Internet property right transaction risk assessment and decision-making system based on artificial intelligence

Abstract

The invention provides an Internet property right transaction risk assessment and decision-making system based on artificial intelligence, and relates to the technical field of data processing, and the system comprises a data collection and preprocessing module which is used for collecting original transaction data in real time, and carrying out the cleaning and standardization processing; the multi-dimensional feature construction module is used for modeling according to a time window and constructing a multi-stage interaction relation graph to obtain time sequence feature data and graph association feature data; normalized weighted splicing is carried out on the features; the risk identification module is used for extracting a suspicious transfer behavior chain, performing benefit flow tracing and executing composite risk judgment; the intelligent decision-making module is used for carrying out hierarchical adaptive discrimination and case migration comparison, summarizing to generate a risk assessment result and pushing an early warning; the processing feedback of the management terminal is received and is reversely used for optimizing identification and decision parameters; according to the invention, the autonomy and accuracy of property right transaction risk assessment and decision making are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to an internet property transaction risk assessment and decision-making system based on artificial intelligence. BACKGROUND

[0002] In the prior art, internet property transaction risk assessment is mostly based on a combination of rule-based risk models and manual review. For example, a knowledge property transaction platform usually relies on historical data to develop risk factors and manually sets thresholds to screen transaction information. When the platform detects abnormal transaction behavior, such as large amount of transfer or frequent changes in the information of the holder within a short period of time, it will trigger a manual review process to determine the risk level of the transaction and achieve preliminary risk prevention and control.

[0003] However, the above method has the problem of insufficient accuracy when facing large-scale and complex transaction scenarios. For example, a platform once used frequent transfer of patent rights between associated companies to evade supervision. Since the rule model cannot effectively identify hidden interests or new fraud patterns, the system fails to provide timely warnings, resulting in the completion of abnormal transactions and seriously affecting the safety and compliance of property circulation. Therefore, improving the intelligent and adaptive capabilities of risk assessment has become a technical problem that needs to be solved. SUMMARY

[0004] The purpose of the present application is to provide an internet property transaction risk assessment and decision-making system based on artificial intelligence, which aims to solve the problems mentioned in the background art.

[0005] To solve the above technical problems, the technical solution of the present application is as follows: An internet property transaction risk assessment and decision-making system based on artificial intelligence, the system comprising: a data acquisition and preprocessing module for acquiring real-time original transaction data from an internet property transaction platform and performing cleaning and standardization processing on the original transaction data to obtain standardized transaction data, the original transaction data including property subject information, transaction party historical records and associated party materials; a multi-dimensional feature construction module for modeling and constructing a multi-level interaction relationship graph according to the standardized transaction data, obtaining time series feature data and graph association feature data, and performing normalized weighted splicing on the features to form fusion feature data; a risk identification module for extracting suspicious transfer behavior chains and performing interest flow tracing according to the fusion feature data and the multi-level interaction relationship graph, executing compound risk discrimination, and generating preliminary risk data; An intelligent decision module is configured to perform hierarchical adaptive discrimination and case migration comparison based on the preliminary risk data, generate a risk assessment result, and push an early warning; and receive a management terminal disposition feedback, which is reversely used to optimize identification and decision parameters.

[0006] Preferably, the multi-dimensional feature construction module comprises: A time window modeling unit is configured to group transaction sequences according to a set time window based on standardized transaction data, and perform sliding statistics and change trend analysis to obtain time series feature data; A relationship graph generation unit is configured to construct a multi-level interaction relationship graph with a property subject as a core based on the standardized transaction data, wherein nodes of the graph represent different participants, edges represent transfer types and time labels, and correlation degrees between the nodes are calculated to generate graph correlation feature data; A feature fusion and weighting unit is configured to splice the time series feature data and the graph correlation feature data, and combine the feature importance weights of abnormal cases in the historical case library to obtain fusion feature data.

[0007] Preferably, the risk identification module comprises: A suspicious behavior chain extraction unit is configured to extract a suspicious transfer behavior chain occurring continuously by using recursive traversal and sequence encoding based on the fusion feature data, and generate behavior chain feature data; An interest flow tracing unit is configured to trace key nodes level by level based on the behavior chain feature data in combination with the multi-level interaction relationship graph, mark a suspected interest flow direction, and generate an abnormal identifier when a cross-domain or cross-platform transfer occurs, to obtain tracing risk data; A composite risk discrimination unit is configured to logically combine and discriminate the behavior chain feature data and the tracing risk data, and execute a risk superposition grading mechanism when multiple high-risk factors are met simultaneously to generate preliminary risk data.

[0008] Preferably, the intelligent decision module comprises: A hierarchical adaptive discrimination unit is configured to calculate a risk score based on the preliminary risk data, automatically match a processing strategy by using a hierarchical threshold system, and update threshold parameters adaptively when a new abnormal pattern is detected to generate a hierarchical risk result; A case migration learning unit is configured to compare the hierarchical risk result with the historical case library one by one, directly output a corresponding disposition result if there is a high-confidence similar case, or store the current risk feature in the case library and generate a to-be-labeled item for the case library, and automatically update the case library to maintain accuracy when receiving external input labeled data to generate a case comparison result; An evaluation aggregation and output unit is configured to aggregate and evaluate the hierarchical risk results and the case comparison results, generate risk evaluation results, and push corresponding early warning information to a management terminal; A feedback closed loop unit is configured to receive disposal feedback from the management terminal, and input the disposal feedback to the risk identification module and the intelligent decision module respectively, so as to continuously optimize the identification and decision parameters, and form a closed loop self-learning mechanism.

[0009] Preferably, the relationship graph generation unit comprises: An entity standardization processing subunit is configured to perform cross-platform entity identification and deduplication on the participants according to the standardized transaction data, and generate a unified node set; A time sequence and domain labeling subunit is configured to write timestamps and regional or platform source identifiers on the edges, and write role types and risk priori labels in the nodes according to the unified node set and the transaction records, and form an initial relationship graph with time sequence and domain attributes; An association degree calculation subunit is configured to calculate the interaction features between the nodes within a preset upper limit of hop numbers, calculate the interaction times, freshness, role similarity and path credibility, and generate association degree scores of node pairs according to the initial relationship graph, and form a weighted relationship graph; A graph incremental updating subunit is configured to perform incremental updating on the weighted relationship graph according to newly arrived transaction data, generate an effective graph snapshot conforming to a sliding time window, and invalidate edges exceeding the window period according to rules, so as to ensure the real-time and effectiveness of the graph.

[0010] Preferably, the feature fusion and weighting unit comprises: A feature weight distribution subunit is configured to distribute learnable weights to the time sequence feature data and the graph association feature data based on transaction types, transaction amounts, data completeness and near-end time features, and generate a weighted feature vector; A case priori guidance subunit is configured to perform similarity comparison with abnormal case feature clusters in a historical case library according to the weighted feature vector, and bias the weight distribution result when the similarity is higher than a threshold, and generate a feature vector guided by priori; A robust normalization processing subunit is configured to perform median scaling and extreme value truncation on abnormal values in the feature vector guided by priori, and obtain robust feature data after normalization; A fusion calibration subunit is configured to splice the time sequence features and the graph association features according to the robust feature data after normalization, and obtain fusion feature data.

[0011] Preferably, the suspicious behavior chain extraction unit comprises: The transaction sequence segmentation subunit is configured to perform time sequence segmentation on continuous transaction events according to the fusion feature data, and divide the continuous transaction events into candidate sub-sequences according to transfer amount change, transaction frequency anomaly and cross-domain occurrence conditions, to obtain candidate behavior chain data. The behavior chain feature encoding subunit is configured to perform progressive encoding on node roles, transfer direction, transaction freshness and node hop length in sequence according to the candidate behavior chain data, to generate a multi-dimensional behavior chain feature vector. The abnormal accumulation determination subunit is configured to calculate a risk accumulation value of continuous transactions step by step according to the multi-dimensional behavior chain feature vector, and mark a corresponding candidate behavior chain as a suspicious behavior chain when the risk accumulation value exceeds a dynamic threshold, and generate behavior chain feature data.

[0012] Preferably, the benefit flow tracing unit comprises: The key node screening subunit is configured to identify nodes participating in multiple transactions in the relationship graph according to the behavior chain feature data, and screen the nodes according to transaction amount concentration and node association degree weight, to generate a key node set. The trajectory backtracking construction subunit is configured to backtrack from each key node as a starting point along transaction links step by step according to the key node set, record node role type, interaction frequency and regional label, and form benefit flow trajectory data. The abnormal trajectory labeling subunit is configured to generate an abnormal identifier for an abnormal path when detecting that a transfer path crosses different regions, different platforms or exists multi-level anonymous transfer according to the benefit flow trajectory data, and take the labeled trajectory as trace risk data.

[0013] Preferably, the hierarchical adaptive discrimination unit comprises: The risk score generation subunit is configured to calculate and output a risk score sequence based on transaction amount abnormality, historical risk label of participants and time span factors according to the preliminary risk data. The hierarchical threshold comparison subunit is configured to set threshold intervals according to at least three risk levels according to the risk score sequence, compare the risk scores layer by layer, and generate a hierarchical discrimination result. The threshold adaptive adjustment subunit is configured to adjust the upper and lower limits of the threshold when detecting that the risk score continuously deviates from the preset threshold interval according to the hierarchical discrimination result, write the updated threshold parameters into a decision parameter library, and generate an adaptively updated hierarchical risk result.

