Large model driven KPI index management and service analysis system and method

Abstract

The invention relates to a large-model-driven KPI (Key Performance Indicator) management and service analysis system and method, and the method comprises the steps: S1, carrying out the definition of an enterprise KPI through a low-code configuration tool, and building an index value network; s2, semantic analysis is carried out on the input natural language instruction of the user, key entities and query intentions in the natural language instruction of the user are recognized, and key indexes comprise a target index and a time range; s3, index data matching is carried out according to the key entities, and correlation analysis and root cause positioning based on an index value network are carried out according to the query intention so as to identify related indexes and potential root causes; and S4, outputting an analysis report and a decision suggestion according to index data matching and correlation analysis and root cause positioning results. According to the invention, the business context can be understood and the demand can be analyzed through natural language dialogues, and more intelligent, flexible and efficient index management and analysis support can be provided for enterprise decision makers.

Description

Technical Field

[0001] The present invention relates to the technical field of enterprise business data processing, and in particular to a large model-driven KPI indicator management and business analysis system and method. Background Art

[0002] As enterprises continue to deepen their digital transformation, efficient and accurate data management and analysis have become the key to improving their core competitiveness. However, there are many pain points in the traditional KPI indicator management and analysis process:

[0003] 1) The definition of indicators and data interfaces are not unified, resulting in a lack of a standardized indicator system within the enterprise;

[0004] 2) The analysis process is complex and time-consuming, and decision makers often need data engineers and analysts to collaborate to gain key insights;

[0005] 3) When business scenarios change rapidly, the cost and difficulty of manually building and maintaining complex indicator systems increase significantly;

[0006] 4) It is difficult to locate the root cause of abnormal fluctuations in indicators, requiring multiple rounds of manual investigation and lack of intelligent auxiliary means.

[0007] 5) Traditional natural language processing and analysis frameworks (such as RNN or LSTM-based models) have problems with parallelism and insufficient global dependency modeling capabilities when processing multi-dimensional and dynamically evolving indicator data.

[0008] The emergence of the Transformer architecture brings a new paradigm for the processing of sequence and relational data. Its self-attention mechanism can simultaneously consider the dependencies of all positions in the input sequence, achieve highly parallel feature extraction, and has achieved remarkable results in machine translation, text summarization, dialogue systems, and information extraction. The large language model built with Transformer as the basic model provides technical possibilities for understanding business context and analysis requirements through natural language dialogue without complex query scripts, and can provide enterprise decision makers with more intelligent, flexible, and efficient indicator management and analysis support.

[0009] Therefore, there is an urgent need for a large model-driven KPI indicator management and business analysis system and method. Summary of the invention

[0010] In order to solve the technical problems existing in the above-mentioned prior art, the purpose of the present invention is to provide a large model-driven KPI indicator management and business analysis system and method, which can understand business context and analysis requirements through natural language dialogue, and provide enterprise decision makers with more intelligent, flexible and efficient indicator management and analysis support.

[0011] To achieve the above-mentioned purpose, the present invention provides a large model-driven KPI indicator management and business analysis method, including:

[0012] Step S1: Define the enterprise KPI indicators through the low-code configuration tool and establish the indicator value network;

[0013] Step S2: semantically parse the input user natural language instruction to identify key entities and query intent in the user natural language instruction, where the key entities include query target indicators and time range;

[0014] Step S3: matching indicator data according to the key entities, and performing correlation analysis and root cause location based on the indicator value network according to the query intent to identify relevant indicators and potential root causes;

[0015] Step S4: output an analysis report and decision-making recommendations based on the results of the indicator data matching and the results of the association analysis and root cause location.

[0016] According to a technical solution of the present invention, it also includes:

[0017] Step S5: According to the output analysis report and decision suggestions, obtain new user natural language instructions and execute steps S2 to S4.

[0018] According to a technical solution of the present invention, the indicator value network includes a plurality of nodes and a plurality of directed edges, each node represents an enterprise KPI indicator, and each directed edge represents the attribute influence relationship between the enterprise KPI indicators corresponding to the node;

[0019] The types of enterprise KPI indicators include:

[0020] Atomic indicators are basic values ​​that directly measure product or business performance;

[0021] Derived indicators are derived from the atomic indicators through derived calculation methods; the derived calculation methods include year-on-year, month-on-month, cumulative, and average values;

[0022] Composite indicators are constructed by mathematical operations on multiple basic indicators; the basic indicators are atomic indicators and / or derived indicators;

[0023] The attribute influence relationship between the enterprise KPI indicators includes positive influence, negative influence, delayed effect, influence weight, and time series characteristics.

[0024] According to a technical solution of the present invention, in step S1, it specifically includes:

[0025] Step S11: Use the low-code configuration tool to define the enterprise KPI indicators;

[0026] Step S12: Generate an indicator catalog according to the defined enterprise KPI indicators;

[0027] Step S13: According to the defined enterprise KPI indicators and enterprise strategic goals, core indicators are determined to form a preliminary framework of the indicator value network; the core indicators include the target KPI indicators and their directly related potential related indicators;

[0028] Step S14: Acquire the historical business data of the core indicators, verify and quantify the correlation between the target KPI indicators and the potential correlation indicators, and form a complete indicator value network.

