or a restricted state when each determination quantity meets the corresponding conditions.

[0031] The back-calculation module is used to perform deviation attribution after settlement is mature, determine the set of affected goods along the graph association edges, and perform partial state back-calculation or full back-calculation based on whether the set of affected goods meets the conditions for full back-calculation.

[0032] The beneficial effects of this invention are:

[0033] This invention determines the dependence of evidence based on the lineage of evidence sources, performs confidence reduction on derived evidence from common upstream sources, and writes numerical differences, trend direction differences, effective time window differences, version differences, settlement status differences, and semantic result differences into a time-series evidence conflict diagram to reduce the impact of duplicate weighting of evidence from common sources and data time misalignment on the product status determination results.

[0034] This invention breaks down AI output into atomic assertions and establishes an assertion-evidence bipartite graph. Based on the overall confidence of the evidence package, the lower bound of the product value, the assertion coverage, the proportion of assertions without evidence, and the intensity of conflict, it performs product status determination with priority. It also employs hysteresis state control and local back-calculation of affected products to reduce the entry of assertions without evidence into subsequent processing, repeated switching of product status, and repeated calculation of irrelevant products. Attached Figure Description

[0035] Figure 1 This is a schematic diagram of the system architecture of the present invention.

[0036] Figure 2 This is a schematic diagram of the main flow of the method of the present invention.

[0037] Figure 3 This is a schematic diagram of the process for generating evidence units and determining source dependence in this invention.

[0038] Figure 4This is a schematic diagram of the product status determination and hysteresis control process of the present invention.

[0039] Figure 5 This is a schematic diagram of the settlement deviation attribution and partial back-calculation process for affected commodities in this invention. Detailed Implementation

[0040] In the description of this invention, it should also be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," and "link" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0041] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0042] The cross-border e-commerce platform refers to a platform that provides services such as product display, order transactions, advertising, warehousing and logistics, or fund settlement. The product data includes product identification, product category, product price, sales volume, inventory, page views, and conversion rate; the fee data includes platform commissions, advertising fees, logistics fees, warehousing fees, taxes, and payment processing fees; the market data includes search popularity, number of competing products, review content, price range, market rules, and exchange rates; and the settlement data includes confirmed revenue, confirmed fees, refund records, chargeback records, logistics deductions, and platform settlement records.

[0043] The AI ​​model is capable of processing natural language text and generating semantic parsing results according to a preset output structure. The AI ​​model can be deployed on a local computing server or connected to a model server via a controlled interface. The output of the AI ​​model is written to the product analysis record after atomic assertion decomposition, evidence matching, and state determination. Key atomic assertions that fail evidence verification are not used to generate executable product processing records.

[0044] Example 1 System

[0045] See Figure 1As shown, a cross-border e-commerce multi-platform product selection and market analysis system based on AI large model includes an evidence processing module, a graph processing module, an output verification module, a status determination module, a deviation attribution module, a back-calculation module, a version snapshot storage module, and a result output module.

[0046] The evidence processing module includes a data access unit, a data standardization unit, an evidence unit generation unit, and a source dependency processing unit. The data access unit connects to interfaces with cross-border e-commerce platforms, advertising platforms, logistics services, payment settlement, and market information, and receives product data, cost data, market data, and settlement data according to a sliding time window.

[0047] The data access unit assigns a data batch identifier to each batch of received data and records the platform identifier, interface identifier, reception time, data generation time, interface return status, data version, and data verification value. After a data batch is received, batch verification is performed on the number of records, field completeness, data verification value, and time range. Data that fails the batch verification is written to the abnormal data set.

[0048] The data standardization unit is used to perform product identifier mapping, currency conversion, time zone alignment, category correspondence, and unified cost standards. When different platforms use different statistical standards for transaction amount, advertising attribution revenue, refund amount, and logistics costs, the data standardization unit converts the corresponding fields into unified fields according to the platform's field mapping table, while retaining the original field names, original field values, and conversion rule versions.

[0049] The evidence unit generation unit is used to encapsulate standardized structured data and semantically parsed unstructured data into evidence units. Each evidence unit includes an evidence unit identifier, evidence object, target market, platform identifier, indicator type, evidence value, unit of measurement, trend direction, original source identifier, derived source identifier, collection link identifier, collection time, data generation time, effective time window, version identifier, settlement maturity, content fingerprint, and original data reference location.

