Attribution analysis method and device, equipment, storage medium and computer program product

Abstract

The invention discloses an attribution analysis method and device, equipment, a storage medium and a computer program product, and relates to the technical field of data analysis, and the method comprises the steps: carrying out the causal structure learning of a target data set, obtaining a target causal relationship in the target data set, and constructing a causal knowledge graph based on the target causal relationship; in response to a selection instruction of a main analysis index of a target user, querying a node set having causal association with the main analysis index from the causal knowledge graph, and determining a recommended association node based on causal strength; performing intervention effect calculation on the main analysis index based on the main analysis index and the recommended association node to generate anti-fact deduction result data; and performing multi-dimensional confidence assessment on the anti-fact deduction result data to obtain an attribution path credibility score. And through causal structure learning and multi-dimensional confidence assessment, the problem that attribution analysis depends on experience hypothesis is solved, and the analysis reliability is improved.

Description

Technical Field

[0001] This application relates to the field of data analysis technology, and in particular to attribution analysis methods, apparatus, devices, storage media, and computer program products. Background Technology

[0002] As enterprise data scales, attribution analysis has become a key tool for understanding operational issues. Current attribution analysis tools generally adopt a "pre-defined, fixed path" model. This means that technical personnel or business experts pre-define fixed analytical dimensions and attribution paths, binding business metrics and related factors to the analytical framework through hard-coding. Users can only select from a limited set of pre-defined dimensions and drill down along a predetermined path. However, this pre-defined, customized analytical framework cannot flexibly expand with changes in business. When business scenarios evolve or users need to explore new potential factors, it's impossible to dynamically add new analytical dimensions or adjust attribution paths, resulting in an analytical perspective that remains fixed within the initially set scope, making it difficult to quickly respond to business innovation needs. This limitation prevents users from incorporating the variables they are interested in into the analytical system when facing emerging business scenarios or needing to verify non-pre-defined factors. They can only passively accept the limited options built into the framework, severely restricting the flexibility and depth of attribution analysis and making it difficult to uncover the true driving factors behind business performance.

[0003] The above content is only used to help understand the technical solution of this application and does not represent an admission that the above content is prior art. Summary of the Invention

[0004] The main objective of this application is to provide an attribution analysis method, apparatus, device, storage medium, and computer program product, which aims to solve the technical problem that attribution analysis relies on empirical assumptions.

[0005] To achieve the above objectives, this application proposes an attribution analysis method, which includes: Perform causal structure learning on the target dataset to obtain the target causal relationships in the target dataset, and construct a causal knowledge graph based on the target causal relationships; In response to the target user's selection instruction for the main analytical indicator, the set of nodes that have a causal relationship with the main analytical indicator is queried from the causal knowledge graph, and the recommended associated nodes are determined based on the causal strength. Based on the main analysis index and the recommended related nodes, the intervention effect of the main analysis index is calculated to generate counterfactual inference result data; A multi-dimensional confidence assessment is performed on the counterfactual inference results data to obtain the attribution path credibility score.

[0006] In one embodiment, the step of performing causal structure learning on the target dataset to obtain target causal relationships in the target dataset, and constructing a causal knowledge graph based on the target causal relationships includes: Construct a business constraint rule base, which includes causal prohibition rules and causal enforcement rules defined by business logic; Perform time-series stationarity detection and segmentation on the target dataset to generate a target data subset; The target data subset is subjected to conditional independence testing using a causal structure learning algorithm to generate an undirected initial causal skeleton. The initial causal framework is pruned based on the causal prohibition rule to remove pseudo-causal relationships that do not conform to business logic. Specific causal relationships are added to the initial causal framework based on the causal enforcement rule to obtain a target causal network with causal direction. The causal knowledge graph is then constructed based on the target causal network.

[0007] In one embodiment, the step of responding to the selection instruction of the target user's main analytics metric further includes: Based on the main analysis indicators, determine the corresponding main analysis data in the target dataset; The master analysis data is filtered based on preset dynamic query conditions to obtain a target scene dataset. The target scene dataset is used to calculate the marginal explanatory power increment of the target node in the set of associated nodes in the causal knowledge graph. The dynamic query conditions include institutional screening conditions, line screening conditions, and time range conditions.

[0008] In one embodiment, the step of querying the set of nodes in the causal knowledge graph that have a causal relationship with the main analysis indicator, and determining the recommended associated nodes based on the causal strength, includes: Extract the set of associated nodes connected to the main analysis index from the causal knowledge graph, and obtain the causal strength weight and user adoption rate of each causal edge in the set of associated nodes; The marginal explanatory power increment of each target node in the set of associated nodes is calculated based on the target scenario dataset, and a recommendation score is generated based on the marginal explanatory power increment, the causal strength weight, and the user adoption rate. The nodes in the associated node set are sorted based on the recommendation score, and the recommended associated nodes in the associated node set are determined based on the sorting results.

[0009] In one embodiment, the step of performing a multi-dimensional confidence assessment on the counterfactual inference result data to obtain an attribution path credibility score includes: A multi-dimensional confidence assessment is performed on the counterfactual inference results data, and the attribution analysis path of the main analysis indicator is determined based on the assessment results; Obtain the data support, effect consistency bias, and user adoption rate for each causal edge in the attribution analysis path; The attribution path credibility score of the attribution analysis path is obtained by weighting the data support, the effect consistency deviation value, and the user adoption rate.

[0010] In one embodiment, the step of performing a multi-dimensional confidence assessment on the counterfactual inference result data and determining the attribution analysis path of the main analysis indicator based on the assessment results includes: The comprehensive confidence of each causal edge in the causal knowledge graph is calculated based on the data support, the effect consistency deviation value, and the user adoption rate. The recommended associated nodes are filtered, and nodes with a comprehensive confidence score greater than a preset confidence score threshold are retained as valid attribution nodes. A set of candidate attribution paths is generated based on the causal direction relationship between the valid attribution nodes. The candidate attribution paths contain at least one causal chain from the recommended associated nodes to the main analysis index. The candidate attribution paths in the candidate attribution path set are weighted and summed to obtain the total confidence of the candidate attribution paths. The candidate attribution path with the highest total confidence is selected as the attribution analysis path of the main analysis indicator.

[0011] Furthermore, to achieve the above objectives, this application also proposes an attribution analysis apparatus, the attribution analysis apparatus comprising: The graph construction module is used to perform causal structure learning on the target dataset, obtain the target causal relationships in the target dataset, and construct a causal knowledge graph based on the target causal relationships. The recommendation node module is used to respond to the selection instruction of the target user's main analysis indicator, query the set of nodes that have a causal relationship with the main analysis indicator from the causal knowledge graph, and determine the recommended associated nodes based on the causal strength. The calculation and inference module is used to calculate the intervention effect of the main analysis indicator based on the main analysis indicator and the recommended related nodes, and generate counterfactual inference result data. The confidence assessment module is used to perform multi-dimensional confidence assessment on the counterfactual inference results data to obtain an attribution path credibility score.

[0012] In addition, to achieve the above objectives, this application also proposes an attribution analysis device, the device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the attribution analysis method as described above.

[0013] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable storage medium, on which a computer program is stored, and which, when executed by a processor, implements the steps of the attribution analysis method as described above.

[0014] In addition, to achieve the above objectives, this application also provides a computer program product, which includes a computer program that, when executed by a processor, implements the steps of the attribution analysis method described above.

[0015] One or more technical solutions proposed in this application have at least the following technical effects: This application proposes an attribution analysis method, apparatus, device, storage medium, and computer program product. It performs causal structure learning on a target dataset to obtain target causal relationships within the dataset and constructs a causal knowledge graph based on these relationships. Responding to a target user's selection instruction for a primary analysis indicator, it queries the causal knowledge graph for a set of nodes causally associated with the primary analysis indicator and determines recommended associated nodes based on causal strength. Based on the primary analysis indicator and the recommended associated nodes, it calculates the intervention effect of the primary analysis indicator, generating counterfactual inference result data. Finally, it performs multi-dimensional confidence assessment on the counterfactual inference result data to obtain an attribution path credibility score. By employing causal structure learning and multi-dimensional confidence assessment, it solves the problem of attribution analysis relying on empirical assumptions, thereby improving the reliability of the analysis. Attached Figure Description

[0016] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating an embodiment of the attribution analysis method of this application. Figure 2This is a schematic diagram of the module structure of the attribution analysis device according to an embodiment of this application; Figure 3 This is a schematic diagram of the device structure of the hardware operating environment involved in the attribution analysis method in the embodiments of this application.

[0019] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0020] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0021] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0022] The main solution of this application embodiment is as follows: causal structure learning is performed on the target dataset to obtain the target causal relationships in the target dataset, and a causal knowledge graph is constructed based on the target causal relationships; in response to the target user's selection instruction for the main analysis indicator, a set of nodes that are causally related to the main analysis indicator is queried from the causal knowledge graph, and recommended associated nodes are determined based on the causal strength; based on the main analysis indicator and the recommended associated nodes, the intervention effect of the main analysis indicator is calculated to generate counterfactual inference result data; and multi-dimensional confidence assessment is performed on the counterfactual inference result data to obtain an attribution path credibility score.

[0023] In this embodiment, for ease of description, the attribution analysis system will be used as the execution subject in the following description.

