Competitive product comparison method based on knowledge graph
By constructing a competitor alignment graph reasoning framework and a relationship combination operator, the problems of inaccurate and unstable competitor comparison results in existing technologies are solved. This enables traceable expression and adaptive propagation of competitor differences, thereby improving the accuracy and stability of competitor comparison.
Patent Information
- Application Number
- CN202610096797.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-01-23
- Publication Date
- 2026-05-01
AI Technical Summary
Existing knowledge graph-based competitor comparison methods are unable to accurately reflect the structural differences between competitors in multi-dimensional competitor comparison scenarios, and the stability and verifiability of the comparison results are insufficient, lacking adaptive modeling methods.
A competitor alignment graph reasoning framework is constructed, and a relation combination operator is introduced for joint modeling. Through structural decomposition and recombination, a set of difference propagation paths is generated and mapped as a comparison feedback signal to modulate the competitor alignment graph reasoning framework, thereby realizing the traceable expression and adaptive propagation of competitor differences.
It improves the accuracy, consistency, and stability of competitor comparison results, enhances traceability, reduces the risk of result instability, and achieves adaptive propagation and updates under multi-dimensional analysis.
Smart Images

Figure CN121960692A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of enterprise competitive intelligence and business analysis, and in particular to a competitor comparison method based on knowledge graphs. Background Technology
[0002] In practical applications such as competitive intelligence analysis, technology solution review, product selection, and patent landscape research, competitor comparison typically involves a large amount of heterogeneous data, including product structure information, functional attribute information, relational information, and supporting evidence. To improve information organization capabilities, existing technologies have begun to introduce knowledge graphs to structurally model competitor-related information and conduct comparative analysis based on this model. However, most existing knowledge graph-based competitor comparison methods still remain at the level of static feature comparison or simple path retrieval, making it difficult to depict the true mechanisms underlying the differences between competitors under multi-dimensional and multi-relational structures.
[0003] On the one hand, existing technologies typically use a single entity as the propagation target in graph reasoning, lacking a joint modeling mechanism for competitor entity pairs. This makes the alignment propagation results susceptible to graph structure bias, making it difficult to accurately reflect the structural differences between competitors. On the other hand, although some methods introduce path analysis or evidence chain interpretation mechanisms, the relevant paths are mostly generated independently as a means of interpreting the results, failing to form an effective closed loop with the reasoning process. The lack of traceable correlation between path contributions and propagation calculations leads to insufficient stability and verifiability of the comparison results.
[0004] Furthermore, existing methods, in multi-dimensional competitor comparison scenarios, typically rely on manually setting weights or empirical rules to weight different dimensions, lacking an adaptive modeling approach oriented towards relationship structures, and thus struggling to cope with the differentiated propagation characteristics of complex competitor relationships across different comparison dimensions.
[0005] Therefore, how to provide a competitive comparison method based on knowledge graphs is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention
[0006] One objective of this invention is to propose a competitor comparison method based on a knowledge graph. This invention uses a competitor comparison knowledge graph as its foundation, constructs a competitor entity pair input structure, introduces a competitor alignment graph inference framework with competitor entity pairs as the propagation object, and utilizes relation combination operators to achieve joint modeling of competitor entity node representations and relation edge representations within this framework. Based on this, by performing structural decomposition and recombination on the relation combination operators, and combining the comparison dimension set to generate dimensional convolution configurations, dimensional alignment convolution propagation under different comparison dimensions is achieved. Simultaneously, this invention generates a set of difference propagation paths based on propagation records and maps them to comparison feedback signals, performing feedback modulation on the competitor alignment graph inference framework, ultimately generating competitor comparison representation results and competitor comparison conclusions. This method achieves structured modeling and traceable expression of the competitor difference formation process, improving the stability and consistency of competitor comparison results.
[0007] A competitive product comparison method based on knowledge graphs according to an embodiment of the present invention includes the following steps:
[0008] Acquire multi-source structured data for competitor comparison, construct a competitor comparison knowledge graph, determine the first and second competitor entities to be compared, and construct the competitor entity pair input structure;
[0009] Using the CompGCN model as the underlying graph convolutional inference model, a competitor alignment graph inference framework is constructed. Combining the competitor comparison knowledge graph and the competitor entity pair input structure, a propagation record is generated. At the same time, a relation combination operator is constructed based on the CompGCN model.
[0010] Structural decomposition is performed on the relation combination operator in the competitor alignment graph inference framework. The relation combination operator is restructured and a dimensional convolution configuration is generated. Dimensional alignment convolution propagation is performed to generate dimensional competitor representation results.
[0011] Based on the dimensional competitor representation results and propagation records, perform path backtracking and contribution accumulation processing to generate a set of differential propagation paths;
[0012] The set of difference propagation paths is mapped to a contrast feedback signal, and the competitor alignment graph inference framework is modulated based on the contrast feedback signal to obtain the modulated competitor alignment graph inference framework.
[0013] The competitor entity pair input structure is fed into the modulated competitor alignment graph inference framework, and feedback alignment convolution propagation is performed to generate competitor comparison representation results.
[0014] Based on the competitor comparison results, dimensional aggregation and path mapping are performed to generate competitor comparison conclusions.
[0015] Optionally, the generation of the competitor comparison knowledge graph and competitor entity pairs input structure includes:
[0016] Acquire multi-source structured data for competitor comparison and standardize it according to competitor identification rules;
[0017] Based on the standardized multi-source structured data of competitor comparison, a competitor comparison knowledge graph is constructed. Competitor entities are mapped to competitor entity nodes, structural attributes directly associated with competitor entities are mapped to attribute nodes, deterministic associations between competitor entity nodes and attribute nodes are mapped to relationship edges, and structured proof information representing the source and validity of relationships is mapped to evidence identifiers.
[0018] In the process of constructing a competitor comparison knowledge graph, the relationship edges are restricted to a set of relationship types that can participate in competitor alignment propagation;
[0019] Based on the competitor comparison knowledge graph, the first and second competitor entities to be compared are determined, and the competitor entity pair input structure is constructed as the alignment object;
[0020] When constructing the competitor entity pair input structure, based on the first competitor entity to be compared and the second competitor entity to be compared, entity uniqueness encoding processing is performed to generate corresponding competitor entity identifiers, which serve as entity identifier elements of the competitor entity pair input structure.