[0050] The source dependency processing unit is used to determine the source dependency relationship between evidence units based on the original source identifier, derived source identifier, acquisition link identifier, content fingerprint and effective time window, and to establish evidence dependency edges between evidence units that meet the dependency judgment conditions.

[0051] The graph processing module includes a conflict graph construction unit, an overall confidence calculation unit, and a commodity value lower bound calculation unit. The conflict graph construction unit is used to write evidence units into the graph data storage structure and establish conflict edges between evidence units that have numerical deviations, opposite trend directions, misaligned time windows, inconsistent versions, inconsistent settlement statuses, or inconsistent semantic results.

[0052] The overall confidence calculation unit is used to read the effective confidence, conflict edge weight, time decay, source historical reliability, settlement maturity, and necessary evidence missing status of the evidence unit, and generate the overall confidence of the evidence package corresponding to the commodity, target market, and analysis time window.

[0053] The commodity value lower bound calculation unit is used to read the confirmed revenue, unvested revenue range, confirmed expenses, unvested expense range, return loss range, and refusal loss range, and determine the commodity value lower bound under the preset credit level.

[0054] The output verification module includes an atomic assertion splitting unit, an assertion-evidence matching unit, and an assertion metric calculation unit. The atomic assertion splitting unit splits the AI ​​output into atomic assertions corresponding to single-judgment content. The assertion-evidence matching unit establishes support or opposition edges between atomic assertions and evidence units. The assertion metric calculation unit determines assertion coverage, the proportion of assertions without evidence, and conflict intensity.

[0055] The state determination module includes a condition judgment unit, a state priority processing unit, and a hysteresis control unit. The condition judgment unit is used to generate one or more candidate states based on the overall confidence of the evidence package, the lower bound of the commodity value, the assertion coverage, the proportion of assertions without evidence, the conflict intensity, and the state of missing necessary evidence.

[0056] The state priority processing unit is used to determine the product state in the order of restricted state, manual review state, supplementary data state, and processable state when a product has multiple candidate states simultaneously. The hysteresis control unit is used to set the entry restriction threshold and exit restriction threshold, and to perform continuous window confirmation and cool-down observation processing for the entry and exit of restricted states.

[0057] The deviation attribution module is used to compare the predicted revenue, predicted expenses, predicted commodity value, and predicted loss with the corresponding actual settlement results after settlement is completed, and to allocate the deviation to price, advertising expenses, logistics expenses, taxes, exchange rates, returns, refusal of payment, or evidence analysis items.

[0058] The back-calculation module is used to take the changed evidence node or parameter node as the starting node, and propagate along the evidence dependency edge, conflict edge, assertion support edge and product association edge to form an affected product set, and perform local state back-calculation or full back-calculation according to whether the affected product set meets the conditions for full back-calculation.

[0059] The version snapshot storage module is used to store the evidence unit version, graph structure version, AI model version, output template version, fee rule version, exchange rate version, and state threshold version used for each product status determination. The result output module is used to output the product status, status generation time, status validity time, overall confidence level of the evidence package, lower bound of product value, atomic assertions that failed verification, reason for status determination, and version snapshot identifier.

[0060] Example 2 Method

[0061] See Figure 2 As shown, a method for product selection and market analysis across multiple platforms in cross-border e-commerce based on an AI big data model includes steps S1 to S9.

[0062] S1, Acquire data from multiple platforms and generate evidence units

[0063] Acquire multi-platform product, fee, market and settlement data within a sliding time window, and generate evidence units with source lineage, validity time and version identifiers.

[0064] In this embodiment, the sliding time window is set according to product category, transaction frequency, and fee confirmation cycle. For products with high order change frequency, hourly or daily time windows can be used; for logistics settlement, refund observation, and platform settlement data, daily, weekly, or time windows corresponding to the platform billing cycle can be used. Adjacent time windows can have overlapping intervals, which are used to identify data changes and trend direction changes.

[0065] For structured data, standard product objects are identified according to the product identifier mapping table, and local times from different platforms are converted to a unified time zone. Product prices, advertising costs, logistics costs, and settlement amounts are converted to the analysis currency according to the exchange rate version of the corresponding time window. For cases where the same cost has both estimated and confirmed records, prediction status evidence units and confirmed status evidence units are generated separately.

[0066] For comment text, platform announcements, market reports, product descriptions, and logistics rule text, construct model inputs that include task identifiers, allowed output indicator types, target products, target markets, effective time windows, original text content, and output field constraints. When the AI ​​large model output lacks target markets, effective time windows, or original data reference locations, the corresponding semantic parsing results are written into a low-completeness candidate set.