[0029] According to a technical solution of the present invention, in step S14, the verification and quantification include:

[0030] Step S141: Based on the historical business data, a statistical method is used to verify whether there is a stable correlation or causal relationship between the target KPI indicator and the potential related indicator at the data level;

[0031] Step S142: According to the verification result, the directed edges in the indicator value network are modified: if the enterprise KPI indicators represented by the nodes have a high correlation and a clear correlation direction, the weight of the directed edge corresponding to the node is increased; if the enterprise KPI indicators represented by the nodes have a weak correlation or no clear causal evidence, the weight of the directed edge corresponding to the node is reduced or the directed edge corresponding to the node is deleted;

[0032] Step S143: According to the correction results and the analysis results of the business experts, the unreasonable directed edges are intervened and reconfirmed.

[0033] According to a technical solution of the present invention, in step S1, it also includes:

[0034] Step S15: define new enterprise KPI indicators according to the new business indicators of the enterprise, and update the indicator catalog;

[0035] Step S16: After the newly added enterprise KPI indicator definition is completed, execute steps S13 to S14 to update the indicator value network.

[0036] According to a technical solution of the present invention, in step S2, it includes:

[0037] Perform semantic understanding on the user's natural language instructions, identify key entities and query intent in the user's natural language instructions, and generate a structured query intent description; the structured query intent description includes query target indicators, time range and query intent, and the query intent includes trend query, anomaly analysis, indicator comparison, attribution query, and risk assessment;

[0038] In step S3, index data matching is performed according to the key entity, specifically including:

[0039] Step S31, searching for matching indicator definitions according to the indicator catalog and the generated structured query intention description;

[0040] Step S32: Generate a data access request according to the matching and retrieval results in step S31;

[0041] Step S33: Generate a data query result according to the data access request.

[0042] According to a technical solution of the present invention, in step S3, correlation analysis and root cause location based on the indicator value network are performed according to the query intent, specifically including:

[0043] Step S34: starting from the query target indicator, exploring upward and / or downward along the directed edges in the indicator value network layer by layer according to the matching result, performing multi-hop reasoning, and obtaining a plurality of potential path chains;

[0044] Step S35: using the self-attention mechanism to score and sort the potential path chains, and output explanatory feedback in natural language form;

[0045] The potential path chain P is defined as a sequence of nodes and directed edges from the query target indicator to a certain associated indicator, expressed as:

[0046] P={n1→n2→…→n k}

[0047] Among them, n1 is the node corresponding to the query target indicator, that is, the target node; n2,…,n k are the intermediate nodes and associated nodes of the potential path chain P;

[0048] The score of the potential path chain P is expressed as:

[0049]

[0050] Among them, Q represents the query vector of the target node, K i represents the key vector of the i-th node in the potential path chain P, V iRepresents the value vector of the i-th node in the potential path chain P; Attention(Q,K i ,V i ) represents the attention weight of the target node and the i-th node multiplied by the value vector.

[0051] According to one aspect of the present invention, a large model-driven KPI indicator management and business analysis system provided by the present invention is used to implement the above-mentioned large model-driven KPI indicator management and business analysis method, including:

[0052] The indicator management module is used to uniformly define and manage KPI indicators through low-code configuration tools;

[0053] Data management module, used for persistent storage of enterprise KPI indicator data and access to business data;

[0054] The AI ​​analysis module uses a pre-trained large model based on the Transformer architecture to semantically understand the user's natural language instructions, and perform indicator data matching, correlation analysis based on the indicator value network, and root cause location based on the semantic understanding results;

[0055] The user interaction module is used to interact with the user in a natural language conversational manner and to display the results of the indicator data matching, the association analysis and the root cause location.

[0056] According to a technical solution of the present invention, the AI ​​analysis module includes:

[0057] A semantic parsing unit, used for mapping the user's natural language instruction into a structured query intent description including key entities and query intent;

[0058] A self-attention association analysis unit, configured to perform multi-head attention analysis on the query target indicator in the indicator value network according to the structured query intent description, and identify the dependency relationship between other enterprise KPI indicators and the query target indicator;

[0059] A root cause location unit is used to automatically attribute abnormal fluctuations of indicators based on the identified dependencies between other enterprise KPI indicators and the query target indicator, and output potential influencing factors and recommended action plans;

[0060] The indicator management module includes an indicator catalog and label management module, an indicator value network construction module, and a data source mapping and logic definition module;

[0061] The data management module includes a Dockerized service deployment module, a MySQL data storage module, a Redis cache acceleration module and an extensible data source access module.

[0062] Compared with the prior art, the present invention has the following beneficial effects:

[0063] 1) Efficient decision support: Users can directly obtain indicator analysis and suggestions through natural language dialogue, which greatly shortens the analysis cycle and lowers the threshold.

