A power user portrait construction method and system based on multi-label classification

Abstract

The application discloses a power user portrait construction method and system based on multi-label classification, relates to the technical field of power user portrait construction, and comprises the following steps: collecting multi-source data of power users; sequentially performing preprocessing and standardization processing on the multi-source data; constructing a power behavior event token set according to the multi-source data after standardization processing; obtaining a cross-source evidence vector corresponding to each candidate label according to the power behavior event token set; mapping the multi-source data after standardization processing and the cross-source evidence vector into a user multi-dimensional feature vector, obtaining initial prediction probabilities by using a multi-label classification model; obtaining calibrated prediction probabilities according to a label business constraint matrix; comparing the calibrated prediction probabilities with an adaptive judgment threshold to obtain a prediction label set and output the prediction label set; and through targeted technical innovation and process optimization, the integrity, accuracy and practicability of the power user portrait are significantly improved, and core support is provided for fine operation of the power industry.

Description

Technical Field

[0001] This invention relates to the field of power user profile construction technology, and more specifically, to a method and system for constructing power user profiles based on multi-label classification. Background Technology

[0002] With the deepening of power system reform and intensified competition in the electricity sales market, electricity user profiling has become a core supporting technology for precision marketing and service optimization. Currently, various electricity user-related analysis technologies have emerged in the industry, including electricity marketing channel operation methods based on big data, cloud computing, and the Internet of Things. Some technologies attempt to mine user characteristics through electricity consumption data to provide a basis for marketing strategy formulation. These technologies have been initially applied in scenarios such as electricity user classification, electricity behavior analysis, and marketing channel optimization, attempting to improve the targeting of electricity services through user data integration and support the operational optimization of electronic channels such as mobile power apps and the 95598 hotline.

[0003] Existing technologies have significant shortcomings in building electricity user profiles, making it difficult to meet the actual needs of precision marketing and service optimization.

[0004] 1. Incomplete coverage of feature data: Existing technologies mostly focus on electricity consumption data, sales data, or single-dimensional electricity consumption data, without fully integrating multi-source core data such as user profiles, daily 96-point load curves, power factor, default records, and marketing query behavior. This results in missing user feature dimensions and an inability to fully depict users' electricity consumption habits and needs.

[0005] 2. Poor adaptability of classification methods: Existing technologies mostly adopt traditional single-label classification methods, which are difficult to handle the multi-dimensional and complex behavioral characteristics of power users. They cannot simultaneously label multi-dimensional characteristics such as user load type, industry attributes, and contribution value, resulting in insufficient accuracy of user profiles.

[0006] 3. User profiles are disconnected from application scenarios: The user profiles built by existing technologies are not deeply integrated with actual application scenarios such as electricity marketing and channel optimization. Feature extraction lacks specificity and cannot provide effective support for assessing the potential of electronic channels and formulating differentiated operation strategies, resulting in limited application value of user profiles.

[0007] 4. Lack of standardized construction process: Existing technologies have not formed a standardized process of "data collection-feature extraction-classification modeling-profile output". The data processing and modeling methods lack uniformity, resulting in low profile construction efficiency and poor reusability, making it difficult to meet the profile construction needs of large-scale power users. Summary of the Invention

[0008] The purpose of this invention is to provide a method and system for constructing power user profiles based on multi-label classification, addressing the problems in existing power user profile construction processes such as simple splicing of multi-source data, lack of business constraints between labels, unstable profile label output due to fixed thresholds, and inability to distinguish between real behavioral changes and short-term abnormal disturbances during profile updates. This invention, based on the collection of multi-source power user data, transforms electricity load, transaction performance, service interaction, credit risk, and auxiliary scenario information into power behavior event tokens. Based on these event tokens, cross-source evidence vectors for candidate labels are formed. Then, user profile data is generated through label business constraint calibration and user-label adaptive threshold determination, thereby improving the accuracy, stability, and interpretability of profile labels.

[0009] The above-mentioned technical objective of the present invention is achieved through the following technical solution.

