An AI-based big data mining assisted decision-making method and system

Abstract

The application relates to the field of artificial intelligence and big data intelligent analysis technology, and particularly discloses an AI-based big data mining auxiliary decision-making method and system, which comprises the following steps: acquiring transaction record time sequences, device running state voltage value sequences and user behavior click frequency sequences in a log collection terminal, calculating adjacent time point difference values to form a change amplitude sequence, aligning change amplitude items in different change amplitude sequences to corresponding common time positions and comparing change directions, pairing variables with consistent change trends and establishing a corresponding relationship to obtain a business variable correlation relationship set; extracting a target variable correlation node based on the business variable correlation relationship set, constructing a disturbance response reasoning model based on the target variable and the target variable correlation node, implementing amplitude disturbance on each node and recording changes before and after the disturbance, sorting change results, extracting response occurrence time positions of each variable and comparing the sequence.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and big data intelligent analysis technology, specifically to an AI-based big data mining-assisted decision-making method and system. Background Technology

[0002] As enterprise informatization continues to advance, enterprise data centers generate a large amount of transaction records, equipment operation monitoring data, and user behavior data during daily operation. Existing big data analysis methods typically collect, clean, store, and extract features from the above data, and then analyze the relationships between different variables through machine learning, association analysis, or rule mining to provide reference for business operation analysis and decision support.

[0003] However, due to the different sources of various data, their sampling periods, statistical calibers, time granularity, and numerical ranges often differ. When conducting joint analysis, it is easily affected by factors such as inconsistent collection times, missing data, normal fluctuations, and abnormal noise, making it difficult to accurately correspond the changes of different variables within the same time period. For example, during peak business periods, the number of user operations, equipment operating parameters, and transaction results may change simultaneously. Based solely on the fact that these variables have similar trends, it is difficult to determine whether there is a stable relationship between them.

[0004] Existing artificial intelligence analysis methods typically determine variable relationships based on the degree of correlation, common trends, or co-occurrence frequency in historical data. These relationships mainly reflect the statistical correlation of variables. When multiple variables are affected by the same external factor, or when there is an indirect relationship between variables, the analysis results may contain accidental, repetitive, or indirect relationships, making it difficult to further determine whether there is a causal relationship between variables, and also difficult to determine the direction of mutual influence.

[0005] In addition, while some artificial intelligence models can provide prediction or classification results, their analytical basis is not intuitive enough, and the degree of influence of different variables on the analysis results is difficult to clarify. When variables with low correlation or repetitive content are used to set decision conditions, it is easy to result in a large number of conditions and complex variable relationships, which in turn affects the stability and credibility of the auxiliary decision-making results.

[0006] Therefore, how to reduce the impact of differences, fluctuations and statistical correlations in multi-source business data on the analysis results, improve the accuracy of judging variable relationships and causal connections, and make the analysis results easier to understand and verify are technical problems that need to be solved in this field.

[0007] Therefore, this invention provides an AI-based big data mining-assisted decision-making method and system. Summary of the Invention

[0008] The purpose of this invention is to provide an AI-based big data mining-assisted decision-making method and system to solve the aforementioned background problems.

[0009] The objective of this invention can be achieved through the following technical solutions:

[0010] An AI-based big data mining-assisted decision-making method includes:

[0011] Obtain the transaction record time series, device operating status voltage value series and user behavior click frequency series from the log collection terminal, calculate the difference between adjacent time points to form a change amplitude series, align the change amplitude items in different change amplitude series to the corresponding common time position and compare the change direction, pair variables with consistent change trends and establish corresponding relationships to obtain a set of business variable association relationships;

[0012] Based on the set of relationships between business variables, target variable association nodes are extracted. Based on the target variable and target variable association nodes, a disturbance response inference model is constructed. Amplitude disturbances are applied to each node and the changes before and after are recorded. The change results are sorted. The response time position of each variable is extracted and the order is compared. The earliest response variable is taken as the starting node and connected in the order of response to form a path structure, thus obtaining a set of causal relationship paths.

[0013] The response changes of each path node are extracted from the causal path set. The continuity of adjacent nodes is compared and the order of change transmission is identified. Paths that meet the characteristics of continuous transmission are retained. The contribution of each node in the retained path to the overall change is decomposed. Low-impact nodes are identified according to the contribution ratio. Adjacent low-impact nodes are merged and the path structure is adjusted to obtain the causal link structure set.

[0014] For the set of causal link structures, complete paths are selected. The adjacency weight of each node is determined based on the connection distance between each node and the target variable. The overall strength of the path is determined based on the contribution ratio of each node and its corresponding adjacency weight. Key causal links are identified by sorting them according to their overall strength. The node combination relationship is extracted and mapped into a rule expression. The rule structure is organized according to the node order, contribution ratio, and path direction to obtain a set of auxiliary decision rules.

[0015] The beneficial effects of this invention are as follows:

[0016] This invention calculates adjacent differences and aligns time between transaction record data, device voltage data, and user click frequency data, enabling data from different sampling periods and time granularities to be compared at the same time position, thus reducing the impact of time misalignment, normal fluctuations, and abnormal noise on variable relationship analysis.

[0017] Based on this, applying amplitude perturbations to the related variables and comparing the response amplitude and response time before and after the perturbations helps to determine the sequential relationship of variable changes and the direction of influence, and reduces the interference of accidental associations, duplicate associations and indirect associations on the analysis results.

[0018] After continuous transmission relationship testing and node contribution calculation, weaker nodes can be identified and continuous low-impact nodes can be merged to make the causal link structure clearer, while preserving the degree of influence of each node on the target variable.

[0019] By further combining the contribution ratio of nodes and the connection distance between nodes and the target variable to calculate the overall path strength, more important links can be selected from multiple causal links. Based on this, auxiliary decision-making rules with information on changing conditions, temporal relationships and weights can be formed, thereby improving the interpretability of the analysis results and the stability and reliability of the auxiliary decision-making basis. Attached Figure Description

[0020] The invention will now be further described with reference to the accompanying drawings.

[0021] Figure 1 This is a flowchart of an AI-based big data mining-assisted decision-making method according to an embodiment of the present invention;

[0022] Figure 2 This is a flowchart illustrating the disturbance response and causal relationship path of an AI-based big data mining-assisted decision-making method according to an embodiment of the present invention.

