Method for constructing domain expert intuition pattern based on interaction graph evolution tracking

Abstract

This invention relates to a method for constructing intuitive patterns for domain experts based on interaction graph evolution tracking. The method includes real-time collection of operations performed by domain experts on a domain knowledge graph when solving domain-specific problems, and atomizing each operation into predefined types of interactive behaviors, forming a timestamped sequence of interactive behaviors. It also involves mining sub-sequences of behaviors associated with positive problem-solving outcomes as candidate intuitive response patterns; encapsulating structured intuitive pattern units; and providing guidance to the user when the current knowledge graph state matches the triggering condition of an intuitive pattern unit, and dynamically adjusting the intuitive pattern unit based on user feedback. The beneficial effect of this invention is that, by atomizing the collection of interaction sequences between experts and the knowledge graph, it achieves for the first time a refined tracking of the dynamic cognitive trajectory of experts.

Description

Technical Field

[0001] This invention belongs to the field of information technology in building engineering, and particularly relates to a method for constructing domain expert intuitive patterns based on interactive graph evolution tracing. Background Technology

[0002] In many professional fields, domain experts (such as senior doctors, experienced engineers, and senior analysts) often demonstrate an "intuition" or "rapid, experiential cognitive" ability that transcends linear logical reasoning when solving problems. This ability manifests as the ability to quickly identify key information nodes when faced with a specific problem situation and follow an efficient, non-explicit cognitive path for exploration, connection, and decision-making. This "intuitive mode" is the culmination of experts' long-term practical experience and has extremely high value. However, how to externalize, structure, and transform this implicit, dynamic expert intuition into computable and reusable digital assets is a core challenge currently facing knowledge engineering and intelligent auxiliary systems.

[0003] Currently, technical solutions for knowledge management and expert experience transfer mainly focus on static knowledge base construction, case-based reasoning, and rule-based expert systems. However, significant bottlenecks exist in capturing and utilizing dynamic intuitive cognitive processes: 1. Black-box interaction process and lack of pattern extraction: Existing systems can record the final decision-making results or operational conclusions of experts (such as diagnostic reports and review opinions), but they cannot track the specific interaction sequences between experts and knowledge graphs (or information systems) throughout the problem-solving process in an atomic manner. How experts browse information, establish connections between nodes, and in what order propose and verify hypotheses—this dynamic "cognitive trajectory"—is not effectively captured. Therefore, the system cannot extract the original material of those recurring and effective cognitive subsequences, i.e., "intuitive patterns," from behavioral data. Essentially, this stems from a lack of refined collection and analysis capabilities for experts' dynamic cognitive behavioral sequences.

[0004] 2. Rigid pattern representation, lacking contextual relevance and evolution: Even when some systems attempt to summarize expert experience, it often exists in the form of static "rules," "checklists," or "templates." These representations fail to capture the essence of intuitive patterns: they are often tightly bound to specific knowledge graph state contexts (such as the appearance of node combinations with a particular attribute) and consist of a series of ordered interactive actions. Static rules cannot flexibly match dynamically changing graph states, nor can they reflect subtle differences in the order of operations. Furthermore, while expert intuition evolves with the absorption of new cases and knowledge, existing static knowledge bodies lack a self-evolving mechanism that can dynamically adjust their intensity and generate new variants based on feedback and data.

[0005] 3. Fragmentation of Pattern Validation and Optimization Loop: Are the constructed patterns universally effective? How should they be adjusted in different scenarios? Existing methods typically rely on one-time manual summaries or limited case validation, lacking a continuous, data-driven validation and optimization loop. When a system attempts to apply a pattern to guide new users, it cannot quantitatively evaluate the pattern's utility and dynamically adjust its confidence level or triggering conditions based on real-time feedback such as actual user adoption and task execution results. This makes the constructed pattern library prone to becoming outdated or detached from reality, failing to achieve the goal of "becoming smarter with use."

[0006] 4. Isolated pattern storage, lacking networked synergy: Multiple intuitive pattern units that have been constructed are often stored and managed in isolation. However, rich semantic relationships such as generalization, specialization, and combination may exist between patterns. For example, a "preliminary heart murmur screening pattern" may be a specialization of a more general "abnormal sign screening pattern." Current technology lacks the ability to mine and utilize these relationships between patterns. When a pattern proves to be highly effective (or ineffective), this credibility cannot be intelligently transferred to other similar patterns, limiting the overall learning and reasoning efficiency of the knowledge base.

[0007] Therefore, there is an urgent need in this field for an innovative technical solution that can overcome the limitations of existing methods in capturing expert intuition, which are "difficult to capture, dead in representation, weak in verification, and isolated". Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for constructing domain expert intuitive patterns based on interactive graph evolution tracing.