[0010] Firstly, this application provides a method for constructing power user profiles based on multi-label classification, including the following specific steps:

[0011] Collect multi-source data from electricity users, including basic profile data, electricity load data, transaction data, behavioral data, credit data, and auxiliary data;

[0012] The multi-source data is preprocessed and standardized sequentially. The preprocessing includes outlier identification and missing value completion.

[0013] Based on the standardized multi-source data, a set of power behavior event tokens for each power user within a preset time window is constructed. The power behavior event tokens include load pattern event tokens, transaction fulfillment event tokens, service interaction event tokens, credit risk event tokens, and auxiliary scenario event tokens.

[0014] Based on the set of power behavior event tokens, calculate the evidence weight of each event token relative to the candidate tags, and generate cross-source evidence vectors corresponding to each candidate tag based on the evidence weights;

[0015] The standardized multi-source data and cross-source evidence vectors are jointly mapped into user multi-dimensional feature vectors, and a multi-label classification model is used to obtain the initial predicted probability of each candidate label.

[0016] Based on the preset tag business constraint matrix, the initial prediction probability is constrained and calibrated to obtain the calibrated prediction probability. The tag business constraint matrix is ​​used to characterize the mutual exclusion relationship, hierarchical relationship and accompaniment relationship between tags.

[0017] Based on the sufficiency of evidence, stability of the label, and aging of the label, the adaptive judgment threshold for each candidate label for each power user is calculated.

[0018] The calibrated predicted probability is compared with the adaptive decision threshold to obtain the predicted label set. The predicted label set, label confidence, and label evidence source are then used as the profile data of the corresponding user and output.

[0019] Furthermore, based on the standardized multi-source data, a set of power behavior event tokens for each power user within a preset time window is constructed. Specifically, this includes: extracting load change rate, peak-valley load, duration, and periodic repetition features from electricity load data to generate load pattern event tokens; extracting payment time, arrears duration, overdue payment frequency, and electricity price execution information from transaction data to generate transaction fulfillment event tokens; extracting query frequency, business processing type, fault reporting, and electricity consultation content from behavior data to generate service interaction event tokens; extracting records of defaulted electricity use, electricity theft, overdue payments, and electricity compliance from credit data to generate credit risk event tokens; and extracting regional, seasonal, meteorological, holiday, and industry production cycle information from auxiliary data to generate auxiliary scenario event tokens.

[0020] Furthermore, the load pattern event token is obtained as follows: Let the normalized load sequence of the i-th user within the t-th time window be: ; In the formula, This represents the standardized load sequence of the i-th user within the t-th time window. Let M represent the standardized load value of the i-th user at the m-th sampling point within the t-th time window, where M represents the total number of sampling points within the time window, i represents the user number, t represents the time window number, and m represents the sampling point number.

[0021] Calculate the load change rate between adjacent sampling points: ; In the formula, This represents the rate of change of load for the i-th user at the m-th sampling point within the t-th time window relative to the previous sampling point. This represents the standardized load value of the i-th user at the m-th sampling point within the t-th time window. This represents the standardized load value of the i-th user at the (m-1)-th sampling point within the t-th time window, where i represents the user number, t represents the time window number, and m represents the sampling point number.

[0022] If q consecutive sampling points satisfy: ; In the formula, This represents the absolute value of the load change rate for user i at the m-th sampling point within the t-th time window. The threshold for determining a step change is represented by i, where i represents the user number, t represents the time window number, and m represents the sampling point number. Then, a step change event token is generated.

[0023] Furthermore, based on the set of power behavior event tokens, the evidentiary weight of each event token relative to the candidate tags is calculated, specifically as follows: ; In the formula, This represents the evidence weight of the r-th event token of the i-th user relative to the k-th candidate tag. Let represent the evidence matching score between the r-th event token of the i-th user and the k-th candidate tag, and exp represent an exponential function with the natural constant as the base. This represents the exponential summation of the evidence matching scores for all event tokens of the i-th user, where i represents the user index, r represents the event token index, and k represents the candidate tag index.

[0024] Furthermore, cross-source evidence vectors corresponding to each candidate label are generated based on the evidence weights, specifically as follows: ; In the formula, This represents the cross-source evidence vector for the i-th user corresponding to the k-th candidate label. This represents the evidence weight of the r-th event token of the i-th user relative to the k-th candidate tag. The vector representation of the r-th event token of the i-th user. This represents a weighted sum of all event tokens for the i-th user according to the evidence weights, where i represents the user sequence number, r represents the event token sequence number, and k represents the candidate tag sequence number.