[0024] As enterprise data scales, attribution analysis has become a key tool for understanding operational issues. Current attribution analysis tools generally adopt a "pre-defined, fixed path" model. This means that technical personnel or business experts pre-define fixed analytical dimensions and attribution paths, binding business metrics and related factors to the analytical framework through hard-coding. Users can only select from a limited set of pre-defined dimensions and drill down along a predetermined path. However, this pre-defined, customized analytical framework cannot flexibly expand with changes in business. When business scenarios evolve or users need to explore new potential factors, it's impossible to dynamically add new analytical dimensions or adjust attribution paths, resulting in an analytical perspective that remains fixed within the initially set scope, making it difficult to quickly respond to business innovation needs. This limitation prevents users from incorporating the variables they are interested in into the analytical system when facing emerging business scenarios or needing to verify non-pre-defined factors. They can only passively accept the limited options built into the framework, severely restricting the flexibility and depth of attribution analysis and making it difficult to uncover the true driving factors behind business performance.

[0025] This application provides a solution that addresses the problem of attribution analysis relying on empirical assumptions by using causal structure learning and multi-dimensional confidence assessment, thereby improving the reliability of the analysis.

[0026] It should be noted that the executing entity in this embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone, or an electronic device or attribution analysis system capable of performing the above functions. The following description uses an attribution analysis system as an example to illustrate this embodiment and the subsequent embodiments.

[0027] Based on this, embodiments of this application provide an attribution analysis method, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the attribution analysis method of this application.

[0028] In this embodiment, the attribution analysis method includes steps S11 to S14: Step S11: Perform causal structure learning on the target dataset to obtain the target causal relationships in the target dataset, and construct a causal knowledge graph based on the target causal relationships.

[0029] It should be noted that the target dataset refers to a collection containing relevant attributes and data records of the object to be analyzed (such as customer transaction data and risk indicator data in financial business). Its role is to provide raw data support for causal relationship mining. Causal structure learning is a technique that uses statistical algorithms and logical reasoning to mine causal relationships between variables in a dataset. The core of it adopts constraint-based causal discovery algorithms, such as the PC algorithm (Peter-Clark algorithm), score-based search algorithms, such as the BIC scoring algorithm (Bayesian Information Criterion Scoring Algorithm), or hybrid algorithms. Its role is to identify potential causal relationships from unlabeled data.

[0030] Additionally, it should be noted that the target causal relationship refers to the association that truly reflects the causal effect between variables after learning and verification (such as the positive causal association between "number of customer delinquencies" and "loan default rate"). This relationship is the core logical foundation for constructing causal analysis.

[0031] Additionally, it should be noted that a causal knowledge graph is a knowledge carrier that visualizes target causal relationships in the form of nodes and edges (nodes represent indicator variables, and edges represent causal relationships and directions). Its function is to intuitively store and display causal logic, providing structured support for subsequent queries and inferences.

[0032] Understandably, the purpose of this step is to address the problems of traditional attribution analysis, which relies on empirical assumptions and lacks data-driven causal logic. The principle is to learn causal structures to avoid being misled by correlations between variables, uncover the true causal relationships hidden in the data, and then solidify the causal relationships through the structured form of knowledge graphs. This provides a rigorous causal logic foundation for subsequent attribution analysis, avoids attribution bias caused by spurious correlations, and improves the scientificity and reliability of attribution analysis.

[0033] Specifically, in one embodiment of this application, a constraint-based PC algorithm is used to perform conditional independence tests on the target dataset, gradually eliminating variable pairs without causal relationships, and determining the causal skeleton and direction; in another embodiment of this application, a score-based BIC algorithm is used to select the optimal causal structure by traversing all possible causal structures and calculating scores; finally, the causal structure obtained by the pure algorithm is corrected by combining prior business knowledge (such as eliminating causal edges that obviously do not conform to business logic).

[0034] For example, suppose the target dataset is personal loan business data from a bank, containing 100,000 records across 20 variables, including customer age, income level, credit history, loan amount, and overdue days. The execution entity uses a PC algorithm to learn the causal structure of this dataset. By using conditional independence tests to eliminate variable pairs with no causal relationship, such as "customer age" and "loan amount," it identifies 15 target causal relationships, such as "income level - loan amount," "credit history - overdue days," and "overdue days - loan default rate." These causal relationships are then used to construct a causal knowledge graph in the form of nodes (variable names) and directed edges (causal directions), with the directed edges labeled as "positive," "negative," etc., indicating the causal action type.

[0035] Step S12: In response to the target user's instruction to select the main analysis indicator, query the set of nodes that have a causal relationship with the main analysis indicator from the causal knowledge graph, and determine the recommended associated nodes based on the causal strength.

[0036] It should be noted that the main analysis indicator selection instruction refers to the operation signal corresponding to the core attribution indicator (such as loan default rate, customer churn rate, etc.) selected by the target user through the interactive interface. Its function is to clarify the core object of attribution analysis.

[0037] Additionally, it should be noted that the set of nodes in the causal knowledge graph refers to all nodes in the causal knowledge graph that have a direct or indirect causal relationship with the main analytical indicator (e.g., when the main analytical indicator is the loan default rate, the set of nodes includes overdue days, credit records, income level, etc.), and its function is to provide the scope of potential influencing factors for attribution analysis.

[0038] Additionally, it should be noted that causal strength refers to a quantitative indicator (such as calculated through mutual information values, regression coefficients, etc.) of the degree of influence of the dependent variable on the independent variable in a causal relationship. Its function is to screen nodes that have a significant impact on the main analysis indicator. Recommended related nodes refer to nodes selected from the node set that have a significant causal impact on the main analysis indicator. Their function is to focus on key attribution factors and improve the efficiency of subsequent analysis.

[0039] Understandably, the purpose of this step is to address the problem of blind and untargeted selection of influencing factors in traditional attribution analysis. Its principle is based on the constructed causal knowledge graph to accurately locate potential influencing factors related to the main analysis indicator, and then to select key factors through quantitative assessment of causal strength, avoiding interference from irrelevant variables, ensuring that subsequent attribution analysis focuses on core influencing factors, and improving the relevance and effectiveness of attribution results.

[0040] Specifically, by using depth-first search or breadth-first search, we query multi-level neighbor nodes that are indirectly connected to the main analysis index to form a node set (the range of nodes can be controlled by setting a search depth threshold); then, we use methods such as mutual information calculation and Pearson correlation coefficient correction to quantify the causal strength of each causal edge; we set a causal strength threshold and select nodes with causal strength higher than the threshold as recommended associated nodes.

[0041] For example, a target user selects "loan default rate" as the main analysis indicator through an interactive interface. After the executing entity responds to this selection instruction, it uses a breadth-first search (setting the search depth to 2) to query the causal knowledge graph, obtaining a set of nodes containing 10 nodes such as overdue days, credit history, income level, debt-to-income ratio, and loan term. The causal strength of each causal edge is calculated through mutual information (e.g., the causal strength of "overdue days - loan default rate" is 0.85, and the causal strength of "loan term - loan default rate" is 0.32). The causal strength threshold is set to 0.5, and four nodes—overdue days, credit history, income level, and debt-to-income ratio—are selected. Then, the nodes are sorted based on their business importance score ("credit history" has the highest business importance score), and these four nodes are finally determined as recommended related nodes.

[0042] Step S13: Based on the main analysis index and the recommended associated nodes, calculate the intervention effect of the main analysis index and generate counterfactual inference result data.

[0043] It should be noted that intervention effect calculation refers to calculating the change in the main analytical indicator under the intervention by setting a virtual intervention value for the recommended related nodes (such as adjusting the "overdue days" from 30 days to 10 days). The core technologies include do calculus and causal effect estimation (such as ATE (Average Treatment Effect) and CATE (Conditional Average Treatment Effect)). Its function is to quantify the degree of causal influence of the recommended related nodes on the main analytical indicator. The calculation is performed through a counterfactual simulation engine, which is a software module with counterfactual reasoning and intervention effect calculation capabilities. Its core uses structural equation modeling, potential outcome framework and other technologies to simulate intervention behavior on the recommended related nodes and calculate its impact on the main analytical indicator.

[0044] Additionally, it should be noted that counterfactual analysis results data refers to datasets containing information such as intervention scenarios, intervention variables, changes in main analytical indicators, and causal effect values. Their purpose is to provide quantitative evidence for subsequent confidence assessments and attribution analyses.

[0045] Understandably, the purpose of this step is to address the problems in traditional attribution analysis, such as the inability to quantify the independent influence of a single factor and the lack of counterfactual verification. Its principle is based on the rigorous causal logic of causal knowledge graphs. By constructing virtual intervention scenarios through a counterfactual simulation engine, the confounding effects of other variables are eliminated, and the independent causal effect of each recommended association node on the main analysis indicator is accurately calculated. This upgrades attribution analysis from qualitative association to quantitative influence, improving the accuracy and persuasiveness of the attribution results.

[0046] Specifically, based on the structural equation model of the causal knowledge graph, a causal model (such as linear structural equation and nonlinear structural equation) is constructed between the main analysis indicator and the recommendation association node. For discrete recommendation association nodes (such as credit records divided into "good", "average", and "poor"), the do-of-fact calculus is used to set intervention values ​​(such as changing "poor" to "good"), and the change in the main analysis indicator is calculated. For continuous recommendation association nodes (such as income level and debt-to-income ratio), the average treatment effect calculation method is used to quantify the impact of unit variable changes on the main analysis indicator. The counterfactual simulation engine is called to process the intervention calculation of multiple recommendation association nodes in parallel, generating counterfactual inference results data containing the intervention effect values ​​of each node. Outlier detection is performed on the intervention effect calculation results to remove unreasonable inference data.

