Data processing method for dynamically updating knowledge graph

Abstract

The invention relates to the field of electric digital data processing, and discloses a data processing method for dynamically updating a knowledge graph, which comprises the following steps of: presetting a basic weight for a knowledge unit, calculating an activity index of the knowledge unit attenuating along with time, and broadcasting an updating pulse signal to a neighbor unit when the activity index exceeds a threshold value, and the neighbor unit autonomously judges whether to trigger updating or continue to propagate the signal according to the signal attribute. According to the method, system resources are naturally focused on a high-value area through a life-driven self-organizing updating mechanism in the knowledge unit, global scanning redundancy is avoided while the knowledge timeliness is guaranteed, efficient transmission of key updating and reliable awakening of a dormant area are achieved by means of signal priority coupling and a neighborhood activity detection mechanism, and the reliability of the system is improved. And the resource efficiency contradiction in dynamic updating of the knowledge graph is solved.

Description

Technical Field

[0001] The present invention relates to a data processing method for dynamically updating a knowledge graph, and belongs to the technical field of electronic digital data processing. Background Art

[0002] The dynamic update of existing knowledge graphs generally relies on periodic full / incremental recalculation or event / rule-driven updates. The core of such methods lies in centralized scheduling logic: update behavior is forcibly triggered by external events or preset rules, requiring global scanning of the graph or frequent polling of data sources to capture changes. For example, in financial risk control scenarios, equity changes need to be tracked in real time, and the system is forced to scan related subgraphs at high frequency to ensure timeliness. However, most data changes are merely noise, resulting in a large amount of redundant computing resources. Sudden research breakthroughs in the medical field may subvert disease treatment knowledge, but low-frequency update mechanisms are unable to quickly respond to such key changes, resulting in the risk of knowledge delay.

[0003] Although the industry has attempted to balance resources and accuracy through dynamic priority scheduling or complex algorithms, its essence is still limited by the static data equality assumption, which ignores the inherent activity differences of knowledge units (such as timeliness and correlation strength). This leads to two deep contradictions: 1. High-frequency updates ensure timeliness but cause waste of computing power, while low-frequency updates lead to knowledge rigidity. Existing scheduling algorithms can only provide partial relief but cannot cure the problem; 2. Centralized mechanisms cannot independently identify the value density of knowledge units, such as the equity change correlation during the policy effectiveness decay period, making it difficult to focus resources on high-value areas.

[0004] More critically, existing technologies lack the ability to model the endogenous laws of the knowledge lifecycle. They rely on external rules or manually preset thresholds, and are unable to dynamically perceive the attenuation of knowledge association strength, such as historical cases that become invalid due to regulatory iterations. They also have difficulty capturing the implicit coupling of cross-domain data, such as the sudden association between drug side effects and gene mutations. The resulting update redundancy or omissions have become a bottleneck restricting the practicality of large-scale knowledge graphs. Therefore, how to establish a decentralized update mechanism that is endogenous to the activity of knowledge units, so that the system can autonomously perceive value density and accurately trigger local updates, fundamentally resolving the contradiction between resource efficiency and knowledge timeliness, has become the technical problem to be solved by this invention. Summary of the Invention

[0005] The present invention provides a data processing method for dynamically updating knowledge graphs, the main purpose of which is to solve the problem of irreconcilable contradiction between resource efficiency and knowledge timeliness caused by the existing centralized update mechanism ignoring the vitality of knowledge.

[0006] To achieve the above-mentioned purpose, the present invention provides a data processing method for dynamically updating a knowledge graph, the method comprising the following steps: step a, initializing and maintaining the knowledge activity index of each knowledge unit in the knowledge graph, the knowledge activity index being calculated based on the determined basic weight of the knowledge unit and the length of time since it was last touched by external data; wherein, the knowledge activity index is calculated based on the inherent importance of the knowledge unit under the knowledge rules of a specific field, and decreases according to a determined attenuation coefficient over time.

[0007] Step b: When any knowledge unit receives external data input and its knowledge activity index is updated instantaneously based on the external data input, if the updated knowledge activity index is greater than a dynamic adaptive threshold, the knowledge unit acts as a source unit to generate and broadcast a lightweight update pulse signal to all its direct neighbor units in the knowledge graph topology structure; the dynamic adaptive threshold is locally adjusted according to the update activity of the direct neighbor units of the source unit in the past preset time window.

[0008] In step c, after receiving the update pulse signal, the neighbor unit determines whether the preset activation conditions are met based on the knowledge activity index of the source unit carried by the pulse signal and the current knowledge activity index of the neighbor unit itself; if the activation conditions are determined to be met, the neighbor unit's own knowledge update operation is triggered, and the update pulse signal is continued to be broadcast to its own direct neighbor units, thereby forming a self-organizing update chain in the local area of ​​the knowledge graph.

[0009] In step d, the identification information of all knowledge units that trigger the knowledge update operation due to the judgment in step c is added to a queue to be updated.

[0010] In step e, the background processing process continuously consumes the queue to be updated and performs the knowledge graph update task on the knowledge units in the queue and their associated graph structures.

[0011] Preferably, in step a, the basic weight of the knowledge unit is obtained from a lookup table according to the preset type of the knowledge unit; the knowledge activity index The calculation method is: ,in, represents the basic weight, represents the attenuation coefficient determined according to the field to which the knowledge unit belongs, Indicates the length of time since the knowledge unit was last touched by external data.