[0014] Preferably, the case migration learning unit comprises: The feature similarity comparison subunit is configured to extract key feature vectors according to the hierarchical risk result, and compare the key feature vectors with feature sets in a historical case library one by one, to form a candidate similar case set. The confidence comprehensive judgment subunit is used for calculating a comprehensive score according to the candidate similar case set, combining the similarity score and the case treatment accuracy, outputting the corresponding treatment result when the comprehensive score exceeds a preset threshold, and forming a case judgment result. The case base expansion and update subunit is used for writing the current risk feature data into the case base to generate a to-be-labeled item when a high-confidence case is not matched according to the case judgment result, expanding and correcting the case base to generate a case comparison result when receiving external input labeled data.

[0015] The above scheme of the present application at least has the following beneficial effects: Firstly, through the data acquisition and preprocessing module, original transaction data can be collected from the Internet property transaction platform in real time, and the data can be cleaned and standardized to obtain unified standardized transaction data. Compared with the prior art which relies on manual setting of threshold and single rule screening, this module can automatically process original data of different platforms and different formats, reduce data redundancy and noise interference, and ensure that the input information for subsequent risk identification is more accurate and complete. For example, when the same transaction subject appears in different platforms with different names or different registration codes, the system can automatically perform normalization processing to avoid risk omission caused by inconsistent data.

[0016] Secondly, through the multi-dimensional feature construction module, time series features and multi-level interaction relationship graphs can be constructed based on standardized transaction data, and they can be normalized and weighted and spliced to form fusion feature data. This method breaks through the limitation of single rule factor in the prior art, and can not only reflect the dynamic changes of transaction behavior in the time dimension, but also reveal the complex correlation between different transaction subjects. For example, when a patent is transferred across domains multiple times in a short period of time, the system can capture both frequency anomalies and cross-platform interactions, thereby enhancing the ability to depict hidden interest chains.

[0017] Thirdly, through the risk identification module, suspicious transfer behavior chains can be automatically extracted and interest flow traced based on fusion feature data and interaction relationship graphs, and preliminary risk data can be generated in combination with composite risk discrimination. Unlike the prior art based on fixed rules to trigger manual review, the present application can automatically identify potential interest flow in the transaction chain, and form a higher level of risk discrimination when there are high-risk factors superimposed. For example, when the transaction shows frequent cross-platform transfer and is accompanied by a final backflow to the same associated group, the system can generate preliminary risk data in a timely manner to achieve early warning.

[0018] Finally, through the intelligent decision module, hierarchical adaptive discrimination and case migration comparison can be carried out on the basis of the preliminary risk data, the risk assessment results are generated and the early warning information is pushed, and the disposal feedback of the management terminal is received to realize parameter optimization and closed-loop self-learning. Compared with the manual review of the prior art, not only the automation level of evaluation and decision is improved, but also the discrimination standard and case library content can be continuously optimized as new risk patterns appear. For example, when the system first identifies a new type of transfer avoidance behavior and confirms it through manual annotation, the model parameters and case library are automatically updated, so that similar risks can be quickly discovered in the future.

[0019] Therefore, the present application can effectively overcome the defects of the prior art in large-scale and complex transaction scenarios, realize real-time monitoring and accurate risk assessment of internet property transactions, improve the security and compliance of property circulation, and significantly enhance the intelligent and adaptive ability of risk assessment. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 is an architecture diagram of an internet property transaction risk assessment and decision system based on artificial intelligence provided by an embodiment of the present application. DETAILED DESCRIPTION

[0021] Exemplary embodiments of the present disclosure will be described in greater detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided so that the present disclosure can be more thoroughly understood and the scope of the present disclosure can be accurately conveyed to those skilled in the art.

[0022] As shown in Figure 1 , an embodiment of the present application proposes an internet property transaction risk assessment and decision system based on artificial intelligence, which comprises: a data acquisition and preprocessing module for acquiring raw transaction data from an internet property transaction platform in real time, and cleaning and standardizing the raw transaction data to obtain standardized transaction data, the raw transaction data including property subject information, transaction party historical records and associated party materials; a multi-dimensional feature construction module for modeling and constructing a multi-level interaction relationship graph according to the standardized transaction data, obtaining time series feature data and graph association feature data, and normalizing and weighting the features to form fusion feature data; a risk identification module for extracting suspicious transfer behavior chains and performing interest flow tracing according to the fusion feature data and the multi-level interaction relationship graph, performing compound risk discrimination, and generating preliminary risk data; An intelligent decision module is configured to perform hierarchical adaptive discrimination and case migration comparison based on the preliminary risk data, generate a risk assessment result, and push a warning; and receive a management terminal disposition feedback, which is reversely used to optimize identification and decision parameters.

[0023] In the embodiments of the present application, the internet property transaction risk assessment and decision system can collect original transaction data from the internet property transaction platform in real time through the data acquisition and preprocessing module, and clean and standardize the collected property subject information, transaction party historical records and associated party materials, to obtain uniform and standardized transaction data. Through this process, the data format differences between different platforms can be eliminated, data redundancy and noise interference can be reduced, and accurate and reliable basic data can be provided for subsequent risk identification. For example, when the same transaction subject uses different recording methods in different platforms, the system can automatically unify them, avoiding missed detection due to inconsistent data.

[0024] The multi-dimensional feature construction module can model according to the standardized transaction data and a set time window, and further construct a multi-level interaction relationship graph, so as to obtain time series feature data and graph associated feature data, and normalize and weight splice multiple features to form fusion feature data. This processing method can not only capture the continuity of transaction behavior in the time dimension, but also depict the complex association relationship between transaction subjects. For example, if a subject frequently transfers patent rights with multiple associated companies in a short period of time, the system can identify the frequency anomaly through the time series feature, and reveal the potential interest connection with other nodes through the graph feature.

[0025] The risk identification module can automatically extract suspicious transfer behavior chains and perform interest flow tracing according to the fusion feature data and the multi-level interaction relationship graph, and generate preliminary risk data in combination with the composite risk discrimination. This function can break through the limitations of traditional rule-based screening and has the ability to discover hidden abnormal patterns. For example, if a transaction chain shows multiple transfers across platforms, and the transferees all have hidden equity relationships with a specific group, the system can automatically identify the behavior chain and mark the potential interest concentration path, thereby realizing early risk warning.

[0026] The intelligent decision module can perform hierarchical adaptive discrimination according to the preliminary risk data, and perform migration comparison combined with the historical case library, to finally generate a risk assessment result and push a warning information. The module can also receive the disposal feedback of the management terminal, and input the feedback result to the risk identification module and the decision module in reverse, so that the parameters are continuously optimized and self-learned. Through this closed-loop mechanism, the system can continuously improve the identification ability of new risk patterns. For example, when the management personnel confirms that a never-appeared avoidance transfer mode is abnormal behavior, the system will automatically absorb the labeling information for adjusting the discrimination threshold and the case library content, so as to quickly identify similar risks in subsequent transactions.

[0027] Through the synergistic effect of the above modules, the embodiment of the present application can effectively improve the intelligentization and self-adaptive ability of the internet property transaction risk assessment, overcome the problem of insufficient identification of complex transaction scenarios based on the traditional rule method, realize real-time monitoring and accurate early warning of large-scale, multi-platform property transactions, and improve the security and compliance of the property circulation process.

[0028] The data acquisition and preprocessing module specifically includes: Firstly, the interface of the internet property transaction platform is called in real time or the crawling mechanism is used to collect original transaction data including property subject information, transaction party historical records and associated party data. During the collection process, the data is cached in chronological order to ensure the time continuity of the subsequent analysis.

[0029] Then, the original data collected is cleaned to eliminate null values, abnormal values and duplicate records. For example, when a transaction entry is missing a transaction amount or has duplicate records twice, the system will supplement or delete the data to ensure data integrity and uniqueness.

[0030] Next, the cleaned data is standardized. The standardization process includes converting transaction amounts of different platforms into the same currency and unit, standardizing the format of the participant names, and adopting consistent storage rules for the contact information, registration code and other fields. For example, if a transaction amount is expressed in US dollars, it is converted into the equivalent RMB and saved; if there is a difference between "Co. Ltd." and "Limited Liability Company" in the company name, they are unified into one expression.

[0031] Finally, the cleaned and standardized data is stored as a standardized transaction data set, so as to be used as the input of the subsequent multi-dimensional feature construction module.

[0032] In a preferred embodiment of the present application, the multi-dimensional feature construction module includes: a time window modeling unit configured to group the transaction sequence according to the standardized transaction data according to a set time window, and perform sliding statistics and change trend analysis to obtain time sequence feature data; a relationship graph generating unit configured to construct a multi-level interaction relationship graph with a property right subject as a core according to the standardized transaction data, wherein nodes of the graph represent different participants, edges represent transfer types and time labels, and the relationship graph generating unit is further configured to calculate correlation degrees between the nodes to generate graph correlation feature data; a feature fusion weighting unit configured to splice the time sequence feature data and the graph correlation feature data, and combine feature importance weights of abnormal cases in a historical case library to perform weighting to obtain fusion feature data.

