Decision-oriented knowledge representation and reasoning processing method and system based on multi-source data

Abstract

The application discloses a decision-oriented knowledge expression and reasoning processing method and system based on multi-source data, relates to the field of electric digital data processing, and realizes automatic conversion and fusion expression of data to knowledge by uniformly accessing and preprocessing multi-source data, constructing a multi-level data semantic mapping model; a knowledge expression mechanism oriented to a decision target is introduced, business rules and constraint conditions are embedded into a knowledge structure in a computable form, a three-layer heterogeneous hypergraph knowledge representation system is formed; based on the knowledge representation system, a hybrid reasoning processing flow combining symbolic reasoning and data-driven reasoning is constructed, and an interpretable decision conclusion is output. The system comprises a data access module, an intelligent analysis module, a collaborative application module and a basic management and control module for executing the above method. The application effectively improves knowledge utilization efficiency and reasoning accuracy, and is suitable for application scenarios such as intelligent decision support, risk assessment, business analysis and automatic auxiliary decision making.

Description

Technical Field

[0001] This invention relates to the field of electronic digital data processing, and in particular to a method and system for decision-oriented knowledge representation and reasoning based on multi-source data. Background Technology

[0002] With the rapid development of big data and artificial intelligence technologies, enterprises and organizations have accumulated massive amounts of multi-source heterogeneous data in their business decision-making processes. This data includes structured data from databases such as transaction records and personnel files, as well as semi-structured data such as system logs and electronic documents, and the ever-increasing amount of unstructured data such as office emails, surveillance videos, and call recordings. Utilizing knowledge graph technology to mine the relationships within this data has become an important tool for supporting intelligent decision-making. However, existing data processing and knowledge reasoning technologies still face significant technical bottlenecks when dealing with complex business decision-making scenarios. The semantic fusion depth of multi-source heterogeneous data is insufficient, resulting in data silos and semantic gaps. Traditional knowledge graphs lack decision-oriented guidance and cannot directly support business decisions, while single reasoning models struggle to balance accuracy and interpretability.

[0003] Patent CN112529187A discloses a knowledge acquisition method that integrates semantics and features from multi-source data, demonstrating excellent performance in industrial process data processing. It effectively achieves semantic labeling and compact feature extraction of data through unsupervised clustering and convolutional autoencoders. However, this method primarily focuses on feature dimensionality reduction and state recognition, and its constructed semantic knowledge base is more of a static mapping of the physical meaning of the data. When facing decision-making scenarios requiring logical deduction based on complex business rules, laws, and regulations, a purely feature-based clustering method may struggle to provide strong interpretable logical support. Patent CN120216706B proposes a cross-scenario question-answering framework for non-performing assets based on knowledge graphs. It constructs a complete system encompassing knowledge fusion, dynamic reasoning, and scenario adaptation, effectively solving the problem of cross-scenario knowledge transfer in the non-performing asset field. Although this framework integrates multiple strategies such as rule-based reasoning and graph reasoning, it is essentially still more of a component-based functional integration framework. Its knowledge representation often follows the traditional entity-relationship graph structure, failing to internalize decision-making objectives and business rules as independent topological levels within the graph structure. This results in room for improvement in the guidance of reasoning paths and computational efficiency when handling multi-objective conflicts or conducting deep attribution analysis. Patent CN120336547A discloses an enterprise-level simulation knowledge graph construction method based on multimodal data integration, which greatly improves the efficiency of cross-modal retrieval and scheme comparison of simulation data by utilizing the collaboration of vector databases and graph databases. However, the core logic of this technical solution lies in semantic expansion search and similar case recall, aiming to assist human decision-making through historical data. For scenarios with clear logical inference requirements, such as requiring the system to autonomously make compliance judgments, risk blocking, or automated approvals, relying solely on vector similarity matching often fails to meet the accuracy and robustness requirements of decision-making. In summary, existing technologies suffer from insufficient decision-making orientation in knowledge representation when dealing with multi-source data decision-making problems. Existing knowledge graphs primarily focus on storing objective facts or associating features, lacking a mechanism for hierarchically and uniformly modeling data facts, business logic, and decision objectives. This makes knowledge retrieval easy but direct support for decision-making difficult. Furthermore, the depth of integration between symbolic and numerical reasoning needs improvement. Existing methods often treat rule-based reasoning and deep learning as independent modules executed in parallel or through simple weighted fusion, failing to achieve deep interaction and complementarity. When faced with complex scenarios containing both fuzzy unstructured data and strict compliance clauses, it is difficult to simultaneously ensure the accuracy and interpretability of reasoning.

[0004] To address the above issues and resolve the common problems existing in current methods, a decision-oriented knowledge representation and reasoning processing method based on multi-source data is needed. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of the existing technology by providing a method for unifying the access of structured, semi-structured, and unstructured data through a multimodal data warehouse and constructing a multi-level, multi-granular data semantic mapping model by combining core feature extraction. This method enables the breakdown of cross-modal data silos and the elimination of semantic gaps between heterogeneous data. By establishing attribute associations between source data feature items and a unified knowledge framework, this invention can automatically transform and deeply integrate multi-source heterogeneous data into a unified knowledge representation, significantly improving the utilization efficiency of complex underlying data and the completeness of information mining.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0007] A decision-oriented knowledge representation and reasoning processing method based on multi-source data is characterized by the following steps:

[0008] Step S1: Utilize a multimodal data warehouse to uniformly access structured data, semi-structured data, and unstructured data;

[0009] Step S2: Preprocess and extract core features from the multi-source data to construct a multi-level, multi-granularity data semantic mapping model, thereby realizing the automatic transformation and fusion of data into knowledge; the multi-level refers to the hierarchical mapping from the data layer, feature layer to the knowledge layer, and the multi-granularity refers to the realization of fine-grained, medium-grained, and coarse-grained semantic expression and calculation in each level.