[0047] Furthermore, when the recommended association node contains discrete dimensional fields, an intervention simulation sample for the discrete dimensional field is generated using an adversarial generative network; when the recommended association node contains continuous indicator fields, a do-calculus sample for the continuous indicator field is generated based on a structural equation model of the causal knowledge graph; and an effect propagation deduction is performed on the intervention simulation sample and / or do-calculus sample along the topological order of the causal knowledge graph using a counterfactual simulation engine to calculate the causal effect value of the recommended association node on the main analysis indicator, generating counterfactual deduction result data containing the causal effect value.

[0048] For example, the main analytical indicator is "loan default rate", and the recommended associated nodes include the discrete node "credit history" (values: good, average, poor) and the continuous node "debt-to-income ratio" (value range: 0-1). A linear structural equation is constructed based on a causal knowledge graph: Loan default rate = 0.2 × Credit history (coded: Good = 0, Average = 1, Poor = 2) + 0.5 × Debt-to-income ratio + 0.03 (constant term); the counterfactual simulation engine is used to intervene in "Credit history" (changing "Poor" to "Good"), resulting in a decrease of 0.4 in the loan default rate; the "Debt-to-income ratio" is intervened in (decreased from 0.8 to 0.5), resulting in a decrease of 0.15 in the loan default rate calculated through the average treatment effect; finally, counterfactual inference results data are generated, including the intervention scenario (credit history optimization, debt-to-income ratio reduction), intervention variables, changes in default rate, and causal effect values ​​(0.4, 0.15), and abnormal inference values ​​are removed according to preset principles (e.g., due to intervention variable values ​​exceeding the reasonable range).

[0049] Step S14: Perform a multi-dimensional confidence assessment on the counterfactual inference result data to obtain an attribution path credibility score.

[0050] It should be noted that the attribution path credibility score refers to a score (e.g., 0-1 points) quantified through multi-dimensional evaluation results, reflecting the reliability of the attribution path. Its purpose is to provide target users with a reference for the credibility of the attribution results. Multi-dimensional confidence assessment refers to the process of verifying counterfactual inference results from multiple dimensions such as data reliability, logical consistency, and business adaptability. It primarily employs techniques such as data support analysis, effect consistency testing, and business rationality verification, and its purpose is to comprehensively evaluate the credibility of the inference results.

[0051] Understandably, the purpose of this step is to address potential biases in counterfactual reasoning results and ensure the reliability of attribution analysis results. The principle is to verify the data support strength, logical consistency, and business adaptability of the reasoning results through multi-dimensional confidence assessment, eliminate unreliable reasoning results and attribution paths, and then use quantitative scoring to intuitively reflect the credibility of the attribution path, avoiding the one-sidedness caused by single-dimensional assessment, and improving the reliability of attribution results and user trust.

[0052] Specifically, the data support of the counterfactual inference results is calculated (e.g., the proportion of samples supporting the result to the total number of samples) to assess the strength of the data support; consistency checks are performed on the results of multiple interventions for the same recommendation-related node (e.g., calculating the coefficient of variation) to assess the consistency of the effect; the rationality of the inference results is verified in conjunction with business rules (e.g., judging whether the change of the main analysis indicator after intervention conforms to common business sense); a weighted evaluation model is constructed based on data support, effect consistency, and business rationality to calculate the credibility score of each attribution path; a credibility score threshold is set, and attribution paths with scores higher than the threshold are selected as the final results.

[0053] For example, the counterfactual inference results data include inference results from two attribution paths: "credit history → loan default rate" and "debt-to-income ratio → loan default rate". The implementing entity calculated the data support: the sample size supporting the inference result of "credit record → loan default rate" was 85%, and the sample size supporting "debt-to-income ratio → loan default rate" was 72%. Consistency tests were performed on the multiple inference results of the two paths, with coefficients of variation of 0.08 and 0.12 respectively (the smaller the coefficient of variation, the higher the consistency). Combined with business rule verification, the inference results of both paths conform to the common business principle that "optimization of credit record and reduction of debt-to-income ratio will reduce loan default rate." A weighted evaluation was performed based on preset weight coefficients (e.g., data support weight 0.4, effect consistency weight 0.3, business rationality weight 0.3, with a weight sum of 1), resulting in a credibility score of 0.83 for "credit record → loan default rate" and 0.75 for "debt-to-income ratio → loan default rate." With a credibility score threshold of 0.7, both paths met the requirements, and the corresponding attribution path credibility scores were finally output.

[0054] This embodiment, through the above-described scheme, mines true causal relationships from the target dataset and constructs a causal knowledge graph using causal structure learning, providing a rigorous causal logical foundation for attribution analysis and solving the problem of traditional attribution relying on empirical assumptions. By responding to the main analysis indicator selection instruction and filtering recommended associated nodes based on causal strength, it focuses on core influencing factors and avoids interference from irrelevant variables. By quantifying the intervention effect of recommended associated nodes on the main analysis indicator, it achieves an upgrade from qualitative association to quantitative influence. Through multi-dimensional confidence assessment, it obtains an attribution path credibility score, ensuring the reliability of the attribution results.

[0055] Based on the above implementation scheme, in one feasible implementation, the steps of performing causal structure learning on the target dataset to obtain the target causal relationships in the target dataset, and constructing a causal knowledge graph based on the target causal relationships include S21~S24: Step S21: Construct a business constraint rule base, which includes causal prohibition rules and causal enforcement rules defined by business logic.

[0056] It should be noted that the business constraint rule base is a database that stores causal relationship constraint rules at the business level. Its purpose is to provide business prior knowledge support for causal structure learning and ensure that the learning results conform to the actual business logic.

[0057] Additionally, it should be noted that causal prohibition rules refer to causal relationships that are explicitly prohibited in the business logic (e.g., if there is no direct causal relationship between customer gender and loan default rate, it is set as a causal prohibition rule). Their purpose is to eliminate false causal relationships that the algorithm may misjudge. Causal mandatory rules, on the other hand, refer to causal relationships that are explicitly required in the business logic (e.g., if the number of overdue days necessarily affects the loan default rate, it is set as a causal mandatory rule). Their purpose is to ensure that core causal relationships are not overlooked by the algorithm.

[0058] Understandably, the purpose of this step is to address the problem that purely algorithm-driven causal structure learning may deviate from business logic, generate false causal relationships that do not conform to reality, or omit core causal relationships. The principle is to transform the experience and knowledge of business experts into structured constraint rules, define a reasonable range for subsequent causal structure learning, and ensure that the learning results are both consistent with data patterns and adapted to business realities, thereby improving the reliability and practicality of causal relationships.

[0059] Specifically, a business constraint rule library is built using a relational database (such as MySQL), and stored in categories such as causal prohibition rules and causal enforcement rules. Each rule contains fields such as variable A, variable B, rule type, and business basis. Unstructured business constraint rules (such as business logic descriptions containing natural language) are stored using a non-relational database (such as MongoDB). The latest business constraint rules are obtained from the business system in real time through a data interface (suitable for scenarios where business rules are frequently updated). Updated business constraint rules are periodically imported from local files (such as Excel and CSV) (suitable for scenarios where business rules are relatively stable).

[0060] Step S22: Perform time-series stationarity detection and segmentation on the target dataset to generate a target data subset.

[0061] It's important to note that time-series stationarity testing verifies the stability of the target dataset's statistical properties (such as mean and variance) over time. It primarily employs techniques like ADF (Augmented Dickey-Fuller Test) and KPSS (Kwiatkowski-Phillips-Schmidt-Shin Test) to determine if the data exhibits trend or periodic fluctuations, thus preventing non-stationary data from interfering with causal structure learning. Time-series segmentation divides the non-stationary target dataset into multiple stationary data segments based on time points. This transforms complex non-stationary data into simpler, stationary data, improving the accuracy of causal structure learning. The target data subset refers to each stationary data segment obtained after segmentation, serving as input data for causal structure learning to ensure the stability of the learning results.

[0062] Understandably, the purpose of this step is to address the problem of unstable and low accuracy in causal structure learning results caused by trend and periodic fluctuations in time series data. The principle is to identify the time characteristics of the data through stationarity detection, and to segment non-stationary data so that each data subset has stable statistical characteristics. This allows causal structure learning to more accurately uncover the inherent causal relationships in the data and avoids the interference of false associations caused by time fluctuations.

[0063] Specifically, the ADF test is used to detect the stationarity of the target dataset. If the p-value of the test result is less than 0.05, the data is considered stationary and no splitting is needed; if the p-value is greater than or equal to 0.05, the data is considered non-stationary and splitting is required, where p is the probability value. The KPSS test can also be used to assist in verifying the stationarity of the data, and cross-validation with the ADF test results is performed. For non-stationary data in datasets with a clear period, the non-stationary data is split by year, quarter, or month (applicable to datasets with a clear time period). For data with an unclear period, the non-stationary data can be split based on the sliding window method (e.g., a window size of 3 months and a step size of 1 month) to ensure that the data within each window is stationary; the stationarity of each data subset after splitting is tested again, and the data subsets that are still non-stationary are removed or the splitting method is readjusted.