[0021] Based on the relationship distribution characteristics between the first and second competitor entities to be compared, dimensional statistics and association mapping are performed to determine the set of comparison dimensions.
[0022] Based on the alignment requirements of the first and second competitor entities to be compared, alignment constraint analysis and mapping consistency judgment are performed on the relationship correspondence structure of the two entities to generate subgraph extraction rules.
[0023] Optionally, the generation of the propagation record includes:
[0024] Based on the competitor comparison knowledge graph and competitor entity pair input structure, and using the CompGCN model as the underlying graph convolutional inference model, a competitor alignment graph inference framework is constructed, and entity alignment encoding unit, alignment propagation unit, and entity pair representation update unit are set up.
[0025] In the competitor alignment graph reasoning framework, based on the CompGCN model, we define the joint mapping calculation rules for competitor entity node representation and relation edge representation, and construct the relation combination operator.
[0026] The alignment propagation unit extracts the first competitor entity adjacency relationship subgraph and the second competitor entity adjacency relationship subgraph from the competitor comparison knowledge graph based on the subgraph extraction rules, and uses them as the structural input for the alignment propagation of competitor entities.
[0027] The entity alignment encoding unit encodes the first competitor entity adjacency relationship subgraph and the second competitor entity adjacency relationship subgraph into an initial alignment representation;
[0028] The entity pair representation update unit performs baseline-aligned convolutional propagation based on the initial alignment representation, records the relation edges and propagation levels involved in the propagation, and generates a propagation record, including the propagation edge index and the propagation level index.
[0029] Optionally, the generation of the dimensional competitor representation results includes:
[0030] In the competitor alignment graph reasoning framework, the relation composition operator is structurally decomposed into three sub-operator structures: semantic sub-operator, dependency sub-operator, and constraint influence sub-operator.
[0031] Based on the set of contrast dimensions, the execution sequence of the calling order of semantic sub-operators, dependency sub-operators and constraint influence sub-operators is processed to generate a relation combination operator reorganization structure. Specifically, for each contrast dimension in the set of contrast dimensions, a relation combination operator reorganization structure uniquely corresponding to that contrast dimension is constructed.
[0032] While constructing the relation combination operator recombination structure, a corresponding set of dimension gating parameters is generated for each pair of relation combination operator recombination structures based on the propagation record;
[0033] Generate dimensional convolution configurations by recombining the structure and dimensional gating parameter set based on relational combination operators;
[0034] Based on the dimensional convolution configuration, dimensional alignment convolution propagation processing is performed to generate dimensional competitive product representation results.
[0035] Optionally, the generation of the differential propagation path set includes:
[0036] Perform reverse traversal processing on the propagation level index and propagation edge index in the propagation record to determine the candidate set of propagation paths to participate in the update during the dimension-aligned convolution propagation process;
[0037] For each path in the candidate propagation path set, perform consistency and continuity checks on the propagation hierarchy order in the path, and retain the propagation paths that satisfy hierarchy continuity and edge connection continuity.
[0038] Based on the dimensional representation results of competitors, path contribution decomposition is performed on the changes in the dimensional representation of each propagation node in the retained propagation path to determine the local contribution components.
[0039] Based on the propagation record and the set of dimensional gating parameters, a weighted accumulation process is performed on the local contribution components of each propagation path to obtain the path contribution value;
[0040] Based on the path contribution value, the propagation paths are filtered for differences in effectiveness, and a set of differential propagation paths is generated.
[0041] Optionally, the modulation of the competitor alignment graph inference framework includes:
[0042] Based on the set of differential propagation paths, normalization mapping is performed on the path contribution value corresponding to each differential propagation path to determine the subset of propagation paths;
[0043] A path selection mask is generated by performing mask mapping on a subset of propagation paths according to the propagation edge index and the propagation level index.
[0044] Based on the path contribution value corresponding to each differential propagation path and the participation of sub-operators in the relation sequence in the path, gating mapping processing is performed to generate a set of feedback gating parameters;
[0045] Based on the distribution characteristics of each differential propagation path on the propagation level index, hierarchical summarization and normalization are performed on the path contribution values corresponding to different propagation levels to generate propagation gain parameters.
[0046] The path selection mask is applied to the alignment propagation unit in the competitor alignment graph inference framework. The set of feedback gating parameters and the propagation gain parameters are applied to the relation combination operator recombination structure and the alignment propagation unit in the competitor alignment graph inference framework, respectively. The sub-operator execution order, sub-operator participation state and propagation strength are modulated to generate the modulated competitor alignment graph inference framework.
[0047] Optionally, the generation of the competitor comparison results includes:
[0048] The competitor entity pair input structure is input into the modulated competitor alignment graph inference framework, the alignment propagation state is initialized, the range of propagation paths that can participate in the propagation update is limited based on the path selection mask, feedback alignment convolution propagation is performed, and the structure is recombined according to the modulated relation combination operator for propagation update.
[0049] During the execution of feedback-aligned convolutional propagation, the participation states of semantic sub-operators, dependent sub-operators, and constraint influence sub-operators are constrained based on the set of feedback gating parameters, and the propagation intensity corresponding to different propagation levels is modulated based on the propagation gain parameter to obtain the propagation result after feedback modulation.
[0050] Based on the propagation results after feedback modulation, the differences in the representation of competitor entity nodes formed under each comparison dimension are extracted, and a dimension difference vector is generated.
[0051] Based on the propagation path that actually participates in the propagation update during the feedback aligned convolution propagation process, and combined with the propagation edge index and propagation level index in the propagation record, the difference path index is determined.
[0052] The dimensional difference vector and the difference path index are combined to represent the competitive comparison results.
[0053] Optionally, the generation of the competitor comparison conclusion includes:
[0054] Based on the competitor comparison results, read the dimension difference vector, perform dimension alignment processing, and generate a set of dimension difference components;
[0055] Based on the set of dimensional difference components, dimensional aggregation processing is performed to obtain the difference score;
[0056] Based on the difference path index in the competitor comparison representation result, a path mapping process is performed on the difference propagation path set to map it into the corresponding difference propagation path sequence.
[0057] Based on the correspondence between the difference path index and the set of dimensional difference components, the difference score under each comparison dimension is bound to the corresponding difference propagation path sequence to generate a dimensional difference path mapping result.
