A method, medium and device for attribution analysis based on a BI platform

Abstract

This invention discloses an attribution analysis method, medium, and device based on a BI platform. It collects multi-dimensional business indicator data and related indicator datasets, employs an adaptive sliding window algorithm to detect abnormal fluctuations, and calculates the absolute growth contribution using an improved contribution rate decomposition model; finally, it generates a composite influence matrix. This technical solution overcomes limitations on chart types, achieving attribution adaptation for most mainstream charts; it eliminates the risk of zero denominator through an improved contribution rate model, and accurately identifies the dimension of percentage mutations using a JS divergence algorithm; through an automated attribution graph model and rule engine, it shortens manual operation time and eliminates the risk of data leakage; the constructed unified data query structure significantly improves system scalability, meeting the real-time analysis needs of complex business scenarios.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and in particular to an attribution analysis method, medium, and device based on a BI platform. Background Technology

[0002] Current attribution analysis techniques in the BI field suffer from significant limitations: In terms of application scenarios, existing solutions (such as time series attribution analysis and traditional contribution rate models) are only compatible with localized time trend charts, failing to adapt to mainstream visualization charts like bar charts and radar charts. Furthermore, their coverage of business scenarios is limited, lacking multi-dimensional cross-analysis and target attribution capabilities. At the algorithm level, traditional contribution rate models are prone to computational failures due to zero denominators and cannot identify dimensions with small but rapidly changing proportions, leading to the omission of key business dimensions. In terms of the analysis process, they heavily rely on manual operation, resulting in inefficiency and data leakage risks. In terms of system architecture, their strong binding to chart types leads to poor scalability, and multiple interface calls cause performance bottlenecks. These shortcomings severely restrict the practical application value of attribution analysis in complex business scenarios. Summary of the Invention

[0003] In view of this, the purpose of this invention is to propose an attribution analysis method, medium, and device based on a BI platform.

[0004] To achieve the above-mentioned technical objectives, the technical solution adopted by this invention is as follows:

[0005] An attribution analysis method based on a BI platform includes:

[0006] Collect multi-dimensional business indicator data and related indicator datasets from the BI platform. The related indicator dataset includes multiple related indicators.

[0007] Fluctuation features are extracted from multi-dimensional business indicator data, an adaptive sliding window algorithm is used to detect abnormal fluctuation ranges of indicators, and the absolute growth contribution of multi-dimensional business indicator data within the target time window is calculated based on the improved contribution rate decomposition model.

[0008] Cross-influence analysis is performed on the dataset of related indicators to calculate the transmission contribution weight of related indicators to the main indicator and generate a composite influence matrix that includes direct and indirect transmission contributions.

[0009] By integrating absolute growth contribution and composite impact matrix, a root cause identification algorithm based on JS divergence is used to quantify the degree of mutation in the proportion of dimension values, and key attribution dimensions are screened by combining contribution ranking and surprise threshold.

[0010] Construct an interactive attribution mapping model based on key attribution dimensions, including:

[0011] The key attribution dimensions are mapped to a directed weighted network according to their contribution intensity, generating potential correlation vectors between the key attribution dimensions.

[0012] Based on the attribution graph model, the pre-set attribution rule engine automatically triggers combined dimensional analysis to generate a dynamic analysis report. The dynamic analysis report includes a dimensional contribution heatmap, an indicator transmission path diagram, and a multi-dimensional drill-down interface.

[0013] In some embodiments, an adaptive sliding window algorithm is used to detect abnormal fluctuation ranges of indicators, and an improved contribution rate decomposition model is used to calculate the absolute growth contribution of multi-dimensional business indicator data within the target time window, including:

[0014] Input the fluctuation feature vector into the dynamic programming optimization module, and automatically adjust the sliding window boundary according to the distribution of local extreme points of the fluctuation feature vector to generate an analysis interval containing the complete fluctuation cycle.

[0015] A dual-channel fluctuation detection algorithm is constructed to simultaneously calculate the absolute change and relative rate of change of the index within the sliding window in the analysis interval. When the change and relative rate of change both exceed the preset linkage threshold, the effective fluctuation interval is marked.

[0016] A dimension weight adjustment factor is added to the improved contribution rate decomposition model. The dimension weight adjustment factor is obtained by weighted fusion calculation of the business importance and historical fluctuation stability of the dimension value.

[0017] The original index sequence is randomly perturbed and then fed into the improved contribution rate decomposition model to output a stable key contribution dimension.

[0018] A contribution propagation tree model is established based on the key contribution dimensions. The contribution accumulation algorithm on the tree structure is used to calculate the direct contribution value of the leaf dimension and the aggregated contribution value of the leaf parent dimension.

[0019] The absolute growth contribution is calculated based on the direct contribution value and the aggregated contribution value.

[0020] In some embodiments, cross-influence analysis is performed on the correlated indicator dataset. Asynchronous fluctuating correlated indicators are aligned using a dynamic time warping algorithm, and the transmission contribution weights of correlated indicators to the main indicator are calculated, including:

[0021] The dynamic correlation coefficients between each related indicator and the main indicator are calculated based on time series data and used as the initial transmission weights.

[0022] The initial transmission weights are adjusted based on the information entropy value of the fluctuation sequence of the associated indicators to obtain the adjusted initial transmission weights, which are denoted as the processed transmission weights.

[0023] Construct a transmission path evaluation matrix based on the processed transmission weights, and output the key transmission paths with statistical significance and their contribution weight values.

[0024] In some embodiments, generating a composite influence matrix that includes direct contributions and indirect transmission contributions includes:

[0025] Construct a transmission contribution network based on key transmission paths and their contribution weight values;

[0026] By traversing the effective propagation paths, the path propagation strength is obtained by multiplying the edge weights on each effective propagation path, and the propagation contribution network is updated based on the path propagation strength.

[0027] The contribution values ​​in the updated transmission contribution network are corrected by using a preset attenuation coefficient and then summed to obtain the cumulative transmission influence value of each related indicator.

[0028] By decomposing the time series data of the correlation index and the main index through wavelet transform, the correlation coefficient of the same frequency band is extracted as the direct correlation strength. Combined with the trend component and periodic component separated by Hodrick-Prescott filtering, the variance explanation ratio of the periodic component is calculated as the direct contribution weight.

[0029] An initial matrix is ​​constructed based on the cumulative transmission impact value and the direct contribution weight, and then linearly combined according to the dynamic adjustment parameter, which is calculated by the sigmoid function based on the time lag between indicators.

