Statistical result dynamic visualization and diagnosis method based on interpretable AI

Abstract

The invention relates to the technical field of artificial intelligence, and discloses an interpretability AI-based statistical result dynamic visualization and diagnosis method, which comprises the following steps of: aiming at each data sample of a machine learning model, applying various interpretability algorithms in parallel, and generating and splicing a high-dimensional integrated interpretation vector; calculating the similarity between vectors through a multi-component interpretation kernel function, carrying out the dimension reduction of a high-dimensional vector set through a manifold learning algorithm, and constructing a low-dimensional interpretation manifold capable of reflecting the internal structure of the model decision logic; density clustering is carried out on the interpretation manifold, and an interpretation stability domain consistent with model decision logic and a logic mutation boundary with logic mutation are automatically recognized; the analysis result is visually presented, and an interactive diagnosis function is provided. According to the method, the global topology view of the model decision logic is constructed, so that the user can intuitively understand and diagnose the behavior pattern of the complex model, and the interpretability, transparency and reliability of the model are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, in particular to a statistical result dynamic visualization and diagnosis method based on explainable AI. BACKGROUND

[0002] In recent years, machine learning models represented by deep neural networks and ensemble models have made remarkable achievements in image recognition, natural language processing, financial risk control and many other fields. However, the strong performance of these models often comes at the cost of the high complexity and opacity of their internal decision-making logic, which makes them often referred to as "black boxes". This "black box" characteristic greatly limits the application of models in medical, autonomous driving, judicial and other high-risk and high-reliability scenarios, because users not only need to know the prediction results of the model, but also need to understand and trust the reasons for obtaining the results.

[0003] In order to open the "black box", explainable artificial intelligence (XAI) technology has emerged. In the prior art, methods such as SHAP (SHapley Additive exPlanations) or LIME (Local Interpretable Model-agnostic Explanations) are widely used to generate local explanations for individual data samples, i.e. to explain how the model makes a specific prediction for a specific input.

[0004] Although these local explanation methods improve the transparency of the model to some extent, there are still deep-seated limitations in their application. These methods essentially provide a microscopic perspective of "looking at a part of an elephant through a small hole in the tube", and users have to check a large number of isolated and discrete local explanations and try to inefficiently and subjectively induce and splice in their minds if they want to understand the global behavior pattern of the model. This process is not only tedious, but also almost impossible to reveal the macro structure, internal relations and transition relationships between different regions of the model decision-making logic in the entire data space.

[0005] Furthermore, the prior art fails to effectively address the stability of the explanation. The decision-making logic of the model is not immutable, and when the input data changes slightly, the explanation result may experience a dramatic, nonlinear "mutation". Existing methods are difficult to systematically identify regions where the model decision-making logic remains stable, and are also unable to clearly demarcate the boundaries where the logic undergoes mutation. Users therefore cannot predict under what circumstances the "way of thinking" of the model will fundamentally change, which directly affects the prediction and trust of the model behavior. In addition, the current technical paradigm mostly provides static and one-time analysis reports, lacking the ability to dynamically interact with users, and users cannot actively probe and diagnose the logic boundaries of the model by adjusting the input in real time, making it difficult to explore and verify the internal mechanism of the model at a deeper level. SUMMARY

[0006] The explainable AI methods provided in the prior art usually generate static and isolated explanation results. This makes it difficult for users to dynamically and interactively diagnose the decision logic of machine learning models, and cannot systematically reveal the stability of the decision patterns within the model and the transition boundaries between different decision patterns.

[0007] The present application aims to solve the above technical problems and provide a method for dynamically visualizing and diagnosing the statistical results of AI models, to realize the automated analysis and display of model decision patterns, stable regions and logic mutation boundaries.

[0008] The first aspect of the present application provides a method for dynamically visualizing and diagnosing the statistical results of explainable AI, comprising the following steps:

[0009] Obtaining a plurality of data samples for a predetermined machine learning model;

[0010] For each data sample in the plurality of data samples, generating a high-dimensional integrated explanation vector by applying at least two explainable AI algorithms and splicing their explanation results;

[0011] Based on the plurality of integrated explanation vectors corresponding to the plurality of data samples, constructing a low-dimensional explanation manifold that shows the relationship between the plurality of integrated explanation vectors by a manifold learning algorithm;

[0012] Topologically analyzing the explanation manifold to identify explanation stable domains representing stable decision patterns, and logic mutation boundaries between the explanation stable domains;

[0013] Visualizing and displaying the explanation manifold, the explanation stable domains and the logic mutation boundaries.

