Artificial intelligence model generalization mechanism evaluation method and system based on cross-sample interaction graph alignment

Abstract

The application discloses a method and system for evaluating generalization mechanism of an artificial intelligence model based on cross-sample interaction graph alignment. Model outputs of multiple samples of the same kind under different occlusion states are obtained, and salient and interaction and salient or interaction are extracted to establish an attribute interaction graph containing semantic unit nodes, interaction nodes and dependent edges. Direct anchors and composite anchors are established according to concept representation, context representation and sample internal adjacency relationship. Under the constraint of the anchors, interaction corresponding relationship and non-corresponding interaction are determined by combining candidate mapping weight, interaction attribute and adjacency topology, and an interaction transfer matrix containing corresponding quality and non-corresponding quality is constructed. The interaction transfer matrices of multiple sample pairs are aggregated, and migratable interaction and sample-specific noise interaction are determined according to cross-sample support, symbol consistency, intensity stability and topology retention, and a model generalization mechanism evaluation value is generated.

Description

Technical Field

[0001] This invention relates to the field of interpretability analysis and generalization ability evaluation of artificial intelligence models, and in particular to a method and system for evaluating the generalization mechanism of artificial intelligence models based on cross-sample interaction graph alignment. Background Technology

[0002] Artificial intelligence models typically learn the mapping relationship between input data and output results through training samples and perform predictions on samples not used in training. The stability of the model's output across different samples, data distributions, or application scenarios is usually used to measure the model's generalization ability. Existing evaluation methods mostly rely on outcome metrics such as accuracy, loss value, or output consistency, which can reflect changes in model performance but are difficult to explain whether the model uses the same or similar internal interaction mechanisms on different samples.

[0003] By setting different occlusion states for input semantic units and observing changes in model output, the interactions of input units that significantly affect the output can be extracted. However, different samples differ in semantic expression, input unit partitioning, and intra-sample positional relationships. The same concept may manifest as a single semantic unit or a combination of multiple adjacent semantic units in different samples. If only one-to-one comparisons are made based on concept identifiers or node positions, it is difficult to handle cross-sample correspondences such as one-to-many, many-to-one, and many-to-many relationships.

[0004] Furthermore, interactions across different samples may differ in interaction type, order, symbol, intensity, and adjacency topology. Matching interactions solely based on semantic similarity makes it difficult to distinguish between stable cross-sample transferable interaction mechanisms and sample-specific noisy interactions, and also makes it difficult to quantify interactions where no correspondence has been established. Therefore, a technical solution is needed that can uniformly represent the interaction structure within samples, establish cross-sample composite semantic correspondences, and combine interaction attributes and topological structure to evaluate the model's generalization mechanism. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the present invention aims to provide a method and system for evaluating the generalization mechanism of artificial intelligence models based on cross-sample interaction graph alignment, thereby overcoming the aforementioned deficiencies in existing technologies.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] Methods for evaluating the generalization mechanism of AI models based on cross-sample interaction graph alignment include:

[0008] The interaction graph construction steps are as follows: obtain the model output response of multiple samples belonging to the same category under different occlusion states, extract salient and interaction and salient or interaction composed of input semantic units, and build the attribute interaction graph of each sample according to the input semantic units, salient interactions and their subordinate relationship.

[0009] The semantic anchoring step establishes cross-sample semantic anchor groups based on the conceptual representation, contextual representation, and intra-sample adjacency of the input semantic units. The semantic anchor groups include direct anchors between individual input semantic units and composite anchors between input semantic units and combinations of semantic units.

[0010] The structural alignment step generates candidate maps within the same semantic anchor group, determines the matching cost based on the candidate map weights, interaction attributes, and adjacency topology, and determines the interaction correspondence and non-corresponding interactions under anchor ownership constraints and subordinate relationship constraints.

[0011] The migration quantization step involves establishing an interaction migration matrix that includes corresponding and non-corresponding qualities based on the interaction correspondence.

[0012] The generalization determination step aggregates the interaction transfer matrices of multiple sample pairs, determines the transferable interactions and sample-specific noise interactions based on the cross-sample support, sign consistency, strength stability, and topology preservation of the interaction correspondence, and generates a model generalization mechanism evaluation value.

[0013] Preferably, the interaction graph construction step includes dividing the sample into multiple input semantic units, forming occlusion states according to different retained combinations of the input semantic units, inputting each occlusion state into the model to be evaluated and recording the corresponding output response; determining the interaction formed by one or more input semantic units and its interaction intensity based on the output response, and determining the interaction whose interaction intensity reaches the significance condition as a significant interaction; establishing an attribute interaction graph including semantic unit nodes, interaction nodes and subordinate edges, recording concept identifiers, context representations and intra-sample position ranges in semantic unit nodes, and recording interaction type, interaction order, interaction symbol and normalized interaction intensity in interaction nodes.