[0033] In the embodiment of the present application, the multi-dimensional feature construction module can group the standardized transaction data according to a set time window by the time window modeling unit, and perform sliding statistics and change trend analysis in each group to generate time sequence feature data. In this way, the dynamic change characteristics of the transaction in the time dimension can be described, for example, when a certain patent is transferred more than three times in a short period of time, the system can highlight the abnormal frequency in the time sequence feature to prompt a possible risk.

[0034] The relationship graph generating unit can construct a multi-level interaction relationship graph with a property right subject as a core according to the standardized transaction data, wherein the nodes are used to represent transaction subjects, and the edges are used to represent transfer relationships and time labels. By calculating the correlation degrees between the nodes, the graph correlation feature data can be obtained. This process can intuitively present the transaction relationships and strengths between different subjects, for example, when several nodes are frequently connected to each other through different types of transfer edges in a short period of time, the system can capture potential interest groups, thereby improving the visualization degree of implicit correlations.

[0035] The feature fusion weighting unit can splice the time sequence feature data and the graph correlation feature data, and combine the feature importance weights of abnormal cases in the historical case library to perform weighting to obtain the fusion feature data. This fusion processing not only ensures the complementarity of the time features and the relationship features, but also highlights the features with higher risk correlation degrees through weight adjustment. For example, if the historical cases show that “cross-border transfer behavior” is a high-risk factor, the system will automatically increase the importance of this type of feature in the fusion process, thereby improving the accuracy of risk identification.

[0036] Through the above processing, the embodiment of the present application can describe the transaction behavior in multiple dimensions, retain the time dynamic characteristics, and strengthen the multi-party interaction characteristics, thereby realizing comprehensive quantitative description of complex transaction behaviors and providing high-quality input data for subsequent risk identification.

[0037] The time window modeling unit specifically includes: Firstly, the standardized transaction data is sorted in chronological order and grouped according to the set time window. The setting of the time window can be by day, week or month, and the system can be flexibly adjusted. For example, under a week time window, all transactions occurring within the time range will be divided into a group.

[0038] Then, statistical analysis is performed within each time window, including the total number of transactions, the sum and average of transaction amounts, the number of participants, etc. Through these statistics, the basic transaction characteristics within the window can be formed.

[0039] Next, the sliding statistics method is used to analyze the differences between adjacent time windows. For example, the number of transactions in the current window is compared with that in the previous window. If the growth rate exceeds a certain proportion of the historical average, it is marked as an abnormal transaction activity.

[0040] Further, the trend of feature changes in the time window is modeled. Trend modeling does not use formulas, but is described in words: the system will judge whether the transaction volume is showing an upward trend, a downward trend, or periodic fluctuations based on the changes in the transaction volume of consecutive windows, thereby generating a trend label.

[0041] Finally, the statistical features and trend features are combined to output time series feature data as input for subsequent feature fusion and risk identification.

[0042] In a preferred embodiment of the present application, the risk identification module comprises: A suspicious behavior chain extraction unit is configured to extract a suspicious transfer behavior chain occurring continuously based on the fusion feature data using a recursive traversal and sequence encoding method, and generate behavior chain feature data. An interest flow tracing unit is configured to trace key nodes level by level based on the behavior chain feature data in combination with a multi-level interaction relationship graph, mark a suspected interest flow direction, and generate an abnormality identification when a cross-domain or cross-platform transfer occurs, to obtain tracing risk data. A composite risk discrimination unit is configured to logically combine and discriminate the behavior chain feature data and the tracing risk data, and when multiple high-risk factors are met simultaneously, execute a risk superposition grading mechanism to generate preliminary risk data.

[0043] In the embodiment of the present application, the suspicious behavior chain extraction unit can extract a suspicious transfer behavior chain occurring continuously based on the fusion feature data using a recursive traversal and sequence encoding method, and generate behavior chain feature data. Through this mechanism, long-term or high-frequency transaction links can be automatically identified, avoiding the low-efficiency problem of manual screening one by one. For example, when a patent is transferred to multiple subordinate companies within a month, the system can extract this series of transfer events as a single behavior chain and mark it as a suspicious pattern.

[0044] The interest flow tracing unit can trace back the key nodes level by level according to the behavior chain feature data, in combination with the multi-level interaction relationship graph, mark the suspected interest flow direction, and generate an abnormal identification when transferring across domains or platforms, so as to obtain the tracing risk data. This function can reveal the flow direction of funds or rights hidden behind surface transactions. For example, when a patent is transferred across multiple jurisdictions multiple times and finally flows back to the original transferor associated company, the system can mark this abnormal backflow path through interest tracing, thereby assisting the supervision party to timely discover evasion behaviors.

[0045] The composite risk discrimination unit can logically combine and discriminate the behavior chain feature data and the tracing risk data, and when multiple high-risk factors are met, perform a risk superposition grading mechanism to generate preliminary risk data. Through this superposition grading method, the system can accurately identify scenarios where a single factor has a relatively small impact but the composite existence has a prominent risk. For example, a single "cross-platform transfer" may have a low risk, but if it is superimposed with "high-frequency transfer" and "backflow path", the system will automatically upgrade the risk level, thereby improving the sensitivity of identification.

[0046] Through the above functions, the embodiment of the present application can effectively identify complex chain transactions and hidden interest transfer paths, significantly improve the discovery ability of new evasion methods, and enhance the depth and comprehensiveness of property transaction risk monitoring.

[0047] The composite risk discrimination unit specifically includes: First, receive the behavior chain feature data, including the node role, transfer direction, frequency and amount of the suspicious transaction chain, and receive the tracing risk data, including cross-domain transfer marking, abnormal backflow path marking and key node concentration.

[0048] Then, logically combine and discriminate the behavior chain feature data and the tracing risk data. The process of logical combination is as follows: when a transaction chain is marked as a high-frequency abnormal transaction, and at the same time, a cross-domain transfer path is found in the tracing result, the system will consider that the risk level of the chain is higher.

[0049] Next, perform a risk superposition grading mechanism. The principle of risk grading is that the more high-risk factors that are met, the higher the risk level. For example, a single factor of "high-frequency transfer" may be determined as low risk; if it is superimposed with "cross-domain transfer" and "associated company backflow", it is upgraded to medium risk or high risk. The system will automatically output the corresponding risk level through the preset grading rules.

[0050] Finally, generate preliminary risk data from the discrimination result, including the risk level, the transaction chain involved and the risk factors triggered, so as to be transmitted to the intelligent decision module for further hierarchical adaptive discrimination and case migration comparison.

[0051] In a preferred embodiment of the present application, the intelligent decision module comprises: a hierarchical adaptive discrimination unit for calculating a risk score according to the preliminary risk data, automatically matching a processing strategy using a hierarchical threshold system, and adaptively updating threshold parameters when a new abnormal pattern is detected to generate a hierarchical risk result; a case migration learning unit for comparing the hierarchical risk result with the historical case library one by one, directly outputting a corresponding disposal result if there is a high-confidence similar case, and if no match is found, storing the current risk features in the case library and generating a to-be-labeled item for the case library, and when receiving external input of labeled data, automatically updating the case library to maintain accuracy and generating a case comparison result; an evaluation summary and output unit for summarizing and evaluating the hierarchical risk result and the case comparison result, generating a risk evaluation result and pushing corresponding early warning information to the management terminal; a feedback closed-loop unit for receiving disposal feedback from the management terminal and inputting the disposal feedback to the risk identification module and the intelligent decision module respectively to continuously optimize the identification and decision parameters, forming a closed-loop self-learning mechanism.

[0052] In an embodiment of the present application, the hierarchical adaptive discrimination unit can calculate a risk score according to the preliminary risk data, and automatically match a processing strategy using a hierarchical threshold system. When the system detects a new abnormal pattern, it can also automatically update the threshold parameters to form an adaptive hierarchical risk result. This dynamic updating mechanism can ensure that the risk discrimination standard is continuously adjusted as the transaction pattern evolves. For example, when a new risk pattern of "cross-border short-term multiple transfers" has never occurred in history, the system can automatically lower the discrimination threshold for this pattern, so as to capture such risks more quickly.

[0053] The case migration learning unit can compare the hierarchical risk result with the historical case library one by one, directly output the corresponding disposal result if there is a high-confidence similar case, and if no match is found, automatically store the new risk features in the case library, and update the case library after receiving external input of labeled data to maintain its integrity and accuracy. This migration learning mechanism can improve the response efficiency of the system to known risk patterns, and at the same time, through continuous expansion of the case library, improve the coverage rate of unknown risks. For example, when a new case similar to the historical "fake associated company transfer" appears, the system can directly output a reference disposal suggestion, saving the time for manual judgment.

