Data interpretable prediction method and device based on tree regular variational encoder

Abstract

The invention relates to the technical field of neural networks, and discloses a data interpretable prediction method and device based on a tree regular variational encoder, and the method comprises the steps: encoding target high-dimensional data to a submerged space through a variational automatic encoder, and obtaining latent variable representation; performing regularized alignment processing on the latent variable representation based on the agent decision-making tree so as to enable a distribution structure of the latent variable representation in the submerged space to be consistent with a rule path of the agent decision-making tree, and generating a regularized submerged space; performing classification prediction based on the regularized subsurface space, and outputting an interpretable prediction result; performing feature attribution analysis processing on the prediction result to generate a feature importance sequence; and performing interpretability verification on the prediction result based on the feature importance ranking and the rule path to obtain a corresponding interpretability evaluation conclusion. According to the method, the prediction performance and the interpretability of the prediction result can be considered at the same time in a high-dimensional data scene, and the credibility of the interpretable result is guaranteed.

Description

Technical Field

[0001] This application relates to the field of neural network technology, specifically to a data interpretable prediction method and apparatus based on a tree regular variational encoder. Background Technology

[0002] With the rapid development of information technology, high-dimensional data is increasingly widely used in various fields such as medicine, bioinformatics, and intelligent decision-making, becoming a core foundation for supporting accurate analysis and judgment. Especially in scenarios such as medical diagnosis and gene analysis, where the reliability, fairness, and security of results are extremely important, models not only need to have excellent predictive performance, but also require that the prediction process and results have clear interpretability so that relevant personnel can understand the model's decision-making logic and conduct compliance reviews and risk management.

[0003] While deep learning methods, such as variational autoencoders, can effectively handle high-dimensional and complex features and achieve high prediction accuracy, their underlying mechanisms are opaque, making it difficult to trace and explain the specific decision-making logic. Traditional interpretable models, such as decision trees, while providing clear rule paths, struggle to maintain structural simplicity and predictive performance in high-dimensional feature spaces. Therefore, existing technologies cannot simultaneously achieve high predictive performance and clear interpretability of prediction results in high-dimensional data scenarios, failing to meet the core requirements of model reliability in key domains.

[0004] The preceding description is intended to provide general background information and does not necessarily constitute prior art. Summary of the Invention

[0005] This application provides a data interpretable prediction method and apparatus based on a tree regular variational encoder, which can simultaneously achieve both prediction performance and interpretability of prediction results in high-dimensional data scenarios, and ensure the credibility of interpretable results.

[0006] In a first aspect, embodiments of this application provide a data interpretable prediction method based on a tree-regular variational encoder, including: The high-dimensional data of the target is encoded into the latent space by a variational autoencoder to obtain a latent variable representation; The latent variable representations are regularized and aligned based on the surrogate decision tree so that the distribution structure of the latent variable representations in the latent space is consistent with the rule path of the surrogate decision tree, thereby generating a regularized latent space. Classification prediction is performed based on the regularized latent space, and interpretable prediction results are output. The prediction results are subjected to feature attribution analysis to generate a ranking of feature importance. Based on the feature importance ranking and the rule path, the interpretability of the prediction results is verified, and the corresponding interpretability evaluation conclusion is obtained.

[0007] Furthermore, in some embodiments of this application, the step of encoding the target high-dimensional data into the latent space using a variational autoencoder to obtain a latent variable representation includes: The target high-dimensional data is input into the encoder of the variational autoencoder to obtain the mean and variance parameters of the latent variables; The latent variable representation is generated by sampling based on the mean and the variance parameter; The latent variable representation is input into the decoder of the variational autoencoder to reconstruct the output data, and the corresponding reconstruction loss is calculated based on the output data and the target high-dimensional data.

[0008] Furthermore, in some embodiments of this application, the step of performing regularization alignment processing on the latent variable representation based on the surrogate decision tree, so that the distribution structure of the latent variable representation in the latent space is consistent with the rule path of the surrogate decision tree, and generating a regularized latent space, includes: Based on the current latent variable representation, the agent decision tree is trained, wherein each leaf node of the agent decision tree corresponds to a rule path from the root node to that leaf node; For each input data sample, calculate the matching degree between the input data sample and the rule paths of each leaf node of the agent decision tree, and normalize it into a soft distribution label; By minimizing the difference between the output of the variational autoencoder and the soft distribution label, the distribution structure represented by the latent variables is aligned with the rule path, thereby generating a regularized latent space.

[0009] Furthermore, in some embodiments of this application, the step of performing rule-based alignment processing on the latent variable representation based on the surrogate decision tree further includes: When a new agent decision tree is used to replace an old agent decision tree, the parameters of the old agent decision tree are transferred to the new agent decision tree based on the path similarity between leaf nodes and the overlap of sample allocation. Before replacement, the new agent decision tree is subjected to a triple gating check based on the joint criteria of feature attribution consistency, leaf node assignment stability, and model fidelity. Replacement is only performed when all criteria are met.

[0010] Furthermore, in some embodiments of this application, the step of migrating the parameters of the old agent decision tree to the new agent decision tree includes: Calculate the path similarity and sample allocation overlap between each new leaf node of the new agent decision tree and each old leaf node of the old agent decision tree; For each new leaf node, based on the calculated path similarity and the sample allocation overlap, the old leaf node that matches the new leaf node is determined from the old leaf nodes; The alignment header parameters of the matched old leaf node are inherited to the new leaf node using a convex combination method. For new leaf nodes that do not match old leaf nodes, initialize them using default parameters.

[0011] Furthermore, in some embodiments of this application, the step of performing rule-based alignment processing on the latent variable representation based on the surrogate decision tree further includes: The average path length of the agent decision tree is calculated as a complexity measure. Based on the average path length and its corresponding model parameters, a proxy neural network is trained to establish a prediction mapping from model parameters to complexity. The complexity value predicted by the surrogate neural network for the current model parameters is added as a regularization term to the overall optimization objective.

[0012] Furthermore, in some embodiments of this application, the step of performing classification prediction based on the regularized latent space and outputting interpretable prediction results includes: The latent variable representations in the regularized latent space are input into a prediction function to obtain initial prediction values; Based on the latent variable representation, the corresponding rule path is retrieved from the leaf node activated in the agent decision tree; The initial predicted value, the activated rule path, and the key features related to the rule path are fused to form the interpretable prediction result.

[0013] Furthermore, in some embodiments of this application, the step of performing feature attribution analysis on the prediction results to generate a feature importance ranking includes: A strategy of weighting by feature subset size is adopted to allocate the sampling budget to generate multiple feature subsets; Estimate the contribution of each feature subset to the prediction result to obtain the corresponding contribution estimate; Aggregate the contribution estimates of all feature subsets, calculate and generate the feature importance ranking.

[0014] Furthermore, in some embodiments of this application, the step of verifying the interpretability of the prediction result based on the feature importance ranking and the rule path to obtain the corresponding interpretability evaluation conclusion includes: The feature importance ranking and the rule path of the proxy decision tree are transformed into structured explanatory data; The structured explanatory data is input into the large language model, which automatically generates a standardized verification questionnaire based on a preset prompt template. Collect feedback answers from domain experts on the validation questionnaire, and compare the feedback answers with the model's own explanatory logic to obtain the comparison results; Based on the comparison results, a quantitative interpretability assessment conclusion is calculated.

[0015] Secondly, embodiments of this application provide a data interpretable prediction apparatus based on a tree-regular variational encoder, comprising: The encoding module is used to encode the target high-dimensional data into the latent space through a variational autoencoder to obtain a latent variable representation. The alignment module is used to perform regular alignment processing on the latent variable representation based on the surrogate decision tree, so that the distribution structure of the latent variable representation in the latent space is consistent with the rule path of the surrogate decision tree, thereby generating a regularized latent space; The prediction module is used to perform classification prediction based on the regularized latent space and output interpretable prediction results; The analysis module is used to perform feature attribution analysis on the prediction results and generate a feature importance ranking. The verification module is used to verify the interpretability of the prediction results based on the feature importance ranking and the rule path, and obtain the corresponding interpretability evaluation conclusion.

[0016] This application provides a data-interpretable prediction method and apparatus based on a tree-regular variational encoder. First, a variational autoencoder encodes the target high-dimensional data into a latent space to obtain latent variable representations, compressing high-dimensional data redundancy and extracting core features, providing a high-quality data foundation for subsequent classification prediction and ensuring the model's predictive ability for high-dimensional data from the source. Second, a surrogate decision tree is used to regularize and align the latent variable representations, ensuring the latent space distribution structure aligns with the tree's rule paths and generating a regularized latent space. This avoids the ambiguity in prediction logic caused by disordered latent spaces, providing a structured framework for interpretability. Next, classification prediction is performed based on this latent space, and interpretable results are output by combining the rule paths, breaking the black-box limitation of traditional models that only know inputs and outputs but not the decision logic. Subsequently, feature attribution analysis generates a feature importance ranking, clarifying the impact of each feature on the prediction results and making the interpretation more targeted. Finally, the prediction results are verified based on feature importance and rule paths, with cross-verification using dual criteria to ensure the rationality of the interpretation. Therefore, the data-interpretable prediction scheme based on tree regular variational encoder provided in this application can not only guarantee the prediction performance of high-dimensional data, but also achieve clear and interpretable prediction results. It can simultaneously achieve excellent prediction performance and highly interpretable decision-making process in high-dimensional data scenarios, and ensure the credibility of interpretable results, providing stable and reliable intelligent decision support for high-risk fields. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application, 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 this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is an application environment diagram of the data interpretable prediction method based on tree regular variational encoder provided in the embodiments of this application; Figure 2 This is a flowchart illustrating the data interpretable prediction method based on a tree-regular variational encoder provided in an embodiment of this application. Figure 3 This is a schematic diagram of the sample distribution of the unexplained guidance mechanism provided in the embodiments of this application; Figure 4 This is a schematic diagram of sample distribution with an explanatory guidance mechanism provided in the embodiments of this application; Figure 5 This is a schematic diagram illustrating the sample attribution stability during the training phase provided in an embodiment of this application; Figure 6 This is a schematic diagram of the structure of a data interpretable prediction device based on a tree regular variational encoder provided in an embodiment of this application; Figure 7 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0019] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of systems and methods consistent with those detailed in the appended claims or with some aspects of this application.