[0014] In one specific embodiment, the at least two explainable AI algorithms include two or more of a method based on game theory, a method based on local proxy model or a method based on gradient. The explanation component vectors generated by this step are spliced into the integrated explanation vector.

[0015] In one specific embodiment, in the step of constructing a low-dimensional explanation manifold, the process of determining the relationship between the plurality of integrated explanation vectors is realized by calculating a multi-component explanation kernel function value. The step of calculating the multi-component explanation kernel function value specifically includes:

[0016] For any two integrated explanation vectors E i and E j , calculate the values of at least two of the following sub-kernel functions and perform weighted summation:

[0017] 1. A direction similarity kernel function K sim (E i ,E j ), whose value is used to represent the direction consistency of the two ensemble explanation vectors. One specific calculation method is:

[0018]

[0019] wherein ||E i -E j || is the Euclidean distance between vectors E i and E j , and σ is a preset bandwidth parameter.

[0020] 2. An order consistency kernel function K rank (E i ,E j ), whose value is used to represent the similarity of the feature contribution degree order in the two ensemble explanation vectors. One specific calculation method is to calculate the Spearman rank correlation coefficient of the element values in the two vectors.

[0021] 3. A key magnitude kernel function K mag (E i ,E j ), whose value is used to represent the closeness of the contribution magnitude of a preset number of key features in the two ensemble explanation vectors. One specific calculation method is:

[0022] K mag ,E j )=(1+||T(E i ,k)-T(E j ,k)||) -1 ;

[0023] wherein T(E i ,k) is a function of extracting a sub-vector composed of the k elements with the highest absolute values in vector E, and k is a preset number of key features.

[0024] Preferably, the manifold learning algorithm comprises at least one of UMAP, Isomap or t-SNE.

[0025] In one specific embodiment, the step of identifying the explanation stable domain is specifically:

[0026] In the low-dimensional space where the explanation manifold is located, a density clustering algorithm is applied to identify the points in the explanation manifold as at least one high-density cluster and boundary points; wherein the high-density cluster is defined as the explanation stable domain.

[0027] Further, the method further comprises calculating a boundary steepness of the logical mutation boundary. The boundary steepness S is defined as:

[0028] The ratio of the distance d exp between two points on either side of the logical mutation boundary, which belong to different explanation stable domains, in the original high-dimensional integrated explanation vector space, to the distance d emb between the two points in the low-dimensional explanation manifold space. The specific calculation is:

[0029]

[0030] where d exp is derived from the value of the multi-component explanation kernel function K, and d emb is the Euclidean distance between the two points in the low-dimensional space.

[0031] Preferably, the step of visualizing further comprises:

[0032] Displaying the numerical value of the boundary steepness on the logical mutation boundary in the visual attribute of color, thickness or transparency.

[0033] In one specific embodiment, the method further comprises:

[0034] In response to a selection operation on any of the explanation stable domains, aggregating the feature statistics and average integrated explanation vector of all data samples in the explanation stable domain to generate a semantic description of the explanation stable domain.

[0035] In one specific embodiment, the method further comprises:

[0036] In response to a modification of the original feature value of any data sample, recalculating the integrated explanation vector of the modified data sample in real time;

[0037] According to the updated integrated explanation vector, determining its new position on the explanation manifold;

[0038] In the visualization display, presenting the moving track of the data sample from the old position to the new position on the explanation manifold.

[0039] The second aspect of the present application provides an interpretable AI-based statistical result dynamic visualization and diagnosis system, comprising:

[0040] a processor;

[0041] a memory, the memory storing computer executable instructions;

[0042] When the processor executes the computer executable instructions, it is configured to perform the method as described in the preceding first aspect.

[0043] The application provides an explainable AI-based statistical result dynamic visualization and diagnosis method. The following beneficial effects are achieved:

[0044] 1. The application can systematically reveal the inherent decision-making mode of a machine learning model. First, an integrated explanation vector is generated for each data sample, then an explanation manifold is constructed to represent the relationship between these vectors, and further, an explanation stable domain and a logical mutation boundary are automatically identified through topological analysis. This process converts a series of discrete and local explanation results into a structured and global decision-making logic view, thereby enabling systematic analysis of the behavior pattern of the model on the entire data set.