[0010] Step S3: Introduce a knowledge representation mechanism oriented towards decision-making objectives, embedding business rules, constraints, and decision preferences into the knowledge structure in a computable form to form a knowledge representation system with a three-layer heterogeneous hypergraph structure; the knowledge representation system will perform hierarchical and unified modeling of data facts, business logic, and decision-making objectives.

[0011] Step S4: Based on the knowledge representation system, construct a hybrid reasoning process that combines symbolic reasoning and data-driven reasoning to deduce, verify, and evaluate the results of decision-related issues;

[0012] Step S5: Output interpretable and traceable decision support conclusions.

[0013] Furthermore, the multimodal data warehouse in step S1 further includes:

[0014] Step S11: Build a unified data access layer to connect with data sources that carry structured, semi-structured, and unstructured data;

[0015] Step S12: Based on the original characteristics of the three types of data, a unified data integration platform is used to collect the three types of data in real time. During the collection process, basic format verification is performed on the three types of data respectively, and the original characteristics of each type of data are preserved throughout the process without any business processing.

[0016] Step S13: Based on the original characteristics of structured data, semi-structured data, and unstructured data, the three types of data that have completed verification are stored in matching storage engines respectively. Structured data is stored in a relational database, semi-structured data is stored in a document database, and unstructured data is stored in object storage and a distributed file system. Each engine achieves linkage query through a unified interface, realizing unified original storage of the three types of data in the multimodal data warehouse, completing the overall access, and constructing the multimodal data warehouse.

[0017] Furthermore, the preprocessing in step S2 includes data deduplication, missing value completion, format standardization and normalization operations. The missing value completion uses a time series data interpolation algorithm to complete time series data, uses an image generation model to complete missing image data, and uses a feature selection algorithm to suppress feature redundancy.

[0018] The core feature extraction includes: for unstructured data such as videos and texts, using deep neural networks combined with time-series modeling to extract spatiotemporal and semantic features; for structured data such as numerical and time-series signals, using a joint frequency domain and time domain feature extraction model to extract multi-scale features; and for qualitative data such as natural language, converting it into quantitative features through language tuples combined with membership degree calculation.

[0019] Furthermore, the data semantic mapping model in step S2 further includes:

[0020] A unified knowledge framework incorporating domain ontology models and business mechanism knowledge is constructed. The metadata structure of multi-source heterogeneous data is parsed, and entity alignment and attribute association rules between source data feature items and the unified knowledge framework are established. Based on the association rules, a semantic mapping algorithm is used to perform alignment calculations to realize the mapping between source data features and the knowledge framework, maintaining the consistency of semantic topology. Multi-feature association metrics and logical transformation operators are used to resolve heterogeneous semantic conflicts. During the mapping process, missing data is filled in by combining feature representation models with interpolation methods to achieve the fusion of cross-type data.

[0021] Furthermore, in step S3, the business rules are configured as domain-specific decision logic, the constraints include data time range and applicable object restrictions, and the decision preferences include priority setting rules; the computable form includes logical encoding representing business rules, constraint equations and penalty functions defining decision boundaries, weight matrices and evaluation functions quantifying decision preferences, and probability distributions and membership parameters for handling uncertain reasoning.

[0022] Furthermore, the knowledge representation system described in step S3 further includes:

[0023] The knowledge representation system is a three-layer heterogeneous hypergraph structure, including a bottom-layer fact subgraph, a middle-layer logic subgraph, a top-layer target subgraph, and an inter-layer association mechanism. The hypergraph structure allows hyperedges to connect multiple heterogeneous nodes. The inter-layer association mechanism includes condition-triggered edges connecting fact nodes and rule nodes, and decision contribution weight edges connecting rule nodes and target nodes. The weight values ​​are set and dynamically updated based on the analytic hierarchy process (AHP) combined with business experience. The bottom-layer fact subgraph consists of entity nodes and attribute edges generated by the data semantic mapping model, used to represent the objective business state of multi-source data. The middle-layer logic subgraph consists of rule nodes, constraint operator nodes, and logical operation edges, used to represent business judgment criteria and compliance verification logic. The top-layer target subgraph consists of decision target nodes, risk level nodes, and utility function nodes, used to represent the final output orientation of reasoning. The inter-layer association mechanism includes condition-triggered edges connecting fact nodes and rule nodes, and decision contribution weight edges connecting rule nodes and target nodes.

