Atlas tracking chain determination method and device, electronic equipment and storage medium

Abstract

The invention discloses an atlas tracking chain determination method and device, electronic equipment and a storage medium. The method comprises the following steps: inputting second data into a second target model, and performing activation probability screening on the second data in each layer of the second target model to obtain a first path set; constructing an interlayer node matching graph based on the first path set, and obtaining a second path set based on the interlayer node matching graph; respectively injecting a structure disturbance signal into each path in the candidate path diagram to form a disturbance sample set; inputting the disturbance sample set into a second target model, determining an output change offset amplitude, and determining a third path set based on the output change offset amplitude; for each path in the third path set, determining a path contribution score based on the output change offset amplitude; and generating an atlas tracking chain corresponding to the first target model based on the first path set, the second path set, the third path set and each path contribution score.

Description

Technical Field

[0001] This invention relates to the field of model interpretability analysis technology, and in particular to a method, apparatus, electronic device and storage medium for determining a spectral tracing chain. Background Technology

[0002] In current deep learning-based decision-making systems, mainstream solutions generally rely on highly complex neural network architectures to complete the end-to-end inference process from input to output. To improve the generalization ability of the model, these systems introduce multi-layer feature extraction, attention mechanisms, or residual connections to progressively compress and reconstruct input features in a high-dimensional space, thereby generating the final decision result.

[0003] While these models offer significant advantages in accuracy, their internal information flow is highly nonlinear and lacks structural labeling, resulting in each intermediate representation, activation path, and attention transfer process being unobservable. Existing solutions are mostly based on posterior computation or local sensitivity assessment, failing to capture structural causal information synchronously during model inference, leading to an information disconnect between the explanation mechanism and the original model. Summary of the Invention

[0004] This invention provides a method, apparatus, electronic device, and storage medium for determining a spectral tracing chain, in order to solve the problem of information disconnection when performing interpretability analysis on models in existing methods.

[0005] According to one aspect of the present invention, a method for determining a spectral tracing chain is provided, the method comprising:

[0006] Based on the first data, the second data is generated. The first data is the data input into the target model when using the first target model, and the second data is based on the first data after graph structure transformation.

[0007] The second data is input into the second target model, and the activation probability of the second data is filtered in each layer of the second target model to obtain the first path set; the second target model is obtained by introducing graph convolution operation and structural gating mechanism into each hidden layer of the first target model.

[0008] Based on the first path set, an inter-layer node matching graph is constructed, and based on the inter-layer node matching graph, the first path set is filtered to obtain the second path set;

[0009] Structural perturbation signals are injected into each path in the candidate path graph to form a perturbation sample set; the candidate path graph is obtained by merging the first path set and the second path set.

[0010] The perturbation sample set is input into the second target model to determine the output change offset magnitude, and the third path set is determined based on the output change offset magnitude.

[0011] For each path in the third path set, the path contribution score is determined based on the magnitude of the output change offset.

[0012] Based on the first path set, the second path set, the third path set, and the contribution scores of each path, a graph tracing chain corresponding to the first target model is generated.

[0013] According to another aspect of the present invention, a map tracing chain determination apparatus is provided, the apparatus comprising:

[0014] The second data generation module is used to generate second data based on the first data. The first data is the data input into the target model when using the first target model. The second data is based on the first data after graph structure transformation.

[0015] The first set construction module is used to input the second data into the second target model and perform activation probability filtering on the second data in each layer of the second target model to obtain the first path set; the second target model is obtained by introducing graph convolution operation and structural gating mechanism into each hidden layer of the first target model.

[0016] The second set construction module is used to construct an inter-layer node matching graph based on the first path set, and to filter the first path set based on the inter-layer node matching graph to obtain the second path set;

[0017] The sample set generation module is used to inject structural perturbation signals into each path in the candidate path graph to form a perturbation sample set; the candidate path graph is obtained by merging the first path set and the second path set.

[0018] The third set construction module is used to input the perturbation sample set into the second target model, determine the output change offset magnitude, and determine the third path set based on the output change offset magnitude;

[0019] The contribution score calculation module is used to determine the path contribution score for each path in the third path set based on the magnitude of the output change offset.

[0020] The tracing chain generation module is used to generate a graph tracing chain corresponding to the first target model based on the first path set, the second path set, the third path set, and the contribution scores of each path.