[0064] Step S23: Perform conditional independence checks on the target data subset using a causal structure learning algorithm to generate an undirected initial causal skeleton.

[0065] It should be noted that causal structure learning algorithms refer to algorithms used to uncover causal relationships between variables. The core algorithms employed include constraint-based PC algorithms and SGS (Spirtes-Glymour-Scheines Algorithm), which identify causal relationships between variables through statistical tests. Conditional independence tests are statistical methods to determine whether two variables are independent given other variables. The core techniques employed include the chi-square test (suitable for discrete variables), mutual information test (suitable for continuous variables), and partial correlation coefficient test, which aim to eliminate spurious associations between variables and retain potential causal relationships.

[0066] Additionally, it should be noted that the initial causal skeleton refers to the network structure formed by connecting variable nodes with potential causal relationships using undirected edges (it only reflects the existence of relationships between variables, without specifying the causal direction), and its function is to provide a basic network structure for subsequent determination of causal direction.

[0067] Understandably, the purpose of this step is to solve the problem of directly identifying potential causal relationships between variables from high-dimensional data. The principle is to use a causal structure learning algorithm combined with conditional independence testing to gradually eliminate pairs of variables without causal relationships, retain potential causal relationships with statistical significance, form an initial causal framework, and lay the foundation for clarifying the causal direction and constructing the target causal network, thus ensuring the systematicness and accuracy of causal structure learning.

[0068] Specifically, for high-dimensional data subsets, the PC algorithm is used to learn the causal structure of the target data subset. By gradually increasing the dimensionality of the condition variable set, conditional independence tests are performed on variable pairs, independent variable pairs are eliminated, and dependent variable pairs are retained to form an initial causal framework. For low-dimensional data, the SGS algorithm is used to perform comprehensive conditional independence tests on variable pairs and directly construct the initial causal framework (suitable for low-dimensional datasets). For discrete variables, the chi-square test is used to test conditional independence, and for continuous variables, the partial correlation coefficient test is used. A significance level for the conditional independence test is set (e.g., a significance level of α=0.05). If the probability value P-value is less than α, the variable pair is determined to be conditionally dependent and the association is retained. Finally, initial causal frameworks are generated for multiple target data subsets, and then merged through a voting method (e.g., majority voting) to obtain a unified initial causal framework (suitable for multi-subset data).

[0069] Step S24: Prune the initial causal skeleton based on the causal prohibition rule to remove pseudo-causal relationships that do not conform to business logic, and add specific causal relationships to the initial causal skeleton based on the causal enforcement rule to obtain a target causal network with causal direction, and construct the causal knowledge graph based on the target causal network.

[0070] It should be noted that pruning refers to the operation of deleting undirected edges (pseudo-causal relationships) in the initial causal skeleton that do not conform to the prohibition of causality. The core of this operation is rule matching technology, which is used to eliminate causal relationships that are misjudged by the algorithm or do not conform to business logic. Adding specific causal relationships refers to supplementing the initial causal skeleton with undirected edges that are required by the mandatory causal rules (if they do not exist) and clarifying their causal direction. This is used to ensure that core causal relationships are not overlooked.

[0071] Additionally, it should be noted that the target causal network refers to a directed causal network (including nodes, directed edges, and causal strength) formed after pruning, edge completion, and causal direction determination. Its function is to provide a precise foundation for causal logic. A causal knowledge graph refers to a visualized and structured knowledge carrier built based on the target causal network. Its function is to intuitively display causal relationships and provide support for subsequent queries and inferences.

[0072] Understandably, the purpose of this step is to address the issues that the initial causal framework is based solely on statistical data patterns, may contain spurious causal relationships or omit core causal relationships, and lacks a clear causal direction. The principle is to optimize the initial causal framework by combining it with a business constraint rule base (pruning spurious causal relationships and supplementing core causal relationships), and then clarify the causal interaction direction between variables through causal direction determination, forming a rigorous target causal network. Finally, a causal knowledge graph that combines data patterns and business logic is constructed, providing reliable causal logic support for subsequent attribution analysis.

[0073] Specifically, in one embodiment of this application, an exact matching algorithm is used to match undirected edges (variable pairs) in the initial causal skeleton with variable pairs in the causal prohibition rule base, and deletes successfully matched undirected edges (pruning). In another embodiment of this application, a fuzzy matching algorithm (such as variable name similarity matching) is used to handle the situation where the variable names in the rule base and the dataset are inconsistent. Then, the variable pairs in the causal enforcement rules are matched with the initial causal skeleton. If no corresponding undirected edge exists, the undirected edge is added. Then, the causal direction is determined based on the chronological order (e.g., if "loan application" occurs before "loan approval", the direction is "loan application → loan approval"). At the same time, the causal direction is determined based on business logic (e.g., if "overdue days" is the cause and "loan default rate" is the effect, the direction is "overdue days → loan default rate"). A scoring-based direction determination algorithm (such as BIC scoring) is used to score the possible directions of undirected edges and select the optimal direction; finally, the nodes, directed edges, causal strength, and other information of the target causal network are imported into a knowledge graph tool (such as Neo4j) to construct a visualized causal knowledge graph.

[0074] This embodiment, through the above-described scheme, obtains a business constraint rule library containing causal prohibition rules and causal enforcement rules, providing prior business knowledge for causal structure learning and avoiding deviations from business logic in causal relationships caused by purely algorithm-driven approaches. Through time series stationarity detection and segmentation, the non-stationary target dataset is transformed into a stationary subset, eliminating the interference of time fluctuations on causal structure learning and improving the stability of the learning results. An undirected initial causal skeleton is generated through the causal structure learning algorithm and conditional independence test, systematically mining potential causal relationships between variables. The initial causal skeleton is pruned and edge-filled using business constraint rules, and the causal direction is clarified, resulting in a target causal network that combines data regularity and business logic, ultimately constructing a causal knowledge graph.

[0075] Based on the above implementation scheme, in one feasible implementation, the step of responding to the selection instruction of the main analysis indicator of the target user further includes S31~S32: Step S31: Determine the corresponding master analysis data in the target dataset based on the master analysis indicators.

[0076] It should be noted that the master analysis data refers to all data records and related fields in the target dataset that are directly related to the master analysis metric. It primarily employs data retrieval and extraction technologies to filter out core data related to the master analysis metric from a massive target dataset, providing a foundational data subset for subsequent data filtering. The corresponding master analysis data includes the value of the master analysis metric, timestamps, and related dimension fields (such as customer ID, organization ID, product ID, etc.), ensuring data integrity and traceability.

[0077] Understandably, the purpose of this step is to address the problem of the large size of the target dataset and the low efficiency of direct processing. The principle is to accurately locate relevant data based on the main analysis indicators, eliminate redundant data that is unrelated to the main analysis indicators, narrow the scope of data processing, improve the efficiency of subsequent data filtering and analysis, and ensure that core data is not missed.

[0078] Specifically, if the target dataset is stored in a relational database (such as MySQL), the master analysis data is extracted using SQL queries (e.g., "SELECT * FROM target_data_table WHERE metric_name='customer churn rate'"). If the target dataset is stored in a non-relational database (such as MongoDB), aggregate queries are used to extract all document data containing the master analysis metric fields. If the target dataset is a local file (such as CSV or Excel), the file can be read using Python's pandas library, and the columns corresponding to the master analysis metrics and related fields can be filtered to form the master analysis data. Finally, the extracted master analysis data is cleaned (e.g., removing null values ​​and outliers) to ensure data quality.

[0079] Step S32: Filter the master analysis data based on preset dynamic query conditions to obtain a target scene dataset. The target scene dataset is used to calculate the marginal explanatory power increment of the target node in the set of associated nodes in the causal knowledge graph. The dynamic query conditions include institutional screening conditions, line screening conditions, and time range conditions.

[0080] It should be noted that dynamic query conditions refer to filtering conditions that can be flexibly configured according to user needs. The core of the technology is to use condition filtering and data matching to filter target data in specific scenarios from the main analysis data, thereby improving the scenario adaptability of attribution analysis.

[0081] Additionally, it should be noted that the "Institutional Filtering Conditions" refer to filtering conditions set according to the institutional dimension (such as branch offices, departments, outlets, etc.) (e.g., "Institution ID=1001", "Institution Type=First-Tier City Branch Office"), which are used to focus on attribution analysis of specific institutions. The "Business Line Filtering Conditions" refer to filtering conditions set according to the business line dimension (such as retail business, corporate business, credit business, etc.) (e.g., "Business Line=Retail Business", "Product Line=Product A"), which are used to focus on attribution analysis of specific business lines. The "Time Range Conditions" refer to filtering conditions set according to the time dimension (such as year, quarter, month, day) (e.g., "Timestamp BETWEEN '2023-01-01' AND '2023-12-31'", "Quarter=Q3 of 2023"), which are used to focus on attribution analysis of a specific time range. The "Target Scenario Dataset" refers to the data set corresponding to a specific business scenario obtained after dynamic query filtering, which is used to provide scenario-based data support for subsequent related node queries and intervention effect calculations.