[0030] Perform SVD decomposition on the initial matrix and extract the first k singular vectors to generate a composite influence matrix.

[0031] In some embodiments, the absolute growth contribution and the composite impact matrix are integrated, and a root cause identification algorithm based on JS divergence is used to quantify the degree of change in the proportion of dimension values, including:

[0032] The absolute growth contribution and the composite influence matrix are tensor-stitched together to construct a three-dimensional analysis cube. The first dimension is the time period, the second dimension is the business dimension, and the third dimension includes the direct contribution value, the transmitted contribution value, and the absolute growth amount.

[0033] Calculate the rate of change of contribution distribution of each business dimension in the composite influence matrix, and use JS divergence to measure the degree of difference between the current contribution distribution and the historical baseline distribution. The historical baseline distribution is the median of the contribution distribution of the most recent N normal periods.

[0034] The degree of mutation is obtained by weighting and fusing the JS divergence value with the absolute growth contribution.

[0035] Perform topological sorting on dimensions whose mutation degree exceeds a preset mutation threshold, and output the mutation results, which include the mutation degree of the dimension value.

[0036] In some embodiments, key attribution dimensions are selected by combining contribution ranking and surprise threshold, including:

[0037] The comprehensive impact score is generated by performing matrix multiplication on the degree of change in the proportion of dimensional values ​​and the absolute growth contribution.

[0038] Based on the percentile distribution of the comprehensive impact scores of each dimension within the historical period, the P95 percentile is taken as the basic threshold, and a seasonal adjustment factor is introduced for dynamic correction.

[0039] Dimensions whose comprehensive impact score exceeds the preset surprise threshold are recorded as candidate dimensions, and a second screening is performed based on the degree centrality of their transmission path network to retain a number of dimensions with the highest network influence ranking.

[0040] Each dimension is accompanied by a quadruple feature description, which includes the overall impact score, mutation degree, transmission path depth, and number of downstream impact dimensions.

[0041] In a second aspect, the present invention also provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the method described in the first aspect.

[0042] In a third aspect, the present invention also provides an electronic device including a memory and a processor, the memory being used to store one or more computer program instructions, wherein the one or more computer program instructions are executed by the processor to implement the method described in the first aspect.

[0043] By adopting the above technical solution, the present invention has the following beneficial effects compared with the prior art:

[0044] By collecting multi-dimensional business indicator data and related indicator datasets, an adaptive sliding window algorithm is used to detect abnormal fluctuations, and an improved contribution rate decomposition model is combined to calculate the absolute growth contribution; a composite influence matrix is ​​then generated. This technical solution can overcome the limitations of chart types and achieve attribution adaptation for most mainstream charts; the improved contribution rate model eliminates the risk of zero denominator, and the JS divergence algorithm accurately identifies the dimension of percentage mutation; through an automated attribution graph model and rule engine, manual operation time is reduced and the risk of data leakage is eliminated; the constructed unified data query structure significantly improves the system's scalability and meets the real-time analysis needs of complex business scenarios. Attached Figure Description

[0045] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 This is a schematic diagram of steps S101 to S106 of the attribution analysis method described in the specific implementation.

[0047] Figure 2 This is a schematic diagram of steps S201 to S206 of the attribution analysis method described in the specific implementation. Detailed Implementation

[0048] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be particularly noted that the following embodiments are for illustrative purposes only and do not limit the scope of the invention. Similarly, the following embodiments are only some, not all, embodiments of the present invention, and all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0049] Please see Figure 1 This embodiment provides an attribution analysis method based on a BI platform, including:

[0050] S101. Collect multi-dimensional business indicator data and related indicator datasets from the BI platform. The related indicator dataset includes multiple related indicators.

[0051] S102. Extract fluctuation features from multi-dimensional business indicator data, use an adaptive sliding window algorithm to detect abnormal fluctuation ranges of indicators, and calculate the absolute growth contribution of multi-dimensional business indicator data within the target time window based on the improved contribution rate decomposition model.

[0052] S103. Perform cross-influence analysis on the dataset of related indicators, calculate the transmission contribution weight of related indicators to the main indicator, and generate a composite influence matrix that includes direct and indirect transmission contributions.

[0053] S104. Integrate absolute growth contribution and composite impact matrix, use JS divergence-based root cause identification algorithm to quantify the degree of change in the proportion of dimension values, and combine contribution ranking and surprise threshold to screen key attribution dimensions.

[0054] S105. Construct an interactive attribution mapping model based on key attribution dimensions, including:

[0055] The key attribution dimensions are mapped to a directed weighted network according to their contribution intensity, generating potential correlation vectors between the key attribution dimensions.

[0056] S106. Based on the attribution graph model, the combined dimensional analysis is automatically triggered by the preset attribution rule engine to generate a dynamic analysis report. The dynamic analysis report includes a dimensional contribution heatmap, an indicator transmission path diagram, and a multi-dimensional drill-down interface.

[0057] In step S101, multi-dimensional business indicator data and related indicator datasets are collected from the BI platform. A standardized data model is set up, including key fields such as filters, indicators, and analysis dimensions, compatible with dozens of chart types in the BI system, including mainstream visualization charts such as bar charts, radar charts, and trend charts. The related indicator dataset contains multiple related indicators, such as clearance sales revenue and procurement rate, which influence other revenues of the main indicator, providing a data foundation for subsequent cross-influence analysis.

[0058] In step S102, fluctuation features are extracted from the multi-dimensional business indicator data, and an adaptive sliding window algorithm is used to detect abnormal fluctuation ranges of the indicators. Based on the improved contribution rate decomposition model, the absolute growth contribution of the multi-dimensional business indicator data within the target time window is calculated. The improved contribution rate algorithm changes the denominator from the total growth value of the indicators to the sum of the absolute values ​​of the dimension growth values, avoiding the calculation failure problem when the denominator is zero. In some preferred embodiments, when the total growth value of the indicators is greater than zero, the dimension value contribution rate is equal to the dimension value growth value divided by the total dimension growth value multiplied by 100%; when the total growth value of the indicators is less than or equal to zero, the dimension value contribution rate is equal to the dimension value growth value divided by the absolute value of the total dimension growth value multiplied by -100%.

[0059] In step S103, a cross-influence analysis is performed on the related indicator dataset to calculate the transmission contribution weight of the related indicators to the main indicator, generating a composite influence matrix that includes both direct and indirect transmission contributions. This solves the problem that traditional techniques can only analyze the influence of dimensions on the main indicator. For example, quantifiable clearance revenue contributes 68.35% to other revenue, with a growth value of 585,599 yuan and a growth rate of 0.78%.

