A federated learning-based automobile extended warranty cross-domain data collaborative analysis method

Abstract

The application provides a kind of based on federated learning's car extended warranty cross-domain data collaborative analysis method, comprising: each participating agency local reading car extended warranty business record, generates standardization business event set, controlled association key and event theme node set;Event theme node set is based on the construction relationship candidate set, local event theme hyperedge structure, local event theme hypergraph and local graph structure grouping;Based on local graph structure grouping, construct local training sample set, train local federated model, generate event theme embedding vector, hyperedge structure signature, training side local risk response parameter and upload summary package;Coordination end is based on upload summary package and executes cross-agency theme alignment, abnormal behavior chain candidate generation and global aggregation, obtains global event theme hypergraph, global collaborative model and global issue rule package;Each participating agency is based on global event theme hypergraph, global collaborative model and global issue rule package and updates local federated model.

Description

Technical Field

[0001] This invention relates to the field of automotive extended warranty data processing and machine learning collaborative analysis technology, and in particular to a cross-domain data collaborative analysis method for automotive extended warranty based on federated learning. Background Technology

[0002] With the increasing connectivity of vehicles, extended warranty business data is typically scattered across different stakeholders, including OEM after-sales systems, 4S dealerships, third-party repair shops, and extended warranty service providers. Existing methods for identifying anomalies have significant limitations: one type of method relies solely on local records from a single institution, making it difficult to reflect the complete anomaly evolution path of the target's cross-institutional flow; another type attempts to aggregate raw data from multiple institutions for centralized analysis, but this involves cross-domain sharing of sensitive information, facing extremely high data compliance pressures and privacy risks.

[0003] To resolve the conflict between data silos and privacy protection, federated learning has been introduced to achieve collaborative modeling across multiple agencies. However, most existing federated learning solutions focus on sharing model parameters or gradients, lacking the ability to structurally represent cross-agency business chains. In real-world scenarios, abnormal behavior of the same extended warranty object is often broken down into scattered actions such as detection, repair, and parts replacement. Observing any single local record in isolation usually only shows normal business operations, making it easy to miss detections. Only by deeply associating records from multiple agencies according to time, mileage, and business semantics can cross-agency behavioral chains with obvious abnormal characteristics be identified. However, existing federated technologies struggle to achieve this reconstruction of associations without sharing detailed information.

[0004] Therefore, this invention proposes a cross-domain data collaborative analysis method for automotive extended warranty based on federated learning. The information disclosed in the background section is only for enhancing understanding of the background of this disclosure and may therefore contain prior art information that is not common knowledge to those skilled in the art. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by providing a cross-domain data collaborative analysis method for automotive extended warranty based on federated learning, thereby solving the technical problems mentioned in the background section.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A collaborative data analysis method for cross-domain automotive extended warranty based on federated learning includes the following steps: S1. Each participating institution locally reads the car extended warranty business records and generates a standardized business event set, controlled association keys, and event theme node set; S2. Construct a candidate set of relations, a local event topic hyperedge structure, a local event topic hypergraph, and a local graph structure based on the event topic node set; S3. Construct a local training sample set based on the local graph structure grouping, train the local federated model, generate event motif embedding vectors, hyperedge structure signatures, training-side local risk response parameters, and upload summary packages; S4. The coordinating end performs cross-organizational topic alignment, abnormal behavior chain candidate generation, and global aggregation based on the uploaded summary package to obtain a global event topic hypergraph, a global collaboration model, and a global rule package. S5. Each participating institution updates its local federated model based on the global event theme hypergraph, global collaboration model, and global rule package, generates local abnormal behavior chain fragments and application-side local risk response results, and performs cross-institutional merging on local abnormal behavior chain fragments corresponding to the same target extended warranty object, outputting cross-institutional abnormal behavior chain analysis results and extended warranty risk assessment results.

[0007] S1 specifically includes: Each participating institution locally reads the vehicle extended warranty business records, performs verification, completion, sorting, and duplicate deletion on the local identifier of the business object, event type, event occurrence time, event mileage value, institution type, action category, result status, and source record type, and generates a standardized business event set and a controlled association key; under the same controlled association key constraint, the standardized business event set is segmented according to the event interval, mileage increment, and event type switching relationship, and a candidate event theme fragment set is generated; the candidate event theme fragment set is categorized and attribute encoded to generate an event theme node set.

[0008] S2 specifically includes: under the same controlled association key constraint, performing bucket verification on the event mother topic node set, and generating a candidate set of relations according to the order of occurrence stage and mileage interval; performing deletion, relation weight calculation and relation shaping on the candidate set of relations to generate a local event mother topic hyperedge structure; constructing a local event mother topic hypergraph based on the event mother topic node set and the local event mother topic hyperedge structure, and generating local graph structure groups according to the controlled association key and connectivity relationship, wherein the local graph structure group includes at least a node identifier set, a hyperedge identifier set, a group start time field, a group end time field, a group start mileage field, a group end mileage field and a group ending status field.

[0009] S3 specifically includes: generating a local training sample set containing business outcome labels and sample weight values ​​based on the local event motif hypergraph, local graph structure grouping, and corresponding case closure records; extracting node feature matrices and hyperedge association matrices based on the local training sample set, training a local federated model, and generating local event motif representation results and local risk response results; performing normalization, pruning, quantization compression, and encapsulation processing on the local event motif representation results and local risk response results to generate an upload summary package including a controlled association key summary, event motif embedding vector, hyperedge structure signature, training-side local risk response parameters, and node boundary information summary.

[0010] S4 specifically includes: the coordinator reading the upload summary packages uploaded by each participating institution, establishing a cross-institutional alignment task record, and generating cross-institutional topic alignment results based on controlled association key summaries, event topic embedding vectors, hyperedge structure signatures, training-side local risk response parameters, and node boundary information summaries; generating a candidate set of abnormal behavior chains based on the cross-institutional topic alignment results; and performing global aggregation based on the cross-institutional topic alignment results, the candidate set of abnormal behavior chains, event topic embedding vectors, hyperedge structure signatures, training-side local risk response parameters, sample quantity values, and sample weight statistics to generate a global event topic hypergraph, a global collaborative model, and a globally distributed rule package.

[0011] S5 specifically includes: each participating institution reads the global event theme hypergraph, the global collaboration model, and the globally distributed rule package; establishes a record of the current extended warranty analysis task; updates the local federated model and extracts the target local event theme subgraph; generates a set of local abnormal behavior chain fragments and a set of application-side local risk response results based on the target local event theme subgraph; and performs cross-institutional merging, conflict resolution, and final risk assessment on the local abnormal behavior chain fragments corresponding to the same target extended warranty object based on the set of local abnormal behavior chain fragments, the set of application-side local risk response results, the global event theme hypergraph, and the globally distributed rule package, outputting cross-institutional abnormal behavior chain analysis results and extended warranty risk assessment results.

[0012] The beneficial effects of this invention are as follows: This invention transforms the vehicle extended warranty business records locally in each participating institution into a standardized set of business events, a set of candidate event motif fragments, and a set of event motif nodes, and further constructs a local event motif hypergraph. This enables dispersed and heterogeneous extended warranty business data to form a unified structured expression without aggregating the original extended warranty business records, thereby improving the alignment and analyzability of cross-institutional business events.

[0013] This invention performs cross-organizational motif alignment through controlled association keys, event motif embedding vectors, hyperedge structure signatures, and node boundary information digests. It can uniformly match similar business events in different participating organizations without directly sharing vehicle identity information and customer identity information, thereby balancing data compliance requirements and cross-domain collaborative analysis capabilities.