[0020] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover descriptions such as non-exclusive inclusion, so that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element. Furthermore, components, features, and elements with the same names in different embodiments of this application may have the same meaning or different meanings, the specific meaning of which must be determined by its interpretation in that specific embodiment or further in conjunction with the context of that specific embodiment.

[0021] It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit this application.

[0022] In the following description, the use of suffixes such as "module," "part," or "unit" to denote elements is solely for the purpose of illustrative purposes and has no specific meaning in itself. Therefore, "module," "part," or "unit" may be used interchangeably.

[0023] To address the aforementioned technical problems and overcome the shortcomings of existing technologies, this application provides a data-interpretable prediction method and apparatus based on a tree-regular variational encoder, which can achieve dynamic pricing that adapts to changes in user behavior and market environment, thereby improving the accuracy and market adaptability of dynamic pricing in cross-border scenarios.

[0024] Figure 1 This is a diagram illustrating the application environment of a data-interpretable prediction method based on a tree-regular variational encoder in one embodiment. (Refer to...) Figure 1This tree-regular variational encoder-based data interpretable prediction method is applied to a tree-regular variational encoder-based data interpretable prediction system. The system includes a terminal 110 and a server 120. The terminal 110 and server 120 are connected via a network. The terminal 110 can be a desktop terminal or a mobile terminal; a mobile terminal can be at least one of a mobile phone, tablet, or laptop. The server 120 can be a standalone server or a server cluster consisting of multiple servers. Server 120 is configured to execute the above-mentioned data interpretable prediction method based on tree regularized variational encoder, including: encoding the target high-dimensional data into a latent space through a variational autoencoder to obtain latent variable representations; performing regularization alignment processing on the latent variable representations based on a surrogate decision tree to make the distribution structure of the latent variable representations in the latent space consistent with the rule path of the surrogate decision tree, generating a regularized latent space; performing classification prediction based on the regularized latent space and outputting interpretable prediction results; performing feature attribution analysis processing on the prediction results to generate a feature importance ranking; and performing interpretability verification on the prediction results based on the feature importance ranking and the rule path to obtain the corresponding interpretability evaluation conclusion.

[0025] Please see Figure 2 , Figure 2 This is a flowchart illustrating a data interpretable prediction method based on a tree-regular variational encoder according to an embodiment of this application. This embodiment primarily uses the application of this data interpretable prediction method based on a tree-regular variational encoder to a computer device as an example. Specifically, the data interpretable prediction method based on a tree-regular variational encoder provided in this embodiment may include the following steps: S1. The high-dimensional data of the target is encoded into the latent space using a variational autoencoder to obtain a latent variable representation; Specifically, for step S1, the target high-dimensional data refers to data with high dimensionality and a large number of features, commonly found in fields such as medicine and bioinformatics. Examples include liver cancer sample data containing thousands of gene expression values ​​and patient health monitoring data containing hundreds of clinical indicators. The core function of the variational autoencoder is to reduce the dimensionality and extract features from this type of high-dimensional data. Through its internal encoder network, redundant information in the high-dimensional data is removed, retaining the core features valuable for subsequent prediction. These core features are then mapped into a low-dimensional latent space, and the resulting low-dimensional feature vector is the latent variable representation. For example, when processing a liver cancer transcriptome sample containing 60,000 gene expression features, the variational autoencoder analyzes the association between these genes and liver cancer pathological features through its encoder network, filtering out gene information crucial for liver cancer diagnosis. This compresses the 60,000-dimensional gene data into a latent space of tens of dimensions. Each of these tens-of-dimensional latent variables accurately represents the core feature indicating whether the corresponding sample has a predisposition to liver cancer.

[0026] S2. Based on the surrogate decision tree, the latent variable representation is regularized and aligned so that the distribution structure of the latent variable representation in the latent space is consistent with the rule path of the surrogate decision tree, thus generating a regularized latent space; Specifically, for step S2, the surrogate decision tree is a model with explicit logical rules. Internally, it contains multiple "rule paths" from the root node to the leaf nodes. Each path corresponds to a decision logic, such as "if feature A > threshold X and feature B < threshold Y, then class P is determined." The rule alignment process adjusts the distribution of latent variables in the latent space, matching the aggregation pattern of latent variables with the rule paths of the surrogate decision tree. For a sample class corresponding to a certain rule path in the surrogate decision tree, latent variables possessing features of that class will aggregate in a specific region of the latent space. Latent variables corresponding to different rule paths are distributed in different regions of the latent space, ultimately forming a rule-aligned latent space where regions correspond one-to-one with rule paths. For example, if the surrogate decision tree has two rule paths for postpartum hemorrhage risk: "gestational age > 37 weeks and systolic blood pressure < 120 mmHg → low risk" and "gestational age < 35 weeks and systolic blood pressure > 140 mmHg → high risk", after alignment processing, two dedicated regions will be formed in the latent space: latent variables of samples that meet the criteria of "gestational age > 37 weeks and systolic blood pressure < 120 mmHg" will be clustered in region 1, and latent variables of samples that meet the criteria of "gestational age < 35 weeks and systolic blood pressure > 140 mmHg" will be clustered in region 2. The distribution structure of the latent variables perfectly matches the rule paths of the surrogate decision tree.

[0027] S3. Perform classification prediction based on the regularized latent space and output interpretable prediction results; Specifically, for step S3, the core advantage of the rule-based latent space is the binding of regions with rule paths. Therefore, during classification prediction, it is only necessary to determine which region of the latent space the latent variable of the sample to be predicted falls into to correspond to a certain rule path in the surrogate decision tree, thereby obtaining the classification result. At the same time, since the prediction result is directly associated with a clear rule path, the rule path information can be output together with the classification result to form an interpretable prediction result. That is, it not only tells the user what the predicted category is, but also tells the user what rule was used to arrive at the prediction. For example, if the latent variable of the postpartum hemorrhage sample to be predicted falls into region 2 (corresponding to the rule path "gestational age <35 weeks and systolic blood pressure >140mmHg → high risk"), the prediction result "this sample is at high risk of postpartum hemorrhage" will be output, along with the explanation "based on the rule 'gestational age <35 weeks and systolic blood pressure >140mmHg', the corresponding latent variable falls into the exclusive region of the high-risk rule," allowing the user to clearly understand the prediction logic.

[0028] S4. Perform feature attribution analysis on the prediction results to generate a ranking of feature importance; Specifically, for step S4, feature attribution analysis analyzes the contribution of each feature in the original high-dimensional target data to the current prediction result. This involves determining which features in the original data are key factors driving the prediction result and which have a smaller impact. By quantifying the contribution value of each feature and arranging them in descending order of contribution value, the feature importance ranking can be obtained, clarifying the correlation strength between each feature in the original high-dimensional data and the prediction result. For example, for the prediction result "a patient is at high risk of liver cancer," feature attribution analysis calculates the contribution value of each gene in the original gene data. If the contribution value of "alpha-fetoprotein (AFP) gene expression level" is the highest, followed by "liver function abnormality-related genes" and "tumor marker CA199 gene," then a feature importance ranking of "AFP gene expression level > liver function abnormality-related genes > CA199 gene > ..." is generated, intuitively presenting the key influencing features.

[0029] S5. Based on feature importance ranking and rule path, the interpretability of the prediction results is verified, and the corresponding interpretability evaluation conclusion is obtained; Specifically, for step S5, the core of interpretability verification is to check whether the explanatory logic of the prediction result is coherent and reasonable. On the one hand, referring to the feature importance ranking, it is confirmed whether the features involved in the surrogate decision tree rule path (such as "gestational age" and "systolic blood pressure" in the rule path) are in a high position in the feature importance ranking (i.e., they are indeed key features affecting the prediction). On the other hand, combining the rule path and feature importance, it is judged whether the generation logic of the prediction result is consistent and unbiased, and finally, an interpretability assessment conclusion is given based on the verification results. For example, if the prediction result is "high risk of postpartum hemorrhage", the rule path involves "gestational age" and "systolic blood pressure", and "gestational age" and "systolic blood pressure" are ranked first and second in the feature importance ranking, it indicates that the prediction logic is coherent, and the assessment conclusion is "the interpretation of the prediction result is reasonable and the interpretability meets the standard"; if "systolic blood pressure" in the rule path is ranked low in the feature importance ranking, it indicates that there is a deviation in the explanatory logic, and the assessment conclusion is "the matching between the rule path and feature contribution needs to be verified, and the interpretability needs to be optimized".

[0030] In processing high-dimensional data, this embodiment ensures the accuracy of classification prediction through variational autoencoders and regularized latent space, while making the prediction logic understandable and verifiable through rule paths, feature importance ranking, and interpretability verification, ultimately achieving a balance between prediction performance and interpretability in high-dimensional data scenarios.