[0082] Understandably, the purpose of this step is to address the problem that traditional attribution analysis lacks contextual adaptation and that the attribution results of the same indicator vary greatly in different scenarios. The principle is to dynamically configure multi-dimensional filtering conditions to limit the main analysis data to a specific organization, business line, and time range, thereby obtaining a target scenario dataset that matches the user's actual analysis scenario. This ensures that the subsequent attribution analysis results can accurately reflect the causal relationship in that scenario, and improve the practicality and relevance of the attribution results.

[0083] Specifically, a visual condition configuration interface is provided, allowing target users to set filtering conditions such as organization (e.g., selecting Beijing branch, Shanghai branch), business line (e.g., selecting "retail business"), and time range (e.g., selecting "July 2023 - September 2023") through checkboxes and input. These user-configured conditions are then converted into data filtering rules. The master analysis data undergoes multi-dimensional joint filtering based on dynamic query conditions: first filtering by organization, then by business line, and finally by time range (or adjusting the filtering order). Conditional logical operations are supported (e.g., "Organization = 1001 OR Organization = 1002" or "Time range: Q3 2023 AND Business line = Retail Business"), enhancing the flexibility of condition configuration. After filtering, the target scenario dataset is validated for sufficient data volume. If the data volume is too small (e.g., less than 1000 records), the user is prompted to adjust the query conditions to avoid unreliable attribution results due to insufficient data.

[0084] This embodiment, through the above-described scheme, clarifies the core objective of attribution analysis by responding to the target user's main analysis indicator selection instruction, ensuring that the analysis direction aligns with user needs; it extracts corresponding main analysis data from the target dataset based on the main analysis indicators, eliminates redundant data, narrows the data processing scope, and improves analysis efficiency; and it filters the main analysis data based on dynamic query conditions including organization, line of business, and time range to obtain the target scenario dataset, thus achieving scenario-based adaptation of attribution analysis.

[0085] Based on the above implementation scheme, in one feasible implementation, the step of querying the set of nodes that have a causal relationship with the main analysis index from the causal knowledge graph and determining the recommended associated nodes based on the causal strength includes S41~S43: Step S41: Extract the set of associated nodes connected to the node where the main analysis index is located from the causal knowledge graph, and obtain the causal strength weight and user adoption rate of each causal edge in the set of associated nodes.

[0086] It should be noted that the set of associated nodes refers to all nodes in the causal knowledge graph that have direct or indirect causal connections with the node containing the main analysis indicator (e.g., when the main analysis indicator is "customer churn rate", associated nodes include customer satisfaction, product usage frequency, customer service response speed, etc.). The core uses graph traversal algorithms (such as depth-first search and breadth-first search), and its function is to delineate the range of potential factors that affect the main analysis indicator.

[0087] Additionally, it should be noted that causal strength weight refers to the quantified weight of the causal edge connecting the node containing the main analysis indicator and related nodes in the causal knowledge graph (e.g., obtained through mutual information value or regression coefficient normalization, with a value range of 0-1). Its function is to quantify the degree of causal influence of each related node on the main analysis indicator. User adoption rate refers to the proportion of target users who adopted the related node as an attribution factor in historical analysis (e.g., if a related node was adopted by users 8 times in the past 10 similar analyses, the user adoption rate is 80%). It primarily uses historical data statistical techniques and its function is to combine users' historical preferences to filter related nodes that better meet actual needs.

[0088] Understandably, the purpose of this step is to address the issues of unclear scope of potential influencing factors, lack of quantitative assessment, and consideration of user preferences. The principle is to accurately extract the set of related nodes with the main analysis indicators from the causal knowledge graph through graph traversal algorithm, then quantify the influence of each node through causal strength weight, and combine it with user adoption rate to reflect the actual application value of the node. This provides a multi-dimensional evaluation basis for determining the related nodes for subsequent recommendations, ensuring that the recommendation results are both in line with data patterns and meet user needs.

[0089] Specifically, Breadth-First Search (BFS) is used to start from the node containing the main analysis metric and traverse all directly or indirectly connected nodes to generate a set of associated nodes (a traversal depth threshold can be set, such as limiting it to 3rd-order neighbor nodes to avoid an excessively large node range). Depth-First Search (DFS) is used to extract the set of associated nodes, suitable for scenarios that require prioritizing the mining of deep indirectly related nodes. The causal strength weights of each causal edge are directly read from the metadata of the causal knowledge graph (e.g., stored in the edge attributes of Neo4j). If the causal strength weights are not preset, the strength of each causal edge is calculated in real time using methods such as mutual information calculation and conditional probability analysis and then normalized (converted into 0-1 weight values). The historical analysis record database is queried to calculate the ratio of the number of times each associated node was adopted by users to the total number of analyses, obtaining the user adoption rate (e.g., if a node participated in 20 analyses and was adopted 15 times, the adoption rate is 75%). The adoption rate is subject to time decay processing (e.g., the weight of adoption records within the most recent year is 1, the weight of 1-2 years is 0.7, and the weight of more than 2 years is 0.3) to ensure that the adoption rate reflects recent user preferences.

[0090] Step S42: Calculate the marginal explanatory power increment of each target node in the set of associated nodes based on the target scene dataset, and generate a recommendation score based on the marginal explanatory power increment, the causal strength weight, and the user adoption rate.

[0091] It's important to note that target nodes refer to each node in the set of associated nodes (e.g., "customer satisfaction," "product usage frequency," etc.), serving as the object for calculating marginal explanatory power increment. Marginal explanatory power increment refers to the increase in explanatory power of a target node on the main analytical indicator after being added to the attribution model within the target scenario dataset. It primarily employs techniques such as regression analysis and decision tree feature importance assessment (e.g., R² increment, information gain), quantifying the actual explanatory power of the target node for the main analytical indicator in a specific scenario. Recommendation score refers to a quantitative score (e.g., 0-100 points) obtained by combining marginal explanatory power increment, causal strength weights, and user adoption rate. It primarily uses a weighted summation algorithm, providing a comprehensive evaluation of each target node and a basis for subsequent ranking and recommendation.

[0092] Understandably, the purpose of this step is to address the problem that a single-dimensional evaluation cannot fully reflect the importance of nodes. The principle is to combine three dimensions—marginal explanatory power increment in a specific scenario (scenario adaptability), global causal strength weight (causal correlation), and user adoption rate (practical application value)—to comprehensively score each target node, avoiding the one-sidedness caused by a single-dimensional evaluation, and ensuring that the recommended nodes not only meet the needs of the scenario but also have strong causal correlation and practical application value.

[0093] Specifically, a linear regression model is used to calculate the marginal explanatory power increment: First, a baseline model containing only the main analysis indicators is constructed, and the baseline R² value is calculated; then, each target node is added to the model sequentially, and a new R² value is calculated. The difference between the new R² value and the baseline R² value is the marginal explanatory power increment of that target node (e.g., if the baseline R² = 0.3, after adding "customer satisfaction" R² = 0.65, the marginal explanatory power increment = 0.35). Alternatively, a decision tree algorithm, such as CART (Classification and Regression Tree), is used to calculate the information gain of each target node, and the normalized information gain is used as the marginal explanatory power increment. Then, a weighted summation formula is used to generate the recommendation score: Recommendation Score = w1 × Marginal Explanatory Power Increment + w2 × Causal Strength Weight + w3 × User Adoption Rate (where w1, w2, and w3 are weight coefficients, satisfying w1 + w2 + w3 = 1, which can be preset by business experts or optimized through grid search, such as w1 = 0.4, w2 = 0.3, and w3 = 0.3). The marginal explanatory power increment, causal strength weight, and user adoption rate are normalized (converted to values ​​between 0 and 1) before being weighted and summed to ensure a consistent score range for each dimension. User-defined weight coefficients are also supported to suit different users' emphasis on each dimension (e.g., if a user is more concerned with scenario adaptability, w1 = 0.5, w2 = 0.2, and w3 = 0.3 can be set).

[0094] Step S43: Sort the nodes in the set of associated nodes based on the recommendation score, and determine the recommended associated nodes in the set of associated nodes based on the sorting results.

[0095] Understandably, the purpose of this step is to address the issue of varying node importance within the set of associated nodes, necessitating the selection of core nodes. The principle is to clarify the overall importance of nodes through recommendation scoring and ranking, and then obtain recommended associated nodes through selection. This avoids irrelevant or secondary nodes interfering with subsequent analysis, improves the efficiency and relevance of attribution analysis, and ensures that subsequent intervention effect calculations focus on core nodes that have a significant impact on the main analysis indicators and align with scenario needs and user preferences.

[0096] Specifically, a quicksort algorithm is used to sort related nodes from highest to lowest recommendation score, yielding the sorted results. Sort by recommendation score in ascending order is supported (suitable for scenarios where secondary nodes need to be excluded). Then, a threshold filtering method is used: a preset recommendation score threshold (e.g., 0.6 points) is set, and nodes with scores higher than the threshold are selected as recommended related nodes. Alternatively, a fixed-number filtering method can be used: a preset number of recommended nodes (e.g., 3, 5) is set, and the top N nodes in the sorted results are selected as recommended related nodes. Both filtering methods can also be combined: first, candidate nodes are filtered by threshold; if the number of candidate nodes is too large (e.g., more than 10), the top N nodes are selected; if the number of candidate nodes is too small (e.g., less than 2), the threshold is lowered and the filtering is repeated.