[0060] In step S104, the absolute growth contribution and the composite impact matrix are integrated, and a root cause identification algorithm based on JS divergence is used to quantify the degree of change in the proportion of dimension values. Key attribution dimensions are then selected by combining contribution ranking and surprise threshold. By calculating the JS entropy value of the current proportion of dimension values ​​and the proportion during the comparison period, abnormal dimensions with small volume but significant changes in proportion are identified, such as the case where the proportion of users in a certain city increases from 1% to 5%.

[0061] In step S105, an interactive attribution graph model is constructed based on the key attribution dimensions. The key attribution dimensions are mapped to a directed weighted network according to their contribution intensity, generating potential correlation vectors between the key attribution dimensions. The attribution graph model supports dimension drill-down functionality; for example, clicking on the city dimension allows viewing the contribution of subdivided financial categories. It also supports combined dimension analysis, such as cross-attribution by city multiplied by supplier multiplied by financial category.

[0062] In step S106, based on the attribution graph model, the combined dimensional analysis is automatically triggered through a preset attribution rule engine to generate a dynamic analysis report. The dynamic analysis report includes a dimensional contribution heatmap, an indicator transmission path diagram, and a multi-dimensional drill-down interface, replacing the cumbersome process of manually configuring multi-dimensional linked charts in the traditional method, reducing the operation time from 10 minutes to 1 second.

[0063] This embodiment achieves compatibility with most chart types through a unified attribution query structure design, improves the contribution rate algorithm to eliminate the risk of zero denominator, accurately identifies abnormal dimensions by combining the JS divergence algorithm, and constructs an automated attribution graph model to generate dynamic analysis reports. It solves the problems of poor chart compatibility, algorithm model defects, and low efficiency of manual operation in traditional technologies, realizing fully automated analysis and deep interactive capabilities, significantly improving the accuracy and practicality of attribution analysis.

[0064] Please see Figure 2 In some embodiments, an adaptive sliding window algorithm is used to detect abnormal fluctuation ranges in indicators, and an improved contribution rate decomposition model is used to calculate the absolute growth contribution of multi-dimensional business indicator data within the target time window, including:

[0065] S201. Input the extracted fluctuation feature vector into the dynamic programming optimization module. The sliding window boundary is automatically adjusted according to the distribution of local extreme points of the fluctuation feature vector to generate an analysis interval containing the complete fluctuation cycle.

[0066] S202. Construct a dual-channel fluctuation detection algorithm to simultaneously calculate the absolute change and relative change rate of the index within the sliding window in the analysis interval. When the change and relative change rate both exceed the preset linkage threshold, output the effective fluctuation interval marker.

[0067] S203. Add a dimension weight adjustment factor to the improved contribution rate decomposition model. The dimension weight adjustment factor is obtained by weighted fusion calculation of the business importance and historical fluctuation stability of the dimension value.

[0068] S204. Randomly perturb the original index sequence and input it into the improved contribution rate decomposition model to output a stable key contribution dimension.

[0069] S205. Establish a contribution propagation tree model based on the key contribution dimensions, and calculate the direct contribution value of the leaf dimension and the aggregated contribution value of the leaf parent dimension through the contribution accumulation algorithm on the tree structure.

[0070] S206. The absolute growth contribution is obtained based on the direct contribution value and the aggregated contribution value.

[0071] In step S201, the fluctuation feature vector is input into the dynamic programming optimization module, which automatically adjusts the sliding window boundary based on the distribution of local extrema of the fluctuation feature vector. The dynamic programming optimization module ensures that each sliding window contains a complete fluctuation cycle (e.g., from trough to trough) by finding the optimal segmentation path between extrema. In some preferred embodiments, a dynamic time warping cost matrix is ​​used to evaluate different window partitioning schemes, selecting the boundary partitioning method that maximizes the consistency of fluctuation patterns within the window, thus avoiding the cycle truncation problem caused by traditional fixed windows.

[0072] In step S202, a dual-channel fluctuation detection algorithm is constructed to simultaneously calculate the absolute change and relative rate of change of the indicator within the sliding window. The absolute change is calculated by the difference between the values ​​at the beginning and end of the window, while the relative rate of change is determined by the ratio of the integral area within the window to the time span. When both exceed a preset linkage threshold (e.g., absolute change > 500 units and relative rate of change > 15%), it is determined to be a valid fluctuation range. This dual-channel mechanism can filter out false alarms from a single indicator, such as small high-frequency oscillations or monotonous but gentle trend changes.

[0073] In step S203, a dimension weight adjustment factor is added to the improved contribution rate decomposition model. This factor is a weighted fusion of two sub-factors: a business importance factor (based on the priority configuration of the dimension in the business rules, with a value of 0.1-1.0) and a historical fluctuation stability factor (calculated by normalizing the inverse of the coefficient of variation over the past 12 periods). For example, if a city dimension is assigned an importance weight of 0.9 due to undertaking core business, and its historical fluctuation stability coefficient is 0.8, then the final adjustment factor is 0.9×0.6+0.8×0.4=0.86.

[0074] In step S204, the original index sequence is randomly perturbed and input into the improved contribution rate decomposition model. 100 sets of perturbed data with added Gaussian noise (σ=0.5%) are generated through Monte Carlo simulation, and the contribution distribution of each dimension is calculated. The final output is the key contribution dimension with a stable ranking within a 95% confidence interval, eliminating the problem of jumps in contribution ranking caused by minor data fluctuations.

[0075] In step S205, a contribution propagation tree model is established based on key contribution dimensions. The tree structure is constructed hierarchically according to business dimensions (e.g., country → province → city). The direct contribution value of the leaf dimensions is calculated using an improved contribution rate formula, while the aggregated contribution value of the parent dimension of the leaf is obtained by weighted summation of the contributions of its child nodes, with the weights dynamically adjusted based on the proportion of business volume of the child nodes. The propagation tree model can clearly show the cross-level contribution transmission path, for example, identifying that the negative contribution of a certain province mainly comes from the three cities under its jurisdiction.

[0076] In step S206, the absolute growth contribution is calculated based on the direct contribution value and the aggregated contribution value. For leaf dimensions, their direct contribution value is used directly; for non-leaf dimensions, their aggregated contribution value is decomposed into each leaf dimension according to the proportion of the business volume of the child nodes and then accumulated. The final output absolute growth contribution includes both the original contribution value and the normalized value after adjustment by the tree structure, supporting both "drill-down" and "summarize-up" analysis perspectives.