[0064] 2) Accurate root cause location: With the help of the indicator value network and self-attention mechanism, the large model can quickly identify the key correlations between enterprise KPI indicators in a complex indicator system, thereby efficiently locating the root causes of abnormal fluctuations.

[0065] 3) Unified indicator management: Use low-code tools to unify the definition and management of enterprise KPI indicators, achieve indicator standardization, and reduce duplicate definitions and data confusion.

[0066] 4) System elasticity and scalability: Using microservices, containerization, and distributed caching technologies, the system can easily cope with the growth in data dimensions and the number of indicators as the business expands.

[0067] 5) Intelligent natural language interaction: The big model performs semantic analysis on the user's natural language instructions and automatically matches indicator data with business scenarios, allowing users to obtain explanatory feedback in natural language without having to master technical details. This reduces the technical difficulty of user interaction with the analysis system and improves the user experience and the efficiency of obtaining business data analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention, and for ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0069] Figure 1 Schematically showing a flow chart of a large model-driven KPI indicator management and business analysis method according to an embodiment of the present invention;

[0070] Figure 2 A schematic diagram schematically shows the structure of a large model-driven KPI indicator management and business analysis system provided in an embodiment of the present invention;

[0071] Figure 3 A schematic diagram schematically shows the structure of an indicator value network according to an embodiment of the present invention;

[0072] Figure 4 The schematic diagram schematically shows the structure of the indicator value network according to another embodiment of the present invention. DETAILED DESCRIPTION

[0073] The description of the embodiments of this specification should be combined with the corresponding drawings, which should be considered as part of the complete specification. In the drawings, the shape or thickness of the embodiments may be enlarged and indicated for simplification or convenience. Furthermore, the parts of each structure in the drawings will be described separately. It is worth noting that the elements not shown in the drawings or not described in words are in a form known to ordinary technicians in the relevant technical field.

[0074] The description of the embodiments herein and any reference to directions and orientations are only for the convenience of description and are not to be construed as any limitation on the scope of protection of the present invention. The following description of the preferred embodiments may involve combinations of features, which may exist independently or in combination, and the present invention is not particularly limited to the preferred embodiments. The scope of the present invention is defined by the claims.

[0075] like Figure 1 As shown, a large model-driven KPI indicator management and business analysis method of the present invention includes the following steps:

[0076] Step S1: Define the enterprise KPI indicators through the low-code configuration tool and establish the indicator value network;

[0077] Step S2: semantically analyze the input user natural language instruction to identify key entities and query intent in the user natural language instruction, where the key entities include query target indicators and time range;

[0078] Step S3: Match indicator data according to key entities, and perform correlation analysis and root cause location based on indicator value network according to query intent to identify relevant indicators and potential root causes;

[0079] Step S4: Output an analysis report and decision-making recommendations based on the results of indicator data matching, correlation analysis, and root cause location.

[0080] Also includes:

[0081] Step S5: Based on the output analysis report and decision suggestions, obtain new user natural language instructions and execute steps S2 to S4, so that the user can further inquire and optimize through natural language interaction.

[0082] The indicator value network includes several nodes and several directed edges. Each node represents an enterprise KPI indicator, and the directed edges represent the attribute influence relationship between enterprise KPI indicators. The types of enterprise KPI indicators include atomic indicators, derived indicators and composite indicators. Atomic indicators are basic values ​​that directly measure product or business performance, such as sales and order volume; derived indicators are derived from atomic indicators through derived calculation methods (such as same ratio, year-on-year, cumulative, average, etc.) to meet more advanced data statistical analysis needs; composite indicators are constructed by mathematical operations of multiple basic indicators (atomic indicators and / or derived indicators) to express more complex business logic; mathematical operations include addition, subtraction, multiplication, division, ratio, etc. The attribute influence relationship between enterprise KPI indicators includes positive influence, negative influence, delayed effect, influence weight, time series characteristics, etc. Through the indicator value network, the query target indicator and its potential key indicators can be found, and then the potential root causes affecting the abnormal fluctuations of enterprise KPI indicators can be determined, thereby providing support for users' indicator management and decision analysis.

[0083] In step S1, it specifically includes:

[0084] Step S11: Use the low-code configuration tool to define the enterprise KPI indicators;

[0085] Step S12: Generate an indicator catalog according to the defined enterprise KPI indicators;

[0086] Step S13: According to the defined enterprise KPI indicators and enterprise strategic goals, core indicators are determined to form a preliminary framework of the indicator value network; the core indicators include the target KPI indicators and their directly related potential related indicators.

[0087] Step S14: Obtain historical business data of core indicators, verify and quantify the correlation between target KPI indicators and potential related indicators, and form a complete indicator value network.

[0088] In step S1, it also includes:

[0089] Step S15: define new enterprise KPIs according to the new business indicators of the enterprise, and update the indicator catalog;

[0090] Step S16: After the newly added enterprise KPI indicator definition is completed, execute steps S13 to S14 to update the indicator value network.