[0067] See Figure 3As shown, the product identifier mapping uses standard product identifiers, platform product identifiers, product variant identifiers, brands, models, specifications, barcodes, and product attributes for matching. When the product identifier mapping result meets the mapping confirmation conditions, the corresponding data is incorporated into the existing product evidence package; when the product identifier mapping result does not meet the mapping confirmation conditions, the corresponding data is written into the mapping set to be confirmed, and the corresponding evidence unit is not incorporated into the existing product evidence package.

[0068] S2, Determine source dependencies

[0069] The source dependency coefficients between evidence units are determined based on the original source identifier, derived source identifier, acquisition link identifier, content fingerprint, and effective time window.

[0070] For evidence unit and evidence unit The source dependence coefficient is determined according to the following formula:

[0071]

[0072] in, For source dependence coefficient, To determine the degree of overlap in source links, To determine the similarity of content fingerprints, To determine the degree of overlap of effective time windows, , and The weight is determined by the decision.

[0073] When the source dependency coefficient reaches the dependency determination condition, an evidence dependency edge is established between the corresponding evidence units, and the corresponding evidence units are assigned to the common upstream source group; when the source dependency coefficient does not reach the dependency determination condition, the corresponding evidence units are used as independent source evidence in the calculation.

[0074] The content fingerprint of textual evidence is generated based on normalized text, entity sets, numerical fields, and time fields. The content fingerprint of structured evidence is generated based on product identifier, indicator type, evidence value, effective time window, and source field. Evidence units with the same source link and identical content fingerprints retain a primary evidence node, and the remaining evidence units are connected to the primary evidence node as derived evidence nodes.

[0075] S3, reducing the confidence level of duplicate evidence.

[0076] Confidence reduction is applied to derived evidence based on the evidence dependency relationships within the common upstream source group.

[0077] No. The valid confidence level of each unit of evidence is determined according to the following formula:

[0078]

[0079] in, The initial confidence level of the evidence unit. To reduce the dependency based on the common upstream source group, This represents the effective confidence level after reduction.

[0080] For evidence units with identical source links and content fingerprints, the dependency reduction amount of the derived evidence is determined according to the direct derivation relationship. For evidence units with partially overlapping source links but containing independent supplementary fields, the common fields are reduced according to the source dependency coefficient, and the independent supplementary fields retain the effective confidence level corresponding to their source reliability.

[0081] For evidence units directly derived from platform settlement bills and without derivation processing, the dependency reduction can be set to zero. For the same page data repeatedly output by multiple aggregation interfaces, the dependency reduction is determined according to the common upstream page identifier and content fingerprint.

[0082] S4, Construct a temporal evidence conflict graph

[0083] A temporal evidence conflict graph is constructed based on the differences in evidence. The temporal evidence conflict graph includes evidence nodes, evidence dependency edges, conflict edges, time continuity edges, and product association edges.

[0084] Evidence units corresponding to the same product, target market, and indicator type are written into the same candidate conflict set, and conflict determination is performed according to the effective time window and version identifier. For numerical evidence, the evidence values ​​are converted to the same unit of measurement and statistical caliber before calculating the numerical differences; for directional evidence, the upward, downward, or stable trend direction is compared; for rule-based evidence, the rule version and effective time are compared; for settlement-based evidence, the predicted cost and confirmed cost are compared; for semantic evidence, the correspondence between semantic parsing results and structured fields or original text is compared.

[0085] The weights of conflicting edges are determined by the following formula:

[0086]

[0087] in, The conflict edge weights between evidence nodes. This refers to the difference in numerical value or trend direction. For version differences, For the effective time window misalignment amount, This represents the semantic or settlement status difference. to Weights for conflict types.

[0088] When the conflict edge weight meets the conflict determination criteria, a valid conflict edge is established, and the conflict edge is included in the reduction term of the overall confidence level of the evidence package. When the conflict edge weight does not meet the conflict determination criteria but is within the candidate conflict interval, a candidate conflict edge is established, and the corresponding candidate conflict state is written into the uncertainty of the evidence package.

[0089] When the evidence version corresponding to a conflicting edge becomes invalid, the invalidation time is written into the conflicting edge, but the corresponding historical conflicting edge is not deleted. When determining the current product status, a valid subgraph corresponding to the current evidence version and the current valid time window is extracted from the temporal evidence conflict graph.

[0090] S5, Calculate the overall confidence level of the evidence package and the lower bound of the commodity value.