[0014] This invention, through the generation of candidate sets of abnormal behavior chains, extraction of local abnormal behavior chain fragments, and merging of cross-organizational abnormal behavior chains, can reconstruct detection, repair, replacement, and compensation actions that appear normal at a single point but have abnormal evolution characteristics as a whole into complete cross-organizational abnormal behavior chains, thereby improving the ability to identify hidden abnormal extended warranty behaviors.

[0015] This invention establishes a closed-loop collaborative mechanism of "local training-global aggregation-local application-result feedback" through local training sample set construction, local federated model training, global collaborative model aggregation, and local federated model updating, which is conducive to improving the stability and continuous optimization capability of extended warranty risk analysis results.

[0016] This invention employs hierarchical calculation of local segment risk values, chain risk support values, and final risk values. It outputs cross-institutional abnormal behavior chain analysis results and extended warranty risk assessment results from both the local institution level and the overall cross-institutional level, thereby improving the accuracy and interpretability of extended warranty risk assessment. It enables multi-institutional cross-domain joint modeling and knowledge sharing without the need for centralized aggregation of original extended warranty business records. This reduces the privacy leakage risk associated with cross-domain transmission of original business data and enhances the practical application value of abnormal behavior identification and risk assessment in automotive extended warranty scenarios. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of a cross-domain data collaborative analysis method for automobile extended warranty based on federated learning, according to the present invention. Detailed Implementation

[0018] 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.

[0019] Example: Figure 1 As shown, this embodiment provides a cross-domain data collaborative analysis method for automobile extended warranty based on federated learning, including the following steps: S1. Each participating institution locally reads the car extended warranty business records and generates a standardized business event set, controlled association keys, and event theme node set; S2. Construct a candidate set of relations, a local event topic hyperedge structure, a local event topic hypergraph, and a local graph structure based on the event topic node set; S3. Construct a local training sample set based on the local graph structure grouping, train the local federated model, generate event motif embedding vectors, hyperedge structure signatures, training-side local risk response parameters, and upload summary packages; S4. The coordinating end performs cross-organizational topic alignment, abnormal behavior chain candidate generation, and global aggregation based on the uploaded summary package to obtain a global event topic hypergraph, a global collaboration model, and a global rule package. S5. Each participating institution updates its local federated model based on the global event theme hypergraph, global collaboration model, and global rule package, generates local abnormal behavior chain fragments and application-side local risk response results, and performs cross-institutional merging on local abnormal behavior chain fragments corresponding to the same target extended warranty object, outputting cross-institutional abnormal behavior chain analysis results and extended warranty risk assessment results.

[0020] S1 specifically includes the following sub-steps: S110. Each participating institution has a local pre-built extended warranty business rule library. The extended warranty business rule library contains triggering logic and prohibition rules between mother question types. The parsing agent in the multi-agent system reads the car extended warranty business records locally in each participating institution. The car extended warranty business records are limited to original records that can represent the changes in the business status of the target extended warranty object during the extended warranty period, including at least code reporting records, inspection records, repair records, parts replacement records, follow-up visit records, and claims application records. The parsing agent first performs field verification on various types of original records, retaining nine core fields: business object local identifier, event record identifier, event type, event occurrence time, event mileage value, organization type, action category, result status, and source record type.

[0021] Among them, the business object local identifier is used to uniquely identify the target extended warranty object within the organization, the event record identifier is used to eliminate duplicate records, the event type is used to distinguish between reporting codes, inspections, repairs, parts replacements, follow-up visits, and claims applications, the event occurrence time and event mileage value are used together to support subsequent segmentation, the organization type is used to characterize the record source entity, the action category and result status are used to characterize the business action and its outcome, and the source record type is used to trace the source of the field.

[0022] The parsing agent deletes records that lack local identifiers for business objects, lack event occurrence times, have an event occurrence time earlier than the extended warranty contract effective time, have an event occurrence time later than the current processing time, or have the same event record identifier and all nine core fields are identical, to prevent invalid records from entering subsequent topic extraction. For the retained records, the parsing agent sorts them in ascending order of event occurrence time. If multiple records exist at the same time, they are arranged in the order of reporting records, inspection records, maintenance records, parts replacement records, follow-up records, and claims application records to unify the sequence of events across organizations.

[0023] Afterwards, the parsing agent performs field completion on the sorted records. For records with missing event mileage values ​​but adjacent valid mileage values, it uses the nearest neighbor interpolation method to complete the field completion. For records with missing organization types but whose source record type uniquely corresponds to the organization type, it completes the field completion according to the mapping table. For example, when the source record type is 4S work order, the organization type is completed as 4S store.

[0024] After completion, the parsing agent converts all kinds of records into standardized business events, forming a standardized business event set. Based on the business object local identifier, extended warranty contract identifier, and vehicle stability identifier fields, it generates a controlled association key by concatenating them in a fixed order, removing null values, and processing restricted summaries. The vehicle stability identifier field is preferably at least one of the following: the chassis number summary field, the engine number summary field, or the unified object number of the extended warranty platform. The purpose of defining the controlled association key is to enable the same target extended warranty object to form a consistent identifier that can be matched in a controlled manner on different participating institutions, while not directly exposing vehicle identity information and customer identity information.

[0025] S110 outputs a standardized business event set and a controlled association key, which S120 reads to perform continuous event segmentation. S130 binds the fragments to the controlled association key to form an event motif node set, and the controlled association key continues to be used by S410 to perform cross-organization motif alignment via subsequent digest encapsulation links.

[0026] S120. The topic extraction agent reads the standardized business event set and identifies candidate event topic fragments according to event continuity under the same controlled association key constraint. The topic extraction agent first verifies whether two adjacent standardized business events belong to the same target extended warranty object, and then calculates the event interval and mileage increment between adjacent events. The calculation formulas are as follows: ; In the formula, Indicates the i-th standardized business event and the i-th... The event interval between standardized business events; Indicates the occurrence time of the i-th standardized business event; Indicates the first The time of occurrence of a standardized business event.

[0027] ; In the formula, Indicates the i-th standardized business event and the i-th... Mileage increments between standardized business events; This represents the event mileage value corresponding to the i-th standardized business event; Indicates the first The event mileage value corresponding to each standardized business event.

[0028] Mother Topic Extraction Agent and Based on the segmentation criteria and combined with the event type switching relationship, it is determined whether to disconnect the event segment: when When the time exceeds a preset time threshold, it is segmented into different candidate event theme fragments. The preset time threshold can be set to 30 days. When the distance exceeds a preset mileage threshold, the event is segmented into different candidate event theme fragments. The preset mileage threshold can be set to 3000 kilometers. When two adjacent standardized business events have event types that span different business chains and whose action categories do not belong to the same maintenance chain, segmentation is also performed. For consecutive events that do not trigger the segmentation condition, the theme extraction agent aggregates them into the same candidate event theme fragment according to business semantics. Among them, the same candidate event theme fragment is limited to consecutive event segments that can jointly represent a single local extended warranty behavior pattern.

[0029] To avoid the accidental deletion of critical abnormal information, the mother topic extraction agent implements retention rules for records within the fragment: for duplicate detection records with the same action category, they are not directly deleted, but the earliest and most recent records are retained to represent the duplicate detection trend; for maintenance records with the result status of processing interruption, the processing interruption record is deleted when there are maintenance records in the same chain with the result status of completion or rejection, so as to avoid the subsequent mother topic merging being interfered with by intermediate states.