[0031] Furthermore, in some embodiments, step S1, "encoding the target high-dimensional data into the latent space using a variational autoencoder to obtain a latent variable representation," may specifically include: S11. Input the target high-dimensional data into the encoder of the variational autoencoder to obtain the mean and variance parameters of the latent variables; Specifically, in step S11, the encoder in the variational autoencoder is a network structure with feature extraction and transformation capabilities. Its core function is to analyze the input high-dimensional data and map the data onto a probability distribution in a low-dimensional latent space. Specifically, the encoder outputs two key parameters—mean and variance—for each dimension of the latent space through multi-layer computation. The mean represents the central location of the data distribution in that dimension, and the variance represents the degree of dispersion of the data in that dimension; together, they describe the probability distribution characteristics of the latent variable.

[0032] S12. Sample based on mean and variance parameters to generate latent variable representations; Specifically, for step S12, since the distribution of latent variables is defined by the mean and variance (usually following a normal distribution), sampling involves randomly selecting a specific value from this normal distribution as the actual value of the corresponding dimension in the latent space. By sampling all dimensions of the latent space separately, the obtained dimensional values ​​are combined into a low-dimensional vector, which is the representation of the latent variable. To ensure that subsequent model training can backpropagate (i.e., the gradient is computable), the sampling process usually employs a reparameterization technique, transferring the randomness of sampling to the input noise, rather than directly randomly sampling the mean and variance, ensuring that the model can optimize parameters through gradient descent.

[0033] S13. Input the latent variable representation into the decoder of the variational autoencoder to reconstruct the output data, and calculate the corresponding reconstruction loss based on the output data and the target high-dimensional data; Specifically, in step S13, the decoder in the variational autoencoder complements the encoder, forming a network structure with data reconstruction capabilities. Its core function is to reverse-map the low-dimensional latent variable representation back to the dimensions of the original high-dimensional data, generating output data (i.e., reconstructed data) that is completely consistent with the original target high-dimensional data structure and dimensions. The reconstruction loss is an indicator used to measure the difference between the reconstructed data and the original target high-dimensional data. Common calculation methods include mean squared error (MSE) and cross-entropy. The smaller the difference, the lower the reconstruction loss value, indicating that the decoder has a better reconstruction effect on the latent variables, and indirectly proving that the latent variables extracted by the encoder accurately retain the core information of the original high-dimensional data.

[0034] This embodiment generates latent variables by sampling the probability distribution parameters of the latent variables output by the encoder, and the decoder reconstructs the data and calculates the loss. This ensures that the latent variables conform to the probability distribution characteristics, and verifies and optimizes the retention effect of the latent variables on the core information of the original high-dimensional data by reconstructing the loss. Finally, a high-quality latent variable representation is generated, which provides reliable data support for subsequent processing steps based on the latent space.

[0035] Furthermore, in some embodiments, step S2, "performing regularization alignment of the latent variable representations based on the surrogate decision tree so that the distribution structure of the latent variable representations in the latent space is consistent with the rule path of the surrogate decision tree, generating a regularized latent space," can specifically include a process of achieving regularization alignment through iterative optimization, specifically as follows: S21. Based on the current latent variable representation, train to obtain an agent decision tree, where each leaf node of the agent decision tree corresponds to a rule path from the root node to that leaf node; Specifically, in step S21, the current latent variable representation refers to the low-dimensional feature vector obtained by encoding the target high-dimensional data, such as gene expression data and clinical indicator data in the medical field, through a variational autoencoder. It contains the core information of the original high-dimensional data. The surrogate decision tree is a model with clear logical judgment rules, obtained by training these latent variable representations. During training, nodes are gradually split according to the feature differences of the latent variables. Starting from the root node, each time the latent variable feature that maximizes the distinction between sample categories and the corresponding threshold are selected as the splitting condition until the generated node meets the stopping condition such as the sample category purity reaching the standard or the lower limit of the number of samples in the node. The final terminal node is the leaf node. Each leaf node corresponds to a unique rule path from the root node to that leaf node. The path contains the judgment conditions of all split nodes, such as "latent variable A > threshold X → latent variable B < threshold Y → leaf node 1". This path is essentially the classification judgment rule of the sample corresponding to the leaf node.

[0036] S22. For each input data sample, calculate the degree of matching between the input data sample and the rule paths of each leaf node of the agent decision tree, and normalize them into soft distribution labels; Specifically, for step S22, the input data sample is the latent variable representation sample corresponding to the original high-dimensional data to be processed. The matching degree is used to measure whether the latent variable features of the sample meet all the judgment conditions of a certain leaf node rule path. If the sample completely meets all the conditions in the path, the matching degree is the highest (the violation degree is the lowest). If some conditions are not met, the specific violation degree value is calculated according to the "number of unmet conditions" and "degree of deviation from the threshold". The lower the violation degree, the higher the matching degree. Normalization is to convert the matching degree of all leaf nodes (or the matching score after violation degree conversion) into a probability distribution form through the Softmax function to obtain a soft distribution label. Each value in the label corresponds to the matching probability of the sample with a certain leaf node rule path. The sum of all values ​​is 1. The leaf node with the highest probability is the rule path that matches the sample best.

[0037] S23. By minimizing the difference between the output of the variational autoencoder and the soft distribution label, the distribution structure represented by the latent variables is aligned with the rule path, thereby generating a regularized latent space; Specifically, for step S23, the output of the variational autoencoder refers to the probability distribution corresponding to the number of leaf nodes in the surrogate decision tree, which is further output by the decoder after reconstructing the latent variable representation. The probability distribution reflects the degree of association between the latent variables and each rule path. The difference is usually measured by loss functions such as KL divergence and cross-entropy. These functions can quantify the similarity between two probability distributions. The smaller the difference, the closer the two distributions are. By optimizing the parameters of the variational autoencoder (such as the weights of the encoder and decoder) through backpropagation, the difference between its output and the soft distribution label can be gradually reduced. This process guides the latent variable representation to adjust its distribution in the latent space, allowing latent variables of samples matching the same leaf node rule path to cluster in the same region, and latent variables of samples matching different rule paths to form clear boundaries in the latent space, ultimately forming a regularized latent space where the latent variable distribution corresponds one-to-one with the rule path.

[0038] This embodiment extracts rule paths by training a proxy decision tree, then calculates soft-match labels between samples and paths, and minimizes the difference between the output of the variational autoencoder and the soft labels. This ensures that the distribution of latent variables in the latent space is precisely aligned with the rule paths, generating a well-structured rule-based latent space. This lays the core foundation for subsequent interpretable prediction based on the latent space.

[0039] Furthermore, in some embodiments, step S2, "performing regularization and alignment of latent variable representations based on surrogate decision trees," may further include, during the iterative optimization process, ensuring the coherence of the latent space structure through a stabilization mechanism when a new surrogate decision tree is introduced. The stabilization mechanism includes: S24. When replacing the old agent decision tree with a new agent decision tree, the parameters of the old agent decision tree are transferred to the new agent decision tree based on the path similarity between leaf nodes and the overlap of sample allocation. Specifically, in step S24, the replacement of the old agent decision tree with the new one typically occurs during model training. This is often due to situations where the old tree structure cannot adapt to new data features or the old tree rules lack sufficient accuracy, necessitating the introduction of a new tree with a superior structure. The core of this step is to achieve a smooth transfer of parameters between the old and new trees using two metrics: path similarity and sample allocation overlap, thus avoiding sudden changes in model performance caused by tree structure replacement.

[0040] Path similarity measures the similarity between a new leaf node in the new agent decision tree and an old leaf node in the old agent decision tree. A rule path consists of all splitting conditions from the root node to the leaf node, such as "latent variable A > threshold X" or "latent variable B < threshold Y". Path similarity is determined by statistically analyzing the proportion of conditions in the paths of the new and old leaf nodes that share the same splitting features and whose threshold differences are within a preset range. The higher the proportion of overlapping conditions, the higher the path similarity. Sample allocation overlap measures the degree of overlap between the samples contained in the new and old leaf nodes. It is calculated by dividing the number of samples commonly contained in the new and old leaf nodes by the total number of samples contained in the new and old leaf nodes (after deduplication). The higher the proportion of overlapping samples, the higher the sample allocation overlap. Based on the above two metrics, the most suitable old leaf node is matched for each new leaf node. Usually, the old leaf node with the highest weighted score of "path similarity + sample allocation overlap" is selected, and the core parameters of the old leaf node (such as the alignment header parameters used for latent space alignment) are transferred to the new leaf node. If there is no suitable old leaf node for the new leaf node (such as the score being lower than the preset threshold), the default parameters (such as the initialization parameters in the early stage of model training) are used for setting.

[0041] S25. Before replacement, a triple gating check is performed on the new agent decision tree based on the joint criteria of feature attribution consistency, leaf node assignment stability, and model fidelity. Replacement is only performed if all criteria are met.

[0042] Specifically, for step S25, the triple gating check is the core verification step before replacement. It is used to ensure the reliability and adaptability of the new agent decision tree and avoid unqualified new trees from causing model performance degradation or confusing interpretation logic. All three criteria must be met simultaneously (i.e., all reach the preset threshold) before the replacement operation is allowed.

[0043] In the feature attribution consistency check, feature attribution consistency refers to the similarity between the importance ranking of the key features (such as latent variables that determine node splits) that the new agent decision tree depends on when splitting and the importance ranking of the splitting features in the old tree. This is usually determined by calculating the correlation coefficient (such as the Pearson correlation coefficient) between the importance rankings of the two trees. If the correlation coefficient is greater than or equal to a preset threshold (such as 0.8), then the splitting logic of the new tree is considered consistent with that of the old tree, satisfying this criterion.