[0091] The overall confidence of the evidence package is calculated based on the effective confidence of each evidence unit, the conflict edge weight, the time decay, the historical reliability of the source, the settlement maturity, and the state of missing necessary evidence.

[0092] No. The overall confidence level of the evidence package is determined by the following formula:

[0093]

[0094] in, For the weight of evidence type, For effective confidence level, To ensure the reliability of the source's history, This is a settlement maturity or time-effectiveness coefficient. This represents the number of valid evidence nodes within the evidence package. This refers to the uncertainty arising from the absence of necessary evidence and the conflicting nature of candidate evidence. and This is for adjusting the coefficient.

[0095] The historical reliability of the source is updated based on the deviation records of the same source in historical mature settlement samples. For cost data sources, it is updated based on the deviation between predicted costs and confirmed costs; for commentary semantic sources, it is updated based on the consistency of manual review results and subsequent structured indicators; for market report sources, it is updated based on the deviation between report indicators and subsequent market data.

[0096] The lower bound of the commodity value is determined based on the recognized revenue, the unvested revenue range, the recognized expenses range, the unvested expense range, and the loss range. The lower bound of the commodity value is determined according to the following formula:

[0097]

[0098] in, As the lower bound of commodity value, For revenue already recognized, For the distribution of immature income, For confirmed fees, For the distribution of unripe costs, Distribution of losses resulting from returns, refusal of payment, or additional deductions. Represents distribution of Quantiles The tail probability parameters are determined based on a preset confidence level; Distribution of immature income of Quantiles Distribution of Unripe Costs of Quantiles For loss distribution of Quantiles.

[0099] When the number of historical mature settlement samples reaches the interval calculation condition, the immature revenue interval, immature expense interval, and loss interval are determined based on the historical settlement residuals of the same platform, the same target market, the same logistics method, and the same product category. When the number of historical mature settlement samples does not reach the interval calculation condition, a preset initial interval is used, and the insufficient sample status is written into the uncertainty of the evidence package.

[0100] S6, Decompose AI output and build assertions—Evidence bipartite diagram

[0101] The AI ​​output is decomposed into atomic assertions, an assertion-evidence bipartite graph is constructed, and assertion coverage, the proportion of assertions without evidence, and the intensity of conflict are determined.

[0102] The AI ​​output includes judgments on product demand, competitive status, estimated value, cost and risk, comment risk, market rules, and product handling suggestions. Each atomic assertion corresponds to a judgment that can be supported or opposed by an evidence unit. The atomic assertion includes an assertion identifier, product subject, target market, attribute type, value or direction, applicable time window, conditions for validity, cited evidence identifier, assertion importance, and model output version.

[0103] When an AI output statement includes both demand changes and commodity value judgments, the AI ​​output statement is split into atomic assertions of demand trends and atomic assertions of commodity value. When an AI output statement includes multiple target markets, atomic assertions are generated separately for each target market.

[0104] When an atomic assertion and an evidence unit meet matching conditions in terms of evidence object, indicator type, target market, effective time window, and version identifier, and the evidence value or trend direction of the evidence unit supports the atomic assertion, a supporting edge is established; when the evidence value or trend direction of the evidence unit is opposite to the atomic assertion, an opposing edge is established.

[0105] Assertion coverage is determined by the following formula:

[0106]

[0107] in, To assert coverage, For the number of atomic assertions, For the first The importance of individual atom assertions This indicates whether the atomic assertion has supporting evidence that satisfies the object, time, and version requirements.

[0108] Atomic assertions lacking valid supporting edges are marked as no-evidence assertions. When the confidence levels of both the supporting and opposing edges meet the conflict assertion criteria, the corresponding atomic assertion is marked as a conflict assertion. The conflict intensity is determined based on the importance of the conflict assertion, the weight of the supporting edge, the weight of the opposing edge, and the weight of the corresponding conflicting evidence edge.

[0109] When the evidence identifier cited by the AI ​​output does not exist in the current evidence package version snapshot, the corresponding atomic assertion is marked as "reference does not exist". When the cited evidence exists but the cited version is earlier than the current valid version, the corresponding atomic assertion is marked as "version mismatch". When the reference does not exist or the version mismatch meets the regeneration condition, the AI ​​model input is reconstructed using the current evidence package version snapshot.

[0110] S7, Perform product status determination

[0111] The determination of the commodity status is performed based on the overall confidence level of the evidence package, the lower bound of the commodity value, the assertion coverage, the proportion of assertions without evidence, the intensity of conflict, and the state of missing necessary evidence.