[0023] Figure 3 This is a functional module diagram of an AI-based big data mining-assisted decision-making method according to an embodiment of the present invention. Detailed Implementation

[0024] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments;

[0025] Example 1

[0026] Please see Figure 1 - Figure 2 As shown in the embodiment of the present invention, a big data mining-based decision-making assistance method based on AI mainly includes the following steps:

[0027] It should be noted that this embodiment is applied to an enterprise data center environment;

[0028] An enterprise data center environment includes business system log collection terminals, data processing servers, distributed storage devices, AI inference computing units, decision rule generation modules, and data communication links for data interaction between the above devices.

[0029] The AI ​​inference computing unit is located inside the data processing server or connected to the data processing server through a data communication interface. It is used to perform functions such as change magnitude calculation, time alignment, change direction determination, trend consistency calculation, disturbance response inference, response time comparison, path construction, continuous transmission verification, node contribution decomposition, node structure adjustment, adjacency weight calculation, and overall path strength calculation.

[0030] The decision rule generation module is used to generate auxiliary decision rules based on key causal links;

[0031] Step 1: Obtain the transaction record time series, equipment operating status voltage value series, and user behavior click frequency series from the enterprise business system log collection terminal. Perform difference processing on the values ​​of adjacent time points in each series to form a change amplitude series. Align the change amplitude items in different change amplitude series to the corresponding common time position, compare the change direction of each variable at the common time position, pair variables with consistent change trends and establish corresponding relationships to obtain a set of business variable association relationships.

[0032] In step one, the process of multi-source business data acquisition and time series formation is as follows:

[0033] In some embodiments, the business system log collection terminal acquires transaction record data, device operating status voltage sampling data, and user behavior click frequency data during a continuous collection period.

[0034] Transaction record data includes at least one of the following: number of transactions, transaction amount, transaction success rate, number of failed transactions, or number of refunds.

[0035] Equipment operating status voltage sampling data includes voltage sampling values ​​generated by data center servers, storage devices, network devices, or business processing devices;

[0036] User behavior click frequency data includes the statistical value of the number of clicks made by users on business system pages, interfaces or function entry points;

[0037] The business system log collection terminal performs field parsing on transaction record data, equipment operating status voltage sampling data, and user behavior click frequency data. It generates variable identifiers based on the data source, business object, and indicator type of each data, extracts or determines the corresponding collection timestamp and variable value, and combines the variable identifier, collection timestamp, and variable value to form a business variable value record.

[0038] Variable identifiers are used to distinguish business variables corresponding to different data sources, different business objects, and different indicator types;

[0039] The data collection timestamp is used to indicate the data collection time or statistical time position of the corresponding variable value;

[0040] The variable value is used to represent the numerical status of the corresponding business variable at the time of data collection or statistical analysis.

[0041] The records of business variable values ​​are classified according to variable identifiers, and the records of the same variable are arranged from earliest to latest according to the collection timestamp, forming transaction record time series, equipment operating status voltage value series, and user behavior click frequency series respectively;

[0042] The data processing server stores each time series data into a distributed storage device and sends it to the artificial intelligence inference computing unit.

[0043] In step one, the process of generating the change amplitude sequence is as follows:

[0044] The AI ​​inference computing unit performs adjacent time point difference processing on each time series;

[0045] For any time series among the transaction record time series, the device operating status voltage value series, and the user behavior click frequency series, starting from the second time position, calculate the difference between the variable value at the current time position and the variable value at the previous adjacent time position, and determine the obtained difference as the change range of the corresponding variable at the current time position;

[0046] When the current time position or the previous adjacent time position lacks variable values, the change range item corresponding to the current time position is not generated;

[0047] The magnitude of change retains the positive or negative attribute of the difference; a positive magnitude of change indicates that the variable has risen relative to the previous adjacent time position, while a negative magnitude of change indicates that it has fallen, and the absolute value of the magnitude of change indicates the degree of change.

[0048] Through differential processing, the transaction record time series, the device operating status voltage value series, and the user behavior click frequency series are respectively converted into the transaction record change amplitude series, the device voltage change amplitude series, and the click frequency change amplitude series;

[0049] Each change range item includes a variable identifier, a timestamp corresponding to the current time position, and the change range;

[0050] In step one, the process of time alignment of the change amplitude sequence is as follows:

[0051] The AI ​​inference computing unit sorts the sequence of changes according to timestamps and determines a uniform time alignment window length based on the statistical period of transaction record data, the sampling period of device voltage data, and the statistical period of click frequency data.

[0052] The time alignment window length is greater than or equal to the largest statistical period or sampling period among all data types;

[0053] The starting timestamp of the continuous collection period is used as the starting point of the first time alignment window. Non-overlapping time alignment windows are continuously divided according to the length of the time alignment window, and common time position identifiers are configured according to the chronological order of time.

[0054] Each time alignment window includes its start time but excludes its end time.

[0055] Group the variation values ​​whose timestamps fall within the same time alignment window into the same common time position;

[0056] When the same variable has multiple magnitude terms within a time alignment window, the sum of these terms is used to determine the representative magnitude of the change at the corresponding common time position.

[0057] When there is no change magnitude term, the variable is marked as missing data at the corresponding common time position, and the common time position is not included in the trend consistency calculation involving the variable;

[0058] Through the above processing, a representative sequence of change magnitudes is formed, with each variable arranged according to its common time position;

[0059] In step one, the process of determining the direction of variable change is as follows:

[0060] The AI ​​inference and computing unit reads the representative change range records of each variable during a historical stable period prior to the current business decision analysis time.

[0061] Historically stable periods are continuous periods during which no known system failures, business rule changes, or abnormal business events have occurred.

[0062] For any variable, the threshold for positive change magnitude is determined based on the distribution of positive values ​​representing the magnitude of change within a historical stable period, and the threshold for negative change magnitude is determined based on the distribution of negative values ​​representing the magnitude of change.

[0063] The positive change threshold and the negative change threshold represent the normal fluctuation boundaries that a variable must exceed to be judged as rising or falling, respectively. The numerical range formed between the two is the range of insignificant change.

[0064] When historical records are insufficient to form a fluctuation boundary in the corresponding direction, the threshold of the change amplitude in that direction is determined based on the minimum acquisition accuracy of the corresponding variable.

[0065] When the magnitude of the change is greater than the threshold for positive change, the direction of change is determined to be upward.

[0066] When the magnitude of the change is less than the threshold for negative change, the direction of change is determined to be downward.

[0067] When the magnitude of change is greater than or equal to the threshold for negative change and less than or equal to the threshold for positive change, the direction of change is determined to be insignificant.

[0068] The direction of change is not determined for common time locations where data is missing.