[0025] Furthermore, a multi-label classification model is used to obtain the initial predicted probability of each candidate label, specifically: ; In the formula, This represents the initial prediction probability that the i-th user is predicted to have the k-th candidate label. This represents the Sigmoid activation function. Let T represent the weight vector corresponding to the k-th candidate label. This represents the multidimensional feature vector of the i-th user. This represents the cross-source evidence vector for the i-th user corresponding to the k-th candidate label. ; ] indicates that and Perform vector concatenation. This represents the bias term corresponding to the k-th candidate label, where i represents the user index and k represents the candidate label index.

[0026] Furthermore, based on the preset tag business constraint matrix, the initial prediction probability is constrained and calibrated, specifically as follows: ; In the formula, This represents the constraint correction amount for the k-th candidate label corresponding to the i-th user. This represents the business constraint coefficient between the k-th candidate label and the q-th candidate label. This represents the initial prediction probability that the i-th user is predicted to have the q-th candidate label. This represents the average predicted probability of the q-th candidate label in historical samples. This represents the summation of the constraint effects on all candidate labels, where i represents the user index, k represents the candidate label index, and q represents the constrained label index.

[0027] ; In the formula, This represents the calibrated prediction probability that the i-th user is predicted to have the k-th candidate label. Let represent the sigmoid activation function, and let logit represent the inverse function of the sigmoid function. This represents the initial prediction probability that the i-th user is predicted to have the k-th candidate label. Indicates the constraint calibration strength coefficient. This represents the constraint correction amount for the i-th user corresponding to the k-th candidate label, where i represents the user index and k represents the candidate label index.

[0028] Furthermore, based on the sufficiency of evidence, stability, and aging degree of each candidate label, an adaptive judgment threshold is calculated for each candidate label corresponding to each electricity user, specifically as follows: ; In the formula, This represents the adaptive decision threshold for the k-th candidate label corresponding to the i-th user. This represents the basic decision threshold for the k-th candidate label. This represents the threshold adjustment coefficient corresponding to the sufficiency of evidence. This represents the sufficiency of evidence for the i-th user corresponding to the k-th candidate label. This represents the threshold adjustment coefficient corresponding to label stability. This represents the label stability of the k-th candidate label for the i-th user. This represents the threshold adjustment coefficient corresponding to the degree of label aging. This represents the tag aging degree of the k-th candidate tag corresponding to the i-th user, where i represents the user index and k represents the candidate tag index.

[0029] Furthermore, the predicted label set is obtained in the following way: ; In the formula, Let i represent the set of predicted labels for the i-th user. This represents the k-th candidate label. This represents the calibrated prediction probability that the i-th user is predicted to have the k-th candidate label. This represents the adaptive judgment threshold for the k-th candidate label corresponding to the i-th user, where K represents the total number of candidate labels, i represents the user index, and k represents the candidate label index.

[0030] Furthermore, the method also includes: reacquiring the multi-source data of the corresponding user and regenerating the predicted label set based on a preset update cycle; calculating the label update error and evidence drift based on the predicted label set before and after the update and the cross-source evidence vector; updating the profile data of the corresponding user when both the label update error and the evidence drift exceed the corresponding threshold; and marking the profile data of the corresponding user as pending review when the label update error exceeds the corresponding threshold but the evidence drift does not exceed the corresponding threshold.

[0031] Furthermore, the label update error is obtained in the following way: ; In the formula, This represents the label update error for the i-th user. Let i represent the set of predicted labels after the i-th user is updated. Let $\mathbf{i}$ represent the set of predicted labels before the update for the $i$-th user, $\mathbf{i}$ denotes the symmetric difference operation between sets, $\mathbf{i}$ denotes the union operation between sets, and $\mathbf{i}$ denotes the set of predicted labels before the update for the $i$-th user. | indicates the number of tags that changed before and after the update, | | indicates the total number of related tags before and after the update, and i represents the user's serial number.