[0012] Preferably, in step b, the dynamic adaptive threshold The calculation method is: ,in, Indicates the preset basic threshold, Indicates the number of direct neighbor units of the source unit that have been activated in the past preset time window, Indicates the total number of direct neighbor cells of the source cell.

[0013] Preferably, in step c, the update pulse signal contains the identification information of the source unit and the knowledge activity index of the source unit; the autonomous judgment of the neighbor unit includes: comparing the received knowledge activity index of the source unit with the knowledge activity index of the neighbor unit itself, and determining whether to trigger its own knowledge update operation or continue to broadcast the update pulse signal according to the preset comparison rules.

[0014] Preferably, in step c, when the neighboring unit receives multiple update pulse signals, it determines the order of response based on the priority parameter attached to the pulse signal; the priority parameter is calculated based on the knowledge activity index of the source unit sending the pulse signal and the number of direct neighbors of the source unit.

[0015] Preferably, the priority parameter The calculation method is: ,in, represents the knowledge activity index of the source unit that sends the pulse signal, Represents the number of direct neighbor units of the source unit; the priority parameter aims to couple the source unit's own activity with the structural sparsity of the unit in the knowledge graph, thereby dynamically evaluating the propagation urgency and structural importance of different update signals.

[0016] Preferably, for a knowledge unit in the knowledge graph whose knowledge activity index is continuously lower than a dynamic adaptive threshold within a preset time period, the method further includes: periodically calculating the variance of the knowledge activity index of the knowledge unit and all its direct neighboring units.

[0017] Preferably, when the neighborhood activity variance calculated by the knowledge unit is greater than a preset entropy change threshold, the knowledge unit is forcibly triggered to broadcast a wake-up pulse signal to its own direct neighbor units.

[0018] Preferably, the external data input includes one or more of real-time transaction records, news event streams, scientific research paper releases, regulatory policy changes, and sensor data.

[0019] Compared with the existing technology, the beneficial effects of the present invention are: 1. By presetting an activity index (KAI) pre-configured based on domain knowledge for each knowledge unit, and relying on its time decay characteristics to autonomously determine the update trigger timing, the knowledge graph system is transformed from the traditional central scheduling mode to a distributed autonomous response mode. When external data injection triggers a local activity jump, the activation unit only needs to broadcast a lightweight pulse signal to its direct neighbors, triggering a chain response rather than a global traversal. This self-organizing update mechanism driven by the intrinsic activity of the knowledge unit naturally focuses computing resources on high-value areas, eliminating resource redundancy of full-graph scanning while maintaining the timeliness of knowledge, thereby resolving the contradiction between high-frequency update requirements and system load.

[0020] 2. When a node broadcasts an update pulse, it dynamically couples its own activity index with the topological sparsity into a priority parameter, superimposing the differentiated propagation weight of the signal. The receiving nodes then autonomously coordinate the response order based on this priority parameter: high-priority signals are processed immediately, while low-priority signals are re-evaluated with random delays. This simple design, which integrates node activity status with network structure characteristics, enables the system to achieve non-blocking transmission of key updates without central scheduling in scenarios with sudden multi-signal concurrency, thus avoiding the risk of signal congestion in decentralized systems from a fundamental level.

[0021] 3. For long-term low-activity nodes, a lightweight entropy change detector is periodically activated to capture local knowledge structure anomalies by calculating the variance of the activity index between the node and its neighbors. When the variance of the imbalanced neighborhood activity distribution exceeds the threshold, the dormant node is forced to broadcast a wake-up signal. This mechanism, which uses spatial distribution characteristics instead of complex time-series monitoring, enables the system to perceive the indirect impact of the external environment on the implicit knowledge area, making the dormant unit sensitive to the environment and avoiding the hidden danger of knowledge association failure caused by long-term silence. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 This is a timing diagram of the wake-up pulse propagation triggered by the neighborhood activity variance of the sleep unit in the present invention.

[0023] Figure 2 This is a comparative curve chart showing how the knowledge activity index of different types of knowledge units changes over time.

[0024] Figure 3 Schematic diagram of the weight level and association strength structure between the knowledge unit and multi-source data nodes of the present invention.

[0025] Figure 4 This is a graph showing the change of the knowledge activity index (KAI) over time.

[0026] Figure 5 This is a relationship diagram between neighborhood activity variance and wake-up triggering in the present invention.

[0027] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0028] It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0029] An embodiment of the present application provides a data processing method for dynamically updating a knowledge graph, the method comprising the following steps: step a, initializing and maintaining the knowledge activity index of each knowledge unit in the knowledge graph, the knowledge activity index being calculated based on a determined basic weight of the knowledge unit and the length of time since it was last touched by external data; wherein, the knowledge activity index is calculated based on the inherent importance of the knowledge unit under the knowledge rules of a specific domain, and decreases over time according to a determined attenuation coefficient.

[0030] Step b: When any knowledge unit receives external data input and its knowledge activity index is updated instantaneously based on the external data input, if the updated knowledge activity index is greater than a dynamic adaptive threshold, the knowledge unit acts as a source unit to generate and broadcast a lightweight update pulse signal to all its direct neighbor units in the knowledge graph topology structure; the dynamic adaptive threshold is locally adjusted according to the update activity of the direct neighbor units of the source unit in the past preset time window.

[0031] In step c, after receiving the update pulse signal, the neighbor unit determines whether the preset activation conditions are met based on the knowledge activity index of the source unit carried by the pulse signal and the current knowledge activity index of the neighbor unit itself; if the activation conditions are determined to be met, the neighbor unit's own knowledge update operation is triggered, and the update pulse signal is continued to be broadcast to its own direct neighbor units, thereby forming a self-organizing update chain in the local area of ​​the knowledge graph.