[0077] This embodiment significantly improves the accuracy of abnormal interval identification through dynamically optimized sliding window partitioning and dual-channel fluctuation detection; the dimensional weight adjustment factor and random perturbation verification mechanism enhance the business rationality and numerical stability of contribution calculation. The introduction of the contribution propagation tree model achieves, for the first time, the visualized quantification of cross-level contribution transmission in attribution analysis, providing an innovative solution for root cause localization in complex business scenarios.

[0078] In some embodiments, cross-influence analysis is performed on the dataset of related indicators to calculate the transmission contribution weight of the related indicators to the main indicator, including:

[0079] The dynamic correlation coefficients between each related indicator and the main indicator are calculated based on time series data and used as the initial transmission weights.

[0080] The initial transmission weights are adjusted based on the information entropy value of the fluctuation sequence of the associated indicators to obtain the adjusted initial transmission weights, which are denoted as the processed transmission weights.

[0081] Construct a transmission path evaluation matrix based on the processed transmission weights, and output the key transmission paths with statistical significance and their contribution weight values.

[0082] In this embodiment, the dynamic correlation coefficients between each related indicator and the main indicator are calculated based on time series data. A sliding window Pearson correlation coefficient is used as the initial propagation weight, and the window size is automatically adapted according to the characteristics of the business cycle. For indicators with obvious seasonality, the window is set to a complete seasonal cycle; for non-periodic indicators, the autocorrelation function is used to determine the optimal window. During the calculation of the dynamic correlation coefficients, the data within each window is standardized to eliminate the influence of unit dimensions.

[0083] In this embodiment, the initial transmission weights are adjusted based on the information entropy value of the associated index fluctuation sequence. Information entropy is calculated using probability distribution estimation within a sliding window, reflecting the uncertainty of the fluctuation pattern. For high-entropy sequences, it indicates that their fluctuation pattern is complex and variable, requiring a reduction in their transmission weights; conversely, low-entropy sequences indicate stable and reliable fluctuation patterns, allowing for an appropriate increase in weights. The adjustment formula uses the normalized result of the reciprocal of the entropy value as the weight adjustment coefficient, ensuring that the processed transmission weights are between 0 and 1.

[0084] In this embodiment, a transmission path evaluation matrix is ​​constructed based on the processing transmission weights. The transmission path evaluation matrix is ​​a symmetric matrix, with rows and columns representing related indicators and the main indicator, respectively, and matrix elements representing processing transmission weights. The statistical significance of each path is evaluated using a permutation test, retaining key transmission paths with p-values ​​less than 0.05. The final output contribution weight values ​​are then normalized using Softmax, representing the relative influence of each related indicator on the fluctuation of the main indicator.

[0085] This embodiment enhances the model's adaptability to complex fluctuation patterns by calculating transmission weights using information entropy weighting, while statistical significance testing ensures the reliability of key transmission paths. It provides a scientific method for quantifying the transmission contribution of indicators, particularly suitable for business scenarios with time delay effects, significantly improving the accuracy of impact analysis of related indicators.

[0086] In some embodiments, generating a composite influence matrix that includes direct contributions and indirect transmission contributions includes:

[0087] Construct a transmission contribution network based on key transmission paths and their contribution weight values;

[0088] By traversing the effective propagation paths, the path propagation strength is obtained by multiplying the edge weights on each effective propagation path, and the propagation contribution network is updated based on the path propagation strength.

[0089] The contribution values ​​in the updated transmission contribution network are corrected by using a preset attenuation coefficient and then summed to obtain the cumulative transmission influence value of each related indicator.

[0090] By decomposing the time series data of the correlation index and the main index through wavelet transform, the correlation coefficient of the same frequency band is extracted as the direct correlation strength. Combined with the trend component and periodic component separated by Hodrick-Prescott filtering, the variance explanation ratio of the periodic component is calculated as the direct contribution weight.

[0091] An initial matrix is ​​constructed based on the cumulative transmission impact value and the direct contribution weight, and then linearly combined according to the dynamic adjustment parameter, which is calculated by the sigmoid function based on the time lag between indicators.

[0092] Perform SVD decomposition on the initial matrix and extract the first k singular vectors to generate a composite influence matrix.

[0093] In this embodiment, a transmission contribution network is constructed based on key transmission paths and their contribution weights. The transmission contribution network is a directed weighted graph structure, where nodes represent indicators and edge weights represent direct transmission contribution values. Optionally, in some preferred embodiments, an adjacency matrix is ​​used to store the network topology, with matrix elements... Indicators For indicators The weight of the weak connection is directly affected by the network strength. During network construction, weak connections with weights below 0.1 are automatically filtered to ensure the saliency of the network structure.

[0094] In this embodiment, the path propagation strength is obtained by traversing the effective propagation paths and multiplying the edge weights of each effective propagation path. The path search employs a depth-first search algorithm, with a maximum path length limited to 5 hops to avoid computational complexity explosion caused by excessively long propagation chains. For each discovered path, its propagation strength is equal to the product of all edge weights on the path; this value reflects the attenuation characteristics of indirect influence. The path propagation strength is proportionally distributed to the terminal nodes of the path, updating the corresponding edge weights of the propagation contribution network.

[0095] In this embodiment, the contribution values ​​in the updated conduction contribution network are corrected using a preset attenuation coefficient and then accumulated. The attenuation coefficient is set as an exponential function of the path length, specifically... ,in Take an empirical value between 0.6 and 0.8. The transmission path length is denoted as . After correction by the attenuation coefficient, the cumulative transmission impact value of each related index is the weighted sum of the correction intensities of all related paths, reflecting the superposition effect of multi-hop transmission.

[0096] In this embodiment, wavelet transform is used to decompose the time-series data of the correlation index and the main index. A five-level decomposition is performed using the Daubechies wavelet basis function, and the frequency bands corresponding to the detail coefficients of the third level are extracted to calculate the correlation coefficient as the direct correlation strength. Simultaneously, Hodrick-Prescott filtering is applied to decompose the sequence into trend and periodic components, and the proportion of the periodic component variance to the overall variance is calculated as the direct contribution weight. This proportion reflects the explanatory power of the index fluctuations on the periodic changes of the main index.