[0091] Wherein, in step S14, verification and quantification include the following steps:

[0092] Step S141: Based on historical business data, a statistical method is used to verify whether there is a stable correlation or causal relationship between the target KPI indicator and the potential related indicators at the data level; the statistical method may be one or a combination of correlation coefficient method, Granger causality test method, and time series lag analysis method;

[0093] Specifically, the Pearson correlation coefficient method is used to measure the linear correlation between two indicators in the time series, and its calculation formula is:

[0094] Among them, r represents the Pearson correlation coefficient between indicator X and indicator Y; X m and Y m Respectively represent the mth data value of indicator X and indicator Y, and Respectively represent the mean of indicator X and indicator Y;

[0095] When r>0, the indicator X is positively correlated with the indicator Y, and the weight of the directed edge between the two nodes corresponding to the indicator X and the indicator Y is positive; when r>0, the indicator X is positively correlated with the indicator Y, and the weight of the directed edge between the two nodes corresponding to the indicator X and the indicator Y is positive; the closer |r| is to 1, the stronger the correlation between the indicator X and the indicator Y, and the larger the weight value.

[0096] Granger causality test can be used to test whether indicator X is the cause of indicator Y. If the lagged value of indicator X in the time series significantly improves the predictive ability of indicator Y, then indicator X is considered to be the cause of indicator Y. The specific steps include:

[0097] a. Construct a lagged time regression model of indicator X and indicator Y;

[0098] b. Calculate the F-statistic and the corresponding p-value in the lagged time regression model of indicator X and indicator Y;

[0099] c. If the p-value is less than the significance level threshold, it is considered that there is a causal relationship between indicator X and indicator Y; otherwise, it is considered that there is no causal relationship between indicator X and indicator Y.

[0100] d. If indicator X is the cause of indicator Y, the weight of the directed edge between the two nodes corresponding to indicator X and indicator Y is assigned a positive value; if indicator X is not a causal relationship with indicator Y, the weight of the directed edge between the two nodes corresponding to indicator X and indicator Y is assigned a value of zero.

[0101] Step S142: According to the verification result, the directed edges in the indicator value network are modified: if the enterprise KPI indicators represented by the nodes have a high correlation and a clear correlation direction, the weight of the directed edge corresponding to the node is increased; if the enterprise KPI indicators represented by the nodes have a weak correlation or no clear causal evidence, the weight of the directed edge corresponding to the node is reduced or the directed edge corresponding to the node is deleted;

[0102] After establishing the preliminary framework, the system will call on historical data to verify and quantify the established relationships, ensuring that the relationships have not only a logical basis but also data support.

[0103] Step S143: According to the correction results and the analysis results of the business experts, the unreasonable directed edges are intervened and reconfirmed.

[0104] Intervention and reconfirmation of unreasonable directed edges can be carried out through collaborative optimization between business experts and big models. Through the review and feedback of business experts, unreasonable associated edges can be manually intervened or reconfirmed based on the statistical results, causal chains and actual business experience explained by the model. At the same time, the big model uses natural language interaction to convert complex statistical or time series analysis results into easy-to-understand explanations (such as "According to historical data, advertising volume has a continuous positive impact on sales, and it is recommended to retain and increase its weight"), so as to significantly speed up expert decision-making and network improvement iteration.

[0105] For example, when constructing an indicator value network with sales as the query target indicator and advertising volume, user retention rate and inventory turnover rate as potential related indicators, the historical quarterly report of the enterprise is used as the data source of the historical business data of the enterprise. First, the target indicator node and the potential related indicator node are defined, and then the weights of the directed edges between advertising volume and sales, user retention rate and sales, and inventory turnover rate and sales are calculated respectively:

[0106] Advertising volume and sales: The Pearson correlation coefficient calculation result is r = 0.75; the lag analysis shows that advertising has a positive lag effect on sales of 1 month. Therefore, the edge weight is set to w = 0.75, marking it as a positive relationship.

[0107] User retention rate and sales: The Granger causality test p value is 0.03, and there is a causal relationship. Therefore, the edge weight is set to w = 0.60, marking it as a positive relationship.

[0108] Inventory turnover rate and sales: The correlation is weak (r = 0.20), and there is no significant causal relationship. Therefore, the edge weight is set to w = 0.10, and it is marked as a weak positive correlation.

[0109] The present invention establishes the following edge weight adjustment rules for the weights of the directed edges of the indicator value network:

[0110] Hierarchical optimization: For the first-level nodes directly related to the target KPI indicators, the weight values ​​of the directed edges are adjusted through data verification; for the second-level and above nodes, the weights of the lagging relationships are adjusted by business experts;

[0111] Dynamic update: Update historical data every quarter and re-verify whether the weights conform to business logic;

[0112] Exception handling: delete edges that have no data support or no clear causal relationship; add newly discovered associations.

[0113] Through the above edge weight adjustment rules, the correlation relationship between the KPI indicators of each enterprise can be efficiently established, thereby providing support for subsequent correlation analysis and root cause location.