[0032] In step d, the identification information of all knowledge units that trigger the knowledge update operation due to the judgment in step c is added to a queue to be updated.

[0033] In step e, the background processing process continuously consumes the queue to be updated and performs the knowledge graph update task on the knowledge units in the queue and their associated graph structures.

[0034] Preferably, in step a, the basic weight of the knowledge unit is obtained from a lookup table according to the preset type of the knowledge unit; the knowledge activity index The calculation method is: ,in, represents the basic weight, represents the attenuation coefficient determined according to the field to which the knowledge unit belongs, Indicates the length of time since the knowledge unit was last touched by external data.

[0035] Preferably, in step b, the dynamic adaptive threshold The calculation method is: ,in, Indicates the preset basic threshold, Indicates the number of direct neighbor units of the source unit that have been activated in the past preset time window, Indicates the total number of direct neighbor cells of the source cell.

[0036] Preferably, in step c, the update pulse signal contains the identification information of the source unit and the knowledge activity index of the source unit; the autonomous judgment of the neighbor unit includes: comparing the received knowledge activity index of the source unit with the knowledge activity index of the neighbor unit itself, and determining whether to trigger its own knowledge update operation or continue to broadcast the update pulse signal according to the preset comparison rules.

[0037] Preferably, in step c, when the neighboring unit receives multiple update pulse signals, it determines the order of response based on the priority parameter attached to the pulse signal; the priority parameter is calculated based on the knowledge activity index of the source unit sending the pulse signal and the number of direct neighbors of the source unit.

[0038] Preferably, the priority parameter The calculation method is: ,in, represents the knowledge activity index of the source unit that sends the pulse signal, Represents the number of direct neighbor units of the source unit; the priority parameter is intended to couple the source unit's own activity with the unit's structural sparsity in the knowledge graph, so as to dynamically evaluate the propagation urgency and structural importance of different update signals; for knowledge units in the knowledge graph whose knowledge activity index is continuously lower than the dynamic adaptive threshold within a preset time period, the method also includes: periodically calculating the variance of the knowledge activity index of the knowledge unit and all its direct neighbor units; when the neighborhood activity variance calculated by the knowledge unit is greater than a preset entropy change threshold, forcibly triggering the knowledge unit to broadcast a wake-up pulse signal to its own direct neighbor units; external data input includes real-time transaction records, news event streams, scientific research paper releases, changes in laws and regulations, and sensor data One or more methods, which realize the transformation of the knowledge graph from a passive response mode to a self-organizing update mode driven by the life within the knowledge unit, thereby reducing the overall consumption of computing resources while ensuring the timeliness of knowledge updates, and improving the system's perception of the importance and timeliness of knowledge. In addition, the data processing process described in the present invention is based on the construction and analysis of structured or semi-structured data streams, and the external data sources involved are all of clear origin, public attributes or authorized content, such as policy releases, scientific research results, industry announcements or desensitized sensor data, etc., to ensure that the knowledge graph update mechanism maintains reasonable constraints on the boundaries of data use while improving efficiency and sensitivity response capabilities. These are all extended implementation methods that can be known to ordinary technicians in this field.

[0039] Example 1: This embodiment provides a data processing method for dynamically updating knowledge graphs. Each knowledge unit is assigned a basic weight during the graph initialization phase. The weight is determined by its preset type and is obtained through a lookup table. The system calculates its current knowledge activity index based on the basic weight and the time interval since the knowledge unit last received external data. The calculation method used is to multiply the basic weight by an exponential function, where the exponential part of the exponential function is the negative product of the time length and a preset attenuation coefficient. The attenuation coefficient is determined by the field category in which the knowledge unit is located, thereby obtaining different knowledge units with reasonable time-sensitive characteristics. During operation, once any knowledge unit receives data input from the outside, the system immediately updates the activity index of the unit and compares it with a dynamic adaptive threshold. When the activity index exceeds the threshold, the unit is regarded as a source unit and begins to broadcast update pulse signals to all direct neighbor nodes in its knowledge graph structure. The dynamic threshold is not a fixed constant, but is calculated based on the proportion of activations of the source unit's neighbor nodes in the past preset time window, that is, a basic threshold multiplied by 1 plus the coefficient of the number of active neighbors divided by the total number of neighbors, so that the threshold has the ability to adapt to the local structural state. When the neighbor node receives the above update, After receiving a new pulse signal, the system does not force it to perform the update operation immediately. Instead, it compares the relationship between the source unit activity index and its own activity index in the signal, and determines whether the activation conditions are met according to the preset rules. If the conditions are met, the knowledge update operation of this unit is executed, and the update pulse is continued to be broadcast to its own direct neighbor units, thereby forming a self-organizing propagation chain in the local area of ​​the knowledge graph. Compared with the whole graph traversal, this method is more computationally efficient and structurally sensitive. In order to further optimize the response management in the update propagation, the present invention designs a response sorting mechanism based on the signal priority parameter. The priority parameter is based on the knowledge activity index of the source unit. The numerator is the number of direct neighbors of the source unit plus one and the logarithm is taken as the denominator to reflect the coupling relationship between the importance of the source unit and the structural sparsity. The receiving node can perform asynchronous sorting processing based on this parameter: give priority to responding to high-priority signals, and implement delayed decision or re-evaluation strategies for low-priority signals, effectively avoiding congestion of system processing resources in emergency situations. The unique identifier of all nodes that trigger update operations due to the above judgments will be recorded in a queue to be updated, and will be continuously consumed and processed by an independent background process. The background update task will combine the graph structure related to the node and perform graph data update calculations on it to ensure that the knowledge state remains synchronized at the semantic and topological levels.