[0021] According to another aspect of the present invention, an electronic device is provided, the electronic device comprising:

[0022] At least one processor; and

[0023] A memory that is communicatively connected to at least one processor; wherein,

[0024] The memory stores a computer program that can be executed by at least one processor, such that the at least one processor is able to perform the map tracing chain determination method according to any embodiment of the present invention.

[0025] According to another aspect of the present invention, a computer-readable storage medium is provided, which stores computer instructions for causing a processor to execute and implement the graph tracing chain determination method of any embodiment of the present invention.

[0026] The technical solution of this invention involves generating second data based on first data; inputting the second data into a second target model, and performing activation probability filtering on the second data in each layer of the second target model to obtain a first path set; the second target model is obtained by introducing graph convolution operations and structural gating mechanisms into each hidden layer of the first target model; constructing an inter-layer node matching graph based on the first path set, and filtering the first path set based on the inter-layer node matching graph to obtain a second path set; injecting structural perturbation signals into each path in the candidate path graph to form a perturbation sample set; the candidate path graph is obtained by merging the first path set and the second path set; and then... The perturbation sample set is input into the second target model to determine the output change offset magnitude, and a third path set is determined based on the output change offset magnitude. For each path in the third path set, a path contribution score is determined based on the output change offset magnitude. Based on the first path set, the second path set, the third path set, and the contribution scores of each path, a graph tracing chain corresponding to the first target model is generated. This avoids the limitations of traditional methods that require retrospective inference or rely solely on single-step gradient information, and achieves full-process modeling of path-level feature propagation, cross-layer semantic jumps, and response trajectories. As a result, a multi-path set with structural continuity and causal verifiability is constructed, forming a unified graph tracing chain.

[0027] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of the present invention, nor is it intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0028] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0029] Figure 1 This is a flowchart of a method for determining a spectral tracing chain according to Embodiment 1 of the present invention;

[0030] Figure 2 This is a flowchart of another method for determining a spectral tracing chain according to Embodiment 2 of the present invention;

[0031] Figure 3 This is a schematic diagram of a spectrum tracing chain determination device according to Embodiment 3 of the present invention;

[0032] Figure 4 This is a schematic diagram of the structure of an electronic device that implements the spectrum tracing chain determination method of the present invention. Detailed Implementation

[0033] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 should fall within the scope of protection of the present invention.

[0034] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0035] Example 1

[0036] Figure 1 This is a flowchart of a method for determining a graph tracing chain according to Embodiment 1 of the present invention. This embodiment is applicable to situations where interpretability analysis of a model is performed. This method can be executed by a graph tracing chain determination device, which can be implemented in hardware and / or software and can be configured in an electronic device with data processing capabilities. Figure 1 As shown, the method includes:

[0037] S110. Generate second data based on the first data.

[0038] The first data is the data input into the target model when using the first target model, and the second data is obtained after transforming the first data into a graph structure.

[0039] The first data can be the data input to the first target model during normal use. The first target model can be a model that requires graph tracing chain determination. The second data can be obtained by transforming the first data into a graph structure. The first target model can be a convolutional neural network (CNN), Transformer, LSTM, etc.

[0040] Obtain the first data X, where, The first data is then normalized and its features are grouped, mapping it to a set of nodes. ;in, ; and edge set The graph structure Each node This indicates the first data in the input. Each feature dimension, each side Representation of features With features There is a structural dependency between them.

[0041] S120. Input the second data into the second target model, and perform activation probability filtering on the second data in each layer of the second target model to obtain the first path set.

[0042] The second target model is obtained by introducing graph convolution operations and structural gating mechanisms into each hidden layer of the first target model.

[0043] The second data is used to replace the first data and is input into the second target model. In each hidden layer of the second target model, the activation probability of the second data is filtered to select the first path set that can represent the structural propagation chain of the input features between the layers of the model.

[0044] Optionally, activation probability filtering is performed on the second data at each layer within the second target model to obtain a first path set, including:

[0045] Based on the graph convolution operation and structure gating mechanism introduced in each hidden layer of the second target model, the activation probability of the edges in the graph structure of the second data is filtered, and the structure propagation path is extracted based on the maximum joint edge activation probability to generate the first path set.