[0009] Firstly, a method for constructing domain expert intuitive patterns based on interaction graph evolution tracing is provided, including: S100: Real-time collection of operations performed by domain experts on the domain knowledge graph when solving domain-specific problems, and atomizes each operation into predefined types of interactive behaviors, forming a time-stamped sequence of interactive behaviors; S200: Based on the interaction behavior sequence of multiple problem-solving sessions, extract the behavioral sub-sequences associated with positive problem-solving effects as candidate intuitive response patterns; S300. Associate the candidate intuitive response patterns with specific knowledge graph state triggering conditions and encapsulate them into structured intuitive pattern units. S400: When it is detected that the current knowledge graph state matches the triggering condition of an intuitive mode unit, guidance is provided to the user, and the intuitive mode unit is dynamically adjusted based on user feedback.

[0010] Preferably, in S100, the predefined types of interactive behaviors include: focusing operation, association operation, questioning operation, and confirmation operation; wherein, focusing operation is used to represent the concentration of attention on a specific graph node, association operation is used to represent the establishment or adjustment of relationships between nodes, questioning operation is used to represent the user raising questions about the authenticity, reliability, or relevance of a node or relationship in the graph, and confirmation operation is used to represent the user verifying or confirming a node or relationship through external evidence, logical reasoning, or authoritative sources.

[0011] As a preferred embodiment, S200 includes: S210. Time-align the interaction behavior sequences of multiple problem-solving sessions and remove noisy operations with operation durations below a preset threshold. S220. Calculate sequence similarity based on at least two dimensions of operation type sequence, target node semantic similarity, and operation time interval distribution, and perform clustering using a hierarchical clustering algorithm; S230. Within each cluster, use a sequence pattern mining algorithm to extract high-frequency behavioral subsequences, and calculate the occurrence frequency, success correlation, and expert consensus index of each subsequence. Combined with the problem-solving effect evaluation, candidate intuitive response patterns are selected.

[0012] Preferably, in S300, the triggering condition is defined by a logical expression based on knowledge graph subgraph matching, wherein the logical expression needs to specify at least one of the following conditions: necessary node set, prohibited node state, node attribute constraint, and relation topology requirement.

[0013] Preferably, in S300, the structured intuitive pattern unit is stored in a machine-readable format, and its content includes pattern identifier, trigger condition logic, behavior sequence array, expected impact description and metadata information; wherein the behavior sequence array records the operation type, target node and operation parameters, and the expected impact description records the expected changes to the weight of knowledge graph nodes and the new relationships.

[0014] Preferably, in S400, the guidance is implemented using a progressive non-intrusive interface. The guidance intensity is increased in stages according to the matching degree between the current map state and the triggering conditions. The guidance forms include visual highlighting, text prompts, and providing a one-click execution option.

[0015] As a preferred option, it also includes: S500: Periodically calculate a health score for each intuitive mode unit. The health score is calculated by combining recent usage frequency, user feedback rating, intensity value and time decay factors. When the health score is lower than the threshold, the mode review process is triggered. S600: Based on intuitive pattern units with health scores higher than a threshold, automatically generate their derivative patterns. The derivative patterns include a parent pattern generated by relaxing constraints in the triggering conditions, a child pattern generated by adding constraints in the triggering conditions, and a combined pattern generated by merging the core behavior sequences of multiple intuitive pattern units. S700. Construct a similarity network among intuitive pattern units. Calculate edge weights based on trigger condition overlap, behavior sequence edit distance, and application scenario overlap. When an intuitive pattern unit is verified as valid, its credibility gain propagates to neighboring nodes according to the similarity weight; when an intuitive pattern unit is disproven, its credibility loss propagates to neighboring nodes according to the similarity weight.

[0016] Secondly, a domain expert intuitive pattern construction system based on interaction graph evolution tracing is provided for performing any of the methods described in the first aspect, including: The sequence acquisition module is used to collect in real time the operations performed by domain experts on the domain knowledge graph when solving domain-specific problems, and to atomize each operation into a predefined type of interactive behavior, forming a time-stamped sequence of interactive behaviors; The pattern discovery module is used to mine behavioral subsequences that are associated with positive problem-solving effects based on the interaction behavior sequences of multiple problem-solving sessions, as candidate intuitive response patterns; The pattern encapsulation module is used to associate the candidate intuitive response patterns with specific knowledge graph state triggering conditions and encapsulate them into structured intuitive pattern units. The pattern verification and adjustment module is used to provide guidance to the user when the current knowledge graph state is detected to match the triggering condition of a certain intuitive pattern unit, and to dynamically adjust the intuitive pattern unit based on user feedback.

[0017] Thirdly, a computer storage medium is provided, wherein a computer program is stored therein; when the computer program is run on a computer, the computer causes the computer to perform any of the methods described in the first aspect.

[0018] Fourthly, an electronic device is provided, comprising: Memory, used to store computer programs; A processor for executing the computer program to implement the method as described in any of the first aspects.

