Machine learning-based element system multi-level contribution deconstruction and aggregation method

Abstract

The invention relates to the field of data analysis and intelligent decision, and particularly discloses a multi-level contribution deconstruction and aggregation method for an element system based on machine learning, which comprises the following steps: extracting element observation data, generating a standardized element feature vector, and forming a system feature matrix; training a relational machine learning model by using the system feature matrix and the overall efficiency data, and automatically generating a multi-level contribution deconstruction tree; selecting nodes in the contribution deconstruction tree for verification, and dynamically adjusting and optimizing the model and the tree structure according to the deviation between a predicted value and an actual value to obtain a final contribution deconstruction tree; based on the final contribution deconstruction tree, contribution value weighted aggregation is carried out from bottom to top, and the comprehensive contribution rate of each element is calculated through recursion. According to the method, the complex contribution relationship is automatically and objectively deconstructed through machine learning, the accuracy and adaptability of the model are ensured by means of verification of a closed loop, the quantitative contribution rate with a clear structure and high interpretability is finally output, and the reliability and decision support value of a calculation result are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of data analysis and intelligent decision-making technology, specifically to a method for multi-level contribution deconstruction and aggregation of element systems based on machine learning. Background Technology

[0002] For complex systems of elements, such as technological architecture, organizational capabilities, and ecosystems, scientifically calculating the system contribution rate is crucial for performance evaluation, resource optimization, and bottleneck diagnosis. Such calculations are widely needed in various fields, including social network impact analysis, environmental system modeling, and engineering structure optimization.

[0003] Traditional methods for calculating system contribution rates primarily rely on a hierarchical decomposition followed by weighted aggregation. However, this method has inherent flaws: First, the establishment of its hierarchical structure and the allocation of element weights are highly dependent on expert experience, resulting in a subjective process and difficulty in standardizing processes. Second, it simplifies complex interactions into linear superposition, failing to capture the prevalent nonlinear, indirect, and hidden contribution relationships among elements; for example, team atmosphere indirectly influences the final output through multiple mediating paths. This often leads to analytical results that lack authority and are unconvincing.

[0004] Existing technologies have attempted to improve these methods by incorporating artificial intelligence models, such as reinforcement learning models for system evaluation. However, these methods heavily rely on high-quality training datasets and simulation environments that can accurately simulate real-world feedback. In many practical application scenarios where complete historical data is lacking or controlled experiments cannot be conducted, they often fail to be implemented due to data scarcity or environmental mismatch, resulting in poor universality.

[0005] The core contradiction in the current technological field lies in two aspects: on the one hand, traditional models are distorted due to their oversimplification and subjectivity; on the other hand, advanced data-driven models are unsuitable due to their stringent requirements for training data. The fundamental problem with existing technologies is the failure to effectively construct a universal method that can automatically learn complex contribution relationships from limited and mixed real-world data, while also possessing inherent verification and optimization mechanisms to ensure the reliability of the results. This leads to long-standing ambiguity in the calculation of contribution rates among elements, and the lack of robust reliability in the calculated system contribution rates. Therefore, a machine learning-based method for multi-level contribution deconstruction and aggregation of element systems is urgently needed to solve these problems. Summary of the Invention

[0006] The purpose of this invention is to provide a multi-level contribution deconstruction and aggregation method for an element system based on machine learning, so as to solve the problems of strong subjectivity and poor universality of existing data-driven models in the prior art.

[0007] To solve the above-mentioned technical problems, the present invention specifically provides the following technical solution:

[0008] A machine learning-based method for multi-level contribution deconstruction and aggregation of an element system includes the following steps:

[0009] S1. Extract element observation data, generate standardized element feature vectors, and construct the system feature matrix;

[0010] S2. Using the system feature matrix and overall performance data, train a relational machine learning model to automatically generate a multi-level contribution deconstruction tree;

[0011] S3. Select nodes in the multi-level contribution deconstruction tree for verification, and optimize based on the deviation between the predicted value and the actual value to obtain the final contribution deconstruction tree;

[0012] S4. Based on the final contribution deconstruction tree, perform weighted aggregation of contribution values ​​from bottom to top, and recursively calculate the comprehensive contribution rate of each element.

[0013] As a preferred embodiment of the present invention, S1 specifically includes:

[0014] S11. For each element in the target element system, collect its static attribute data and dynamic time-series status data from multiple preset data sources;

[0015] S12. Preprocess each collected element data, including at least missing value imputation, outlier detection and correction, and data standardization to eliminate dimensional differences;

[0016] S13. The preprocessed multidimensional data of each element are spliced ​​and encoded according to predetermined rules to generate a fixed-dimensional, numerical element feature vector.

[0017] S14. Use the feature vectors of all elements as row vectors, arrange them in the order of element index, and construct a complete system feature matrix.

[0018] As a preferred embodiment of the present invention, S2 specifically includes:

[0019] S21. Using the system feature matrix as input features and the overall system performance output data as supervision labels, supervise the training of the relational machine learning model so that the model learns the complex mapping relationship from element features to overall performance and the internal correlation pattern between element features.

[0020] S22. From the converged relational machine learning model, extract the contribution weights of each input feature dimension of each element to the predicted output, as the initial correlation strength between elements and between elements and the overall effectiveness.

[0021] S23. Taking the overall system efficiency as the root node, based on the initial correlation strength, a threshold-based correlation screening and recursive hierarchical algorithm is used to organize the elements with significant correlation strength into child nodes layer by layer, and assign corresponding weights to the directed edges between nodes, thereby automatically constructing a multi-level contribution deconstruction tree.

[0022] As a preferred embodiment of the present invention, S22 specifically includes:

[0023] S221. The relational machine learning model adopts a gradient boosting decision tree. By analyzing the model feature coefficients, it directly quantifies the direct contribution weight of each feature to the final prediction result. The direct contribution weight serves as the main component of the initial correlation strength of the feature to the overall effectiveness.

[0024] S222. Normalize and fuse the direct contribution weights and interaction influence strengths to form an initial correlation strength matrix, wherein the rows and columns of the matrix correspond to each element and the overall effectiveness, and the matrix element values ​​represent the correlation strength from one node to another.

[0025] As a preferred embodiment of the present invention, S23 specifically includes:

[0026] S231. Apply a dynamic hierarchical threshold to the element values ​​in the initial association strength matrix to filter and retain associations with strength values ​​exceeding the threshold, forming a significant association set;

[0027] S232. Taking the overall system effectiveness node as the starting layer, select all associations with the current layer node as the target from the significant association set, and organize the source element nodes corresponding to the associations into the next layer of sub-nodes;

[0028] S233. Normalize the initial association strength values ​​retained after filtering into relative contribution weights of child nodes to their parent nodes, and use these weight values ​​to generate directed edges from child nodes to parent nodes;

[0029] S234. Using the newly generated child node layer as the current layer, repeat the operation recursively to expand the tree structure downwards until a certain layer can no longer match new element nodes that have not been organized into the tree from the significant association set, and finally generate a complete multi-level contribution deconstruction tree.

[0030] As a preferred embodiment of the present invention, S3 specifically includes:

[0031] S31. Based on the topology of the multi-level contribution deconstruction tree and prior business knowledge, select several nodes as verification nodes;

[0032] S32. For each verification node, after masking the feature data of the corresponding element of the node in the system feature matrix, input the relational machine learning model to obtain the model prediction contribution value of the node, and obtain the actual observed contribution value of the node, and calculate the deviation between the two.