[0014] Preferably, the semantic anchoring step includes: grouping semantic unit nodes with the same concept representation and contextual representation similarity reaching a first threshold into the same direct anchor group; for semantic unit nodes not grouped into the direct anchor group, forming semantic unit combinations while maintaining positional continuity or spatial adjacency within the sample; comparing the aggregated representation of the semantic unit combination with a single semantic unit node or semantic unit combination in another sample; grouping the comparison objects with similarity reaching a second threshold into the same composite anchor group; and recording the anchor confidence of the direct anchor group and the composite anchor group respectively.

[0015] Preferably, the candidate mapping generation in the structure alignment step includes: establishing one-to-one candidate mappings within the same direct anchor group, establishing one-to-many, many-to-one, or many-to-many candidate mappings within the same composite anchor group, determining candidate mapping weights based on anchor confidence, concept marker consistency, context representation similarity, and adjacency completeness of semantic unit combinations, and normalizing the candidate mapping weights when the same semantic unit node belongs to multiple candidate mappings, and recording the unassigned weights after normalization as node non-correspondence quality.

[0016] Preferably, the matching cost determination in the structure alignment step includes: determining the semantic cost based on the candidate mapping weights between the semantic unit nodes connected to the node to be assigned; determining the interaction attribute cost based on the interaction type compatibility, interaction order difference, interaction symbol consistency, and interaction strength difference between the interaction nodes to be matched; determining the topology cost based on the degree of repetition of the adjacent interaction nodes of the interaction node to be matched on the semantic anchor group; and configuring the combined weights of the semantic cost, interaction attribute cost, and topology cost according to the anchor confidence to form the matching cost.

[0017] Preferably, the structure alignment step further includes setting the candidate mappings corresponding to the direct anchor group as fixed mappings, setting the candidate mappings corresponding to the composite anchor group as allocatable mappings, initializing the correspondence between interactive nodes according to the fixed mappings, updating the topology cost using the determined adjacent interactive nodes in each round of alignment, and ensuring that the sum of the mapping weights of the same semantic unit node is not greater than one. The correspondence between the interactive nodes to be matched is solved under the constraints that the interactive type compatibility relationship is satisfied and the allocatable mapping does not cross the composite anchor group. When the adjacent correspondence no longer changes or the decrease in the total matching cost is lower than the stopping threshold, the remaining interactive nodes are recorded as uncorresponding interactions.

[0018] Preferably, in the migration quantization step, the significant interactions of the first and second samples are used as row and column indices, respectively. Empty corresponding indices are set in the row and column indices, and the corresponding quality of the interaction correspondence is written into the corresponding matrix elements. The corresponding quality is determined based on semantic anchor coverage, interaction type compatibility, interaction order proximity, interaction symbol consistency, normalized interaction strength stability, and topology preservation. The unallocated remaining quality of each significant interaction is written into the corresponding empty corresponding index to form the interaction migration matrix.

[0019] Preferably, the interaction trajectory construction in the generalization determination step includes: using semantic anchor groups as indexes, listing the significant interactions with corresponding quality in the interaction migration matrices of different sample pairs as interaction trajectories according to sample identifiers; when the same sample has multiple corresponding interactions in the same semantic anchor group, selecting interactions according to corresponding quality or merging the normalized interaction strengths of interactions according to corresponding quality; recording the interaction type, interaction order, interaction symbol, normalized interaction strength, and adjacency relationship of each sample in the interaction trajectory; and recording the non-corresponding quality at the sample positions where no corresponding interaction exists.

[0020] Preferably, the generalization determination step further includes determining cross-sample support, symbol consistency, strength stability, and topology preservation based on the proportion of samples covered by the interaction trajectory, the proportion of identical interaction symbols, the dispersion of normalized interaction strength, and the preservation ratio of adjacency relationships before and after alignment. Interaction trajectories that meet the corresponding determination conditions for each quality assurance are determined as transferable interactions, while interactions that do not meet the noise threshold and interaction trajectories that do not meet the cross-sample support conditions are determined as sample-specific noise interactions. The normalized interaction strengths of the two types of interactions are accumulated according to the interaction order and interaction symbol to generate a model generalization mechanism evaluation value.

[0021] An AI model generalization mechanism evaluation system based on cross-sample interaction graph alignment includes:

[0022] The interaction graph construction module obtains the model output response of multiple samples belonging to the same category under different occlusion states, extracts salient and interaction and salient or interaction composed of input semantic units, and builds the attribute interaction graph of each sample according to the input semantic units, salient interactions and their subordinate relationship.