[0114] In step S2, it specifically includes: semantic understanding of the user's natural language instructions, identifying the key entities and query intent in the user's natural language instructions, and generating a structured query intent description; the structured query intent description includes query target indicators, time range and query intent, and the query intent includes trend query, anomaly analysis, indicator comparison, attribution query, and risk assessment. The query target indicator and time range are used to locate the indicator business data. According to the type of query intent, the large model can perform multi-hop reasoning and root cause location analysis to output decision analysis for the user.

[0115] In step S3, indicator data matching is performed according to key entities, specifically including:

[0116] Step S31, searching for matching indicator definitions according to the indicator catalog and the generated structured query intent description;

[0117] Step S32: Generate a data access request according to the matching and retrieval results in step S31;

[0118] Step S33: Generate data query results according to the data access request.

[0119] In step S31, the index definition for retrieval matching specifically includes: locating the standard name and data scope in the index catalog based on the query intent index; judging the calculation logic, data source and whether there is a derivative or compound relationship of the query intent index, so that subsequent data acquisition is more accurate; selecting a data acquisition strategy based on the type of query intent, query complexity and data characteristics; based on the index matching and retrieval results, the model generates a corresponding data access request; after the generated data access request is executed, the corresponding result data will be returned to the large model for subsequent processing.

[0120] The data acquisition strategy is: when the query intention type is a trend query and the query intention indicator is an atomic indicator, select the SQL query strategy, generate an SQL statement corresponding to the query intention indicator, directly access the underlying database, and obtain the business data of the query intention indicator; when the query intention indicator is a derived indicator or a composite indicator, select the API query strategy, call the indicator API of the query intention indicator and its related indicators to quickly return the results.

[0121] If the query intent indicator can be directly queried from the underlying data source (such as existing data warehouse tables and wide table structures), the model will automatically construct the query statement. For example, based on the information of "this quarter", "sales" and the basic data table, a SQL statement with time filtering and aggregation calculation is generated (such as SELECT SUM(sales)FROM sales_table WHEREdate BETWEEN'2023-01-01'AND'2023-03-31').

[0122] If the query intent indicator has been defined as a standardized derived indicator or composite indicator in the indicator catalog, the model can obtain the pre-calculated results by calling the unified indicator API. This reduces the user's attention to the underlying data logic and ensures the consistency of the indicator caliber.

[0123] The data query results can be presented in a variety of ways, such as by displaying the changing trends of indicators over time through visual charts (such as line charts, bar charts, pie charts, etc.), helping users to observe the rising or falling trends of sales at a glance; or summarizing the data results through natural language generation technology to generate a text analysis report; multi-hop reasoning can also be performed based on the indicator value network to infer and locate possible influencing factors, potential risks and improvement space, providing users with further cause analysis or related indicator information.

[0124] In step S3, correlation analysis and root cause location based on the indicator value network are performed according to the query intent, including:

[0125] Step S34: Based on the matching results, starting from the query target indicator, explore upward and / or downward along the directed edges in the indicator value network layer by layer, perform multi-hop reasoning, and obtain several potential path chains;

[0126] The large model starts from querying the target indicator, exploring upward or downward layer by layer along the directed edges in the indicator value network, automatically discovering potential influencing indicators hidden behind secondary, tertiary or even higher-level relationships, and forming multiple potential path chains.

[0127] Step S35: Use the self-attention mechanism to score and sort several potential path chains, and output explanatory feedback in natural language form.

[0128] The calculation formula of the self-attention mechanism can be expressed as:

[0129]

[0130] Among them, Q is the representation vector of the query target indicator, K is the representation vector of the potential correlation indicator, and V is the output value of the potential correlation indicator combined with the attention weight; QK T To query the correlation score between the target indicator and the associated indicator; is a scaling factor used to avoid gradients that are too large or too small, d k represents the dimensions of Q and K; softmax means normalizing the correlation scores into weights.

[0131] Multiple inferences can be performed based on a multi-head mechanism. Through parallel calculations of multiple independent attention heads, information of different dimensions can be captured, which can be expressed as:

[0132] MultiHead(Q,K,V)=Concat(head1,head2,…,head h )W O

[0133]

[0134] in, are the linear transformation matrices of the j-th attention head, W O is the output linear transformation matrix; h is the number of attention heads.

[0135] Therefore, when performing multi-hop reasoning to obtain several potential path chains, we can first define the potential path chain P as a sequence consisting of a set of nodes and directed edges from the query target indicator to a certain associated indicator, expressed as:

[0136] P={n1→n2→…→n k}

[0137] Among them, n1 is the target node, n2,…,n k are intermediate nodes and associated nodes, and k is the total number of nodes on the potential path chain P.

[0138] Each potential path chain represents a set of potential causal relationships, and the goal is to find the importance score of the path to the query target metric.