[0044] For leaf node assignment stability checks, leaf node assignment stability refers to the similarity between the new tree and the old tree in the leaf node assignment results for the same batch of samples, that is, the degree of overlap of the leaf node categories (such as "low risk" and "high risk") assigned to the samples. It is usually judged by calculating the standardized mutual information (NMI) of the leaf node assignment results of the two trees. If NMI ≥ a preset threshold (such as 0.7), the classification logic of the new tree for the samples is considered stable, and this criterion is met.

[0045] For model fidelity checks, model fidelity refers to the degree of agreement between the new tree's prediction results for samples and the true labels of the original data. It is usually judged by calculating the prediction error (such as mean squared error, classification error rate). If the prediction error of the new tree is less than or equal to the prediction error of the old tree (or the error decrease is greater than or equal to the preset value), then the prediction accuracy of the new tree is considered to be no less than that of the old tree, and this criterion is met.

[0046] This embodiment achieves smooth migration of parameters between the old and new agent decision trees by using leaf node similarity and overlap, and then uses triple gating checks to ensure the reliability of the new tree. This avoids model clustering drift caused by tree replacement and ensures the stability and continuity of the training process.

[0047] Furthermore, in some embodiments, step S24, "migrating the parameters of the old agent decision tree to the new agent decision tree," may specifically include: S241. Calculate the path similarity and sample assignment overlap between each new leaf node of the new agent decision tree and each old leaf node of the old agent decision tree; Specifically, in step S241, a new leaf node is the terminal node of the new agent decision tree. Each new leaf node corresponds to a rule path from the root node of the new tree to that node, such as "latent variable A > threshold X → latent variable B ≤ threshold Y". An old leaf node is the terminal node of the old agent decision tree, and also corresponds to a rule path of the old tree. This step needs to quantify the correlation between the new and old leaf nodes through two core indicators. The rule path consists of three core information categories: splitting features, branching directions, and thresholds. Path similarity is used to measure the similarity of the rule logic between the new and old leaf nodes. During calculation, first, the total number of splitting conditions in the old leaf node path is counted, such as "latent variable A > threshold X" and "latent variable B ≤ threshold Y" being two conditions. Then, the number of conditions in the new leaf node path that are "same as the old leaf node in terms of splitting features, have the same branching direction (such as both being ">" or both being "≤"), and have threshold differences within a preset range (such as ±0.2)" is counted. Finally, the path similarity is obtained by "the number of conditions that meet the criteria ÷ the total number of conditions in the old leaf node", with a value of 0-1. The closer to 1, the more similar the rule logic. Sample assignment overlap is used to measure the degree of overlap between samples contained in new and old leaf nodes. First, count the number of samples that are simultaneously assigned to both new and old leaf nodes (i.e., sample intersection). Then, count the sum of the number of samples in the new leaf node and the number of samples in the old leaf node, minus the number of samples in the intersection (i.e., sample union, to avoid double counting). Finally, the sample assignment overlap is obtained by "intersection sample count ÷ union sample count", with a value of 0-1. The closer it is to 1, the higher the sample overlap.

[0048] S242. For each new leaf node, based on the calculated path similarity and sample allocation overlap, determine the old leaf node that matches the new leaf node from the old leaf nodes. Specifically, for step S242, for each new leaf node, the most suitable matching object needs to be selected from all old leaf nodes by comprehensively considering two indicators: path similarity and sample allocation overlap. A weighted summation is typically used to calculate the comprehensive matching score, with weights set according to actual needs, such as a path similarity weight of 0.5 and a sample allocation overlap weight of 0.5. The formula is "Comprehensive Score = Path Similarity × 0.5 + Sample Allocation Overlap × 0.5". Then, the old leaf node with the highest comprehensive score that is not lower than a preset threshold is selected as the matching old leaf node for the new leaf node. If multiple old leaf nodes have the same highest comprehensive score, the old leaf node with higher sample allocation overlap can be prioritized to ensure a closer sample association.

[0049] S243. Inherit the alignment header parameters of the matched old leaf node to the new leaf node using a convex combination method; Specifically, for step S243, the alignment head parameters are the core parameters in the old leaf nodes used to align the latent variable representation with the regular path. These parameters include weight coefficients and bias terms used when calculating the latent variable and path matching degree, directly affecting the latent space regularization effect. Convex combination is a smooth parameter inheritance method that can avoid model fluctuations caused by parameter mutations. The formula is "New leaf node alignment head parameter = Path similarity × Old leaf node alignment head parameter + (1 - Path similarity) × Default base parameter", where the default base parameter is the model's preset initial alignment head value. The higher the path similarity, the closer the new node parameter is to the old node parameter, and the more thorough the inheritance. When the path similarity is low, the parameter will be adjusted in combination with the base value, balancing inheritance and adaptability.

[0050] S244. For new leaf nodes that do not match old leaf nodes, initialize them using default parameters; Specifically, for step S244, "no matching with old leaf nodes" means that the overall matching score of a new leaf node with all old leaf nodes is lower than the preset threshold. This indicates that the rule path and sample distribution of the new leaf node are too different from those of the old leaf nodes, and effective parameters cannot be obtained through inheritance. In this case, the alignment head parameters need to be initialized directly using the model's preset default parameters. The default parameters are usually verified general initial values ​​to ensure that the new leaf node has basic alignment capabilities to support subsequent model training and avoid training interruption or abnormal results due to missing parameters.

[0051] This embodiment achieves a smooth migration of head parameters between the old and new agent decision trees by accurately quantifying the correlation between new and old leaf nodes and reasonably matching nodes, smoothly inheriting parameters and standardizing the initialization of unmatched nodes. This ensures that the parameters of new nodes have a reliable basis and avoids parameter mutations, thus guaranteeing the continuity and stability of the model's latent space regularization effect.

[0052] Furthermore, in some embodiments, step S2, "performing regularization alignment of latent variable representations based on the surrogate decision tree," may further include: during the iterative optimization process, introducing regularization constraints based on the complexity of the surrogate decision tree to assist in achieving the regularization alignment, as follows: S26. Calculate the average path length of the agent decision tree as a measure of complexity; Specifically, in step S26, the surrogate decision tree is a tree-like model used to guide the regularization of the latent space. Its structural complexity directly affects the readability of the rules and the model's generalization ability. The longer the tree path and the more complex the branches, the more difficult the rules are to understand, and it may also lead to an increase in generalization error due to overfitting the data. The average path length (APL) is the core indicator for quantifying the structural complexity of the tree. First, the number of edges (path length, e.g., the path length of the root node directly connecting to a leaf node is 1, and the path length of the root node connecting to a leaf node through one intermediate node is 2) is counted in the surrogate decision tree. Then, the path lengths of all leaf nodes are summed, and finally, the average is divided by the total number of leaf nodes. The average is the average path length. The larger the average path length, the more branches the tree has and the more complex the structure; the smaller the value, the simpler the tree structure and the clearer the rules.

[0053] S27. Based on the average path length and its corresponding model parameters, train the surrogate neural network to establish a prediction mapping from model parameters to complexity; Specifically, for step S27, the model parameters refer to the core structural parameters of the surrogate decision tree, including the tree's splitting features (such as latent variables corresponding to "gestational age" and "systolic blood pressure" used to determine risk), the thresholds of each splitting node (such as "37" in "gestational age > 37 weeks"), the number of leaf nodes, etc. These parameters directly determine the tree's structure, thus affecting the average path length (complexity). The surrogate neural network is a small, lightweight network (such as a shallow multilayer perceptron MLP), and its core function is to learn the correlation between model parameters and average path length, that is, to establish a mapping between input model parameters and output predicted average path length.

[0054] The training process first requires collecting multiple sets of "model parameter-average path length" data pairs. By adjusting parameters such as the splitting features and thresholds of the surrogate decision tree, trees with different structures are generated. The average path length corresponding to each set of parameters is calculated to form a training dataset. Then, these data pairs are input into the surrogate neural network. The difference between the average path length predicted by the neural network and the actual calculated average path length (such as mean squared error) is used as the loss function to optimize the neural network parameters until the neural network can accurately predict the corresponding average path length based on the input model parameters (prediction error is lower than a preset threshold). At this point, the prediction mapping from model parameters to complexity is established.

[0055] S28. The complexity value predicted by the surrogate neural network for the current model parameters is added as a regularization term to the overall optimization objective; Specifically, for step S28, the overall optimization objective is the objective function that needs to be minimized during the joint training of the surrogate decision tree and the associated model (such as a variational autoencoder). Its core function is to ensure the prediction accuracy of the model (such as reducing classification error). The regularization term is essentially a constraint on the model's complexity. By adding the complexity value (average path length) of the surrogate neural network prediction as a regularization term, the overall optimization objective will simultaneously consider prediction accuracy and structural simplicity, avoiding the model from forming an overly complex tree structure in pursuit of high accuracy, which could lead to overfitting and difficult-to-understand rules.

[0056] In practice, a weight coefficient is set for the regularization term (which can be adjusted according to actual needs). The overall optimization objective can be expressed as "Overall optimization objective = original prediction error objective + λ × prediction complexity value". During training, the model needs to minimize both the prediction error and the prediction complexity value. If the prediction complexity value corresponding to a certain model parameter is too high (the tree structure is too complex), even if its prediction error is low, the overall optimization objective value will increase due to the contribution of the regularization term. The model will tend to select model parameters with lower prediction error and lower complexity value, ultimately achieving a balance between accuracy and simplicity.