[0097] This embodiment uses the above-described scheme to extract a set of related nodes connected to the main analysis index from the causal knowledge graph, quantify the degree of causal association of the nodes by combining causal strength weights, and reflect the actual application value of the nodes by combining user adoption rates; calculate the marginal explanatory power increment of each target node based on the target scenario dataset to quantify the explanatory power of the node in a specific scenario; and sort the nodes by combining the recommendation scores of the three dimensions to obtain recommended related nodes.

[0098] Based on the above implementation scheme, in one feasible implementation, the step of performing a multi-dimensional confidence assessment on the counterfactual inference result data to obtain the attribution path credibility score includes S51~S53: Step S51: Perform a multi-dimensional confidence assessment on the counterfactual inference result data, and determine the attribution analysis path of the main analysis indicator based on the assessment results.

[0099] It should be noted that the attribution analysis path refers to the effective causal chain from the recommended association node to the main analysis indicator (such as "customer satisfaction → product usage frequency → customer churn rate"). The core of the path filtering algorithm is to clarify the key attribution logic link of the main analysis indicator.

[0100] Understandably, the purpose of this step is to address potential issues such as randomness, logical contradictions, or business inconsistencies in the back-inference results, ensuring the reliability of the attribution analysis path. The principle is to verify the data support strength, causal logic consistency, and business adaptability of the inference results through multi-dimensional confidence assessment, eliminate paths with low confidence, and select effective attribution analysis paths that combine data regularity and business logic, laying the foundation for subsequent credibility score calculation.

[0101] Specifically, the data support assessment includes: calculating the proportion of samples supporting a particular attribution path in the counterfactual inference results (e.g., if 21,000 samples support a certain path and the total number of samples is 23,700, the data support is 88.6%); the effect consistency assessment involves calculating the coefficient of variation (standard deviation / mean) for multiple inferences of the same attribution path (e.g., inferences from different batches of samples). A smaller coefficient of variation indicates higher consistency (e.g., setting a coefficient of variation threshold of 0.1; values ​​below this value are considered acceptable); and the business rationality assessment involves verifying whether the causal logic of the attribution path conforms to common business sense (e.g., "increased customer satisfaction → increased product usage frequency → decreased customer churn rate"). (If it conforms to business logic, it is deemed reasonable); Calculation of comprehensive confidence score: The comprehensive confidence score of each attribution path is calculated using a weighted summation formula (data support weight 0.4, effect consistency weight 0.3, business reasonableness weight 0.3) (e.g., if a path has data support of 0.88, effect consistency of 0.92, and business reasonableness of 1.0, the comprehensive confidence score is 0.4×0.88+0.3×0.92+0.3×1.0=0.91); Finally, path selection is performed: A comprehensive confidence score threshold is set (e.g., 0.7), and attribution paths with scores higher than the threshold are selected as attribution analysis paths for the main analysis indicator.

[0102] Step S52: Obtain the data support, effect consistency deviation, and user adoption rate of each causal edge in the attribution analysis path.

[0103] It should be noted that a "causal edge" refers to a directed edge connecting two nodes in the attribution analysis path (such as "customer satisfaction → product usage frequency" and "product usage frequency → customer churn rate" in the path "customer satisfaction → product usage frequency"). Its function is to form the basic unit of the attribution path.

[0104] Additionally, it should be noted that data support refers to the proportion of samples in the target scenario dataset that support the existence of the causal edge (e.g., for the causal edge "customer satisfaction → product usage frequency", 85% of the samples in the target scenario data show both "increased customer satisfaction" and "increased product usage frequency"). It primarily employs sample statistical techniques and serves as the data analysis basis for quantifying causal edges. Effect consistency deviation refers to the degree of fluctuation in the effect value of the causal edge across multiple inferences (e.g., calculated using standard deviation; the smaller the deviation, the higher the consistency). It primarily employs statistical deviation analysis techniques and serves to reflect the stability of the causal edge effect. User adoption rate refers to the proportion of times the causal edge was adopted by users as part of the attribution path in historical analyses (e.g., a causal edge was adopted 16 times in the past 20 similar analyses, with an adoption rate of 80%). It primarily employs historical data retrieval and statistical techniques and serves to evaluate the practical application value of the causal edge in conjunction with users' historical preferences.

[0105] Understandably, the purpose of this step is to obtain multi-dimensional quantitative indicators for each basic unit in the attribution analysis path, providing a fine-grained basis for subsequent path credibility score calculation. The principle is to obtain the data support of the causal edge, the effect consistency deviation value, and the user adoption rate by statistically analyzing target scenario data, analyzing and inferring stability, and retrieving historical records, respectively, to ensure that subsequent score calculation can comprehensively reflect the reliability and practicality of each link in the path.

[0106] Specifically, the data support is first obtained by: for each causal edge, the number of samples in the target scenario dataset that satisfy "change in dependent variable → corresponding change in effect variable" is counted, and the ratio to the total number of samples is the data support (e.g., for the causal edge "customer service response speed → customer churn rate", the number of samples that "shortened customer service response speed" and "decreased customer churn rate" is counted, and the proportion is calculated). Then, the effect consistency deviation is calculated: the effect value of the causal edge in multiple back-inferences (e.g., the effect value in 10 batch inferences) is extracted, and the standard deviation is calculated as the effect consistency deviation (e.g., effect values ​​of -0.08, -0.09, and -0.07, with a standard deviation of 0.008, and a deviation of 0.008). Finally, the user adoption rate is obtained by: querying the historical analysis record database and counting the number of times the attribution path containing the causal edge was adopted by users, and the ratio to the total number of analyses containing the causal edge is the user adoption rate (e.g., if a causal edge participated in 15 analyses and was adopted 12 times, the adoption rate is 80%). Next, the consistency deviation value is normalized: the deviation value is converted into a value between 0 and 1 (the smaller the deviation value, the higher the score after normalization, such as a deviation value of 0.008 being normalized to 0.92), which facilitates subsequent weighted calculation. Time decay is applied to the user adoption rate: historical adoption records are weighted according to time (e.g., records within 1 year have a weight of 1.0, records from 1-2 years have a weight of 0.7, and records older than 2 years have a weight of 0.3), and the weighted adoption rate is recalculated to ensure that it reflects recent user preferences.

[0107] Step S53: Based on the data support, the effect consistency deviation value, and the user adoption rate, a weighted calculation is performed to obtain the attribution path credibility score of the attribution analysis path.

[0108] Understandably, the purpose of this step is to address the problem that multi-dimensional evaluation indicators are difficult to intuitively reflect the overall reliability of the path. The principle is to integrate the indicators of data support strength, effect stability, and user preference into a unified credibility score through weighted calculation, so as to realize the quantitative evaluation of the reliability of the attribution path, allowing users to quickly judge the credibility of the attribution results and improve the practicality of attribution analysis.

[0109] Specifically, first, preset weight coefficients are established: weights are set based on business needs and / or expert experience (e.g., data support 0.4, normalization effect consistency deviation 0.3, user adoption rate 0.3) to ensure the reasonable contribution of each dimension indicator. Then, dynamic weight adjustments are performed: weight coefficients are optimized based on historical data using the Analytic Hierarchy Process (AHP) or machine learning algorithms (e.g., automatically increasing the weight when data support is low). Finally, a weighted summation calculation is performed: Attribution path credibility score = w1 × data support + w2 × normalization effect consistency deviation + w3 × user adoption rate (e.g., if the indicators for a certain path are 0.89, 0.94, and 0.86, with weights of 0.4, 0.3, and 0.3, the score = 0.4 × 0.89 + 0.3 × 0.94 + 0.3 × 0.86 = 0.89). After calculation, the score is normalized: the calculation result is converted into a 0-100 point scale (e.g., 0.89 × 100 = 89 points) to improve user readability. Path total score calculation: For attribution paths containing multiple causal edges (such as path 2 containing E2 and E3), first calculate the credibility score of each causal edge, and then sum them up according to the causal strength weight of the causal edges to obtain the path total score (such as E2 score 0.84, causal strength 0.7; E3 score 0.88, causal strength 0.9, path total score = (0.84×0.7+0.88×0.9) / (0.7+0.9)≈0.86).

[0110] This embodiment, through the above-described scheme, performs multi-dimensional confidence assessment on the counterfactual reasoning results data, verifying the reliability of the attribution path from three dimensions: data support, effect consistency, and business rationality, ensuring that the attribution path combines data regularity and business logic; by obtaining the data support degree, effect consistency deviation value, and user adoption rate of each causal edge in the attribution path, it provides fine-grained, multi-dimensional quantitative basis for the path credibility score; and by weighted calculation, it integrates the multi-dimensional indicators into a unified attribution path credibility score, achieving intuitive quantification of the attribution path reliability.

[0111] Based on the above implementation scheme, in one feasible implementation, the step of performing a multi-dimensional confidence assessment on the counterfactual inference result data and determining the attribution analysis path of the main analysis indicator based on the assessment results includes S61~S63: Step S61: Calculate the comprehensive confidence of each causal edge in the causal knowledge graph based on the data support, the effect consistency deviation value, and the user adoption rate.

[0112] It should be noted that the comprehensive confidence level refers to a comprehensive index (range 0-1) that quantifies the reliability of causal edges. The core of it uses a weighted summation algorithm, and its function is to integrate the multi-dimensional evaluation indexes of causal edges to provide a basis for subsequent node selection and path generation.