[0097] In this embodiment, an initial matrix is ​​constructed based on the cumulative transmission impact value and the direct contribution weight. The matrix rows represent related indicators, and the columns contain two feature dimensions: transmission impact value and direct contribution weight. The dynamic adjustment parameter is calculated using the Sigmoid function, with the input being the time lag length between indicators and the output being an adjustment coefficient between 0 and 1. The linear combination formula is: Composite Weight = α × Transmission Impact + (1-α) × Direct Contribution, where α is the dynamic adjustment parameter, achieving time-lag-sensitive adaptive fusion.

[0098] In this embodiment, SVD decomposition is performed on the initial matrix to extract the first k singular vectors. The value of k is determined based on the cumulative contribution rate of the eigenvalues, retaining dimensions that achieve a cumulative contribution rate of 85% or higher. The generated composite influence matrix represents the indicator feature pattern in the row space and the influence type distribution in the column space. The matrix element values ​​comprehensively reflect the influence strength of each indicator on the main indicator through different pathways.

[0099] This embodiment comprehensively captures the combined effects of direct and indirect transmission between indicators through dual calculations of transmission networks and direct contributions. Wavelet analysis and trend decomposition enhance the ability to distinguish effects at different time scales, while SVD dimensionality reduction extracts key influence patterns. It solves the problem of traditional methods' difficulty in quantifying indirect effects, provides an interpretable mathematical model for multi-level attribution analysis in complex systems, and significantly improves the comprehensiveness and accuracy of attributing fluctuations in business indicators.

[0100] In some embodiments, the absolute growth contribution and the composite impact matrix are integrated, and a root cause identification algorithm based on JS divergence is used to quantify the degree of change in the proportion of dimension values, including:

[0101] The absolute growth contribution and the composite influence matrix are tensor-stitched together to construct a three-dimensional analysis cube. The first dimension is the time period, the second dimension is the business dimension, and the third dimension includes the direct contribution value, the transmitted contribution value, and the absolute growth amount.

[0102] Calculate the rate of change of contribution distribution of each business dimension in the composite influence matrix, and use JS divergence to measure the degree of difference between the current contribution distribution and the historical baseline distribution. The historical baseline distribution is the median of the contribution distribution of the most recent N normal periods.

[0103] The degree of mutation is obtained by weighting and fusing the JS divergence value with the absolute growth contribution.

[0104] Perform topological sorting on dimensions whose mutation degree exceeds a preset mutation threshold, and output the mutation results, which include the degree of mutation as a percentage of the dimension value.

[0105] In this embodiment, the absolute growth contribution and the composite influence matrix are tensor-concatenated to construct a three-dimensional analysis cube. The first dimension of this cube is a time period slice, supporting rolling time window analysis; the second dimension is the business dimension hierarchy, including analysis dimensions such as geographical region and product category; the third dimension is the indicator features, integrating three key features: direct contribution value, transmitted contribution value, and absolute growth amount. Optionally, in some preferred embodiments, the Z-score normalization method is used to normalize the three-dimensional data to eliminate the dimensional differences between different features.

[0106] In this embodiment, the rate of change of contribution distribution for each business dimension in the composite influence matrix is ​​calculated. For each time period, the distribution of each dimension in two feature dimensions—direct contribution value and transmitted contribution value—is extracted to form the current contribution distribution vector. The historical baseline distribution is dynamically updated through a sliding window mechanism, taking the median of the contribution distribution during non-abnormal periods in the most recent 12 periods (or adjusted according to business characteristics). The rate of change is calculated using the element-by-element relative difference method, preserving information on the direction of change.

[0107] In this embodiment, JS divergence is used to measure the degree of difference between the current contribution distribution and the historical baseline distribution. Before calculating the JS divergence, the contribution distribution vector is smoothed, and small values ​​are added to avoid the zero-probability problem. Optionally, in some preferred embodiments, JS divergence is calculated separately for the direct contribution distribution and the transmitted contribution distribution, and then linearly combined using fixed weights of 0.6 and 0.4 to reflect the sensitivity of the difference between the two types of influence.

[0108] In this embodiment, the JS divergence value and the absolute growth contribution are weighted and fused. The absolute growth contribution is transformed to the 0-1 range using the Sigmoid function, and the fusion weight with the JS divergence is dynamically configured according to the business scenario: for scenarios with high stability requirements, the JS divergence weight is set to 0.7; for growth-oriented scenarios, the absolute growth contribution weight is set to 0.6. The fusion formula is: mutation degree = w × JS + (1-w) × absolute contribution, where w is a configurable weight parameter.

[0109] In this embodiment, dimensions whose mutation severity exceeds a preset mutation threshold are topologically sorted. The mutation threshold is determined by the 95th quantile of historical data distribution and can be manually fine-tuned. The sorting algorithm comprehensively considers the magnitude and persistence of mutation severity, assigning a 1.2-fold sorting weight to dimensions that exceed the threshold for three consecutive periods. The final output mutation result includes the dimension name, mutation severity value, direction of change (positive / negative), and main impact type (direct / transductive).

[0110] This embodiment integrates the spatiotemporal characteristics of multi-dimensional indicators by constructing a three-dimensional analysis cube; the fusion measurement of JS divergence and absolute contribution captures both the changes in distribution pattern and the impact of absolute magnitude; the dynamic threshold mechanism and topological sorting ensure the accuracy and interpretability of root cause identification. It effectively solves the problem of insufficient detection of latent anomalies in traditional methods, accurately identifying key dimensions with significant changes in proportion but small absolute magnitudes, providing more comprehensive root cause analysis support for business decisions.

[0111] In some embodiments, key attribution dimensions are selected by combining contribution ranking and surprise threshold, including:

[0112] The comprehensive impact score is generated by performing matrix multiplication on the degree of change in the proportion of dimensional values ​​and the absolute growth contribution.

[0113] Based on the percentile distribution of the comprehensive impact scores of each dimension within the historical period, the P95 percentile is taken as the basic threshold, and a seasonal adjustment factor is introduced for dynamic correction.

[0114] Dimensions whose comprehensive impact score exceeds the preset surprise threshold are recorded as candidate dimensions, and a second screening is performed based on the degree centrality of their transmission path network to retain a number of dimensions with the highest network influence ranking.

[0115] Each dimension is accompanied by a quadruple feature description, which includes the overall impact score, mutation degree, transmission path depth, and number of downstream impact dimensions.