[0030] After segmentation, the topic extraction agent labels each segment as a detection precursor segment, maintenance transition segment, replacement response segment, repeat visit segment, or compensation trigger segment according to the dominant event type and action evolution order. Each candidate event topic segment is then bound to its corresponding controlled association key, forming a candidate event topic segment set. For example, if the same target extended warranty object experiences one code report, two detections, and one repeat visit without replacement within 12 days, it can be classified as a detection precursor segment. S120 outputs the candidate event topic segment set for S130 to perform category merging, attribute encoding, and node mapping.

[0031] S130. The topic extraction agent reads the set of candidate event topic fragments. First, it verifies whether each candidate event topic fragment has the following characteristics: fragment type, fragment start time field, fragment end time field, fragment start mileage field, fragment end mileage field, dominant action category, and controlled association key. Candidate event topic fragments that are missing any of these characteristics are not merged and are directly deleted to ensure that event topic nodes are built on fragments with complete fields. After that, the topic extraction agent performs category merging and attribute encoding on the retained fragments and maps the candidate event topic fragments that meet the unified business semantic conditions to event topic nodes.

[0032] To ensure the merge rule is executable, the mother topic extraction agent calculates the merge consistency between any two candidate event mother topic fragments, using the following formula: ; In the formula, This represents the merge consistency between candidate event mother theme fragment a and candidate event mother theme fragment b; Indicates the consistency of the parent question type; Indicates the consistent value at the occurrence stage; This indicates the consistency of the action intensity value; , and These represent the consistency weights for the parent topic type, the occurrence stage, and the action intensity, respectively.

[0033] Among them, the consistency value of the mother question type The value is 1 when the two candidate segments have the same theme type and 0 when they do not; the consistency value is determined by the occurrence stage. The consistency value of motion intensity is obtained by calculating the intersection-over-union ratio (IoU) between the time intervals of the two segments. The ratio of the number of underlying standardized business events contained in the two segments is used to determine the outcome.

[0034] In this embodiment, it can be , , Set to 0.4, 0.35, 0.25; when When the value is greater than or equal to the preset merging threshold of 0.85, it is determined that two candidate event theme fragments are mapped to the same event theme node; when When the value is less than 0.85, it is mapped to different event mother node.

[0035] If the parent topic type is the same but the occurrence stage is different, then it is forcibly mapped to different event parent topic nodes; if the parent topic type and occurrence stage are the same but the institution type is different, then it is mapped to the same event parent topic node, and the source institution type set is recorded in the institution type label to retain cross-institution source information.

[0036] After merging, the topic extraction agent writes the following information to each event topic node: event topic node identifier, topic type label, occurrence stage label, mileage interval label, organization type label, controlled association key, node start time field, node end time field, node start mileage field, node end mileage field, and number of source fragments, forming an event topic node set. The number of source fragments is used to characterize the support strength of the event topic node in the local event stream and is used by S220 to determine the relation weight.

[0037] S130 outputs a set of event motif nodes, which S210 uses to identify temporal relationships, co-occurrence relationships, and causal relationships, and which S510 can directly call when extracting the target local event motif subgraph.

[0038] S2 specifically includes the following sub-steps: S210: The graphing agent in the multi-agent system reads the event topic node set output by S130. First, it performs bucket verification on the event topic nodes according to the controlled association key. Only event topic nodes with the same controlled association key and simultaneously possessing the event topic node identifier, topic type label, occurrence stage label, mileage interval label, organization type label, node start time field, node end time field, node start mileage field, node end mileage field, and number of source fragments are retained. Event topic nodes with missing controlled association keys, duplicate event topic node identifiers and identical attribute fields, or zero number of source fragments are deleted to prevent invalid nodes from entering relation construction.

[0039] After verification, the mapping agent, under the same controlled association key constraint, rearranges the nodes according to the occurrence stage order and mileage interval order of the event theme nodes, and generates candidate relationship records in three combinations: adjacent node pairs, co-occurrence node pairs, and preceding-following driving node pairs. Adjacent node pairs are used to identify temporal relationships, co-occurrence node pairs are used to identify co-occurrence relationships, and preceding-following driving node pairs are used to identify causal relationships. Co-occurrence node pairs are limited to two event theme nodes under the same controlled association key, with a time interval not exceeding a preset time window, and overlapping or connecting mileage intervals. Preceding-following driving node pairs are limited to those where the theme type label of the preceding node can trigger the theme type label of the following node in the extended warranty business rule base; for example, detecting a leading segment can trigger a maintenance transition segment, and a maintenance transition segment can trigger a parts replacement response segment.

[0040] The graphing agent writes the starting event mother node identifier, the target event mother node identifier, the candidate relation type, the candidate relation direction, the time interval value, the mileage jump variable, the combination value of the number of source fragments, and the controlled association key for each group of nodes to form a relation candidate set.

[0041] To ensure a consistent calculation method for time interval values ​​and mileage jump variables, the graph-constructing agent calculates the temporal adjacency score and mileage adjacency score for node pairs, respectively. The temporal adjacency score is calculated using the following formula: ; In the formula, This represents the temporal adjacency score between the starting event parent node i and the target event parent node j; This represents the time interval between the boundary times corresponding to the two event parent nodes; This represents the upper limit of the preset time window. A time adjacency score less than 0 is recorded as 0. The mileage adjacency score is calculated using the following formula: ; In the formula, This represents the mileage adjacency score between the starting event parent node i and the target event parent node j; This represents the mileage jump variable between the boundary mileages corresponding to the two event parent nodes; This indicates the upper limit of the preset mileage window. A mileage adjacency score less than 0 is recorded as 0.

[0042] For example, when a target extended warranty object first experiences a detection pilot segment and then a maintenance transition segment within 18 days, and the corresponding mileage jump variable is 1200 kilometers, the preset time window upper limit is 30 days, and the preset mileage window upper limit is 3000 kilometers, a candidate time sequence relationship record can be formed and entered into S220.

[0043] S210 outputs a candidate set of relations, which is then used by S220 to perform relation filtering, relation shaping, and relation weight calculation.

[0044] S220: The graphing agent reads the candidate relationship set output by S210 and first verifies whether each candidate relationship record simultaneously possesses the following: starting event mother node identifier, target event mother node identifier, candidate relationship type, candidate relationship direction, time adjacency score, mileage adjacency score, source fragment quantity combination value, and controlled association key. Candidate relationship records with empty starting event mother node identifier, empty target event mother node identifier, inconsistent controlled association key, candidate relationship direction reversed with the occurrence stage order, or marked as prohibited by the extended warranty business rule base are deleted. Among them, prohibited direct connection includes at least two situations: the claim trigger fragment directly points back to the detection lead fragment, and the repeat visit fragment directly crosses the maintenance transition fragment and points back to the replacement response fragment.

[0045] After deletion, the graphing agent calculates the relation weights for the remaining candidate relation records. The relation weights are calculated using the following formula: ; In the formula, This represents the weight of the relationship between the starting event node i and the target event node j; Indicates the temporal adjacency score; Indicates the mileage adjacency score; The source fragment indicates the supporting score; , and These represent the weights for time adjacency score, mileage adjacency score, and source segment support score, respectively. In this embodiment, it can be , and Set to 0.4, 0.35, and 0.25.

[0046] The source fragment support score is obtained by normalizing the number of source fragments for the starting event mother node and the target event mother node, and is used to characterize whether the candidate relationship is supported by a sufficient number of original fragments. The graph-constructing agent deletes candidate relationship records whose relationship weights are less than a preset relationship weight threshold of 0.6 to avoid occasional co-occurrence or weakly associated nodes from entering subsequent hyperedge construction.