[0024] Furthermore, the symbolic reasoning described in step S4 is based on preset business rules and a knowledge representation system for logical deduction, specifically including decision-related logical operations; monitoring changes in entity state in the underlying fact subgraph, automatically triggering associated rule nodes in the middle-level logical subgraph through a pattern matching mechanism, employing a forward linking reasoning strategy, performing transitive closure calculations along logical operation edges, verifying whether the current fact satisfies the constraint conditions of the constraint operator node, generating a logical blocking signal when the reasoning path touches mutually exclusive business rules or violates rigid constraint nodes, and marking the evidence subgraph that causes the conflict; if the verification passes, generating a definite logical judgment result, and outputting a symbolic execution path containing a complete triggering rule chain, as... The logical basis for decision interpretation; the data-driven reasoning specifically includes using a graph neural network model to map discrete nodes and edges in the underlying fact subgraph to high-dimensional continuous vectors, generating relation embedding vectors based on knowledge representation learning algorithms, calculating similarity metrics based on the embedding vectors, predicting undefined potential association edges between fact nodes, realizing link prediction for hidden risk groups and related transactions, inputting target entity vectors aggregating neighborhood features into a pre-trained deep learning classifier, calculating the confidence probability distribution of their belonging to nodes of different risk levels in the top-level target subgraph, and finally, based on the newly flowing data stream, fine-tuning the model parameters in real time through an incremental learning mechanism to update the predicted scores of potential risk points in real time.

[0025] Furthermore, the hybrid inference processing flow described in step S4 includes cross-validation, conflict resolution, and consistency verification steps for symbolic inference results and data-driven inference results; when the symbolic inference encounters rigid business constraints and generates a logic blocking signal, the symbolic inference is given absolute priority, and the blocking signal is directly used as the consistency verification result, thus rejecting the prediction output of data-driven inference.

[0026] Furthermore, the interpretable and traceable decision support conclusions in step S5 include a logical chain of reasoning, data source annotations, and explanations of the basis for the conclusions, clarifying the derivation process of the decision conclusions. The conclusions are presented in the form of a hierarchical path diagram of fact nodes, rule nodes, and target nodes.

[0027] The decision-oriented knowledge representation and reasoning processing system based on multi-source data specifically includes a data access module, an intelligent analysis module, a collaborative application module, and a basic management and control module.

[0028] The data access module is used to realize the batch access, management, and automated parsing of multimodal and multi-format data. The data access module includes a file access unit, a communication data access unit, and a terminal device data access unit. Specifically, the file access unit is used to realize the batch import, parsing monitoring, and sharing management of multi-format documents and multimedia files; the communication data access unit is used to realize the batch extraction, basic information parsing, and status monitoring of network communication data; and the terminal device data access unit is used to realize the collection and standardized parsing of terminal device operating status data and network connection data.

[0029] The intelligent analysis module, communicatively connected to the data access module, is used for in-depth analysis and information mining of standardized and parsed data. The intelligent analysis module includes a semantic and behavioral analysis unit, an entity profiling analysis unit, a relationship network construction unit, a spatiotemporal geographic analysis unit, and a multilingual processing unit. Specifically, the semantic and behavioral analysis unit performs semantic clustering, behavioral feature statistics, and AI-based content summarization of communication data; the entity profiling analysis unit constructs multi-dimensional device and network feature profiles to analyze and display relationships and behavioral trajectories; the relationship network construction unit constructs multi-level interaction relationship graphs of target entities based on multi-source data; the spatiotemporal geographic analysis unit performs target spatial location retrieval, cluster feature heatmap analysis, and temporal trajectory reconstruction; and the multilingual processing unit performs bidirectional translation processing of cross-language text.

[0030] The collaborative application module is bidirectionally connected to the intelligent analysis module to realize collaborative data analysis, thematic management, and analysis model construction. The collaborative application module is equipped with a collaborative analysis unit and a data modeling and analysis unit. The collaborative analysis unit is used to create business topics, conduct multi-user collaborative review and annotation, and perform in-depth analysis and summarization of analysis content based on artificial intelligence. The data modeling and analysis unit is used to manage data sources, perform visual correlation modeling, schedule data tasks, and manage model processing resources, supporting the visual construction and reuse of business decision-making models.

[0031] The basic control module is communicatively connected to the data access module, intelligent analysis module, and collaborative application module, respectively, and is used to realize system process control, permission configuration, and security auditing. The basic control module is equipped with a tag management unit, a process approval unit, a result export unit, a log auditing unit, and a permission control unit. Among them, the tag management unit is used to build and maintain a multi-dimensional entity and content tag system; the process approval unit is used to realize the process-oriented approval of system business processing; the result export unit is used to realize the generation and export management of analysis reports and business data; the log auditing unit is used to realize the recording and auditing of the entire system operation; and the permission control unit is used to realize role-based hierarchical access control.

[0032] Compared with the prior art, the present invention, employing the above technical solution, has the following technical effects:

[0033] 1. This invention provides a decision-oriented knowledge representation and reasoning processing method based on multi-source data. By constructing a multi-level data semantic mapping model and utilizing a multi-feature association measurement method, it effectively solves the semantic gap problem between structured, semi-structured, and unstructured data, realizes automatic alignment and complementarity of cross-modal data under a unified knowledge framework, and significantly improves data utilization.

[0034] 2. This invention provides a decision-oriented knowledge representation and reasoning processing method based on multi-source data. By introducing a knowledge representation mechanism oriented towards decision-making objectives, it innovatively constructs a three-layer heterogeneous hypergraph structure containing a bottom-level fact subgraph, a middle-level logic subgraph, and a top-level target subgraph. This structure directly internalizes and embeds business rules, rigid constraints, and decision preferences into the knowledge system in a computable form, breaking the limitations of traditional knowledge graphs that can only perform static fact storage and feature association. It achieves deep topological binding between objective data facts and subjective business judgment objectives, greatly enhancing the decision-oriented capability and computational efficiency of the knowledge model in complex business scenarios.