[0033] S33. If the deviation of any verification node exceeds the preset tolerance threshold, the optimization process is triggered;

[0034] S34. Repeat the operation until the deviation of all selected verification nodes is lower than the tolerance threshold. The resulting contribution deconstruction tree is the final contribution deconstruction tree.

[0035] As a preferred embodiment of the present invention, S32 specifically includes:

[0036] S321. For the selected verification node, create its corresponding contribution evaluation dataset. The contribution evaluation dataset is generated by setting all the feature vectors of the elements corresponding to the nodes in the system feature matrix to zero or baseline values.

[0037] S322. Input the contribution evaluation dataset into the relational machine learning model. The model re-infers based on the feature data of the remaining elements and implicitly outputs a temporary local deconstruction tree for the current data state according to its internal correlation logic.

[0038] S323. Based on the structure and weights of the temporary local deconstruction tree, calculate the change in the overall system performance prediction value caused by the occlusion of target elements. The change is used as the model prediction contribution value of the verification node.

[0039] S324. Compare the model-predicted contribution value of the verification node with its actual observed contribution value measured independently, and calculate the specific deviation value using the root mean square error formula.

[0040] As a preferred embodiment of the present invention, S33 specifically includes:

[0041] S331. The original feature vector of the verification node that generates the excess deviation, its occluded feature vector, and the calculated deviation value are encapsulated together into a high-priority training sample and injected into the original training dataset of the relational machine learning model.

[0042] S332. Using the expanded training dataset, focus on high-priority training samples with higher learning weights, and retrain the relational machine learning model to correct the model's cognitive bias regarding the contribution relationship of nodes, thereby obtaining a retrained model.

[0043] S333. Extract the updated association patterns from the retrained model to generate a new initial association strength matrix, which reflects the optimized association perception between elements and the elements' impact on overall effectiveness;

[0044] S334. Using the new initial association strength matrix as input, re-execute the tree structure generation step to reconstruct the structure of the multi-level contribution deconstruction tree and generate an updated tree structure for the next round of verification.

[0045] As a preferred embodiment of the present invention, S4 specifically includes:

[0046] S41. Set the contribution value of all leaf nodes in the final contribution deconstruction tree as the preset initial unit contribution value of their corresponding feature vectors;

[0047] S42. Starting from the bottom-level node, for any non-leaf node, multiply the contribution values ​​of all its direct child nodes by the weights of the directed edges between the node and each child node, and sum them up. The result is used as the aggregated contribution value of the non-leaf node.

[0048] S43. Contribution rates are calculated and transmitted synchronously during the aggregation process;

[0049] S44. Repeat the aggregation calculation and contribution rate transfer calculation process, recursively calculating upwards layer by layer until the contribution value of the root node is calculated, and simultaneously obtain the comprehensive contribution rate of each leaf node to the root node as the final output.

[0050] As a preferred embodiment of the present invention, S43 specifically includes:

[0051] S431. For any leaf node, in the final contribution deconstruction tree, starting from the leaf node, traverse upwards along the direction pointing to the parent node to the root node, recording the weight of each directed edge traversed; the initial contribution value of the leaf node is multiplied by the weights of all edges on the traversal path, which is taken as the path contribution rate of the leaf node to the root node through a specific path; a leaf node may indirectly affect the root node through multiple paths, such as through different intermediate convergence nodes, and the path contribution rate of all possible paths needs to be calculated.

[0052] S432. Sum the contribution rates of all paths calculated for the same leaf node, and the sum is the comprehensive contribution rate of the leaf node element to the overall system effectiveness.

[0053] S433. After each layer of aggregation calculation in step S42 is completed, the contribution value of non-leaf nodes is updated. The comprehensive contribution rate of a node to the root node is equal to the sum of the comprehensive contribution rates of all its child nodes. The direct contribution rate of a node to its direct parent node is defined by the weight of the corresponding edge when the tree is generated and is directly used in the aggregation calculation.

[0054] Compared with the prior art, the present invention has the following advantages:

[0055] This invention uses a machine learning model to automatically learn the relationships between elements from data, replacing the traditional model that relies on manual subjective weight setting. It can effectively identify and quantify indirect, multi-path nonlinear contributions, thus solving the analytical blind spots in this field.

[0056] This invention uses a masking verification-bias-driven tuning mechanism to automatically verify and correct the initial model, ensuring the adaptability and long-term reliability of the analysis results as the system evolves, and overcoming the problem of static models being prone to failure.

[0057] The contribution deconstruction tree generated by this invention intuitively reveals the contribution path, intensity, and transmission chain of each element, providing not only a quantitative contribution rate but also a clearer view of the contribution source, greatly enhancing the decision support depth of the analysis results. Attached Figure Description

[0058] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0059] Figure 1 This is a schematic diagram summarizing the process of the method described in Embodiment 1 of the present invention.

[0060] Figure 2 This is a detailed flowchart of the method described in Embodiment 1 of the present invention. Detailed Implementation

[0061] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0062] The concepts involved in this application will first be described with reference to the accompanying drawings. It should be noted that the following descriptions of various concepts are only for the purpose of making the content of this application easier to understand and do not constitute a limitation on the scope of protection of this application; furthermore, the embodiments and features in the embodiments of this application can be combined with each other unless otherwise specified. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0063] Example 1

[0064] like Figure 1 - Figure 2 As shown, this invention provides a multi-level contribution deconstruction and aggregation method for an element system based on machine learning, including the following steps:

[0065] S1. Extract element observation data, generate standardized element feature vectors, and construct the system feature matrix; specifically including:

[0066] S11. For each element in the target element system, implement comprehensive collection of multi-source heterogeneous data. Specifically, for a specific element, two types of key data need to be simultaneously acquired from multiple pre-defined data sources associated with it:

[0067] The first category is static attribute data, which covers features that do not change over time, such as the inherent physical attributes, design specifications, and technical performance indicators of elements.

[0068] The second category is dynamic time-series state data, which is serialized data with timestamps that reflects the state evolution, performance fluctuations, and environmental responses of elements during actual operation.

[0069] The data collection process must ensure the diversity and coverage of data sources to support the completeness of the subsequent contribution relationship deconstruction.

[0070] S12. Systematically preprocess each element data collected in S11 to eliminate noise and inconsistencies in the raw data and improve data quality. This preprocessing includes at least three progressive steps:

[0071] First, missing value imputation is performed. For missing values ​​caused by sensor failure or packet loss during data acquisition, interpolation algorithms, statistical mean imputation, or collaborative filtering methods based on similar elements are used to fill in the missing values.

[0072] Secondly, outlier detection and correction are carried out. Outliers that deviate from the normal distribution are identified through box plot analysis, isolated forest algorithm or 3σ criterion, and smoothed or removed by combining physical rationality judgment.

[0073] Finally, data standardization is implemented using Z-score standardization, minimum-maximum scaling, or vector normalization methods to eliminate dimensional differences between different feature dimensions and ensure that the features of each element are comparable within a unified numerical space.