[0019] The beneficial effects of this invention are as follows: Firstly, by atomizing the interaction sequences between experts and knowledge graphs, this invention achieves, for the first time, refined tracking of the dynamic cognitive trajectory of experts; secondly, by mining high-frequency effective subsequences from multi-conversation behavioral data, it automatically discovers implicit intuitive patterns from the data; thirdly, by binding patterns with the graph context and encapsulating them into structured units, it transforms static experience into "living" knowledge assets that can be understood, scheduled, and executed by the system; and fourthly, by establishing a dynamic adjustment and verification closed loop based on feedback from practical applications, it ensures the continuous evolution and practicality of the pattern library. Ultimately, this method effectively solves the fundamental problem of the difficulty in capturing, representing, verifying, and reusing expert intuition, laying a key technological foundation for building truly "expert-level" cognitive assistance systems. Attached Figure Description

[0020] Figure 1 The overall flowchart of the domain expert intuitive pattern construction method based on interactive graph evolution tracing provided by the present invention is shown below. Figure 2 A schematic diagram of the architecture of the domain expert intuitive mode construction system provided by the present invention; Figure 3 A schematic diagram of the network propagation of intuitive pattern credibility provided by the present invention. Detailed Implementation

[0021] The present invention will be further described below with reference to embodiments. The description of the embodiments below is only for the purpose of helping to understand the present invention. It should be noted that those skilled in the art can make several modifications to the present invention without departing from the principle of the present invention, and these improvements and modifications also fall within the protection scope of the claims of the present invention.

[0022] Example 1: To address the problems in the prior art, Embodiment 1 of this application provides a method that can track the evolution of the interaction graph throughout the process, mine high-frequency effective patterns from the behavioral sequence, and encapsulate them into structured units that are bound to the context and have self-evolution capabilities. This enables the explicit, assetized, and intelligent reuse of domain expert intuition, thereby significantly improving the efficiency and level of professional training, intelligent decision support, and collaborative problem solving.

[0023] Specifically, such as Figure 1 As shown, the domain expert intuitive pattern construction method based on interaction graph evolution tracing provided in this application includes: S100 collects in real time the operations performed by domain experts on the domain knowledge graph when solving domain-specific problems, and atomizes each operation into a predefined type of interactive behavior, forming a time-stamped sequence of interactive behaviors.

[0024] In S100, the predefined types of interactive behaviors include: focusing operation, association operation, questioning operation, and confirmation operation; wherein, focusing operation is used to characterize the concentration of attention on a specific graph node, and association operation is used to characterize the establishment or adjustment of relationships between nodes.

[0025] S200: Based on the interaction behavior sequences of multiple problem-solving sessions, extract behavioral sub-sequences that are associated with positive problem-solving effects as candidate intuitive response patterns.

[0026] In S200, the interaction behavior sequences of multiple problem-solving sessions are clustered, and high-frequency behavior subsequences are mined in each cluster. Based on the positive correlation between the behavior subsequences and the problem-solving effect, candidate intuitive response patterns are selected. Specifically, S200 includes: S210. Time-align the interaction behavior sequences of multiple problem-solving sessions and remove noisy operations with operation durations below a preset threshold. S220. Calculate sequence similarity based on at least two dimensions: operation type sequence, target node semantic similarity, and operation time interval distribution, and perform clustering using a hierarchical clustering algorithm. The operation type sequence refers to a series of predefined interactive behaviors executed by the user in a single interactive session in chronological order, such as [focus, associate, question, confirm]. The target node semantic similarity refers to the semantic similarity between the knowledge graph nodes affected by corresponding operations in two interaction sequences, obtained by calculating cosine similarity after inputting the node text description into a pre-trained language model to obtain embedding vectors. The operation time interval distribution refers to the statistical distribution formed by the time differences between adjacent operations in the same interaction sequence, used to characterize the user's operation rhythm.

[0027] For example, sequence similarity calculation takes into account both the edit distance of the operation type sequence and the cosine similarity of the semantic embedding vector of the target node.

[0028] S230. Within each cluster, use a sequence pattern mining algorithm to extract high-frequency behavioral subsequences, and calculate the occurrence frequency, success correlation, and expert consensus index of each subsequence. Combined with the problem-solving effect evaluation, candidate intuitive response patterns are selected.

[0029] S300. Associate the candidate intuitive response patterns with specific knowledge graph state triggering conditions and encapsulate them into structured intuitive pattern units.

[0030] In S300, the triggering condition is defined by a logical expression based on knowledge graph subgraph matching. The logical expression needs to specify at least one of the following conditions: necessary node set, prohibited node state, node attribute constraint, and relation topology requirement. The necessary node set refers to the set of knowledge graph nodes that must exist to trigger a certain intuitive pattern. The prohibited node state refers to the node and its state combination that must not appear in the current graph context. The node attribute constraint refers to the restriction conditions on the specific attribute values ​​of the graph node, including numerical range, enumerated value, or existence. The relation topology requirement refers to the graph pattern constraint on the connection structure between nodes, including path length, relation type sequence, or subgraph isomorphism condition.