[0035] 3. This invention provides a decision-oriented knowledge representation and reasoning processing method based on multi-source data. It constructs a hybrid reasoning processing flow combining symbolic reasoning and data-driven reasoning, achieving deep interaction and complementarity between the two through cross-validation and conflict resolution mechanisms. On the one hand, this method utilizes deep learning models such as graph neural networks to uncover hidden potential connections, ensuring high prediction accuracy for unknown risks. On the other hand, it relies on a rigorous logical deduction mechanism to generate symbolic execution paths containing complete trigger rule chains and data source annotations. This effectively overcomes the technical bottleneck of a single computational model's inability to simultaneously achieve high accuracy and strong interpretability, ultimately outputting highly reliable and traceable business decision conclusions. Attached Figure Description

[0036] Figure 1This is a flowchart of the method of the present invention;

[0037] Figure 2 This is a flowchart illustrating the construction process of the multimodal data warehouse of the present invention.

[0038] Figure 3 Flowchart for constructing the data semantic mapping model of this invention;

[0039] Figure 4 This is a schematic diagram of the knowledge representation system of the present invention;

[0040] Figure 5 This is a flowchart of the hybrid reasoning process of the present invention. Detailed Implementation

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

[0042] A decision-oriented knowledge representation and reasoning processing method based on multi-source data includes the following steps:

[0043] Step S1: Utilize a multimodal data warehouse to uniformly access structured data, semi-structured data, and unstructured data;

[0044] Step S2: Preprocess and extract core features from the multi-source data to construct a multi-level, multi-granularity data semantic mapping model, thereby realizing the automatic transformation and fusion of data into knowledge; the multi-level refers to the hierarchical mapping from the data layer, feature layer to the knowledge layer, and the multi-granularity refers to the realization of fine-grained, medium-grained, and coarse-grained semantic expression and calculation in each level.

[0045] Step S3: Introduce a knowledge representation mechanism oriented towards decision-making objectives, embedding business rules, constraints, and decision preferences into the knowledge structure in a computable form to form a knowledge representation system with a three-layer heterogeneous hypergraph structure; the knowledge representation system will perform hierarchical and unified modeling of data facts, business logic, and decision-making objectives.

[0046] Step S4: Based on the knowledge representation system, construct a hybrid reasoning process that combines symbolic reasoning and data-driven reasoning to deduce, verify, and evaluate the results of decision-related issues;

[0047] Step S5: Output interpretable and traceable decision support conclusions.

[0048] Furthermore, such as Figure 2 As shown, the multimodal data warehouse in step S1 further includes:

[0049] Step S11: Establish a unified data access layer to connect to data sources that carry structured, semi-structured, and unstructured data; a dynamic message queue is configured as a buffer pool in the unified data access layer. To prevent system overload caused by instantaneous data surges, the system dynamically adjusts the allocated capacity of the buffer pool based on the real-time data inflow rate;

[0050] The buffer capacity calculation model is set as follows:

[0051]

[0052] in, The total capacity of the buffer dynamically allocated at time t; The total number of concurrently accessing data sources (including structured, semi-structured, and unstructured data sources). This represents the data arrival rate (in MB / s) of the i-th data source at time t. This represents the maximum peak data volume in the system's historical records. and For smoothing adjustment coefficient;

[0053] In this embodiment, the empirical value is set. [1.2, 1.5] Reserved margin, [0.1, 0.2] serves as the safety redundancy base for sudden peak values.

[0054] Step S12: Based on the original characteristics of the three types of data, a unified data integration platform is used to collect the three types of data in real time. During the collection process, basic format verification is performed on the three types of data respectively, and the original characteristics of each type of data are preserved throughout the process without any business processing.

[0055] Step S13: Based on the original characteristics of structured data, semi-structured data, and unstructured data, the three types of data that have completed verification are stored in matching storage engines respectively. Structured data is stored in a relational database, semi-structured data is stored in a document database, and unstructured data is stored in object storage and a distributed file system. Each engine achieves linkage query through a unified interface, realizing unified original storage of the three types of data in the multimodal data warehouse, completing the overall access, and constructing the multimodal data warehouse.

[0056] Furthermore, the preprocessing in step S2 includes data deduplication, missing value completion, format standardization, and normalization. Missing value completion uses a time-series data interpolation algorithm to complete the time-series data. When performing time-series data interpolation and completion, if a missing value is detected at timestamp t, the system first extracts a dynamic sliding window centered at t containing N valid sampling points. To reflect the higher reference value of nearest-neighbor data, the system introduces a time correlation decay mechanism using an exponential decay formula:

[0057]

[0058] Calculate each valid data point within the window The weights (where the coefficient controlling the decay rate) Preferred range is Then, the valid data within the window is weighted and summed using normalized weights, calculated using the formula:

[0059]

[0060] The final interpolation completion result is obtained. Furthermore, when the system detects that the missing rate of valid data points within the dynamic sliding window exceeds a preset threshold of 50%, it will automatically switch to and use the piecewise cubic Hermite spline interpolation (PCHIP) algorithm for completion, thereby effectively ensuring the monotonicity and smoothness of the interpolation curve in the case of continuous large segments of missing data, and preventing abnormal oscillations.