[0113] Understandably, the purpose of this step is to address the problem that a single indicator is insufficient to comprehensively assess the reliability of causal edges. The principle is to integrate indicators from three dimensions—data support strength, effect stability, and user preference—through weighted calculation to obtain a comprehensive confidence level that reflects the overall reliability of causal edges. This lays the foundation for subsequent screening of effective attribution nodes and generation of candidate attribution paths, ensuring that subsequent path construction is based on highly reliable causal edges.

[0114] Specifically, in one embodiment of this application, a fixed-weighted summation is used: preset weight coefficients (data support 0.4, normalized effect consistency deviation 0.3, user adoption rate 0.3), and the overall confidence score = 0.4 × data support + 0.3 × normalized effect consistency deviation + 0.3 × user adoption rate (e.g., for a causal edge with data support of 0.89, normalized deviation 0.94, and user adoption rate of 0.86, the overall confidence score = 0.4 × 0.89 + 0.3 × 0.94 + 0.3 × 0.86 = 0.89). In another embodiment of this application, a dynamic-weighted summation is used: the weights are adjusted based on the type of causal edge (e.g., direct causal edge, indirect causal edge) (the weight of data support for direct causal edges is increased to 0.5, and the weight of effect consistency deviation for indirect causal edges is increased to 0.4). Simultaneously, the user adoption rate can be adjusted to suit specific business scenarios (e.g., the weight of user adoption rate is increased to 0.4 in a specific business scenario). After calculation, the overall confidence level is normalized: if the calculation result exceeds the range of 0-1, Min-Max normalization is used to ensure that the value meets the standard. Then, abnormal causal edge filtering is performed: a minimum threshold for the overall confidence level is set (e.g., 0.5), and causal edges with scores below the threshold are directly removed to avoid low-reliability causal edges affecting subsequent analysis.

[0115] Step S62: Filter the recommended associated nodes and retain the nodes with a comprehensive confidence level greater than a preset confidence threshold as valid attribution nodes. Generate a set of candidate attribution paths based on the causal direction relationship between the valid attribution nodes. The candidate attribution paths contain at least one causal chain from the recommended associated nodes to the main analysis index.

[0116] It should be noted that the preset reliability threshold refers to the overall confidence threshold (e.g., 0.7) used to filter effective attribution nodes. The core technology employed is threshold setting (which can be user-defined or preset by business experts), and its purpose is to filter out highly reliable recommendation nodes. Effective attribution nodes refer to recommendation nodes with an overall confidence level greater than the preset reliability threshold (e.g., "customer satisfaction," "product usage frequency," "customer service response speed"), and their function is to serve as the starting point for candidate attribution paths.

[0117] Additionally, it's important to clarify that causal direction refers to the cause-effect relationship between valid attribution nodes and the main analytical indicator, as well as the "cause → effect" relationship between each node (e.g., "customer satisfaction → product usage frequency → customer churn rate"). Its purpose is to ensure that candidate attribution paths conform to causal logic. The candidate attribution path set refers to the collection of all valid causal chains pointing from valid attribution nodes to the main analytical indicator. It utilizes a path search algorithm to provide a comprehensive range of potential attribution paths. A causal chain refers to a logical link from the starting node to the main analytical indicator, consisting of continuous causal edges (e.g., "customer service response speed → customer churn rate" is a causal chain composed of a single causal edge, while "customer satisfaction → product usage frequency → customer churn rate" is a causal chain composed of multiple causal edges).

[0118] Understandably, the purpose of this step is to address the issues of low-reliability nodes and unclear attribution path ranges in the recommended association nodes. The principle is to filter high-reliability effective attribution nodes by setting a pre-set reliability threshold, eliminate the interference of low-reliability nodes on the attribution results, and then search for all possible effective causal chains based on causal direction relationships to generate a set of candidate attribution paths, ensuring that subsequent path selection is based on comprehensive and highly reliable potential paths.

[0119] Specifically, in one embodiment of this application, a pre-set confidence threshold is used: a fixed threshold (e.g., 0.7) is set by business experts based on historical data and business needs, suitable for situations where the business scenario is stable. In another embodiment of this application, a dynamic threshold adjustment is used: the threshold is automatically adjusted based on the comprehensive confidence distribution of recommended associated nodes (e.g., mean + 1 standard deviation), suitable for scenarios where the data distribution fluctuates greatly. Then, effective attribution node screening is performed: all recommended associated nodes are traversed, and nodes whose comprehensive confidence of all causal edges associated with them is greater than the threshold are retained (e.g., E1 and E2 associated with "customer satisfaction" both have a comprehensive confidence greater than 0.7, and are retained as effective attribution nodes). After screening, path search is performed: a depth-first search (DFS) algorithm is used to start from each effective attribution node, traversing along the causal direction relationship to the main analysis indicator, generating all possible causal chains (e.g., starting from "customer satisfaction", two causal chains are obtained: "customer satisfaction → customer churn rate" and "customer satisfaction → product usage frequency → customer churn rate"). Then, path deduplication is performed: the searched causal chains are deduplicated (e.g., to avoid generating the same path repeatedly). Simultaneously, path length limits are imposed: a maximum length for the causal chain is set (e.g., 3 causal edges) to avoid generating excessively long indirect paths (the effect of indirect paths decays significantly and has low reliability); finally, a candidate attribution path set is constructed: all valid causal chains after deduplication are integrated to form a candidate attribution path set.

[0120] Step S63: Perform a weighted summation on the candidate attribution paths in the candidate attribution path set to obtain the total confidence of the candidate attribution paths, and select the candidate attribution path with the highest total confidence as the attribution analysis path of the main analysis indicator.

[0121] It should be noted that the total confidence of a path refers to the weighted sum of the combined confidence of each causal edge in the candidate attribution path (e.g., if path C2 contains E2 and E3, with causal strengths of 0.7 and 0.9 respectively, the total confidence of the path = 0.84 × 0.7 + 0.88 × 0.9 = 1.344), which directly reflects the overall reliability of the candidate path. Causal strength refers to the quantified value of the causal influence of a causal edge (ranging from 0 to 1), stored in the edge attributes of the causal knowledge graph, and its function is to reflect the contribution weight of each causal edge in the path (the higher the causal strength, the greater the contribution to the total confidence of the path).

[0122] Understandably, the purpose of this step is to address the issue of inconsistent path reliability in the candidate attribution path set and the need to select the optimal path. The principle is to calculate the total confidence of each candidate path by weighted summation of causal strength. The higher the total confidence, the stronger the overall reliability of the path and the more significant its causal contribution. The path with the highest total confidence is selected as the attribution analysis path to ensure that the attribution results are based on the most reliable and core causal logic.

[0123] Specifically, the causal strength is first obtained by directly reading the causal strength of each causal edge from the edge attributes of the causal knowledge graph (e.g., the causal strength of E1 is 0.85, E2 is 0.7, E3 is 0.9, and E4 is 0.75). Then, the total confidence of the path is calculated: for each candidate attribution path, all causal edges it contains are traversed, and the sum of the products of the comprehensive confidence of each causal edge and the corresponding causal strength is calculated, that is, the total confidence of the path = Σ(comprehensive confidence of causal edge i × causal strength of causal edge i) (e.g., if path C1 contains E1, the total confidence = 0.89 × 0.85 = 0.7565; if path C2 contains E2 and E3, the total confidence = 0.84 × 0.7 + 0.88 × 0.9 = 0.588 + 0.792 = 1.38). Then, the total confidence score is normalized: the total confidence score of all candidate paths is normalized (converted to a value between 0 and 1) for easy comparison. Method four, optimal path selection: paths are sorted from highest to lowest total confidence score, and the candidate path with the highest score is selected as the attribution analysis path for the main analysis indicator. If multiple paths have the same and highest total confidence score, the path with the fewest causal edges is selected (direct paths are more reliable than indirect paths), or a manual selection is made based on business needs. Finally, the path results are output: the optimal attribution analysis path and its total confidence score are output to provide users with decision-making references.

[0124] This embodiment, through the above-described scheme, integrates data support, effect consistency deviation, and user adoption rate using weighted calculation to obtain the comprehensive confidence of each causal edge, thus comprehensively evaluating the reliability of the causal edges. It filters effective attribution nodes by setting a pre-set confidence threshold, eliminating interference from low-reliability nodes to ensure that the attribution analysis is based on highly reliable potential influencing factors. Based on the causal direction relationship of effective attribution nodes, it generates a candidate attribution path set containing all potentially effective causal chains through path search, ensuring the comprehensiveness of the path range. Finally, it calculates the total confidence of the candidate paths by weighted summation of causal strength, selecting the path with the highest total confidence as the attribution analysis path, ensuring that the attribution results are based on the most reliable and core causal logic.

[0125] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the attribution analysis method of this application. Any simple modifications based on this technical concept are within the protection scope of this application.