[0074] S13. The multidimensional data of each element after preprocessing in S12 is transformed into a structured numerical representation. This process follows predetermined rules for splicing and encoding:

[0075] The splicing operation connects the processing results of static attribute data and dynamic time-series state data into a unified vector according to feature category priority or time series order;

[0076] The encoding operation uses one-hot encoding or embedding encoding to map categorical features to numerical values, while for continuous features, it retains their floating-point values.

[0077] Finally, a fixed-dimensional feature vector is generated. The number of dimensions of this vector is predefined by the complexity of the feature system, and each dimension has a clear physical or statistical meaning.

[0078] S14. Organize the feature vectors of all elements into a matrix. In practice, each feature vector is used as an independent row vector of the matrix, and arranged vertically according to the feature index order such as feature number, functional category, or hierarchical attribute to construct a complete system feature matrix.

[0079] The row dimension of this matrix corresponds to the total number of elements in the element system, and the column dimension corresponds to the fixed feature dimension defined in S13. Each element value in the matrix has been standardized by S12, thus forming a high-dimensional input tensor with regular numerical values, clear structure, and usable for training machine learning models.

[0080] S2. Train a relational machine learning model using the system feature matrix and overall performance data to automatically generate a multi-level contribution deconstruction tree; specifically including:

[0081] S21. As the modeling foundation for relation deconstruction, its core lies in constructing a supervised learning framework capable of simultaneously capturing linear, nonlinear, and higher-order interaction effects. In practical implementation:

[0082] S211. Use the system feature matrix output from step S14 as the model input feature tensor. This tensor has a clear physical structure: its row axes correspond to the N independent elements in the element system, and its column axes correspond to the M-dimensional standardized feature dimensions. Each element value reflects the quantization state of a specific element in a specific feature dimension. The overall system performance output data, used as the supervision label, needs to be strictly aligned with the feature matrix in the sample dimension. If it is multi-scenario batch training, it forms a dataset containing K observation samples. Each sample contains an N×M feature matrix block and a performance label in scalar or vector form.

[0083] S212. The selection of a relational machine learning model needs to balance prediction accuracy and interpretability. Gradient Boosting Decision Tree (GBDT) is adopted. During model initialization, the first decision tree is constructed, and its root node receives the flattened system feature matrix as input. A greedy algorithm selects the feature dimension that minimizes the sum of squared residuals of the performance prediction and the splitting threshold for node splitting. Subsequently, the second tree fits the prediction residuals of the first tree, and this process is iterated until the preset number of trees or the residual convergence threshold is reached. During training, the model not only learns the monotonic / non-monotonic mapping from single-feature to performance, but also implicitly captures the synergistic enhancement effect between features through the co-occurrence splitting patterns of different feature features on the same type of node in the tree structure. For example, when two feature features exceed the threshold at the same time, the performance is significantly improved; or the antagonistic inhibition effect is achieved, such as when one feature improves while the other decreases, the performance is offset.

[0084] S213. To prevent overfitting and ensure the rationality of physical meaning, the training process needs to embed multiple regularization mechanisms: L1 / L2 regularization terms are added to the loss function to constrain the sparsity of feature weights; an early stopping strategy is adopted to terminate training when the prediction error of the validation set performance begins to rise; and hyperparameters such as limiting tree depth and the minimum number of samples in leaf nodes are used to prevent the model from excessively mining noisy correlations. After sufficient training, the ensemble structure formed by the model can approximate arbitrarily complex continuous mapping functions. Its internal node splitting paths essentially constitute a contribution transmission network from element features to overall performance, laying an analytical topological foundation for contribution extraction in stage S22.

[0085] S22. Deconstruct the trained black-box model, transforming its implicit complex mapping relationships into explicit, measurable association strength indicators. This stage requires penetrating the ensemble structure of the gradient boosting decision tree, systematically extracting the independent contributions of elements to overall effectiveness and the interactive synergistic effects between elements, ultimately constructing a structured representation network. Specifically:

[0086] S221. The primary level of association strength extraction focuses on quantifying the direct contribution weight of a single feature. In the gradient boosting decision tree architecture, each decision tree consists of a series of split nodes forming a non-linear decision boundary. Each split node divides the sample space into two sub-regions based on a specific feature dimension threshold of a certain feature. To quantify this process, a multi-dimensional contribution measurement fusion strategy is required: First, calculate the feature splitting frequency index by traversing all non-leaf nodes of all decision trees and counting the frequency with which each feature dimension is selected as a splitting criterion. High-frequency features indicate that they are identified as key criteria for performance prediction in most tree structures. Second, calculate the information gain contribution value by calculating the reduction in Gini impurity or mean square error before and after splitting for each node using feature f. This reduction is the information gain of that split. The gain values ​​of all relevant nodes of the feature in all trees are summed to obtain the total amount of uncertainty reduction of the feature in the predicted output. Finally, introduce the SHAP value for marginal contribution actuarial calculation. Based on the Shapley value principle in cooperative game theory, calculate the average marginal contribution of each feature to the model prediction value in a specific sample context. The stable contribution weight of the feature is obtained by averaging the absolute values ​​of the SHAP values ​​of all training samples. The three types of metrics mentioned above are weighted and fused after Z-score standardization to form the main component of the initial correlation strength of factors with overall effectiveness. This value increases monotonically between 0 and 1, intuitively reflecting the importance ranking of factors.

[0087] The second level of association strength extraction delves deeper into the intensity of interactions between elements. Gradient boosting decision trees naturally possess the ability to capture interaction effects. When a split in a node based on element A leads to element B becoming a subsequent key splitting feature in a child node, an interaction path between A and B is constructed. Specifically, quantification requires identifying all node pairs that satisfy the condition "parent node splitting feature is element A, child node splitting feature is element B," calculating the transfer product of information gain along this path, and using this as the unidirectional interaction strength from A to B. Further, by traversing all such parent-child node relationships in the tree structure and accumulating the cross-tree interaction transfer amount, a bidirectional interaction strength matrix between elements is obtained. This matrix maintains the same dimension as the direct contribution weight matrix, but its physical meaning characterizes the nonlinear coupling effect between elements, such as how the state change of element A modulates the contribution efficiency of element B to the overall effectiveness through the tree structure path.

[0088] S222. The final generation of the initial correlation strength requires a rigorous normalization and fusion calculation process:

[0089] First, Min-Max normalization is applied to the direct contribution weight matrix and the interaction influence strength matrix respectively, scaling their numerical range to the [0,1] interval to eliminate dimensional bias caused by different calculation methods. Then, a dynamic weighted fusion strategy is adopted, and the fusion coefficient is determined based on the model prediction error analysis. If the model residuals mainly come from nonlinear interactions, the weight of the interaction matrix is ​​increased; otherwise, the direct contribution matrix dominates. The fusion calculation can be performed using the matrix Hadamard product, i.e., element-wise multiplication, to generate the initial correlation strength matrix.

[0090] This matrix is ​​an N+1 dimensional square matrix, where N rows / columns correspond to N elements, and an additional row and column correspond to the overall system effectiveness nodes. The matrix element value is α. ij This matrix represents the directed association strength from source node i to target node j, satisfying nonnegativity and asymmetry. Diagonal elements are defined as 1, representing node self-associations. This matrix serves as the sole input for generating the tree structure in stage S23, and its sparsity and numerical distribution directly determine the topological shape and hierarchical depth of the deconstructed tree.