[0061] When performing missing image data completion, the system first constructs a binary mask matrix M through abnormal pixel detection or boundary extraction, where pixel values ​​of missing or damaged areas are marked as 1, and effectively preserved areas are marked as 0. Subsequently, the system processes the incomplete original image containing missing areas... The mask matrix M is fed together with the generator G of a pre-trained context-aware generative adversarial network (GAN). The generator uses multiple receptive fields to extract texture and high-level semantic features of surrounding effective pixels, infers and outputs a preliminary completed image. To ensure a natural transition between the generated region and the original image boundary, as well as consistency in global semantics, the system employs a joint optimization model combining spatial reconstruction loss and adversarial loss. The joint loss function is defined as:

[0062]

[0063] in, To constrain pixel-level structural consistency Norm loss, To evaluate the adversarial loss for the discriminator network to improve realism, the reconstruction weights are preferably set in the model inference parameters. Set at Between these elements, a smooth constraint on the local boundary is strengthened to counteract the weighting. Fixed at 1.

[0064] Ultimately, the system uses the mask fusion formula:

[0065]

[0066] in The Hadamard product represents the process of stitching the generated repaired area pixel-by-pixel with the real, effective area of ​​the original image, outputting high-fidelity complete image data. This provides structurally complete unstructured data support for subsequent spatiotemporal feature extraction, and suppresses feature redundancy through feature selection algorithms.

[0067] The core feature extraction includes: for unstructured data such as videos and texts, using deep neural networks combined with time-series modeling to extract spatiotemporal and semantic features; for structured data such as numerical and time-series signals, using a joint frequency domain and time domain feature extraction model to extract multi-scale features; and for qualitative data such as natural language, converting it into quantitative features through language tuples combined with membership degree calculation.

[0068] Furthermore, such as Figure 3 As shown, the data semantic mapping model in step S2 further includes:

[0069] A unified knowledge framework incorporating domain ontology models and business mechanism knowledge is constructed. The metadata structure of multi-source heterogeneous data is parsed, and entity alignment and attribute association rules between source data feature items and the unified knowledge framework are established. Based on the association rules, a semantic mapping algorithm is used to perform alignment calculations to realize the mapping between source data features and the knowledge framework, maintaining the consistency of semantic topology. Multi-feature association metrics and logical transformation operators are used to resolve heterogeneous semantic conflicts. During the mapping process, missing data is filled in by combining feature representation models with interpolation methods to achieve the fusion of cross-type data.

[0070] Furthermore, in step S3, the business rules are configured as domain-specific decision logic, the constraints include data time range and applicable object restrictions, and the decision preferences include priority setting rules; the computable form includes logical encoding representing business rules, constraint equations and penalty functions that limit decision boundaries, weight matrices and evaluation functions that quantify decision preferences, and probability distributions and membership parameters for handling uncertain reasoning.

[0071] For the constraint equations and penalty functions that define the decision boundary, assuming the system's decision output vector is X, the rigid constraints of the business (such as funding caps and compliance red lines) can be expressed as inequality constraints:

[0072]

[0073] To introduce these constraints into hybrid reasoning, a target evaluation function with an external penalty term can be constructed:

[0074]

[0075] in, Based on the expected returns of fundamental decisions, The penalty coefficient is extremely high; once the decision path reaches the red line (i.e., The penalty will be amplified exponentially, forcing the system to abandon the reasoning path.

[0076] For feature indicators that are difficult to classify into binary "black and white" categories (For example, the degree of deviation from abnormal behavior) is quantified using a Gaussian fuzzy membership function:

[0077]

[0078] in This is the standard reference center value for this feature. This is the tolerance distribution width. Calculated. It is directly used as the decay multiplier for passing the confidence level down to the next level in the "logical subgraph" of this node.

[0079] Furthermore, the knowledge representation system described in step S3 further includes:

[0080] The knowledge representation system is a three-layer heterogeneous hypergraph structure, including a bottom-layer fact subgraph, a middle-layer logic subgraph, a top-layer target subgraph, and inter-layer association mechanisms, such as... Figure 4 As shown, the hypergraph structure allows hyperedges to connect multiple heterogeneous nodes; the inter-layer association mechanism includes conditional trigger edges connecting fact nodes and rule nodes, and decision contribution weight edges connecting rule nodes and target nodes, where the weight values ​​are set and dynamically updated based on the analytic hierarchy process combined with business experience; specifically, business experts construct a judgment matrix A, calculate the eigenvector corresponding to the largest eigenvalue, and perform a consistency check (…). Establish the initial prior weights for each decision node. .

[0081] With the continuous input of data streams and feedback from actual business results, a time smoothing factor is introduced. Perform real-time adaptive fine-tuning:

[0082]

[0083] in, The objective weight distribution at time t is obtained based on data-driven reasoning (such as neural network attention weight extraction). It is used to balance expert experience with objective data.

[0084] The underlying fact subgraph consists of entity nodes and attribute edges generated by the data semantic mapping model, used to represent the objective business state of multi-source data; the middle-layer logic subgraph consists of rule nodes, constraint operator nodes, and logical operation edges, used to represent business judgment criteria and compliance verification logic; the top-layer target subgraph consists of decision target nodes, risk level nodes, and utility function nodes, used to represent the final output orientation of reasoning; the inter-layer association mechanism includes condition triggering edges connecting fact nodes and rule nodes, and decision contribution weight edges connecting rule nodes and target nodes.