[0058] Based on the dimensional difference path mapping results, a competitor comparison conclusion is generated, which includes the difference score corresponding to each comparison dimension in the comparison dimension set and the path sequence of the difference propagation path set bound to it.
[0059] The beneficial effects of this invention are:
[0060] First, this invention constructs a competitor alignment graph reasoning framework with competitor entity pairs as the propagation object, and introduces a relation combination operator and its structural decomposition and recombination mechanism into this framework. This enables the joint modeling of competitor entity node representations and relation edge representations in the same propagation process, thereby achieving a refined alignment expression of competitors under a multi-relation structure. This avoids the structural bias problem caused by propagation centered on a single entity in the prior art, and significantly improves the accuracy and consistency of competitor comparison results at the structural level.
[0061] Secondly, by generating propagation records during the alignment propagation process and constructing a set of difference propagation paths based on these records, the present invention enables the formation process of competitor differences to be fully traced back and quantified. The relevant differences are no longer just output results, but are established with the propagation process in the form of path contribution values, thereby enhancing the traceability and verifiability of competitor comparison results and reducing the risk of unstable results caused by unexplainable paths in comparative analysis.
[0062] Furthermore, this invention further maps the set of difference propagation paths into contrast feedback signals, and implements feedback modulation on the alignment propagation unit and relation combination operator recombination structure in the competitor alignment graph inference framework. This enables the subsequent feedback alignment convolution propagation to be specifically constrained at the structural level, thereby achieving difference-aware driven adaptive propagation update in multi-contrast dimension scenarios and improving the stability and consistency of competitor comparison representation results under multi-dimensional analysis. Attached Figure Description
[0063] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:
[0064] Figure 1 This is an overall flowchart of a knowledge graph-based competitor comparison method proposed in this invention.
[0065] Figure 2 This is a schematic diagram of the competitor alignment graph reasoning framework and relation combination operator structure in this invention;
[0066] Figure 3 This is a schematic diagram of the process of generating the differential propagation path set and modulating the comparison feedback signal in this invention. Detailed Implementation
[0067] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.
[0068] refer to Figure 1-3 A knowledge graph-based competitive analysis method includes the following steps:
[0069] Acquire multi-source structured data for competitor comparison, construct a competitor comparison knowledge graph based on the multi-source structured data for competitor comparison, the competitor comparison knowledge graph consists of competitor entity nodes, attribute nodes, relationship edges, and evidence identifiers, determine the first competitor entity to be compared and the second competitor entity to be compared based on the competitor comparison knowledge graph, construct a competitor entity pair input structure, the competitor entity pair input structure includes the first competitor entity identifier, the second competitor entity identifier, the comparison dimension set, and the subgraph extraction rules;
[0070] Based on the competitor comparison knowledge graph and competitor entity pair input structure, a competitor alignment graph inference framework is constructed using the CompGCN model as the underlying graph convolutional inference model. Furthermore, a relation combination operator for joint entity and relation modeling is built based on the CompGCN model. The competitor alignment graph inference framework includes an entity alignment encoding unit, an alignment propagation unit, and an entity pair representation update unit. Specifically, the alignment propagation unit extracts a first competitor entity adjacency relation subgraph and a second competitor entity adjacency relation subgraph based on subgraph extraction rules. The entity alignment encoding unit encodes the first and second competitor entity adjacency relation subgraphs into initial alignment representations. The entity pair representation update unit performs baseline alignment convolutional propagation based on the initial alignment representations and generates propagation records, which include propagation edge indices and propagation level indices.
[0071] Structural decomposition is performed on the relation combination operator in the competitor alignment graph inference framework, decomposing it into three sub-operator structures: semantic sub-operators, dependency sub-operators, and constraint influence sub-operators. Based on the three sub-operator structures and the comparison dimension set, a relation combination operator recombination structure is constructed and a dimensional convolution configuration is generated. The dimensional convolution configuration is a unique sub-operator recombination sequence and a set of dimensional gating parameters for each comparison dimension configuration in the comparison dimension set. Based on the dimensional convolution configuration, dimensional alignment convolution propagation is performed in the competitor alignment graph inference framework to generate dimensionalized competitor representation results.
[0072] Based on the dimensional competitor representation results and propagation records, a set of differential propagation paths is generated by performing path backtracking and contribution accumulation processing on the dimensional competitor representation results and propagation records. Each differential propagation path in the set of differential propagation paths consists of a starting node identifier, a relation sequence, an ending node identifier, and a path contribution value.
[0073] The set of differential propagation paths is mapped to a contrast feedback signal, which includes a path selection mask, a set of feedback gating parameters, and a propagation gain parameter. The alignment propagation unit and relation combination operator in the competitor alignment graph inference framework are modulated based on the contrast feedback signal to obtain the modulated competitor alignment graph inference framework.
[0074] The competitor entity pair input structure is input into the modulated competitor alignment graph inference framework. Feedback alignment convolution propagation is performed in the modulated competitor alignment graph inference framework to generate competitor comparison representation results. The competitor comparison representation results include dimension difference vector and difference path index.
[0075] Based on the competitor comparison results, a competitor comparison conclusion is generated by performing dimension aggregation and path mapping on the dimension difference vector and difference path index. The competitor comparison conclusion includes the difference score of each comparison dimension in the comparison dimension set and the path sequence of the corresponding difference propagation path set.
[0076] In this embodiment, the generation of the competitor comparison knowledge graph and competitor entity pairs input structure includes:
[0077] Obtain multi-source structured data for competitor comparison, and standardize the multi-source structured data for competitor comparison according to competitor identification rules. The multi-source structured data for competitor comparison is used to characterize the structural attribute information, relational information and evidence association information of competitors.
[0078] A competitor comparison knowledge graph is constructed based on standardized multi-source structured data for competitor comparison. Competitor entities are mapped to competitor entity nodes, structural attributes directly associated with competitor entities are mapped to attribute nodes, deterministic associations between competitor entity nodes and attribute nodes are mapped to relationship edges, and structured proof information used to characterize the source and validity of relationships is mapped to evidence identifiers. The evidence identifiers are configured as propagation control elements bound to relationship edges in the competitor comparison knowledge graph. They are used to participate in propagation reachability determination in the subsequent competitor alignment graph reasoning framework and serve as constraints for determining whether relationship edges can be included in the propagation path when generating propagation records and performing path backtracking processing. Evidence identifiers are not used as static annotation fields.