[0045] 2. The application can quantitatively evaluate the stability of the decision-making logic of the model. This is achieved by calculating the boundary steepness after identifying the logical mutation boundary between different explanation stable domains; the boundary steepness is defined as the ratio of the distance change in the high-dimensional integrated explanation vector space to the distance change in the low-dimensional explanation manifold space. This ratio provides a specific and comparable numerical value for the degree of logical transition of the model at the decision-making boundary, thereby realizing quantitative evaluation of the robustness of the model.

[0046] 3. The application provides an intuitive method for dynamic diagnosis of the decision-making logic of the model. When the user modifies the original feature value of any data sample, the system can recalculate its integrated explanation vector and update its position on the explanation manifold in real time. By presenting this position change as a visible moving track, the user can directly observe how the data sample point moves between different explanation stable domains and logical mutation boundaries, thereby intuitively understanding the causal relationship between the change of the input data and the transition of the decision-making logic of the model. BRIEF DESCRIPTION OF DRAWINGS

[0047] Figure 1 The figure is a hardware structure diagram of an explainable AI-based statistical result dynamic visualization and diagnosis system according to an embodiment of the application.

[0048] Figure 2 The figure is a functional module diagram of an explainable AI-based statistical result dynamic visualization and diagnosis system according to an embodiment of the application.

[0049] Figure 3 The figure is a flowchart of an explainable AI-based statistical result dynamic visualization and diagnosis method according to an embodiment of the application.

[0050] Figure 4 The figure is a schematic diagram of the explanation manifold construction process according to an embodiment of the application.

[0051] Wherein, 101, processor; 102, memory; 103, system bus; 104, input / output interface; 105, communication interface; 10, data acquisition module; 20, integrated interpretation vector generation module; 30, interpretation manifold construction module; 40, topology analysis module; 50, visualization and interaction module. DETAILED DESCRIPTION

[0052] In order to make the objects, technical solutions and advantages of the present application clearer, the specific embodiments of the present application will be described in detail below with reference to the drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0053] Referring to the drawings Figure 1 , Figure 1 is a hardware structure schematic diagram of an interpretable AI-based statistical result dynamic visualization and diagnosis system according to an embodiment of the present application. The system can be a server, a personal computer, a workstation or any device with computing capability. The system includes: at least one processor 101, such as a central processing unit (CPU); a memory 102 for storing program codes and data; a system bus 103 for connecting the processor 101 and the memory 102 and other components; an input / output interface 104 for connecting external devices such as a display, a keyboard, a mouse, etc.; and a communication interface 105 for data communication with other devices. The memory 102 stores computer program instructions, and when the processor 101 executes these instructions, the method described later is implemented.

[0054] Referring to the drawings Figure 2 , Figure 2 is a functional module diagram of an interpretable AI-based statistical result dynamic visualization and diagnosis system according to an embodiment of the present application. In this embodiment, the system can include: a data acquisition module 10, an integrated interpretation vector generation module 20, an interpretation manifold construction module 30, a topology analysis module 40 and a visualization and interaction module 50.

[0055] The data acquisition module 10 is configured to acquire a preset, trained machine learning model from the memory 102 or an external data source via the communication interface 105, and acquire a plurality of data samples for diagnosing the model. After acquisition, the machine learning model and the plurality of data samples are provided to the integrated interpretation vector generation module 20.

[0056] An integrated explanation vector generation module 20 receives the machine learning model and the data samples provided by the data acquisition module 10. For each data sample, the module calls at least two different types of explainable AI algorithm libraries, respectively calculates the corresponding explanation component vectors, and then splices the explanation component vectors in a predetermined order to form a high-dimensional integrated explanation vector. The module outputs the set of integrated explanation vectors corresponding to all data samples to the explanation manifold construction module 30.

[0057] The explanation manifold construction module 30 receives the set of integrated explanation vectors output by the integrated explanation vector generation module 20. The module first calculates the similarity values between any two integrated explanation vectors according to a preset multi-component explanation kernel function, thereby constructing a similarity matrix. Subsequently, the module inputs the similarity matrix into a preset manifold learning algorithm for dimension reduction calculation, and finally generates a set of low-dimensional coordinate points representing the relationship between the integrated explanation vectors. The set of low-dimensional coordinate points is defined as an explanation manifold and is output to the topology analysis module 40 and the visualization and interaction module 50.