[0046] After obtaining the second data, the second data is input into the second target model. Graph convolution operation and structure gating mechanism are introduced into each hidden layer of the second target model. The activation probability of the edges in each graph structure is filtered. The structure propagation path is extracted based on the maximum joint edge activation probability to generate the first path set. The first path set represents the structure propagation chain of the input features between each layer of the model.

[0047] After obtaining the second data, namely the graph structure The input is fed into each hidden layer of the second objective model, and the first... Layer diagram representation Perform a graph convolution update operation, and update the graph using the following formula: Layer node representation matrix :

[0048] ;

[0049] in Indicates the first The feature vector of each node in the layer; It is an adjacency matrix. Indicates the existence of the first The first in the layer , Edges connecting nodes; for The degree matrix; For trainable weight matrix, For activation functions;

[0050] For each edge Introducing the first structural gating factor And calculate its gate value according to the following formula:

[0051] ;

[0052] in and They represent the first Layer , The feature vector of each node; These are trainable weight vectors; For bias terms; It is a logical sigmoid activation function; This represents a vector concatenation operation;

[0053] According to the edge existence index With the first structural gating factor Calculate the first joint activation probability for each edge. Defined as:

[0054] ;

[0055] In all hidden layers of the model, construct all possible paths from the first input data to the model output using depth-first search or dynamic programming. For each path... Joint activation score The calculation is performed and defined as follows:

[0056] ;

[0057] Select the paths with the highest joint activation scores from all paths to form the first path set. Each path represents a high-strength structure propagation channel.

[0058] S130. Based on the first path set, construct an inter-layer node matching graph, and based on the inter-layer node matching graph, filter the first path set to obtain a second path set.

[0059] Interlayer node matching graphs typically consist of multiple layers, each containing a set of nodes. Layers are connected by edges to represent the matching relationships between nodes.

[0060] A layer-to-layer node matching graph is constructed based on all paths in the first path set. Based on this graph, node pairs satisfying cross-layer semantic jump features are identified, and a second path set is generated. This second path set represents cross-layer semantic dependency chains formed by discontinuous activation paths, thus aiding in the identification of potential spurious structures within the model.

[0061] Optionally, based on the inter-layer node matching graph, the first path set is filtered to obtain a second path set, including:

[0062] Based on the inter-layer node matching graph, the semantic embedding similarity and activation direction consistency between each node pair are determined.

[0063] Based on the semantic embedding similarity and activation direction consistency between node pairs, determine whether each node pair meets the preset jump threshold.

[0064] If the conditions are met, the node pair is retained, and a second path set is constructed based on the retained node pairs.

[0065] Semantic embedding similarity is an important indicator for measuring the semantic similarity of text. It converts text into semantic vectors through an embedding model, and then uses methods such as cosine similarity and Euclidean distance to calculate the similarity between vectors, thus determining the degree of similarity in text meaning. Activation direction consistency refers to the degree of consistency in the direction of activation outputs from different neurons or different layers. If two neurons have the same activation direction, it indicates that their response patterns to the input are similar, and they may have captured similar features or semantic information.

[0066] Since semantic embedding similarity and activation direction consistency can provide a basis for determining cross-layer semantic jump features, an inter-layer node matching graph is constructed for all paths in the first path set, and the semantic embedding similarity and activation direction consistency between each node pair are calculated. If a node pair satisfies the cross-layer semantic jump feature, then node pairs that meet the conditions are selected according to a preset jump threshold, and the second path set is generated accordingly.

[0067] For the first path set Each path in Extract all cross-level node combinations in the path, let the first... The first in the layer The activation of a node is represented as , No. The first in the layer The activation of a node is represented as ,in Construct a set of node pairs:

[0068] ;

[0069] For sets Any node pair Define the semantic jump consistency scoring function for this node pair. It is composed of the product of the following three indicators:

[0070] ;

[0071] ;

[0072] ;

[0073] ;

[0074] The meanings of each item are as follows:

[0075] ;

[0076] Cosine similarity is used to represent the similarity between node pairs, reflecting semantic direction consistency.

[0077] ;

[0078] This represents the squared difference of the Euclidean distance between node pairs in the activation space, reflecting the semantic span strength.