[0079] In the process of constructing a competitor comparison knowledge graph, the relationship edges are type-limited. The relationship edges are limited to a set of relationship types that can participate in competitor alignment propagation. The relationship edges in the set of relationship types are used to support alignment propagation and dimension alignment convolution propagation in the subsequent competitor alignment graph inference framework.
[0080] Based on the competitor comparison knowledge graph, the first competitor entity to be compared and the second competitor entity to be compared are determined, and the first competitor entity to be compared and the second competitor entity to be compared are used as alignment objects to construct the competitor entity pair input structure;
[0081] When constructing the competitor entity pair input structure, based on the first competitor entity to be compared and the second competitor entity to be compared, entity uniqueness encoding processing is performed to generate corresponding competitor entity identifiers, and the first competitor entity identifier and the second competitor entity identifier are used as entity identifier elements of the competitor entity pair input structure;
[0082] Based on the relationship distribution characteristics between the first and second competitor entities to be compared in the competitor comparison knowledge graph, dimensional statistics and association mapping are performed to determine the comparison dimension set. The comparison dimension set is automatically determined by the relationship distribution characteristics between the first and second competitor entities to be compared in the competitor comparison knowledge graph. The determination process of the comparison dimensions does not rely on manually preset classification labels, but is formed based on the distribution of relationship types, relationship hierarchy structure and evidence identifier distribution, which is used to limit the dimension alignment convolution propagation mode in the subsequent competitor alignment graph reasoning framework.
[0083] Based on the alignment requirements of the first and second competitor entities to be compared, alignment constraint analysis and mapping consistency determination are performed on the relationship correspondence structure of the two entities to generate subgraph extraction rules. The subgraph extraction rules are bound to the competitor entity input structure. The subgraph extraction rules are used to extract the first competitor entity adjacency relationship subgraph and the second competitor entity adjacency relationship subgraph for alignment propagation from the competitor comparison knowledge graph.
[0084] In this embodiment, the generation of the propagation record includes:
[0085] Based on the competitor comparison knowledge graph and competitor entity pair input structure, a competitor alignment graph inference framework is constructed using the CompGCN model as the underlying graph convolutional inference model. The graph convolutional propagation object of the CompGCN model is limited to competitor entity pairs, used to jointly represent the structural alignment state of the first and second competitor entities to be compared during the same propagation process. The competitor alignment graph inference framework includes an entity alignment encoding unit, an alignment propagation unit, and an entity pair representation update unit. The entity alignment encoding unit performs alignment encoding processing on the structural representation of competitor entity pairs, the alignment propagation unit performs graph convolutional propagation under relational constraints at the competitor entity pair level, and the entity pair representation update unit updates the alignment representation of competitor entity pairs during the propagation process.
[0086] In the competitor alignment graph reasoning framework, based on the CompGCN model, a joint mapping calculation rule for competitor entity node representation and relation edge representation is defined, and a relation combination operator for joint modeling of entities and relations is constructed. The relation combination operator is called by the alignment propagation unit when performing alignment propagation calculation of competitor entity pairs. It is used to perform joint mapping processing on competitor entity node representation and relation edge representation during the alignment propagation calculation process. The relation combination operator is configured with a decomposable structure to support subsequent structural decomposition of the relation combination operator and construction of relation combination operator reorganization structure.
[0087] The relation combination operator is used only as a computational component in the alignment propagation process within the competitor alignment graph reasoning framework. Its output is not stored as an independent entity representation or relation representation. The existence of the relation combination operator is centered on the propagation computation rules, and is used to support the synchronization constraints between entity node representations and relation edge representations during the alignment propagation process.
[0088] In the competitor alignment graph reasoning framework, the alignment propagation unit extracts the first competitor entity adjacency relationship subgraph and the second competitor entity adjacency relationship subgraph from the competitor comparison knowledge graph based on the subgraph extraction rules in the competitor entity pair input structure. The adjacency relationship subgraph serves as the structural input for competitor entity pair alignment propagation.
[0089] In the competitor alignment graph reasoning framework, the entity alignment encoding unit encodes the first competitor entity adjacency relationship subgraph and the second competitor entity adjacency relationship subgraph into an initial alignment representation, which is used to characterize the initial alignment state of competitor entity pairs at the relation structure level.
[0090] In the competitor alignment graph reasoning framework, the entity pair representation update unit performs baseline alignment convolution propagation based on the initial alignment representation. During the baseline alignment convolution propagation process, the relation edges and propagation levels involved in the propagation are recorded to generate a propagation record. The propagation record includes a propagation edge index and a propagation level index. The propagation record serves as the structural input for the subsequent generation of the difference propagation path set and path backtracking processing.
[0091] The propagation record is saved as an immutable structure after it is generated. Its propagation edge index and propagation level index are used to limit the range of legal paths in subsequent path backtracking, path contribution accumulation and feedback modulation processes. The propagation record does not participate in the model training parameter update, but is only used for the traceability control of the propagation structure.
[0092] In this embodiment, the generation of the dimensional competitor representation result includes:
[0093] In the competitor alignment graph reasoning framework, the relation composition operator is structurally decomposed into three sub-operator structures: semantic sub-operator, dependency sub-operator, and constraint influence sub-operator. The semantic sub-operator is configured to model the semantic consistency of relation edges to determine the alignment direction of relation semantics in the alignment propagation of competitor entity pairs. The dependency sub-operator is configured to model relation dependencies across propagation levels to determine the dependency propagation relationship between competitor entity node representations at different propagation levels. The constraint influence sub-operator is configured to impose propagation restrictions on incomparable relation paths to suppress propagation participation that does not meet the alignment conditions.
[0094] The semantic sub-operators, dependency sub-operators, and constraint influence sub-operators are structurally independent of each other. Their division is based on the difference in propagation responsibilities rather than mathematical decomposition. The three sub-operators correspond to semantic consistency constraints, cross-level propagation dependency constraints, and propagation path feasibility constraints, respectively.