[0023] The semantic anchoring module establishes cross-sample semantic anchor groups based on the conceptual representation, contextual representation, and intra-sample adjacency of the input semantic units. The semantic anchor groups include direct anchors between individual input semantic units and composite anchors between input semantic units and combinations of semantic units.

[0024] The structural alignment module generates candidate maps within the same semantic anchor group, determines the matching cost based on the candidate map weight, interaction attributes, and adjacency topology, and determines the interaction correspondence and non-corresponding interactions under anchor ownership constraints and subordinate relationship constraints.

[0025] The migration quantization module establishes an interaction migration matrix that includes corresponding and non-corresponding qualities based on the interaction correspondence.

[0026] The generalization determination module aggregates the interaction transfer matrices of multiple sample pairs, determines the transferable interactions and sample-specific noise interactions based on the cross-sample support, sign consistency, strength stability, and topology preservation of the interaction correspondence, and generates a model generalization mechanism evaluation value.

[0027] The beneficial effects of this invention are as follows: It extracts salient and interactive elements and salient or interactive elements based on the model output responses of similar samples under different occlusion states, and establishes an attribute interaction graph including semantic unit nodes, interaction nodes, and dependent edges. The graph records concept identifiers, contextual representations, location ranges, interaction types, interaction orders, interaction symbols, and normalized interaction strengths, providing a unified structured representation of the interaction mechanisms of different samples. It establishes cross-sample semantic anchor groups based on concept representations, contextual representations, and intra-sample adjacency relationships. Direct anchors handle the correspondence between individual semantic units, while composite anchors handle the correspondence between individual semantic units and combinations of semantic units, adapting to differences in semantic expression granularity and input unit partitioning methods among different samples. It comprehensively determines candidate mapping weights based on anchor confidence, concept identifier consistency, contextual representation similarity, and adjacency completeness, and solves the interaction correspondence by combining semantic cost, interaction attribute cost, and topological cost, ensuring that cross-sample alignment is simultaneously affected by... Constraints on semantic anchors, interaction attributes, adjacency topology, and dependency relationships are applied. Candidate mapping weights are normalized during structural alignment, and the non-corresponding quality of nodes and interactions is recorded. By setting an empty correspondence index, the corresponding and non-corresponding quality of interactions are jointly written into the interaction transfer matrix, ensuring that interaction information that fails to establish a correspondence across samples is retained and participates in subsequent generalization judgment. Interaction transfer matrices of multiple sample pairs are aggregated to construct interaction trajectories recording interaction type, interaction order, interaction symbol, normalized interaction strength, and adjacency relationship. Cross-sample support, symbol consistency, strength stability, and topology preservation are calculated respectively to determine the stability of the interaction mechanism among similar samples from multiple dimensions. Based on various indicators of the interaction trajectory, transferable interactions and sample-specific noise interactions are distinguished. The normalized interaction strength is accumulated according to the interaction order and interaction symbol to generate a model generalization mechanism evaluation value, ensuring that the evaluation results reflect whether the model maintains stable input interactions across different samples. Attached Figure Description

[0028] Figure 1 This is an overall flowchart of the present invention;

[0029] Figure 2 This is a flowchart of the attribute interaction graph construction process of the present invention;

[0030] Figure 3 This is a flowchart of the cross-sample semantic anchoring and candidate mapping process of the present invention;

[0031] Figure 4This is a flowchart of the structural alignment and interactive migration matrix construction process of the present invention;

[0032] Figure 5 This is a flowchart of the interactive trajectory construction and generalization determination process of the present invention. Detailed Implementation

[0033] 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.

[0034] It should be noted that when a component is described as "fixed to" another component, it can be directly on the other component or may have a component in between. When a component is considered "connected to" another component, it can be directly connected to the other component or may have a component in between. When a component is considered "set on" another component, it can be directly set on the other component or may have a component in between. The terms "vertical," "horizontal," "left," "right," and similar expressions used in this document are for illustrative purposes only.