[0032] Furthermore, the evidence drift was obtained in the following way: ; In the formula, Let K represent the evidence drift for the i-th user, and K represent the total number of candidate labels. This indicates that the summation is performed on all candidate labels. This represents the updated cross-source evidence vector for the i-th user corresponding to the k-th candidate label. This represents the cross-source evidence vector before the update for the i-th user and the k-th candidate label, where · denotes the vector dot product operation. This represents the norm of the updated cross-source evidence vector. Let represent the norm of the cross-source evidence vector before the update, i represent the user index, and k represent the candidate label index.

[0033] Secondly, this application provides a power user profile construction system based on multi-label classification, applicable to any of the power user profile construction methods based on multi-label classification in the first aspect, including:

[0034] The multi-source data acquisition module is used to collect multi-source data from electricity users. The multi-source data includes basic profile data, electricity load data, transaction data, behavioral data, credit data, and auxiliary data.

[0035] The multi-source data processing module is used to preprocess and standardize multi-source data sequentially. The preprocessing includes outlier identification and missing value completion.

[0036] The event token construction module is used to construct a set of power behavior event tokens for each power user within a preset time window based on standardized multi-source data. Power behavior event tokens include load pattern event tokens, transaction fulfillment event tokens, service interaction event tokens, credit risk event tokens, and auxiliary scenario event tokens.

[0037] The cross-source evidence generation module is used to calculate the evidence weight of each event token relative to the candidate label based on the set of power behavior event tokens, and generate the cross-source evidence vector corresponding to each candidate label.

[0038] The multi-label prediction module is used to map standardized multi-source data and cross-source evidence vectors together into user multi-dimensional feature vectors, and to obtain the initial prediction probability of each candidate label using a multi-label classification model.

[0039] The constraint calibration module is used to perform constraint calibration on the initial prediction probability according to the preset tag service constraint matrix to obtain the calibrated prediction probability. The tag service constraint matrix is ​​used to characterize the mutual exclusion relationship, hierarchical relationship and accompaniment relationship between tags.

[0040] The adaptive threshold determination module is used to calculate the adaptive determination threshold for each candidate label for each power user based on the evidence sufficiency, label stability and label aging degree of each candidate label, and generate a predicted label set.

[0041] The profile output module is used to compare the calibrated prediction probability with the adaptive judgment threshold to obtain the predicted label set, and output the predicted label set as the profile data of the corresponding user.

[0042] Thirdly, this application provides an electronic device, including: at least one processor, at least one memory, and a data bus; wherein the processor and the memory communicate with each other through the data bus; the memory stores program instructions that can be executed by the processor, and the processor calls the program instructions to execute a method for constructing a power user profile based on multi-label classification as described in any of the first aspects.

[0043] Fourthly, this application provides a non-transitory computer-readable storage medium that stores computer instructions, which cause a computer to execute any one of the methods for constructing a power user profile based on multi-label classification in the first aspect.

[0044] Compared with the prior art, the present invention has at least the following beneficial effects:

[0045] First, this invention transforms multi-source power data into power behavior event tokens, enabling raw load, transaction, behavior, credit, and auxiliary data to be uniformly expressed as event units with business meaning, thus avoiding the problems of scattered feature meanings and unclear label basis caused by directly splicing multi-source features in the prior art.

[0046] Second, the present invention generates cross-source evidence vectors by using event tokens relative to the evidence weights of candidate tags, enabling different profile tags to aggregate related load evidence, transaction evidence, behavioral evidence, and credit evidence respectively, thereby improving the interpretability of the profile tag generation process.

[0047] Third, this invention calibrates the multi-label classification results through a label business constraint matrix, so that the predicted label set conforms to the mutual exclusion relationship, hierarchical relationship and accompaniment relationship in the power business, thereby reducing the probability of the multi-label classification model outputting conflicting labels.

[0048] Fourth, this invention uses an adaptive judgment threshold at the user-tag granularity to replace a uniform fixed threshold, enabling the tag output to be dynamically adjusted based on the sufficiency of evidence, tag stability, and tag aging, thereby improving the stability of user profiles during continuous updates.