[0069] In step one, the process of obtaining trend consistency and pairing variables is as follows:

[0070] For any two variables, when both are marked as rising or falling at the same common time position, the common time position is determined as the valid time position;

[0071] When any variable is marked as having little change or missing data, the common time position is not determined as a valid time position;

[0072] In the valid time position, when both variables are either increasing or both are decreasing, the corresponding valid time position is determined as the time position with the same direction.

[0073] When one variable increases and another variable decreases, the corresponding valid time position is not determined as a time position with the same direction.

[0074] The number of valid time locations and the number of time locations with consistent direction are counted separately. The ratio of the number of time locations with consistent direction to the total number of valid time locations is determined as the trend consistency ratio between the two variables.

[0075] The preset trend consistency threshold represents the minimum trend consistency ratio required for two variables to be determined to have a stable, same-direction relationship. It is determined based on the distribution of trend consistency ratios of different variable pairs in historical business data and the minimum requirements for the stability of the association in business decision analysis tasks.

[0076] When the total number of valid time locations is zero, it is determined that the two variables do not meet the pairing condition.

[0077] When the total number of valid time locations is not zero and the trend consistency ratio is greater than or equal to the preset trend consistency threshold, a variable pairing relationship is established between the two.

[0078] Each variable pairing relationship includes the variable identifiers and sources of the two variables, the effective time position, the time position with consistent direction, the change direction and the representative change magnitude corresponding to each effective time position, and the proportion of consistent trends;

[0079] In step one, the process of generating the set of business variable relationships is as follows:

[0080] The AI ​​inference computing unit aggregates all variable pairing relationships that meet the pairing conditions, forms a set of business variable association relationships, and stores them in a distributed storage device by the data processing server.

[0081] The business variable association set is used to represent the candidate trend association between different business variables, but does not directly indicate that a causal relationship has been formed between the variables.

[0082] Step 2: Extract the target variable association nodes that correspond to the target variable based on the business variable association set. Construct a perturbation response inference model based on the target variable and its association nodes. Perform amplitude perturbation on each target variable association node and record the changes before and after the perturbation. Arrange the change results in chronological order, extract the response time position of each variable and compare the order of occurrence. Take the variable with the earliest response as the starting node and connect them one by one according to the response order to form a path structure, thus obtaining a set of causal relationship paths.

[0083] In step two, the target variable and the set of nodes associated with the target variable are determined.

[0084] In some embodiments, the data processing server reads the pre-specified business variables in the current business decision analysis task and determines them as target variables;

[0085] The AI ​​inference computing unit traverses the set of business variable relationships, searches for variable pairings that contain the target variable, and extracts another business variable other than the target variable from each variable pairing.

[0086] Each extracted business variable is identified as a target variable association node, and the representative change range sequence of the corresponding business variable, the variable pairing relationship between it and the target variable, and the corresponding trend consistency ratio are recorded.

[0087] Summarize all nodes associated with the target variable to form a set of nodes associated with the target variable;

[0088] The target variable itself is not considered a target variable associated node in the target variable associated node set;

[0089] When the set of nodes associated with the target variable is empty, stop the current path reasoning and record the processing result for the absence of nodes associated with the target variable.

[0090] In step two, the process of constructing the disturbance response inference model is as follows:

[0091] The AI ​​inference computing unit reads the target variable and the continuous historical representative change sequence of each target variable's associated nodes from the distributed storage device before the current business decision analysis time, and arranges them according to their common time position;

[0092] Training samples are constructed using consecutively arranged common time positions as units;

[0093] For each training sample, the representative change magnitude, variable pairing relationship, trend consistency ratio and common time position of the target variable and the associated nodes of each target variable within multiple consecutive common time positions are determined as training input, and the actual representative change magnitude of the target variable and the associated nodes of each target variable within the subsequent preset response time range are determined as training output.

[0094] The preset response time range is a time interval that starts from the first common time position after the end of the training input and continuously contains a preset number of common time positions. It is used to limit the range of model output and the recognition range of effective response time positions.

[0095] The preset quantity is determined based on the number of common time locations that the associated variables experienced after significant changes in historical business data to complete the main response, and remains unchanged in the same business decision analysis task;

[0096] The disturbance response inference model defines the connections between variables that can directly transmit changes based on the set of relationships between business variables, and adjusts the intensity of change transmission of the corresponding connections according to the trend consistency ratio; based on the change status of each variable within the preceding common time position, it outputs the predicted change magnitude of the target variable and the associated nodes of each target variable within the preset response time range.

[0097] Input the training samples into the model, and adjust the model parameters according to the difference between the predicted change and the actual representative change, so that the difference is gradually reduced.

[0098] When the model no longer reduces the prediction difference on historical samples that have not been adjusted, stop adjusting the parameters and save the model parameters to obtain the perturbation response inference model.

[0099] When historical data is insufficient to form training samples containing complete training inputs and outputs, pause model building and record the lack of historical data.

[0100] In step two, the process of handling single-node amplitude perturbations is as follows:

[0101] After the perturbation response inference model is constructed, the AI ​​inference computing unit sequentially applies amplitude perturbations to the nodes associated with each target variable, processing only one node associated with a target variable at a time.

[0102] For the currently disturbed node, select the representative change magnitude term at the end of the current inference input time range from its representative change magnitude sequence as the disturbance object, and determine their common time position as the disturbance time position;

[0103] The disturbance amplitude is determined based on the historical representative change amplitude records of the disturbed node, the normal change amplitude range, and the minimum acquisition accuracy.

[0104] The magnitude of the change after the disturbance should be within the normal range of the corresponding variable, and should form a identifiable non-zero difference from the original magnitude of the change.

[0105] The disturbance amplitude is superimposed on the original representative change amplitude;

[0106] When the positive superposition exceeds the normal range of variation, a disturbance amplitude with the same value but opposite direction is used;

[0107] When neither direction can generate a non-zero change within the normal range of change, do not perform perturbation inference on the current node and record that there is no perturbation range that can be implemented.

[0108] The amplitude perturbation is performed in the computed copy;

[0109] Except for the change in the representative change magnitude of the disturbed node at the disturbance time position, the representative change magnitude sequence of the target variable and the nodes associated with the target variable remains unchanged;

[0110] In step two, the reasoning process for the response before and after the disturbance is as follows:

[0111] The AI ​​inference computing unit calls the disturbance response inference model to perform inference based on the input data before and after the disturbance, respectively.