[0057] This embodiment quantifies the complexity of the surrogate decision tree by using the average path length, establishes a prediction mapping between parameters and complexity using the surrogate neural network, and then uses the prediction complexity as a regularization term to constrain the optimization objective. This effectively controls the structural complexity of the surrogate decision tree, avoids overfitting, and makes the tree rules simpler and easier to understand, providing a clear structural foundation for subsequent latent space regularization and interpretable prediction.

[0058] Furthermore, in some embodiments, step S3, "performing classification prediction based on the regularized latent space and outputting interpretable prediction results," may specifically include: S31. Input the latent variable representations in the regularized latent space into a prediction function to obtain the initial prediction values; Specifically, in step S31, the latent variable representation in the regularized latent space is a low-dimensional feature vector aligned with the rule path of the surrogate decision tree. It retains the core information of the original high-dimensional data (such as tumor gene data and clinical indicator data in medical scenarios, and transcriptome data in bioinformatics scenarios), and its distribution structure corresponds one-to-one with the rule path, providing a structured data foundation for accurate prediction. The prediction function is a computational model used for classification prediction (such as multilayer perceptron, logistic regression, support vector machine, etc.). Its core function is to output a quantitative result (such as probability value, category score) reflecting the category to which the sample belongs based on the feature information represented by the latent variables. This result is the initial prediction value, which only reflects the classification conclusion and does not yet include explanatory logic.

[0059] S32. Based on the latent variable representation of the activated leaf node in the agent decision tree, retrieve the corresponding rule path; Specifically, in step S32, each leaf node of the surrogate decision tree is bound to a unique rule path. The path consists of all splitting conditions from the root node to the leaf node, such as "latent variable A > threshold X → latent variable B < threshold Y". Each path corresponds to the judgment logic for a class of samples. Activating a leaf node means that the features represented by the latent variable completely meet all the splitting conditions of a certain leaf node, and the model determines that the node belongs to that leaf node. Retrieving the rule path means extracting the complete sequence of splitting conditions corresponding to that leaf node, clarifying the specific judgment rules satisfied by the latent variable.

[0060] S33. The initial predicted value, the activated rule path, and the key features related to the rule path are fused to form an interpretable prediction result; Specifically, for step S33, the key features related to the rule path refer to the original high-dimensional data features corresponding to the latent variables involved in the rule path. For example, latent variable 1 corresponds to "alpha-fetoprotein concentration" in the original data, and latent variable 2 corresponds to "EGFR gene expression value, whose original feature name and actual measurement value need to be clearly defined." This establishes the association between the latent variable rules and the original data. The categorical meaning of the initial predicted value, the judgment logic of the rule path, and the specific information of the key features are integrated into an intuitive and easy-to-understand expression, allowing users to not only know the prediction conclusion but also clearly trace the basis for its generation.

[0061] This embodiment not only relies on the rule-based latent space and prediction function to ensure the accuracy of classification prediction, but also integrates the initial predicted value, rule path and key features to make the prediction results accompanied by clear judgment logic and original data basis. This effectively breaks the black box limitation of traditional models that "only give results without explaining the basis", and realizes the understanding and traceability of prediction results.

[0062] Furthermore, in some embodiments, step S4, "performing feature attribution analysis on the prediction results to generate a feature importance ranking," may specifically include: S41. A sampling budget is allocated to generate multiple feature subsets by using a strategy that weights the sampling budget by the size of the feature subset; Specifically, for step S41, a feature subset refers to a set of features selected from all features of the original high-dimensional data. For example, selecting 3 clinical indicators from 50 indicators to form a subset. The sampling budget is the total number of samples used to generate the feature subset (i.e., the total number of subsets to be generated). Different sampling numbers are allocated according to the number of features contained in the feature subset (i.e., size, such as 1 feature, 2 features, ... up to the total number of features - 1 feature). Very small subsets (e.g., containing only 1-2 features) and very large subsets (e.g., containing the total number of features - 2 to the total number of features - 1 feature) that are easily ignored in high-dimensional data are given higher weights and allocated more sampling slots. Medium-sized subsets are given lower weights and allocated fewer sampling slots. This ensures that subsets of all sizes are adequately covered and avoids analytical bias caused by sampling bias.

[0063] S42. Estimate the contribution of each feature subset to the prediction result and obtain the corresponding contribution estimate; Specifically, for step S42, contribution refers to the degree of influence of a certain feature subset on the model's prediction results. If the model's prediction results are more accurate when the subset is included, or if the subset can significantly distinguish different categories of samples, then its contribution is higher. To compare the model's prediction differences when the feature subset is included and when it is not, first input the feature subset into the model to obtain the prediction result including the subset (e.g., "high risk of postpartum hemorrhage, probability 0.92"); then remove the feature subset using random replacement or masking (e.g., replace the feature values ​​within the subset with the mean) to obtain the prediction result without the subset (e.g., "high risk of postpartum hemorrhage, probability 0.65"); finally, calculate the difference between the two prediction results (e.g., probability difference 0.27). This difference is the estimated contribution of the feature subset; the larger the difference, the higher the contribution.

[0064] S43. Aggregate the contribution estimates of all feature subsets, calculate and generate the feature importance ranking; Specifically, for step S43, aggregation refers to summarizing the contributions of all feature subsets, extracting the total contribution capability of each individual feature, and for each original feature, statistically estimating the contribution of all feature subsets containing that feature, and calculating the total contribution score of that feature by weighted average. The weights can be consistent with the weights of the subset size to ensure that the contributions of extremely small / extremely large subsets are fully considered. Subsequently, all original features are arranged in descending order of total contribution score to form a feature importance ranking. The higher the score of a feature, the more critical its impact on the prediction result.

[0065] This embodiment covers various feature subsets by using a scale-weighted sampling strategy, accurately estimates the contribution of subsets and aggregates them to obtain the total contribution of a single feature. It can efficiently generate accurate feature importance rankings in high-dimensional data scenarios, clearly locate key features that affect the prediction results, and provide core evidence for the interpretability of the prediction results.

[0066] Furthermore, in some embodiments, step S5”” performs interpretability verification on the prediction results based on feature importance ranking and rule paths, and obtains the corresponding interpretability evaluation conclusion, which may specifically include: S51. Transform the rule paths of feature importance ranking and surrogate decision tree into structured explanatory data; Specifically, for step S51, the feature importance ranking is an ordered list reflecting the degree of influence of each feature in the original high-dimensional data on the prediction result, such as "gestational age > systolic blood pressure > platelet count > prothrombin time". The rule path of the surrogate decision tree is a judgment logic composed of splitting conditions and corresponding conclusions, such as "gestational age < 35 weeks and systolic blood pressure > 140 mmHg → high risk of postpartum hemorrhage". The above two types of information are organized into a data format with a unified format and clear logic, such as key-value pairs, tables, and structured text, to eliminate the fragmentation and ambiguity of information and ensure that the subsequent large language model can accurately identify and generate verification content based on this. During the transformation, the core elements of the two types of information need to be clearly defined: for feature importance ranking, the feature name, importance score, and ranking position need to be labeled; for rule paths, the splitting conditions (including feature name, threshold, and branch direction) and corresponding prediction conclusions need to be labeled.

[0067] S52. Input the structured explanatory data into the large language model, which will automatically generate a standardized verification questionnaire based on the preset prompt template; Specifically, for step S52, the large language model is an intelligent model with text understanding and generation capabilities, capable of generating text content that meets the requirements based on input information and guidance instructions. The preset prompt template is an instruction framework used to constrain the direction, format, and content scope of questionnaire generation. It needs to clearly define the purpose of the questionnaire (verifying the rationality of the explanatory logic), question types (such as multiple choice, true / false, and short answer questions), number of questions, and key points of assessment (such as the consistency between rule paths and feature importance, and the clinical rationality of conclusions). After inputting the structured explanatory data and the prompt template into the large language model, the model will automatically extract key information and generate a standardized questionnaire. Standardization means that the proportion of question types, score settings, and expression styles of all questionnaires are consistent, avoiding the impact of differences in questionnaire format on the objectivity of expert feedback.

[0068] S53. Collect feedback answers from domain experts on the validation questionnaire, and compare the feedback answers with the model's own explanatory logic to obtain the comparison results; Specifically, for step S53, domain experts refer to personnel with professional knowledge in the domain to which the prediction task belongs, such as obstetricians in postpartum hemorrhage prediction scenarios and oncologists in tumor prediction scenarios. Their feedback answers can reflect the domain rationality of the explanatory logic. The explanatory logic of the model itself is the core logic contained in the structured explanation data, such as the rule path corresponding to a specific conclusion and the features involved in the rule being ranked high in importance. The comparison needs to be carried out question by question. First, determine the standard answer of the model's explanatory logic for each question. For example, the standard answer for question 1 is B, corresponding to the conclusion of the rule path; the standard answer for question 2 is A, corresponding to the gestational age and systolic blood pressure involved in the rule being ranked 1st and 2nd respectively. Then, count the consistency between the expert feedback answers and the standard answers, including the single-question consistency rate (number of experts who answered the question correctly ÷ total number of experts who participated in the feedback) and the overall consistency rate (number of experts who answered all questions correctly ÷ total number of experts who participated in the feedback). These statistical results are the comparison results.