[0085] Furthermore, such as Figure 5 As shown, the symbolic reasoning described in step S4 is based on preset business rules and a knowledge representation system for logical deduction. Specifically, it includes decision-related logical operations; monitoring changes in entity states in the underlying fact subgraph; automatically triggering associated rule nodes in the middle-level logical subgraph through a pattern matching mechanism; employing a forward linking reasoning strategy; performing transitive closure calculations along logical operation edges; verifying whether the current fact satisfies the constraint conditions of the constraint operator node; generating a logical blocking signal and marking the evidence subgraph causing the conflict when the reasoning path touches mutually exclusive business rules or violates rigid constraint nodes; if the verification passes, generating a definite logical judgment result and outputting a symbolic execution path containing a complete triggering rule chain as the decision. The logical basis for the policy explanation; the data-driven reasoning specifically includes using a graph neural network model to map discrete nodes and edges in the underlying fact subgraph to high-dimensional continuous vectors, generating relation embedding vectors based on knowledge representation learning algorithms, calculating similarity based on the embedding vectors, predicting undefined potential related edges between fact nodes, realizing link prediction for hidden risk groups and related transactions, inputting target entity vectors aggregating neighborhood features into a pre-trained deep learning classifier, calculating the confidence probability distribution of their belonging to nodes of different risk levels in the top-level target subgraph, and finally, based on the newly flowing data stream, fine-tuning the model parameters in real time through an incremental learning mechanism to update the predicted scores of potential risk points in real time.

[0086] Furthermore, the hybrid inference processing flow described in step S4 includes cross-validation, conflict resolution, and consistency verification steps for symbolic inference results and data-driven inference results; when the symbolic inference encounters rigid business constraints and generates a logic blocking signal, the symbolic inference is given absolute priority, and the blocking signal is directly used as the consistency verification result, thus rejecting the prediction output of data-driven inference.

[0087] Furthermore, the interpretable and traceable decision support conclusions in step S5 include a logical chain of reasoning, data source annotations, and explanations of the basis for the conclusions, clarifying the derivation process of the decision conclusions. The conclusions are presented in the form of a hierarchical path diagram of fact nodes, rule nodes, and target nodes.

[0088] The decision-oriented knowledge representation and reasoning processing system based on multi-source data specifically includes a data access module, an intelligent analysis module, a collaborative application module, and a basic management and control module.

[0089] The data access module is used to realize the batch access, management, and automated parsing of multimodal and multi-format data. The data access module includes a file access unit, a communication data access unit, and a terminal device data access unit. Specifically, the file access unit is used to realize the batch import, parsing monitoring, and sharing management of multi-format documents and multimedia files; the communication data access unit is used to realize the batch extraction, basic information parsing, and status monitoring of network communication data; and the terminal device data access unit is used to realize the collection and standardized parsing of terminal device operating status data and network connection data.

[0090] The intelligent analysis module, communicatively connected to the data access module, is used for in-depth analysis and information mining of standardized and parsed data. The intelligent analysis module includes a semantic and behavioral analysis unit, an entity profiling analysis unit, a relationship network construction unit, a spatiotemporal geographic analysis unit, and a multilingual processing unit. Specifically, the semantic and behavioral analysis unit performs semantic clustering, behavioral feature statistics, and AI-based content summarization of communication data; the entity profiling analysis unit constructs multi-dimensional device and network feature profiles to analyze and display relationships and behavioral trajectories; the relationship network construction unit constructs multi-level interaction relationship graphs of target entities based on multi-source data; the spatiotemporal geographic analysis unit performs target spatial location retrieval, cluster feature heatmap analysis, and temporal trajectory reconstruction; and the multilingual processing unit performs bidirectional translation processing of cross-language text.

[0091] The collaborative application module is bidirectionally connected to the intelligent analysis module to realize collaborative data analysis, thematic management, and analysis model construction. The collaborative application module is equipped with a collaborative analysis unit and a data modeling and analysis unit. The collaborative analysis unit is used to create business topics, conduct multi-user collaborative review and annotation, and perform in-depth analysis and summarization of analysis content based on artificial intelligence. The data modeling and analysis unit is used to manage data sources, perform visual correlation modeling, schedule data tasks, and manage model processing resources, supporting the visual construction and reuse of business decision-making models.

[0092] The basic control module is communicatively connected to the data access module, intelligent analysis module, and collaborative application module, respectively, and is used to realize system process control, permission configuration, and security auditing. The basic control module is equipped with a tag management unit, a process approval unit, a result export unit, a log auditing unit, and a permission control unit. Among them, the tag management unit is used to build and maintain a multi-dimensional entity and content tag system; the process approval unit is used to realize the process-oriented approval of system business processing; the result export unit is used to realize the generation and export management of analysis reports and business data; the log auditing unit is used to realize the recording and auditing of the entire system operation; and the permission control unit is used to realize role-based hierarchical access control.

[0093] Those skilled in the art should understand that, unless otherwise specified, the meanings of the technical and scientific terms used herein are consistent with the general understanding of the relevant technical field. Furthermore, terms defined in general dictionaries should be understood in the context of the technical background in this field and should not be interpreted in an overly idealized or formalistic manner divorced from practical application scenarios.