[0112] Input the target variable without perturbation and the representative change magnitude sequence of each target variable's associated node, variable pairing relationship, trend consistency ratio, and common time position into the model to obtain the baseline change magnitude output within the preset response time range;

[0113] Replace the original representative change magnitude of the perturbed node at the perturbed time position with the representative change magnitude after the perturbation, keep other input data and model parameters unchanged, and perform inference again to obtain the intervention change magnitude output within the same preset response time range;

[0114] In step two, the process of identifying the disturbance response amplitude and effective response time is as follows:

[0115] For the target variable and any node associated with the target variable, calculate the difference between the output of the intervention change magnitude and the output of the baseline change magnitude of the corresponding variable at the same common time position, and determine the obtained difference as the disturbance response magnitude;

[0116] The disturbance response amplitudes are arranged from earliest to latest according to their common time position to form a response amplitude sequence;

[0117] The AI ​​inference and computing unit determines the threshold for judging the perturbation response magnitude for the target variable and the associated nodes of each target variable.

[0118] The disturbance response amplitude judgment threshold represents the minimum judgment boundary that the absolute value of the disturbance response amplitude must reach when the corresponding variable is determined to generate a valid disturbance response. It is determined based on the minimum acquisition accuracy of the corresponding variable, the range of historical normal variation amplitude, and the range of prediction differences formed by the model on historical samples without artificial disturbances.

[0119] For the disturbed node, the disturbance time location is determined as its response occurrence time location, and the actual disturbance magnitude is determined as its direct response magnitude;

[0120] For the target variable and each target variable associated node except the disturbed node, starting from the first common time position after the disturbance time position, compare the absolute value of the disturbance response amplitude with the corresponding disturbance response amplitude judgment threshold in chronological order within the preset response time range.

[0121] When the absolute value of the disturbance response amplitude is greater than or equal to the corresponding disturbance response amplitude judgment threshold for the first time, it is determined that the corresponding variable has generated a valid response, and the common time position that first meets the condition is determined as the first valid response time position, and the disturbance response amplitude at the common time position is determined as the first valid disturbance response amplitude.

[0122] Variables that do not reach the corresponding disturbance response magnitude judgment threshold within the preset response time range are not considered as valid response variables for this disturbance.

[0123] When the target variable does not produce a valid response, the current node perturbation does not form a causal path pointing to the target variable;

[0124] In step two, the process of determining the starting node, response order, and response delay is as follows:

[0125] The disturbed nodes and all valid response variables are uniformly used as variables for response time comparison;

[0126] The time position of the response of the disturbed node is the disturbance time position, and the time position of the response of other valid response variables is the corresponding first valid response time position;

[0127] Arrange the variables in order of their response times from earliest to latest, and determine the variable with the earliest response time as the starting node;

[0128] When the responses of multiple variables occur at the same time, they are recorded as synchronous response variables, and no directed connections indicating the order of occurrence are established between synchronous response variables.

[0129] For two variables whose responses occur at different times, the response delay value between the corresponding variables is determined based on the number of common time positions between their response occurrence times and the length of the time alignment window.

[0130] In step two, the process of constructing the causal relationship path is as follows:

[0131] The AI ​​inference computing unit uses the starting node and the effective response variable as path nodes, and establishes directed connections based on the response occurrence time and location and the set of business variable relationships.

[0132] For any two path nodes, a directed connection from the former node to the latter node is established only if the response of the latter node occurs later than that of the former node, and if there is a variable pairing relationship between the two in the business variable association set.

[0133] No directed connections are established between synchronous response nodes;

[0134] Starting from the initial node, search for subsequent nodes that can be connected in order of the time the response appears, from earliest to latest.

[0135] When a node corresponds to multiple subsequent nodes that meet the conditions, the path is constructed along each subsequent node.

[0136] Do not add the same node repeatedly along the same path;

[0137] When a path reaches the target variable, the corresponding node sequence is identified as a candidate causal path; if the path cannot be continued and the target variable has not been reached, the corresponding path branch is terminated.

[0138] By summarizing all candidate causal relationship paths formed after perturbing the nodes associated with each target variable, a set of causal relationship paths is obtained.

[0139] Each candidate causal path includes the starting node, target variable, other valid response nodes, node order, response occurrence time and location of each node, response delay value between adjacent nodes, direct response magnitude or first valid disturbance response magnitude of each node, and path connection direction.

[0140] The causal path set only represents candidate directional paths formed after perturbation response reasoning and response time comparison;

[0141] Step 3: Extract the response changes of each path node based on the causal path set, perform continuous comparison of the change process of adjacent nodes and identify the change transmission order, and retain the paths that meet the continuous transmission characteristics; decompose the contribution of each node in the retained path in the overall change process, identify low-impact nodes according to the contribution ratio, merge adjacent low-impact nodes and adjust the path structure to obtain the causal link structure set.

[0142] In step three, the process of determining the responsiveness of path nodes is as follows:

[0143] In some embodiments, the AI ​​inference computing unit reads the direct response magnitude or the first effective perturbation response magnitude of the nodes in each candidate causal path;

[0144] Based on the historical normal range of change of each variable, its response magnitude is converted into the nodal response degree.

[0145] The node response level is the relative degree of the absolute value of the corresponding response amplitude within the historical normal range of the variable;

[0146] When the historical normal variation range is zero, the minimum acquisition accuracy of the corresponding variable is used as the scale benchmark.

[0147] The node response level is used to characterize the strength of responses to different variables within the same comparison space, and serves as a common input for continuous propagation verification and node contribution decomposition.

[0148] In step three, the process of verifying the continuous path transmission features is as follows:

[0149] For any candidate causal relationship path, the AI ​​inference computing unit checks the relationship between each adjacent node one by one in the order of nodes pointing from the starting node to the target variable;

[0150] The business variable association set retains only the same-direction related variable pairs whose trend consistency ratio reaches the preset trend consistency threshold. Therefore, the expected change transmission direction between adjacent nodes is determined to be same-direction transmission.

[0151] A set of adjacent nodes is considered to have a continuous transmission relationship when both of the following conditions are met:

[0152] The response of the later node appears later than that of the previous node;

[0153] The preceding and following nodes have a variable pairing relationship in the business variable association set;

[0154] The absolute value of the first effective disturbance response amplitude of the subsequent node is greater than or equal to its disturbance response amplitude determination threshold.

[0155] The response amplitudes of the preceding and following nodes have the same positive and negative directions;

[0156] The responsiveness of the preceding node is greater than that of the following node;

[0157] The relationship between adjacent nodes that does not simultaneously meet the above conditions is determined to be a discontinuous transmission relationship;

[0158] In step three, the process of selecting effective causal paths is as follows:

[0159] The number of continuously propagated relationships in candidate causal relationship paths is counted, and the ratio of this number to the number of relationships between all adjacent nodes in the path is determined as the continuous propagation ratio.