[0069] S54. Based on the comparison results, a quantitative interpretability assessment conclusion is calculated; Specifically, for step S54, the quantitative interpretability assessment conclusion refers to converting the qualitative comparison results into specific numerical values ​​(such as scores, percentages) or clear grades (such as "excellent," "qualified," and "unqualified") to intuitively reflect the reasonableness of the interpretation of the prediction results. Quantitative rules need to be set during the calculation. For example, the overall consistency rate is the core indicator, and the individual question consistency rate is the auxiliary indicator. If the overall consistency rate is ≥80% and the individual question consistency rate is ≥70%, the assessment score is 90-100 points, and the interpretability is excellent; if the overall consistency rate is 60%-79% and the individual question consistency rate is ≥60%, the assessment score is 70-89 points, and the interpretability is qualified; if the overall consistency rate is <60% or the consistency rate of any individual question is <60%, the assessment score is <70 points, and the interpretability is unqualified.

[0070] This embodiment, through structured data transformation, large-scale model-generated standardized questionnaires, expert feedback comparison, and quantitative calculation, can objectively evaluate the rationality of the prediction results interpretation, form actionable and interpretable conclusions, ensure that the model interpretation conforms to domain professional understanding, and improve the credibility of the prediction results.

[0071] To facilitate understanding of the data interpretable prediction method based on a tree-regularized variational encoder provided in this embodiment, this embodiment will provide a detailed explanation in conjunction with a data interpretable prediction system based on a tree-regularized variational encoder. The core idea of ​​this data interpretable prediction system based on a tree-regularized variational encoder is to combine the representation learning capability of a VAE with the interpretable structure of a decision tree, supplemented by a feature attribution mechanism based on SHAP. The system consists of four parts: a VAE encoder-decoder structure, a tree guidance module, a feature attribution module, and a tree regularization module. The VAE encoder-decoder structure is used to learn the low-dimensional latent space representation of high-dimensional data; the tree guidance module is used to introduce an interpretable rule partitioning structure into the latent space; the feature attribution module is used to evaluate the contribution of each dimension of the input features to the distribution of latent variables using SHAP; and the tree regularization module is used to control model complexity through average path length (APL) constraints.

[0072] For VAE modules: The training set is denoted as ,in VAE uses the standard ELBO as the reconstruction and prior constraint objective: This creates a compact potential space for subsequent interpretable alignment.

[0073] For explainable bootstrapping mechanisms: A surrogate decision tree T is introduced on top of the latent representation of the VAE as structural supervision. Each leaf node of the tree corresponds to a decision path rule from the root to the leaf; the model assigns a soft distribution about all leaves to each sample and forces the network output to align with this soft distribution, so that samples with similar path rules naturally cluster in the representation space.

[0074] For any sample Record his relationship with Ye The path violation degree is: in The first The characteristics, threshold, and left and right branch directions of each split. The soft label p(x) for the leaf cluster is obtained by softmax normalization, and the network output is... Align it with KL divergence: . This loss aligns the latent representation with tree-guided rule partitioning, improving the interpretability and separability of the representation.

[0075] Training proceeds iteratively in a "tree-network" manner: in the t-th iteration, the proxy tree is distilled from the current network. Its leaves form soft clusters of tags. The tree structure is replaced with... First, the cluster head is reset according to the number of new leaves. Then, the old leaves are matched with the new leaves based on the path similarity and the IoU of the sample assignment. The alignment head parameters are inherited in a convex combination manner. Unmatched ones are initialized with default values, so as to maintain the consistency of cluster identity and optimization stability under small structural changes.

[0076] To avoid cluster drift caused by frequent replacements, candidate trees We need to first minimize the fidelity target after the penalty: , Subsequently, a triple-gating admission process is implemented: first, feature attribution consistency (the correlation between the split importance of the tree and the attribution score reaches a threshold); second, assignment stability (the leaf assignment NMI of two consecutive trees reaches a threshold); and third, fidelity reduction (J strictly decreases). This strategy is lenient in the early stages of training and strict in the later stages, balancing adaptability and stability.

[0077] For high-dimensional SHAP sampling and feature subset selection, directly applying KernelSHAP in high dimensions often results in high overhead and high variance. This solution proposes a sampling strategy weighted by subset size, first assigning weights to all sizes |U|: , Where d represents the dimension of the data. The total sampling budget is then allocated according to the normalized weights. To each scale, Based on this, sample weights are assigned to each sampling instance. This improves the coverage and estimation stability of minimal / maximal subsets while maintaining Shapley consistency.

[0078] It should be noted that this embodiment does not simply use SHAP for post-hoc interpretation, but rather embeds SHAP attribution as a feature selection and dimensionality reduction step into the entire training process, and uses this to train the agent decision tree. The specific process is as follows: In the original high-dimensional feature space, SHAP (an implementation based on Kernel SHAP) is used for each sample to estimate the contribution of each feature to the model prediction, thus obtaining the feature attribution vector. , where the i-th dimension represents the marginal contribution of the i-th feature to the current prediction.

[0079] During training, the feature attributions obtained in each round are smoothly aggregated using exponential moving average (EMA) to form a stable global feature importance score vector: By reducing noise and variance from a single sampling, a more robust feature importance ranking can be obtained.

[0080] All features are sorted according to the global score s, and the top k important features are selected to form the feature subset for the current round. The feature subset size is set to: , in d For the original feature dimension, The proportionality coefficient (typically taken as...) That is, at most about 25% of the original features are retained for explicit feature selection and dimensionality reduction steps.

[0081] Subsequently only in the feature subset The agent decision tree is trained on the above, and each split of the decision tree only allows the use of The features in the model are fitted to and approximate the decision boundary of the original model in a significantly compressed feature subspace. By discarding redundant or invalid features with low contribution, the search space can be reduced while maintaining predictive performance, thus alleviating the overfitting and structural expansion problems caused by directly building trees in high-dimensional cases.

[0082] It is important to note that the surrogate decision tree in this embodiment does not directly use the real labels as supervision signals. Instead, it uses the current neural network's predicted output on the samples as the distillation target to construct the surrogate dataset for training: for each sample... Only its projection on the feature subset is retained. As input, and the original model in Continuous prediction on As the regression target, thus obtaining sample pairs The proxy tree is obtained by minimizing the difference between the proxy tree and the proxy tree on this proxy dataset. The prediction error is used for training, so that in the feature subset The output of the tree model is designed to closely approximate the decision-making behavior of the original model while maintaining the interpretability of the rule structure.

[0083] For tree regularization and complexity control, this embodiment adopts a tree regularization and complexity control strategy proposed in existing research during model training. This method is used to constrain the structural depth and balance of the surrogate decision tree to prevent model overfitting and rule redundancy, and to control the complexity and balance of the tree. Specifically, during the training phase, the average path length is introduced as a metric for tree structural complexity and used as a regularization term. This is incorporated into the overall optimization objective. Since tree growth is non-differentiable, a small MLP is trained as the APL agent. Based on the most recent parameter—APL observations Fitting a 2-norm regression: The agent only uses samples from the most recent 50 epochs to fit the current model state.

[0084] This constraint mechanism is derived from mature methods in publicly available literature. In this invention, it is only invoked as part of the optimization process and does not involve new algorithm principles or structural improvements.

[0085] The final optimization objective for the overall loss is: For the end-to-end training process: First, the VAE is trained with all features to obtain the initial latent representation; then, improved SHAP sampling is used to estimate feature attributions on the current model, and a subset is selected after EMA smoothing. .exist Distillation of proxy trees and generation of leaf-level soft distributions The network output is matched with the target tree using a KL alignment term, while APL metrics are collected to update the APL agent and inject tree complexity regularization. Whenever a candidate tree is distilled, a target score of fidelity + APL is first applied, followed by a three-gating system of "attribution consistency—assignment stability—fidelity reduction" to determine whether to replace it. If replaced, the cluster head parameters are transferred based on path similarity and IoU. This process is repeated until the joint loss converges and a model with path rule explanations is obtained.

[0086] Specifically, the steps of the interpretable method based on large models are as follows: In the interpretability assessment phase, this technical solution designs a clinical interpretability verification mechanism based on a locally deployed Large Language Model (LLM) for the diagnosis of PPH (postpartum hemorrhage). The core objective of this mechanism is to transform the model output into a form that is more in line with the reading and understanding habits of clinical experts, thereby achieving clinical verification and trust assessment of the model's inference results.

[0087] This system employs an open-source large language model and is deployed in a local secure environment to ensure data privacy and controllability. The input to the large language model is strictly limited to the structured output content generated by the model, including SHAP-based feature importance ranking and symbolic logical descriptions of the surrogate decision tree. In this way, the model avoids the risk of fabrication inherent in free generation, ensuring that the LLM's output is completely consistent with the interpretable results of the TRI-VAE model itself. In other words, the LLM acts as a readability enhancer, rather than an independent reasoning entity, thus guaranteeing the objectivity and consistency of the output content.

[0088] In the actual evaluation process, the system generated 20 model-derived questions based on the structured output described above. These questions included multiple-choice and judgment-based clinical reasoning questions. All questions were generated by the LLM according to a standard prompt template. Subsequently, these questions and their answer choices were submitted to nine clinicians from a local hospital (including seven attending physicians, one associate chief physician, and one chief physician) to assess the clinical rationality of the model's reasoning logic. The physicians completed the questionnaire based solely on the explanatory output provided by the model, without directly accessing its internal mechanisms, thus forming a "black box explanation verification" model.

[0089] During the scoring process, doctors' answers were compared with the model's output, and different weights were assigned according to the doctor's level (attending physician weight 1, associate chief physician weight 2, chief physician weight 3). The final weighted average interpretability score was 86.9 / 100. This result demonstrates that the interpretable content generated by the TRI-VAE model not only statistically aligns highly with expert judgment but also possesses high comprehensibility and verifiability. The controlled use of LLM ensures the stability and intra-domain consistency of the generated interpretations, making the model's interpretability verification process both scientific and compliant with clinical application standards. This process design fully embodies the organic integration of "algorithm interpretability – model credibility – clinical validation," providing a reference path for the standardized evaluation of interpretability of subsequent medical AI models.