[0112] See Figure 4 and Figure 5 As shown, when the overall confidence level of the evidence package reaches the confidence level processing threshold, the lower bound of the product value reaches the value processing threshold, the assertion coverage reaches the coverage processing threshold, and the proportion of assertions without evidence is lower than the assertion restriction threshold and the conflict intensity is lower than the conflict review threshold, the candidate status of the product is set to the processable status.

[0113] When the evidence package lacks the pre-set necessary evidence, and the product does not meet the judgment conditions for restricted status and manual review status, the candidate status of the product is set to supplementary data status. The supplementary data status records the type of evidence to be supplemented, the target data source, the target time window, the data version requirements, and the supplementation period.

[0114] When the conflict intensity reaches the conflict review threshold, or when a key atomic assertion simultaneously has both supporting and opposing edges that meet the conditions, the candidate state of the product is set to manual review state. The manual review state generates a review task associated with the conflict evidence node, the conflict atomic assertion, and the current evidence package version snapshot.

[0115] When a commodity's lower bound is below the value limit threshold, the proportion of assertions without evidence reaches the assertion limit threshold, there is a confirmed market rule restriction, or the immature cost range exceeds the allowable range, the commodity's candidate status is set to a restricted status. The restricted status records the reason for the restriction, the entry time, the effective time of the status, and the conditions required to lift the restriction.

[0116] When a product has multiple candidate states, the product state is determined in the following order: restricted state, manual review state, supplementary data state, and processable state. If the restricted state is active, lower-priority candidate states are no longer used to override it. If the manual review state is active but the restricted state is not active, supplementary data state or processable state is no longer used to override the manual review state.

[0117] S8, Execution state hysteresis control

[0118] An entry restriction threshold and an exit restriction threshold are set respectively, and the entry restriction threshold and the exit restriction threshold form a hysteresis interval.

[0119] When a product meets the entry restrictions and continuously reaches the preset number of confirmation windows, the product is switched to a restricted state. If a product only meets the entry restrictions within one sliding time window and does not reach the preset number of confirmation windows, the product's previous state remains unchanged, and the pending confirmation restricted state is recorded.

[0120] When a product in a restricted state meets the exit restriction conditions, it does not directly switch to a processable state, but instead enters a cooling-off observation state. The cooling-off observation state records the time of entering the cooling-off observation state, the number of windows that continuously meet the recovery conditions, the source of recovery evidence, and the number of unprocessed conflict edges.

[0121] When two or more platforms simultaneously output recovery evidence with the same direction, it is determined whether the recovery evidence has a common upstream source based on the evidence dependency edges. If the recovery evidence has a common upstream source, it is treated as evidence from a single source in the recovery determination; if the recovery evidence does not have a common upstream source, it is treated as recovery evidence from independent sources in the recovery determination.

[0122] When the exit constraint remains valid for multiple consecutive sliding time windows, independent evidence for recovery exists, and there are no unprocessed high-weight conflict edges, the product is switched from the cool-down observation state to the processable state or the supplementary data state. The processable state or the supplementary data state is determined based on the current evidence completeness, the overall confidence of the evidence package, and the assertion coverage.

[0123] When a high-weight conflict edge reappears during the cooling-off observation period, the lower bound of the commodity value falls below the exit restriction condition, the proportion of assertions without evidence reaches the assertion restriction threshold, or the confirmed settlement fee exceeds the corresponding unripe fee range, the number of continuous recovery windows will be reset to zero, and the commodity will be returned to the restricted state.

[0124] S9, Perform deviation attribution and state recalculation.

[0125] Once the settlement is complete, deviation attribution is performed to identify affected goods and conduct local state recalculation.

[0126] When the order completion status, refund observation period, logistics cost confirmation status, and platform settlement confirmation status meet the maturity conditions, the version snapshot saved when the product status was determined is read, and the predicted revenue, predicted expenses, predicted losses, and predicted product value are compared with the actual settlement results.

[0127] The deviation between actual and predicted results is allocated to the price, advertising costs, logistics costs, taxes, exchange rates, returns, chargebacks, or evidence analysis items. When actual logistics costs differ from predicted logistics costs, the logistics cost deviation is allocated based on the logistics method, chargeable weight, logistics surcharge version, and target market. Deviations caused by product mapping errors, text parsing errors, or AI assertion citation errors are written into the evidence analysis item.