[0058] The topology analysis module 40 receives the explanation manifold output by the explanation manifold construction module 30. The module first performs a density clustering algorithm on the set of low-dimensional coordinate points of the explanation manifold to identify high-density point clusters and boundary points, wherein the high-density point clusters are defined as explanation stable domains and the boundary points are defined as logical mutation boundaries. Subsequently, the module calculates the boundary steepness of each point on the logical mutation boundary. The module outputs the identified explanation stable domains, logical mutation boundaries, and calculated boundary steepness data to the visualization and interaction module 50.

[0059] The visualization and interaction module 50 receives the explanation manifold data from the explanation manifold construction module 30 and the analysis result data from the topology analysis module 40. The module is used to render these data into a graphical user interface and display it on a display device. At the same time, the module is also used to receive user operation instructions transmitted through an input device and perform corresponding interaction functions according to the instructions, for example, when receiving a user's modification instruction for a certain data sample's original feature value, the module sends the modified data sample to the integrated explanation vector generation module 20 for triggered recalculation.

[0060] Referring to the accompanying drawings Figure 3 , Figure 3 is a flowchart of an explainable AI-based statistical result dynamic visualization and diagnosis method according to an embodiment of the present application. The specific process of the method can include the following steps:

[0061] Step S301, a preset trained machine learning model and a plurality of data samples for diagnosing the machine learning model are obtained.

[0062] Step S302: For each data sample among multiple data samples, apply at least two interpretability AI algorithms to calculate the corresponding interpretability component vectors, and concatenate the interpretability component vectors to generate a high-dimensional integrated interpretability vector.

[0063] Step S303: Based on the integrated explanatory vector set corresponding to all data samples, calculate the pairwise similarity by applying a preset multi-component explanatory kernel function, and use the manifold learning algorithm to reduce dimensionality, thereby constructing a low-dimensional explanatory manifold.

[0064] Step S304: Perform topological analysis on the explanatory manifold, identify the explanatory stable region and the logical mutation boundary by applying the density clustering algorithm, and further calculate the boundary steepness of the logical mutation boundary.

[0065] Step S305: Integrate and render the interpretive manifold, as well as the interpretive stable regions and logical mutation boundaries marked on the interpretive manifold, and finally generate and display a visual diagnostic interface.

[0066] The following will provide a detailed explanation of each step in the above process.

[0067] exist Figure 3 In the illustrated process, step S301 is the preparatory work before performing the diagnosis. This step first requires obtaining a pre-trained machine learning model. The machine learning model can be any type of supervised learning model, such as a logistic regression model, support vector machine, gradient boosting decision tree (GBDT), or deep neural network (DNN) for classification tasks, or a linear regression model or regression tree for regression tasks. The model's state is fixed, meaning its internal parameters, such as weights, biases, or tree structure, have been determined through the training process, and the model already possesses interface functions for predicting new data.

[0068] This step requires acquiring multiple data samples for diagnostics. These samples can originate from training sets used to train the model, validation sets used to tune hyperparameters, test sets used to evaluate model performance, or production datasets encountered in real-world applications. The structure and format of each data sample must match the input requirements of the machine learning model. For example, for a model processing tabular data, each data sample is a feature vector, with each dimension corresponding to a specific business or physical feature.

[0069] The acquired data sample set should be sufficient in quantity to support subsequent topology analysis, so as to be able to discover statistically significant patterns therefrom. In addition, the content of the data sample set should cover various input situations that the model may encounter, such as containing samples correctly predicted by the model, samples incorrectly predicted by the model, and samples with low prediction confidence of the model. This provides the necessary data basis for subsequent comprehensive analysis of the decision-making behavior of the model.

[0070] In Figure 3 The step S302 is to generate, for each data sample, an integrated high-dimensional explanation vector that can comprehensively describe the basis on which the machine learning model makes decisions.