[0040] To address the hidden danger of long-term low-activity units being difficult to respond to environmental changes, this method introduces a lightweight entropy change detection mechanism. The system will periodically calculate the variance of the activity index of a unit whose activity index is always lower than the dynamic threshold in all continuous time periods and that of all neighboring nodes. When the variance exceeds a preset entropy change threshold, the system automatically triggers the unit to broadcast a wake-up pulse signal to its neighbors. This enables the silent unit to perceive external local disturbances without the need for continuous energy consumption and enter an activated state in a timely manner, further improving the overall timeliness and structural resilience of the knowledge graph.

[0041] Example 2: In a dynamic update mechanism driven by the knowledge activity index, during the system initialization phase, a basic weight is configured for each node by querying a preset knowledge unit type weight table. Usually, policy nodes are assigned a higher initial value due to their importance to the global structure of the graph; and industry event nodes are assigned a relatively low value due to their frequent updates. Subsequently, the system calculates the current knowledge activity index based on the timestamp of the last time each knowledge unit was touched by external data, combined with the basic weight and the attenuation coefficient set for the corresponding field. The index is calculated according to the exponential decay formula, that is, the basic weight is multiplied by the attenuation factor that decreases exponentially with time, where the time variable represents the length of time since the last active moment. During operation, when the system receives a new external data stream, such as a new product released by the regional competent department, Industry policy is first structured and mapped to the corresponding policy node in the graph through the entity recognition and relationship extraction module. The node triggers the recalculation of the activity index due to data update. If the index exceeds the dynamic adaptive threshold, which is calculated based on the number of activations of its neighboring nodes within the set time window, an update pulse signal is generated and broadcast to its direct neighboring nodes. The update pulse signal contains the identification information of the source node and the current activity index. After the neighboring node receives the signal, it compares the activity value in the signal with the current activity index of the node and determines whether to trigger the local knowledge update operation according to the preset activation rules. If it is determined that the activation conditions are met, the node immediately loads the external related data, updates its own attributes or graph relationship structure, and also broadcasts the update pulse to its direct neighboring nodes to form a chain propagation path.

[0042] To enhance the system's processing efficiency under multiple concurrent signals, each update pulse signal is accompanied by a priority parameter. This parameter is calculated by dividing the current activity index of the source node by the logarithm of the number of its neighboring nodes. It reflects the signal's importance and sparsity in the structure. The system prioritizes signal processing, responding first to high-priority signals and introducing delayed evaluation for low-priority signals to avoid concurrency congestion and ensure the continuity of the update chain. Furthermore, the system incorporates an entropy change detection mechanism to detect state changes in chronically inactive nodes. This mechanism periodically calculates the variance of the knowledge activity index between the target node and its neighboring nodes. If the entropy change threshold is exceeded, the system triggers the node to actively broadcast a wake-up signal. This wake-up mechanism enables passive detection and local reactivation of regions with potential hidden structural changes without continuous high-frequency polling, enhancing the comprehensiveness and flexibility of graph updates. The system manages all update requests in a queue. All knowledge units triggered by local judgments are added to the pending update queue. Background processing threads continuously consume this queue and sequentially perform operations such as attribute refresh, relationship reconstruction, or edge weight adjustment on the corresponding units in the graph, thereby achieving local dynamic updates of the graph structure.

[0043] Example 3: This example aims to deeply verify a data processing method for dynamically updating knowledge graphs, its operating mechanism in complex and highly timely financial market risk event tracking scenarios and the technical advantages it can achieve. In the field of financial risk monitoring, such as real-time tracking of changes in the equity structure of listed companies and risk transmission paths of related transactions, existing knowledge graph update mechanisms generally face challenges. Traditional methods usually rely on periodic full or incremental recalculations, or updates are forced to be triggered by external rules. Such centralized scheduling strategies are prone to redundant consumption of computing resources when faced with massive, high-frequency, and noisy market data. At the same time, when dealing with sudden, highly correlated events, response lags may occur, and key risk transmission paths may not be captured in a timely manner. This inherent contradiction between resource efficiency and knowledge timeliness is a key factor restricting the practicality of large-scale financial risk knowledge graphs. This experiment specifically constructs a simulated financial market operating environment, which contains a large number of knowledge units (representing enterprises, individuals, transactions, policies, etc.) and their complex relationships, in order to verify how the life-driven self-organizing update mechanism within the knowledge units proposed in this invention can effectively cope with the above challenges in such practices. Specifically, the experiment will focus on quantitatively analyzing the method's ability to intelligently allocate computing resources under different market activities, and how its chain update and wake-up mechanism ensures the rapid and accurate transmission of key risk information and reduces scanning redundancy in non-core areas when sudden risk events occur. This experiment is conducted on a high-performance distributed knowledge graph processing platform that can simulate the storage, interconnection and real-time input of external data streams of large-scale knowledge units. The core processor cluster configured by the platform is oriented to high-concurrency computing environments and can be used to accurately measure resource consumption. Each knowledge unit in the knowledge graph is preset with a basic weight ,This weight is obtained from a preset lookup table based on the inherent importance of the knowledge unit in the financial field, for example, core entities such as large financial institutions or listed companies, Can be set to 1.0; important events, such as major asset restructuring or equity pledge, Can be set to 0.8; related parties, including executives or related companies, It can be set to 0.6; while secondary information, such as general market announcements or historical trading records, It can be set to 0.4. This weight setting is based on the long-term practical experience and consensus in the field of financial risk analysis on the sensitivity of different information elements. Knowledge Activity Index The calculation method is: , where the attenuation coefficient The setting of is crucial, as it reflects the speed at which a specific financial knowledge unit becomes obsolete over time. For knowledge units with strong timeliness, such as changes in financial policies or regulations, The value can be set to 0.05, indicating that its activity decays relatively quickly; while for relatively stable knowledge units such as equity structure or historical financial reports, The value can be set to 0.01, indicating that its activity decays relatively slowly.