[0035] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0036] The embodiments of the present invention will be further described in detail below with reference to the accompanying drawings:

[0037] like Figures 1-5As shown, this embodiment provides a method for evaluating the generalization mechanism of artificial intelligence models based on cross-sample interaction graph alignment, including: an interaction graph construction step, which obtains the model output response of multiple samples belonging to the same category under different occlusion states, extracts salient and interaction and salient or interaction composed of input semantic units, and establishes attribute interaction graphs for each sample according to the input semantic units, salient interactions and their subordinate relationships; a semantic anchoring step, which establishes cross-sample semantic anchor groups according to the conceptual representation, contextual representation and intra-sample adjacency relationship of the input semantic units, the semantic anchor group including direct anchors between individual input semantic units and anchors between input semantic units and semantic units. The method involves several steps: First, it establishes composite anchors between meta-groups. Second, a structural alignment step generates candidate maps within the same semantic anchor group, determines the matching cost based on candidate map weights, interaction attributes, and adjacency topology, and identifies corresponding and non-corresponding interactions under anchor attribution and dependency constraints. Third, a transfer quantization step establishes an interaction transfer matrix containing corresponding and non-corresponding qualities based on the interaction correspondences. Fourth, a generalization judgment step aggregates the interaction transfer matrices of multiple sample pairs, distinguishes between transferable interactions and sample-specific noise interactions based on cross-sample support, sign consistency, strength stability, and topology preservation of the interaction correspondences, and generates a model generalization mechanism evaluation value. This method is used to quantitatively evaluate whether a trained artificial intelligence model relies on common, transferable discriminative logic across multiple samples of the same category. Here, a sample refers to a data instance input into the model to be evaluated for processing, such as image, text, or speech data; a category refers to the prediction target or label to which the sample belongs, and multiple samples belonging to the same category refer to multiple samples with the same labeling result or the same prediction target; occlusion state refers to the sample variant formed after partially preserving or masking the input content in the sample, and by applying different combinations of preservation or masking to the same sample, multiple variants of the sample under different occlusion states can be obtained; the model output response refers to the output result obtained by the model to be evaluated after inferring the sample variant under a certain occlusion state, such as classification confidence, prediction probability, or the value of a specific output channel.

[0038] The interaction graph construction steps include dividing the samples into multiple input semantic units, forming occlusion states according to different retained combinations of input semantic units, inputting each occlusion state into the model to be evaluated and recording the corresponding output response; determining the interaction formed by one or more input semantic units and its interaction strength based on the output response, and identifying interactions whose interaction strength reaches the significance condition as significant interactions; and establishing an attribute interaction graph including semantic unit nodes, interaction nodes, and dependent edges, recording concept identifiers, context representations, and intra-sample location ranges in semantic unit nodes, and recording interaction type, interaction order, interaction symbol, and normalized interaction strength in interaction nodes. Here, an input semantic unit refers to the smallest input segment with independent semantics obtained after dividing the sample, such as an image patch in an image, a word or phrase in text, or a phoneme or syllable in speech. A sample is usually divided into multiple input semantic units. Interaction refers to the contribution of one or more input semantic units to the model output. When an interaction requires multiple input semantic units to significantly affect the model output, the interaction is represented by a logical AND relation, denoted as a significant AND interaction. When any one of the multiple input semantic units significantly affects the model output, the interaction is represented by a logical OR relation, denoted as a significant OR interaction. Interaction strength refers to the quantified value of the contribution of the interaction to the model output response. An interaction is defined as significant when its strength reaches a pre-set significance condition; otherwise, it is not included in subsequent processing. The interaction order refers to the number of input semantic units participating in the interaction. The interaction sign refers to the directional indicator indicating whether the interaction has a positive promoting effect or a negative inhibiting effect on the model output response. Normalized interaction strength refers to the value obtained by normalizing the interaction strength according to a unified dimension range, which facilitates the comparison of interaction strength between different samples. Attribute interaction graph is a graph structure used to describe the interaction relationships between input semantic units within a sample. Semantic unit nodes are used to record the conceptual identifier, contextual representation, and position range of the corresponding input semantic unit within the sample. The conceptual identifier indicates the semantic content referred to by the input semantic unit, and the contextual representation indicates the feature representation of the input semantic unit in its sample context environment. Interaction nodes are used to record the interaction type, interaction order, interaction symbol, and normalized interaction strength of the corresponding interaction. The interaction type is either salience and interaction or salience and interaction. Subordinate edges are used to connect semantic unit nodes with the interaction nodes formed by them, indicating whether the input semantic unit participates in the formation of the corresponding interaction, so that the attribute interaction graph can fully reflect the subordinate relationship between the input semantic units and salience interactions within the sample.