[0079] ;

[0080] Represents the change in energy integral along the cross-layer embedding path, where Based on and The intermediate projection representation obtained by linear interpolation;

[0081] Represents a semantic similarity mapping function. This is the standard Sigmoid function;

[0082] This indicates the activation span penalty function;

[0083] Represents the structural perturbation energy function;

[0084] in Positive real number parameters set for experience, Used to control the influence weight of each indicator in the scoring function;

[0085] Set a rating threshold Define the node pair jump determination function. :

[0086] ;

[0087] in Represents node pairs It was determined to have a valid semantic jump relationship and constitute part of a potential discontinuous path;

[0088] Based on all satisfying Node pairs are used to construct a sequence of cross-layer jump paths:

[0089] ;

[0090] For any They all ,and This ensures that the path is a cross-level semantic jump between non-adjacent levels.

[0091] For each path Calculate its cross-layer propagation span Defined as:

[0092] ;

[0093] like ,in If a set jump depth threshold is set, the path is considered a non-explicit jump path with long-term dependency characteristics.

[0094] All satisfied path Forming the second path set: , which serves as the set of cross-layer discontinuous structural dependency chains in the second objective model.

[0095] S140. Inject structural perturbation signals into each path in the candidate path graph to form a perturbation sample set; the candidate path graph is obtained by merging the first path set and the second path set.

[0096] Structural disturbance signals refer to external excitation signals or internal interference signals that cause vibration or other dynamic response changes in a structural system, which can cause the structure to deviate from its normal static or dynamic working state.

[0097] The first path set and the second path set are merged to obtain a candidate path map. Structural perturbation signals are injected into each path in the candidate path map to form a perturbation sample set.

[0098] Optionally, a structural perturbation signal is injected into each path in the candidate path graph to form a perturbation sample set, including:

[0099] Based on the disturbance response path identification algorithm, structural disturbance signals are injected into each path in the candidate path graph to form a disturbance sample set.

[0100] Perturbation Response Path Identification (PRPI) is a class of algorithms used to determine the propagation path of perturbations from the source to the response in a system. It can inject structural perturbations into the fusion path and cause the model output to shift.

[0101] S150. Input the perturbation sample set into the second target model, determine the output change offset magnitude, and determine the third path set based on the output change offset magnitude.

[0102] To identify causal relationships within the model, the association between input variables and output is analyzed using a perturbation sample set to verify the true causal impact of variables on the output. Therefore, a perturbation sample set is needed and input into the second target model. Since the perturbation sample set contains injected structural perturbation signals, it will cause changes in the output of the second target model. Based on the magnitude of these changes, the output shift can be determined.

[0103] Optionally, the perturbation sample set can be input into the second target model to determine the output changes caused by the perturbation of each path in the perturbation sample set;

[0104] Based on the output changes caused by disturbances along each path, determine the offset magnitude of the output change for each path.

[0105] If the output change offset is greater than the preset disturbance threshold, the corresponding path will be stored in the third path set.

[0106] The perturbation sample set is input into the second target model, the output change caused by the perturbation of each path is recorded, and the corresponding perturbation response trajectory is generated by backpropagation.

[0107] If the output change caused by the disturbance exceeds the preset disturbance threshold, the corresponding path will be included in the third path set; the third path set is used to characterize the deviation behavior of path-level structural disturbance on the model prediction results.

[0108] Merge all paths from the first path set and the second path set to construct a candidate path set, denoted as . Each path It is a directed structural dependency chain consisting of several nodes and edges connected in sequence;

[0109] For each path Perform a structural perturbation injection operation, defining the perturbation function as follows: ,in Represents the disturbance intensity constant. Representing a path The Middle The edge weight vectors of the layer, after perturbation, are represented as follows: ;

[0110] The graph structure after perturbation The input is fed into the second target model to obtain the perturbated model output. and compared with the original output results Compare and calculate the output offset value .

[0111] S160. For each path in the third path set, determine the path contribution score based on the output change offset magnitude.

[0112] Optionally, for each path in the third path set, a path contribution score is determined based on the magnitude of the output change offset, including:

[0113] For each path in the third path set, a perturbation integral function is constructed based on the output change offset magnitude;

[0114] Based on the perturbation integral function, the path perturbation integral value of each path is determined;

[0115] The path contribution score is obtained by determining the ratio of the path disturbance integral value to the original output prediction.