[0031] The triggering condition logic is defined using expressions based on graph query languages ​​(such as Cypher and SPARQL), which can flexibly specify the necessary node types, attribute constraints, relationship paths, and prohibited subgraph structures.

[0032] The structured intuitive pattern unit is stored in a machine-readable format, which facilitates semantic representation, exchange and integration. Its content includes pattern identifier, trigger condition logic, behavior sequence array, expected impact description and metadata information. The behavior sequence array records the operation type, target node and operation parameters, and the expected impact description records the expected changes to the weight of knowledge graph nodes and the new relationships.

[0033] S400: When it is detected that the current knowledge graph state matches the triggering condition of an intuitive mode unit, guidance is provided to the user, and the intuitive mode unit is dynamically adjusted based on user feedback.

[0034] In S400, the guidance is implemented using a progressive non-intrusive interface. The guidance intensity is increased in stages according to the matching degree between the current map state and the triggering conditions. The guidance forms include visual highlighting, text prompts, and providing a one-click execution option.

[0035] For example, the relevant graph nodes are first visually highlighted; if the user does not take any further action, a brief contextual prompt is displayed; if the system has a very high level of confidence and the user is still confused, a shortcut button for the core behavior sequence of the one-click execution mode is provided.

[0036] Furthermore, the dynamic adjustment in S400 is based on a preset reinforcement learning-inspired formula to update the pattern strength, which takes into account historical strength, time decay, positive feedback gain, and negative ignore penalty.

[0037] For example, the default formula is: Wherein, strength_new: represents the confidence strength value of the intuitive pattern after the current task is completed, used to measure the likelihood that the pattern will be recommended or activated in the future; strength_old: represents the historical cumulative confidence strength of the intuitive pattern before the start of the current task, the initial value can be set to 0 or set based on prior knowledge; decay factor: refers to the value that gradually increases over time, used to simulate the memory forgetting effect; its value range is [0,1], usually proportional to the time interval from the triggering of the intuitive pattern to the completion of the task; feedback gain: a scalar value set according to the positive or negative and degree of the final result of the task, used to quantify the contribution of the intuitive pattern to the success of the task; for example, if the task is successful and efficiency is significantly improved, the feedback gain is... Set to +2; if the task fails but does not completely deviate from the path, set to -1; Adoption coefficient: indicates whether the user actively adopts the operation path suggested by the intuitive pattern during the task, with a value of 0 or 1; when the user performs a key operation guided by the pattern, the adoption coefficient is 1, otherwise it is 0; Ignore penalty: used to suppress the problem of users maintaining high confidence after frequently ignoring an intuitive pattern, its value is a non-negative real number, which can be dynamically adjusted according to the number of times or duration of ignoring; for example, each ignoring increases the penalty value by 0.5; Ignore coefficient: indicates the frequency or severity of the user's ignoring behavior of the intuitive pattern, with a value range of [0,1], which can be calculated by using a sliding window to count the proportion of ignoring the pattern in the most recent N interactions.

[0038] Example 2: Based on Example 1, Example 2 of this application provides a more specific method for constructing domain expert intuitive patterns based on interactive graph evolution tracing.

[0039] This embodiment uses the scenario of "medical image-assisted diagnosis" as an example for illustration. Experts work through an image reading system that integrates a medical knowledge graph (containing nodes such as diseases, symptoms, image features, and anatomical structures). The method specifically includes the following steps: S100 collects in real time the operations performed by domain experts on the domain knowledge graph when solving domain-specific problems, and atomizes each operation into a predefined type of interactive behavior, forming a time-stamped sequence of interactive behaviors.

[0040] When a radiologist uses the system to analyze a CT image suspected of containing a "pulmonary nodule," the sequence acquisition module in the system's background begins to work.

[0041] Atomization: All interactions between experts and the knowledge graph are decomposed and recorded in real time. For example: Focusing on the operation: The expert used the mouse to select a ground-glass nodule (GGO) with a diameter of 8mm on the image. The system recorded: [Operation: FOCUS, Target Node: “GGO_Node_001”, Timestamp: t1]. Subsequently, the expert clicked on the “Lung Cancer” disease node in the knowledge graph sidebar and recorded: [Operation: FOCUS, Target Node: “Disease_LungCancer”, Timestamp: t2].

[0042] Association operation: The expert dragged the selected GGO nodule node onto the "Lung Cancer" node and selected the relationship type "Possibly an early manifestation". System log: [Operation: RELATE, Source Node: “GGO_Node_001”, Relation Type: “suspicious_for”, Target Node: “Disease_LungCancer”, Timestamp: t3].