[0049] Fifth, this invention uses a combination of tag update error and evidence drift to determine whether to update the user profile, which can distinguish between real changes in electricity consumption behavior and short-term abnormal disturbances, thereby improving the reliability of dynamic profile updates. Attached Figure Description

[0050] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and form part of this application, do not constitute a limitation thereof. In the drawings:

[0051] Figure 1 This is a flowchart of the construction method in an embodiment of the present invention;

[0052] Figure 2 This is a schematic diagram of the connection of the system in an embodiment of the present invention;

[0053] Figure 3 This is a schematic diagram of the connection of an electronic device in an embodiment of the present invention. Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention are described clearly and completely below. The described embodiments are only a part of the embodiments of the present invention, and not all of them. Other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are all within the scope of protection of the present invention.

[0055] Example 1: This example provides a method for constructing a power user profile based on multi-label classification, including the following specific steps.

[0056] S1 collects multi-source data from electricity users, including basic profile data, electricity load data, transaction data, behavioral data, credit data, and auxiliary data.

[0057] The basic data includes the user's industry, electricity capacity, electricity price type, registered address, user category, and contracted capacity; the electricity load data includes daily 96-point load curves, power factor, monthly electricity consumption, quarterly electricity consumption, peak-valley-normal-day electricity consumption, and maximum demand; the transaction data includes electricity price standards, payment records, arrears records, supplementary payment records, and historical bills; the behavioral data includes marketing inquiry frequency, inquiry content, online business processing type, online business processing frequency, fault repair records, and electricity consultation records; the credit data includes electricity theft records, electricity violation records, overdue payment records, and electricity compliance records; and the auxiliary data includes regional electricity consumption characteristics, seasonal electricity consumption fluctuation patterns, holiday information, meteorological information, and industry production cycle information.

[0058] Let the number of data source types be n, and the number of sub-items to be collected for the j-th data type be n. The actual number of valid sub-items collected was Then the data collection completion rate for the j-th type of data is: ; In the formula, This represents the data collection completion rate of the j-th type of data. This represents the number of valid sub-items actually collected for the j-th data category. This indicates the number of sub-items that need to be collected for the j-th data category, where j represents the data source type index. When When the completion rate falls below the preset threshold, the j-th type of data is re-collected or marked as low-reliability data.

[0059] S2 performs preprocessing and standardization on multi-source data in sequence. Preprocessing includes outlier identification and missing value completion.

[0060] For continuous data such as electricity load, electricity consumption, payment amount, and query frequency, let a certain data sample set be: ;

[0061] In the formula, X represents a sample set of a certain type of continuous data. Let represent the a-th sample value in sample set X, n represent the total number of samples in sample set X, and a represent the sample index.

[0062] Calculate the sample mean: ; In the formula, Let represent the sample mean of sample set X, and n represent the total number of samples in sample set X. This represents summing all sample values ​​in the sample set X. Let represent the a-th sample value in the sample set X, where a represents the sample index.

[0063] Calculate the sample standard deviation: ; In the formula, Let represent the sample standard deviation of sample set X, and let n represent the total number of samples in sample set X. This represents the summation of the squares of the differences between all sample values ​​in the sample set X and the sample mean. This represents the value of the a-th sample in the sample set X. Let represent the sample mean of the sample set X, and 'a' represent the sample index.

[0064] If the sample satisfies: ; In the formula, This represents the absolute value of the difference between the a-th sample value and the sample mean in the sample set X. This represents the value of the a-th sample in the sample set X. This represents the sample mean of the sample set X. Represents the outlier detection coefficient. Let denot X represent the sample standard deviation of the sample set X, and let a represent the sample index. Then, the sample is determined to be an outlier.

[0065] For missing values, multiple valid data points within the neighborhood of the missing value are selected, and imputation is performed using a weighted average. Specifically: ; In the formula, This represents the calculated result of missing value completion. This represents the weight of the b-th valid data point in the neighborhood. This represents the value of the b-th valid data point in the neighborhood. This represents the summation of the weighted values ​​of all valid data points within the neighborhood. This represents the summation of the weights of all valid data points in the neighborhood, where b represents the index of a valid data point in the neighborhood.