[0054] The evaluation aggregation and output unit can aggregate and evaluate the hierarchical risk results and the case comparison results, generate a final risk evaluation result, and push the result to the management terminal to realize rapid early warning and intervention. At the same time, the feedback closed-loop unit can receive the disposal feedback of the management terminal and input it to the risk identification module and the intelligent decision module in a reverse direction to form a closed-loop self-learning mechanism, so that the parameter configuration and identification ability of the system are continuously optimized according to the actual operation situation. For example, when the management personnel confirms that a low-risk label is a misjudgment, the system will automatically correct the relevant parameters to avoid subsequent repetitive misjudgments.

[0055] Through the above mechanism, the embodiment of the present application can realize the intelligentization of the whole process from risk discovery, risk evaluation to disposal feedback, has high efficiency and self-adaptability, and effectively improves the accuracy and real-time performance of the risk prevention and control of the Internet property transaction.

[0056] The evaluation aggregation and output unit specifically includes: First, the hierarchical risk results output from the hierarchical self-adaptive discrimination unit are received, and the results include scores for different risk levels and corresponding risk labels. At the same time, the case comparison results from the case migration learning unit are received, and the results include whether similar historical cases are found and cited disposal suggestions.

[0057] Then, the hierarchical risk results and the case comparison results are aggregated. The aggregation process is as follows: when the hierarchical risk results show a high risk level and the case comparison results do not find similar cases, the system will increase the weight of manual review in the evaluation; when the hierarchical risk results show a medium risk level and the case comparison results are consistent with high-confidence historical cases, the system will preferentially adopt the disposal method of the historical cases. In this way, the system can integrate discrimination results from different sources to form a more comprehensive risk evaluation.

[0058] Next, the final risk evaluation result is generated according to the aggregated results. The generated content includes: risk level, triggered main risk factors, matched historical case number, and disposal suggestions, etc. If the result is high risk, a priority label will also be automatically set to facilitate rapid processing on the management terminal.

[0059] Finally, the generated risk evaluation result is pushed to the management terminal through an interface and displayed in a visual manner, including text prompts, chart displays, and risk link diagrams. This can help management personnel intuitively understand the risk sources and risk levels, thereby improving the response efficiency of early warning.

[0060] The feedback closed-loop unit specifically includes: First, receive the handling feedback information from the management terminal, and the feedback information includes risk confirmation results (such as confirming abnormality or misjudgment), handling measures taken (such as suspending transaction, manual review, and releasing transaction), and feedback time.

[0061] Then, the handling feedback is converted into structured feedback data. The structuring process includes extracting the risk factors confirmed in the feedback, labeling whether it is a real risk, and recording the final handling result. For example, when the management terminal confirms that a certain transaction chain is a "false transfer", the system will label the chain as high risk in the feedback data and add the handling measure "transaction suspension".

[0062] Next, the feedback data is input into the risk identification module in reverse, for adjusting the risk factor weight and the identification strategy of abnormal links. For example, if a certain type of behavior chain is labeled as a misjudgment in multiple feedbacks, the system will reduce the sensitivity of this type of factor to reduce subsequent false positives.

[0063] Further, the feedback data is input into the intelligent decision-making module in reverse, for optimizing the hierarchical threshold and the case library content. For example, when a new risk pattern first appears and is confirmed by the management terminal, the system will automatically write its features into the case library and add corresponding discrimination conditions in the hierarchical threshold, so that similar patterns can be quickly identified in the future.

[0064] Finally, a closed-loop self-learning mechanism is formed. This mechanism ensures that the system can continuously optimize the model and parameters according to actual feedback after each risk assessment and handling, thereby improving the adaptability and accuracy of the system in new scenarios.

[0065] In a preferred embodiment of the present application, the relationship graph generation unit comprises: An entity standardization processing subunit is configured to perform cross-platform entity identification and deduplication on the participants according to the standardized transaction data, and generate a unified node set. A time sequence and domain labeling subunit is configured to write timestamps and regional or platform source identifiers on the edges and write role types and risk priori labels in the nodes according to the unified node set and transaction records, and form an initial relationship graph with time sequence and domain attributes. An association degree calculation subunit is configured to calculate the interaction frequency, freshness, role similarity, and path credibility between nodes within a preset upper limit of hop count according to the initial relationship graph, and generate association degree scores of node pairs to form a weighted relationship graph. A graph incremental updating subunit is configured to perform incremental updating on the weighted relationship graph according to newly arrived transaction data, generate an effective graph snapshot conforming to a sliding time window, and invalidate edges that exceed the window period according to rules, so as to ensure the real-time and effectiveness of the graph.

[0066] In the embodiment of the present application, the entity standardization processing subunit can perform cross-platform entity recognition and deduplication on the participants according to the standardized transaction data, and generate a unified node set. Through this processing, the system can eliminate the problem of repeated records caused by differences in name writing, differences in registration code, or inconsistencies in contact information. For example, when the same enterprise appears as "XX Technology Co., Ltd." and "XX Technology Co., Ltd." on different platforms, the system can identify them as the same subject through multi-feature comparison, thereby avoiding the problem of node dispersion and inaccurate identification in the risk analysis process.

[0067] The time sequence and domain labeling subunit can write timestamps and regional or platform source identifiers on the edges of the relationship graph according to the unified node set and transaction records, and write role types and risk priori labels in the nodes, forming an initial relationship graph with time sequence and domain attributes. In this way, the system can intuitively reflect the time sequence and jurisdiction of the transaction in the graph. For example, when the same patent is first transferred domestically and then transferred on an overseas platform, the system can clearly record the cross-regional transaction path, facilitating subsequent risk tracing.

[0068] The correlation degree calculation subunit can calculate the interaction features between nodes within the preset upper limit of hop count, including the number of interactions, the freshness of the latest interaction time, role similarity, and multi-hop path credibility, and generate a correlation score for the node pair based on the initial relationship graph, thereby forming a weighted relationship graph. Through this weighting mechanism, the system can more accurately reveal the closeness between transaction subjects. For example, although two companies have few direct transactions, they frequently transact indirectly through multiple intermediate nodes. The system will give a high correlation score through path credibility calculation, thereby revealing hidden interests.

[0069] The graph incremental update subunit can update the weighted relationship graph incrementally based on newly arrived transaction data, and generate an effective graph snapshot that meets the sliding time window, while invalidating edges that exceed the window period according to the rules, to ensure the real-time and effectiveness of the graph. In this way, the system can avoid the inefficiency of rebuilding the full graph every time new data arrives, and achieve rapid response to real-time transaction dynamics. For example, on a large-scale intellectual property transaction platform, tens of thousands of transfer records may be added every day. The system can only update the newly added part and generate the latest snapshot, ensuring the timeliness and efficiency of risk assessment.

[0070] The entity standardization processing subunit specifically includes: First, the participant information in the transaction data is extracted, including name, registration code, contact information, and registered address.

[0071] Then, according to the record differences of different platforms, entity recognition is performed by name fingerprint matching, character similarity comparison, unified code verification and the like. For example, when "XX Technology Co., Ltd." and "XX Technology Co., Ltd." appear at the same time, by removing redundant words and comparing the registration codes, the same principal can be identified.

[0072] Next, the repeatedly identified participants are de-duplicated, and a unique identification code is uniformly assigned to form a consistent node information across platforms.

[0073] Finally, a unified node set is generated and stored in the node library of the relationship graph, for subsequent labeling and calculation.

[0074] Among them, the time sequence and domain labeling subunit specifically includes: First, the time information of each transfer event is extracted from the transaction record, and written into the transaction edge as a timestamp label, reflecting the order of occurrence of the transaction.

[0075] Then, according to the transaction source platform or the jurisdiction of the transaction, the region or platform identifier is added to the transaction edge. For example, when a transfer occurs on a domestic platform, the edge is labeled as "domestic"; when it occurs on an overseas platform, it is labeled as "overseas".

[0076] Next, the role type is written for the participating nodes, such as "transferor", "transferee" and "intermediary institution", and according to historical data or existing labels, some nodes are assigned with risk priori labels, such as "high-risk principal" or "abnormal frequent transaction party".

[0077] Finally, the graph with time attributes, regional attributes and node labels is output to form an initial relationship graph, providing input data for subsequent correlation calculation.

[0078] Among them, the correlation degree calculation subunit specifically includes: First, select any two nodes in the graph, and within the set number of hops, count the number of direct interactions and the number of paths between them.

[0079] Then, according to the time interval between the last interaction and the current interval, the freshness of the interaction is determined. The shorter the time interval, the higher the interaction weight; the longer the interval, the interaction weight gradually decays.

[0080] Next, the role similarity of the two nodes is compared. For example, if both are transferors, the role similarity score is low; if one is a transferor and the other is a transferee, the role complementarity is strong and the score is high.

[0081] Further, the credibility of the multi-hop path is analyzed. When there are multiple intermediate nodes between two nodes, the longer the path, the credibility is reduced by a proportional step by step. For example, when two nodes are connected through three layers of intermediaries, the contribution of this path to the final correlation degree will be significantly less than that of the directly connected path.

[0082] Finally, the weighted results of the interaction times, freshness, role similarity and path credibility are normalized to generate a correlation score of the node pair, and a weighted relationship graph is formed accordingly.

[0083] The graph incremental updating subunit specifically comprises: First, when new transaction data arrives, it is mapped to the existing weighted relationship graph. If it involves a new node, the node is added. If it involves an existing node, the interaction attributes of its edges are updated.