[0128] Update the corresponding source historical reliability, cost range parameters, conflict edge weights, or state threshold versions based on the deviation attribution results. After the parameters are updated, the changed evidence nodes or parameter nodes are identified as starting nodes and propagated along evidence dependency edges, conflict edges, assertion support edges, and product association edges.

[0129] During the propagation process, the current propagation level, visited nodes, and associated products are recorded. When a node with no further associated edges is reached, an associated edge weight lower than the propagation condition, or the maximum propagation level is reached, the corresponding propagation branch is stopped. The product nodes traversed by the propagation are merged into an affected product set, and duplicate product identifiers are removed.

[0130] When the affected product set does not meet the conditions for full recalculation, only the currently valid evidence subgraph is re-extracted from the affected product set, the overall confidence of the evidence package, the lower bound of product value, the assertion coverage, the proportion of assertions without evidence, and the conflict intensity are recalculated, and the product status determination is re-executed.

[0131] A full recalculation is performed when the affected product set meets the conditions for a full recalculation, or when the product identifier mapping rules, basic exchange rate rules, or platform category mapping rules undergo a complete version switch. After a partial or full recalculation is completed, the new state results and version snapshot are saved.

[0132] Example 3 Verification

[0133] This embodiment employs de-identified historical sample replay and perturbation data injection to verify source dependency reduction, AI output verification, product status determination, status hysteresis control, and local status backcalculation. Verification samples include multi-platform product records, order records, advertising invoices, logistics invoices, refund records, review texts, market rule texts, and settlement records.

[0134] The verification equipment includes a data access server, relational data storage nodes, graph data storage nodes, AI model calling nodes, and result verification nodes. Logs record the number of evidence units, the number of evidence-dependent edges, the number of conflicting edges, the number of atomic assertions, the number of assertions without evidence, the number of conflicting assertions, the product status, the state transition time, the number of affected products, and the state backcalculation results.

[0135] Under the condition of common source evidence, data from a platform page enters the system through both the data supply interface and the internal transfer interface, forming demand trend evidence with identical or similar content. Without source dependency reduction, this evidence is separately included in the evidence package; with source dependency reduction, evidence dependency edges are established based on the common page identifier, acquisition link identifier, content fingerprint, and effective time window, and the confidence level of derived evidence is reduced.

[0136] The common source evidence accounts for approximately 18.6% to 21.3% of the candidate requirement evidence. Without source dependency reduction, the common source evidence increases the overall confidence level of the evidence package for some items by approximately 0.071 to 0.094; after performing source dependency reduction, the corresponding increase is approximately 0.014 to 0.027. The calculation results fluctuate by approximately 0.006 to 0.011 between different data batches.

[0137] In the AI ​​output verification scenario, no-evidence demand assertions, old version logistics cost references, and market rule assertions applicable to market mismatches are set in the AI ​​output samples. The AI ​​output is decomposed into atomic assertions of product demand, competitive state, product value, logistics cost, and market rule, and support or opposition edges are established according to product object, target market, effective time window, and version identifier.

[0138] Each product analysis result contains approximately 9 to 16 atomic assertions, with the proportion of assertions without evidence ranging from approximately 6.8% to 11.5%, and the proportion of old version evidence citations ranging from approximately 3.1% to 5.4%. After evidence verification, products corresponding to key assertions without evidence enter a manual review or restricted state, and old version evidence citations trigger AI output regeneration. After manual review, the proportion of atomic assertions with erroneous evidence matching is still approximately 1.7% to 2.6%.

[0139] Under the state priority determination condition, a sample of goods is set up where the following conditions are simultaneously met: insufficient evidence completeness, conflict intensity reaching the review condition, and the lower bound of the goods value falling below the value limit threshold. Without setting a state priority order, different processing threads may output supplementary data status, manual review status, and restricted status respectively. After adopting state priority processing, the goods form a current state in the order of restricted status, manual review status, supplementary data status, and processable status.

[0140] Under short-term market fluctuations, a single-threshold state control method is used as a control, while entry restriction threshold, exit restriction threshold, and continuous window confirmation methods are used as the processing methods in this embodiment. During continuous time window replay, the average number of state transitions per item in the single-threshold method is approximately 4.6 to 5.3 times, while the average number of state transitions per item in the processing method of this embodiment is approximately 1.8 to 2.4 times.