[0043] Questioning action: The expert clicked the "Question" icon on the automatically labeled attribute "Smooth Edge" in the attribute panel of the nodule node. System record: [Operation: QUESTION, Target node: "GGO_Node_001", Questioned attribute: "Edge feature", Timestamp: t4].

[0044] Confirmation procedure: To verify the margins, experts used a "3D reconstruction" tool to observe the nodule from different angles and linked it to a literature article on "GGO lobulation sign". System log: [Operation: CONFIRM, Target node: "GGO_Node_001", Confirmation method: "3D view", Related evidence: "Literature ID: Ref_123", Timestamp: t5].

[0045] After the entire diagnostic session concludes, a timestamped sequence of original interaction behaviors is generated: [FOCUS(GGO) ->FOCUS(LungCancer) ->RELATE(GGO ->LungCancer) ->QUESTION(GGO edge) ->CONFIRM(GGO via 3D)]. This sequence, along with the final diagnostic conclusion (e.g., "early-stage lung adenocarcinoma is highly probable; follow-up is recommended"), is stored in the session log database.

[0046] S200: Based on the interaction behavior sequences of multiple problem-solving sessions, extract behavioral sub-sequences that are associated with positive problem-solving effects as candidate intuitive response patterns.

[0047] Specifically, the conversation logs of hundreds of experts and thousands of diagnoses of pulmonary nodules were analyzed, and the following steps were performed: S210. Sequence Alignment and Cleaning: Using the Dynamic Time Warping (DTW) algorithm, the sequences of different experts analyzing "ground-glass nodules" cases are aligned on the time axis, and random clicks (noise) with extremely short durations are filtered out.

[0048] S220, Multi-dimensional Clustering: Calculates the comprehensive similarity between sequences. For example, it combines the edit distance of operation type sequences (such as the difference between [FOCUS, RELATE, QUESTION] and [FOCUS, FOCUS, RELATE]) with the semantic similarity of target nodes (calculated using word vectors of node names). Hierarchical clustering is used to group sessions with similar behavioral patterns into several clusters.

[0049] S230. Frequent Subsequence Mining and Evaluation: Within the session cluster with "high diagnostic accuracy," sequence pattern mining algorithms such as PrefixSpan were used to discover a high-frequency subsequence: [FOCUS (specific type of nodule) -> RELATE (nodule -> suspected disease) -> QUESTION (key nodule features such as margins and density) -> CONFIRM (via advanced tools or literature)]. Statistical analysis shows that sessions following this pattern have significantly higher diagnostic efficiency (average time) and accuracy than other patterns. This subsequence was selected as a candidate intuitive pattern for "deep differential diagnosis of pulmonary nodules."

[0050] S300. Associate the candidate intuitive response patterns with specific knowledge graph state triggering conditions and encapsulate them into structured intuitive pattern units.

[0051] Specifically, the above candidate patterns are encapsulated into a structured intuitive pattern unit (illustrated in JSON-LD format).

[0052] { "@id": "Pattern_GGN_Differential_Diagnosis", “name”: “Ground-glass nodule depth differential diagnosis pattern” “description”: “The standard deep analysis path initiated by experts after discovering GGO”. “triggerCondition”: { "type": "GraphPattern", "query": "` / / SPARQL example" PREFIX med:<http: / / medical-ontology.org / > ASK { ?session a med:DiagnosisSession. ?session med:hasFocusedNode ?nodule . ?nodule a med:PulmonaryNodule . ?nodule med:hasDensity “GroundGlass” . FILTER NOT EXISTS { ?nodule med:hasVerifiedCharacteristic ?anyChar .} }`” / / Triggering condition: The user focuses on a lung nodule node with a density of "ground glass" during the session, and the nodule does not yet have any verified features.

[0053] }, “coreActionSequence”: [ { "order": 1, "action": "SUGGEST_RELATE", "params": { "source": "?nodule", "suggestedTargetType": "med:Disease"}}, { “order”: 2, “action”: “SUGGEST_QUESTION”, “params”: { “target”: “?nodule”, “aspect”: [“edge features”, “internal density uniformity”]}}, { “order”: 3, “action”: “SUGGEST_CONFIRM”, “params”: { “target”: “?nodule”, “method”: [“3D Reconstruction”, “Multi-phase Phase Enhanced Scan Contrast”]}} ], “expectedImpact”: { "nodeWeightChange": { "?nodule": {"confidence": "increase"}}, “potentialNewRelations”: [ “?nodule med:hasDefinitiveCharacteristic ?char” ] }, “metadata”: { "creationDate": "2023-10-26", "initialStrength": 0.75, “healthScore”: 0.85 } } S400: When it is detected that the current knowledge graph state matches the triggering condition of an intuitive mode unit, guidance is provided to the user, and the intuitive mode unit is dynamically adjusted based on user feedback.