[0084] Then, according to the sliding time window mechanism, only the transaction edges within the window range are retained. For example, when the time window is set to six months, the system will automatically delete the transaction edges that have not occurred for more than six months.

[0085] Next, the retained edges are recalculated, including updating the interaction freshness, readjusting the path credibility, and generating a new correlation score to ensure that the score always reflects the current transaction state.

[0086] Finally, the updated effective graph snapshot is output as input data for subsequent risk identification and decision analysis, avoiding the reconstruction of the full graph each time the data is updated, and improving the system running efficiency and real-time performance.

[0087] In a preferred embodiment of the present application, the feature fusion weighting unit comprises: The feature weight distribution subunit is used to distribute learnable weights to the time series feature data and the graph association feature data based on transaction type, transaction amount, data completeness and near-end time features, and generate a weighted feature vector; The case prior guidance subunit is used to compare the similarity of the weighted feature vector with the abnormal case feature cluster in the historical case library, and when the similarity is higher than the threshold, the weight distribution result is biased and adjusted to generate a feature vector guided by priori. The robust normalization processing subunit is used to perform median scaling and extreme value truncation on the abnormal values in the feature vector guided by priori to obtain normalized robust feature data. The fusion calibration subunit is used to splice the time series features and the graph association features according to the normalized robust feature data to obtain the fusion feature data.

[0088] In the embodiment of the present application, the feature weight distribution subunit can assign learnable weights to the time series feature data and the graph association feature data based on transaction type, transaction amount, data integrity, and near-time features, and generate a weighted feature vector. In this way, the system can automatically highlight features that have a greater impact on risk identification. For example, for a particularly high cross-border transfer, the system can automatically assign a higher weight to the amount feature, thereby improving risk perception.

[0089] The case prior guidance subunit can compare the weighted feature vector with the abnormal case feature clusters in the historical case library for similarity, and when the similarity is higher than a threshold, bias the weight distribution result to generate a prior-guided feature vector. Through this mechanism, the system can fully utilize historical knowledge and integrate experiential information into current feature processing. For example, when a certain transaction feature combination is highly similar to a historical "shell company patent transfer" case, the system will automatically amplify the impact of such features, thereby quickly prompting potential risks.

[0090] The robust normalization processing subunit can perform median scaling and extreme value truncation on the abnormal values in the prior-guided feature vector to obtain normalized robust feature data. This processing can reduce the interference of noise or extreme data on the overall feature fusion result. For example, when a transaction amount is abnormally high but only an isolated instance, the system will weaken its excessive impact on the overall model through extreme value truncation, maintaining the stability of risk assessment.

[0091] The fusion calibration subunit can splice the time series features and the graph association features based on the normalized robust feature data to obtain fusion feature data. At the same time, the output distribution of the fusion feature data is subjected to monotonic transformation calibration to keep the risk score consistent with the actual occurrence frequency between different batches, and an explainability record containing feature contribution and impact direction is generated. This mechanism not only improves the stability of the model, but also enhances the explainability of the results. For example, when the system determines that a transaction is high-risk, it can output "frequent cross-border transfer" and "high association with high-risk nodes" as the main basis, facilitating the understanding and acceptance of regulatory personnel.

[0092] Through the synergistic effect of the above-mentioned subunits, the embodiment of the present application can realize dynamic weight distribution, prior knowledge guidance, robust normalization, and output calibration in the process of multi-dimensional feature fusion, significantly improving the accuracy and explainability of risk assessment.

[0093] The feature weight distribution subunit specifically includes: First, the time series feature data and the graph association feature data are collected and integrated to ensure that the time information, transaction amount, and transaction subject quantity of each transaction event are extracted.

[0094] Then, according to the preset learnable model parameters, the importance of different features is calculated in turn. In the calculation process, the transaction type (such as cross-border transfer, frequent transfer), the transaction amount, the data completeness and the time characteristics are taken as the main factors, and different initial weights are given.

[0095] Next, in the training or running stage, the system continuously adjusts the weight value, so that the weight distribution result can be closer to the historical risk identification experience. For example, for cross-border high-amount transactions, the system will gradually learn that higher weights should be given to transaction amount and regional characteristics.

[0096] Finally, the assigned weight is multiplied with the original feature vector to obtain a new weighted feature vector, which is output to the next subunit.

[0097] Among them, the case prior guidance subunit specifically includes: First, the stored abnormal case feature set is retrieved from the historical case library, and each case cluster contains important feature combinations that have shown high risk in previous transactions.

[0098] Then, the weighted feature vector is compared with these historical case clusters one by one to calculate the similarity. The calculation process uses a similarity-based feature dimension measurement method, which is described in words as follows: When the difference between several key features in the weighted vector and the corresponding features in the historical case is small, the similarity score is high.

[0099] Next, when the similarity score exceeds the preset threshold, the system will refer to the importance ranking of high-risk features in the historical case to bias adjust the weight distribution result of the current features. For example, if the case shows that "frequent cross-platform transfer in a short time" is a high-risk factor, the system will increase the proportion of this factor in the weighted vector.

[0100] Finally, the prior-guided feature vector after bias adjustment is output to play a stronger role in risk identification in subsequent processing steps.

[0101] Among them, the robust normalization processing subunit specifically includes: First, collect each feature component in the current feature vector and calculate its overall distribution, focusing on analyzing the extreme values and deviation values.

[0102] Then, taking the median of the feature components as the center reference, all feature values are scaled to make the overall feature distribution more balanced, avoiding the influence of individual abnormal values on the overall result.

[0103] Next, the extreme values at the uppermost and lowermost ends of the distribution are identified and subjected to mild truncation. The truncation method is described in words: values exceeding the upper percentile are reduced to the vicinity of the upper percentile, and values below the lower percentile are raised to the vicinity of the lower percentile, thereby limiting the interference of extreme values on the overall vector.

[0104] Finally, the results of the scaling and truncation process are output to form normalized robust feature data, ensuring the stability and reliability of the subsequent risk assessment process.

[0105] The fusion calibration subunit specifically includes: First, the normalized time series features and the normalized graph-related features are spliced according to the feature dimensions to obtain a complete fusion feature dataset.

[0106] Then, the fusion feature data is subjected to distribution calibration. The calibration method is: comparing the risk score distribution of the current batch of fusion features with the actual risk frequency in the historical batch, if there is a deviation, then a monotonic transformation is performed on the output distribution to make it more consistent with the occurrence probability of the actual risk.

[0107] Next, in the calibration process, the system generates an explainability record, which includes the feature items with the highest contribution and their influence direction on risk determination. For example, when the cross-border transfer amount plays a major role in the risk result, the record will explicitly identify the contribution of this feature.

[0108] Finally, the calibrated fusion feature data is output as the final output, which is provided to the risk identification module and the intelligent decision module for calling, ensuring that the risk assessment is both stable and interpretable.

[0109] In a preferred embodiment of the present application, the suspicious behavior chain extraction unit includes: The transaction sequence segmentation subunit is used to segment continuous transaction events in chronological order according to the fusion feature data, and divide them into candidate sub-sequences according to transfer amount changes, transaction frequency abnormalities, and cross-domain occurrence conditions to obtain candidate behavior chain data. The behavior chain feature encoding subunit is used to progressively encode node roles, transfer directions, transaction freshness, and node hop lengths according to the candidate behavior chain data to generate a multi-dimensional behavior chain feature vector. The abnormal accumulation determination subunit is used to gradually calculate the risk accumulation value of continuous transactions according to the multi-dimensional behavior chain feature vector, and when the risk accumulation value exceeds the dynamic threshold, the corresponding candidate behavior chain is marked as a suspicious behavior chain, and behavior chain feature data is generated.

[0110] In the embodiment of the present application, the transaction sequence segmentation subunit can perform time sequence segmentation on continuous transaction events according to the fused feature data, and divide them into candidate sub-sequences according to the change of transfer amount, transaction frequency anomaly and cross-domain occurrence condition to obtain candidate behavior chain data. In this way, the system can decompose complex transaction flow into several behavior chain segments with clear characteristics, which is convenient for subsequent anomaly identification. For example, when a patent is continuously transferred for multiple times in a very short time, the system will automatically divide these transfers into a candidate chain segment as a potential suspicious object.

[0111] The behavior chain feature encoding subunit can progressively encode the node role, transfer direction, transaction freshness and node hop length according to the candidate behavior chain data to generate a multi-dimensional behavior chain feature vector. This step-by-step encoding method can capture the complex relationships within the chain, not only reflecting single events, but also showing the chain transmission pattern. For example, the system can discover through the feature vector that a certain behavior chain presents a circular pattern of "from parent company to multiple subsidiaries and then backflow".

[0112] The abnormal accumulation determination subunit can calculate the risk accumulation value of continuous transactions step by step according to the multi-dimensional behavior chain feature vector, and when the risk accumulation value exceeds the dynamic threshold, mark the corresponding candidate behavior chain as a suspicious behavior chain and generate behavior chain feature data. This determination mechanism can effectively discover chain risks with multiple factors superimposed. For example, a single cross-platform transfer may not be abnormal, but if it is accompanied by high-frequency transfers of high amounts and forms a backflow link, the risk accumulation value will quickly rise, triggering a high-risk label, thereby achieving accurate early warning.