[0090] Furthermore, this embodiment also provides experimental verification and performance of a data-interpretable prediction method based on a tree-regular variational encoder.

[0091] (1) Experimental setup To verify the performance and interpretability of TRI-VAE (Tree-Regularized Interpretable Variational Autoencoder) on high-dimensional data, this study selected three representative datasets: TCGA-LIHC: Contains 424 liver cancer transcriptome samples with a total of 60,661 features; TUEP: A subset derived from the TUH EEG database, containing 186,200 samples and 798 features after preprocessing; PPH: A private obstetric clinical dataset containing 5,935 samples and 48 clinical variables.

[0092] (2) Validation of the effectiveness of the interpretability guidance mechanism The structured properties of TRI-VAE in the latent space were verified using t-SNE dimensionality reduction visualization. For example... Figure 3As shown, the standard VAE without a tree-guided mechanism exhibits a relatively diffuse sample distribution and significant overlap between different decision paths; however, with the addition of explanatory guidance, such as Figure 4 As shown, the samples form tight clusters according to the rule paths, with clear boundaries between different paths. This indicates that the latent space of TRI-VAE is highly aligned with the rule partitioning of the decision tree, significantly improving the interpretability of the representation.

[0093] In addition, such as Figure 5 As shown, the surrogate decision tree on the PPH dataset achieves a 96.1% fit on the validation set, indicating that the extracted decision paths highly reflect the true behavior of the VAE classifier. Further NMI metrics show that the tree structure fluctuates significantly within the first 50 epochs of training, but after enabling the controlled update mechanism, the cluster allocation of the model gradually stabilizes, ensuring cluster consistency and training controllability.

[0094] (3) Evaluation of the effect of the tree regularization module To evaluate the effect of tree regularization, models with and without tree path length constraints were compared. The results show that regularization not only effectively reduces the depth and number of nodes in the decision tree, making the tree structure more compact and balanced, but also achieves performance improvements on all datasets: AUC increases of 0.059 (PPH), 0.011 (TCGA-LIHC), and 0.011 (TUEP), respectively. This indicates that controlling path complexity contributes to model generalization and structural readability.

[0095] (4) Efficiency analysis of SHAP sampling strategy To address the high cost of SHAP estimation in high-dimensional feature spaces, TRI-VAE proposes a sampling strategy based on subset size weighting. Experimental results show that, with sampling sizes ranging from 10 to 5000, the model AUC improves from 0.978 to 0.996, and the interpretability consistency increases from 0.882 to 0.976; while the running time increases only linearly. This result demonstrates that the proposed sampling method achieves a good balance between accuracy and efficiency, exhibiting significant advantages in high-dimensional data interpretation tasks.

[0096] (5) Overall performance In multi-dataset validation, the TRI-VAE model demonstrated excellent overall performance in prediction accuracy, precision, recall, F1 score, and AUC.

[0097] On the TCGA-LIHC liver cancer gene expression dataset, TRI-VAE achieved Accuracy = 0.987, Precision = 0.978, Recall = 0.958, F1 Score = 0.965, and AUC = 0.996. These results demonstrate that the model possesses extremely high discriminative ability and stability in a high-dimensional gene feature environment, maintaining excellent classification performance while preserving generalization ability.

[0098] On the TUEP EEG dataset, the model also demonstrated robust performance across the board, with Accuracy = 0.906, Precision = 0.944, Recall = 0.940, F1 Score = 0.943, and AUC = 0.936. Even in scenarios with high signal-to-noise ratios, TRI-VAE maintained high consistency and recognition capabilities, indicating that its latent space structure possesses good robustness.

[0099] On the PPH postpartum hemorrhage clinical dataset, the model demonstrated highly reliable prediction results, specifically Accuracy = 0.961, Precision = 0.967, Recall = 0.961, F1 Score = 0.963, and AUC = 0.912. These results validate the superior performance of TRI-VAE in clinical multivariate data analysis, particularly its balance between precision and recall, reflecting the model's high sensitivity to key features and its stable interpretability.

[0100] In summary, TRI-VAE achieved stable performance with high accuracy (Accuracy ≥ 0.90) and high AUC (≥ 0.91) on datasets from three different domains, fully demonstrating the model's generalization ability and reliability in multimodal data.

[0101] In summary, compared with existing technologies, the data interpretable prediction method based on a tree-regular variational encoder provided in this embodiment introduces a surrogate decision tree structure into the latent space of the variational autoencoder (VAE). By applying soft distribution alignment constraints to the leaf node rule paths, it ensures that the model output is consistent with the tree partitioning rules, achieving regularized clustering of samples in the latent space. This mechanism simultaneously optimizes the reconstruction objective and the interpretation alignment objective during the training phase, significantly improving the structural interpretability and rule consistency of the latent space. Addressing the issues of large variance and high computational cost in SHAP estimation under high-dimensional feature spaces, a subset-size-weighted sampling strategy is proposed. This method improves the coverage and estimation stability of minimal and maximal feature subsets while maintaining Shapley consistency, achieving efficient high-dimensional feature attribution. To prevent clustering drift due to frequent tree replacements during training, a triple gating criterion is proposed, using feature attribution consistency, leaf node assignment stability, and model fidelity degradation as joint criteria to dynamically control the candidate tree replacement process. This mechanism ensures the coherence of the model's clustering structure and the stability of training convergence. For high-credibility scenarios such as healthcare, a clinical interpretation validation mechanism combining a locally deployed Large Language Model (LLM) is proposed. By inputting the structured interpretation results generated by the TRI-VAE model into the Large Language Model, a standardized expert validation questionnaire is automatically generated in a controlled environment to evaluate the clinical rationality of the model's interpretation.

[0102] To facilitate better implementation of the data interpretability prediction method based on a tree-regular variational encoder according to the embodiments of this application, this application also provides a data interpretability prediction device based on a tree-regular variational encoder, which is based on the aforementioned data interpretability prediction method based on a tree-regular variational encoder. The meanings of the terms used are the same as in the aforementioned data interpretability prediction method based on a tree-regular variational encoder, and specific implementation details can be found in the descriptions in the method embodiments.

[0103] Please see Figure 6 , Figure 6 The schematic diagram below illustrates the structure of a data interpretable prediction device based on a tree-regular variational encoder, as provided in this embodiment of the application. This data interpretable prediction device specifically includes an encoding module 201, an alignment module 202, a prediction module 203, an analysis module 204, and a verification module 205, as follows: Encoding module 201 is used to encode the target high-dimensional data into the latent space through a variational autoencoder to obtain a latent variable representation; Alignment module 202 is used to perform regularization alignment processing on the latent variable representation based on the surrogate decision tree, so that the distribution structure of the latent variable representation in the latent space is consistent with the rule path of the surrogate decision tree, thereby generating a regularized latent space; The prediction module 203 is used to perform classification prediction based on the regularized latent space and output interpretable prediction results; Analysis module 204 is used to perform feature attribution analysis on the prediction results and generate a ranking of feature importance. The verification module 205 is used to verify the interpretability of the prediction results based on feature importance ranking and rule path, and obtain the corresponding interpretability evaluation conclusion.

[0104] For specific limitations regarding the data interpretable prediction device based on tree regular variational encoders, please refer to the limitations of the data interpretable prediction method based on tree regular variational encoders mentioned above, which will not be repeated here. Each module in the aforementioned data interpretable prediction device based on tree regular variational encoders can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in hardware or independently of the processor in a computer device, or stored in software in the memory of a computer device, so that the processor can call and execute the corresponding operations of each module.

[0105] The data interpretable prediction device based on tree regular variational encoder provided in this embodiment can not only ensure the prediction performance of high-dimensional data, but also achieve clear and interpretable prediction results. It can simultaneously achieve excellent prediction performance and highly interpretable decision-making process in high-dimensional data scenarios, and ensure the credibility of interpretable results, providing stable and reliable intelligent decision support for high-risk fields.

[0106] Furthermore, embodiments of this application also provide an electronic device, such as... Figure 7 As shown, it illustrates a structural schematic diagram of the electronic device involved in the embodiments of this application, specifically: The electronic device may include components such as a processor 301 with one or more processing cores, a memory 302 with one or more computer-readable storage media, a power supply 303, and an input unit 304. Those skilled in the art will understand that... Figure 7 The electronic device structure shown does not constitute a limitation on the electronic device and may include more or fewer components than shown, or combine certain components, or have different component arrangements. Wherein: The processor 301 is the control center of the electronic device. It connects various parts of the electronic device via various interfaces and lines, and performs various functions and processes data by running or executing software programs and / or modules stored in the memory 302, and by calling data stored in the memory 302, thereby providing overall monitoring of the electronic device. Optionally, the processor 301 may include one or more processing cores; preferably, the processor 301 may integrate an application processor and a modem processor, wherein the application processor mainly handles the operating system, user interface, and applications, and the modem processor mainly handles wireless communication. It is understood that the modem processor may not be integrated into the processor 301.

[0107] Memory 302 can be used to store software programs and modules. Processor 301 executes various functional applications and data-interpretable prediction methods based on tree regular variational encoders by running the software programs and modules stored in memory 302. Memory 302 may mainly include a program storage area and a data storage area. The program storage area may store the operating system, application programs required for at least one function (such as sound playback, image playback, etc.), etc.; the data storage area may store data created based on the use of the electronic device. In addition, memory 302 may include high-speed random access memory and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other volatile solid-state storage device. Accordingly, memory 302 may also include a memory controller to provide processor 301 with access to memory 302.