[0054] Specifically, when a resident physician uses the system, the pattern validation and dynamic adjustment steps include: S410, Guided Trigger: The resident physician focuses on a GGO nodule. The system detects whether the current atlas state (focused GGO, unverified features) matches the trigger condition of Pattern_GGN_Differential_Diagnosis.

[0055] Gradual guidance: First threshold (matching degree > 60%): The system will visually highlight relevant disease nodes such as "lung cancer" and "inflammatory pseudotumor" in the interface.

[0056] Second threshold (matching degree > 80%): The system displays a brief prompt in the sidebar: "In-depth analysis tip: For ground-glass nodules, it is recommended to associate them with differential diagnoses and carefully evaluate their margin and density characteristics." The third threshold (matching degree > 95%, and no key user operation within 30 seconds): The system displays a "One-click deep analysis" button floating next to the nodule.

[0057] S402, Dynamic Adjustment: If a resident physician clicks the "One-Click In-Depth Analysis" button, the system automatically executes the suggested steps in the mode, such as association, questioning, and reviewing 3D reconstruction, and records the final accurate completion of the task. Based on positive feedback, the system calculates the feedback gain according to the formula (as described in claim 7) to increase the intensity value of the mode.

[0058] If a resident physician ignores all guidance, quickly diagnoses the condition as "benign" and ends the session, and subsequent events prove that the diagnosis was hasty, the system will impose an ignore penalty, reducing the intensity of that mode.

[0059] The intensity of a pattern decays over time, and patterns that are not used effectively for a long period of time will gradually weaken.

[0060] S500: Periodically calculate a health score for each intuitive mode unit. The health score is calculated by combining recent usage frequency, user feedback rating, intensity value and time decay factors. When the health score is lower than the threshold, the mode review process is triggered.

[0061] Specifically, a health score is calculated periodically (e.g., weekly) for each intuitive pattern unit, using the following formula: When the health level is lower than the preset threshold, the mode review process is triggered.

[0062] Calculation example: For Pattern_GGN_Differential_Diagnosis, Recent usage frequency (w1=0.3): Triggered 20 times this week, frequency score 0.8.

[0063] Average feedback rating (w2=0.4): This week, the average task completion quality rating of users who adopted the suggestions was 4.2 / 5.0, which is equivalent to 0.84.

[0064] Intensity value (w3=0.2): The current intensity is 0.88.

[0065] Age penalty (w4=0.1): The pattern has been created for 12 months, and the penalty is 0.1. log(12) ≈ 0.11.

[0066] Health level = 0.3 0.8 + 0.4 0.84 + 0.2 0.88 - 0.11 = 0.806.

[0067] Review Process: A health threshold of 0.7 is set. When a pattern's health falls below this threshold, the system automatically marks it as "pending review" and notifies the knowledge administrator. The administrator can then check whether the pattern has become invalid due to medical guideline updates or whether its triggering conditions are too broad / narrow, and decide whether to revise or archive it.

[0068] S600: Based on intuitive pattern units with health scores higher than a threshold, automatically generate their derivative patterns. The derivative patterns include parent patterns generated by relaxing constraints in trigger conditions, child patterns generated by adding constraints in trigger conditions, and combined patterns generated by merging the core behavior sequences of multiple intuitive pattern units.

[0069] Specifically, based on patterns with high health (e.g., >0.85), the system can automatically generate derivative and combined patterns.

[0070] Generate parent pattern (generalization): Relax the trigger condition of Pattern_GGN_Differential_Diagnosis, and change ?nodule a med:PulmonaryNodule . ?nodule med:hasDensity “GroundGlass” . to ?nodule a med:PulmonaryNodule ., thus obtaining a universal differential diagnosis pattern applicable to all pulmonary nodules, whose initial strength inherits 0.8 from the parent pattern.

[0071] Generate sub-mode (specialization): Add trigger conditions, such as requiring nodule size to be greater than 10mm, to generate an "emergency assessment mode for GGOs larger than 10mm", whose initial behavior sequence may add the operation "recommend PET-CT assessment".

[0072] Generate a combined model: Combine the core behavioral sequences of the "In-depth Differential Diagnosis Model for Lung Nodules" with another high-healthy "Lung Cancer Staging Assessment Model" to form a combined model of "Integrated Workflow for Lung Nodule Diagnosis and Staging" for highly suspicious cases.

[0073] S700. Construct a similarity network among intuitive pattern units. Calculate edge weights based on trigger condition overlap, behavior sequence edit distance, and application scenario overlap. When an intuitive pattern unit is verified as valid, its credibility gain propagates to neighboring nodes according to the similarity weight; when an intuitive pattern unit is disproven, its credibility loss propagates to neighboring nodes according to the similarity weight.

[0074] Specifically, a similarity network between patterns is constructed. Taking two patterns in the medical field as an example, such as... Figure 3 As shown: Mode A: Emergency treatment mode for "chest pain accompanied by ST segment elevation on electrocardiogram".