[0091] S23. The initial association strength matrix is ​​transformed into a multi-level contribution deconstruction tree with hierarchical semantics through dynamic thresholding and a recursive hierarchical algorithm. This process must strictly follow the logical chain of saliency screening, hierarchical tracing, weight normalization, and recursive termination to ensure that the generated tree structure is both faithful to the original association strength distribution and possesses clear physical interpretability. Specifically:

[0092] S231. The dynamic hierarchical threshold filtering mechanism is key to tree structure sparsity and noise reduction. This mechanism performs layer-by-layer adaptive filtering on the N×(N+1) elements in the initial association strength matrix. Specifically:

[0093] At the root node level, i.e., the 0th level, the threshold is set to the upper quartile of the matrix intensity distribution plus 1.5 times the interquartile range to ensure that only the top-level correlations that have the most significant impact on the overall system performance are retained.

[0094] As the tree depth extends downwards to the Lth level, the threshold Threshold(L) is dynamically adjusted to a decay coefficient Mean(Parent) that is 0.7 times the average association strength of the parent nodes. Strengths (and the global median intensity) Median The weighted combination, i.e., Threshold(L) = w×Mean(Parent) Strengths ) + (1-w)×Global Median The weight w decreases linearly with depth. This dynamic strategy effectively handles the natural decay of association strength among deep nodes, avoiding over-pruning of deep associations due to a fixed threshold.

[0095] The strength values ​​retained after filtering constitute a significant association set S, whose elements are in the form of a three-way group: source node ID, target node ID, and association strength value. This set will serve as the only valid edge pool for subsequent tree structure generation.

[0096] S232. Initiate the topology construction of the tree structure, with the system's overall performance node E0 as the root node and the starting layer;

[0097] In the first iteration, all three-way groups with target node IDs equal to E0 are retrieved from the significant association set S. The source node IDs of these three-way groups are the element nodes that directly and significantly contribute to the overall effectiveness of the system, and they are organized into the first-level child node set N1. This process essentially achieves reverse tracing through the directionality of association strength, that is, pointing from the effectiveness result to the contribution source, ensuring that the tree structure conforms to causal logic.

[0098] In the second and subsequent iterations, the newly generated child node layer N will be... i As the current target layer, we iterate through the significant association set S again, retrieving all target node IDs belonging to N. i The three-way groups, whose source node IDs are the secondary elements that significantly contribute to the current layer's elements, are organized into a set of child nodes N for the next layer. i+1 This recursive retrieval process continues, with each node generated based on its significant association strength with the target node in the previous layer, thus naturally forming a hierarchical chain of contribution propagation.

[0099] S233. Responsible for the precise assignment and relativistic normalization of directed edge weights. For each parent-child node relationship generated by S232, the parent node N... i Child node N i+1 Extract the original association strength value α from the significant association set S.

[0100] To ensure horizontal comparability of weights, local normalization needs to be implemented at the parent node level: for parent node N i All corresponding child nodes N i+1 The set of association strength values ​​is subjected to Softmax transformation or L1 normalization to generate relative contribution weights. These weights directly reflect the proportion of contribution of child nodes to their parent nodes within the same level. Their values ​​range from (0,1) and satisfy the constraint that the sum of all weights is 1.

[0101] Subsequently, the weight values ​​are assigned as the edge weights of the directed edges from the child nodes to the parent nodes, with the direction of the edges strictly following the contribution flow, i.e., the lower-level elements point to the higher-level efficiency, thereby embedding quantitative transmission relationships in the tree structure.

[0102] S234. Implement recursive termination checks and integrity verification for the tree structure. Each time, a new child node layer N is generated. i+1Then, two termination conditions need to be checked:

[0103] First, determine N i+1 Whether it is an empty set, that is, whether a new source node targeting the current layer node could not be retrieved from the significant association set S. If it is empty, it means that the contribution transmission chain has reached the bottom layer of the system elements, and the recursion will terminate naturally.

[0104] Secondly, determine whether there are any unassigned free element nodes, that is, elements that have not yet been organized into the tree structure and whose association strength is lower than the current layer threshold. If so, trigger the threshold backtracking mechanism, appropriately reduce Threshold(L) and re-execute S231-S233 to ensure that all elements are reasonably included in the tree structure.

[0105] The resulting multi-level contribution deconstruction tree has strict single-root characteristics, i.e., only E0 is the root, acyclic and directional, with edges always pointing upwards, as well as weight normalization properties. Its tree depth, number of branches, and node distribution directly map the true contribution hierarchy of the element system, providing a topologically complete and weighted input structure for bottom-up aggregation calculation in the S4 stage.

[0106] S3. Select nodes in the multi-level contribution deconstruction tree for verification, and fine-tune the tree based on the deviation between the predicted and actual values ​​to obtain the final contribution deconstruction tree; specifically including:

[0107] S31. Select a set of representative, high-value, and risk-sensitive validation nodes from the multi-level contribution deconstruction tree as probes for subsequent model accuracy evaluation and structure tuning. In practice, the topological characteristics of the deconstruction tree and prior knowledge of the business domain must be comprehensively considered.

[0108] At the topology level, nodes located at different depths of the tree are prioritized, including core elements at the first layer that are directly related to the root node, such as key performance indicator elements; high-connectivity hub nodes in the intermediate transmission layer, such as aggregated elements that multiple child nodes point to; and basic elements at the leaf nodes at the end of the tree, such as bottom-level sensor data elements, to ensure the full path of verification coverage contribution transmission.

[0109] At the business level, by combining historical operation logs, expert experience databases, and failure case databases, we identify key elements that have a decisive impact on the overall system performance, high-risk elements that have experienced performance degradation or data drift, and newly introduced or structurally changed elements, and include them in the verification node set.

[0110] The final number of selected verification nodes usually accounts for 15% to 25% of the total number of elements, and their distribution in the tree structure must be balanced to avoid concentration in a certain subtree leading to verification blind spots.

[0111] S32. Construct a counterfactual reasoning framework to quantify the model's cognitive accuracy regarding the contribution relationship of a node by systematically masking the observational information of that node. This stage requires accurately isolating the influence of the target element while maintaining the integrity of the characteristics of other elements in the system, and reconstructing a temporary contribution transmission path within the model. Finally, the deviation is computable by comparing the predicted perturbation with the actual perturbation; specifically:

[0112] S321. Refined Construction of the Contribution Evaluation Dataset:

[0113] For each selected verification node, its precise coordinates in the system feature matrix must first be located: if the feature index corresponding to the verification node is i, then the original feature vector v of that feature is extracted. i .

[0114] Masking strategies should be flexibly selected based on the data type of the elements and the business scenario: For sparse binary features representing switch status or existence indications, such as equipment start / stop flags, zero-based masking is used to simulate complete element failure; for dense features representing continuous physical quantities, such as temperature, pressure, and vibration amplitude, the values ​​are filled with the statistical baseline values ​​of that dimension in the training set, typically the mean μ. j or median med j To ensure that the feature vector v after occlusion i_masked The distribution is significantly different from normal observations, thus avoiding the introduction of abnormal disturbances.