[0139] The scoring of potential path chains is based on the self-attention mechanism and is obtained by weighted summing of the relevance scores of each node, expressed as:

[0140]

[0141] Among them, Q represents the query vector of the query target indicator, K i represents the key vector of the i-th node in the potential path chain P, V i Represents the value vector of the i-th node in the path; Attention(Q,K i ,V i ) represents the attention weight of the target node and the i-th node multiplied by the value vector.

[0142] Step S352: Calculate the attention branch for each node in each potential path chain in turn, and add up the scores of all nodes to obtain the score of the corresponding potential path chain;

[0143] Step S353: Generate natural language feedback through a predefined template according to the scoring numerical results of the potential path chain.

[0144] Specifically, Figure 4 Taking the indicator value network shown as an example, setting the query target indicator (i.e. node n1) to sales, the potential path chain P1 is: advertising volume (node ​​n2) → user retention rate (node ​​n3) → sales (node ​​n4).

[0145] The feature vectors of each node are expressed as:

[0146] Query vectors for sales: Q = [0.8, 0.2, 0.5], Q = [0.8, 0.2, 0.5];

[0147] Key vector and value vector of ad delivery: K2 = [0.6, 0.1, 0.4], V2 = [0.7, 0.3, 0.6];

[0148] Key vector and value vector of user retention rate: K3 = [0.5, 0.3, 0.8], V3 = [0.6, 0.4, 0.7];

[0149] Key vector and value vector of sales: K4 = [0.9, 0.2, 0.3], V4 = [0.8, 0.5, 0.4];

[0150] The attention score of computing node n2 (i.e., the amount of advertising delivered) is:

[0151]

[0152] in,

[0153] Apply softmax (assuming the result of softmax for multiple nodes is 0.5).

[0154] Then the attention score is: Attention(Q,K2,V2)=0.5·[0.7,0.3,0.6]=[0.35,0.15,0.3].

[0155] The attention scores of nodes n3 and n4 are calculated using the same formula, and then the link scores are accumulated according to the following formula:

[0156] Score(P)=Attention(Q,K2,V2)+Attention(Q,K3,V3)+Attention(Q,K4,V4)

[0157] Path P1: advertising volume → user retention rate → sales, score 0.85.

[0158] Path P2: advertising volume → conversion rate → sales, score 0.78.

[0159] The generated natural language feedback is:

[0160] The score of the path [advertising volume → user retention rate → sales] is 0.85, indicating that the importance of advertising to sales is reflected by improving user retention rate.

[0161] The self-attention mechanism is used to perform weighted summation of the relationship between the query target indicator and the related indicators layer by layer. The higher the score, the more significant the impact on the query target indicator, and the low-scoring path can be ignored, thereby helping users understand the complex indicator association and providing support for decision-making.

[0162] The large model uses the self-attention mechanism to score and sort the multiple potential path chains obtained according to the weights and attribute information of nodes and directed edges, and informs users of the scoring and sorting results in the form of intuitive explanatory feedback (such as "In the data of the past three quarters, the link that average order value affects sales by affecting user retention rate and ultimately affects sales has steadily emerged. It is recommended to focus on this path"), thereby providing users with auxiliary decision support.

[0163] As the business environment and business data continue to change, the indicator value network also needs to be updated periodically. Data-driven analysis, business expert feedback and multi-hop reasoning of large models work together to form a closed-loop iteration mechanism: when new quarterly goals (such as improving profitability or reducing inventory turnover days) appear, the indicator catalog can be updated to add or reshape associations based on the existing indicator value network. By using updated business data, the key paths in the indicator value network are re-evaluated, so that users do not need to manually modify the indicator value network, and can be reminded to optimize decision-making strategies and pay attention to new influencing factors.

[0164] like Figure 2As shown, a large model-driven KPI indicator management and business analysis system of the present invention includes:

[0165] The indicator management module is used to uniformly define and manage KPI indicators through low-code configuration tools to form an indicator value network. The structure of the indicator value network in one embodiment of the present invention is as follows: Figure 3 As shown;

[0166] The data management module is used for persistent storage of enterprise KPI indicator data and access to business data; business data is structured data and / or semi-structured data;

[0167] The AI ​​analysis module uses a pre-trained large model based on the Transformer architecture to perform semantic understanding of user natural language queries, and performs indicator data matching, correlation analysis based on the indicator value network, and root cause location based on the semantic understanding results;

[0168] The user interaction module is used to interact with users in a natural language conversational manner and display the results of indicator data matching, correlation analysis, and root cause location.

[0169] The data management module includes a Dockerized service deployment module, a MySQL data storage module, a Redis cache acceleration module, and an extensible data source access module. The MySQL data storage module is used to store the basic indicator definitions and calculation logic of the enterprise KPI indicators, and the Redis cache acceleration module is used to cache hot indicator data. The system supports access to multiple types of data sources, including real-time data (such as PLC, Modbus, OPCUA, Socket, MQTT subscription or polling mode) and non-real-time data (such as DAMO, PostgreSQL, Kingdee database, SQL Server, Oracle, MySQL, etc.).