[0113] Through the above processing, the embodiment of the present application can convert scattered transaction events into organic behavior chains, and reveal potential hidden risk patterns through step-by-step encoding and accumulation determination, significantly enhancing the ability to discover complex abnormal transaction chains.

[0114] The transaction sequence segmentation subunit specifically includes: First, sort the transaction events in the fused feature data by timestamp to ensure the continuity of the transaction events in time.

[0115] Then, according to the set time window and event interval, divide the continuous transaction events into different segments. For example, when multiple transaction events occur within a week and involve the same or related subjects, the system will classify these events into the same sequence.

[0116] Next, detect the change trend of the transfer amount for the divided transaction sequence, and if it is found that the transaction amount increases or decreases significantly in a short period of time, it is marked as an abnormal condition.

[0117] Further, the frequency of events in the sequence is analyzed. When the transaction frequency exceeds the preset threshold of the historical average level, the sequence is marked as a high-frequency anomaly. If there is a transfer behavior across platforms or regions in the same chain, it is also recorded as a cross-domain anomaly condition in the sequence features.

[0118] Finally, the transaction sequences that meet the above conditions are extracted as candidate sub-sequences and stored uniformly as candidate behavior chain data for subsequent encoding sub-unit processing.

[0119] Among them, the behavior chain feature encoding sub-unit specifically includes: First, the role type of each node in the candidate behavior chain is extracted, such as "transferor", "transferee" or "intermediary", and encoded as the corresponding role label.

[0120] Then, the transfer direction of the transaction event is encoded. The basis for judging the direction is the transfer path of the property right subject, for example, "from parent company to subsidiary" is encoded as downward transmission, and "from subsidiary back to parent company" is encoded as backflow transmission.

[0121] Next, the transaction freshness is calculated and encoded. In words: the system will give higher freshness values to recent transactions according to the interval between the transaction time and the current time, and lower values to transactions with longer intervals, to reflect the timeliness of the transaction.

[0122] Further, the node hop count length is encoded. The hop count refers to the number of intermediate nodes that a transaction passes through from the starting node to the target node. The more hops, the more complex the path, and the system will accordingly reduce the credibility score when encoding.

[0123] Finally, the above role encoding, direction encoding, freshness value and hop count features are spliced into a multi-dimensional feature vector and output as input data for the subsequent judgment sub-unit.

[0124] Among them, the abnormal accumulation judgment sub-unit specifically includes: First, the input multi-dimensional behavior chain feature vector is scored item by item. During the scoring process, the system will set risk contribution values for high-amount transfers, frequent transactions and cross-domain paths.

[0125] Then, according to the time sequence of the transaction, these risk contribution values are sequentially accumulated to form a risk accumulation value sequence. In words: the system will gradually add the risk values of early transactions to the risk values of subsequent transactions to obtain a cumulative risk curve that changes over time.

[0126] Then, the cumulative risk value is compared with a dynamic threshold. The dynamic threshold is not a fixed value, but is adjusted in real time according to historical statistical distribution and current trading environment, for example, in a period when market transactions are abnormally active, the threshold is automatically increased to avoid false positives.

[0127] Finally, when the cumulative risk value exceeds the dynamic threshold, the system marks the corresponding candidate behavior chain as a suspicious behavior chain, and generates structured behavior chain feature data, including the nodes in the chain, the path, the triggered high-risk factors and their cumulative results. This data will be passed as input to the subsequent interest flow tracing unit and composite risk discrimination unit.

[0128] In a preferred embodiment of the present application, the interest flow tracing unit comprises: a key node screening subunit for identifying nodes participating in multiple transactions in the relationship graph according to the behavior chain feature data, and screening according to the transaction amount concentration and node association degree weight to generate a key node set; a trajectory backtracking construction subunit for backtracking from each key node as a starting point along the transaction link to record the node role type, interaction frequency and regional label, and form interest flow trajectory data; an abnormal trajectory labeling subunit for generating an abnormal label for abnormal paths when detecting that the transfer path crosses different regions, different platforms or has multi-level anonymous transfer according to the interest flow trajectory data, and taking the labeled trajectory as the trace risk data.

[0129] In the embodiment of the present application, the key node screening subunit can identify nodes participating in multiple transactions in the relationship graph according to the behavior chain feature data, and screen out key nodes as the starting point and end point of interest flow by combining the transaction amount concentration and node association degree weight. In this way, the system can quickly locate the participants with control or core influence in the transaction network. For example, when a company repeatedly appears in multiple chains and occupies a major proportion of transaction amount, the system will mark it as a key node for subsequent key tracking.

[0130] The trajectory backtracking construction subunit can backtrace from each key node as a starting point along the transaction link to record the node role type, interaction frequency and regional label, and form complete interest flow trajectory data. In this way, the system can reveal the flow path of funds or property rights among different nodes and reconstruct the potential interest transfer mode. For example, when a key node has continuous transfer relationship with multiple overseas companies, the system can clearly draw the cross-border interest transfer path.

[0131] The abnormal trajectory labeling subunit can label abnormal trajectories according to the interest flow trajectory data, generate an abnormal label when detecting that the transfer path crosses different regions, different platforms or there is multi-level hidden transfer, and output the labeled trajectory as the traceability risk data. Through this mechanism, the system can automatically identify and label the transfer path that is not reasonable or has hidden. For example, when a patent is transferred from a domestic enterprise to an overseas associated enterprise through multiple shell companies, the system can label this path as an abnormal trajectory and output it for the supervision or auditing personnel to focus on.

[0132] Through the above functions, the embodiment of the present application can realize the whole process of traceability from key node identification to complete trajectory construction, and then to abnormal path labeling, which significantly improves the ability of the system to reveal hidden interest flow and cross-domain risks.

[0133] The key node screening subunit specifically includes: First, receive the behavior chain feature data, and map the transaction subjects involved in it to the relationship graph.

[0134] Then, count the number of occurrences of each node in the transaction chain and the total amount of transaction amount involved. The more the number of occurrences, the more active the node is in the chain; the higher the transaction amount ratio, the more important the node is in the transfer of funds or property rights.

[0135] Next, according to the interaction intensity of the node and other nodes, the correlation degree of the node is calculated. The literal description is: if there is frequent and recent interaction between a node and multiple other nodes, the correlation degree score of the node is higher.

[0136] Further, the transaction amount concentration and the node correlation degree are weighted and combined to obtain a comprehensive score. Nodes with higher comprehensive scores are screened out and included in the key node set.

[0137] Finally, the key node set is output as the starting set for subsequent trajectory backtracking construction.

[0138] The trajectory backtracking construction subunit specifically includes: First, select a node from the key node set as the backtracking starting point.

[0139] Then, step back along the time sequence of the transaction chain, and retrieve the historical transaction path of the node in turn. Every time a node is passed, its role type (such as the transferor, the transferee or the intermediary), the number of interactions with the previous node, and the regional or platform information to which the node belongs are recorded.

[0140] Next, if a new key node is encountered during the backtracking process, it is included in the path, and the backtracking range is continued to be expanded until the set backtracking depth is reached or there is no more historical transaction data.

[0141] Further, all path information is integrated to generate complete benefit flow trajectory data. The trajectory not only contains path order, but also carries role information, frequency information and regional label of each node, thereby forming a traceable chain record.

[0142] Finally, the generated trajectory data is outputted to provide input for subsequent abnormal annotation and risk identification.

[0143] The abnormal trajectory annotation subunit is used, and specifically comprises: First, the benefit flow trajectory data is received, and the transfer path in each piece of data is searched.

[0144] Then, the regional label of the path is checked, and if it is found that the path crosses different jurisdictions or different trading platforms, it is determined that there is an abnormal situation of cross-domain or cross-platform transfer.

[0145] Next, the hierarchical structure of the path is analyzed. The textual description is as follows: when a property transfer path passes through multiple intermediate nodes, and the intermediate nodes are all low-activity or hollow subjects, the path will be marked as hidden transfer.

[0146] Further, an abnormal identifier is added to the path determined to be abnormal. The abnormal identifier includes the path abnormal type (cross-domain, cross-platform or hidden) and the corresponding node sequence.

[0147] Finally, the complete trajectory after identification is outputted as traceability risk data for subsequent processing of the composite risk discrimination unit.

[0148] In a preferred embodiment of the present application, the hierarchical adaptive discrimination unit comprises: The risk score generation subunit is used to calculate and output a risk score sequence based on transaction amount abnormality, participant historical risk label and time span factor according to the preliminary risk data; The hierarchical threshold comparison subunit is used to compare the risk score in layers according to at least three risk level threshold intervals based on the risk score sequence, to generate a hierarchical discrimination result; The threshold adaptive adjustment subunit is used to adjust the upper and lower limits of the threshold when it is detected that the risk score continuously deviates from the preset threshold interval based on the hierarchical discrimination result, and write the updated threshold parameters into the decision parameter library to generate an adaptive updated hierarchical risk result.