[0115] A more refined approach is to use a multi-dimensional occlusion matrix to apply differentiated occlusion strengths to different feature dimensions. For example, the core dimension is completely occluded with a weight of 0, while the auxiliary dimension is partially occluded with a weight of 0.5, thereby simulating a gradual scenario of feature performance degradation rather than complete failure.

[0116] The masked feature vectors replace the i-th row of the original matrix, generating a contribution evaluation dataset X that only lacks target feature information. masked This dataset has the same dimensions as the original matrix, but its information entropy is significantly reduced.

[0117] S322. Drive relational machine learning models to perform counterfactual reasoning and implicitly generate temporary local deconstruction trees. Specifically:

[0118] X maskedAfter inputting into the model, each tree in the gradient boosting decision tree will undergo forward propagation based on the features of the remaining N-1 elements. Due to the lack of information about the target element, nodes that originally relied on that element for key splits will be forced to activate alternative split paths. The model completes inference by adjusting the split threshold or shifting to secondary features. During this process, a temporary local deconstruction tree is automatically formed internally within the model. Its generation mechanism is not an explicit call to the S23 algorithm, but rather a reverse reconstruction by capturing the changing patterns of the model's output gradient and intermediate layer activation values. Specifically:

[0119] The computational model outputs a Jacobian matrix for the remaining feature characteristics, where each element J in the matrix... kj The sensitivity of the j-th dimension feature of the k-th remaining element to the predicted output is represented. Based on the sensitivity threshold, a set of elements whose importance significantly increases after the target element is masked is identified. These elements constitute compensatory contribution nodes in the temporary deconstruction tree. Simultaneously, by tracking the feature usage frequency changes of the split nodes in GBDT, the dynamic association paths between the remaining elements can be plotted, forming a temporary tree structure with the overall system effectiveness as the root and compensatory elements as child nodes. This structure maps the adaptive reconstruction logic of the model under conditions of incomplete information.

[0120] S323. Accurately calculate the change in the predicted overall system performance caused by the obscuring of target elements. In practice, two independent forward inference steps are required:

[0121] The first use of the unmasked original system feature matrix X original Input the model to obtain the baseline predicted value P. base This value reflects the model's performance under complete information;

[0122] Second use of contribution evaluation dataset X masked Input the model to obtain the predicted attenuation value P. masked This value reflects the model's effectiveness assessment after missing target element information.

[0123] The two inferences must be performed under the premise of fixed model parameters, locked random seeds, and consistent batch size to ensure that the differences only stem from the masking operation of the input data.

[0124] Change ΔP = P base - P masked Directly defined as the model prediction contribution value of the validation node, its physical meaning is the magnitude of the marginal effect of the target element within the model's cognitive framework. To improve computational robustness, an integrated perturbation strategy can be adopted: for the masked feature vector v i_masked A small random noise ε is applied, which follows a normal distribution with a mean of 0 and a variance of σ², where σ is 5% of the standard deviation of the feature dimension, generating multiple perturbation versions X.masked k Calculate multiple perturbation versions X respectively masked k The change ΔP k Then take the mean μ ΔP With standard deviation σ ΔP Construct confidence intervals [μ] for the predicted contribution values. ΔP - 2σ ΔP ,μ ΔP + 2σ ΔP This interval can be used to assess the degree of uncertainty in the model's understanding of the relationship between the contribution of the node and the node.

[0125] S324. Perform the predicted contribution value ΔP pred Compared with the actual observed contribution value ΔP obs The deviation quantification is as follows:

[0126] Obtaining the actual observed contribution value requires objective measurement methods independent of the model: In physically isolated systems, the controlled variable method can be used to adjust the operating parameters of this element individually under stable operating conditions, such as changing the equipment power setpoint by ±10%, while simultaneously recording the actual response of the overall system efficiency. The true marginal contribution ΔP of this element can then be extracted through linear regression or difference calculation. obs In complex systems that cannot be isolated, statistical analysis is performed on natural fluctuations in historical data. Time segments with significant changes in the characteristic values ​​of the target element while other elements remain relatively stable are selected to construct a local regression model to estimate ΔP. obs .

[0127] After obtaining two sets of values, the deviation is calculated using the root mean square error (RMSE) formula: RMSE = √[

[0128] (ΔP pred -ΔP obs [ )²], this indicator has a penalizing amplification effect on large deviations. To enhance the comprehensiveness of the assessment, the mean absolute percentage error (MAPE) can be calculated simultaneously: MAPE = |ΔP pred -ΔP obs | / |ΔP obs When MAPE > 20% or RMSE exceeds the preset tolerance threshold, it is usually set to 0.5 times the standard deviation of all validation node observations. This indicates that the model has a systematic cognitive bias at that validation node, triggering the optimization process in stage S33.

[0129] S33. When the deviation of a validation node exceeds the tolerance threshold, an adaptive tuning closed loop is triggered. Its core logic is to transform model prediction errors into targeted training signals. Through a progressive process of sample reweighting, incremental retraining, and structural reconstruction, a cascading effect is achieved, moving from local error correction to global cognitive optimization. Specifically:

[0130] S331. Construct high-priority training samples to accurately target the model's cognitive weaknesses. The specific encapsulation process requires extracting three types of key information:

[0131] One is to verify the complete element eigenvector v of the node in the original system feature matrix. original , as the input benchmark that the model should learn correctly;

[0132] Secondly, the feature vector v after S321 masking processing. masked , as negative sample input for model prediction inaccuracies;

[0133] Thirdly, the deviation value Δerror calculated by S324 serves as a strongly supervised signal to represent the cognitive gap between the model output and the actual contribution.

[0134] The three components are combined to form a structured training sample, which is then labeled with a high-priority tag. The injection strategy for this sample is not simply appending it to the end of the original training set, but rather employing head insertion or a dedicated buffer mechanism to ensure its high-frequency sampling in subsequent retraining batches. To further enhance the guiding nature of the sample, the topological position encoding of the validation node in the deconstruction tree can be injected simultaneously, such as a list of parent node IDs, the number of child nodes, and business semantic tags, such as key element identifiers. This allows the model to not only perceive numerical biases during training but also understand their importance and weight within the system architecture.

[0135] S332. Incremental retraining of the model requires a careful balance between retaining old knowledge and correcting new biases.

[0136] At the data sampling level, high-priority samples are given a significantly increased sampling weight γ, typically set to 5 to 10 times the weight of ordinary samples. This is achieved by introducing a sample-level weighting factor into the loss function. This forces the gradient descent process to impose a greater penalty on the prediction error of the validation nodes when updating parameters, accelerating the correction of cognitive biases.

[0137] At the optimization strategy level, a smaller learning rate η is used. fine_tune For example, the initial training learning rate η initial The learning rate is increased by 0.05 to 0.1 times, and is used in conjunction with a learning rate warmup and a cosine annealing scheduler to prevent parameter oscillations from disrupting the learned general correlation patterns.

[0138] At the parameter freezing level, the parameters of the top fully connected layer or output layer of the model are frozen, while only the bottom feature interaction layer is unfrozen, such as the first 50% of the tree structure of GBDT, to ensure that the optimization focuses on the local correction of contribution relationships rather than global reconstruction.