[0039] The semantic anchoring steps include: grouping semantic unit nodes with the same concept representation and contextual representation similarity reaching a first threshold into the same direct anchor group; forming semantic unit combinations for semantic unit nodes not grouped into direct anchor groups while maintaining positional continuity or spatial adjacency within the samples; comparing the aggregated representation of the semantic unit combinations with a single semantic unit node or semantic unit combination in another sample; grouping the comparison objects with similarity reaching a second threshold into the same composite anchor group; and recording the anchor confidence of the direct anchor group and the composite anchor group respectively. Semantic anchor sets refer to sets of correspondences that span different samples, indicating semantic comparability between input semantic units or combinations of input semantic units in different samples. Semantic anchor sets include two types: direct anchors and composite anchors. Direct anchors refer to the correspondences established between individual input semantic units in different samples, while composite anchors refer to the correspondences established between input semantic units and combinations of semantic units. A semantic unit combination refers to a combination formed by multiple input semantic units while maintaining positional continuity or spatial adjacency within the sample. Anchor confidence is a quantitative value used to indicate the reliability of the correspondences within a semantic anchor set. Through these semantic anchoring steps, a unified semantic correspondence foundation can be established between the attribute interaction graphs of different samples, providing a basis for subsequent structural alignment steps.

[0040] The candidate mapping generation in the structural alignment step includes: establishing one-to-one candidate mappings within the same direct anchor group, and establishing one-to-many, many-to-one, or many-to-many candidate mappings within the same composite anchor group. Candidate mapping weights are determined based on anchor confidence, concept identifier consistency, contextual representation similarity, and the adjacency completeness of semantic unit combinations. When a semantic unit node belongs to multiple candidate mappings, the candidate mapping weights are normalized, and the unassigned weights after normalization are recorded as node non-correspondence quality. Here, a candidate mapping refers to a mapping relationship that may have corresponding relationships between semantic unit nodes in different samples within the same semantic anchor group; the candidate mapping weight is a quantified value indicating the probability of a candidate mapping being valid; and the node non-correspondence quality refers to the weight of a semantic unit node that, after normalization, fails to be assigned to any candidate mapping, indicating the degree to which the semantic unit node is not sufficiently matched in cross-sample correspondence relationships.

[0041] The matching cost determination in the structural alignment step includes: determining the semantic cost based on the candidate mapping weights between the semantic unit nodes connected to the node to be assigned; determining the interaction attribute cost based on the interaction type compatibility, interaction order difference, interaction symbol consistency, and interaction strength difference between the interaction nodes to be matched; determining the topological cost based on the degree of repetition of the adjacent interaction nodes of the node to be matched on the semantic anchor group; and configuring the combined weights of the semantic cost, interaction attribute cost, and topological cost according to the anchor confidence to form the matching cost. The semantic cost indicates the reasonableness of the correspondence of semantic unit nodes at the semantic level; the interaction attribute cost indicates the reasonableness of the correspondence of interaction nodes at the interaction attribute level (interaction type, interaction order, interaction symbol, and interaction strength); and the topological cost indicates the reasonableness of the correspondence of interaction nodes at the level of their adjacency structure. The matching cost is the total cost obtained by weighting the semantic cost, interaction attribute cost, and topological cost according to their respective combined weights. The lower the matching cost, the more reasonable the correspondence, which can be used as the basis for subsequent determination of interaction correspondence.

[0042] The structural alignment step also includes setting the candidate mappings corresponding to direct anchor groups as fixed mappings and the candidate mappings corresponding to composite anchor groups as assignable mappings. The correspondence between interaction nodes is initialized based on the fixed mappings. In each round of alignment, the topological cost is updated using the determined adjacent interaction nodes, and the sum of the mapping weights of nodes in the same semantic unit is no greater than one. The correspondence between the interaction nodes to be matched is solved under the constraints that the interaction type compatibility relationship is satisfied and the assignable mapping does not cross the composite anchor group. When the adjacent correspondence no longer changes or the decrease in the total matching cost is lower than the stopping threshold, the remaining interaction nodes are recorded as non-corresponding interactions. Here, a fixed mapping refers to a candidate mapping whose correspondence remains unchanged during the alignment process, an assignable mapping refers to a candidate mapping whose correspondence can be adjusted during the alignment iteration process, and a non-corresponding interaction refers to an interaction node that still fails to establish a correspondence with the interaction nodes of other samples after the above iterative solution. By repeatedly iterating under anchor attribution and dependency constraints, the interaction correspondence and non-corresponding interactions between different sample attribute interaction graphs within the same semantic anchor group can be determined.