[0116] For each path in the third path set, a disturbance integral function is constructed based on the disturbance amplitude in the path disturbance response trajectory;

[0117] The offset ratio between the path perturbation integral value and the original output prediction is calculated, and a path contribution score is generated accordingly. The path contribution score is used to quantify the causal effect of the current path on the final model output. The path contribution score value corresponds to the path node structure.

[0118] For the third path set Each path in Extract its disturbance response trajectory ,in Indicates the first in the path Layer node output offset value Gradient contribution;

[0119] Based on disturbance response trajectory Construct the path perturbation integral function Defined as:

[0120] ;

[0121] in This represents the normalized interval of the perturbation evolution step size. For the first Layer nodes in the perturbation evolution process Activation weight at any given moment;

[0122] path integral value Original predicted offset corresponding to the path The normalized ratio between them constructs the path contribution score. Defined as:

[0123] ;

[0124] in To prevent extremely small positive numbers from being divided by zero, This indicates the overall magnitude of change in the predicted output caused by path disturbances;

[0125] Path contribution score Corresponding to the positions of each node in the path, a path-node contribution distribution mapping structure is formed, denoted as... ,in For the first The local contribution score of the layer node;

[0126] Contribute mapping to all path nodes Perform a structure normalization operation to generate a node structure causal weight graph. ,in Represents a node Cumulative contribution weighting across all paths:

[0127] ;

[0128] in For indicator functions, As a normalization factor, it ensures that the sum of all weights is 1;

[0129] Finally, the causal weight graph of the node structure will be generated. With path contribution score set A combined causal contribution evaluation structure for the pathway.

[0130] S170. Based on the first path set, the second path set, the third path set, and the contribution scores of each path, generate the graph tracing chain corresponding to the first target model.

[0131] A fusion operation is performed on the first path set, the second path set, and the third path set to construct a fused path set, which is used to uniformly represent the structural information, jump characteristics, and disturbance response behavior from different path sources.

[0132] For each fusion path in the fusion path set, extract its node sequence between each layer of the black box model, identify whether there is a cross-layer jump relationship, and structurally label the cross-layer jump state with jump identifiers to generate a graph tracing chain.

[0133] Using the technical solution of this application, second data is generated based on first data; the second data is input into a second target model, and activation probability filtering is performed on the second data in each layer of the second target model to obtain a first path set; the second target model is obtained by introducing graph convolution operations and structural gating mechanisms into each hidden layer of the first target model; based on the first path set, an inter-layer node matching graph is constructed, and based on the inter-layer node matching graph, the first path set is filtered to obtain a second path set; structural perturbation signals are injected into each path in the candidate path graph to form a perturbation sample set; the candidate path graph is obtained by merging the first path set and the second path set; the... A perturbation sample set is input into the second target model to determine the output change offset magnitude, and a third path set is determined based on the output change offset magnitude. For each path in the third path set, a path contribution score is determined based on the output change offset magnitude. Based on the first path set, the second path set, the third path set, and the contribution scores of each path, a graph tracing chain corresponding to the first target model is generated. This avoids the limitations of traditional methods that require retrospective inference or rely solely on single-step gradient information, and achieves full-process modeling of path-level feature propagation, cross-layer semantic jumps, and response trajectories. As a result, a multi-path set with structural continuity and causal verifiability is constructed, forming a unified graph tracing chain.

[0134] Example 2

[0135] Figure 2 This invention provides a flowchart of another method for determining a spectral tracing chain. This embodiment further optimizes the process of generating second data based on first data in the aforementioned embodiments, building upon the previous embodiments. This embodiment can be combined with various optional solutions in one or more of the above embodiments. Figure 2 As shown, the spectrum tracing chain determination method of this embodiment may include the following steps:

[0136] S210. Normalize and group the first data to obtain the initial data.

[0137] S220. Map the initial data to a graph structure consisting of a set of nodes and a set of edges to obtain the second data. The nodes in the second data are the feature dimensions of the first data, and the edges represent the structural dependencies between features.

[0138] Get the first input data X, It performs normalization and feature grouping on the first input data, mapping the data to a set of nodes. With edge set The graph structure Each node Represents the first in the input data sequence Each feature dimension, each side Representation of features With features There is a structural dependency between them.

[0139] S230. Input the second data into the second target model, and perform activation probability filtering on the second data in each layer of the second target model to obtain the first path set; the second target model is obtained by introducing graph convolution operation and structural gating mechanism into each hidden layer of the first target model.

