CN-122020594-A - Element system multi-level contribution deconstructing and aggregation method based on machine learning

CN122020594ACN 122020594 ACN122020594 ACN 122020594ACN-122020594-A

Abstract

The invention relates to the field of data analysis and intelligent decision, in particular to a machine learning-based element system multi-level contribution deconstructing and aggregation method, which comprises the steps of extracting element observation data, generating standardized element feature vectors and forming a system feature matrix; the method comprises the steps of training a relational machine learning model by utilizing a system characteristic matrix and overall efficiency data, automatically generating a multi-level contribution solution tree, selecting nodes in the contribution solution tree for verification, dynamically adjusting the model and the tree structure according to the deviation between a predicted value and an actual value to obtain a final contribution solution tree, carrying out weighted aggregation of the contribution value from bottom to top based on the final contribution solution tree, and recursively calculating the comprehensive contribution rate of each element. According to the invention, the complex contribution relation is automatically and objectively deconstructed through machine learning, the accuracy and the adaptability of the model are ensured by means of verification closed loop, and the quantitative contribution rate with clear structure and strong interpretability is finally output, so that the reliability and the decision support value of a calculation result are obviously improved.

Inventors

DING LIANGLIANG
ZHOU XU
LU LIANCHENG
CUI XIAOXIAO
XUE XIAOGUANG

Assignees

中国人民解放军63921部队

Dates

Publication Date: 20260512
Application Date: 20260203

Claims (10)