[0170] The indicator management module includes the indicator catalog and label management module, the indicator value network construction module, and the data source mapping and logic definition module. Based on the access of multiple types of data sources in the data management module, the indicator management module uses low-code configuration tools to allow users to quickly create, modify and manage enterprise KPI indicators. Through standardized and structured storage, as well as an extensible indicator catalog and label mechanism, enterprises can flexibly expand various business indicators to ensure data consistency and traceability. Through a unified indicator management platform, data consistency and traceability can be ensured, laying a solid foundation for subsequent large model analysis.

[0171] The AI ​​analysis module uses a pre-trained large model based on the Transformer architecture as the core engine, and inputs the user's natural language query (such as "Why did sales decline this quarter?" or "Which indicators are closely related to user retention rate?") into the model that has been fine-tuned with domain knowledge. Users do not need to master complex query syntax and can directly obtain accurate indicator analysis results and decision-making recommendations through natural language dialogue.

[0172] After receiving user input, the AI ​​analysis module first performs semantic understanding and intent recognition, then automatically searches the indicator library and selects the appropriate data acquisition method (generating SQL or calling indicator API) based on the intent. Finally, the model presents the results to the user in a clear and easy-to-understand form, including charts, text analysis reports, etc., to help users quickly obtain the information needed for decision-making.

[0173] The AI ​​analysis module includes a semantic parsing unit, a self-attention association analysis unit and a root cause location unit. The semantic parsing unit is used to map the user's natural language instructions into indicator query instructions; the self-attention association analysis unit is used to perform multi-head attention analysis on the query target indicator in the indicator value network according to the structured query intent description, and identify the dependency relationship between other enterprise KPI indicators and the query target indicator; the root cause location unit is used to automatically attribute abnormal fluctuations of indicators based on the identified dependency relationship between other enterprise KPI indicators and the query target indicator, and output potential influencing factors and recommended action plans.

[0174] The system can be deployed in the internal cloud environment of the enterprise. The backend uses MySQL to store basic indicator definitions and calculation logic, and Redis caches hot indicator data. Users access the system through the Web front end (Vue.js framework) and enter natural language queries in the dialog box, such as: "Why did sales decline this quarter? Is it related to the inventory turnover of new product lines?"

[0175] The system receives the request and passes the natural language input into the Transformer large model fine-tuned with economic field data for semantic analysis. The large model retrieves the enterprise KPI indicator library and value network.

[0176] The big model analyzes the output results, including key KPIs (such as sales, inventory turnover, advertising volume, customer feedback scores, etc.) and their correlations, and gives possible root causes (such as changes in market demand or supply chain delays leading to lower inventory turnover, thereby affecting sales).

[0177] The system presents the analysis results in a visual form and recommends increasing the delivery to a specific market, optimizing supply chain scheduling, or adjusting product pricing strategies. Users can continue to ask questions in natural language to request more detailed reasons or alternatives.

[0178] It should be noted that the above is a preferred embodiment of the present invention. It should be pointed out that although the preferred embodiment of the present invention has been described, once the basic creative concept of the present invention is known, a number of improvements and modifications can be made by a person skilled in the art without departing from the principles of the present invention, and these improvements and modifications should also be regarded as the protection scope of the present invention. Therefore, the attached claims are intended to be interpreted as including the preferred embodiment and all changes and modifications that fall within the scope of the embodiments of the present invention.

Claims

1. A large model-driven KPI indicator management and business analysis method, characterized in that: include: Step S1: Define the enterprise KPI indicators through the low-code configuration tool and establish the indicator value network; Step S2: semantically parse the input user natural language instruction to identify key entities and query intent in the user natural language instruction, where the key entities include query target indicators and time range; Step S3: matching indicator data according to the key entities, and performing correlation analysis and root cause location based on the indicator value network according to the query intent to identify relevant indicators and potential root causes; Step S4: output an analysis report and decision-making recommendations based on the results of the indicator data matching and the results of the association analysis and root cause location.

2. The large model-driven KPI indicator management and business analysis method according to claim 1 is characterized in that: Also includes: Step S5: According to the output analysis report and decision suggestions, obtain new user natural language instructions and execute steps S2 to S4.

3. The large model-driven KPI indicator management and business analysis method according to claim 1 is characterized in that: The indicator value network includes a number of nodes and a number of directed edges, each node represents an enterprise KPI indicator, and each directed edge represents the attribute influence relationship between the enterprise KPI indicators corresponding to the node; The types of enterprise KPI indicators include: Atomic indicators are basic values ​​that directly measure product or business performance; Derived indicators are derived from the atomic indicators through derived calculation methods; the derived calculation methods include year-on-year, month-on-month, cumulative, and average values; Composite indicators are constructed by mathematical operations on multiple basic indicators; the basic indicators are atomic indicators and / or derived indicators; The attribute influence relationship between the enterprise KPI indicators includes positive influence, negative influence, delayed effect, influence weight, and time series characteristics.