[0141] Under the condition of shared source of recovery evidence, two data interfaces simultaneously output the required recovery evidence, and the two data interfaces correspond to the same upstream page. When source dependency judgment is not performed, the two recovery evidences are treated as two independent sources in the recovery determination; after source dependency judgment is performed, the two recovery evidences are treated as evidence from one source in the recovery determination, and the product remains in a cooling-off observation state until recovery evidence that does not have a common upstream source is received.

[0142] Under delayed settlement conditions, the commodity status is determined using both a fixed estimated cost method and a commodity value lower bound method. The absolute deviation of the commodity value prediction corresponding to the fixed estimated cost method is approximately 8.9% to 12.7%, and the actual settlement coverage rate corresponding to the commodity value lower bound is approximately 91.8% to 94.6%. Samples not covered by the commodity value lower bound include samples with changes in logistics billing weight and cross-period advertising deductions.

[0143] In the partial back-calculation scenario, the logistics surcharge versions corresponding to one platform, two target markets, and some logistics methods are updated. The full back-calculation method recalculates the current set of goods, while the partial back-calculation method starts from the changed fee rule node and determines the set of affected goods along evidence dependency edges, conflict edges, assertion support edges, and goods association edges.

[0144] The affected product set accounts for approximately 12.4% to 15.1% of the current product set. The average processing time for partial back-calculation is approximately 19.6% to 24.8% of the average processing time for full back-calculation, and the number of AI model calls for partial back-calculation is approximately 17.9% to 22.6% of the number of AI model calls for full back-calculation. The consistency rate between the product status of partial back-calculation results and full back-calculation results is approximately 98.4% to 99.1%.

[0145] The verification results are summarized in the table below.

[0146] Increase in confidence due to duplicate sources 0.071～0.094 0.014～0.027 Proportion of assertions without evidence 6.8%～11.5% Block, review or restrict Old version evidence citation ratio 3.1%～5.4% Regenerate or review Average number of status transitions per product 4.6 to 5.3 times 1.8 to 2.4 times Back-calculation of product proportions All products currently available 12.4%～15.1% Relative value of average processing time 1.00 0.196～0.248 Relative value of AI model call count 1.00 0.179～0.226 Consistency ratio with full recalculation status Comparison results 98.4%～99.1%

[0147] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for product selection and market analysis across multiple platforms in cross-border e-commerce based on an AI-powered large-scale model, characterized in that: Includes the following steps: Acquire multi-platform product, fee, market, and settlement data within a sliding time window, and generate evidence units with source lineage, validity period, and version identifiers; reduce the confidence of duplicate evidence based on common upstream sources, construct a time-series evidence conflict graph based on evidence differences, and calculate the overall confidence of the evidence package and the lower bound of product value; decompose the AI ​​output into atomic assertions, establish an assertion-evidence bipartite graph, and determine assertion coverage, the proportion of assertions without evidence, and conflict intensity; determine a processable state when each judgment quantity meets the processable condition, a supplementary data state when the evidence completeness is insufficient, a manual review state when the conflict intensity meets the review condition, and a restricted state when the lower bound of product value or the proportion of assertions without evidence meets the restriction condition; perform deviation attribution after settlement is mature, identify affected products, and perform local state backcalculation.

2. The method according to claim 1, characterized in that, The generation of evidence units with source lineage, validity period, and version identifiers includes: For structured data, we perform commodity identification mapping, currency conversion, time zone alignment, and fee standardization. For unstructured data, we perform semantic parsing and encapsulate the processing results into evidence units that include evidence object, target market, indicator type, evidence value, original source identifier, derived source identifier, collection link identifier, effective time window, version identifier, settlement maturity, and original data reference location. When the commodity identification mapping result does not meet the mapping confirmation conditions, the corresponding data is written into the mapping set to be confirmed, and the corresponding evidence unit is not merged into the existing commodity evidence package.

3. The method according to claim 1, characterized in that, The confidence level for reducing duplicate evidence based on a common upstream source includes: The source dependency coefficient between evidence units is determined based on the original source identifier, derived source identifier, acquisition link identifier, content fingerprint, and effective time window. When the source dependency coefficient reaches the dependency judgment condition, an evidence dependency edge is established between the corresponding evidence units, and confidence reduction is performed on the derived evidence within the common upstream source group. When the source dependency coefficient does not reach the dependency judgment condition, the corresponding evidence unit is treated as independent source evidence in the calculation.