[0116] In this embodiment, a matrix multiplication operation is performed on the degree of change in the proportion of dimensional values ​​and the absolute growth contribution to generate a comprehensive impact score. Specifically, the rows of the matrix representing the degree of change in the proportion of dimensional values ​​represent business dimensions, and the columns represent time windows. Its element values ​​are the fusion result of JS divergence and absolute growth contribution. The absolute growth contribution matrix is ​​calculated using an improved contribution rate algorithm to ensure effective calculation even when the total growth value of the indicator is zero. Optionally, in some preferred embodiments, the matrix multiplication uses Hadamard product (element-by-element multiplication) instead of standard matrix multiplication to preserve the independent characteristics of each time window.

[0117] In this embodiment, the P95 quantile is used as the base threshold based on the percentile distribution of the comprehensive impact scores for each dimension within the historical period. The historical period is selected by default as the most recent 12 business periods, and the baseline data is dynamically updated through a sliding window mechanism. Optionally, in some preferred embodiments, for business indicators with obvious seasonality, the STL decomposition algorithm is used to extract the seasonal components, and the window length for percentile calculation is automatically adjusted based on the seasonality intensity.

[0118] In this embodiment, dimensions whose overall impact score exceeds a preset surprise threshold are designated as candidate dimensions. The surprise threshold is a dynamically adjusted factor superimposed on the base P95 threshold, and this factor is determined by both the indicator volatility and the business importance coefficient. Optionally, in some preferred embodiments, for KPI-type core indicators, the threshold sensitivity is automatically increased by 20% to reduce the false positive rate.

[0119] In this embodiment, a secondary screening is performed using the degree centrality of the transmission path network. The degree centrality is calculated by weighting the sum of the in-degree and out-degree, and the edge weights are taken from the transmission contribution values ​​of the composite influence matrix. Optionally, in some preferred embodiments, the network influence ranking retains the top 5 dimensions, but when there are parallel dimensions with a degree centrality difference of less than 15%, the ranking is expanded to a maximum of 8 dimensions to ensure critical path coverage.

[0120] In this embodiment, each dimension is accompanied by a quadruple feature description. The overall impact score is mapped to the 0-100 range through Min-Max normalization; the mutation degree retains the original JS divergence value; the transmission path depth is taken as the maximum number of hops in all transmission paths of that dimension; the downstream impact dimension count is the total number of dimensions directly or indirectly affected. Optionally, in some preferred embodiments, for dimensions with a transmission path depth exceeding 3, a second-level drill-down analysis is automatically triggered to verify the reliability of the transmission chain.

[0121] This embodiment integrates the degree of mutation and absolute contribution through matrix operations, overcoming the limitations of traditional methods that rely on only a single indicator. A dynamic threshold mechanism combined with network topology analysis ensures stable identification of the true root cause amidst business fluctuations. The four-tuple feature system provides multi-dimensional evaluation criteria, supporting a comprehensive assessment from statistical significance to network influence. This significantly improves the accuracy and interpretability of attribution analysis in complex business scenarios, and is particularly suitable for analyzing indicator fluctuations with multi-dimensional cross-influences and transmission effects.

[0122] To facilitate understanding, the following examples are provided for further application:

[0123] The existing technology has the following problems:

[0124] The application scenarios are severely limited, and the coverage is narrow, specifically in the following ways:

[0125] Poor chart type compatibility: It only supports time trend charts (accounting for about 15%-20% of report charts) and cannot adapt to mainstream visualization charts such as bar charts, radar charts, and number trend charts. As a result, more than 80% of the report data cannot be directly used for attribution analysis, which greatly reduces its practical application value.

[0126] Insufficient coverage of business scenarios: It can only handle the analysis of time-dimensional indicator fluctuations, and cannot meet the needs of multi-dimensional (such as city, supplier, financial category) cross-analysis or target attribution (such as KPI achievement rate analysis), making it difficult to adapt to complex business scenarios.

[0127] The algorithm model has fundamental flaws and insufficient computational power, specifically manifested in the following ways:

[0128] Risk of contribution rate algorithm failure: The general formula "Dimension value contribution rate = Dimension value growth / Total indicator growth" cannot be calculated when the total indicator growth is 0 (e.g., positive and negative growth cancel each other out). This may miss key anomalies (e.g., significant fluctuations within each dimension but overall flatness). It requires calling two backend interfaces: the dimension module and the overall indicator module. Query performance is limited by the interface response speed, and delays are likely to occur with large datasets.

[0129] Weak root cause identification capability: Traditional algorithms rely solely on absolute growth values ​​for sorting, failing to identify dimensions that are "small in size but have significant changes in proportion" (e.g., the proportion of users in a certain city increases from 1% to 5%, but the absolute growth value is lower than that of other cities), resulting in key dimensions that the business focuses on being ignored.

[0130] The analysis process relies on manual operation, which is inefficient and risky, specifically in the following ways:

[0131] Manual processing is costly: When indicators are abnormal, the dashboard configurator needs to manually configure multi-dimensional linked charts or export detailed data, which takes more than 10 minutes for users. In addition, periodic reports need to be processed repeatedly, resulting in a huge investment of manpower.

[0132] Data error and leakage risks: Data calculation errors are prone to occur during manual processing (such as incorrect formula references or dimension matching deviations), and offline data transmission may lead to the leakage of sensitive information, making security impossible to guarantee.

[0133] High dependence on configurators: Non-technical users cannot complete attribution analysis independently and need to rely on the support of the configurator on the dashboard. The process is cumbersome and the response speed is slow, making it difficult to meet the real-time analysis needs of the business.

[0134] The report is simplistic in format and lacks in-depth analytical capabilities, specifically manifested in the following ways:

[0135] Lack of modular content: Existing technology cannot automatically generate multi-level reports that include "report overview, dimension attribution, and indicator attribution," requiring business personnel to manually organize data, resulting in low efficiency in outputting conclusions.

[0136] Insufficient interactivity: It does not support drill-down to dimensions (such as clicking on the city dimension to view the contribution of subdivided financial categories) or combined dimension analysis (such as cross-attribution of "city × supplier"), making it difficult to uncover deep correlations in the data.

[0137] Indicator attribution gap: It can only analyze the impact of dimensions on the main indicator, and cannot break down the contribution of related indicators (such as the impact of "clearance revenue" and "purchase rate" on "other revenue"), resulting in a single analysis dimension.

[0138] The system has bottlenecks in compatibility and performance, specifically manifested in:

[0139] Poor architectural scalability: The attribution function is strongly bound to the chart type, and a unified data query structure has not been built. When adding a new chart type, the adaptation logic needs to be redeveloped, resulting in high development costs and long development cycles.

[0140] Insufficient performance stability: Multiple interface calls and complex calculation logic result in slow loading speeds under large data volumes, causing sluggish user experience and failing to meet real-time analysis needs.