[0043] The migration quantization step uses the significant interactions of the first and second samples as row and column indices, respectively. Empty corresponding indices are set in the row and column indices, and the corresponding quality of the interaction correspondence is written into the corresponding matrix elements. The corresponding quality is determined based on semantic anchor coverage, interaction type compatibility, interaction order proximity, interaction symbol consistency, normalized interaction strength stability, and topology preservation. The remaining unassigned quality of each significant interaction is written into the corresponding empty corresponding index to form an interaction migration matrix. The interaction migration matrix is ​​a matrix used to record the significant interaction correspondence quality and non-correspondence quality between sample pairs. Correspondence quality refers to the quantitative value of the reasonableness of the correspondence between an interaction node and another sample interaction node. Semantic anchor coverage is used to indicate the coverage of the semantic anchor group on which the correspondence is based. Interaction type compatibility, interaction order proximity, and interaction symbol consistency are used to indicate the degree of consistency between two interaction nodes in terms of interaction type, interaction order, and interaction symbol, respectively. Normalized interaction strength stability is used to indicate the proximity of the normalized interaction strength of two interaction nodes. Topology preservation is used to indicate the degree of preservation of the adjacency structure of two interaction nodes. The empty correspondence index is used to carry the remaining quality that failed to establish a correspondence. The unassigned remaining quality is the non-correspondence quality, which is used to indicate the degree to which the interaction node has not been assigned to any correspondence.

[0044] The interaction trajectory construction in the generalization determination step includes: using semantic anchor groups as indices, arranging salient interactions of corresponding quality in the interaction transfer matrices of different sample pairs according to their sample identifiers to form interaction trajectories. When multiple corresponding interactions exist for the same sample within the same semantic anchor group, interactions are selected based on their corresponding quality, or the normalized interaction intensities of interactions are merged according to their corresponding quality. The interaction type, interaction order, interaction symbol, normalized interaction intensity, and adjacency relationship of each sample are recorded sequentially in the interaction trajectory. For samples where no corresponding interaction exists, the non-corresponding quality is recorded. Here, the interaction trajectory refers to the record sequence formed by arranging salient interactions of corresponding quality in the interaction transfer matrices of different sample pairs according to their sample identifiers, using a certain semantic anchor group as an index. This sequence indicates the continuity and changes of the interaction content corresponding to the same semantic anchor group across different samples.

[0045] The generalization determination step also includes determining cross-sample support, symbol consistency, strength stability, and topology preservation based on the proportion of samples covered by the interaction trajectory, the proportion of identical interaction symbols, the dispersion of normalized interaction strength, and the proportion of adjacency relationship preservation before and after alignment. Interaction trajectories that meet the corresponding determination conditions for each indicator are determined as transferable interactions, while interactions that do not meet the noise threshold and interaction trajectories that do not meet the cross-sample support condition are determined as sample-specific noise interactions. The normalized interaction strength of the two types of interactions is accumulated according to the interaction order and interaction symbol to generate the model generalization mechanism evaluation value. Among them, cross-sample support indicates the degree to which the interaction content involved in the interaction trajectory is commonly observed in multiple samples participating in the evaluation; sign consistency indicates whether the direction of action of the interaction trajectory remains consistent across different samples; strength stability indicates whether the strength of action of the interaction trajectory remains stable across different samples; topology preservation indicates whether the adjacency structure of the input semantic units involved in the interaction trajectory remains consistent before and after cross-sample alignment; transferable interaction indicates that the interaction content reflects the discriminative logic commonly relied upon by the model under evaluation across different samples; sample-specific noise interaction indicates that the interaction content is merely a random response generated by the model under evaluation on individual samples and does not represent that the model has common discriminative logic; the model generalization mechanism evaluation value comprehensively indicates the relative proportion of the normalized cumulative interaction strength between transferable interaction and sample-specific noise interaction in the discriminative logic relied upon by the model under evaluation across multiple samples of the same class. The greater the proportion leaning towards transferable interaction, the more the generalization mechanism of the model under evaluation relies on common discriminative logic, and vice versa, the more the generalization mechanism of the model under evaluation relies on sample-specific random responses.

[0046] The aforementioned method for evaluating the generalization mechanism of AI models based on cross-sample interaction graph alignment can be implemented by a corresponding system for evaluating the generalization mechanism of AI models based on cross-sample interaction graph alignment. The system includes: an interaction graph construction module, which acquires the model output responses of multiple samples belonging to the same category under different occlusion states, extracts salient and salient or interactions composed of input semantic units, and establishes attribute interaction graphs for each sample based on the input semantic units, salient interactions, and their subordinate relationships; and a semantic anchoring module, which establishes cross-sample semantic anchor groups based on the conceptual representation, contextual representation, and intra-sample adjacency relationships of the input semantic units. The semantic anchor groups include direct connections between individual input semantic units. The system comprises four modules: an anchor point and a composite anchor point between input semantic units and combinations of semantic units; a structure alignment module, which generates candidate mappings within the same semantic anchor point group, determines the matching cost based on the candidate mapping weights, interaction attributes, and adjacency topology, and determines the interaction correspondence and non-corresponding interactions under anchor point ownership and dependency constraints; a transfer quantization module, which establishes an interaction transfer matrix containing corresponding and non-corresponding qualities based on the interaction correspondence; and a generalization determination module, which aggregates the interaction transfer matrices of multiple sample pairs, distinguishes between transferable interactions and sample-specific noise interactions based on the cross-sample support, sign consistency, strength stability, and topology preservation of the interaction correspondence, and generates a model generalization mechanism evaluation value. The specific processing content performed by the interaction graph construction module, semantic anchoring module, structure alignment module, transfer quantization module, and generalization determination module corresponds to the descriptions of the aforementioned interaction graph construction steps, semantic anchoring steps, structure alignment steps, transfer quantization steps, and generalization determination steps, respectively. These modules work collaboratively in sequence to achieve a quantitative evaluation of the generalization mechanism of the model to be evaluated.