[0094] The above embodiments have described in detail the main concept, technical solution, and technical effects of the present invention. It should be noted that the above content is merely illustrative and not intended to limit the scope of protection of the present invention. Any equivalent modifications, substitutions, or optimizations based on the present invention without departing from its core principles are within the scope of the present invention.

Claims

1. A decision-oriented knowledge representation and reasoning processing method based on multi-source data, characterized in that, Includes the following steps: Step S1: Utilize a multimodal data warehouse to uniformly access structured data, semi-structured data, and unstructured data; Step S2: Preprocess and extract core features from the multi-source data to construct a multi-level, multi-granularity data semantic mapping model, thereby realizing the automatic transformation and fusion of data into knowledge; the multi-level refers to the hierarchical mapping from the data layer, feature layer to the knowledge layer, and the multi-granularity refers to the realization of fine-grained, medium-grained, and coarse-grained semantic expression and calculation in each level. Step S3: Introduce a knowledge representation mechanism oriented towards decision-making objectives, embedding business rules, constraints, and decision preferences into the knowledge structure in a computable form to form a knowledge representation system with a three-layer heterogeneous hypergraph structure; The knowledge representation system will model data facts, business logic, and decision-making objectives in a hierarchical and unified manner. Step S4: Based on the knowledge representation system, construct a hybrid reasoning process that combines symbolic reasoning and data-driven reasoning to deduce, verify, and evaluate the results of decision-related issues; Step S5: Output interpretable and traceable decision support conclusions.

2. The decision-oriented knowledge representation and reasoning processing method based on multi-source data according to claim 1, characterized in that, The multimodal data warehouse in step S1 further includes: Step S11: Build a unified data access layer to connect with data sources that carry structured, semi-structured, and unstructured data; Step S12: Based on the original characteristics of the three types of data, a unified data integration platform is used to collect the three types of data in real time. During the collection process, basic format verification is performed on the three types of data respectively, and the original characteristics of each type of data are preserved throughout the process without any business processing. Step S13: Based on the original characteristics of structured data, semi-structured data, and unstructured data, the three types of data that have completed verification are stored in matching storage engines respectively. Structured data is stored in a relational database, semi-structured data is stored in a document database, and unstructured data is stored in object storage and a distributed file system. Each engine achieves linkage query through a unified interface, realizing unified original storage of the three types of data in the multimodal data warehouse, completing the overall access, and constructing the multimodal data warehouse.

3. The decision-oriented knowledge representation and reasoning processing method based on multi-source data according to claim 1, characterized in that, The preprocessing in step S2 includes data deduplication, missing value completion, format standardization and normalization operations. The missing value completion uses a time series data interpolation algorithm to complete time series data, uses an image generation model to complete missing image data, and uses a feature selection algorithm to suppress feature redundancy. The core feature extraction includes: for unstructured data such as videos and texts, using deep neural networks combined with time-series modeling to extract spatiotemporal and semantic features; for structured data such as numerical and time-series signals, using a joint frequency domain and time domain feature extraction model to extract multi-scale features; and for qualitative data such as natural language, converting it into quantitative features through language tuples combined with membership degree calculation.

4. The decision-oriented knowledge representation and reasoning processing method based on multi-source data according to claim 1, characterized in that, The data semantic mapping model in step S2 further includes: A unified knowledge framework incorporating domain ontology models and business mechanism knowledge is constructed. The metadata structure of multi-source heterogeneous data is parsed, and entity alignment and attribute association rules between source data feature items and the unified knowledge framework are established. Based on the association rules, a semantic mapping algorithm is used to perform alignment calculations to realize the mapping between source data features and the knowledge framework, maintaining the consistency of semantic topology. Multi-feature association metrics and logical transformation operators are used to resolve heterogeneous semantic conflicts. During the mapping process, missing data is filled in by combining feature representation models with interpolation methods to achieve the fusion of cross-type data.

5. The decision-oriented knowledge representation and reasoning processing method based on multi-source data according to claim 1, characterized in that, In step S3, the business rules are configured as domain-specific decision logic. The constraints include data time range and applicable object restrictions. The decision preferences include priority setting rules. The computable form includes logical encoding representing business rules, constraint equations and penalty functions that limit decision boundaries, weight matrices and evaluation functions that quantify decision preferences, and probability distributions and membership parameters for handling uncertain reasoning.

6. The decision-oriented knowledge representation and reasoning processing method based on multi-source data according to claim 1, characterized in that, The knowledge representation system described in step S3 further includes: The knowledge representation system is a three-layer heterogeneous hypergraph structure, including a bottom-layer fact subgraph, a middle-layer logic subgraph, a top-layer target subgraph, and an inter-layer association mechanism. The hypergraph structure allows hyperedges to connect multiple heterogeneous nodes. The inter-layer association mechanism includes condition-triggered edges connecting fact nodes and rule nodes, and decision contribution weight edges connecting rule nodes and target nodes. The weight values ​​are set and dynamically updated based on the analytic hierarchy process (AHP) combined with business experience. The bottom-layer fact subgraph consists of entity nodes and attribute edges generated by the data semantic mapping model, used to represent the objective business state of multi-source data. The middle-layer logic subgraph consists of rule nodes, constraint operator nodes, and logical operation edges, used to represent business judgment criteria and compliance verification logic. The top-layer target subgraph consists of decision target nodes, risk level nodes, and utility function nodes, used to represent the final output orientation of reasoning. The inter-layer association mechanism includes condition-triggered edges connecting fact nodes and rule nodes, and decision contribution weight edges connecting rule nodes and target nodes.