[0140] S240. A set of paths is used to construct an inter-layer node matching graph. Based on the inter-layer node matching graph, the first set of paths is filtered to obtain a second set of paths.

[0141] S250. Inject structural perturbation signals into each path in the candidate path graph to form a perturbation sample set; the candidate path graph is obtained by merging the first path set and the second path set.

[0142] S260. Input the perturbation sample set into the second target model, determine the output change offset magnitude, and determine the third path set based on the output change offset magnitude.

[0143] S270. For each path in the third path set, determine the path contribution score based on the output change offset magnitude.

[0144] S280. Based on the first path set, the second path set, the third path set, and the contribution scores of each path, generate the graph tracing chain corresponding to the first target model.

[0145] By adopting the technical solution of this application, the first data is normalized and feature grouped to obtain the initial data. The initial data is then mapped to a graph structure composed of a set of nodes and a set of edges to obtain the second data. The nodes in the second data are the feature dimensions of the first data, and the edges represent the structural dependencies between features. The transformation method of the first data into the second data is clarified, and the relationship between nodes and edges in the second data is also clarified.

[0146] Example 3

[0147] Figure 3 This invention provides a structural block diagram of a spectrogram tracing chain determination device, applicable to situations involving interpretability analysis of models. This spectrogram tracing chain determination device can be implemented in hardware and / or software and can be configured in an electronic device with data processing capabilities. Figure 3 As shown, the spectral tracing chain determination device of this embodiment may include: a second data generation module 310, a first set construction module 320, a second set construction module 330, a sample set generation module 340, a third set construction module 350, a contribution score calculation module 360, and a tracing chain generation module 370. Wherein:

[0148] The second data generation module 310 is used to generate second data based on the first data. The first data is the data input into the target model when using the first target model. The second data is based on the first data after graph structure transformation.

[0149] The first set construction module 320 is used to input the second data into the second target model and perform activation probability filtering on the second data in each layer of the second target model to obtain the first path set; the second target model is obtained by introducing graph convolution operation and structural gating mechanism into each hidden layer of the first target model.

[0150] The second set construction module 330 is used to construct an inter-layer node matching graph based on the first path set, and to filter the first path set based on the inter-layer node matching graph to obtain the second path set.

[0151] The sample set generation module 340 is used to inject structural perturbation signals into each path in the candidate path graph to form a perturbation sample set; the candidate path graph is obtained by merging the first path set and the second path set;

[0152] The third set construction module 350 is used to input the perturbation sample set into the second target model, determine the output change offset magnitude, and determine the third path set based on the output change offset magnitude;

[0153] The contribution score calculation module 360 ​​is used to determine the path contribution score for each path in the third path set based on the output change offset magnitude.

[0154] The tracing chain generation module 370 is used to generate a graph tracing chain corresponding to the first target model based on the first path set, the second path set, the third path set, and the contribution scores of each path.

[0155] Based on the above embodiments, optionally, the first set construction module 320 includes:

[0156] Based on the graph convolution operation and structure gating mechanism introduced in each hidden layer of the second target model, the activation probability of the edges in the graph structure of the second data is filtered, and the structure propagation path is extracted based on the maximum joint edge activation probability to generate the first path set.

[0157] Based on the above embodiments, optionally, the second set construction module 330 includes:

[0158] Based on the inter-layer node matching graph, the semantic embedding similarity and activation direction consistency between each node pair are determined.

[0159] Based on the semantic embedding similarity and activation direction consistency between node pairs, determine whether each node pair meets the preset jump threshold.

[0160] If the conditions are met, the node pair is retained, and a second path set is constructed based on the retained node pairs.

[0161] Based on the above embodiments, optionally, the third set construction module 350 includes:

[0162] The perturbation sample set is input into the second target model to determine the output changes caused by the perturbation of each path in the perturbation sample set;

[0163] Based on the output changes caused by disturbances along each path, determine the offset magnitude of the output change for each path.

[0164] If the output change offset is greater than the preset disturbance threshold, the corresponding path will be stored in the third path set.

[0165] Based on the above embodiments, optionally, the contribution scoring calculation module 360 ​​includes:

[0166] For each path in the third path set, a perturbation integral function is constructed based on the output change offset magnitude;

[0167] Based on the perturbation integral function, the path perturbation integral value of each path is determined;

[0168] The path contribution score is obtained by determining the ratio of the path disturbance integral value to the original output prediction.