[0160] The total number of relationships between all adjacent nodes is equal to the total number of nodes in the path minus one.

[0161] When the continuous transmission ratio is greater than or equal to the preset continuous transmission ratio threshold, the path is determined as a valid causal path;

[0162] When the continuous transmission ratio is less than the preset continuous transmission ratio threshold, the contribution decomposition of subsequent nodes is not performed on the path.

[0163] The preset continuous transmission ratio threshold represents the minimum degree of continuous transmission required to preserve candidate causal relationship paths, and is determined based on the distribution of continuous transmission ratios in historical paths and the minimum degree of continuous connectivity required to form auxiliary decision rules.

[0164] In step three, the process of decomposing the contribution of path nodes is as follows:

[0165] For any valid causal path, read the node response degree of each active node in the path, excluding the target variable;

[0166] The total path response level is obtained by summing the response levels of all affected nodes along the same path.

[0167] The ratio of the node's response level to the total path response level is determined as the node's contribution ratio.

[0168] Starting from the first effective response time of the target variable, the absolute values ​​of the disturbance response amplitudes of the target variable at each common time position within the preset response time range are accumulated to obtain the total response change amplitude of the target variable.

[0169] Multiply the contribution ratio of any node by the total change in the target variable to obtain the node contribution value.

[0170] When the total response degree of a path is zero or the total response change of the target variable is zero, the node contribution decomposition is not performed on the path and it is excluded from the causal link structure set.

[0171] In step three, the process of identifying low-impact nodes is as follows:

[0172] The preset low impact judgment threshold represents the minimum contribution ratio required for an active node to be retained as an independent node, and is determined based on the distribution of node contribution ratios in historical effective causal paths and the minimum degree of node activity required to be independently retained when forming auxiliary decision rules.

[0173] When the contribution ratio of an active node is less than the preset low-impact judgment threshold, it is marked as a low-impact node and its node position in the original path is recorded.

[0174] When the contribution ratio is greater than or equal to the preset low-impact judgment threshold, it is retained as an independent node.

[0175] In step three, the process of merging consecutive low-impact nodes and reconstructing paths is as follows:

[0176] When there are two or more consecutively arranged low-impact nodes in an effective causal path, the consecutive low-impact nodes are merged into one merge node, and the node contribution values ​​of each merged low-impact node are summed to obtain the merge contribution value.

[0177] The connection between the merged low-impact node closest to the starting node and its upstream node is determined as the connection between the merged node and its upstream node.

[0178] The connection between the merged low-impact node closest to the target variable and its downstream nodes is determined as the connection between the merged node and the downstream node, and the corresponding response delay values ​​are retained respectively.

[0179] When consecutive low-impact nodes contain the original path start node, the merged node is determined as the start node of the reconstructed path, and the disturbance time position and direct response magnitude of the original start node are retained.

[0180] When consecutive low-impact nodes are directly connected to the target variable, the merged node maintains its connection with the target variable;

[0181] The merge node records the variable identifier, original order, direction of change, response occurrence time and location, response delay value between adjacent original nodes, node contribution value and contribution ratio of each original low-impact node; the merge node is only used to simplify the path structure and is not used as a new business variable.

[0182] In step three, the process of generating the causal link structure set is as follows:

[0183] For independent nodes that did not participate in the merger, their node contribution value is used as the contribution value after structural adjustment; for merged nodes, their merged contribution value is used as the contribution value after structural adjustment.

[0184] The total adjusted contribution value is obtained by summing the structural adjustments of all independent and merged nodes in the same path.

[0185] The ratio of the adjusted contribution value of any node to the total adjusted contribution value is determined as the adjusted contribution ratio of that node.

[0186] When the total adjusted contribution value is zero, the corresponding path will be excluded from the causal link structure set.

[0187] The effective causal paths that have completed continuous transmission verification, node contribution decomposition, low-impact node identification, and path reconstruction are summarized to form a set of causal link structures.

[0188] Each causal link includes the starting node, the target variable, the adjusted node order, the node connection relationship, the node contribution value of the independently acting node, the combined contribution value of the merging node, the adjusted contribution ratio of each node, the response delay value between adjacent nodes, the original node composition of the merging node, and the continuous transmission of verification results.

[0189] The data processing server simultaneously stores the node arrangement order, node change direction, representative change magnitude, response latency value, and contribution ratio of each valid causal path as historical valid causal path records in a distributed storage device.

[0190] Step 4: Filter complete paths for the causal link structure set, determine the adjacency weight of each node based on the connection distance between each node and the target variable, determine the overall strength of the path based on the contribution ratio of each node and the corresponding adjacency weight, and sort the complete paths according to the overall strength of the path to identify key causal links, extract the node combination relationship in the key causal links and map it into a rule expression form, organize the rule structure according to the node arrangement order, contribution ratio and path direction to obtain the set of auxiliary decision rules;

[0191] In step four, the process of screening for the integrity of the causal link structure is as follows:

[0192] In some embodiments, the AI ​​inference computing unit sequentially examines each causal link in the causal link structure set;

[0193] When a causal link has a clear starting node and target variable, the node connection relationship between the starting node and the target variable is continuous, each independent node or merging node has a clear path position and adjusted contribution ratio, and each adjacent node has a response delay value, the causal link is determined to be a structurally complete causal link.

[0194] Causal links that do not meet any of the above conditions will not be included in the subsequent adjacent weight and overall path strength calculation.

[0195] In step four, the process of determining the adjacency order and adjacency weight is as follows:

[0196] For a structurally complete causal link, the adjacency order is determined based on the number of connections between each independent acting node or merging node and the target variable.

[0197] Nodes directly connected to the target variable are first-order adjacency nodes, nodes connected to the target variable via an intermediate node are second-order adjacency nodes, and the adjacency order of the remaining nodes increases sequentially according to the connection distance from the target variable to the starting node.

[0198] Calculate the reciprocal of the adjacency order of each node in the same causal link, and sum the reciprocals to obtain the total adjacency reciprocal;

[0199] The ratio of the reciprocal of the adjacency order of any node to the total number of adjacency reciprocals is determined as the adjacency weight of that node.

[0200] In step four, the process of obtaining the overall path strength is as follows:

[0201] For any structurally complete causal link, the independent action nodes and merging nodes in the causal link, excluding the target variable, are numbered sequentially from 1 to n.