[0044] Table 1: Basic weight and attenuation parameter configuration table of knowledge units in the financial field.

[0045]

[0046] At the same time, the attenuation coefficient involved in the dynamic evolution of the above knowledge activity index The specific value is determined based on the technical trade-off between the updating rhythm of knowledge units in practical application fields and the characteristics of information obsolescence. The fundamental purpose of the setting is to achieve an optimal engineering balance between the sensitivity of the knowledge state to changes over time and the stability of the overall computing resource allocation of the system. Specifically, if If the value is too large, the activity index of the knowledge unit will drop rapidly, making the system more sensitive to the freshness of the data, so that it can timely perceive the local disturbance caused by external emergencies. However, too fast decay may also cause the information to be judged as low activity before it is outdated, resulting in frequent misjudgment of long-term effective knowledge and waste of resources. On the contrary, if A value that is too small can avoid frequent system updates, but it may also lead to a slow response to outdated information, resulting in knowledge lag. Therefore, in this embodiment, for policy or news nodes with short response cycles and frequent changes driven by real-time events, the set The value is based on the active change cycle reflected in the historical financial market dynamic data, and its engineering experience value is set to 0.05; for equity relations or historical financial report nodes with stable structure and slow evolution, a relatively slow decay rate is used, and the corresponding The value is set to 0.01; the above The determination of the value is based on the statistical analysis of the characteristics of historical financial market data and the risk transmission cycle, and the typical empirical value selected after the engineering trade-off between the timeliness of knowledge and the computing cost. This setting is intended to make the update frequency of the knowledge graph match the actual life cycle of financial information. The dynamic adaptive threshold The calculation method is: , where the basic threshold It can be set to 0.3, which is an empirical value determined based on the typical distribution characteristics of the knowledge unit activity index and after multiple rounds of preliminary experiments, after balancing the update sensitivity and the system stable load. Indicates the number of direct neighboring units of the source unit that have been activated in the past preset time window (set to 24 hours in this experiment); Represents the total number of direct neighboring units of the source unit. The engineering consideration of this dynamic adjustment mechanism is that when the activity of the source unit's neighbors is low, the update threshold is relatively low, which is conducive to the activation of dormant areas. When the neighbor activity is high, the threshold will be increased accordingly to avoid overactivation, thereby achieving self-regulation of the load in the local area and effectively suppressing the excessive diffusion of update pulses, ensuring that resources are concentrated in highly active and high-value update areas.

[0047] The update pulse signal contains the identification information of the source unit that sends the pulse signal and the knowledge activity index of the source unit. After receiving the update pulse signal, the neighboring unit makes a judgment based on the following preset activation conditions: Activation condition 1: The knowledge activity index of the received source unit is greater than 1.2 times the current knowledge activity index of the neighboring unit itself. This condition is intended to give priority to responding to update signals from sources with higher activity or greater importance to ensure the priority dissemination of key information. Activation condition 2: The current knowledge activity index of the neighboring unit itself is lower than the preset minimum activation threshold (set to 0.2 in this experiment), and the knowledge activity index of the received source unit is greater than the threshold. This condition is intended to awaken long-term Dormant knowledge units in a low-activity state are activated to prevent missing potentially important information associated with them due to long-term silence. The third activation condition is that the priority parameter of the received pulse signal is higher than the lowest priority acceptable to the neighboring unit itself (set to 0.5 in this experiment), and the update queue load rate of the neighboring unit is less than 80%. This condition is intended to ensure that in the face of high-priority signals, even if the unit's own activity is acceptable, it should be responded to first; at the same time, taking into account its own processing capabilities to avoid system congestion due to overload. The core is: while ensuring the rapid and efficient dissemination of important information, avoid excessive processing of low-value information, and achieve refined control of update diffusion through multi-dimensional criteria. The priority parameter The calculation method is: ,in, represents the knowledge activity index of the source unit that sends the pulse signal, This parameter, representing the number of direct neighboring units of a source unit, aims to couple the activity of the source unit with the structural sparsity of the unit in the knowledge graph, thereby dynamically evaluating the propagation urgency and structural importance of different update signals. For example, a unit with a high activity index but a relatively sparse number of neighbors, that is, a unit located at the edge or in an isolated area of ​​the knowledge graph, will have its update signal given a relatively high priority, as it may represent critical, emergent information that cannot be ignored and requires rapid attention and dissemination. Conversely, a unit with a medium activity index but a large number of neighbors will have a relatively low pulse priority, as its information may have been widely disseminated or the influence of its update has been initially diffused. This design effectively avoids priority bias caused by judging a single activity metric and improves the system's ability to perceive critical updates in complex network structures. For knowledge units in the knowledge graph whose knowledge activity index remains below the dynamic adaptive threshold for a preset time period (set to 72 consecutive hours in this experiment), the method also includes periodically calculating the variance of the knowledge activity index of the knowledge unit and all its direct neighboring units (called the neighborhood activity variance).