[0047] For records where the candidate relation type is co-occurrence, the graphing agent further calculates the co-occurrence support, which is expressed by the following formula: ; In the formula, This represents the co-occurrence support of the starting event mother node i and the target event mother node j in the k-th local graph structure group; This indicates the number of times two event parent nodes appear together in the same object's event chain; This represents the total number of event parent nodes in the event chain of the k-th object. A candidate relation type is retained as a co-occurrence relation only if the co-occurrence support is greater than or equal to the preset co-occurrence support threshold of 0.3; otherwise, the candidate relation record is deleted.

[0048] Subsequently, the graphing agent defines the relationships for the retained records based on the mother topic type label, occurrence stage label, and candidate relationship type: when the detection precursor segment precedes the maintenance transition segment or replacement response segment, it is preferentially defined as a temporal relationship; when the preceding node can trigger the subsequent node in the business rule base, it is defined as a causal relationship; when two repeated segments occur together multiple times in a short period of time, it is defined as a co-occurrence relationship. If the mother topic node of the same target event corresponds to multiple preceding candidate relationship records, the one with the higher relationship weight is preferred; if the relationship weights are the same, the one with the smaller time interval value is preferred; if they are still the same, the one with the larger combination value of the number of source segments is preferred.

[0049] After the modeling is finalized, the graphing agent merges multiple modeled relation records that are interconnected under the same controlled association key and can jointly represent a single local extended protection behavior chain into a local event theme hyperedge structure, and writes a hyperedge identifier, node identifier set, relation type set, relation direction set, weight set, and controlled association key for each local event theme hyperedge structure.

[0050] S220 outputs the local event mother hyperedge structure, which is used by S230 to construct the local event mother hypergraph and local graph structure grouping.

[0051] S230: The graphing agent reads the event motif node set output by S130 and the local event motif hyperedge structure output by S220. First, it verifies whether the node identifier set in each local event motif hyperedge structure can find a unique corresponding node in the event motif node set, and whether all nodes in the node identifier set have a consistent controlled association key. Then, it deletes local event motif hyperedge structures that have dangling node identifiers, inconsistent controlled association keys, empty relation type sets, or empty weight sets.

[0052] After verification, the graph-constructing agent calculates the hyperedge confidence of each local event parent hyperedge structure. The hyperedge confidence is expressed by the following formula: ; In the formula, This represents the superedge confidence of the superedge structure of the e-th local event mother topic; This represents the number of stereotyped relation records that constitute the hyperedge structure of the e-th local event mother topic; This represents the relationship weight between the starting event node i and the target event node j in the e-th local event parent hyperedge structure. The graph-constructing agent deletes local event parent hyperedge structures with a hyperedge confidence score less than the preset hyperedge confidence score threshold of 0.65 to prevent low-confidence local associations from entering the training samples.

[0053] Subsequently, the graph-constructing agent writes the retained event motif nodes and local event motif hyperedge structures into a unified graph structure container to construct a local event motif hypergraph. The local event motif hypergraph is limited to a set of object-level graph structures formed locally within the same participating institution, with event motif nodes as node objects and local event motif hyperedge structures as hyperedge objects. Its function is to provide the S310 with directly trainable structured input.

[0054] The graph-building agent further performs object-level aggregation on the local event theme hypergraph according to the controlled association key and connectivity relationship, generating local graph structure groups. Among them, the set of event theme nodes and their corresponding hyperedge sets that can be connected through the effective local event theme hyperedge structure under the same controlled association key are merged into one local graph structure group. If there are multiple disconnected connected subgraphs under the same controlled association key, multiple local graph structure groups are generated respectively. If the time break length inside a connected subgraph exceeds the preset time break threshold, such as 45 days, it is split into multiple local graph structure groups according to the time break position.

[0055] For isolated event mother topic nodes that do not contain any valid local event mother topic hyperedge structures, the graph-constructing agent does not directly write them into the local graph structure group, but writes them into the isolated node set; only when the mother topic type label of the isolated event mother topic node is a compensation trigger fragment or a replacement response fragment is it retained as a single node local graph structure group, so as to retain high-risk final state samples.

[0056] Finally, the graph-building agent writes the local graph structure group identifier, controlled association key, node identifier set, super-edge identifier set, group start time field, group end time field, group start mileage field, group end mileage field, and group final status field for each local graph structure group.

[0057] For example, when a target extended warranty object forms a connected subgraph of "detection pilot segment - maintenance transition segment - replacement response segment" under the same controlled association key, the connected subgraph can generate a local graph structure group; if a claim trigger segment appears independently after 60 days, the claim trigger segment can be retained as a single-node local graph structure group.

[0058] S230 outputs the local event theme hypergraph and local graph structure groupings, which are then read by S310 to construct the local training sample set, and provided by S510 to extract the target local event theme subgraph.

[0059] S3 specifically includes the following sub-steps: S310: The training agent in the multi-agent system reads the local event theme hypergraph and local graph structure groups output by S230, and reads the local completed compensation conclusion, rejection conclusion, and normal case closure conclusion corresponding to each local graph structure group. The training agent first verifies whether each local graph structure group simultaneously possesses a local graph structure group identifier, controlled association key, node identifier set, hyperedge identifier set, group start time field, group end time field, group start mileage field, group end mileage field, and group outcome status field; for local graph structure groups with missing identifiers, missing controlled association keys, empty node identifier sets, empty hyperedge identifier sets, and whose outcome status field does not belong to the compensation trigger fragment or replacement response fragment, deletion is performed.

[0060] The training agent re-verifies the correspondence between case closure records and local graph structure groups, retaining only those records that can be uniquely matched through controlled association keys, time intervals, and mileage intervals. Records that are not yet closed, lack a final effective conclusion, or contain both payout and rejection conclusions within the same local graph structure group without a specified final effective time are deleted. For retained samples, the training agent generates business outcome labels based on the final effective conclusion, with payout conclusions corresponding to high-risk labels, rejection conclusions to disputed labels, and normal case closure conclusions to low-risk labels. If a payout revocation record exists, the business outcome label is reset based on the final effective conclusion after revocation.

[0061] The training agent then writes a sample weight value for each local graph structure group. The sample weight value is used to characterize the participation intensity of the local graph structure group in this round of training. When the payout amount is higher than the preset amount threshold or the payout trigger segment corresponding to the group's outcome status field, the sample weight value is increased. For example, the basic weight value of 1 can be adjusted to 1.5.

[0062] After completing the above processing, the training agent merges the local graph structure groupings, business outcome labels, and sample weight values ​​into a local training sample set. The local training sample set is defined as the sole input object for this round of local federated model training. Each local training sample includes at least the local graph structure grouping identifier, controlled association key, node identifier set, hyperedge identifier set, business outcome label, and sample weight value.

[0063] S310 outputs a local training sample set for S320 to read and execute local federated model training.

[0064] S320: The training agent reads the local training sample set output by S310, and first verifies whether the node identifier set and hyperedge identifier set in each local training sample can find a unique corresponding node object and hyperedge object in the local event topic hypergraph; deletes local training samples with dangling node identifiers, dangling hyperedge identifiers, empty business outcome labels, or sample weight values ​​exceeding the preset range.

[0065] The training agent extracts node feature matrices, hyperedge correlation matrices, and business outcome labels from the retained local training samples. The node feature matrix is ​​encoded by the mother topic type label, occurrence stage label, mileage interval label, institution type label, and number of source segments. The hyperedge correlation matrix is ​​used to characterize the connection relationship between nodes and hyperedges. The business outcome labels are used to constrain the output direction of the local federated model.