[0126] This application also provides an attribution analysis device, please refer to... Figure 2 The attribution analysis device includes: The graph construction module 201 is used to perform causal structure learning on the target dataset, obtain the target causal relationship in the target dataset, and construct a causal knowledge graph based on the target causal relationship. The recommendation node module 201 is used to respond to the selection instruction of the target user's main analysis indicator, query the set of nodes that have a causal relationship with the main analysis indicator from the causal knowledge graph, and determine the recommended associated nodes based on the causal strength. The calculation and inference module 203 is used to calculate the intervention effect of the main analysis index based on the main analysis index and the recommended related nodes, and generate counterfactual inference result data. The confidence assessment module 204 is used to perform multi-dimensional confidence assessment on the counterfactual inference result data to obtain the attribution path credibility score.

[0127] The attribution analysis apparatus provided in this application, employing the attribution analysis method in the above embodiments, can solve the technical problem of attribution analysis relying on empirical assumptions. Compared with the prior art, the beneficial effects of the attribution analysis apparatus provided in this application are the same as those of the attribution analysis method provided in the above embodiments, and other technical features in the attribution analysis apparatus are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0128] This application provides an attribution analysis device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, which are executed by the at least one processor to enable the at least one processor to perform the attribution analysis method in Embodiment 1 above.

[0129] The following is for reference. Figure 3 The diagram illustrates a structural schematic of an attribution analysis device suitable for implementing embodiments of this application. The attribution analysis device in the embodiments of this application may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), in-vehicle terminals (e.g., in-vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. Figure 3 The attribution analysis device shown is merely an example and should not impose any limitations on the functionality and scope of the embodiments of this application.

[0130] like Figure 3 As shown, the attribution analysis device may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.) that can perform various appropriate actions and processes according to a program stored in a read-only memory 1002 or a program loaded from a storage device 1003 into a random access memory 1004. The random access memory 1004 also stores various programs and data required for the operation of the attribution analysis device. The processing unit 1001, the read-only memory 1002, and the random access memory 1004 are interconnected via a bus 1005. An input / output interface 1006 is also connected to the bus. Typically, the following systems can be connected to the input / output interface 1006: input devices 1007 including, for example, a touchscreen, touchpad, keyboard, mouse, image sensor, microphone, accelerometer, gyroscope, etc.; output devices 1008 including, for example, a liquid crystal display (LCD), speaker, vibrator, etc.; storage devices 1003 including, for example, magnetic tape, hard disk, etc.; and communication devices 1009. Communication device 1009 allows the attribution analysis device to communicate wirelessly or wiredly with other devices to exchange data. While the figures show attribution analysis devices with various systems, it should be understood that implementation or possession of all the systems shown is not required. More or fewer systems may be implemented alternatively.

[0131] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from read-only memory 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0132] The attribution analysis device provided in this application, employing the attribution analysis method in the above embodiments, can solve the technical problem of attribution analysis relying on empirical assumptions. Compared with the prior art, the beneficial effects of the attribution analysis device provided in this application are the same as those of the attribution analysis method provided in the above embodiments, and other technical features in this attribution analysis device are the same as those disclosed in the method of the previous embodiment, and will not be repeated here.

[0133] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0134] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0135] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, the computer-readable program instructions being used to execute the attribution analysis method in the above embodiments.

[0136] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems or devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0137] The aforementioned computer-readable storage medium may be included in the attribution analysis device; or it may exist independently and not assembled into the attribution analysis device.

[0138] The aforementioned computer-readable storage medium carries one or more programs. When these programs are executed by the attribution analysis device, the attribution analysis device performs the following actions: performs causal structure learning on the target dataset to obtain target causal relationships within the target dataset, and constructs a causal knowledge graph based on the target causal relationships; responds to a target user's selection instruction for a primary analysis indicator, queries the causal knowledge graph for a set of nodes that are causally related to the primary analysis indicator, and determines recommended associated nodes based on causal strength; calculates the intervention effect of the primary analysis indicator based on the primary analysis indicator and the recommended associated nodes, generating counterfactual inference result data; and performs multi-dimensional confidence assessment on the counterfactual inference result data to obtain an attribution path credibility score.

[0139] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a Local Area Network (LAN) or a Wide Area Network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0140] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0141] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0142] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., a computer program) for executing the above-described attribution analysis method, thereby solving the technical problem that attribution analysis relies on empirical assumptions. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the attribution analysis method provided in the above embodiments, and will not be repeated here.

[0143] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the attribution analysis method described above.

[0144] The computer program product provided in this application can solve the technical problem that attribution analysis relies on empirical assumptions. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as the beneficial effects of the attribution analysis method provided in the above embodiments, and will not be repeated here.

[0145] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. An attribution analysis method, characterized in that, The attribution analysis method includes: Perform causal structure learning on the target dataset to obtain the target causal relationships in the target dataset, and construct a causal knowledge graph based on the target causal relationships; In response to the target user's selection instruction for the main analytical indicator, the set of nodes that have a causal relationship with the main analytical indicator is queried from the causal knowledge graph, and the recommended associated nodes are determined based on the causal strength. Based on the main analysis index and the recommended related nodes, the intervention effect of the main analysis index is calculated to generate counterfactual inference result data; A multi-dimensional confidence assessment is performed on the counterfactual inference results data to obtain the attribution path credibility score.

2. The attribution analysis method as described in claim 1, characterized in that, The steps of performing causal structure learning on the target dataset, obtaining the target causal relationships in the target dataset, and constructing a causal knowledge graph based on the target causal relationships include: Construct a business constraint rule base, which includes causal prohibition rules and causal enforcement rules defined by business logic; Perform time-series stationarity detection and segmentation on the target dataset to generate a target data subset; The target data subset is subjected to conditional independence testing using a causal structure learning algorithm to generate an undirected initial causal skeleton. The initial causal framework is pruned based on the causal prohibition rule to remove pseudo-causal relationships that do not conform to business logic. Specific causal relationships are added to the initial causal framework based on the causal enforcement rule to obtain a target causal network with causal direction. The causal knowledge graph is then constructed based on the target causal network.

3. The attribution analysis method as described in claim 1, characterized in that, Following the step of responding to the target user's instruction to select the main analytics metric, the following is also included: Based on the main analysis indicators, determine the corresponding main analysis data in the target dataset; The master analysis data is filtered based on preset dynamic query conditions to obtain a target scene dataset. The target scene dataset is used to calculate the marginal explanatory power increment of the target node in the set of associated nodes in the causal knowledge graph. The dynamic query conditions include institutional screening conditions, line screening conditions, and time range conditions.

4. The attribution analysis method as described in claim 3, characterized in that, The step of querying the set of nodes in the causal knowledge graph that have a causal relationship with the main analysis indicator, and determining the recommended associated nodes based on the causal strength, includes: Extract the set of associated nodes connected to the main analysis index from the causal knowledge graph, and obtain the causal strength weight and user adoption rate of each causal edge in the set of associated nodes; The marginal explanatory power increment of each target node in the set of associated nodes is calculated based on the target scenario dataset, and a recommendation score is generated based on the marginal explanatory power increment, the causal strength weight, and the user adoption rate. The nodes in the associated node set are sorted based on the recommendation score, and the recommended associated nodes in the associated node set are determined based on the sorting results.

5. The attribution analysis method as described in claim 1, characterized in that, The step of performing a multi-dimensional confidence assessment on the counterfactual inference results data to obtain the attribution path credibility score includes: A multi-dimensional confidence assessment is performed on the counterfactual inference results data, and the attribution analysis path of the main analysis indicator is determined based on the assessment results; Obtain the data support, effect consistency bias, and user adoption rate for each causal edge in the attribution analysis path; The attribution path credibility score of the attribution analysis path is obtained by weighting the data support, the effect consistency deviation value, and the user adoption rate.

6. The attribution analysis method as described in claim 5, characterized in that, The step of performing a multi-dimensional confidence assessment on the counterfactual inference results data and determining the attribution analysis path of the main analysis indicator based on the assessment results includes: The comprehensive confidence of each causal edge in the causal knowledge graph is calculated based on the data support, the effect consistency deviation value, and the user adoption rate. The recommended associated nodes are filtered, and nodes with a comprehensive confidence score greater than a preset confidence score threshold are retained as valid attribution nodes. A set of candidate attribution paths is generated based on the causal direction relationship between the valid attribution nodes. The candidate attribution paths contain at least one causal chain from the recommended associated nodes to the main analysis index. The candidate attribution paths in the candidate attribution path set are weighted and summed to obtain the total confidence of the candidate attribution paths. The candidate attribution path with the highest total confidence is selected as the attribution analysis path of the main analysis indicator.

7. An attribution analysis device, characterized in that, The attribution analysis device includes: The graph construction module is used to perform causal structure learning on the target dataset, obtain the target causal relationships in the target dataset, and construct a causal knowledge graph based on the target causal relationships. The recommendation node module is used to respond to the selection instruction of the target user's main analysis indicator, query the set of nodes that have a causal relationship with the main analysis indicator from the causal knowledge graph, and determine the recommended associated nodes based on the causal strength. The calculation and inference module is used to calculate the intervention effect of the main analysis indicator based on the main analysis indicator and the recommended related nodes, and generate counterfactual inference result data. The confidence assessment module is used to perform multi-dimensional confidence assessment on the counterfactual inference results data to obtain an attribution path credibility score.

8. An attribution analysis device, characterized in that, The device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the attribution analysis method as described in any one of claims 1 to 6.

9. A storage medium, characterized in that, The storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, it implements the steps of the attribution analysis method as described in any one of claims 1 to 6.

10. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the steps of the attribution analysis method as described in any one of claims 1 to 6.

Images