[0048] When the neighborhood activity variance calculated by a knowledge unit is greater than a preset entropy change threshold (set to 0.1 in this experiment), the knowledge unit is forced to broadcast a wake-up pulse signal to its own direct neighbor units. The determination of this entropy change threshold is based on a statistical analysis of knowledge units in the financial field that have been dormant for a long time but suddenly become associated. It was found that their neighborhood activity distribution usually shows a significant imbalance, that is, a few neighbors are activated by external events while the knowledge unit itself is still dormant. This threshold has been tuned to accurately capture this local structural anomaly to avoid excessive awakening and ensure the accuracy and low energy consumption of the awakening mechanism. This mechanism can awaken related knowledge through indirect associations outside the core risk event. For example, when an associated enterprise suddenly becomes active due to policy changes, even if the data of the dormant enterprise it invested in has not been updated, it can be awakened through the neighborhood activity variance, thereby assessing potential chain risks.

[0049] This experiment simulated a 30-day financial market data flow, during which external data inputs included normal market fluctuations, financial report updates of medium-sized enterprises, and several sudden major financial risk events, such as equity freezes of listed companies, investigations involving senior executives, and major changes in laws and policies. During the simulation process, three types of representative knowledge units were selected for observation: Class A units simulated a large listed financial institution with a high basic weight and encountered sudden major negative news on the fifth day of the experiment; Class B units simulated a local small and medium-sized enterprise that had been in a low-activity state for a long time with a low basic weight and received no external data updates in the 20 days before the experiment; and Class C units simulated a regular transaction record unit with medium activity fluctuations.

[0050] Table 2: Evolution of activity index and update-related indicators of typical knowledge units during the observation period.

[0051]

[0052] The knowledge activity decay and instantaneous update characteristics have been verified. When the three types of knowledge units A, B, and C are not touched by external data, their knowledge activity indexes all show a natural decay trend, which is consistent with the The rules reflected by the model are consistent. On the fifth day of the experiment, after receiving external data, the knowledge activity index of Class A unit instantly increased significantly to 0.95, and this value is higher than its dynamic adaptive threshold, indicating that its update mechanism is sensitive to response. After Class C unit received the update pulse from Class A unit on the sixth day, the knowledge activity index also instantly increased to 0.65, proving the effective transmission of the update pulse.

[0053] The dynamic adaptive threshold adjustment function was demonstrated. The dynamic threshold of a Class A unit increased with the increase in its neighboring update activity (from 0.315 to 0.435), effectively suppressing the potential for update flooding during the initial outbreak of a core event and ensuring the system's refined control over updates in hotspot areas. Once updates in hotspot areas stabilized, the threshold dropped as activity decreased, demonstrating its adaptive nature. The effectiveness of the dormant wakeup mechanism was also verified. After a long period of low activity, a Class B unit's neighborhood activity variance reached 0.15 on the 25th day of the experiment, exceeding the preset entropy change threshold of 0.1. This successfully triggered a wakeup pulse and led to an instantaneous increase in its knowledge activity index. This demonstrates that the neighborhood activity variance detection mechanism can effectively awaken dormant areas, resolving the problem of long-dormant knowledge units being unable to respond to changes in their indirect connections. Update pulse priority scheduling and propagation paths: In a simulation of a sudden financial risk event involving a company facing acquisition rumors, knowledge units for multiple related companies and transactions simultaneously received external data input, resulting in multiple concurrent update pulses. The experiment tracked the response and propagation of pulses of different priorities.

[0054] Table 3: Comparison of priority parameters and propagation response characteristics for different source unit types.

[0055]

[0056] The effectiveness of the priority parameter has been verified. As the knowledge unit of the core M&A target, it obtained a higher priority parameter of 0.86 due to its higher knowledge activity index and moderate number of neighbors. Its update pulse received the highest priority immediate processing at the receiving neighbor and quickly propagated 5 layers deep, which ensured that the core risk information could be quickly spread to key related parties. The coupling effect of knowledge activity and structural sparsity was reflected. Although the knowledge activity index of the secondary business partner was lower than that of the related investment institution, due to its smaller number of neighbors and relatively higher structural sparsity, its priority parameter of 0.65 was higher than that of the related investment institution of 0.38. Therefore, its update pulse obtained a faster response and a deeper propagation depth, which shows that the priority parameter The calculation method can ensure that even units that are not absolutely highly active can receive priority attention if they are in a critical but sparse conduction path in the knowledge graph. This mechanism has significant benefits in resource scheduling. It avoids indiscriminate processing of all concurrent signals. High-priority signals are processed immediately, while low-priority signals may introduce moderate delays depending on the load of the receiving node. This ensures the timeliness of key information while achieving intelligent scheduling and optimized allocation of computing resources.

[0057] Resource consumption and redundancy elimination effect evaluation,To quantify the resource consumption and redundancy elimination effects,,this experiment compared the average CPU utilization and update response time,of the method of the present invention with the traditional,centralized full-volume scanning method when processing the same amount of data and the,same risk events.

[0058] Table 4: Comparison of update performance indicators between the method of the present invention and traditional full scan.