[0169] Based on the above embodiments, optionally, the sample set generation module 340 includes:

[0170] Based on the disturbance response path identification algorithm, structural disturbance signals are injected into each path in the candidate path graph to form a disturbance sample set.

[0171] Based on the above embodiments, optionally, the second data generation module 310 includes:

[0172] The first data is normalized and feature-grouped to obtain the initial data.

[0173] The initial data is mapped to a graph structure consisting of a set of nodes and a set of edges to obtain the second data. The nodes in the second data are the feature dimensions of the first data, and the edges represent the structural dependencies between features.

[0174] The spectrum tracing chain determination device provided in the embodiments of the present invention can execute the spectrum tracing chain determination method provided in any embodiment of the present invention, and has the corresponding functional modules and beneficial effects of the execution method.

[0175] Example 4

[0176] Figure 4 A schematic diagram of an electronic device 10 that can be used to implement embodiments of the present invention is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices (such as helmets, glasses, watches, etc.), and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0177] like Figure 4As shown, the electronic device 10 includes at least one processor 11 and a memory, such as a read-only memory (ROM) 12 or a random access memory (RAM) 13, communicatively connected to the at least one processor 11. The memory stores computer programs executable by the at least one processor. The processor 11 can perform various appropriate actions and processes based on the computer program stored in the ROM 12 or loaded into the RAM 13 from storage unit 18. The RAM 13 can also store various programs and data required for the operation of the electronic device 10. The processor 11, ROM 12, and RAM 13 are interconnected via a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14.

[0178] Multiple components in electronic device 10 are connected to I / O interface 15, including: input unit 16, such as keyboard, mouse, etc.; output unit 17, such as various types of displays, speakers, etc.; storage unit 18, such as disk, optical disk, etc.; and communication unit 19, such as network card, modem, wireless transceiver, etc. Communication unit 19 allows electronic device 10 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0179] Processor 11 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of processor 11 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. Processor 11 performs the various methods and processes described above, such as the graph tracing chain determination method.

[0180] In some embodiments, the graph tracing chain determination method may be implemented as a computer program tangibly contained in a computer-readable storage medium, such as storage unit 18. In some embodiments, part or all of the computer program may be loaded and / or installed on electronic device 10 via ROM 12 and / or communication unit 19. When the computer program is loaded into RAM 13 and executed by processor 11, one or more steps of the graph tracing chain determination method described above may be performed. Alternatively, in other embodiments, processor 11 may be configured to perform the graph tracing chain determination method by any other suitable means (e.g., by means of firmware).

[0181] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0182] Computer programs used to implement the methods of the present invention may be written in any combination of one or more programming languages. These computer programs may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the processor, the computer programs cause the functions / operations specified in the flowcharts and / or block diagrams to be performed. The computer programs may be executed entirely on a machine, partially on a machine, or as a standalone software package, partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0183] In the context of this invention, a computer-readable storage medium can be a tangible medium that may contain or store a computer program for use by or in conjunction with an instruction execution system, apparatus, or device. A computer-readable storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination thereof. Alternatively, a computer-readable storage medium may be a machine-readable signal medium. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0184] To provide interaction with a user, the systems and techniques described herein can be implemented on an electronic device having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the electronic device. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0185] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as data servers), or middleware components (e.g., application servers), or frontend components (e.g., user computers with graphical user interfaces or web browsers through which users can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., communication networks). Examples of communication networks include local area networks (LANs), wide area networks (WANs), blockchain networks, and the Internet.

[0186] A computing system can include clients and servers. Clients and servers are generally located far apart and typically interact through communication networks. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. The server can be a cloud server, also known as a cloud computing server or cloud host, which is a hosting product within the cloud computing service system to address the shortcomings of traditional physical hosts and VPS services, such as high management difficulty and weak business scalability.