[0149] In the embodiment of the present application, the risk score generation subunit can calculate and output a risk score sequence according to the preliminary risk data, taking into account the transaction amount abnormality, the historical risk label of the participant and the time span factor. In this way, the system can quantitatively express the transaction risk, providing an objective basis for subsequent multi-layer discrimination. For example, when the transaction amount is much higher than the average level of the platform and the participant has a history of high-risk records, the risk score will be significantly increased, facilitating priority disposal.

[0150] The hierarchical threshold comparison subunit can set threshold intervals according to at least three risk levels according to the risk score sequence, and compare the risk score layer by layer to generate a hierarchical discrimination result. Through this mechanism, the system can automatically match different processing strategies for transactions of different risk levels. For example, low-risk transactions only need to be recorded by the system, medium-risk transactions are pushed for early warning, and high-risk transactions are directly put into the manual review process, thereby realizing the grading and efficiency of risk disposal.

[0151] The threshold adaptive adjustment subunit can dynamically adjust the upper and lower limits of the threshold when it is detected that the risk score deviates from the preset threshold interval, and write the updated threshold parameters into the decision parameter library to generate an adaptively updated hierarchical risk result. This mechanism can ensure that the discrimination standard is continuously optimized with the emergence of new transaction patterns. For example, when a new evasive method appears recently, causing frequent missed judgments under the traditional threshold, the system will automatically lower the threshold for this type of pattern to identify it in time, thereby improving the adaptability to new risk forms.

[0152] Through the above functions, the embodiment of the present application can realize a complete process of risk quantification, hierarchical discrimination and adaptive update, so that the risk assessment has both fine-grained levels and the ability to dynamically adjust with the scene, greatly improving the accuracy and flexibility of transaction risk prevention and control.

[0153] The risk score generation subunit specifically includes: First, receive preliminary risk data, which includes the transaction amount, occurrence time, participant basic information and risk label of the transaction chain.

[0154] Then, calculate the abnormality of the transaction amount. In words, the system compares the current transaction amount with the average level of similar transactions, and gives a higher abnormality score when the transaction amount is much higher or lower than the average value.

[0155] Next, score according to the historical risk label of the participant. For example, if a transaction subject has been marked as high risk in past transactions, its risk label score will be higher; if it is a new subject appearing for the first time, its risk label score will be lower.

[0156] Further, the time span of transaction occurrence is modified. When the time interval of consecutive transactions is too short, the system will increase the risk score to reflect the risk characteristics of high-frequency transactions; when the interval time is longer, the score will be reduced accordingly.

[0157] Finally, the above multiple factors are weighted and summed to generate the risk score of a single transaction, and the risk scores of consecutive transactions are arranged in sequence to form a risk score sequence, which is output for subsequent discrimination.

[0158] Among them, the hierarchical threshold comparison subunit specifically includes: First, at least three risk levels are set, such as low risk, medium risk and high risk, and each level is assigned a corresponding threshold interval. In words: low risk corresponds to a score below the first threshold, medium risk corresponds to a score between the first threshold and the second threshold, and high risk is above the second threshold.

[0159] Then, each score in the risk score sequence is compared with the threshold interval one by one. If a score falls in the low risk interval, it is marked as low risk; if it falls in the middle interval, it is marked as medium risk; if it exceeds the highest threshold, it is marked as high risk.

[0160] Next, all the marked results are integrated to form a hierarchical discrimination result. The integration method is: when a transaction chain contains multiple high-risk scores, the chain as a whole is identified as high risk; if most of the scores fall in the low risk interval, the whole is identified as low risk.

[0161] Finally, the generated hierarchical discrimination result is output for subsequent combination with the case comparison module to form a more accurate risk assessment.

[0162] Among them, the threshold adaptive adjustment subunit specifically includes: First, the hierarchical discrimination result is received, and the scores falling into the boundary region are counted. The boundary region refers to the score interval close to the upper and lower limits of the threshold.

[0163] Then, it is determined whether there are a large number of risk scores concentrated near a certain boundary. If it is found that a large number of transactions have scores concentrated near the upper limit of the threshold, it means that the current threshold is set too low; if they are concentrated near the lower limit, it means that the threshold is set too high.

[0164] Next, the threshold is adjusted according to the concentration. In words: when the risk score frequently approaches the high threshold, the system will appropriately increase the upper limit of the high threshold; when the score is long-term at the low threshold boundary, the system will lower the lower limit of the low threshold, so that the risk stratification is more in line with the actual situation.

[0165] Further, the adjusted threshold parameter is written into the decision parameter library and applied as a new threshold standard in subsequent risk assessment tasks.

[0166] Finally, an updated hierarchical risk result is generated to ensure that the risk level division can be adaptively adjusted with changes in market environment and transaction patterns.

[0167] In a preferred embodiment of the present application, the case migration learning unit comprises: a feature similarity comparison subunit configured to extract a key feature vector from the hierarchical risk result and compare it with feature sets in the historical case library one by one to form a candidate similar case set; a confidence score comprehensive determination subunit configured to calculate a comprehensive score according to the candidate similar case set, in combination with the similarity score and the case handling accuracy, and output a corresponding handling result when the comprehensive score exceeds a preset threshold, and form a case determination result; a case library expansion and update subunit configured to write current risk feature data into the case library to generate a to-be-labeled item when a high-confidence case is not matched, and expand and correct the case library to generate a case comparison result when receiving externally input labeled data.

[0168] In the embodiment of the present application, the feature similarity comparison subunit can extract a key feature vector from the hierarchical risk result and compare it with feature sets in the historical case library one by one to form a candidate similar case set. In this way, the system can quickly find similar historical cases of the current risk event to provide a reference basis for subsequent judgment. For example, when a transaction is highly consistent with a typical fraud case in terms of amount, region, and associated principal, the system will automatically include the case in the candidate set.

[0169] The confidence score comprehensive determination subunit can calculate a comprehensive score according to the candidate similar case set, in combination with the similarity score and the historical case handling accuracy, and output a corresponding handling result when the comprehensive score exceeds a preset threshold, and form a case determination result. In this way, the system can reuse historical experience with high confidence to speed up the handling of similar risks. For example, when the comprehensive score shows that a transaction is highly consistent with a known "fake associated transfer" case, the system can directly output a handling suggestion to shorten the review time.

[0170] The case base extension updating subunit can write the current risk feature data into the case base to generate a to-be-labeled item when a high-confidence case is not matched according to the case determination result, and extend and correct the case base after receiving the externally input labeled data to generate a case comparison result. In this way, the system can continuously supplement and correct the case base content, improving its coverage and accuracy. For example, when a new risk pattern first appears and is confirmed by manual labeling, the system will automatically store the pattern in the case base, so that it can be quickly identified in subsequent transactions.

[0171] Through the above functions, the embodiment of the present application can realize automatic retrieval of similar cases, high-confidence determination and self-extension of the case base in risk assessment, which not only improves the response speed of the system to known risks, but also enhances the learning and adaptation ability of the system to emerging risks.

[0172] The feature similarity comparison subunit specifically includes: First, the key feature vectors used for comparison are extracted from the hierarchical risk results, including transaction amount, risk label of the participant, transaction frequency, cross-platform transfer situation and complexity of the behavior chain, etc.

[0173] Then, the stored case feature sets in the historical case base are called, and each case set corresponds to one or more typical risk patterns.

[0174] Next, the extracted key feature vectors are compared with the features in the case base one by one. The comparison method is described as follows: if the key features and the feature values in the case have small differences in numerical value or attribute, the similarity score is high; if the differences are large, the similarity score is low.

[0175] Further, according to the comparison result, the cases with similarity scores exceeding a set threshold are screened out and form a candidate similar case set.

[0176] Finally, the candidate similar case set is output as an output for subsequent analysis by the confidence comprehensive determination subunit.

[0177] The confidence comprehensive determination subunit specifically includes: First, the candidate similar case set is received, and the similarity score and the historical disposition record of each candidate case are obtained.

[0178] Then, the disposition result accuracy of each candidate case is counted. The textual representation is as follows: the accuracy rate represents the proportion of cases that are confirmed to be correctly identified and properly disposed of in the past application process.

[0179] Next, the similarity score and the historical disposition accuracy are combined by weighting to obtain a comprehensive score. The higher the similarity and the higher the disposition accuracy, the higher the comprehensive score.

[0180] Further, the comprehensive score is compared with a preset threshold value. When the score exceeds the threshold value, the system directly outputs the treatment result corresponding to the case and generates a case determination result; when the score does not reach the threshold value, the next subunit processing is entered.

[0181] Finally, the case determination result is taken as an output, which is used for the final summary of the risk assessment result.

[0182] Among them, the case library expansion and update subunit specifically includes: First, when the confidence comprehensive determination subunit fails to output a high-confidence case, the risk feature data of the current transaction is extracted and stored as a to-be-labeled item.

[0183] Then, the to-be-labeled item is added to the temporary storage area of the case library, and waits for the management personnel or external system to provide labeling information. The labeling information includes whether the risk event is abnormal, the final treatment method, etc.

[0184] Next, after receiving the externally input labeling data, the system will update the case library, convert the item into a formal case, and compare it with the original case to avoid repeated records.