[0139] The retraining process continues until the validation node bias no longer decreases for three consecutive epochs or reaches the maximum number of iterations, such as 50 epochs. Then, the retrained model is output, and its internal parameters have been adaptively adjusted to address the weak points.

[0140] S333. Extracting the updated association patterns from the retrained model requires repeating the entire contribution analysis process from S22, but this analysis needs to focus on the differences in intensity changes. Specifically:

[0141] The SHAP value distribution and tree splitting gain matrix are recalculated on the retrained model to generate the updated initial association strength matrix A. new .

[0142] Compare this matrix with the original matrix A original Element-level comparison can identify key element pairs with intensity value change rates exceeding 15%. These element pairs typically contain the verification node and its parent / child nodes, as well as other associated paths adjusted due to cascading effects.

[0143] Further analysis of the direction of change in intensity values ​​reveals that if the intensity value of a verification node increases towards the actual observed contribution value, it indicates that the model has enhanced its understanding of the node's importance; if the weights of its parent nodes to child nodes are redistributed, it indicates that the model has corrected the priority order of contribution transmission. This differential analysis not only verifies the effectiveness of the tuning but also provides a basis for the structural reconstruction of S334.

[0144] S334. Based on A new To reconstruct the multi-level contribution deconstruction tree, the recursive hierarchical algorithm of S23 needs to be fully reproduced, but structural stability constraints need to be introduced.

[0145] When using dynamic threshold filtering in S231, if a significant association that exists in the original tree is found to have weakened in strength in the new matrix but not completely disappeared (e.g., if the value decreases from 0.8 to 0.6 but is still higher than the threshold), then the association is retained in the tree structure but its edge weight is reduced to avoid frequent pruning that could cause topological oscillations.

[0146] After hierarchical tracing (S232) and weight normalization (S233) are completed, an updated tree structure is generated. This structure may undergo local adjustments: existing relationships may be weakened and pruned, new significant relationships may emerge and new nodes may be added, and the relative weights between child nodes will also be rebalanced. For example, the weight of a certain element in the child nodes of a validation node may increase from 30% to 45%, reflecting its prominent contribution after optimization.

[0147] The reconstructed tree structure is used for the next round of S31-S32 verification, forming a closed-loop iteration of verification-bias detection-retraining-reconstruction, which is continuously optimized until global convergence.

[0148] S34. The system repeats the closed-loop process from S31 to S33. After each iteration, the deviation of all verification nodes is recalculated. If there are still nodes whose deviation exceeds the tolerance threshold, the next round of optimization continues. If the deviation of all verification nodes is lower than the threshold in two consecutive iterations, or the deviation decrease is less than the convergence tolerance, such as 1%, then the model and deconstruction tree are determined to have reached a stable optimal state.

[0149] The contribution deconstruction tree obtained at this point is the final contribution deconstruction tree. Its node level, edge weight, and association strength have all been calibrated and verified by actual observation data, and it has high credibility and strong interpretability. It can be used as the authoritative input for contribution rate aggregation calculation in the S4 stage.

[0150] S4. Based on the final contribution deconstruction tree, perform bottom-up weighted aggregation of contribution values ​​to recursively calculate the comprehensive contribution rate of each element; specifically including:

[0151] S41. Assign a baseline contribution value to all leaf nodes in the final contribution deconstruction tree. Specifically:

[0152] Leaf nodes correspond to the most basic elements at the bottom level of the element system. Their contribution values ​​cannot be obtained by aggregating child nodes, so a preset initial unit contribution value must be set. The setting of this initial value follows a unified quantification principle: usually, the initial contribution value of each leaf node is assigned to 1.0, representing the unit baseline contribution, or differentiated weights are assigned based on the business importance of the element, such as an initial value of 1.5 for key elements and 1.0 for ordinary elements.

[0153] This initial value is not an absolute measure of the actual contribution, but rather serves as a benchmark anchor for subsequent weighted aggregation. In stage S43, it is transformed into a relative contribution rate through a multiplication operation with the path edge weights. After all leaf nodes in the tree are initialized, a set of leaf node contribution values ​​{C} is formed. leaf_i This provides a starting point for the bottom-up recursive computation of S42.

[0154] S42. Initiate hierarchical aggregation calculation for non-leaf nodes, the logic of which strictly follows the directed edge weights and hierarchical topology of the destructuring tree. Specifically:

[0155] Starting from the bottom-most non-leaf node, that is, a node whose parent node is an intermediate-level node and whose child nodes are all leaf nodes, for each non-leaf node N... parent First, retrieve the set of all its direct child nodes in the deconstruction tree, and obtain the current contribution value C of each child node. child_j The initial value set for S41 in the first round of calculation, and the updated value passed up from the lower-level aggregation in subsequent iterations.

[0156] Simultaneously, the relative contribution weight w of each parent-child edge is extracted from the edge weight attribute of the deconstructed tree. child_j→parentThe weights were normalized during the generation in stage S23, ensuring that the sum of all weights is 1.

[0157] Aggregate calculations perform weighted summation operations: C parent = Σ(C child_j ×w child_j→parent This formula essentially weights and aggregates the contributions of child nodes according to their relative importance to the parent node, resulting in C. parent This serves as the input contribution value for the non-leaf node to its parent node. This process recursively ascends layer by layer. After each layer of calculation is completed, the contribution values ​​of all nodes in that layer are updated and passed to the next layer, until all non-leaf nodes are assigned the aggregated contribution value.

[0158] S43. During the aggregation calculation process, the hierarchical transfer of contribution rate and path accumulation calculation are performed simultaneously, specifically as follows:

[0159] S431. Calculation of path contribution rate from leaf node to root node:

[0160] For any given leaf node Leaf k In the final contribution deconstruction tree, starting from the leaf node, traverse upwards along the directed edges pointing to the parent node until the root node, i.e., the overall system efficiency node, is reached. The weight of each directed edge traversed during the traversal is recorded, forming the path weight sequence {w1, w2, ..., w...}. L}, where L is the number of layers in the path.

[0161] The initial contribution value C of the leaf node leaf_k The product of the weights of all edges along the path, i.e., C leaf_k ×w1,×w2×...×w L , is defined as the path contribution rate of the leaf node to the root node via this specific path.

[0162] Because a deconstruction tree contains intermediate converging nodes (i.e., multiple child nodes sharing the same parent node), a leaf node can indirectly influence the root node through multiple independent paths formed by different combinations of intermediate nodes. For example, leaf node A can contribute to root node E0 through two paths: A→B→E0 and A→C→D→E0. Therefore, it is necessary to exhaustively enumerate all possible simple paths (i.e., acyclic paths) and calculate the path contribution rate for each path, forming a set of path contribution rates {R}. path_m}

[0163] S432. Complete the synthesis of the overall contribution rate of the leaf nodes:

[0164] The same leaf node calculated by S431 k Contribution rate of all paths {R path_m Perform algebraic summation, i.e., Rtotal_k =ΣR path_m The sum represents the overall contribution rate of leaf node element k to the overall system effectiveness. Its value is between 0 and 1, reflecting the degree of global influence of this element on the top-level effectiveness through all direct or indirect transmission paths.

[0165] This calculation mechanism fully considers the cross-layer jump associations and multi-path convergence effects existing in the deconstruction tree, ensuring that the contribution rate calculation is neither repeated nor omitted, and realizing a precise quantitative mapping from local benchmarks to global performance.