[0047] The above description is merely a specific embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0048] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principle of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for evaluating the generalization mechanism of artificial intelligence models based on cross-sample interaction graph alignment, characterized in that, include: The interaction graph construction steps are as follows: obtain the model output response of multiple samples belonging to the same category under different occlusion states, extract salient and interaction and salient or interaction composed of input semantic units, and build the attribute interaction graph of each sample according to the input semantic units, salient interactions and their subordinate relationship. The semantic anchoring step establishes cross-sample semantic anchor groups based on the conceptual representation, contextual representation, and intra-sample adjacency of the input semantic units. The semantic anchor groups include direct anchors between individual input semantic units and composite anchors between input semantic units and combinations of semantic units. The structural alignment step generates candidate maps within the same semantic anchor group, determines the matching cost based on the candidate map weights, interaction attributes, and adjacency topology, and determines the interaction correspondence and non-corresponding interactions under anchor ownership constraints and subordinate relationship constraints. The migration quantization step involves establishing an interaction migration matrix that includes corresponding and non-corresponding qualities based on the interaction correspondence. The generalization determination step aggregates the interaction transfer matrices of multiple sample pairs, determines the transferable interactions and sample-specific noise interactions based on the cross-sample support, sign consistency, strength stability, and topology preservation of the interaction correspondence, and generates a model generalization mechanism evaluation value.

2. The method for evaluating the generalization mechanism of artificial intelligence models based on cross-sample interaction graph alignment according to claim 1, characterized in that, The interaction graph construction step includes dividing the sample into multiple input semantic units, forming occlusion states according to different retention combinations of the input semantic units, inputting each occlusion state into the model to be evaluated, and recording the corresponding output response. Based on the output response, determine the interaction formed by one or more input semantic units and its interaction strength, and determine the interaction whose interaction strength reaches the salience condition as a salient interaction; establish an attribute interaction graph including semantic unit nodes, interaction nodes and subordinate edges, record concept identifiers, context representations and in-sample position ranges in semantic unit nodes, and record interaction type, interaction order, interaction symbol and normalized interaction strength in interaction nodes.

3. The method for evaluating the generalization mechanism of artificial intelligence models based on cross-sample interaction graph alignment according to claim 1, characterized in that, The semantic anchoring step includes grouping semantic unit nodes with the same concept representation and contextual representation similarity reaching a first threshold into the same direct anchor group. For semantic unit nodes not grouped into the direct anchor group, semantic unit combinations are formed while maintaining positional continuity or spatial adjacency within the sample. The aggregated representation of the semantic unit combination is compared with a single semantic unit node or semantic unit combination in another sample, and the comparison objects with similarity reaching a second threshold are grouped into the same composite anchor group. The anchor confidence of the direct anchor group and the composite anchor group are recorded respectively.

4. The method for evaluating the generalization mechanism of artificial intelligence models based on cross-sample interaction graph alignment according to claim 1, characterized in that, The candidate mapping generation in the structural alignment step includes: establishing one-to-one candidate mappings within the same direct anchor group, establishing one-to-many, many-to-one, or many-to-many candidate mappings within the same composite anchor group, determining candidate mapping weights based on anchor confidence, concept marker consistency, context representation similarity, and adjacency completeness of semantic unit combinations, and normalizing the candidate mapping weights when the same semantic unit node belongs to multiple candidate mappings, and recording the unassigned weights after normalization as nodes not corresponding to the quality.