[0187] It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0188] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for determining a graph trace chain, characterized in that, The method comprises the following steps: generating second data based on first data, the first data being data input into a target model when using a first target model, the second data being based on graph structure transformation on the first data; inputting the second data into a second target model and performing activation probability screening on the second data in each layer of the second target model to obtain a first path set; the second target model being obtained by introducing graph convolution operation and structure gating mechanism into each hidden layer in the first target model; constructing an inter-layer node matching graph based on the first path set, and screening the first path set based on the inter-layer node matching graph to obtain a second path set; injecting a structure disturbance signal into each path in the candidate path graph to form a disturbance sample set; the candidate path graph is obtained by merging the first path set and the second path set; inputting the disturbance sample set into the second target model to determine an output change offset amplitude, and determining a third path set based on the output change offset amplitude; for each path in the third path set, determining a path contribution score based on the output change offset amplitude; generating a graph atlas tracking chain corresponding to the first target model based on the first path set, the second path set, the third path set, and the path contribution score of each path.

2. The method of claim 1, wherein, performing activation probability screening on the second data in each layer of the second target model to obtain a first path set, comprising: performing activation probability screening on the edges in the graph structure of the second data based on the introduction of graph convolution operation and structure gating mechanism into each hidden layer in the second target model, and extracting a structure propagation path based on the maximum joint edge activation probability to generate a first path set.

3. The method of claim 1, wherein, screening the first path set based on the inter-layer node matching graph to obtain a second path set, comprising: determining semantic embedding similarity and activation direction consistency between each node pair based on the inter-layer node matching graph; determining whether each node pair meets a preset jump threshold based on the semantic embedding similarity and the activation direction consistency between the node pairs; if so, retaining the node pair, and constructing a second path set based on the retained node pairs.

4. The method of claim 1, wherein, inputting the disturbance sample set into the second target model to determine an output change offset amplitude, and determining a third path set based on the output change offset amplitude, comprising: inputting the disturbance sample set into the second target model to determine the output change caused by each path disturbance in the disturbance sample set; determining the output change offset amplitude of each path based on the output change caused by each path disturbance; if the output change offset amplitude is greater than a preset disturbance threshold, the corresponding path is stored in the third path set.

5. The method of claim 1, wherein, for each path in the third path set, determining a path contribution score based on the output change offset amplitude, comprising: for each path in the third path set, constructing a disturbance integral function based on the output change offset amplitude; determining a path disturbance integral value of each path based on the disturbance integral function; determining the ratio of the path disturbance integral value to the original output prediction to obtain the path contribution score.

6. The method of claim 1, wherein, The structural perturbation signal is injected into each path in the candidate path graph respectively to form a perturbation sample set, including: The structural perturbation signal is injected into each path in the candidate path graph respectively based on the perturbation response path identification algorithm to form a perturbation sample set.

7. The method of claim 1, wherein, Based on the first data, the second data is generated, including: The first data is normalized and feature grouped to obtain initial data; The initial data is mapped to a graph structure composed of a node set and an edge set to obtain the second data, wherein the node in the second data is the feature dimension in the first data, and the edge represents the structural dependency relationship between the features.

8. A graph trace chain determination apparatus characterized by comprising: Including: The second data generation module is configured to generate the second data based on the first data, wherein the first data is the data input into the target model when the first target model is used, and the second data is obtained after graph structure transformation on the first data; The first set construction module is configured to input the second data into the second target model, and perform activation probability screening on the second data at each layer in the second target model to obtain a first path set; the second target model is obtained by introducing graph convolution operation and structural gating mechanism into each hidden layer in the first target model; The second set construction module is configured to construct an inter-layer node matching graph based on the first path set, and screen the first path set based on the inter-layer node matching graph to obtain a second path set; The sample set generation module is configured to inject a structural perturbation signal into each path in the candidate path graph to form a perturbation sample set; the candidate path graph is obtained by merging the first path set and the second path set; The third set construction module is configured to input the perturbation sample set into the second target model, determine an output change offset amplitude, and determine a third path set based on the output change offset amplitude; The contribution score calculation module is configured to determine a path contribution score based on the output change offset amplitude for each path in the third path set; The tracking chain generation module is configured to generate a graph atlas tracking chain corresponding to the first target model based on the first path set, the second path set, the third path set, and the path contribution score.

9. An electronic device, comprising: The electronic device includes: At least one processor; and The memory is in communication with the at least one processor; wherein The memory stores a computer program executable by the at least one processor, and the computer program is executed by the at least one processor to enable the at least one processor to execute the graph atlas tracking chain determination method of any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium stores computer instructions for enabling the processor to execute the graph atlas tracking chain determination method of any one of claims 1-7 when executed.

Images