[0166] S433. Synchronously update the overall contribution rate and direct contribution rate of non-leaf nodes:

[0167] After S42 completes the aggregation calculation for non-leaf nodes at a certain level, the direct contribution rate of that node to its direct parent node is explicitly defined by the edge weights of the edges connecting it to its child nodes, i.e., w. node→parent This value remains unchanged after the deconstruction tree is generated and is used directly in the aggregation calculation.

[0168] Meanwhile, the overall contribution rate of this node, that is, its total contribution R to the root node. total_node Equal to the combined contribution rate R of all its child nodes total_child_j Sum of: R total_node =ΣR total_child_j .

[0169] This recursive relationship ensures the bottom-up conservation of contribution rates, meaning the root node's overall contribution rate ultimately equals the sum of all leaf node contribution rates, satisfying the normalization constraint for all-factor contribution rates. This sub-step is executed immediately after each aggregation layer is completed, achieving simultaneous calculation of contribution values ​​and contribution rates through dual channels, avoiding repeated traversal.

[0170] S44. Repeat the aggregation and propagation operations of S42 and S43, progressing upwards from the bottom layer of the destructuring tree. The specific process is as follows:

[0171] After completing the first-level aggregation from the bottom leaf node to its direct parent node, the calculation result is passed to the intermediate node of the next higher level. S42 is performed on the intermediate node to obtain its aggregate contribution value, and S433 is performed to accumulate the comprehensive contribution rate of its child nodes to obtain its own comprehensive contribution rate. Then the updated contribution value and contribution rate are passed to the next higher level, and so on in a loop.

[0172] When the recursive calculation reaches the root node, the contribution value C of the root node is... root This is the aggregated predicted value of the overall system effectiveness, and its comprehensive contribution rate R total_root Under normalization conditions, the value is always equal to 1, representing 100% system efficiency. At this point, the algorithm terminates the recursion and simultaneously outputs the comprehensive contribution rate {R} of each leaf node to the root node.total_leaf_i This result set serves as the final calculation result. It represents a quantitative ranking of the contribution rate of each basic element in the element system to the overall effectiveness, and can be directly used for business decision-making, element optimization, and resource allocation.

[0173] As can be seen from the above description, the embodiments of the present invention achieve the following technical effects:

[0174] This invention achieves intelligent and precise identification and quantification of complex, nonlinear, and hidden contribution relationships, solving the obscure technical challenges of traditional methods that rely on subjective experience and are difficult to model. Traditional methods, such as the analytic hierarchy process (AHP) and expert scoring, heavily rely on prior human knowledge to define contribution relationships and weights between elements when analyzing a system of elements, failing to effectively handle indirect, nonlinear, and hidden interactions. This invention introduces a relational machine learning model, training it using the system's feature matrix and overall performance data, enabling it to automatically learn and extract complex correlation patterns between elements from historical data. This model can capture multi-level transmission and nonlinear superposition effects that traditional linear models cannot detect, such as team psychological safety indirectly affecting code quality through improved knowledge sharing. Based on these automatically learned correlation patterns, a contribution deconstruction tree is generated, achieving data-driven, automated, and precise quantification of such hidden contribution relationships. This fundamentally overcomes the inherent shortcomings of traditional methods in this field, making contribution analysis closer to the real operating mechanisms of complex systems.

[0175] A closed-loop intelligent system of learning, verification, and dynamic optimization is constructed to ensure the continuous reliability of the contribution model and its strong adaptability to system evolution. Traditional static weight allocation models, once established, are fixed and cannot be adjusted to changes in the system's internal and external environments, making them prone to failure. This invention designs a rigorous verification and dynamic optimization mechanism. This method not only generates an initial deconstruction tree but also objectively verifies the model's accuracy by selecting key verification nodes, masking their feature data, and comparing the model's predicted contribution values ​​with actual observations. When the deviation exceeds a tolerance threshold, the system automatically triggers model retraining and tree structure reconstruction, using the deviation information as feedback signals to optimize the model. This closed-loop mechanism transforms the entire analysis method from a one-time static snapshot into an intelligent system capable of dynamic development, continuous self-calibration, and evolution alongside the system. This significantly improves the timeliness, robustness, and long-term practical value of contribution rate analysis results, enabling it to effectively cope with the evolution of the factor system over time.

[0176] This invention provides structured analysis results that combine high accuracy and interpretability, greatly enhancing the depth and practicality of decision support. Many complex machine learning models, while highly accurate in prediction, are black boxes, making it difficult for decision-makers to understand their underlying logic. This invention cleverly solves this contradiction. Its final output is not a set of incomprehensible weighting coefficients or black-box predictions, but a clearly structured, multi-level contribution deconstruction tree. This tree intuitively reveals the value contribution chain, path, and intensity from bottom-level elements to top-level goals. Decision-makers can not only obtain a precisely quantified comprehensive contribution rate for each element, but also clearly understand how that contribution is generated through direct or indirect paths. For example, they can clearly see the contribution of toolchain proficiency to final performance, how much is direct, and how much is indirectly achieved through improving code quality and the clarity of task decomposition. This deep interpretability transforms the analysis results from simple rankings or scores into a powerful map for root cause analysis, bottleneck identification, and resource optimization, greatly enhancing the practical value of this method in strategic decision-making and operational management.

[0177] The embodiments and / or implementation methods described above are merely preferred embodiments and / or implementation methods for implementing the technology of the present invention, and are not intended to limit the implementation methods of the technology of the present invention in any way. Any person skilled in the art can make some modifications or alterations to other equivalent embodiments without departing from the scope of the technical means disclosed in the present invention, but these should still be regarded as the technology or embodiments that are substantially the same as the present invention.

[0178] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. The above descriptions are only preferred embodiments of this application. It should be noted that due to the limitations of written expression, while there are objectively infinite specific structures, those skilled in the art can make several improvements, modifications, or changes without departing from the principles of this application, and can also combine the above technical features in an appropriate manner. These improvements, modifications, changes, or combinations, or the direct application of the inventive concept and technical solution to other situations without modification, should all be considered within the scope of protection of this application.

Claims

1. A multi-level contribution deconstruction and aggregation method for an element system based on machine learning, characterized in that, include: Extract the element observation data, generate standardized element feature vectors, and construct the system feature matrix; Using the system feature matrix and overall performance data, a relational machine learning model is trained to automatically generate a multi-level contribution deconstruction tree; Nodes in the multi-level contribution deconstruction tree are selected for verification, and the final contribution deconstruction tree is obtained by adjusting the values ​​based on the deviation between the predicted and actual values. Based on the final contribution deconstruction tree, the contribution values ​​are weighted and aggregated from bottom to top, and the comprehensive contribution rate of each element is recursively calculated.

2. The multi-level contribution deconstruction and aggregation method for an element system based on machine learning according to claim 1, characterized in that, The extraction of element observation data, generating standardized element feature vectors, and constructing a system feature matrix specifically includes: For each element in the target element system, static attribute data and dynamic time-series status data are collected from multiple preset data sources; Each element of data collected is preprocessed, including at least missing value imputation, outlier detection and correction, and data standardization. Each element's preprocessed multidimensional data is spliced ​​and encoded according to predetermined rules to generate a fixed-dimensional, numerical element feature vector. The feature vectors of all elements are used as row vectors and arranged in the order of element index to construct a complete system feature matrix.