[0066] The training agent inputs the node feature matrix and hyperedge association matrix into the local federated model to obtain the node representation vectors corresponding to each event topic node. Then, it aggregates the node representation vectors within the same local graph structure group to form a grouped representation vector. The aggregation method is average aggregation, and the calculation formula is: ; In the formula, The grouping representation vector represents the k-th local graph structure group; This represents the number of event topic nodes in the k-th local graph structure group; Let represent the node representation vector corresponding to the i-th event mother node. The trained agent then calculates the local risk response value based on the grouped representation vectors, using the following formula: ; In the formula, This represents the local risk response value of the k-th local graph structure group; This represents the Sigmoid activation function; This represents the risk mapping weight vector in the local federated model; The term "bias" is indicated by the superscript "T", which represents the transpose of the vector. The closer the local risk response value is to 1, the closer the local graph structure group is to the high-risk payout evolution path. For example, when a local graph structure group contains a continuous structure of "detection pilot segment - maintenance transition segment - replacement response segment" and corresponds to a high-weight sample, its local risk response value can be higher than 0.8.

[0067] The training agent calculates the training loss based on the difference between the local risk response value and the business outcome label, and adjusts the loss contribution according to the sample weight value to ensure that high-payout low-frequency samples are not overwhelmed by normal settlement samples. When the loss change in two consecutive local iterations is less than the preset loss threshold, or when the preset local training round limit is reached, the local training in this round is stopped.

[0068] After training, the trained agent outputs local event topic representation results and local risk response results. The local event topic representation results are defined as the set of node representation vectors corresponding to each event topic node, as well as the controlled association key summary, node start time field, node end time field, node start mileage field, node end mileage field, and occurrence stage label summary corresponding to each node representation vector. The local risk response results are defined as the set of local risk response values, risk levels, and prediction confidence scores corresponding to each local graph structure group.

[0069] S320 outputs local event motif representations and local risk response results for S330 to perform digest compression and restricted sharing processing.

[0070] Specifically, the local federated model employs a hypergraph neural network structure, which extracts high-order association features of nodes through a multi-layer message passing mechanism. In each iteration, the node representation vector is aggregated based on its associated hyperedges and subjected to a nonlinear transformation in conjunction with the hyperedge weights. Its computational logic follows: ; Where H is the node representation matrix, A is the hyperedge incidence matrix, D and B are the degree matrices of the nodes and hyperedges, respectively, and W is the hyperedge weight matrix. These are learnable parameters between layers. It is a feature transformation, representing the feature matrix of all nodes in the l-th layer. Multiply by the inter-layer learnable parameter .

[0071] The training loss is calculated using the weighted cross-entropy loss function, and the formula is as follows: ; In the formula, Loss represents the training loss value in this round; K represents the number of local training samples participating in this iteration; This represents the sample weight value corresponding to the k-th sample; This indicates the corresponding business outcome tag value; This indicates the corresponding local risk response value.

[0072] S330: The training agent reads the local event motif representation results and local risk response results output by S320. First, it verifies whether the node representation vector dimension corresponding to each event motif node is consistent, whether the local risk response value corresponding to each local graph structure group is within the legal range of 0 to 1, and whether each prediction confidence meets the requirements of the current model version. Result items with inconsistent vector dimensions, empty local risk response values, missing prediction confidence, or model version identifiers inconsistent with the current training round are deleted.

[0073] The training agent first normalizes the retained results, then performs pruning and quantization compression to form shared results that can be compared across institutions without revealing the original extended warranty business records. The pruning process uses the following formula: ; In the formula, This represents the original vector to be shared. The original vector can be any one of the following: node representation vector, hyperedge statistical vector, or training-side risk statistical vector. Represents the clipped shared vector; Indicates the preset clipping threshold; This represents the 2-norm of the original vector.

[0074] The training agent aggregates the pruned node representation vectors according to the event motif node identifier to form an event motif embedding vector. It compresses and encodes the relation type set, relation direction set, and weight set of each hyperedge within the local graph structure group to form a hyperedge structure signature. It aggregates the local risk response value, risk level, and prediction confidence according to the controlled association key to form training-side local risk response parameters. These parameters characterize the risk intensity and judgment credibility of each local graph structure group during the training phase. Simultaneously, the training agent extracts the controlled association key summary, node start time field, node end time field, node start mileage field, node end mileage field, and occurrence stage label summary from the local event motif representation results to form a node boundary information summary.

[0075] The event motif embedding vector is used to represent the local structural position and business semantics of each event motif node; the hyperedge structural signature is used to represent the relational organization method in the local event motif hypergraph; the node boundary information digest is used to provide time, mileage and stage boundary basis for subsequent cross-institutional motif alignment and cross-institutional abnormal behavior chain splicing.

[0076] Finally, the training agent encapsulates the training round identifier, participating institution identifier summary, controlled association key summary, event motif embedding vector, hyperedge structure signature, training-side local risk response parameters, node boundary information summary, sample count, sample weight statistics, and model version identifier into an upload summary package. The upload summary package is defined as a unique data object for the participating institution in this round to perform restricted sharing with the coordinator.

[0077] The S330 outputs an uploaded summary packet, which is then read by the S410 and used for cross-agency event motif alignment.

[0078] S4 specifically includes the following sub-steps: S410. The alignment agent in the multi-agent system reads the upload summary packets uploaded by each participating institution in this round, and writes the training round identifier, model version identifier, participating institution identifier summary, controlled association key summary, event motif embedding vector, hyperedge structure signature, training-side local risk response parameters, node boundary information summary, sample quantity value, and sample weight statistics into the cross-institutional alignment task record. The cross-institutional alignment task record is defined as the only input object for performing cross-institutional event motif alignment in this round. Its function is to organize the originally scattered upload summary packets into an alignment input set with unified fields, verifiability, deletability, and computability.

[0079] The alignment agent first verifies whether the training round identifier of each cross-agency alignment task record is consistent with the current coordinating round, whether the model version identifier is consistent with the current global collaborative model version, whether the event motif embedding vector dimension is consistent with the coordinating end's preset dimension, and whether the hyperedge structure signature, training-side local risk response parameters, and node boundary information summary are empty. Cross-agency alignment task records with inconsistent training round identifiers, inconsistent model version identifiers, inconsistent event motif embedding vector dimensions, empty hyperedge structure signatures, empty training-side local risk response parameters, or empty node boundary information summaries are deleted to avoid data from different rounds, different versions, or different dimensions from being mixed into the same round of alignment process.

[0080] After verification, the alignment agent, under the same controlled association key summary constraint, generates candidate alignment node pairs from the event motif embedding vectors of different participating institutions, and synchronously associates the hyperedge structure signature, training-side local risk response parameters, and node boundary information summaries corresponding to the candidate alignment node pairs with the same candidate alignment record; then, the alignment similarity of each candidate alignment record is calculated using the following formula: ; In the formula, This represents the alignment similarity between the event topic node of participating institution i and the event topic node of participating institution j; This represents the embedding similarity between the event embedding vectors corresponding to two event topic nodes; This indicates the structural similarity between the hyperedge structure signatures corresponding to two event mother node nodes; This indicates the consistent value of the controlled association key summary corresponding to the two event parent nodes; , and 1 represents the embedding similarity weight, structural similarity weight, and controlled association key consistency weight, respectively.

[0081] Among them, embedding similarity Cosine similarity was used for calculation; structural similarity. Jaccard similarity coefficients are calculated using a hyperedge-structured signature set; consistency values ​​are determined for controlled association key summaries. Set to 1 if the two are exactly a match, otherwise set to 0.

[0082] In this embodiment, it can be , and Set the values ​​to 0.45, 0.35, and 0.20 respectively; when the alignment similarity is less than the preset alignment threshold of 0.80, delete the corresponding candidate alignment record; when the alignment similarity is greater than or equal to the preset alignment threshold of 0.80, retain it as a valid alignment relationship.