[0071] For each data sample x i obtained from step S301, the system applies a preset algorithm set composed of at least two different types of interpretable AI algorithms in parallel or in series. The algorithm set can include:

[0072] An algorithm based on game theory, such as SHAP (Shapley Additive exPlanations). This algorithm assigns a contribution degree of each feature to the prediction result of the sample by calculating the marginal contribution of each feature in all possible feature combinations and taking the Shapley value as the result. After applying this algorithm, an explanation component vector V shap (x i ) is obtained, which has the same dimension as the feature dimension of the data sample.

[0073] An algorithm based on local proxy model, such as LIME (Local Interpretable Model-agnostic Explanations). This algorithm generates a set of perturbed samples within the neighborhood of the data sample x i , and uses a preset machine learning model to make predictions on these perturbed samples. Then, a simple model with interpretability (such as a linear regression model or a decision tree) is used to fit these perturbed samples and their prediction results. The parameters of this simple model (such as the weights of the linear model) are taken as the explanation of the prediction result of the original sample x i , forming an explanation component vector V lime (x i ).

[0074] When the preset machine learning model is a differentiable model (such as a deep neural network), the algorithm set can also include a gradient-based method. This method calculates the gradient of the model output value with respect to the input data sample x iThe gradient vector is obtained by taking the partial derivatives of each feature dimension. This gradient vector, or its derivatives (e.g., the product of the gradient and the input), can be used as the explanatory component vector V. gard (x i The value represents the sensitivity of the model output to small changes in the input features.

[0075] After calculating the explanatory component vectors corresponding to all preset algorithms, this step concatenates these vectors in a fixed, predefined order to form a higher-dimensional, more comprehensive integrated explanatory vector E. i This splicing operation can be represented as:

[0076]

[0077] in, This represents the vector concatenation operation. By performing this step on all data samples, a final set {E1, E2, ..., E...} of multiple ensemble interpretation vectors is generated. N}, where N is the total number of data samples. This set will be used for subsequent interpretative manifold construction.

[0078] See attached document Figure 4 , Figure 4 This is a schematic diagram of the explanatory manifold construction process according to an embodiment of the present invention. In step S303, this step receives the integrated explanatory vector set {E1, E2, ..., E...} generated in step S302. N This process maps the high-dimensional space to a low-dimensional space to construct the interpretable manifold. Specifically, this step involves two stages: calculating similarity values ​​and performing dimensionality-reduced embedding.

[0079] During the similarity calculation stage, the system uses a pre-defined multi-component interpretation kernel function K to measure the similarity between any two integrated interpretation vectors E. i and E j The degree of similarity between them. The value of the kernel function is defined as a weighted sum of the values ​​of multiple sub-kernel functions, in the form of:

[0080] K(E i E j ) = w sim ·K sim (E i E j )+w rank ·K rank (E i E j )+w mag ·K mag (E i E j );

[0081] where w sim ,w rank ,w mag are preset weight coefficients greater than zero, and their sum is 1.

[0082] The first sub-kernel function is a direction similarity kernel K sim , which is used to measure the overall direction consistency of two integrated explanation vectors in the vector space. In the embodiment, it is calculated by a Gaussian radial basis function:

[0083]

[0084] where ||E i -E j || represents the L2 norm, i.e., the Euclidean distance, between vectors E i and E j ; and σ is a preset bandwidth hyperparameter greater than zero.

[0085] The second sub-kernel function is an order consistency kernel K rank , which is used to measure the consistency of the ordering of the numerical values of the elements in two integrated explanation vectors. The calculation process is as follows: first, generate the ordering vectors R(E i ) and R(E j ) corresponding to the two integrated explanation vectors E i and E j , respectively. Each element in the ordering vector is the rank of the numerical value of the corresponding element in the original vector. Then, calculate the Spearman rank correlation coefficient of the two ordering vectors R(E i ) and R(E j ), and the coefficient value is K rank (E i ,E j ).

[0086] The third sub-kernel function is a key magnitude kernel K mag , which is used to measure the proximity of the contribution magnitude of the preset number of key features with the highest absolute values in two integrated explanation vectors. The calculation process is as follows: first, according to a preset integer k, extract the sub-vectors T(E i ,k) and T(E j ,k) composed of the k elements with the highest absolute values from vectors E i and E j , respectively. Then, calculate the kernel function value by the following formula:

[0087] K mag (E i ,E j ) = (1 + ||T(E i ,k) - T(Ej ,k)||) -1 ;

[0088] Among them, ||T(E i ,k)-T(E j ,k)|| denotes the L2 norm between two key feature vectors.