3. The multi-level contribution deconstruction and aggregation method for a feature system based on machine learning according to claim 1, characterized in that, Using the system feature matrix and overall performance data, a relational machine learning model is trained to automatically generate a multi-level contribution deconstruction tree, specifically including: Using the system feature matrix as input features and the overall system performance output data as supervision labels, the relational machine learning model is trained under supervision, enabling the model to learn the complex mapping relationship from element features to overall performance and the internal correlation patterns between element features. From the converged relational machine learning model, the contribution weights of each input feature dimension of each element to the predicted output are extracted as the initial correlation strength between elements and between elements and the overall effectiveness. Using the overall system efficiency as the root node, and based on the initial correlation strength, a threshold-based correlation screening and recursive hierarchical algorithm is used to organize elements with significant correlation strength into child nodes layer by layer, and assign corresponding weights to the directed edges between nodes, thereby automatically constructing a multi-level contribution deconstruction tree.

4. The multi-level contribution deconstruction and aggregation method for a feature system based on machine learning according to claim 3, characterized in that, The extraction of the contribution weights of each input feature dimension to the prediction output for each corresponding element specifically includes: The relational machine learning model adopts a gradient boosting decision tree. By analyzing the model feature coefficients, it directly quantifies the direct contribution weight of each feature to the final prediction result. The direct contribution weight serves as the main component of the initial correlation strength of the feature to the overall effectiveness. The direct contribution weights and interaction influence strengths are normalized and fused to form an initial association strength matrix, where the rows and columns of the matrix correspond to each element and the overall effectiveness, and the matrix element values ​​represent the association strength from one node to another.

5. The multi-level contribution deconstruction and aggregation method for a feature system based on machine learning according to claim 4, characterized in that, The specific steps for constructing the multi-level contribution deconstruction tree are as follows: The element values ​​in the initial association strength matrix are filtered by applying a dynamic hierarchical threshold, retaining associations with strength values ​​exceeding the threshold to form a significant association set; Starting with the overall system effectiveness node as the initial layer, select all associations with the current layer node as the target from the significant association set, and organize the source element nodes corresponding to the associations into the next layer of child nodes; The initial association strength values ​​retained after filtering are normalized to the relative contribution weight of child nodes to their parent nodes, and directed edges from child nodes to parent nodes are generated using these weight values. The newly generated child node layer is used as the current layer. The tree structure is recursively expanded downwards by repeating this operation until a certain layer can no longer match a new element node that has not been organized into the tree from the set of significant associations. Finally, a complete multi-level contribution deconstruction tree is generated.

6. The multi-level contribution deconstruction and aggregation method for a feature system based on machine learning according to claim 1, characterized in that, Nodes in the multi-level contribution deconstruction tree are selected for verification, and the tree is optimized based on the deviation between the predicted and actual values ​​to obtain the final contribution deconstruction tree, specifically including: Based on the topology of the multi-level contribution deconstruction tree and prior business knowledge, several nodes are selected as verification nodes. For each verification node, after masking the feature data of the corresponding element of the node in the system feature matrix, the data is input into the relational machine learning model to obtain the model prediction contribution value of the node and the actual observed contribution value of the node, and the deviation between the two is calculated. If the deviation of any verification node exceeds the preset tolerance threshold, the optimization process will be triggered. Repeat the operation until the deviation of all selected verification nodes is lower than the tolerance threshold. The resulting contribution deconstruction tree is the final contribution deconstruction tree.

7. The multi-level contribution deconstruction and aggregation method for a feature system based on machine learning according to claim 6, characterized in that, The process of obtaining the model-predicted contribution value of a node, acquiring the actual observed contribution value of the node, and calculating the deviation between the two specifically includes: For the selected verification node, a corresponding contribution evaluation dataset is created. The contribution evaluation dataset is generated by setting all the feature vectors of the elements corresponding to the nodes in the system feature matrix to zero or baseline values. The contribution evaluation dataset is input into a relational machine learning model. The model re-infers based on the feature data of the remaining elements and implicitly outputs a temporary local deconstruction tree for the current data state according to its internal correlation logic. Based on the structure and weights of the temporary local deconstruction tree, the change in the overall system performance prediction value caused by the occlusion of target elements is calculated, and the change is used as the model prediction contribution value of the verification node. The model-predicted contribution value of the validation node is compared with its actual observed contribution value measured independently, and the specific deviation value is calculated using the root mean square error formula.

8. The multi-level contribution deconstruction and aggregation method for a feature system based on machine learning according to claim 7, characterized in that, The optimization process specifically includes: The original feature vector of the verification node that generates the excess deviation, its occluded feature vector, and the calculated deviation value are encapsulated together into a high-priority training sample and injected into the original training dataset of the relational machine learning model. Using the expanded training dataset, the relational machine learning model is retrained by focusing on high-priority training samples with higher learning weights to correct the model's cognitive bias in the relationship of node contribution, thus obtaining a retrained model. The updated association patterns are extracted from the retrained model to generate a new initial association strength matrix. The matrix reflects the optimized association perception between elements and the elements' impact on overall effectiveness. Using the new initial association strength matrix as input, the tree structure generation step is re-executed to reconstruct the multi-level contribution deconstruction tree and generate an updated tree structure for the next round of verification.

9. The multi-level contribution deconstruction and aggregation method for a feature system based on machine learning according to claim 1, characterized in that, Based on the final contribution deconstruction tree, a weighted aggregation of contribution values ​​is performed from bottom to top, and the comprehensive contribution rate of each element is recursively calculated, specifically including: Set the contribution value of all leaf nodes in the final contribution deconstruction tree as the preset initial unit contribution value of their corresponding feature vectors. Starting from the bottom-level node, for any non-leaf node, multiply the contribution values ​​of all its direct child nodes by the weights of the directed edges between the node and each child node, and sum them up. The result is used as the aggregated contribution value of the non-leaf node. Contribution rates are calculated and transmitted synchronously during the aggregation process; The process of repeated aggregation calculation and contribution rate transfer calculation is carried out recursively upwards layer by layer until the contribution value of the root node is calculated, and the comprehensive contribution rate of each leaf node to the root node is obtained simultaneously as the final output.

10. The multi-level contribution deconstruction and aggregation method for a feature system based on machine learning according to claim 9, characterized in that, The contribution rate is calculated and transmitted synchronously during the aggregation process, specifically including: For any leaf node, in the final contribution deconstruction tree, starting from the leaf node, traverse upwards along the direction pointing to the parent node to the root node, recording the weight of each directed edge traversed; the initial contribution value of the leaf node is multiplied by the weights of all edges on the traversal path, which is taken as the path contribution rate of the leaf node to the root node via a specific path. The sum of all path contribution rates calculated for the same leaf node is the comprehensive contribution rate of the leaf node element to the overall system effectiveness. After each layer of aggregation calculation is completed, the contribution value of non-leaf nodes is updated. The overall contribution rate of a node to the root node is equal to the sum of the overall contribution rates of all its child nodes. The direct contribution rate of a node to its direct parent node is defined by the weight of the corresponding edge when the tree is generated and is used directly in the aggregation calculation.

Images