[0095] After completing the structural decomposition of relational composition operators, the execution sequence of the calling order of semantic sub-operators, dependency sub-operators, and constraint influence sub-operators is constructed based on the set of comparison dimensions to generate a relational composition operator reorganization structure. The relational composition operator reorganization structure is limited to a structural sequence that arranges the execution order of sub-operator structures, rather than numerically weighting the sub-operator output results. For each comparison dimension in the set of comparison dimensions, a relational composition operator reorganization structure uniquely corresponding to that comparison dimension is constructed, so that different comparison dimensions correspond to different sub-operator reorganization sequence structures, thereby realizing the structural differences in relational semantic modeling, relational dependency modeling, and constraint influence modeling under different comparison dimensions.
[0096] The relation combination operator recombination structure does not introduce new trainable parameters during the construction process. Its difference is only reflected in the execution order of sub-operators and the combination method of structure. The propagation difference corresponding to different comparison dimensions is generated by structural differences, rather than by changes in numerical weights.
[0097] While constructing the relation combination operator recombination structure, based on the propagation record generated in the competitor alignment graph reasoning framework, a corresponding set of dimension gate parameters is generated for each pair of relation combination operator recombination structures. The set of dimension gate parameters is used to limit the propagation participation state of the sub-operator structure under the corresponding comparison dimension, and establishes a mapping relationship with the propagation edge index and the propagation level index.
[0098] The specific steps for generating the dimensional gating parameter set are as follows: Based on the propagation edge index and propagation level index in the propagation record, perform structural expansion processing on the sub-operator recombination sequence corresponding to each pair of relation combination operator recombination structures to obtain the dimensional propagation constraints for each comparison dimension in the comparison dimension set; based on the dimensional propagation constraints, perform relation type consistency determination and evidence identification validity determination on the relation edges corresponding to the propagation edge index; based on the determination results, assign a propagation allowed state or a propagation inhibited state to each propagation edge index under the corresponding comparison dimension, generate propagation edge gating parameters; based on the dimensional propagation constraints, perform a level reachability determination on the propagation level corresponding to the propagation level index, and based on the determination... The result is that each propagation level index is assigned a level participation state or a level suppression state under the corresponding contrast dimension, forming propagation level gating parameters. Based on the propagation edge gating parameters and the propagation level gating parameters, combined with the structural responsibilities of semantic sub-operators, dependency sub-operators, and constraint influence sub-operators, the effective propagation participation of each sub-operator under the corresponding contrast dimension is summarized to generate sub-operator gating parameters. The propagation edge gating parameters, propagation level gating parameters, and sub-operator gating parameters are combined according to the contrast dimension to form a set of dimension gating parameters. A definite mapping relationship is established between the set of dimension gating parameters and the propagation edge index and the propagation level index, so as to provide subsequent dimension convolution configuration generation and dimension-aligned convolution propagation calls.
[0099] Based on the relation combination operator, the structure is reorganized and the set of dimensional gating parameters is regenerated to generate a dimensional convolution configuration. The dimensional convolution configuration is used to limit the execution order of sub-operators, propagation dependency path and propagation participation constraints of dimensional alignment convolution propagation in the competitor alignment graph inference framework.
[0100] The steps for generating the dimensional convolution configuration include: reorganizing the structure based on relation combination operators; determining the execution order of semantic sub-operators, dependency sub-operators, and constraint-influencing sub-operators under the corresponding contrast dimension to form a sub-operator execution sequence; filtering the relation edges and propagation levels allowed to participate in propagation in the sub-operator execution sequence based on the propagation edge gating parameters and propagation level gating parameters in the dimensional gating parameter set to determine the propagation dependency path under the corresponding contrast dimension; determining the propagation participation constraints of each sub-operator under the corresponding contrast dimension based on the sub-operator execution sequence and propagation dependency path, combined with the sub-operator gating parameters; and combining the sub-operator execution sequence, propagation dependency path, and propagation participation constraints to form a dimensional convolution configuration for limiting the propagation behavior of dimensional alignment convolution.
[0101] Based on the dimensional convolution configuration, dimensional alignment convolution propagation processing is performed in the competitor alignment graph inference framework. The structure is reorganized by combining the corresponding relation combination operator. The competitor entity node representation is propagated and updated under different comparison dimensions to generate dimensional competitor representation results. The dimensional competitor representation results serve as the input basis for subsequent path backtracking and difference propagation path set generation.
[0102] In this embodiment, the generation of the differential propagation path set includes:
[0103] Based on the dimensionality-based competitor representation results and propagation records, reverse traversal processing is performed on the propagation level index and propagation edge index in the propagation records to determine the candidate set of propagation paths to participate in the update during the dimension-aligned convolution propagation process;
[0104] The reverse traversal process is only performed within the scope of the propagation edge index and propagation level index defined by the propagation record, and traversal is not performed on the non-participating propagation areas of the competitor comparison knowledge graph, so as to ensure that the set of different propagation paths strictly corresponds to the actual propagation process;
[0105] For each path in the candidate propagation path set, the consistency of the propagation hierarchy order in the path is checked according to the propagation hierarchy index, and the continuity of the connection relationship of the relation edges in the path is checked according to the propagation edge index. Propagation paths that satisfy the hierarchy continuity and edge connection continuity are retained.
[0106] Based on the dimensional competitor representation results, path contribution decomposition is performed on the dimensional representation changes of each propagation node in the retained propagation path to determine the local contribution component of each propagation path under the corresponding comparison dimension.
[0107] Based on the propagation level index, propagation edge index and dimension gating parameter set in the propagation record, a weighted accumulation process is performed on the local contribution components of each propagation path to obtain the path contribution value of the corresponding propagation path under the corresponding comparison dimension.
[0108] Based on the path contribution value, the propagation path is filtered for difference validity to generate a set of difference propagation paths. Each difference propagation path consists of a starting node identifier, a relationship sequence, an ending node identifier, and a path contribution value. The set of difference propagation paths is output and used as the sole source for constructing the comparison feedback signal. It is not combined with external rule paths or manually selected paths for use in the feedback modulation processing of the competitor alignment graph inference framework.
[0109] In this embodiment, the modulation of the competitor alignment graph inference framework includes:
[0110] Based on the set of differential propagation paths, normalization mapping is performed on the path contribution value corresponding to each differential propagation path to determine the subset of propagation paths with differential discrimination significance in the alignment propagation process under each comparison dimension.