[0202] For the i-th node, its adjacency order di is determined based on the number of connections between the node and the target variable. Specifically, the adjacency order di of a node directly connected to the target variable is 1; the adjacency order di of a node connected to the target variable via an intermediate node is 2; and the adjacency order of the remaining nodes is determined sequentially according to the number of connections between them and the target variable.

[0203] Calculate the reciprocal of the adjacency order of each node, 1 / di, and sum the reciprocals of the adjacency orders of all nodes in the same causal link to obtain the total number of adjacency reciprocals H:

[0204] The ratio of the reciprocal of the adjacency order of the i-th node to the total number of adjacency reciprocals is determined as the adjacency weight wi of that node:

[0205] In this context, the sum of the adjacency weights of all nodes in the same causal link is 1; the smaller the connection distance between a node and the target variable, the smaller the corresponding adjacency order and the larger the adjacency weight.

[0206] Read the adjusted contribution ratio pi of the i-th node, multiply the adjusted contribution ratio pi by the corresponding adjacent weight wi, and obtain the weighted contribution value qi of the i-th node:

[0207] The overall path strength S of the causal link is obtained by summing the weighted contributions of all independent and merging nodes in the same causal link:

[0208] Since qi is obtained by multiplying the adjusted contribution ratio pi of the i-th node with the corresponding adjacency weight wi, the overall path strength S represents the comprehensive value formed by weighting the adjusted contribution ratios of each node in the same causal link with the adjacency weight.

[0209] Therefore, the overall path strength represents the comprehensive value formed by weighting the contribution ratio of each node in the same causal link after considering the connection distance between the node and the target variable.

[0210] When any node lacks an adjusted contribution ratio or its adjacency order cannot be determined, the overall path strength of the corresponding causal link is not calculated, and the causal link is excluded from the candidate range of critical causal links.

[0211] In step four, the process of screening key causal links is as follows:

[0212] All structurally complete causal links whose overall path strength can be calculated are arranged from high to low according to their overall path strength.

[0213] Based on the number of auxiliary decision-making rules required for the current business decision analysis task, select the corresponding number of causal links from the sorting results and identify them as key causal links.

[0214] When the number of available causal links is less than the required number, all available causal links are identified as critical causal links;

[0215] When there are multiple causal links with the same overall path strength at the sorting boundary, the causal links with the same overall path strength as the boundary causal links will be retained together.

[0216] For each key causal link, record its starting node, target variable, independent or merged nodes, node order, node connection relationship, direction of change of each original business variable node, contribution ratio of each node after adjustment, adjacency order, adjacency weight, response delay value and overall path strength, and associate it with the corresponding historical effective causal path records.

[0217] In step four, the process of mapping key causal links to auxiliary decision rules is as follows:

[0218] The data processing server sends records of each key causal link and its associated historical valid causal paths to the decision rule generation module.

[0219] For any key causal link, the decision rule generation module reads the node data corresponding to the starting node, intermediate independent nodes, merging nodes, and target variable in sequence, following the path direction from the starting node to the target variable.

[0220] For any original business variable node in the critical causal link, the node data read includes at least the variable identifier, direction of change, node arrangement position, response occurrence time and position, and response delay value between the node and its neighboring nodes.

[0221] The decision rule generation module extracts representative change magnitudes of the corresponding original business variable nodes under the same change direction from historical valid causal path records that have the same node arrangement relationship and the same node change direction as the key causal link. The minimum value among the extracted representative change magnitudes is determined as the lower limit of the change magnitude range of the node, and the maximum value is determined as the upper limit of the change magnitude range of the node.

[0222] When there is no historically valid causal path record that can be used to determine the range of change of any original business variable node, auxiliary decision-making rules are not generated using the corresponding key causal link.

[0223] For any original business variable node, the variable identifier, change direction, lower limit of change range, and upper limit of change range of that node are combined to form the corresponding node change conditions.

[0224] When there is a business variable in the current business data that has the same variable identifier as the variable in the node change condition, and the direction of change of the business variable is the same as the direction of change in the node change condition, and the magnitude of change of the business variable is greater than or equal to the lower limit of the magnitude range and less than or equal to the upper limit of the magnitude range, the corresponding node change condition is determined to be true.

[0225] The node change conditions corresponding to the starting node and the intermediate independent action node located between the starting node and the target variable in the key causal link are determined as the rule antecedent conditions.

[0226] The variable identifier, direction of change, lower limit of the range of change, and upper limit of the range of change of the target variable are combined to form the rule consequent.

[0227] The decision rule generation module arranges the antecedent conditions of each rule according to the order of nodes pointing from the starting node to the target variable in the key causal chain.

[0228] For any two adjacent original business variable nodes, the node change condition corresponding to the former node is determined as the first occurrence condition, and the node change condition corresponding to the latter node is determined as the second occurrence condition, and the response delay value recorded between the two is read.

[0229] In the auxiliary decision-making rule, the timing constraint between the two node change conditions is defined as the condition that the subsequent condition is established later than the condition that the preceding condition is established, and the time interval between the establishment of the subsequent condition and the preceding condition is equal to the corresponding response delay value.

[0230] According to the node arrangement order, the preconditions of each rule and the temporal constraints between adjacent preconditions are organized sequentially to form the rule preconditions.

[0231] The rule antecedent is determined to be valid only if all node change conditions in the rule antecedent are met, and the order in which the node change conditions are met and the time interval between the meeting of adjacent node change conditions both conform to the corresponding timing constraints.

[0232] For independent nodes that did not participate in the merging, record the adjusted contribution ratio of the independent node in relation to its node change conditions.

[0233] When there is a merging node in the critical causal link, the decision rule generation module reads the variable identifier, original order, change direction, change range, and response delay value between adjacent original low-impact nodes of each original low-impact node contained in the merging node, and generates the node change conditions corresponding to each original low-impact node.

[0234] According to the original arrangement order of each original low-impact node in the causal path before merging, the node change conditions corresponding to each original low-impact node are arranged, and the order and time interval of the establishment of the adjacent node change conditions are limited according to the response delay value between adjacent original low-impact nodes, forming a combined condition unit.

[0235] When all the original low-impact node change conditions in the combined condition unit are met, and the order in which the original low-impact node change conditions are met and the time interval between adjacent conditions are met, the combined condition unit is determined to be met.

[0236] In the rule antecedent, the combined condition unit represents the merging node, and the adjusted contribution ratio of the merging node is associated with the combined condition unit and recorded.