1. The element system multi-level contribution deconstructing and aggregating method based on machine learning is characterized by comprising the following steps of: extracting element observation data, generating standardized element feature vectors, and forming a system feature matrix; Training a relational machine learning model by utilizing the system feature matrix and the overall efficiency data, and automatically generating a multi-level contribution solution tree; selecting nodes in the multi-level contribution solution tree for verification, and optimizing according to the deviation between the predicted value and the actual value to obtain a final contribution solution tree; And carrying out weighted aggregation of contribution values from bottom to top based on the final contribution deconstructing tree, and recursively calculating the comprehensive contribution rate of each element.
2. The machine learning-based element system multi-level contribution deconstructing and aggregating method of claim 1, wherein the extracting element observation data generates normalized element feature vectors, and forms a system feature matrix, specifically comprising: for each element in the target element system, collecting static attribute data and dynamic time sequence state data of the element from a plurality of preset data sources; preprocessing each piece of collected element data, wherein the preprocessing at least comprises filling of a missing value, detection and correction of an abnormal value and data standardization; Splicing and encoding the preprocessed multidimensional data of each element according to a preset rule to generate a fixed-dimension and numeric element feature vector; and taking element feature vectors of all elements as row vectors, and arranging according to element index sequences to construct a complete system feature matrix.
3. The machine learning-based element system multi-level contribution deconstructing and aggregating method of claim 1, wherein the training of the relational machine learning model with the system feature matrix and the overall performance data automatically generates a multi-level contribution deconstructing tree, comprising: taking the system feature matrix as an input feature, taking system overall efficiency output data as a supervision tag, and performing supervision training on a relational machine learning model to enable the model to learn a complex mapping relation from element features to overall efficiency and an internal association mode between the element features; Extracting contribution degree weight of each input characteristic dimension corresponding to each element to prediction output from the relation type machine learning model after training convergence as initial association strength between the elements and the overall efficiency of the elements; And taking the overall efficiency of the system as a root node, adopting a correlation screening and recursion layering algorithm based on a threshold value according to the initial correlation strength, organizing elements with obvious correlation strength into child nodes layer by layer, and giving corresponding weights to directed edges among the nodes, so as to automatically construct a multi-level contribution deconstructed tree.
4. The machine learning-based element system multi-level contribution deconstructing and aggregating method of claim 3, wherein the extracting the contribution degree weight of each input feature dimension corresponding to each element to the predicted output specifically comprises: the relation type machine learning model adopts a gradient lifting decision tree, and the direct contribution weight of each element feature to the final prediction result is directly quantized through analyzing the model feature coefficient and is taken as a main component part of the initial association strength of the element to the overall efficiency; And normalizing and fusing the direct contribution weight and the interaction influence intensity to form an initial association intensity matrix, wherein the rows and columns of the matrix correspond to each element and the overall efficiency, and the matrix element values represent the association intensity from one node to the other node.
5. The machine learning-based element system multi-level contribution deconstructing and aggregating method of claim 4, wherein the constructing of the multi-level contribution deconstructing tree specifically comprises the following steps: Screening element values in the initial association strength matrix by using a dynamic layering threshold value, and reserving association of strength values exceeding the threshold value to form a significant association set; Taking the system overall efficiency node as an initial layer, selecting all associations with the current layer node as a target from the obvious association set, and organizing the source element nodes corresponding to the associations as sub-nodes of the next layer; Normalizing the initial association strength value reserved after screening into the relative contribution weight of the child node to the father node, and generating a directed edge pointing to the father node from the child node according to the weight value; And taking the newly generated sub-node layer as the current layer, repeating the operation, recursively expanding the tree structure downwards until a certain layer cannot be matched with a new element node which is not organized into a tree from the obvious association set, and finally generating a complete multi-level contribution deconstructing tree.
6. The machine learning-based element system multi-level contribution deconstructing and aggregating method of claim 1, wherein nodes in the multi-level contribution deconstructing tree are selected for verification, and the final contribution deconstructing tree is obtained according to deviation of a predicted value and an actual value, and the method specifically comprises: Selecting a plurality of nodes as verification nodes according to the topological structure and the service priori knowledge of the multi-level contribution deconstructed tree; For each verification node, after feature data of corresponding elements of the nodes are shielded in the system feature matrix, inputting the relational machine learning model to obtain model prediction contribution values of the nodes, obtaining actual observation contribution values of the nodes, and calculating deviation between the model prediction contribution values and the actual observation contribution values of the nodes; If the deviation of any verification node exceeds a preset tolerance threshold, triggering a tuning flow; Repeating the operation until the deviation of all the selected verification nodes is lower than the tolerance threshold, and obtaining the contribution solution tree at the moment, namely the final contribution solution tree.
7. The machine learning-based element system multi-level contribution deconstructing and aggregating method of claim 6, wherein the obtaining the model predicted contribution value of the node and obtaining the actual observed contribution value of the node, calculating the deviation between the two, specifically includes: for the selected verification node, creating a corresponding contribution evaluation data set, wherein the contribution evaluation data set is generated by setting element feature vectors of corresponding elements of nodes in the system feature matrix to zero values or reference values; Inputting the contribution evaluation data set into a relational machine learning model, wherein the model carries out reasoning again based on the characteristic data of the residual elements and implicitly outputs a temporary local deconstructing tree aiming at the current data state according to the internal association logic; Calculating the variation of the system overall efficiency predicted value caused by the shielding target element according to the structure and the weight of the temporary local deconstructed tree, wherein the variation is taken as the model predicted contribution value of the verification node; Comparing the model prediction contribution value of the verification node with the actual observation contribution value measured in an independent mode, and calculating a specific deviation value by adopting a root mean square error formula.
8. The machine learning-based element system multi-level contribution deconstructing and aggregating method of claim 7, wherein the tuning flow specifically comprises: Packaging the original element feature vector corresponding to the verification node generating the overrun deviation, the feature vector which is shielded and the calculated deviation value together into a high-priority training sample, and injecting the high-priority training sample into an original training data set of the relational machine learning model; using the expanded training data set, paying attention to a high-priority training sample with higher learning weight, and retraining the relational machine learning model to correct the cognitive deviation of the model on the node contribution relationship so as to obtain a retraining model; Extracting an updated association mode from the retraining model, generating a new initial association strength matrix, wherein the matrix reflects the adjusted inter-element association cognition and the association cognition of the elements on the overall efficiency; And re-executing the tree structure generation step by taking the new initial association strength matrix as input, so as to reconstruct the structure of the multi-level contribution deconstructed tree, and generating an updated tree structure for the next round of verification.
9. The machine learning-based element system multi-level contribution deconstructing and aggregating method of claim 1, wherein the performing the weighted aggregation of the contribution values from bottom to top based on the final contribution deconstructing tree recursively calculates the comprehensive contribution rate of each element, and specifically comprises: Setting the contribution values of all leaf nodes in the final contribution deconstructed tree as the preset initial unit contribution values of the corresponding element feature vectors; Starting from the bottommost node, for any non-leaf node, multiplying the contribution values of all direct child nodes by the directed edge weights between the node and each child node respectively, and summing, wherein the obtained result is used as the aggregated contribution value of the non-leaf node; synchronously calculating and transmitting contribution rate in the polymerization process; repeating the aggregation calculation and contribution rate transfer calculation processes, recursively calculating layer by layer upwards until the contribution value of the root node is calculated, and synchronously obtaining the comprehensive contribution rate of each leaf node to the root node as a final output.
10. The machine learning-based element system multi-level contribution deconstructing and aggregating method of claim 9, wherein the aggregating process synchronously calculates and transmits the contribution rate, specifically comprising: For any leaf node, traversing upwards from the leaf node to a root node along the direction pointing to a father node in the final contribution deconstructed tree, and recording the weight of each passing directed edge; taking the product of the initial contribution value of the leaf node and all edge weights on the traversal path as the path contribution rate of the leaf node to the root node through the specific path; summing all path contribution rates calculated by the same leaf node, wherein the sum is the comprehensive contribution rate of leaf node elements to the overall efficiency of the system; After the aggregation calculation of each layer is completed, the contribution value of the non-leaf node is updated, the comprehensive contribution rate of the node to the root node is equal to the sum of the comprehensive contribution rates of all the child nodes, and the direct contribution rate of the node to the 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.