[0185] Further, the existing cases in the case library are corrected. For example, when a case is confirmed to be misjudged for multiple times subsequently, the system will reduce the weight of the case in the comparison process, to ensure the accuracy of the case library.

[0186] Finally, the updated case comparison result is generated and used as a new case set for subsequent risk event comparison, so as to continuously improve the coverage and practicability of the case library.

[0187] The above is the preferred embodiment of the present application. It should be noted that for ordinary skilled persons in the technical field, several improvements and refinements can be made without departing from the principles of the present application, and these improvements and refinements should also be considered as the protection scope of the present application.

Claims

1. An artificial intelligence-based internet property rights transaction risk assessment and decision-making system, characterized in that, The system includes: The data acquisition and preprocessing module is used to collect raw transaction data from the Internet property rights trading platform in real time, and to clean and standardize the raw transaction data to obtain standardized transaction data. The raw transaction data includes information on the property rights, historical records of the trading parties, and information on related parties. The multi-dimensional feature construction module is used to model and construct a multi-level interaction relationship graph based on standardized transaction data according to time windows, thereby obtaining time series feature data and graph-related feature data; and to normalize and weight the features to form fused feature data. The risk identification module is used to extract suspicious transfer behavior chains and trace the flow of benefits based on fused feature data and multi-level interaction relationship graphs, perform composite risk judgment, and generate preliminary risk data. The intelligent decision-making module is used to perform hierarchical adaptive discrimination and case migration comparison based on preliminary risk data, summarize and generate risk assessment results and push early warnings; it also receives feedback from the management terminal and uses it to optimize identification and decision parameters.

2. The Internet property rights transaction risk assessment and decision-making system based on artificial intelligence according to claim 1, characterized in that, The multidimensional feature construction module includes: The time window modeling unit is used to group transaction sequences according to a set time window based on standardized transaction data, and to perform sliding statistics and trend analysis to obtain time series feature data. The relationship graph generation unit is used to construct a multi-level interactive relationship graph with the property rights object as the core based on standardized transaction data. The nodes of the graph represent different participants, the edges represent transfer types and time tags, and the unit calculates the correlation between nodes to generate graph association feature data. The feature fusion weighting unit is used to concatenate time series feature data with map-related feature data, and then weight the data by combining the feature importance weights of abnormal cases in the historical case library to obtain fused feature data.

3. The Internet property rights transaction risk assessment and decision-making system based on artificial intelligence according to claim 1, characterized in that, The risk identification module includes: The suspicious behavior chain extraction unit is used to extract continuously occurring suspicious transfer behavior chains based on the fused feature data, using recursive traversal and sequence encoding methods, and generate behavior chain feature data. The profit flow tracing unit is used to trace key nodes step by step based on behavioral chain characteristic data and multi-level interaction relationship graphs, mark suspected profit flows, and generate anomaly markers when transferring across domains or platforms to obtain tracing risk data. The composite risk discrimination unit is used to logically combine behavioral chain feature data and source tracing risk data for discrimination. When multiple high-risk factors are met at the same time, a risk superposition and grading mechanism is executed to generate preliminary risk data.

4. The Internet property rights transaction risk assessment and decision-making system based on artificial intelligence according to claim 1, characterized in that, The intelligent decision-making module includes: The hierarchical adaptive discrimination unit is used to calculate risk scores based on preliminary risk data, adopts an automatic matching processing strategy based on a hierarchical threshold system, and adaptively updates threshold parameters when a new abnormal pattern is detected to generate hierarchical risk results. The case transfer learning unit is used to compare the hierarchical risk results with the historical case library one by one. If a high-confidence similar case exists, the corresponding treatment result is directly output. If no match is found, the current risk characteristics are stored in the case library, and an entry to be labeled is generated for the case library. When externally input labeled data is received, the case library is automatically updated to maintain accuracy and a case comparison result is generated. The assessment summary and output unit is used to summarize and assess the stratified risk results and the case comparison results, generate risk assessment results, and push corresponding early warning information to the management terminal. The feedback loop unit is used to receive the handling feedback from the management terminal and input the handling feedback back to the risk identification module and the intelligent decision-making module respectively, so as to continuously optimize the identification and decision parameters and form a closed-loop self-learning mechanism.

5. The Internet property rights transaction risk assessment and decision-making system based on artificial intelligence according to claim 2, characterized in that, The relation graph generation unit includes: The entity standardization processing subunit is used to perform cross-platform entity identification and deduplication of participants based on standardized transaction data, and generate a unified set of nodes; The temporal and domain labeling subunit is used to write timestamps and regional or platform source identifiers on the edges based on a unified set of nodes and transaction records, and to write role type and risk prior labels in the nodes to form an initial relationship graph with temporal and domain attributes. The correlation calculation subunit is used to statistically analyze the interaction characteristics between nodes within a preset hop limit based on the initial relationship graph, calculate the number of interactions, freshness, role similarity and path credibility, and generate correlation scores for node pairs to form a weighted relationship graph. The graph incremental update subunit is used to incrementally update the weighted relation graph based on newly arrived transaction data, generate a valid graph snapshot that conforms to the sliding time window, and invalidate edges that exceed the window period according to rules, so as to ensure the real-time performance and effectiveness of the graph.

6. The Internet property rights transaction risk assessment and decision-making system based on artificial intelligence according to claim 2, characterized in that, The feature fusion weighting unit includes: The feature weight allocation subunit is used to assign learnable weights to time series feature data and graph-related feature data based on transaction type, transaction amount, data completeness and recent time features, and generate a weighted feature vector. The case prior guidance subunit is used to compare the similarity of the weighted feature vector with the feature clusters of abnormal cases in the historical case library. When the similarity is higher than the threshold, the weight allocation result is biased and adjusted to generate a prior-guided feature vector. The robust normalization processing subunit is used to perform median scaling and extreme value truncation on outliers in the feature vector guided by prior knowledge, so as to obtain normalized robust feature data. The fusion calibration subunit is used to concatenate time series features and spectral correlation features based on the normalized robust feature data to obtain fused feature data.

7. The Internet property rights transaction risk assessment and decision-making system based on artificial intelligence according to claim 3, characterized in that, The suspicious behavior chain extraction unit includes: The transaction sequence segmentation subunit is used to divide continuous transaction events into time sequences based on the fused feature data, and divide them into candidate subsequences according to changes in transfer amount, abnormal transaction frequency, and cross-domain occurrence conditions to obtain candidate behavior chain data; The behavior chain feature encoding subunit is used to progressively encode the node role, transfer direction, transaction freshness and node hop length according to the candidate behavior chain data, and generate a multi-dimensional behavior chain feature vector. The anomaly accumulation determination subunit is used to calculate the risk accumulation value of continuous transactions step by step based on the multi-dimensional behavior chain feature vector. When the risk accumulation value exceeds the dynamic threshold, the corresponding candidate behavior chain is marked as a suspicious behavior chain, and behavior chain feature data is generated.

8. The Internet property rights transaction risk assessment and decision-making system based on artificial intelligence according to claim 3, characterized in that, The benefit flow tracing unit includes: The key node filtering subunit is used to identify nodes that participate in multiple transactions in the relationship graph based on behavioral chain feature data, and to filter them according to the concentration of transaction amount and the weight of node correlation to generate a set of key nodes. The trajectory backtracking construction subunit is used to backtrack along the transaction link step by step based on the set of key nodes, starting from each key node, and record the node role type, interaction frequency and regional label to form the trajectory data of the flow of benefits. The abnormal trajectory labeling subunit is used to generate abnormal labels for abnormal paths based on the profit flow trajectory data. When it is detected that the transfer path crosses different regions, different platforms or has multiple levels of concealed transfer, the abnormal path is used as source tracing risk data.

9. The Internet property rights transaction risk assessment and decision-making system based on artificial intelligence according to claim 4, characterized in that, The hierarchical adaptive discrimination unit includes: The risk score generation subunit is used to calculate and output a risk score sequence based on preliminary risk data, the abnormality of transaction amount, the historical risk labels of participants, and time span factors. The stratified threshold comparison subunit is used to compare risk scores layer by layer according to the risk score sequence, set threshold intervals according to at least three risk levels, and generate stratified discrimination results. The threshold adaptive adjustment subunit is used to adjust the upper and lower limits of the threshold based on the stratification judgment results. When the risk score is detected to continuously deviate from the preset threshold range, the unit writes the updated threshold parameters into the decision parameter library to generate adaptively updated stratified risk results.

10. The Internet property rights transaction risk assessment and decision-making system based on artificial intelligence according to claim 4, characterized in that, The case transfer learning unit includes: The feature similarity comparison subunit is used to extract key feature vectors based on the hierarchical risk results and compare them one by one with the feature set in the historical case library to form a candidate similar case set; The confidence level comprehensive judgment subunit is used to calculate a comprehensive score based on the candidate similar case set, combined with the similarity score and the case handling accuracy. When the comprehensive score exceeds the preset threshold, the corresponding handling result is output and a case judgment result is formed. The case library expansion and update subunit is used to write the current risk characteristic data into the case library and generate entries to be labeled when no high-confidence cases are matched based on the case judgment results. When externally input labeled data is received, the case library is expanded and corrected to generate case comparison results.

Images