7. The decision-oriented knowledge representation and reasoning processing method based on multi-source data according to claim 1, characterized in that, The symbolic reasoning described in step S4 is based on preset business rules and a knowledge representation system for logical deduction. Specifically, it includes decision-related logical operations; monitoring changes in entity states in the underlying fact subgraph; automatically triggering associated rule nodes in the middle-level logical subgraph through a pattern matching mechanism; employing a forward linking reasoning strategy to perform transitive closure calculations along logical operation edges; verifying whether the current fact satisfies the constraints of the constraint operator nodes; generating a logical blocking signal and marking the evidence subgraph that caused the conflict when the reasoning path encounters mutually exclusive business rules or violates rigid constraint nodes; if the verification passes, generating a definite logical judgment result and outputting a symbolic execution path containing a complete triggering rule chain as the decision. The logical basis for the explanation; the data-driven reasoning specifically includes using a graph neural network model to map discrete nodes and edges in the underlying fact subgraph to high-dimensional continuous vectors, generating relation embedding vectors based on knowledge representation learning algorithms, calculating similarity metrics based on the embedding vectors, predicting undefined potential association edges between fact nodes, realizing link prediction for hidden risk groups and related transactions, inputting target entity vectors aggregating neighborhood features into a pre-trained deep learning classifier, calculating the confidence probability distribution of their belonging to nodes of different risk levels in the top-level target subgraph, and finally, based on the newly incoming data flow, fine-tuning the model parameters in real time through an incremental learning mechanism to update the predicted scores of potential risk points in real time.

8. The decision-oriented knowledge representation and reasoning processing method based on multi-source data according to claim 1, characterized in that, The hybrid inference processing flow described in step S4 includes cross-validation, conflict resolution, and consistency verification steps for symbolic inference results and data-driven inference results. When the symbolic inference encounters rigid business constraints and generates a logic blocking signal, the symbolic inference is given absolute priority, and the blocking signal is directly used as the consistency verification result, thus rejecting the prediction output of data-driven inference.

9. The decision-oriented knowledge representation and reasoning processing method based on multi-source data according to claim 1, characterized in that, The interpretable and traceable decision support conclusions in step S5 include a logical chain of reasoning, data source annotations, and explanations of the basis for the conclusions. The derivation process of the decision conclusions is clearly defined, and the conclusions are presented in the form of a hierarchical path diagram of fact nodes, rule nodes, and target nodes.

10. A decision-oriented knowledge representation and reasoning processing system based on multi-source data, characterized in that, The dynamic system is applied to the decision-oriented knowledge representation and reasoning processing method based on multi-source data as described in any one of claims 1-9, specifically including a data access module, an intelligent analysis module, a collaborative application module, and a basic control module: The data access module is used to realize the batch access, management, and automated parsing of multimodal and multi-format data. The data access module includes a file access unit, a communication data access unit, and a terminal device data access unit. Specifically, the file access unit is used to realize the batch import, parsing monitoring, and sharing management of multi-format documents and multimedia files; the communication data access unit is used to realize the batch extraction, basic information parsing, and status monitoring of network communication data; and the terminal device data access unit is used to realize the collection and standardized parsing of terminal device operating status data and network connection data. The intelligent analysis module, communicatively connected to the data access module, is used for in-depth analysis and information mining of standardized and parsed data. The intelligent analysis module includes a semantic and behavioral analysis unit, an entity profiling analysis unit, a relationship network construction unit, a spatiotemporal geographic analysis unit, and a multilingual processing unit. Specifically, the semantic and behavioral analysis unit performs semantic clustering, behavioral feature statistics, and AI-based content summarization of communication data; the entity profiling analysis unit constructs multi-dimensional device and network feature profiles to analyze and display relationships and behavioral trajectories; the relationship network construction unit constructs multi-level interaction relationship graphs of target entities based on multi-source data; the spatiotemporal geographic analysis unit performs target spatial location retrieval, cluster feature heatmap analysis, and temporal trajectory reconstruction; and the multilingual processing unit performs bidirectional translation processing of cross-language text. The collaborative application module is bidirectionally connected to the intelligent analysis module to realize collaborative data analysis, thematic management, and analysis model construction. The collaborative application module is equipped with a collaborative analysis unit and a data modeling and analysis unit. The collaborative analysis unit is used to create business topics, conduct multi-user collaborative review and annotation, and perform in-depth analysis and summarization of analysis content based on artificial intelligence. The data modeling and analysis unit is used to manage data sources, perform visual correlation modeling, schedule data tasks, and manage model processing resources, supporting the visual construction and reuse of business decision-making models. The basic control module is communicatively connected to the data access module, intelligent analysis module, and collaborative application module, respectively, and is used to realize system process control, permission configuration, and security auditing. The basic control module is equipped with a tag management unit, a process approval unit, a result export unit, a log auditing unit, and a permission control unit. Among them, the tag management unit is used to build and maintain a multi-dimensional entity and content tag system; the process approval unit is used to realize the process-oriented approval of system business processing; the result export unit is used to realize the generation and export management of analysis reports and business data; the log auditing unit is used to realize the recording and auditing of the entire system operation; and the permission control unit is used to realize role-based hierarchical access control.

Images