Description

Element system multi-level contribution deconstructing and aggregation method based on machine learning Technical Field The invention relates to the technical field of data analysis and intelligent decision making, in particular to a machine learning-based element system multi-level contribution deconstructing and aggregation method. Background The scientific system contribution rate calculation is a key for performance evaluation, resource optimization and bottleneck diagnosis on complex element systems such as technical architecture, organizational capacity and ecosystem. Such computing has wide needs in many fields such as social network impact analysis, environmental system modeling, engineering structure optimization, and the like. The traditional system contribution rate calculation method mainly depends on the thought of weighting and aggregation after hierarchical decomposition. However, this method has inherent drawbacks in that, firstly, the hierarchical structure is built and the element weights are assigned extremely depending on expert experiences, the process is highly subjective and the standards are difficult to unify, and secondly, the complex interaction is simplified into linear superposition, and the contribution relationship of non-linearity, indirection and secrecy which are commonly existed among the elements cannot be described, for example, the team atmosphere indirectly influences the final output through various intermediate paths. This results in analysis results that are often not authoritative enough to be convincing. Attempts have been made in the art to introduce artificial intelligence models for improvement, such as using reinforcement learning models for system evaluation. However, such methods rely heavily on high quality training data sets and simulation environments that can accurately simulate real feedback, and in many practical application scenarios that lack complete historical data or cannot perform controllable experiments, it is often difficult to land due to data starvation or environmental mismatch, and the universality is poor. The core contradiction in the current technical field is that on one hand, the traditional model is too simple and subjective to be distorted, and on the other hand, the advanced data driving model cannot be applied due to the severe requirements on training data. The fundamental problem of the prior art is that a generic method which not only can automatically learn complex contribution relations from limited and mixed real data, but also has an internal verification and tuning mechanism to ensure the reliability of results cannot be effectively constructed. This results in a long-term ambiguity in the calculation of the contribution rates between the elements, and a lack of firm reliability in the calculation of the system contribution rates. Therefore, a method for multi-level contribution deconstructing and aggregation of element systems based on machine learning is needed to solve the above problems. Disclosure of Invention The invention aims to provide a machine learning-based element system multi-level contribution deconstructing and aggregating method, which aims to solve the problems of strong subjectivity and poor universality of the existing data driving model in the prior art. In order to solve the technical problems, the invention specifically provides the following technical scheme: The element system multi-level contribution deconstructing and aggregating method based on machine learning comprises the following steps: s1, extracting element observation data, generating standardized element feature vectors, and forming a system feature matrix; s2, training a relational machine learning model by utilizing the system feature matrix and the overall efficiency data, and automatically generating a multi-level contribution deconstructing tree; S3, selecting nodes in the multi-level contribution solution tree for verification, and optimizing according to the deviation between the predicted value and the actual value to obtain a final contribution solution tree; And S4, carrying out weighted aggregation on the contribution values from bottom to top based on the final contribution deconstructed tree, and recursively calculating the comprehensive contribution rate of each element. As a preferred embodiment of the present invention, the S1 specifically includes: S11, collecting static attribute data and dynamic time sequence state data of each element in a target element system from a plurality of preset data sources; S12, preprocessing each acquired element data, wherein the preprocessing at least comprises filling of a missing value, detection and correction of an abnormal value and data standardization, and dimension difference is eliminated; s13, splicing and encoding the preprocessed multidimensional data of each element according to a preset rule to generate a fixed-dimension and numerical element feature vector; S14, taking el