[0108] The electronic device also includes a power supply 303 that supplies power to various components. Preferably, the power supply 303 can be logically connected to the processor 301 through a power management system, thereby enabling functions such as charging, discharging, and power consumption management through the power management system. The power supply 303 may also include one or more DC or AC power supplies, recharging systems, power fault detection circuits, power converters or inverters, power status indicators, and other arbitrary components.

[0109] The electronic device may also include an input unit 304, which can be used to receive input digital or character information and generate keyboard, mouse, joystick, optical or trackball signal inputs related to user settings and function control.

[0110] Although not shown, the electronic device may also include a display unit, etc., which will not be described in detail here. Specifically, in this embodiment, the processor 301 in the electronic device loads the executable files corresponding to the processes of one or more applications into the memory 302 according to the following instructions, and the processor 301 runs the applications stored in the memory 302 to realize various functions, as follows: The target high-dimensional data is encoded into a latent space using a variational autoencoder to obtain latent variable representations. These latent variable representations are then regularized and aligned using a surrogate decision tree to ensure their distribution structure within the latent space aligns with the rule paths of the surrogate decision tree, generating a regularized latent space. Classification prediction is then performed based on this regularized latent space, and interpretable prediction results are output. Feature attribution analysis is then performed on the prediction results to generate a feature importance ranking. Finally, based on the feature importance ranking and the rule paths, the interpretability of the prediction results is verified, yielding corresponding interpretability evaluation conclusions.

[0111] For details on the implementation of each of the above operations, please refer to the previous examples, which will not be repeated here.

[0112] The embodiments of this application can both guarantee the predictive performance of high-dimensional data and achieve clear and interpretable prediction results. In high-dimensional data scenarios, it can simultaneously achieve excellent predictive performance and a highly interpretable decision-making process, and ensure the credibility of interpretable results, providing stable and reliable intelligent decision support for high-risk fields.

[0113] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be performed by instructions, or by instructions controlling related hardware. These instructions can be stored in a computer-readable storage medium and loaded and executed by a processor.

[0114] To this end, embodiments of this application provide a storage medium storing a plurality of instructions that can be loaded by a processor to execute steps in any of the data interpretable prediction methods based on tree regularized variational encoders provided in embodiments of this application. For example, the instructions can execute the following steps: The target high-dimensional data is encoded into a latent space using a variational autoencoder to obtain latent variable representations. These latent variable representations are then regularized and aligned using a surrogate decision tree to ensure their distribution structure within the latent space aligns with the rule paths of the surrogate decision tree, generating a regularized latent space. Classification prediction is then performed based on this regularized latent space, and interpretable prediction results are output. Feature attribution analysis is then performed on the prediction results to generate a feature importance ranking. Finally, based on the feature importance ranking and the rule paths, the interpretability of the prediction results is verified, yielding corresponding interpretability evaluation conclusions.

[0115] For details on the implementation of each of the above operations, please refer to the previous examples, which will not be repeated here.

[0116] The storage medium may include: read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0117] Since the instructions stored in the storage medium can execute the steps in any of the data interpretable prediction methods based on tree regular variational encoders provided in the embodiments of this application, the beneficial effects that any of the data interpretable prediction methods based on tree regular variational encoders provided in the embodiments of this application can achieve can be realized. For details, please refer to the previous embodiments, which will not be repeated here.

[0118] The above provides a detailed description of a data interpretable prediction method and apparatus based on a tree regular variational encoder provided in the embodiments of this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The description of the above embodiments is only for the purpose of helping to understand the method and its core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A data-interpretable prediction method based on a tree-regular variational encoder, characterized in that, include: The high-dimensional data of the target is encoded into the latent space by a variational autoencoder to obtain a latent variable representation; The latent variable representations are regularized and aligned based on the surrogate decision tree so that the distribution structure of the latent variable representations in the latent space is consistent with the rule path of the surrogate decision tree, thereby generating a regularized latent space. Classification prediction is performed based on the regularized latent space, and interpretable prediction results are output. The prediction results are subjected to feature attribution analysis to generate a ranking of feature importance. Based on the feature importance ranking and the rule path, the interpretability of the prediction results is verified, and the corresponding interpretability evaluation conclusion is obtained.

2. The data interpretable prediction method based on a tree-regular variational encoder according to claim 1, characterized in that, The process of encoding the target high-dimensional data into the latent space using a variational autoencoder to obtain a latent variable representation includes: The target high-dimensional data is input into the encoder of the variational autoencoder to obtain the mean and variance parameters of the latent variables; The latent variable representation is generated by sampling based on the mean and the variance parameter; The latent variable representation is input into the decoder of the variational autoencoder to reconstruct the output data, and the corresponding reconstruction loss is calculated based on the output data and the target high-dimensional data.

3. The data interpretable prediction method based on a tree-regular variational encoder according to claim 1, characterized in that, The step of performing regularization alignment processing on the latent variable representation based on the surrogate decision tree, so that the distribution structure of the latent variable representation in the latent space is consistent with the rule path of the surrogate decision tree, and generating a regularized latent space, includes: Based on the current latent variable representation, the agent decision tree is trained, wherein each leaf node of the agent decision tree corresponds to a rule path from the root node to that leaf node; For each input data sample, calculate the matching degree between the input data sample and the rule paths of each leaf node of the agent decision tree, and normalize it into a soft distribution label; By minimizing the difference between the output of the variational autoencoder and the soft distribution label, the distribution structure represented by the latent variables is aligned with the rule path, thereby generating a regularized latent space.

4. The data interpretable prediction method based on a tree-regular variational encoder according to claim 3, characterized in that, The rule-based alignment process for the latent variable representation based on the surrogate decision tree further includes: When a new agent decision tree is used to replace an old agent decision tree, the parameters of the old agent decision tree are transferred to the new agent decision tree based on the path similarity between leaf nodes and the overlap of sample allocation. Before replacement, the new agent decision tree is subjected to a triple gating check based on the joint criteria of feature attribution consistency, leaf node assignment stability, and model fidelity. Replacement is only performed when all criteria are met.

5. The data interpretable prediction method based on a tree-regular variational encoder according to claim 4, characterized in that, The step of migrating the parameters of the old agent decision tree to the new agent decision tree includes: Calculate the path similarity and sample allocation overlap between each new leaf node of the new agent decision tree and each old leaf node of the old agent decision tree; For each new leaf node, based on the calculated path similarity and the sample allocation overlap, the old leaf node that matches the new leaf node is determined from the old leaf nodes; The alignment header parameters of the matched old leaf node are inherited to the new leaf node using a convex combination method. For new leaf nodes that do not match old leaf nodes, initialize them using default parameters.

6. The data interpretable prediction method based on a tree-regular variational encoder according to claim 3, characterized in that, The rule-based alignment process for the latent variable representation based on the surrogate decision tree further includes: The average path length of the agent decision tree is calculated as a complexity measure. Based on the average path length and its corresponding model parameters, a proxy neural network is trained to establish a prediction mapping from model parameters to complexity. The complexity value predicted by the surrogate neural network for the current model parameters is added as a regularization term to the overall optimization objective.

7. The data interpretable prediction method based on a tree-regular variational encoder according to claim 1, characterized in that, The classification prediction based on the regularized latent space, and the output of interpretable prediction results, includes: The latent variable representations in the regularized latent space are input into a prediction function to obtain initial prediction values; Based on the latent variable representation, the corresponding rule path is retrieved from the leaf node activated in the agent decision tree; The initial predicted value, the activated rule path, and the key features related to the rule path are fused to form the interpretable prediction result.

8. The data interpretable prediction method based on a tree-regular variational encoder according to claim 1, characterized in that, The step of performing feature attribution analysis on the prediction results to generate a feature importance ranking includes: A strategy of weighting by feature subset size is adopted to allocate the sampling budget to generate multiple feature subsets; Estimate the contribution of each feature subset to the prediction result to obtain the corresponding contribution estimate; Aggregate the contribution estimates of all feature subsets, calculate and generate the feature importance ranking.

9. The data interpretable prediction method based on a tree-regular variational encoder according to claim 1, characterized in that, The process of verifying the interpretability of the prediction results based on the feature importance ranking and the rule path, and obtaining the corresponding interpretability evaluation conclusion, includes: The feature importance ranking and the rule path of the proxy decision tree are transformed into structured explanatory data; The structured explanatory data is input into the large language model, which automatically generates a standardized verification questionnaire based on a preset prompt template. Collect feedback answers from domain experts on the validation questionnaire, and compare the feedback answers with the model's own explanatory logic to obtain the comparison results; Based on the comparison results, a quantitative interpretability assessment conclusion is calculated.

10. A data-interpretable prediction device based on a tree-regular variational encoder, characterized in that, include: The encoding module is used to encode the target high-dimensional data into the latent space through a variational autoencoder to obtain a latent variable representation. The alignment module is used to perform regular alignment processing on the latent variable representation based on the surrogate decision tree, so that the distribution structure of the latent variable representation in the latent space is consistent with the rule path of the surrogate decision tree, thereby generating a regularized latent space; The prediction module is used to perform classification prediction based on the regularized latent space and output interpretable prediction results; The analysis module is used to perform feature attribution analysis on the prediction results and generate a feature importance ranking. The verification module is used to verify the interpretability of the prediction results based on the feature importance ranking and the rule path, and obtain the corresponding interpretability evaluation conclusion.

Images