[0089] After defining the multi-component interpretation kernel function, this step applies the integrated interpretation vector set {E1, E2, ..., E...} N Applying this kernel function to each pair of vectors in the matrix results in an N×N symmetric similarity matrix M. The elements M in matrix M... ij The value is equal to K(E) i E j ).

[0090] In this embodiment, the weight coefficient w in the multi-component interpretation kernel function sim ,w rank ,w mag The specific values ​​of the bandwidth hyperparameter σ in the directional similarity kernel function and the number of key features k in the key magnitude kernel function can be predetermined based on experience or determined by experimental methods such as grid search on the validation dataset with the goal of obtaining the best visualization effect.

[0091] During the dimensionality reduction embedding stage, the similarity matrix M is used as input to the manifold learning algorithm. In this embodiment, the Uniform Manifold Approximation and Projection (UMAP) algorithm can be employed. The UMAP algorithm constructs a weighted graph structure in the high-dimensional space based on the similarity matrix M, and then searches for an optimal low-dimensional (e.g., two-dimensional or three-dimensional) embedding representation that best preserves the topological structure of the original high-dimensional space. The output of this algorithm is a set of low-dimensional coordinate points {p1, p2, ..., p...}. N}, where each coordinate point p i Corresponding to the original integrated interpretation vector E i This set of coordinate points in low-dimensional space together constitutes the interpretive manifold for visualization.

[0092] It should be understood that although this embodiment uses the UMAP algorithm as an example, any manifold learning algorithm that can accept a similarity matrix or distance matrix as input and perform nonlinear dimensionality reduction is applicable to this invention. Other feasible algorithms include, but are not limited to: Isomap, t-distributed random neighborhood embedding (t-SNE), or locally linear embedding (LLE).

[0093] In step S304, this step receives the low-dimensional interpreted manifold generated in step S303 and performs automated analysis of its topology.

[0094] This step will first explain the manifold, i.e. the low-dimensional coordinate point set {p1, p2, …, pN} as input, and apply a preset density clustering algorithm. In this embodiment, the algorithm used is HDBSCAN (Hierarchical Density-Based Spatial Clustering of Applications with Noise). The HDBSCAN algorithm processes the input data point set, and divides it into several clusters and a set of noise points according to the density distribution of the points. This algorithm does not need to specify the number of clusters in advance, and can identify clusters of any shape. N} will be assigned a label indicating the cluster to which it belongs or marked as noise. This step defines the region composed of all data points assigned to any one cluster label as the explanation stable domain of the explanation manifold. Each independent cluster corresponds to an independent explanation stable domain. The set of points marked as noise by the algorithm is defined as the logical mutation boundary between explanation stable domains.

[0095] After the HDBSCAN algorithm is executed, each data point p i will be assigned a label indicating the cluster to which it belongs or marked as noise. This step defines the region composed of all data points assigned to any one cluster label as the explanation stable domain of the explanation manifold. Each independent cluster corresponds to an independent explanation stable domain. The set of points marked as noise by the algorithm is defined as the logical mutation boundary between explanation stable domains.

[0096] Subsequently, this step quantitatively analyzes the logical mutation boundary identified, i.e. calculates its boundary steepness. The calculation process is as follows: for any pair of adjacent points p a and p b belonging to two different explanation stable domains C i and C a , the boundary steepness S j is calculated. b ij

[0097] First, the distance d emb between the two points in the low-dimensional explanation manifold space is calculated, which is the Euclidean distance in this embodiment:

[0098] d emb = ||p i -p j ||;

[0099] Then, according to the indices i and j, the original high-dimensional integrated explanation vectors E i and E j generated in step S302 without dimension reduction are retrieved. The distance d exp between the two high-dimensional vectors in the original space is calculated, which is also the Euclidean distance in this embodiment:

[0100] d exp = ||E i -E​​j ||;

[0101] Finally, the boundary steepness S is calculated using the following formula. ij :

[0102]

[0103] In this embodiment, the high-dimensional distance d exp and low-dimensional distance d emb The calculations are all illustrated using Euclidean distance (L2 norm) as an example. Depending on the specific distribution characteristics of the data, other distance metrics can also be used, such as Manhattan distance (L1 norm) or Minkowski distance.