[0141] This example can overcome the limitations of application scenarios and achieve compatibility with all chart types, specifically in the following ways:

[0142] Unified Attribution Query Structure Design: A standardized data model is designed, incorporating key fields such as filters, metrics, and analysis dimensions. It is compatible with dozens of chart types in BI systems (such as bar charts, radar charts, and trend charts), breaking away from the traditional reliance on time-based trend charts. The front-end obtains all chart data through a unified API interface. Adding new chart types eliminates the need for redeveloping adaptation logic, increasing the coverage of attribution functionality from 15%-20% to over 80% of chart types in reports.

[0143] Multi-scenario attribution type support: Provides two modes: fluctuation attribution (comparing data fluctuations at different times) and target attribution (analyzing the reasons for KPI achievement / failure), adapting to the daily fluctuation analysis and target management needs of the business.

[0144] This example reconstructs the algorithm model to address computational bottlenecks and root cause identification problems, specifically in the following ways:

[0145] Improved Dimension Contribution Rate Algorithm: Formula Optimization: The denominator was changed from "total growth value of indicators" to "sum of absolute values ​​of dimension growth values" to avoid the inability to calculate when the denominator is 0. The specific formula is as follows:

[0146] When the total growth value of the indicator is greater than 0, the contribution rate of the dimension value = (the growth value of the dimension value / the total growth value of the dimension value) × 100%;

[0147] When the total growth value of the indicator is ≤0, the contribution rate of the dimension value = (the growth value of the dimension value / the absolute value of the total growth value of the dimension value) × (-1) × 100%; Note: the total growth value of the dimension value = the sum of the absolute values ​​of the growth values ​​of the dimension values.

[0148] Performance improvement: Only the dimension module interface needs to be called, reducing one backend query and improving data loading efficiency.

[0149] Introducing a root cause analysis algorithm (explanatory power + surprise factor): Combining "explanatory power" (the impact of dimensional value growth on the overall fluctuation of the indicator) and "surprise factor" (the change in the proportion of dimensions calculated using JS entropy), this algorithm accurately identifies anomalous dimensions that are small in size but have significant differences in proportion. For example, by calculating the JS entropy value of the current proportion (P) of a dimension value and the proportion (Q) of the comparison period, the greater the difference between P and Q, the closer the entropy value is to 1, indicating that the change in that dimension is more noteworthy.

[0150] This example improves efficiency and reduces risk through full-process automation and modular design, specifically in the following ways:

[0151] One-click attribution report generation reduces manual operations: Users can trigger automated analysis by clicking on chart data points, generating an attribution report within 1 second, replacing the traditional manual processing workflow that takes more than 10 minutes. The report includes:

[0152] Overall data: current value, comparison value, growth value, and growth rate of the indicator;

[0153] Report Overview: Automatically summarizes the top dimensions and metrics of both positive and negative contributions;

[0154] Dimensional attribution: Ranking and details of contribution of each dimension;

[0155] Indicator attribution: Analysis of the impact of related indicators on the main indicator.

[0156] Data security and accuracy assurance: All calculations are completed within the platform, avoiding offline data transmission and reducing the risk of leakage; the algorithm model standardizes the calculation logic and eliminates human calculation errors.

[0157] This example enhances the report's interactivity and in-depth analytical capabilities, specifically in the following ways:

[0158] Dimensional drill-down and combined dimensional analysis:

[0159] Drill down to attribution: Click on any dimension value (such as "City X") to analyze the contribution of its subdivided dimensions (such as "Maternal and Infant Products" and "Fresh Food") and dig out the reasons for fluctuations layer by layer.

[0160] Combined Dimension Attribution: Supports configuration of multi-dimensional combinations such as "city × supplier × financial category" to identify cross-dimensional correlation factors (such as revenue fluctuations in the "city × supplier × fresh produce" category).

[0161] Indicator attribution module expansion:

[0162] For the main indicator (such as "other income"), analyze the contribution of related indicators (such as "clearance income" and "purchase rate"). For example, clearance income contributes 68.35% to other income, with an increase of RMB 585,599 and a growth rate of 0.78%.

[0163] This example optimizes system compatibility and performance, specifically in the following ways:

[0164] Unified architecture and caching mechanism: Seamless compatibility of chart types is achieved through attribution query structure, and data can be cached based on the structure to improve the efficiency of repeated queries.

[0165] Lightweight computing logic: Reduce the number of interface calls (from 2 to 1), optimize algorithm computational complexity, and ensure loading speed under large data volumes.

[0166] This example has the following outstanding benefits:

[0167] It is compatible with over 80% of chart types in BI systems (such as bar charts and radar charts), breaking through industry limitations; it improves the contribution rate algorithm (the denominator is the sum of the absolute values ​​of dimension growth), supporting calculations when the total growth value of the indicator is 0; it combines root cause analysis algorithms (explanatory power + surprise factor) to identify abnormal dimensions; it generates attribution reports with a single click on data points, automatically completing data calculation and analysis, reducing operation time to 1 second; it is divided into four modules: overall data, report overview, dimension attribution, and indicator attribution, supporting drill-down and combined dimension analysis; and it features a unified query structure design, reducing interface calls and improving query efficiency and compatibility.

[0168] In a second aspect, this embodiment also provides a computer-readable storage medium storing computer program instructions thereon, which, when executed by a processor, implement the method described in the first aspect.

[0169] In a third aspect, this embodiment also provides an electronic device, including a memory and a processor, the memory being used to store one or more computer program instructions, wherein the one or more computer program instructions are executed by the processor to implement the method described in the first aspect.

[0170] By collecting multi-dimensional business indicator data and related indicator datasets, an adaptive sliding window algorithm is used to detect abnormal fluctuations, and an improved contribution rate decomposition model is combined to calculate the absolute growth contribution; a composite influence matrix is ​​then generated. This technical solution can overcome the limitations of chart types and achieve attribution adaptation for most mainstream charts; the improved contribution rate model eliminates the risk of zero denominator, and the JS divergence algorithm accurately identifies the dimension of percentage mutation; through an automated attribution graph model and rule engine, manual operation time is reduced and the risk of data leakage is eliminated; the constructed unified data query structure significantly improves the system's scalability and meets the real-time analysis needs of complex business scenarios.