4. The large model-driven KPI indicator management and business analysis method according to claim 3 is characterized in that: In the step S1, it specifically includes: Step S11: Use the low-code configuration tool to define the enterprise KPI indicators; Step S12: Generate an indicator catalog according to the defined enterprise KPI indicators; Step S13: According to the defined enterprise KPI indicators and enterprise strategic goals, core indicators are determined to form a preliminary framework of the indicator value network; the core indicators include the target KPI indicators and their directly related potential related indicators; Step S14: Acquire the historical business data of the core indicators, verify and quantify the correlation between the target KPI indicators and the potential correlation indicators, and form a complete indicator value network.

5. The large model-driven KPI indicator management and business analysis method according to claim 4 is characterized in that: In the step S14, the verification and quantification include: Step S141: Based on the historical business data, a statistical method is used to verify whether there is a stable correlation or causal relationship between the target KPI indicator and the potential related indicator at the data level; Step S142: According to the verification result, the directed edges in the indicator value network are modified: if the enterprise KPI indicators represented by the nodes have a high correlation and a clear correlation direction, the weight of the directed edge corresponding to the node is increased; if the enterprise KPI indicators represented by the nodes have a weak correlation or no clear causal evidence, the weight of the directed edge corresponding to the node is reduced or the directed edge corresponding to the node is deleted; Step S143: According to the correction results and the analysis results of the business experts, the unreasonable directed edges are intervened and reconfirmed.

6. The large model-driven KPI indicator management and business analysis method according to claim 4 is characterized in that: In the step S1, it also includes: Step S15: define new enterprise KPI indicators according to the new business indicators of the enterprise, and update the indicator catalog; Step S16: After the newly added enterprise KPI indicator definition is completed, execute steps S13 to S14 to update the indicator value network.

7. The large model-driven KPI indicator management and business analysis method according to claim 1 is characterized in that: In the step S2, it includes: Perform semantic understanding on the user's natural language instructions, identify key entities and query intent in the user's natural language instructions, and generate a structured query intent description; the structured query intent description includes query target indicators, time range and query intent, and the query intent includes trend query, anomaly analysis, indicator comparison, attribution query, and risk assessment; In step S3, index data matching is performed according to the key entity, specifically including: Step S31, searching for matching indicator definitions according to the indicator catalog and the generated structured query intention description; Step S32: Generate a data access request according to the matching and retrieval results in step S31; Step S33: Generate a data query result according to the data access request.

8. The large model-driven KPI indicator management and business analysis method according to claim 7 is characterized in that: In the step S3, correlation analysis and root cause location based on the indicator value network are performed according to the query intent, specifically including: Step S34: starting from the query target indicator, exploring upward and / or downward along the directed edges in the indicator value network layer by layer according to the matching result, performing multi-hop reasoning, and obtaining a plurality of potential path chains; Step S35: using the self-attention mechanism to score and sort the potential path chains, and output explanatory feedback in natural language form; The potential path chain P is defined as a sequence of nodes and directed edges from the query target indicator to a certain associated indicator, expressed as: P={n1→n2→…→n k } Among them, n1 is the node corresponding to the query target indicator, that is, the target node; n2,…,n k are the intermediate nodes and associated nodes of the potential path chain P; The score of the potential path chain P is expressed as: Among them, Q represents the query vector of the target node, K i represents the key vector of the i-th node in the potential path chain P, V i Represents the value vector of the i-th node in the potential path chain P; Attention(Q,K i ,V i ) represents the attention weight of the target node and the i-th node multiplied by the value vector.

9. A large model driven KPI indicator management and business analysis system, used to implement the large model driven KPI indicator management and business analysis method according to any one of claims 1 to 8, characterized in that: include: The indicator management module is used to uniformly define and manage KPI indicators through low-code configuration tools; Data management module, used for persistent storage of enterprise KPI indicator data and access to business data; The AI ​​analysis module uses a pre-trained large model based on the Transformer architecture to semantically understand the user's natural language instructions, and perform indicator data matching, correlation analysis based on the indicator value network, and root cause location based on the semantic understanding results; The user interaction module is used to interact with the user in a natural language conversational manner and to display the results of the indicator data matching, the association analysis and the root cause location.

10. The large model-driven KPI indicator management and business analysis method according to claim 9, characterized in that: The AI ​​analysis module includes: A semantic parsing unit, used for mapping the user's natural language instruction into a structured query intent description including key entities and query intent; A self-attention association analysis unit, configured to perform multi-head attention analysis on the query target indicator in the indicator value network according to the structured query intent description, and identify the dependency relationship between other enterprise KPI indicators and the query target indicator; A root cause location unit is used to automatically attribute abnormal fluctuations of indicators based on the identified dependencies between other enterprise KPI indicators and the query target indicator, and output potential influencing factors and recommended action plans; The indicator management module includes an indicator catalog and label management module, an indicator value network construction module, and a data source mapping and logic definition module; The data management module includes a Dockerized service deployment module, a MySQL data storage module, a Redis cache acceleration module and an extensible data source access module.

Images