[0104] This step repeats the calculation for all points on the logistic mutation boundary or multiple pairs of nearby points to obtain a set of boundary steepness values. These values, along with the results of interpreting the stable region and identifying the logistic mutation boundary, are output for subsequent visualization.

[0105] In step S305, the purpose of this step is to integrate all the analysis results generated in the previous steps, generate a graphical user interface, and realize interactive diagnostic functions with the user.

[0106] First, this step generates a static visualization interface by receiving the set of low-dimensional coordinate points {p1, p2, ..., p} of the interpretive manifold generated in step S303. N} and render these points as a scatter plot in a two-dimensional or three-dimensional coordinate system to form a basic view for interpreting the manifold.

[0107] Next, the interpretable stability region information identified in step S304 is received, and an independent visual identifier is assigned to each interpretable stability region, for example, by using different colors or symbols to mark points belonging to different stability regions.

[0108] Simultaneously, it receives the identification results of logical abrupt change boundaries and the calculated boundary steepness values. For points defined as logical abrupt change boundaries, the system uses a different unified visual identifier (e.g., gray) distinct from the interpreted stable region for labeling.

[0109] Furthermore, based on the boundary steepness value of each boundary point or neighboring point pair, it is rendered into different visual attributes through a preset visual mapping function. For example, the boundary steepness value is linearly or logarithmically mapped to the color depth, thickness, or transparency of the boundary line, where a higher boundary steepness value corresponds to a darker, thicker, or more opaque visual appearance.

[0110] Secondly, this step implements two core interactive diagnostic functions.

[0111] The first interactive function is the generation of semantic descriptions of interpretable stable regions. When the system detects that a user has selected a specific interpretable stable region on the visual interface (e.g., by clicking with the mouse) through input / output interface 104, the system will perform the following operations:

[0112] First, retrieve the indices of all data sample points belonging to this interpretation's stability region;

[0113] Then, based on these indices, the corresponding subset of data samples is extracted from the original dataset;

[0114] Next, statistical analysis is performed on each feature dimension of the data sample subset to calculate its mean, variance, median, mode, and other statistical measures.

[0115] At the same time, the system also calculates the average vector of all integrated interpretation vectors within the stability region.

[0116] Finally, the system combines these statistical information and key features from the average explanatory vector into a structured text as a semantic description of the explanatory stability domain, and displays it in a preset area of ​​the interface.

[0117] The second interactive function is dynamic trajectory tracking of data samples. When the system detects that the user has selected a data sample point on the interface and modified one or more feature values ​​of the corresponding original data sample, the system will trigger a real-time recalculation process. This process will pass the modified new data sample back to step S302 and execute some calculations in S302 and S303 in sequence. That is, for only this modified sample, its integrated interpretation vector will be regenerated, and its new position coordinates on the low-dimensional interpretation manifold will be calculated using the constructed manifold model.

[0118] After obtaining the new location coordinates, the visualization interface executes an animation process that smoothly moves the point representing the data sample from its old location to the new location, leaving a visual trajectory line along its movement path. This trajectory line visually demonstrates how changes in the input data cause the model's interpretation to shift its location on the interpretive manifold.

[0119] It is worth noting that when calculating the new position coordinates of the modified data samples, this embodiment utilizes the transform function of the manifold learning model (e.g., the trained UMAP model) that has been built and solidified in step S303. This process does not require retraining the manifold model for all data points; it only projects the new high-dimensional ensemble interpretation vector onto the existing low-dimensional interpretation manifold space. Therefore, the computational overhead is small, and it can support real-time or near-real-time interactive feedback.

[0120] In summary, the method and system provided by the embodiments of the present invention generate integrated explanatory vectors for data samples, construct a low-dimensional explanatory manifold based on the vector set, analyze the topological structure of the manifold to identify explanatory stable regions and logical mutation boundaries, and finally visualize and interactively diagnose the analysis results, thereby realizing a systematic analysis of the internal decision-making logic of machine learning models.

[0121] Finally, it should be noted that the above descriptions are merely specific embodiments of the present invention, intended to illustrate the technical solutions of the present invention rather than to limit them. For those skilled in the art, any modifications to the technical solutions in the above embodiments, equivalent substitutions of some technical features, or combinations based thereon, made without departing from the spirit and principles of the present invention, should fall within the protection scope of the appended claims.