[0171] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0172] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods of various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0173] The above description is only a part of the embodiments of the present invention and does not limit the scope of protection of the present invention. Any equivalent device or equivalent process transformation made based on the content of the present invention specification and drawings, or direct or indirect application in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. An attribution analysis method based on a BI platform, characterized in that, include: Collect multi-dimensional business indicator data and related indicator datasets from the BI platform, wherein the related indicator dataset includes multiple related indicators; Fluctuation features are extracted from the multi-dimensional business indicator data, and an adaptive sliding window algorithm is used to detect abnormal fluctuation ranges of the indicators. Based on an improved contribution rate decomposition model, the absolute growth contribution of the multi-dimensional business indicator data within the target time window is calculated, including: The extracted fluctuation feature vectors are input into the dynamic programming optimization module, which automatically adjusts the sliding window boundary based on the distribution of local extreme points of the fluctuation feature vectors to generate an analysis interval containing the complete fluctuation cycle. A dual-channel fluctuation detection algorithm is constructed to simultaneously calculate the absolute change and relative rate of change of the index within the sliding window in the analysis interval. When the change and relative rate of change both exceed the preset linkage threshold, the effective fluctuation interval is marked. A dimension weight adjustment factor is added to the improved contribution rate decomposition model. The dimension weight adjustment factor is obtained by weighted fusion calculation of the business importance and historical fluctuation stability of the dimension value. The original index sequence is randomly perturbed and then fed into the improved contribution rate decomposition model to output a stable key contribution dimension. A contribution propagation tree model is established based on the key contribution dimensions. The contribution accumulation algorithm on the tree structure is used to calculate the direct contribution value of the leaf dimension and the aggregated contribution value of the leaf parent dimension. The absolute growth contribution is obtained based on the direct contribution value and the aggregated contribution value; Cross-influence analysis is performed on the dataset of related indicators to calculate the transmission contribution weight of the related indicators to the main indicator and generate a composite influence matrix that includes direct and indirect transmission contributions. By integrating the absolute growth contribution and the composite impact matrix, a root cause identification algorithm based on JS divergence is used to quantify the degree of mutation in the proportion of dimension values, and key attribution dimensions are screened by combining contribution ranking and surprise threshold. Construct an interactive attribution mapping model based on key attribution dimensions, including: The key attribution dimensions are mapped to a directed weighted network according to their contribution intensity, generating potential correlation vectors between the key attribution dimensions. Based on the attribution graph model, the pre-set attribution rule engine automatically triggers combined dimensional analysis to generate a dynamic analysis report, which includes a dimensional contribution heatmap, an indicator transmission path diagram, and a multi-dimensional drill-down interface.

2. The attribution analysis method based on a BI platform according to claim 1, characterized in that, Cross-influence analysis is performed on the aforementioned dataset of related indicators to calculate the transmission contribution weights of the related indicators to the main indicator, generating a composite influence matrix that includes both direct and indirect transmission contributions, including: The dynamic correlation coefficients between each related indicator and the main indicator are calculated based on time series data and used as the initial transmission weights. The initial transmission weights are adjusted based on the information entropy value of the fluctuation sequence of the associated indicators to obtain the adjusted initial transmission weights, which are denoted as the processed transmission weights. Based on the processed transmission weights, a transmission path evaluation matrix is ​​constructed, and the key transmission paths with statistical significance and their contribution weight values ​​are output.

3. The attribution analysis method based on a BI platform according to claim 2, characterized in that, Generate a composite influence matrix that includes both direct and indirect contributions, including: Construct a transmission contribution network based on the key transmission paths and their contribution weight values; By traversing the effective propagation paths, the path propagation strength is obtained by multiplying the edge weights on each effective propagation path, and the propagation contribution network is updated according to the path propagation strength. The contribution values ​​in the updated transmission contribution network are corrected by using a preset attenuation coefficient and then summed to obtain the cumulative transmission influence value of each related indicator. By decomposing the time series data of the correlation index and the main index through wavelet transform, the correlation coefficient of the same frequency band is extracted as the direct correlation strength. Combined with the trend component and periodic component separated by Hodrick-Prescott filtering, the variance explanation ratio of the periodic component is calculated as the direct contribution weight. An initial matrix is ​​constructed based on the cumulative transmission influence value and the direct contribution weight, and then linearly combined according to the dynamic adjustment parameter, which is calculated by the sigmoid function based on the time lag between indicators. Perform SVD decomposition on the initial matrix and extract the first k singular vectors to generate a composite influence matrix.

4. The attribution analysis method based on a BI platform according to claim 1, characterized in that, By integrating the absolute growth contribution and the composite impact matrix, a root cause identification algorithm based on JS divergence is used to quantify the degree of change in the proportion of dimensional values, including: The absolute growth contribution and the composite influence matrix are tensor-concatenated to construct a three-dimensional analysis cube. The first dimension is the time period, the second dimension is the business dimension, and the third dimension includes the direct contribution value, the transmitted contribution value, and the absolute growth amount. Calculate the rate of change of contribution distribution of each business dimension in the composite influence matrix, and use JS divergence to measure the degree of difference between the current contribution distribution and the historical benchmark distribution, wherein the historical benchmark distribution is the median of the contribution distribution of the most recent N normal periods; The degree of mutation is obtained by weighting and fusing the JS divergence value with the absolute growth contribution. Dimensions with mutation levels exceeding a preset mutation threshold are topologically sorted, and mutation results are output, including the mutation degree as a percentage of the dimension value.

5. The attribution analysis method based on a BI platform according to claim 4, characterized in that, Key attribution dimensions were selected by combining contribution ranking and surprise threshold, including: The degree of abrupt change in the proportion of the aforementioned dimension values ​​and the absolute growth contribution are multiplied by a matrix to generate a comprehensive impact score. Based on the percentile distribution of the comprehensive impact scores of each dimension within the historical period, the P95 percentile is taken as the basic threshold, and a seasonal adjustment factor is introduced for dynamic correction. Dimensions whose comprehensive impact score exceeds the preset surprise threshold are recorded as candidate dimensions, and a second screening is performed based on the degree centrality of their transmission path network to retain a number of dimensions with the highest network influence ranking. Each dimension is accompanied by a quadruple feature description, which includes the overall impact score, mutation degree, transmission path depth, and number of downstream impact dimensions.

6. A computer-readable storage medium storing computer program instructions thereon, characterized in that, The computer program instructions, when executed by a processor, implement the method as described in any one of claims 1 to 5.

7. An electronic device comprising a memory and a processor, characterized in that, The memory is used to store one or more computer program instructions, wherein the one or more computer program instructions are executed by the processor to implement the method as described in any one of claims 1 to 5.

Images