[0083] When multiple valid alignment relationships correspond to the same event node, the alignment agent prioritizes retaining the valid alignment relationship with the highest alignment similarity; if the alignment similarity is the same, it prioritizes retaining the valid alignment relationship with the smaller difference in the local risk response parameters on the training side.

[0084] After the screening is completed, the alignment agent outputs the cross-institutional topic alignment results. The cross-institutional topic alignment results include at least the set of aligned node pairs, the set of alignment similarities, the combination of source participating institutions, the corresponding controlled association key summary, and the node start time field, node end time field, node start mileage field, node end mileage field and occurrence stage label summary corresponding to each aligned node pair. These are used by S420 to generate a candidate set of abnormal behavior chains and by S430 to calculate the dynamic aggregation weights at the participating institutions level.

[0085] S420: The splicing agent in the multi-agent system reads the cross-agency topic alignment results and the training-side local risk response parameters of each participating agency output by S410. First, it performs sequence rearrangement on the aligned event topic nodes according to the controlled association key summary. Then, based on the node start time field, node end time field, node start mileage field, node end mileage field, and occurrence stage label summary carried in the cross-agency topic alignment results, it splices the detection lead fragment, maintenance transition fragment, replacement response fragment, revisit repeat fragment, and compensation trigger fragment into a candidate event chain according to the time sequence and mileage advancement direction. The candidate event chain is defined as a cross-agency node sequence that has not yet passed the chain integrity verification and risk support verification, and is not directly used as the subsequent global aggregation object.

[0086] The splicing agent writes the abnormal behavior chain candidate identifier, the aligned event mother node sequence, the hyperedge relationship sequence, the training side local risk response parameter sequence, the chain start time field, the chain end time field, the chain start mileage field, and the chain end mileage field for each candidate event chain, forming an abnormal behavior chain candidate record.

[0087] Subsequently, the splicing agent verifies and deletes candidate records of abnormal behavior chains: it deletes candidate records of abnormal behavior chains where the number of aligned event mother node nodes is less than the preset minimum number of nodes (3); it deletes candidate records of abnormal behavior chains where the chain start time field is later than the chain end time field, the chain start mileage field is greater than the chain end mileage field, or the node sequence contains a compensation trigger segment located before the detection lead segment; it breaks the chain for connection segments where there is no hyperedge relationship between two adjacent aligned event mother node nodes and the corresponding two training-side local risk response parameter values ​​are both less than the preset training-side local risk threshold of 0.50, and generates new candidate records of abnormal behavior chains at the broken chain position.

[0088] After deletion, the splicing agent calculates the chain risk support value for each candidate record of an abnormal behavior chain. The calculation formula is as follows: ; In the formula, This represents the chain risk support value of the candidate record for the c-th abnormal behavior chain; This represents the number of local risk response parameters on the training side in the candidate record of the c-th abnormal behavior chain; This represents the training-side risk response value corresponding to the u-th training-side local risk response parameter. The chain risk support value characterizes whether a cross-institutional event chain exhibits a continuous abnormal evolution trend, rather than being triggered accidentally by a single high-risk node. When the chain risk support value is less than the preset chain threshold of 0.65, the corresponding abnormal behavior chain candidate record is deleted; when the chain risk support value is greater than or equal to the preset chain threshold of 0.65, it is retained and included in global aggregation. If the same target extended insurance object corresponds to multiple abnormal behavior chain candidate records terminating at the same payout trigger segment, the one with the higher chain risk support value is prioritized for retention; if the chain risk support values ​​are the same, the one with a shorter chain span and a more continuous hyperedge relationship sequence is prioritized for retention.

[0089] For example, if the same target extended warranty object forms a "testing pilot segment - repair transition segment" and a "parts replacement response segment - claim trigger segment" in two participating institutions respectively, and the two segments have the same controlled association key summary and a chain risk support value of 0.72, they can be retained as a valid abnormal behavior chain candidate record.

[0090] S420 outputs a candidate set of abnormal behavior chains; the candidate set of abnormal behavior chains includes at least valid candidate records of abnormal behavior chains and their corresponding chain risk support values, which are then used by S430 to perform global aggregation.

[0091] S430: The aggregation agent in the multi-agent system reads the cross-institutional motif alignment results output by S410, the abnormal behavior chain candidate set output by S420, and the event motif embedding vector, hyperedge structure signature, training-side local risk response parameters, sample quantity value, and sample weight statistics corresponding to each participating institution. First, it establishes participating institution aggregation records according to the participating institution identifier summary, and verifies whether each participating institution aggregation record has the number of valid uploaded summary packages, the number of valid abnormal behavior chain candidate records, the sample quantity value, and the sample weight statistics value. Participating institution aggregation records with missing fields, zero valid abnormal behavior chain candidate records, or zero sample quantity value are deleted to avoid participating in global aggregation without supporting institutions.

[0092] After verification, the aggregation agent calculates the dynamic aggregation weight of each participating institution based on the sample quantity, sample weight statistics, and candidate quality of the abnormal behavior chain. The calculation formula is as follows: ; In the formula, This represents the dynamic aggregation weight of the s-th participating institution; This represents the normalized value of the sample size corresponding to the s-th participating institution; This represents the statistically normalized value of the sample weight corresponding to the s-th participating institution; This represents the normalized quality value of the candidate abnormal behavior chain corresponding to the s-th participating institution; , and These represent the normalized weights of sample quantity, statistical normalized weights of sample weights, and normalized weights of candidate quality for the abnormal behavior chain, respectively. In this embodiment, it can be , and The values ​​were set to 0.30, 0.30, and 0.40 respectively to give a higher weight to the candidate quality of the abnormal behavior chain in the global aggregation.

[0093] The aggregated agent performs weighted fusion on the event motif embedding vectors of each participating institution according to the dynamic aggregation weights to generate a global event motif embedding space; it performs weighted fusion on the hyperedge structure signatures of each participating institution to generate a global event motif hypergraph; and it performs weighted fusion on the training-side local risk response parameters of each participating institution to generate risk mapping parameters in the global collaborative model.

[0094] Subsequently, the aggregated agent extracts the topic mapping rules, anomaly chain concatenation threshold, and risk score correction rules from the global event topic hypergraph and the candidate set of abnormal behavior chains. These rules, along with the training round identifier and model version identifier, are then encapsulated into a globally distributed rule package. The topic mapping rules guide participating institutions to map their local event topic nodes to the global unified space in subsequent steps. The anomaly chain concatenation threshold guides the cross-institutional merging of local abnormal behavior chain segments. The risk score correction rules unify the risk score scale among different participating institutions.

[0095] S430 outputs a global event hypergraph, a global collaboration model, and a global rule package, which are then read by S510 to update the local federated model. S530 also reads the global rule package to execute the final generation of cross-organizational abnormal behavior chain analysis results and extended warranty risk assessment results.

[0096] S5 specifically includes the following sub-steps: S510: The application agent in the multi-agent system reads the global event theme hypergraph, global collaboration model, and globally distributed rule package output by S430, and simultaneously reads the standardized business event set, event theme node set, local event theme hypergraph, and the previous round rule correction result summary written to the rule correction interface by S530 in the previous round, all stored locally in the participating institution. If this is the first round of cross-institutional collaborative analysis, the previous round rule correction result summary is empty or uses a preset initial correction value. The application agent first writes the fields that the target extended warranty object needs to use in this round of analysis into the current extended warranty analysis task record. The current extended warranty analysis task record is defined as the only input object for the target extended warranty object to perform this round of abnormal behavior chain analysis on the current participating institution side. It includes at least the target extended warranty object identifier summary, controlled association key, global event theme hypergraph version identifier, global collaboration model version identifier, globally distributed rule package version identifier, standardized business event set summary, event theme node set summary, local event theme hypergraph summary, and previous round rule correction result summary.