[0059]

[0060] Significant savings in computing resources have been confirmed. The average CPU utilization of the method of the present invention is 32%, which is much lower than the 85% of the traditional full-scan method. This directly proves that through the self-organizing update mechanism driven by the vitality within the knowledge unit, system resources can naturally focus on high-value areas, significantly reducing the computational redundancy caused by global scanning, and significantly improving the response speed of key information. In terms of the response time for key event updates, the method of the present invention only takes 3 seconds, while the traditional method takes 15 seconds. This is due to the self-organizing update chain and priority scheduling mechanism, which ensures the efficient delivery of key updates. The effective elimination of redundancy in non-core areas has been verified. The scanning volume of the method of the present invention in non-core areas is only 0.15 of that of the traditional method, which means that a large number of knowledge units that are not related to the current update event are no longer frequently and unnecessarily traversed and processed, fundamentally solving the resource efficiency contradiction in the dynamic update of the knowledge graph.

[0061] Example 4: This example combines Figures 1 to 5 , a data processing method for dynamically updating knowledge graph is described. Figure 1 As shown, first, the timer starts to trigger the periodic detection operation, and the sleeping unit B then sends the acquisition information to all its direct neighbor units (ie, neighbor unit 1, neighbor unit 2, neighbor unit 3 and neighbor unit 4). Value request and receive the returned activity index in turn , where neighbor unit 1 returns , neighbor unit 2 returns , neighbor unit 3 returns , neighbor unit 4 returns ; The sleeping unit B calculates the 9 activity variances of the knowledge activity index returned by the above four neighboring nodes, and obtains the variance = 0.15, and determines whether the value is greater than the preset entropy change threshold 0.1. Since the variance = 0.15> the entropy change threshold 0.1, the trigger condition is met, and the alt branch structure is entered to start the wake-up operation. The sleeping unit B first broadcasts the broadcast wake-up pulse signal to the neighbor unit 1, and then broadcasts the broadcast wake-up pulse signal to the neighbor unit 2, neighbor unit 3 and neighbor unit 4 in turn. After broadcasting the wake-up signal, the sleeping unit B will broadcast its own The value is updated to 0.42, indicating that it has entered the state where it can be activated.

[0062] like Figure 2 As shown in the figure, there are three types of nodes: type A units (solid lines, circle marks), type B units (dashed lines, square marks) and type C units (dotted lines, triangle marks). Value trajectory, type A units are represented by solid lines, data points are represented by circular marks, and their initial The value is close to 1.0 and gradually decreases over time, reflecting that its activity index decays naturally over a long period of time; Class B units are represented by dotted lines and data points are represented by square marks. The overall value remained at a low level, and there was a significant fluctuation on the 25th day, corresponding to the index change before and after the activation of the awakening mechanism; Class C units are represented by dotted lines, and data points are represented by triangle marks. The value fluctuates at a moderate level, and overall shows a slow decline but maintains a certain stability.

[0063] like Figure 3As shown in the figure, the knowledge unit is located in the center of the figure, represented by a black solid circle and marked as a high-weight node. It is connected to the four medium-weight nodes of policy node, enterprise node, event node and transaction node through solid lines. These four nodes are represented by gray diagonal filling and marked as medium-weight nodes according to the legend. Each medium-weight node is connected to a low-weight node through a dotted line weak association, in order: policy node connects to regulations, enterprise node connects to shareholders, event node connects to news, and transaction node connects to the market. The above low-weight nodes are all hollow circles and marked as low-weight nodes. In addition, the announcement node is also connected to the policy node through a weakly associated dotted line, reflecting how the multi-dimensional information composed of data sources such as regulations, announcements, news, shareholders, and markets in the knowledge graph forms a hierarchical and differentiated transmission path structure around the core knowledge unit. And Figure 4 Shows the decay and update of the activity index of different types of knowledge units during the 30-day observation period. Figure 5 Show the variance changes of neighborhood activity of dormant units and the timing of awakening triggers.

[0064] Example 5: During the operation of the knowledge graph system, when a knowledge unit meets the preset activation conditions, the system enters the update operation stage. Specifically, this stage includes three types of update subtasks: one is to synchronize the attribute information corresponding to the unit with the latest external data received. The attributes involved may include node labels, semantic types or identification fields related to event time; the second is to recalculate the edge weights between the unit and the nodes directly connected to it. The basis for weight adjustment includes the current activity index difference between the unit and the associated nodes, the historical interaction frequency and the similarity between the semantic categories to which the nodes belong; the third is to process new or invalid semantic relationships between entities mentioned in external data. If the corresponding connection has not been established in the graph, a new edge relationship is added. If there is a connection but it has expired, the edge is cleared or the status is marked; the above three types of operations are executed in sequence in the background processing thread in the unit-level order. During the execution process, the system maintains the consistency of the topological structure and semantic data to avoid conflicts.

[0065] Regarding the basic threshold setting used in the dynamic adaptive threshold, the system uses the historical distribution mean of the activity index of typical knowledge units in the field as the initial baseline during the initialization phase, and smoothly corrects it in combination with the fluctuation range of the update frequency. The fine-tuning part of the threshold changes dynamically according to the update status of the current node's direct neighbors, so that when the overall activity of the neighbors is low, the node is more likely to be activated; conversely, when the neighbors are generally active, the system automatically raises its threshold to avoid redundant activation, thereby maintaining the compactness of the update chain and the balance of resource consumption; in terms of the wake-up mechanism, when the activity index of a knowledge unit continues to be lower than the dynamic adaptive threshold within a continuous preset period, the system starts the entropy change detection process, and the detection method is periodic counting. The sample variance of the current activity index of the unit and all its direct neighbors is calculated and compared with a preset entropy change threshold. The variance reflects whether the activity distribution in the area where the unit is located is balanced. If the current value exceeds the threshold, the system immediately triggers the unit to broadcast a wake-up pulse signal to all direct neighbor nodes. The wake-up behavior does not directly lead to an update, but instead puts the unit into a latent activation state. In this state, the system sets an observation window period. During this period, if the unit receives new external data input or an update pulse broadcast by a neighboring node again, it is regarded as an activation event. The system will put it into the queue to be updated and perform subsequent update operations. If the above situation does not occur before the end of the observation period, the unit state remains dormant and does not enter the update process.