[0237] The merged node serves only as a structural identifier for the combined condition unit and is not used as a new business variable for variable identification matching with the current business data; the current business data is matched with the node change conditions corresponding to each original low-impact node in the combined condition unit.

[0238] The overall path strength of key causal links is associated with and recorded with the corresponding auxiliary decision-making rules.

[0239] A decision support rule includes the rule antecedent, rule consequent, the order of the changing conditions of each node in the rule antecedent, the response delay constraints between the changing conditions of adjacent nodes, the adjusted contribution ratio of each independent node or merged node, and the overall path strength of the key causal link.

[0240] The auxiliary decision rule is expressed as follows: when the business variables corresponding to the starting node and each intermediate node change in the corresponding direction according to the node arrangement order of the key causal link, the representative change amplitude of each business variable is within the corresponding change amplitude range, and the time interval between the establishment of the change conditions of adjacent nodes meets the corresponding response delay value, the change direction and change amplitude range of the target variable are output.

[0241] During the invocation of auxiliary decision-making rules, the data processing server generates, according to step one, representative changes and directions of the current business variables at each common time position.

[0242] The data processing server compares the variable identifier, direction of change, and magnitude of change of the current business variable with the corresponding node change conditions in the order of the nodes in the rule antecedents, and records the common time position that satisfies the node change conditions.

[0243] When all node change conditions or combination conditions in the rule antecedent are met, the data processing server further compares the order in which the node change conditions are met and calculates the time interval between the meeting of adjacent node change conditions based on their corresponding common time positions.

[0244] When the order in which the change conditions of each node are met is consistent with the order of the nodes in the critical causal link, and the time interval between the meeting of the change conditions of each adjacent node is equal to the corresponding response delay value, the auxiliary decision rule is triggered.

[0245] Once the auxiliary decision-making rule is triggered, the data processing server outputs the variable identifier, direction of change, and range of change of the target variable in the rule's consequent, and simultaneously outputs the adjusted contribution ratio of each independent or merged node and the overall path strength of the corresponding key causal link, as auxiliary decision-making basis for the current business decision analysis task.

[0246] Example 2

[0247] like Figure 3 As shown in the embodiment of the present invention, a big data mining-based decision-making auxiliary method and system based on AI includes the following modules:

[0248] Variable association module: Obtain transaction record time series, device operating status voltage value series and user behavior click frequency series from the log collection terminal, calculate the difference between adjacent time points to form a change amplitude series, align the change amplitude items in different change amplitude series to the corresponding common time position and compare the change direction, pair variables with consistent change trends and establish corresponding relationships to obtain a set of business variable association relationships;

[0249] Causal reasoning module: Extract the target variable association nodes based on the business variable association set, construct a disturbance response reasoning model based on the target variable and the target variable association nodes, implement amplitude disturbances on each node and record the changes before and after, sort the change results, extract the response time position of each variable and compare the order, take the earliest response variable as the starting node, connect them according to the response order to form a path structure, and obtain the causal relationship path set;

[0250] Link reconstruction module: Extract the response changes of each path node according to the causal path set, compare the continuity of adjacent nodes and identify the order of change transmission, retain the path that meets the characteristics of continuous transmission, decompose the contribution of each node in the retained path to the overall change, identify low-impact nodes according to the contribution ratio, merge adjacent low-impact nodes and adjust the path structure to obtain the causal link structure set.

[0251] Rule construction module: Filter complete paths for the causal link structure set, determine the adjacency weight of each node based on the connection distance between each node and the target variable, determine the overall strength of the path based on the contribution ratio of each node and the corresponding adjacency weight, sort the key causal links according to the overall strength of the paths, extract the node combination relationship and map it into rule expression, organize the rule structure according to the node order, contribution ratio and path direction, and obtain the set of auxiliary decision rules.

[0252] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the present invention should still fall within the scope of the present invention.

Claims

1. An AI-based big data mining-assisted decision-making method, characterized in that: Includes the following steps: Obtain the transaction record time series, device operating status voltage value series and user behavior click frequency series from the log collection terminal, calculate the difference between adjacent time points to form a change amplitude series, align the change amplitude items in different change amplitude series to the corresponding common time position and compare the change direction, pair variables with consistent change trends and establish corresponding relationships to obtain a set of business variable association relationships; Based on the set of relationships between business variables, target variable association nodes are extracted. Based on the target variable and target variable association nodes, a disturbance response inference model is constructed. Amplitude disturbances are applied to each node and the changes before and after are recorded. The change results are sorted. The response time position of each variable is extracted and the order is compared. The earliest response variable is taken as the starting node and connected in the order of response to form a path structure, thus obtaining a set of causal relationship paths. The response changes of each path node are extracted from the causal path set. The continuity of adjacent nodes is compared and the order of change transmission is identified. Paths that meet the characteristics of continuous transmission are retained. The contribution of each node in the retained path to the overall change is decomposed. Low-impact nodes are identified according to the contribution ratio. Adjacent low-impact nodes are merged and the path structure is adjusted to obtain the causal link structure set. For the set of causal link structures, complete paths are selected. The adjacency weight of each node is determined based on the connection distance between each node and the target variable. The overall strength of the path is determined based on the contribution ratio of each node and its corresponding adjacency weight. Key causal links are identified by sorting them according to their overall strength. The node combination relationship is extracted and mapped into a rule expression. The rule structure is organized according to the node order, contribution ratio, and path direction to obtain a set of auxiliary decision rules.

2. The AI-based big data mining-assisted decision-making method according to claim 1, characterized in that: The process of assigning the change magnitude items to a common time position and determining the representative change magnitude is as follows: The length of the time alignment window is determined based on the statistical period of transaction record data, the sampling period of equipment operating status voltage sampling data, and the statistical period of user behavior click frequency data. The time alignment window is continuously divided with the starting timestamp of the continuous collection period as the starting point. The change amplitude items whose timestamps fall within the same time alignment window are grouped into the same common time position, and the cumulative result of multiple change amplitude items of the same variable within the common time position is determined as the representative change amplitude.

3. The AI-based big data mining-assisted decision-making method according to claim 2, characterized in that: The process of determining the direction of change and variable pairing is as follows: Based on the representative change range records of each variable within a historical stable period, thresholds for positive and negative change ranges are determined. Based on these thresholds, the representative change ranges at each common time position are marked as increasing, decreasing, or showing no significant change. Common time positions where both variables are marked as increasing or decreasing are identified as valid time positions. Valid time positions where the two variables change in the same direction are identified as direction-consistent time positions. The trend consistency ratio is determined based on the ratio of the number of direction-consistent time positions to the number of valid time positions. Variable pairing relationships are established when the trend consistency ratio reaches the trend consistency threshold.