[0111] Based on a subset of propagation paths, a path selection mask is generated by performing a mask mapping process on the subset of propagation paths according to the propagation edge index and the propagation level index. The path selection mask is used to limit the range of propagation paths participating in the propagation update in subsequent alignment propagation units and to shield unselected propagation paths.
[0112] The mask mapping process includes: mapping the relation sequence of each differential propagation path in the propagation path subset to a propagation edge index sequence, mapping the propagation hierarchy distribution of each differential propagation path in the propagation path subset to a propagation hierarchy index sequence, initializing a path selection mask on the complete set of propagation edge indices and the complete set of propagation hierarchy indices defined by the propagation record, setting the propagation edge indices covered by the propagation path subset to an enabled state and setting the uncovered propagation edge indices to a masked state, setting the propagation hierarchy indices covered by the propagation path subset to an enabled state and setting the uncovered propagation hierarchy indices to a masked state, thereby forming a path selection mask for aligning propagation units;
[0113] Based on the path contribution value corresponding to each differential propagation path in the differential propagation path set and the participation status of sub-operators in the relation sequence in the path, gating mapping processing is performed to generate a feedback gating parameter set. The feedback gating parameter set is used to perform feedback modulation on the participation status of semantic sub-operators, dependent sub-operators and constraint influence sub-operators in the relation combination operator recombination structure.
[0114] The gating mapping process specifically involves: based on the set of difference propagation paths, statistically analyzing the participation of semantic sub-operators, dependency sub-operators, and constraint influence sub-operators in each difference propagation path, and recording the path contribution value of the corresponding difference propagation path; based on the statistical results, summarizing the path contribution values of each sub-operator in the difference propagation path to obtain the cumulative contribution of each sub-operator under the corresponding comparison dimension; based on the cumulative contribution of each sub-operator, determining the participation status of semantic sub-operators, dependency sub-operators, and constraint influence sub-operators under the corresponding comparison dimension, determining the feedback participation intensity of each sub-operator, mapping the feedback participation intensity to the gating value of the sub-operator participation status, generating a set of feedback gating parameters, and establishing a corresponding relationship between the set of feedback gating parameters and the corresponding sub-operators in the relation combination operator recombination structure;
[0115] Based on the distribution characteristics of each differential propagation path in the differential propagation path set on the propagation level index, the path contribution values corresponding to different propagation levels are subjected to hierarchical summarization and normalization processing to generate propagation gain parameters. The propagation gain parameters are used to perform differential modulation on the propagation intensity of different propagation levels in the alignment propagation unit.
[0116] The path selection mask is applied to the alignment propagation unit in the competitor alignment graph inference framework to limit the propagation path that can be selected during the alignment propagation process. The set of feedback gating parameters and the propagation gain parameters are applied to the relation combination operator recombination structure and the alignment propagation unit in the competitor alignment graph inference framework, respectively, to modulate the sub-operator execution order, sub-operator participation state and propagation strength, thereby generating the modulated competitor alignment graph inference framework.
[0117] In this embodiment, the generation of the competitor comparison result includes:
[0118] The competitor entity pair input structure is input into the modulated competitor alignment graph inference framework. The modulated competitor alignment graph inference framework initializes the alignment propagation state under the common constraints of path selection mask, feedback gating parameter set and propagation gain parameter. In the modulated competitor alignment graph inference framework, the range of propagation paths that can participate in propagation update is limited based on the path selection mask. Feedback alignment convolution propagation is performed within the limited range, so that the competitor entity node representation is reorganized into a structure according to the modulated relation combination operator in each comparison dimension for propagation update.
[0119] The feedback-aligned convolutional propagation differs from the baseline-aligned convolutional propagation at the propagation structure level. Its propagation path, sub-operator participation state, and propagation level strength all change, resulting in the competitive comparison representation results originating from different propagation structures.
[0120] During the execution of feedback-aligned convolutional propagation, the participation states of semantic sub-operators, dependent sub-operators, and constraint influence sub-operators are constrained based on the set of feedback gating parameters, and the propagation intensity corresponding to different propagation levels is modulated based on the propagation gain parameter to obtain the propagation result after feedback modulation.
[0121] Based on the propagation results after feedback modulation, the differences in the representation of competitor entity nodes formed under each comparison dimension are extracted, and a dimension difference vector is generated.
[0122] Based on the propagation path that actually participates in the propagation update during the feedback aligned convolution propagation process, and combined with the propagation edge index and propagation level index in the propagation record, the difference path index is determined.
[0123] The dimensional difference vector and the difference path index are combined as the competitive comparison representation output, wherein the competitive comparison representation output serves as the input basis for generating subsequent competitive comparison conclusions.
[0124] In this embodiment, the generation of the competitor comparison conclusion includes:
[0125] Based on the competitor comparison results, the dimension difference vector is read, and dimension alignment processing is performed on each dimension component in the dimension difference vector according to the comparison dimension set to generate a dimension difference component set that corresponds one-to-one with the comparison dimension set.
[0126] Based on the set of dimensional difference components, dimensional aggregation processing is performed on the dimensional difference components under each comparison dimension to obtain the difference score of the corresponding comparison dimension. The difference score is obtained entirely by the aggregation of dimensional difference vectors.
[0127] Based on the difference path index in the competitor comparison representation results, a path mapping process is performed on the difference propagation path set to map each difference path index to the corresponding difference propagation path sequence.
[0128] Based on the correspondence between the difference path index and the set of dimensional difference components, the difference score under each comparison dimension is bound to the corresponding difference propagation path sequence to generate a dimensional difference path mapping result.
[0129] Based on the dimensional difference path mapping results, a competitor comparison conclusion is generated. The competitor comparison conclusion includes the difference score corresponding to each comparison dimension in the comparison dimension set and the path sequence of the difference propagation path set bound to it.
[0130] Example 1:
[0131] To verify the feasibility of this invention in practice, it was applied to a competitive product comparison analysis scenario conducted by enterprises during product planning and technology decision-making. In this scenario, enterprises often need to systematically compare multiple candidate product solutions, focusing not only on differences in functionality and performance but also on the structural reasons for these differences and the supporting evidence. However, existing competitive product analysis methods often remain at the level of attribute alignment or overall similarity, failing to answer how these differences are progressively transmitted through relational structures. The analysis results lack traceability, introducing uncertainty into decision-making. This invention addresses this problem by using knowledge graphs and alignment propagation mechanisms to achieve structured and interpretable analysis of competitive product differences.