[0066] To limit the information redundancy and resource usage that may occur during the update propagation process, the system calculates the adjacency density index of the unit each time it generates an update pulse signal. This index is measured as the inverse of the ratio of the number of direct neighbors to the number of secondary neighbors, and is used to approximately characterize the degree of clustering of the node in the graph structure. The system dynamically sets the maximum propagation depth of the signal based on this index. For example, in areas with relatively sparse structures, the allowed propagation depth is relatively high to ensure signal coverage in edge areas; in areas with dense structures, the system controls the propagation level at a low level to prevent the disorderly spread of update paths.

[0067] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0068] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not limiting. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A data processing method for dynamically updating a knowledge graph, characterized in that: The method includes the following steps: step a, initializing and maintaining the knowledge activity index of each knowledge unit in the knowledge graph, the knowledge activity index is calculated based on the determined basic weight of the knowledge unit and the length of time since it was last touched by external data; wherein, the knowledge activity index is calculated based on the inherent importance of the knowledge unit under the knowledge rules of a specific domain, and decreases according to a determined attenuation coefficient over time; step b, when any knowledge unit receives external data input and causes its knowledge activity index to be instantaneously updated based on the external data input, if the updated knowledge activity index is greater than a dynamic adaptive threshold, the knowledge unit is used as a source unit to generate and broadcast a lightweight update pulse signal to all its direct neighbor units in the knowledge graph topology structure; the dynamic adaptive threshold is based on the direct neighbor units of the source unit. The update activity of the neighbor unit in the past preset time window is locally adjusted; in step c, after receiving the update pulse signal, the neighbor unit determines whether the preset activation condition is met based on the knowledge activity index of the source unit carried by the pulse signal and the current knowledge activity index of the neighbor unit itself; if it is determined that the activation condition is met, the neighbor unit's own knowledge update operation is triggered, and the update pulse signal is continued to be broadcast to its own direct neighbor unit, thereby forming a self-organizing update chain in the local area of ​​the knowledge graph; in step d, the identification information of all knowledge units that trigger the knowledge update operation due to the judgment in step c is added to a queue to be updated; in step e, the background processing process continues to consume the queue to be updated, and performs the knowledge graph update task on the knowledge units in the queue and their associated graph structures.

2. A data processing method for dynamically updating a knowledge graph according to claim 1, characterized in that: In step a, the basic weight of the knowledge unit is obtained from a lookup table according to the preset type of the knowledge unit; the knowledge activity index The calculation method is: ,in, represents the basic weight, represents the attenuation coefficient determined according to the field to which the knowledge unit belongs, Represents the length of time since the knowledge unit was last touched by external data.

3. A data processing method for dynamically updating a knowledge graph according to claim 1, characterized in that: In step b, the dynamic adaptive threshold The calculation method is: ,in, Indicates the preset basic threshold, Indicates the number of direct neighbor units of the source unit that have been activated in the past preset time window, Indicates the total number of direct neighbor cells of the source cell.

4. A data processing method for dynamically updating a knowledge graph according to claim 1, characterized in that: In step c, the update pulse signal includes identification information of the source unit and the knowledge activity index of the source unit; The autonomous judgment of the neighbor unit includes: comparing the knowledge activity index of the received source unit with the knowledge activity index of the neighbor unit itself, and determining whether to trigger its own knowledge update operation or continue to broadcast the update pulse signal according to the preset comparison rules.

5. A data processing method for dynamically updating a knowledge graph according to claim 1, characterized in that: In step c, when a neighbor unit receives multiple update pulse signals, it determines the order of response based on the priority parameter attached to the pulse signal; the priority parameter is calculated based on the knowledge activity index of the source unit sending the pulse signal and the number of direct neighbors of the source unit.

6. A data processing method for dynamically updating a knowledge graph according to claim 5, characterized in that: Priority parameter The calculation method is: ,in, represents the knowledge activity index of the source unit that sends the pulse signal, Represents the number of direct neighbor units of the source unit; the priority parameter aims to couple the source unit's own activity with the structural sparsity of the unit in the knowledge graph, thereby dynamically evaluating the propagation urgency and structural importance of different update signals.

7. A data processing method for dynamically updating a knowledge graph according to claim 1, characterized in that: For a knowledge unit in the knowledge graph whose knowledge activity index is continuously lower than a dynamic adaptive threshold within a preset time period, the method further includes: periodically calculating the variance of the knowledge activity index of the knowledge unit and all its direct neighboring units.

8. A data processing method for dynamically updating a knowledge graph according to claim 7, characterized in that: When the neighborhood activity variance calculated by the knowledge unit is greater than a preset entropy change threshold, the knowledge unit is forcibly triggered to broadcast a wake-up pulse signal to its own direct neighbor units.

9. A data processing method for dynamically updating a knowledge graph according to claim 1, characterized in that: External data input includes one or more of real-time transaction records, news event streams, scientific research paper releases, regulatory policy changes, and sensor data.

Images