4. The AI-based big data mining-assisted decision-making method according to claim 1, characterized in that: The process of constructing the disturbance response inference model is as follows: Before perturbating the magnitude of the nodes associated with the target variable, obtain the historical representative change magnitude sequence of the target variable and the nodes associated with the target variable arranged according to their common time position. Use the representative change magnitude, variable pairing relationship, trend consistency ratio and common time position of each variable in the preceding consecutive common time positions as training input, and use the actual representative change magnitude of each variable in the subsequent response time range as training output to construct training samples. The direct connections between variables are defined by the set of relationships between business variables. The intensity of the change propagation of the direct connections is adjusted according to the trend consistency ratio. The model parameters are adjusted according to the difference between the training output and the predicted change magnitude to obtain the disturbance response inference model.

5. The AI-based big data mining-assisted decision-making method according to claim 4, characterized in that: The process for determining the amplitude of a single-node disturbance and its response is as follows: One target variable associated node is selected sequentially as the perturbed node. In the computation replica, the perturbation amplitude is superimposed on the representative change amplitude of the perturbed node at the end of the inference input time range, while keeping the data of the target variable and other target variable associated nodes unchanged. By inputting the pre-disturbance data and post-disturbance data into a disturbance response inference model with the same parameters, the baseline change magnitude and the intervention change magnitude are obtained. The difference between the intervention change magnitude and the baseline change magnitude of the same variable at the same common time position is determined as the disturbance response magnitude.

6. The AI-based big data mining-assisted decision-making method according to claim 5, characterized in that: The process of determining the effective response and candidate paths is as follows: Based on the minimum acquisition accuracy of each variable, the range of historical normal variation, and the range of prediction differences of the disturbance response inference model when no disturbance is applied, the threshold for judging the disturbance response amplitude of each variable is determined. The time position of the disturbance of the disturbed node is determined as the time position of its response occurrence, and the common time position when the absolute value of the disturbance response amplitude of other variables first reaches the corresponding judgment threshold is determined as the time position of the response occurrence. The variable with the earliest response time is identified as the starting node. Directed connections are established between path nodes with variable pairing relationships in order of response time from earliest to latest. The sequence of nodes reaching the target variable is identified as the candidate causal relationship path.

7. The AI-based big data mining-assisted decision-making method according to claim 1, characterized in that: The verification process for continuously propagated features is as follows: The degree of node response is determined based on the response amplitude of each path node and the historical normal range of the corresponding variable. When the response of the next node in an adjacent node occurs later than that of the previous node, there is a variable pairing relationship between the two, the response amplitude of the next node reaches the corresponding judgment threshold, the response amplitudes of the two nodes are in the same direction, and the node response degree of the previous node is greater than that of the next node, the adjacent nodes are determined to be in a continuous transmission relationship. The continuous propagation ratio is determined by the ratio of the number of continuous propagation relationships to the number of relationships between adjacent nodes in the path, and candidate causal relationship paths that reach the continuous propagation ratio threshold are determined as valid causal paths.

8. The AI-based big data mining-assisted decision-making method according to claim 7, characterized in that: The process of determining node contributions and path reconstruction is as follows: The total path response is obtained by summing the node response of each node in the effective causal path except for the target variable. The ratio of the node response of each node to the total path response is determined as the contribution ratio. The node contribution value is determined based on the contribution ratio and the change in the total response of the target variable. Nodes with a contribution ratio lower than the low impact judgment threshold are identified as low impact nodes. Merge two or more consecutively arranged low-impact nodes into a merge node, sum the node contribution values ​​of the merged low-impact nodes, and reconstruct the connection relationship between the merge node and upstream and downstream nodes according to the original arrangement of the merged low-impact nodes, and redetermine the contribution ratio of each independent node and the merge node.

9. The AI-based big data mining-assisted decision-making method according to claim 1, characterized in that: The process of determining the overall path strength based on adjacency weights and generating auxiliary decision rules based on key causal links is as follows: The adjacency order of a node is determined by the number of connections between each independent or merged node and the target variable. The adjacency weight of the corresponding node is determined by the ratio of the reciprocal of the adjacency order of each node to the sum of the reciprocals of the adjacency orders of all nodes in the same path. Multiply the contribution ratio of each node by its corresponding adjacency weight to obtain the weighted contribution value of the corresponding node, and sum the weighted contribution values ​​of all nodes in the same path to obtain the overall strength of the path. The structurally complete paths are sorted from high to low based on their overall strength to identify key causal links. The variable identifiers, change directions, and change ranges corresponding to the starting and intermediate nodes in the key causal links are organized as node change conditions, and the node change conditions are organized as rule antecedents according to the node arrangement order. The variable identifier, direction of change, and range of change corresponding to the target variable are organized into rule consequents, and the order and time interval of the corresponding node change conditions are limited according to the response delay value between adjacent nodes.

10. An AI-based big data mining-assisted decision-making system, characterized in that: Includes the following modules: Variable association module: Obtain transaction record time series, device operating status voltage value series and user behavior click frequency series from the log collection terminal, calculate the difference between adjacent time points to form a change amplitude series, align the change amplitude items in different change amplitude series to the corresponding common time position and compare the change direction, pair variables with consistent change trends and establish corresponding relationships to obtain a set of business variable association relationships; Causal reasoning module: Extract the target variable association nodes based on the business variable association set, construct a disturbance response reasoning model based on the target variable and the target variable association nodes, implement amplitude disturbances on each node and record the changes before and after, sort the change results, extract the response time position of each variable and compare the order, take the earliest response variable as the starting node, connect them according to the response order to form a path structure, and obtain the causal relationship path set; Link reconstruction module: Extract the response changes of each path node according to the causal path set, compare the continuity of adjacent nodes and identify the order of change transmission, retain the path that meets the characteristics of continuous transmission, decompose the contribution of each node in the retained path to the overall change, identify low-impact nodes according to the contribution ratio, merge adjacent low-impact nodes and adjust the path structure to obtain the causal link structure set. Rule construction module: Filter complete paths for the causal link structure set, determine the adjacency weight of each node based on the connection distance between each node and the target variable, determine the overall strength of the path based on the contribution ratio of each node and the corresponding adjacency weight, sort the key causal links according to the overall strength of the paths, extract the node combination relationship and map it into rule expression, organize the rule structure according to the node order, contribution ratio and path direction, and obtain the set of auxiliary decision rules.

Images