[0075] Pattern B: Differential diagnosis pattern for “chest pain accompanied by elevated myocardial enzyme levels”.

[0076] Edge weight calculation: Both triggering conditions involve nodes related to "chest pain" and "cardiac emergencies" (high overlap); both behavioral sequences include "associated with cardiovascular diseases" and "recommendation of emergency examination" (small edit distance); both application scenarios are in the emergency department (overlapping scenarios). Therefore, the similarity weight between the two is high (e.g., 0.8).

[0077] Credibility propagation: If Pattern A proves extremely effective in multiple real-world scenarios, its strength value increases significantly by 0.1. Then, the credibility of Pattern B will receive a weighted gain: New strength of Pattern B = Old strength of Pattern B + 0.1 0.8 * Propagation coefficient.

[0078] Conversely, if pattern A is disproven by new clinical evidence, its strength drops sharply, and the strength of pattern B will also be negatively affected to a corresponding degree, prompting the administrator to review pattern B as well.

[0079] It should be noted that the parts in this embodiment that are the same as or similar to those in Embodiment 1 can be referred to each other, and will not be repeated in this application.

[0080] Example 3: Building upon Example 2, Example 3 of this application provides a domain expert intuitive pattern construction system based on interaction graph evolution tracing. This system adopts a microservice architecture and specifically includes: Interactive Client: Provides a graphical knowledge graph interactive interface (such as a web application) for domain experts, with a built-in behavior acquisition probe responsible for capturing raw interaction events and sending them to the backend.

[0081] The sequence acquisition module is used to collect in real time the operations performed by domain experts on the domain knowledge graph when solving domain-specific problems, and to atomize each operation into a predefined type of interactive behavior, forming a time-stamped sequence of interactive behaviors.

[0082] Specifically, the sequence acquisition module receives the client event stream, performs real-time atomication and formatting based on a predefined dictionary of atomic operation types, adds timestamps and session IDs, and stores the data in the interactive sequence time-series database.

[0083] The pattern discovery module is used to mine behavioral subsequences that are associated with positive problem-solving effects based on the interaction behavior sequences of multiple problem-solving sessions, as candidate intuitive response patterns.

[0084] Specifically, the pattern discovery module periodically (e.g., daily) pulls batches of session data from a time-series database and performs cleaning, alignment, clustering, and frequent subsequence mining algorithms. This service calls a natural language processing service to calculate node semantic similarity and accesses the task effect library to obtain session result labels for evaluation. The mined candidate patterns are output to a pattern library.

[0085] The pattern encapsulation module is used to associate the candidate intuitive response patterns with specific knowledge graph state triggering conditions and encapsulate them into structured intuitive pattern units.

[0086] Specifically, the schema encapsulation module provides graphical tools for knowledge engineers to review and adjust candidate schemas, and encapsulates them into a standardized JSON-LD format, which is stored in the schema library (which can use a document database that supports graph queries, such as Neo4j or ArangoDB).

[0087] The pattern verification and adjustment module is used to provide guidance to the user when the current knowledge graph state is detected to match the triggering condition of a certain intuitive pattern unit, and to dynamically adjust the intuitive pattern unit based on user feedback.

[0088] Specifically, the pattern verification and adjustment module monitors the client's graph state changes in real time. This service includes a trigger condition matching engine that continuously matches the current graph state with trigger conditions (compiled into executable queries) in the pattern library. When a match is successful, it interacts with the bootstrapping strategy engine to determine the strength and form of bootstrapping and sends the bootstrapping instructions back to the interactive client.

[0089] In addition, the pattern verification and adjustment module is also used to receive feedback events from the client regarding guided adoption and task results, dynamically update the strength value, health and other indicators of the corresponding patterns in the pattern library according to predefined formulas, and trigger the pattern derivation service (pattern derivation module) and the credibility network propagation service (credibility propagation module) to perform corresponding derivation and propagation calculations.

[0090] The various services communicate loosely via message queues or RPC interfaces, sharing a unified domain ontology and user session context. This system achieves a complete closed loop from data acquisition, pattern mining, encapsulation and storage, online matching, guided execution, to feedback learning.

[0091] It should be noted that the system provided in this embodiment is the corresponding system of the method provided in embodiment 2. Therefore, the parts that are the same as or similar to those in embodiment 2 in this embodiment can be referred to each other, and will not be described again in this application.

Claims

1. A method for constructing intuitive patterns of domain experts based on interaction graph evolution tracing, characterized in that, include: S100: Real-time collection of operations performed by domain experts on the domain knowledge graph when solving domain-specific problems, and atomizes each operation into predefined types of interactive behaviors, forming a time-stamped sequence of interactive behaviors; S200: Based on the interaction behavior sequence of multiple problem-solving sessions, extract the behavioral sub-sequences associated with positive problem-solving effects as candidate intuitive response patterns; S300. Associate the candidate intuitive response patterns with specific knowledge graph state triggering conditions and encapsulate them into structured intuitive pattern units. S400: When it is detected that the current knowledge graph state matches the triggering condition of an intuitive mode unit, guidance is provided to the user, and the intuitive mode unit is dynamically adjusted based on user feedback.