Claims

1. A method for dynamic visualization and diagnosis of statistical results based on interpretable AI, characterized in that, Includes the following steps: Obtain multiple data samples for a predefined machine learning model; For each of the plurality of data samples, a high-dimensional integrated interpretation vector is generated by applying at least two interpretability AI algorithms and concatenating their interpretation results. Based on the multiple ensemble explanatory vectors corresponding to the multiple data samples, a low-dimensional explanatory manifold that displays the relationship between the multiple ensemble explanatory vectors is constructed through a manifold learning algorithm; A topological analysis is performed on the explanatory manifold to identify explanatory stable regions representing stable decision-making patterns, as well as logical abrupt boundary boundaries between the explanatory stable regions; The explanatory manifold, the explanatory stable region, and the logical mutation boundary are visualized.

2. The method for dynamic visualization and diagnosis of statistical results based on interpretable AI according to claim 1, characterized in that, The steps of applying at least two interpretable AI algorithms specifically include: The application includes at least two of the following methods: game theory-based methods, local proxy model-based methods, or gradient-based methods, to generate corresponding explanatory component vectors, and then concatenates the explanatory component vectors into the integrated explanatory vector.

3. The method for dynamic visualization and diagnosis of statistical results based on interpretable AI according to claim 1, characterized in that, In the step of constructing a low-dimensional interpretive manifold that demonstrates the relationship between the multiple integrated interpretive vectors, the process of determining the relationship between the multiple integrated interpretive vectors is achieved by calculating the multi-component interpretive kernel function value; The steps for calculating the multi-component interpretation kernel function value specifically include: For any two of the aforementioned ensemble interpretation vectors, calculate the values ​​of at least two of the following sub-kernel functions and perform a weighted sum: The orientation similarity kernel function, whose value is used to characterize the orientational consistency of the two integrated interpretation vectors; The ranking consistency kernel function is used to characterize the similarity of the feature contribution ranking in the two integrated explanatory vectors; A key magnitude kernel function, the value of which is used to characterize the proximity of the contribution magnitudes of a predetermined number of key features in the two integrated interpretation vectors.

4. The method for dynamic visualization and diagnosis of statistical results based on interpretable AI according to claim 1, characterized in that, The manifold learning algorithm includes at least one of UMAP, Isomap, or t-SNE.

5. The method for dynamic visualization and diagnosis of statistical results based on interpretable AI according to claim 1, characterized in that, The specific steps for identifying the explanatory stability region representing the stable decision-making pattern are as follows: In the low-dimensional space where the explanatory manifold is located, a density clustering algorithm is applied to identify the points in the explanatory manifold as at least one high-density cluster and boundary points; wherein, the high-density cluster is defined as the explanatory stable region.

6. The method for dynamic visualization and diagnosis of statistical results based on interpretable AI according to claim 5, characterized in that, The method further includes: Calculate the boundary steepness of the logical mutation boundary; The boundary steepness is defined as the ratio of the distance between two points belonging to different interpretive stable domains on either side of the logical mutation boundary in the original high-dimensional integrated interpretive vector space to the distance between the two points in the low-dimensional interpretive manifold space.

7. The method for dynamic visualization and diagnosis of statistical results based on interpretable AI according to claim 6, characterized in that, The visualization process also includes: The numerical value of the boundary steepness is displayed on the logically abrupt boundary using visual attributes such as color, thickness, or transparency.

8. The method for dynamic visualization and diagnosis of statistical results based on interpretable AI according to claim 5, characterized in that, The method further includes: In response to the selection operation of any of the aforementioned interpretive stable regions, the feature statistics and average ensemble interpretive vector of all data samples within the interpretive stable region are aggregated to generate a semantic description of the interpretive stable region.

9. The method for dynamic visualization and diagnosis of statistical results based on interpretable AI according to claim 1, characterized in that, The method further includes: In response to a modification to the original feature value of any data sample, the integrated interpretation vector of the modified data sample is recalculated in real time. Based on the updated integrated interpretation vector, determine its new position on the interpretation manifold; The visualization presents the trajectory of the data sample as it moves from an old position to a new position on the explanatory manifold.

10. A dynamic visualization and diagnostic system for statistical results based on interpretable AI, characterized in that, include: processor; A memory that stores computer-executable instructions; When the processor executes the computer-executable instructions, it implements the method as described in any one of claims 1-9.

Images