5. The method for evaluating the generalization mechanism of artificial intelligence models based on cross-sample interaction graph alignment according to claim 4, characterized in that, The matching cost determination in the structure alignment step includes: determining the semantic cost based on the candidate mapping weights between the semantic unit nodes connected to the node to be assigned; determining the interaction attribute cost based on the interaction type compatibility, interaction order difference, interaction symbol consistency, and interaction strength difference between the interaction nodes to be matched; determining the topology cost based on the degree of repetition of the adjacent interaction nodes of the interaction node to be matched on the semantic anchor group; and configuring the combined weights of the semantic cost, interaction attribute cost, and topology cost according to the anchor confidence to form the matching cost.

6. The method for evaluating the generalization mechanism of artificial intelligence models based on cross-sample interaction graph alignment according to claim 5, characterized in that, The structure alignment step further includes setting the candidate mappings corresponding to the direct anchor group as fixed mappings and the candidate mappings corresponding to the composite anchor group as allocatable mappings. The correspondence between interactive nodes is initialized according to the fixed mappings. In each round of alignment, the topology cost is updated using the determined adjacent interactive nodes. The sum of the mapping weights of the same semantic unit node is not greater than one. The correspondence between the interactive nodes to be matched is solved under the constraints that the interactive type compatibility relationship is satisfied and the allocatable mapping does not cross the composite anchor group. When the adjacent correspondence no longer changes or the decrease in the total matching cost is lower than the stopping threshold, the remaining interactive nodes are recorded as uncorresponding interactions.

7. The method for evaluating the generalization mechanism of artificial intelligence models based on cross-sample interaction graph alignment according to claim 1, characterized in that, The migration quantization step uses the significant interactions of the first and second samples as row and column indices, respectively. Empty corresponding indices are set in the row and column indices, and the corresponding quality of the interaction correspondence is written into the corresponding matrix elements. The corresponding quality is determined based on semantic anchor coverage, interaction type compatibility, interaction order proximity, interaction symbol consistency, normalized interaction strength stability, and topology preservation. The remaining unallocated quality of each significant interaction is written into the corresponding empty corresponding index to form the interaction migration matrix.

8. The method for evaluating the generalization mechanism of artificial intelligence models based on cross-sample interaction graph alignment according to claim 1, characterized in that, The interaction trajectory construction in the generalization determination step includes: using semantic anchor groups as indexes, listing the significant interactions with corresponding quality in the interaction transfer matrices of different sample pairs as interaction trajectories according to sample identifiers; when the same sample has multiple corresponding interactions in the same semantic anchor group, selecting interactions according to corresponding quality or merging the normalized interaction strengths of interactions according to corresponding quality; recording the interaction type, interaction order, interaction symbol, normalized interaction strength, and adjacency relationship of each sample in the interaction trajectory; and recording the non-corresponding quality at the sample positions where no corresponding interaction exists.

9. The method for evaluating the generalization mechanism of artificial intelligence models based on cross-sample interaction graph alignment according to claim 8, characterized in that, The generalization determination step further includes determining cross-sample support, symbol consistency, strength stability, and topology preservation based on the proportion of samples covered by the interaction trajectory, the proportion of identical interaction symbols, the dispersion of normalized interaction strength, and the preservation ratio of adjacency relationships before and after alignment. Interaction trajectories that meet the corresponding determination conditions for each quality assurance are determined as transferable interactions, while interactions that do not meet the noise threshold and interaction trajectories that do not meet the cross-sample support conditions are determined as sample-specific noise interactions. The normalized interaction strengths of the two types of interactions are accumulated according to the interaction order and interaction symbol to generate a model generalization mechanism evaluation value.

10. An evaluation system for the generalization mechanism of artificial intelligence models based on cross-sample interaction graph alignment, characterized in that, include: The interaction graph construction module obtains the model output response of multiple samples belonging to the same category under different occlusion states, extracts salient and interaction and salient or interaction composed of input semantic units, and builds the attribute interaction graph of each sample according to the input semantic units, salient interactions and their subordinate relationship. The semantic anchoring module establishes cross-sample semantic anchor groups based on the conceptual representation, contextual representation, and intra-sample adjacency of the input semantic units. The semantic anchor groups include direct anchors between individual input semantic units and composite anchors between input semantic units and combinations of semantic units. The structural alignment module generates candidate maps within the same semantic anchor group, determines the matching cost based on the candidate map weight, interaction attributes, and adjacency topology, and determines the interaction correspondence and non-corresponding interactions under anchor ownership constraints and subordinate relationship constraints. The migration quantization module establishes an interaction migration matrix that includes corresponding and non-corresponding qualities based on the interaction correspondence. The generalization determination module aggregates the interaction transfer matrices of multiple sample pairs, determines the transferable interactions and sample-specific noise interactions based on the cross-sample support, sign consistency, strength stability, and topology preservation of the interaction correspondence, and generates a model generalization mechanism evaluation value.

Images