2. The method for constructing domain expert intuitive patterns based on interaction graph evolution tracing according to claim 1, characterized in that, In S100, the predefined types of interactive behaviors include: focusing operation, association operation, questioning operation, and confirmation operation; wherein, focusing operation is used to represent the concentration of attention on a specific graph node, association operation is used to represent the establishment or adjustment of relationships between nodes, questioning operation is used to represent the user raising questions about the authenticity, reliability, or relevance of a node or relationship in the graph, and confirmation operation is used to represent the user verifying or confirming a node or relationship through external evidence, logical reasoning, or authoritative sources.

3. The method for constructing domain expert intuitive patterns based on interaction graph evolution tracing according to claim 2, characterized in that, S200 includes: S210. Time-align the interaction behavior sequences of multiple problem-solving sessions and remove noisy operations with operation durations below a preset threshold. S220. Calculate sequence similarity based on at least two dimensions of operation type sequence, target node semantic similarity, and operation time interval distribution, and perform clustering using a hierarchical clustering algorithm; S230. Within each cluster, use a sequence pattern mining algorithm to extract high-frequency behavioral subsequences, and calculate the occurrence frequency, success correlation, and expert consensus index of each subsequence. Combined with the problem-solving effect evaluation, candidate intuitive response patterns are selected.

4. The method for constructing domain expert intuitive patterns based on interaction graph evolution tracing according to claim 3, characterized in that, In S300, the triggering condition is defined by a logical expression based on knowledge graph subgraph matching. The logical expression needs to specify at least one of the following conditions: necessary node set, prohibited node state, node attribute constraint, and relation topology requirement.

5. The method for constructing domain expert intuitive patterns based on interaction graph evolution tracing according to claim 4, characterized in that, In S300, the structured intuitive pattern unit is stored in a machine-readable format, and its contents include pattern identifier, trigger condition logic, behavior sequence array, expected impact description and metadata information. The behavior sequence array records the operation type, target node, and operation parameters, while the expected impact description records the expected changes to the weights of knowledge graph nodes and the newly added relationships.

6. The method for constructing domain expert intuitive patterns based on interaction graph evolution tracing according to claim 5, characterized in that, In S400, the guidance is implemented using a progressive non-intrusive interface. The guidance intensity is increased in stages according to the matching degree between the current map state and the triggering conditions. The guidance forms include visual highlighting, text prompts, and providing a one-click execution option.

7. The method for constructing domain expert intuitive patterns based on interaction graph evolution tracing according to claim 6, characterized in that, Also includes: S500: Periodically calculate a health score for each intuitive mode unit. The health score is calculated by combining recent usage frequency, user feedback rating, intensity value and time decay factors. When the health score is lower than the threshold, the mode review process is triggered. S600: Based on intuitive pattern units with health scores higher than a threshold, automatically generate their derivative patterns. The derivative patterns include a parent pattern generated by relaxing constraints in the triggering conditions, a child pattern generated by adding constraints in the triggering conditions, and a combined pattern generated by merging the core behavior sequences of multiple intuitive pattern units. S700. Construct a similarity network among intuitive pattern units. Calculate edge weights based on trigger condition overlap, behavior sequence edit distance, and application scenario overlap. When an intuitive pattern unit is verified as valid, its credibility gain propagates to neighboring nodes according to the similarity weight; when an intuitive pattern unit is disproven, its credibility loss propagates to neighboring nodes according to the similarity weight.

8. A domain expert intuitive pattern construction system based on interaction graph evolution tracing, characterized in that, For performing the method according to any one of claims 1 to 7, comprising: The sequence acquisition module is used to collect in real time the operations performed by domain experts on the domain knowledge graph when solving domain-specific problems, and to atomize each operation into a predefined type of interactive behavior, forming a time-stamped sequence of interactive behaviors; The pattern discovery module is used to mine behavioral subsequences that are associated with positive problem-solving effects based on the interaction behavior sequences of multiple problem-solving sessions, as candidate intuitive response patterns; The pattern encapsulation module is used to associate the candidate intuitive response patterns with specific knowledge graph state triggering conditions and encapsulate them into structured intuitive pattern units. The pattern verification and adjustment module is used to provide guidance to the user when the current knowledge graph state is detected to match the triggering condition of a certain intuitive pattern unit, and to dynamically adjust the intuitive pattern unit based on user feedback.

9. A computer storage medium, characterized in that, The computer storage medium stores a computer program; when the computer program is run on the computer, it causes the computer to perform the method described in any one of claims 1 to 7.

10. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor for executing the computer program to implement the method as described in any one of claims 1 to 7.

Images