4. The method according to claim 1, characterized in that, The step of constructing a temporal evidence conflict diagram based on evidence differences and calculating the overall confidence level of the evidence package and the lower bound of the commodity value includes: Establish conflict edges between evidence units that correspond to the same product and target market but have numerical deviations, opposite trend directions, misaligned time windows, inconsistent versions, inconsistent settlement statuses, or inconsistent semantic results; when the weight of a conflict edge reaches the conflict determination condition, the corresponding conflict edge is included in the reduction item of the overall confidence of the evidence package; determine the lower bound of the product value based on the confirmed revenue, unvested revenue range, confirmed expenses, unvested expense range, and loss range.

5. The method according to claim 1, characterized in that, The process of decomposing AI output into atomic assertions and establishing an assertion-evidence bipartite graph includes: Atomic assertions, including product subject, target market, attribute type, value or direction, applicable time window, establishment conditions, and referenced evidence identifier, are extracted from the AI ​​output. When the atomic assertion and the evidence unit meet the matching conditions in terms of evidence object, indicator type, target market, effective time window, and version identifier, a supporting edge or an opposing edge is established. When a key atomic assertion does not have a valid supporting edge, the corresponding key atomic assertion is marked as an assertion without evidence.

6. The method according to claim 1, characterized in that, The determination of the product status includes: When the overall confidence level of the evidence package, the lower bound of the commodity value, and the assertion coverage rate reach the corresponding processing thresholds, and the proportion of assertions without evidence and the intensity of conflict are lower than the corresponding restriction thresholds, the status is determined to be processable. When necessary evidence is missing and the restriction conditions and manual review conditions are not met, the status is determined to be supplementary data. When the intensity of conflict reaches the review threshold or the key atomic assertion has both supporting and opposing edges that meet the conditions, the status is determined to be manually reviewed. When the lower bound of the commodity value is lower than the value restriction threshold or the proportion of assertions without evidence reaches the assertion restriction threshold, the status is determined to be restricted. When there are multiple determination results, the commodity status is determined in the order of restricted status, manual review status, supplementary data status, and processable status.

7. The method according to claim 6, characterized in that, The method further includes a state hysteresis control step, which includes: An entry restriction threshold and an exit restriction threshold are set respectively, and the entry restriction threshold and the exit restriction threshold form a hysteresis interval; when a product meets the entry restriction conditions and continuously reaches the preset number of confirmation windows, it enters the restriction state; when a product in the restriction state meets the exit restriction conditions, it enters the cooling observation state; when the exit restriction conditions are continuously met within multiple consecutive sliding time windows, and there is independent source recovery evidence and no unprocessed high-weight conflict edges, it exits the cooling observation state.

8. The method according to claim 6, characterized in that, The product status determination also includes: When two or more platforms simultaneously output recovery evidence in the same direction, it is determined whether the recovery evidence has a common upstream source based on the evidence dependency edge. If the recovery evidence has a common upstream source, the recovery evidence is used as evidence from one source in the recovery determination. If the recovery evidence does not have a common upstream source, and the recovery evidence meets the recovery conditions within a continuous sliding time window, the corresponding product is switched from the cooling observation state to the processable state or the supplementary data state.

9. The method according to claim 1, characterized in that, The process of identifying the affected goods and performing local state recalculation includes: When settlement deviations, evidence versions, or fee rules change, the changed evidence node or parameter node is identified as the starting node, and propagation proceeds along evidence dependency edges, conflict edges, assertion support edges, and product association edges. Propagation stops when it reaches a preset termination condition, and the product nodes traversed by the propagation are identified as the affected product set. If the affected product set does not meet the full recalculation condition, the product status is recalculated only for the affected product set. When the affected product set meets the full recalculation condition, a full recalculation is performed.

10. A cross-border e-commerce multi-platform product selection and market analysis system based on AI large-scale models, characterized in that, include: The evidence processing module is used to acquire data from multiple platforms, generate evidence units, and reduce the confidence level of duplicate evidence based on common upstream sources. The graph processing module is used to construct a temporal evidence conflict graph and calculate the overall confidence of the evidence package and the lower bound of the commodity value. The output verification module is used to decompose the AI ​​output into atomic assertions, establish an assertion-evidence bipartite graph, and determine assertion coverage, the proportion of assertions without evidence, and conflict intensity. The status determination module is used to determine the product as a processable state, a supplementary data state, a manual review state, or a restricted state when each determination quantity meets the corresponding conditions. The back-calculation module is used to perform deviation attribution after settlement is mature, determine the set of affected goods along the graph association edges, and perform partial state back-calculation or full back-calculation based on whether the set of affected goods meets the conditions for full back-calculation.

Images