[0132] In this application scenario, the multi-source structured data for competitor comparison accessed by the system includes product functional module information, technical dependency information, performance indicator association information, and corresponding evidence data. After unified and standardized processing, this data is mapped to competitor entity nodes, attribute nodes, relationship edges, and evidence identifiers, constructing a competitor comparison knowledge graph. Based on this knowledge graph, the system selects a first and a second competitor entity to be compared, constructs a competitor entity pair input structure, and clarifies the comparison dimension set and subgraph extraction rules. Subsequently, using the competitor entity pair as the propagation object, the CompGCN model is introduced to construct a competitor alignment graph inference framework. Within this framework, alignment propagation is performed and propagation records are generated, ensuring that the relationship edges and propagation levels actually involved in the calculation during the propagation process are fully recorded.
[0133] Based on the propagation results, the system performs structural decomposition of the relation combination operator and constructs corresponding relation combination operator recombination structures and dimensional convolution configurations for different comparison dimensions. Under the constraints of the dimensional convolution configuration, dimensional alignment convolution propagation is performed to obtain dimensionalized competitor representation results. Further, the system combines the dimensionalized competitor representation results with the propagation records to perform path backtracking and contribution accumulation processing, generating a set of difference propagation paths, ensuring that each difference propagation path clearly corresponds to a starting node, relation sequence, ending node, and path contribution value. Subsequently, the system maps the set of difference propagation paths to a comparison feedback signal, modulates the competitor alignment graph inference framework, and performs feedback alignment convolution propagation again within the modulated framework, ultimately generating competitor comparison representation results and further forming competitor comparison conclusions.
[0134] To quantitatively evaluate the effectiveness of the method of this invention, it was compared with two common competitive analysis methods: a competitive analysis method based on attribute similarity and a competitive analysis method based on graph embedding. Under the condition of unified competitive entity pairs and comparison dimension set, the accuracy of difference identification, the interpretability and completeness of difference paths, the stability of multi-round analysis results, and the number of effective difference paths were statistically analyzed. The experimental results are shown below.
[0135] Table 1. Comparison of Experimental Results for Different Competitive Product Comparison Methods
[0136] Comparison indicators Attribute similarity-based methods Graph embedding-based methods Method of the present invention Difference recognition accuracy 0.72 0.79 0.85 Differential paths can explain completeness 0.38 0.56 0.83 Stability of multi-round analysis results 0.70 0.76 0.88 Average number of effective difference paths 3.2 5.6 8.9
[0137] As shown in Table 1, attribute similarity-based methods primarily rely on attribute matching results. While simple to implement, they tend to overlook cross-level dependency information in complex structural relationship scenarios, resulting in low interpretability and stability of the difference path. Graph embedding-based methods improve overall difference identification accuracy, but due to the lack of explicit propagation path constraints in the embedding representation itself, their output is difficult to directly correlate with specific relationship sequences, limiting their interpretability. The method presented in this invention improves difference identification accuracy and demonstrates a significant advantage in interpretability of the difference path. This is primarily because the invention binds difference scores to specific propagation paths through propagation records, path backtracking, and contribution accumulation mechanisms, giving the analysis results a clear structural origin. Furthermore, through feedback modulation processing of the competitor alignment graph inference framework, this invention maintains high result stability in multiple rounds of analysis, avoiding inconsistencies in conclusions caused by fluctuations in a single propagation.
[0138] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. A competitive product comparison method based on knowledge graphs, characterized in that, Includes the following steps: Acquire multi-source structured data for competitor comparison, construct a competitor comparison knowledge graph, determine the first and second competitor entities to be compared, and construct the competitor entity pair input structure; Using the CompGCN model as the underlying graph convolutional inference model, a competitor alignment graph inference framework is constructed. Combining the competitor comparison knowledge graph and the competitor entity pair input structure, a propagation record is generated. At the same time, a relation combination operator is constructed based on the CompGCN model. Structural decomposition is performed on the relation combination operator in the competitor alignment graph inference framework. The relation combination operator is restructured and a dimensional convolution configuration is generated. Dimensional alignment convolution propagation is performed to generate dimensional competitor representation results. Based on the dimensional competitor representation results and propagation records, perform path backtracking and contribution accumulation processing to generate a set of differential propagation paths; The set of difference propagation paths is mapped to a contrast feedback signal, and the competitor alignment graph inference framework is modulated based on the contrast feedback signal to obtain the modulated competitor alignment graph inference framework. The competitor entity pair input structure is fed into the modulated competitor alignment graph inference framework, and feedback alignment convolution propagation is performed to generate competitor comparison representation results. Based on the competitor comparison results, dimensional aggregation and path mapping are performed to generate competitor comparison conclusions.
2. The competitive comparison method based on knowledge graphs according to claim 1, characterized in that, The generation of the competitor comparison knowledge graph and competitor entity pairs input structure includes: Acquire multi-source structured data for competitor comparison and standardize it according to competitor identification rules; Based on the standardized multi-source structured data of competitor comparison, a competitor comparison knowledge graph is constructed. Competitor entities are mapped to competitor entity nodes, structural attributes directly associated with competitor entities are mapped to attribute nodes, deterministic associations between competitor entity nodes and attribute nodes are mapped to relationship edges, and structured proof information representing the source and validity of relationships is mapped to evidence identifiers. In the process of constructing a competitor comparison knowledge graph, the relationship edges are restricted to a set of relationship types that can participate in competitor alignment propagation; Based on the competitor comparison knowledge graph, the first and second competitor entities to be compared are determined, and the competitor entity pair input structure is constructed as the alignment object; When constructing the competitor entity pair input structure, based on the first competitor entity to be compared and the second competitor entity to be compared, entity uniqueness encoding processing is performed to generate corresponding competitor entity identifiers, which serve as entity identifier elements of the competitor entity pair input structure. Based on the relationship distribution characteristics between the first and second competitor entities to be compared, dimensional statistics and association mapping are performed to determine the set of comparison dimensions. Based on the alignment requirements of the first and second competitor entities to be compared, alignment constraint analysis and mapping consistency judgment are performed on the relationship correspondence structure of the two entities to generate subgraph extraction rules.