[0097] The application agent performs field verification on the current extended warranty analysis task record, retaining only the current extended warranty analysis task record where the global event topic hypergraph version identifier, global collaboration model version identifier, and global rule package version identifier are consistent, and the target extended warranty object identifier summary, controlled association key, and event topic node set summary are all non-empty; delete the current extended warranty analysis task record where the version identifier is inconsistent, the target extended warranty object identifier summary is empty, the controlled association key is empty, the event topic node set summary is empty, or the target extended warranty object cannot be matched in the topic mapping rules of the global rule package, to prevent invalid tasks from entering the local application stage.

[0098] After verification, the application agent updates the local federated model based on the risk mapping parameters, topic mapping rules, risk scoring correction rules, and the summary of the previous round of rule correction results in the global collaborative model. If the model parameter write-back fails, the model version switch fails, or the correction rule loading fails, the current model update record is deleted and the previous version of the local federated model remains unchanged.

[0099] Subsequently, under controlled association key constraints, the application agent extracts the set of nodes and hyperedges corresponding to the target extended protection object from the event motif node set and the local event motif hypergraph. Combining this with the motif mapping rules in the globally distributed rule package, it identifies boundary nodes with cross-institutional mapping relationships to the target extended protection object, thus constructing a target local event motif subgraph. The target local event motif subgraph is defined as a local graph structure unit used by the target extended protection object on the current participating institution side to generate local abnormal behavior chain fragments. It consists of the event motif nodes corresponding to the target extended protection object, the effective local event motif hyperedge structure, and boundary nodes with mapping relationships to the globally distributed rule package.

[0100] The application agent re-executes the deletion rules on the target local event mother topic subgraph: delete isolated nodes that are not connected by any valid local event mother topic hyperedge structure and whose mother topic type label does not belong to the compensation trigger fragment or replacement response fragment; delete hyperedge structures with reversed time order or obvious conflicting mileage intervals.

[0101] S510 outputs the current extended warranty analysis task record and target local event mother topic subgraph, which are then used by S520 to generate local abnormal behavior chain fragments and application-side local risk response results.

[0102] S520: The application agent reads the current extended warranty analysis task record and target local event mother topic subgraph output by S510. First, it verifies whether each event mother topic node in the target local event mother topic subgraph has a mother topic type label, occurrence stage label, mileage interval label and organization type label, and whether each valid local event mother topic hyperedge structure has a relation type set, relation direction set and weight set. Nodes or hyperedge structures that are missing any necessary fields are deleted.

[0103] The application agent then expands the execution path of nodes in the target local event mother theme subgraph according to the time sequence, mileage advancement direction, and mother theme mapping rules in the globally distributed rule package. Only nodes that meet the following order are allowed to form candidate local paths: "detection leading segment first, maintenance transition segment or replacement response segment in the middle, and compensation trigger segment last". If there is a time reversal or mileage reversal in the candidate local path, or if there is no effective local event mother theme hyperedge structure to support two adjacent nodes, then a chain break is performed at that position, and new candidate local paths are formed respectively.

[0104] The application agent writes a local segment identifier, segment node sequence, segment hyperedge sequence, segment start time field, segment end time field, segment start mileage field, and segment end mileage field for each candidate local path, forming a local abnormal behavior chain segment record. A local abnormal behavior chain segment record is defined as a fragmented path record that can characterize the local abnormal evolution process of the target extended warranty object within the scope of a single participating institution. The application agent executes deletion rules on the local abnormal behavior chain segment records: local abnormal behavior chain segment records with fewer than a preset minimum of 2 nodes are deleted; local abnormal behavior chain segment records with a segment start time field later than a segment end time field or a segment start mileage field greater than a segment end mileage field are deleted; local abnormal behavior chain segment records containing only repeat visit segments and not maintenance transition segments, parts replacement response segments, or claims trigger segments are deleted.

[0105] After the deletion is completed, the application agent calls the updated local federated model to calculate the application-side local risk response results for the node sequence and fragment hyperedge sequence in each retained local abnormal behavior chain fragment record.

[0106] To ensure a unified calculation method for the retention rules of local anomalous behavior chain segments, the application agent records and calculates the local segment risk value for each local anomalous behavior chain segment. The calculation formula is as follows: ; In the formula, This represents the local segment risk value of the f-th local abnormal behavior chain segment; This represents the number of application-side local risk response results in the f-th segment of the local abnormal behavior chain; This represents the application-side local risk response value corresponding to the v-th application-side local risk response result. The local segment risk value is used to characterize the abnormal intensity of the local path on the current participating institution side. When the local segment risk value is less than the preset local segment threshold of 0.60, the corresponding local abnormal behavior chain segment record is deleted; when the local segment risk value is greater than or equal to the preset local segment threshold of 0.60, the local abnormal behavior chain segment record is retained.

[0107] The application agent further compares the local segment risk value with the preset classification threshold to generate the application-side local risk level, and generates the segment confidence based on the number of effective nodes, the number of effective hyperedges, and the consistency of model prediction. For example, a local abnormal behavior chain segment contains three nodes in sequence: a detection pilot segment, a maintenance transition segment, and a replacement response segment, and the corresponding application-side local risk response values ​​are 0.58, 0.69, and 0.81, respectively. Then, the local segment risk value of this segment is 0.693, and it can be retained as a valid local abnormal behavior chain segment.

[0108] S520 outputs a set of local abnormal behavior chain fragments and a set of application-side local risk response results, which are then used by S530 to perform cross-agency merging, conflict resolution, and final risk assessment.

[0109] S530: The splicing agent and the aggregation agent in the multi-agent system jointly read the set of local abnormal behavior chain fragments output by each participating agency, the set of local risk response results on the application side, the global event mother topic hypergraph and the global distributed rule package, and establish cross-agency merged task records.

[0110] The cross-institutional merge task record is defined as the unique input object for performing cross-institutional abnormal behavior chain merging and extended insurance risk assessment for the same target extended insurance object. It includes at least the target extended insurance object identifier summary, controlled association key, local abnormal behavior chain fragment set, application-side local risk response result set, global event mother hypergraph version identifier, and global rule package version identifier.

[0111] The concatenating agent first verifies the cross-agency merging task records, retaining only those cross-agency merging task records where the controlled association key is consistent, the version identifier of the global event topic hypergraph is consistent with the version identifier of the globally issued rule package, the set of local abnormal behavior chain segments is not empty, and the application-side local risk response values ​​in the application-side local risk response result set are all within the range of 0 to 1; cross-agency merging task records where the controlled association key is inconsistent, the version identifier is inconsistent, the set of local abnormal behavior chain segments is empty, or the application-side local risk response values ​​contain null values ​​or are out of range are deleted.

[0112] After verification, the splicing agent performs cross-agency matching on local anomalous behavior chain fragments from different participating agencies based on the unified motif mapping result in the global event motif hypergraph, and calculates the cross-agency merging compatibility value between any two local anomalous behavior chain fragments. The calculation formula is as follows: ; In the formula, This represents the cross-mechanism merging compatibility value between local abnormal behavior chain segment a and local abnormal behavior chain segment b. Indicates the score for continuity of time; The score indicates the consistency of the parent question order; This indicates the consistency score of local risk response on the application side; , and These represent the weights for time continuity, consistency of the parent topic order, and consistency of application-side local risk response, respectively. .