3. The competitive comparison method based on knowledge graphs according to claim 1, characterized in that, The generation of the propagation record includes: Based on the competitor comparison knowledge graph and competitor entity pair input structure, and using the CompGCN model as the underlying graph convolutional inference model, a competitor alignment graph inference framework is constructed, and entity alignment encoding unit, alignment propagation unit, and entity pair representation update unit are set up. In the competitor alignment graph reasoning framework, based on the CompGCN model, we define the joint mapping calculation rules for competitor entity node representation and relation edge representation, and construct the relation combination operator. The alignment propagation unit extracts the first competitor entity adjacency relationship subgraph and the second competitor entity adjacency relationship subgraph from the competitor comparison knowledge graph based on the subgraph extraction rules, and uses them as the structural input for the alignment propagation of competitor entities. The entity alignment encoding unit encodes the first competitor entity adjacency relationship subgraph and the second competitor entity adjacency relationship subgraph into an initial alignment representation; The entity pair representation update unit performs baseline-aligned convolutional propagation based on the initial alignment representation, records the relation edges and propagation levels involved in the propagation, and generates a propagation record, including the propagation edge index and the propagation level index.
4. The competitive comparison method based on knowledge graphs according to claim 1, characterized in that, The generation of the dimensional competitor representation results includes: In the competitor alignment graph reasoning framework, the relation composition operator is structurally decomposed into three sub-operator structures: semantic sub-operator, dependency sub-operator, and constraint influence sub-operator. Based on the set of contrast dimensions, the execution sequence of the calling order of semantic sub-operators, dependency sub-operators and constraint influence sub-operators is processed to generate a relation combination operator reorganization structure. Specifically, for each contrast dimension in the set of contrast dimensions, a relation combination operator reorganization structure uniquely corresponding to that contrast dimension is constructed. While constructing the relation combination operator recombination structure, a corresponding set of dimension gating parameters is generated for each pair of relation combination operator recombination structures based on the propagation record; Generate dimensional convolution configurations by recombining the structure and dimensional gating parameter set based on relational combination operators; Based on the dimensional convolution configuration, dimensional alignment convolution propagation processing is performed to generate dimensional competitive product representation results.
5. The competitive comparison method based on knowledge graphs according to claim 1, characterized in that, The generation of the differential propagation path set includes: Perform reverse traversal processing on the propagation level index and propagation edge index in the propagation record to determine the candidate set of propagation paths to participate in the update during the dimension-aligned convolution propagation process; For each path in the candidate propagation path set, perform consistency and continuity checks on the propagation hierarchy order in the path, and retain the propagation paths that satisfy hierarchy continuity and edge connection continuity. Based on the dimensional representation results of competitors, path contribution decomposition is performed on the changes in the dimensional representation of each propagation node in the retained propagation path to determine the local contribution components. Based on the propagation record and the set of dimensional gating parameters, a weighted accumulation process is performed on the local contribution components of each propagation path to obtain the path contribution value; Based on the path contribution value, the propagation paths are filtered for differences in effectiveness, and a set of differential propagation paths is generated.
6. The competitive comparison method based on knowledge graphs according to claim 1, characterized in that, The modulation of the competitor alignment graph inference framework includes: Based on the set of differential propagation paths, normalization mapping is performed on the path contribution value corresponding to each differential propagation path to determine the subset of propagation paths; A path selection mask is generated by performing mask mapping on a subset of propagation paths according to the propagation edge index and the propagation level index. Based on the path contribution value corresponding to each differential propagation path and the participation of sub-operators in the relation sequence in the path, gating mapping processing is performed to generate a set of feedback gating parameters; Based on the distribution characteristics of each differential propagation path on the propagation level index, hierarchical aggregation and normalization are performed on the path contribution values corresponding to different propagation levels to generate propagation gain parameters. The path selection mask is applied to the alignment propagation unit in the competitor alignment graph inference framework. The set of feedback gating parameters and the propagation gain parameters are applied to the relation combination operator recombination structure and the alignment propagation unit in the competitor alignment graph inference framework, respectively. The sub-operator execution order, sub-operator participation state and propagation strength are modulated to generate the modulated competitor alignment graph inference framework.
7. The competitive comparison method based on knowledge graphs according to claim 1, characterized in that, The generation of the competitor comparison results includes: The competitor entity pair input structure is input into the modulated competitor alignment graph inference framework, the alignment propagation state is initialized, the range of propagation paths that can participate in the propagation update is limited based on the path selection mask, feedback alignment convolution propagation is performed, and the structure is recombined according to the modulated relation combination operator for propagation update. During the execution of feedback-aligned convolutional propagation, the participation states of semantic sub-operators, dependent sub-operators, and constraint influence sub-operators are constrained based on the set of feedback gating parameters, and the propagation intensity corresponding to different propagation levels is modulated based on the propagation gain parameter to obtain the propagation result after feedback modulation. Based on the propagation results after feedback modulation, the differences in the representation of competitor entity nodes formed under each comparison dimension are extracted, and a dimension difference vector is generated. Based on the propagation path that actually participates in the propagation update during the feedback aligned convolution propagation process, and combined with the propagation edge index and propagation level index in the propagation record, the difference path index is determined. The dimensional difference vector and the difference path index are combined to represent the competitive comparison results.
8. The competitive comparison method based on knowledge graphs according to claim 1, characterized in that, The generation of the competitor comparison conclusions includes: Based on the competitor comparison results, read the dimension difference vector, perform dimension alignment processing, and generate a set of dimension difference components; Based on the set of dimensional difference components, dimensional aggregation processing is performed to obtain the difference score; Based on the difference path index in the competitor comparison representation result, a path mapping process is performed on the difference propagation path set to map it into the corresponding difference propagation path sequence. Based on the correspondence between the difference path index and the set of dimensional difference components, the difference score under each comparison dimension is bound to the corresponding difference propagation path sequence to generate a dimensional difference path mapping result. Based on the dimensional difference path mapping results, a competitor comparison conclusion is generated, which includes the difference score corresponding to each comparison dimension in the comparison dimension set and the path sequence of the difference propagation path set bound to it.