[0113] In this embodiment, it can be , and The values ​​are set to 0.30, 0.40, and 0.30 respectively. When the cross-organizational merging compatibility value is less than the preset merging threshold of 0.75, merging is not performed. When the cross-organizational merging compatibility value is greater than or equal to the preset merging threshold of 0.75, the corresponding local abnormal behavior chain segments are allowed to be merged into the same cross-organizational abnormal behavior chain. If two local abnormal behavior chain segments overlap in time but conflict in the order of the parent themes, the segment with higher confidence is retained first. If the segment confidence values ​​are the same, the segment with higher risk value is retained first. If they are still the same, the segment with more nodes and more continuous super-edge sequence is retained first.

[0114] The aggregation agent calculates the final risk value for the merged cross-agency abnormal behavior chain using the following formula: ; In the formula, Indicates the final risk value of the target extended warranty recipient; F represents the number of local abnormal behavior chain segments in the merged cross-organizational abnormal behavior chain; This represents the local segment risk value of the f-th segment of the local anomalous behavior chain. This represents the segment confidence level of the f-th local abnormal behavior chain segment. The final risk value is used to uniformly characterize the intensity of abnormal risk of extended insurance claims for the target extended insurance recipient within the scope of multiple participating institutions; when the final risk value is greater than or equal to 0.80, a high risk level is output; when the final risk value is less than 0.80 but greater than or equal to 0.50, a medium risk level is output; when the final risk value is less than 0.50, a low risk level is output.

[0115] Finally, the splicing agent and the aggregation agent output the cross-organizational abnormal behavior chain analysis results and the extended warranty risk assessment results. The cross-organizational abnormal behavior chain analysis results include at least the complete abnormal path, the sequence of key abnormal institutions, and the set of triggering motifs. The extended warranty risk assessment results include at least the final risk value, the final risk level, and the corresponding target extended warranty object identifier summary. The cross-organizational abnormal behavior chain analysis results and the extended warranty risk assessment results are written into the extended warranty historical analysis record pool. At the same time, the abnormal chain splicing failure record, conflict resolution record, and final risk level are written into the rule correction interface so that the next round of S510 can continue to call them as the summary of the previous round of rule correction results.

[0116] All the above formulas are performed using dimensionless numerical calculations; the relevant formulas are based on empirical models that approximate the real situation, obtained through extensive data collection and software simulation fitting. The preset parameters and thresholds involved in the formulas can be conventionally set and adjusted by those skilled in the art according to the physical constraints of the actual application scenario.

[0117] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0118] In conclusion, the above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A cross-domain data collaborative analysis method for automobile extended warranty based on federated learning, characterized in that, Includes the following steps: S1. Each participating institution locally reads the car extended warranty business records and generates a standardized business event set, controlled association keys, and event theme node set; S2. Construct a candidate set of relations, a local event topic hyperedge structure, a local event topic hypergraph, and a local graph structure based on the event topic node set; S3. Construct a local training sample set based on the local graph structure grouping, train the local federated model, generate event motif embedding vectors, hyperedge structure signatures, training-side local risk response parameters, and upload summary packages; S4. The coordinating end performs cross-organizational topic alignment, abnormal behavior chain candidate generation, and global aggregation based on the uploaded summary package to obtain a global event topic hypergraph, a global collaboration model, and a global rule package. S5. Each participating institution updates its local federated model based on the global event hypergraph, the global collaboration model, and the globally distributed rule package, generating local abnormal behavior chain fragments and application-side local risk response results.

2. The method for cross-domain data collaborative analysis of automobile extended warranty based on federated learning according to claim 1, characterized in that, S5 also includes: performing cross-organizational merging of local abnormal behavior chain segments corresponding to the same target extended warranty object, and outputting cross-organizational abnormal behavior chain analysis results and extended warranty risk assessment results.

3. The method for cross-domain data collaborative analysis of automobile extended warranty based on federated learning according to claim 1, characterized in that, S1 specifically includes: Each participating institution reads the vehicle extended warranty business records locally and performs verification, supplementation, sorting, and duplicate deletion on the local identifier of the business object, event type, event occurrence time, event mileage value, institution type, action category, result status, and source record type, generating a standardized business event set and controlled association key; Under the same controlled association key constraint, the standardized business event set is segmented according to the event interval, mileage increment and event type switching relationship to generate a candidate event theme fragment set; Perform category merging and attribute encoding on the candidate event motif fragment set to generate an event motif node set.

4. The method for cross-domain data collaborative analysis of automobile extended warranty based on federated learning according to claim 1, characterized in that, S2 specifically includes: Under the same controlled association key constraint, perform bucket verification on the event mother node set, and generate a candidate set of relations according to the order of occurrence stage and mileage interval; The candidate set of relations is deleted, relation weights are calculated, and relation types are defined to generate a local event topic hyperedge structure. A local event hypergraph is constructed based on the event motif node set and the local event motif hyperedge structure, and local graph structure groups are generated according to controlled association keys and connectivity relationships.

5. The method for cross-domain data collaborative analysis of automobile extended warranty based on federated learning according to claim 4, characterized in that, This map structure grouping includes at least a set of node identifiers, a set of super-edge identifiers, a group start time field, a group end time field, a group start mileage field, a group end mileage field, and a group final status field.

6. The method for cross-domain data collaborative analysis of automobile extended warranty based on federated learning according to claim 1, characterized in that, S3 specifically includes: Based on the local event theme hypergraph, local graph structure grouping and corresponding case closure records, a local training sample set containing business outcome labels and sample weight values ​​is generated. Based on the local training sample set, node feature matrices and hyperedge correlation matrices are extracted to train a local federated model, generating local event motif representations and local risk response results. Normalization, pruning, quantization compression, and encapsulation are performed on the local event motif representation results and local risk response results.

7. The method for cross-domain data collaborative analysis of automobile extended warranty based on federated learning according to claim 6, characterized in that, Also includes: Generate an upload digest package that includes a controlled association key summary, an event motif embedding vector, a hyperedge structure signature, training-side local risk response parameters, and a node boundary information summary.

8. The method for cross-domain data collaborative analysis of automobile extended warranty based on federated learning according to claim 1, characterized in that, S4 specifically includes: The coordinator reads the uploaded summary packages from each participating institution, establishes a cross-institutional alignment task record, and generates cross-institutional topic alignment results based on controlled association key summaries, event motif embedding vectors, hyperedge structure signatures, training-side local risk response parameters, and node boundary information summaries. Generate a candidate set of anomalous behavior chains based on cross-institutional theme alignment results; Global aggregation is performed based on cross-institutional topic alignment results, candidate sets of abnormal behavior chains, event topic embedding vectors, hyperedge structure signatures, local risk response parameters on the training side, sample quantity values, and sample weight statistics to generate a global event topic hypergraph, a global collaborative model, and a global rule package.

9. A cross-domain data collaborative analysis method for automobile extended warranty based on federated learning as described in claim 1, characterized in that, S5 specifically includes: Each participating institution reads the global event hypergraph, the global collaboration model, and the globally distributed rule package, establishes a record of the current extended warranty analysis task, updates the local federated model, and extracts the target local event subgraph. Generate a set of local abnormal behavior chain fragments and a set of application-side local risk response results based on the target local event mother topic subgraph; Based on a set of local abnormal behavior chain fragments, a set of application-side local risk response results, a global event mother topic hypergraph, and a global rule package.

10. A cross-domain data collaborative analysis method for automobile extended warranty based on federated learning according to claim 1, characterized in that, It also includes: performing cross-institutional merging, conflict resolution, and final risk assessment on local abnormal behavior chain segments corresponding to the same target extended warranty object, and outputting cross-institutional abnormal behavior chain analysis results and extended warranty risk assessment results.

Images