[0066] S3 constructs a set of power behavior event tokens for each power user within a preset time window based on standardized multi-source data. The power behavior event tokens include load pattern event tokens, transaction fulfillment event tokens, service interaction event tokens, credit risk event tokens, and auxiliary scenario event tokens.

[0067] Let the set of power behavior event tokens for the i-th user within the t-th time window be: ; In the formula, This represents the set of power behavior event tokens for the i-th user within the t-th time window. This represents the r-th electricity behavior event token generated by the i-th user within the t-th time window. This represents the total number of power behavior event tokens generated by the i-th user within the t-th time window, where i represents the user number, t represents the time window number, and r represents the event token number.

[0068] Among them, the load pattern event token is generated from electricity load data and is used to characterize step changes, peak load, valley load, periodic start-stop, and power factor anomalies; the transaction fulfillment event token is generated from transaction data and is used to characterize overdue payments, arrears payment, payment fluctuations, and electricity price changes; the service interaction event token is generated from behavioral data and is used to characterize high-frequency queries, repeated processing, fault reporting, electricity consultation, and channel preferences; the credit risk event token is generated from credit data and is used to characterize electricity theft, default on electricity use, long-term overdue payments, and abnormal electricity compliance; and the auxiliary scenario event token is generated from auxiliary data and is used to characterize regions, seasons, weather, holidays, and industry production cycles.

[0069] S4. Based on the set of power behavior event tokens, calculate the evidence weight of each event token relative to the candidate label, and generate the cross-source evidence vector corresponding to each candidate label based on the evidence weight.

[0070] Vectorize the r-th event token of the i-th user to obtain the event token vector: ; In the formula, The vector representation of the r-th event token of the i-th user. This represents the event token vectorization function. This represents the r-th event token of the i-th user, where i represents the user sequence number and r represents the event token sequence number.

[0071] Calculate the evidence matching score of the r-th event token relative to the k-th candidate tag: ; In the formula, This represents the evidence matching score between the r-th event token of the i-th user and the k-th candidate tag. Let T represent the label vector of the k-th candidate label, and let T denote the vector transpose. The vector representation of the r-th event token of the i-th user. This represents the business prior association value between the r-th event token of the i-th user and the k-th candidate tag, where i represents the user number, r represents the event token number, and k represents the candidate tag number.

[0072] Calculate the evidence weight based on the evidence matching score: ; In the formula, Let represent the evidence weight of the r-th event token of the i-th user relative to the k-th candidate tag, and exp represent an exponential function with the natural constant as the base. This represents the evidence matching score between the r-th event token of the i-th user and the k-th candidate tag. This represents the exponential summation of the evidence matching scores for all event tokens of the i-th user, where i represents the user index, r represents the event token index, and k represents the candidate tag index.

[0073] Generate cross-source evidence vectors based on evidence weights: ; In the formula, This represents the cross-source evidence vector for the i-th user corresponding to the k-th candidate label. This represents the evidence weight of the r-th event token of the i-th user relative to the k-th candidate tag. The vector representation of the r-th event token of the i-th user. This represents a weighted sum of all event tokens for the i-th user according to the evidence weights, where i represents the user sequence number, r represents the event token sequence number, and k represents the candidate tag sequence number.

[0074] S5 maps the standardized multi-source data and cross-source evidence vectors together into a user multi-dimensional feature vector, and uses a multi-label classification model to obtain the initial prediction probability of each candidate label.

[0075] Let the multidimensional feature vector of the i-th user be... The cross-source evidence vector for the i-th user corresponding to the k-th candidate label is: Then the initial prediction probability of the k-th candidate label is: ; In the formula, This represents the initial prediction probability that the i-th user is predicted to have the k-th candidate label. This represents the Sigmoid activation function. Let T represent the weight vector corresponding to the k-th candidate label. This represents the multidimensional feature vector of the i-th user. This represents the cross-source evidence vector for the i-th user corresponding to the k-th candidate label. ; ] indicates that and Perform vector concatenation. This represents the bias term corresponding to the k-th candidate label, where i represents the user index and k represents the candidate label index.

[0076] S6. Based on the preset tag business constraint matrix, the initial prediction probability is constrained and calibrated to obtain the calibrated prediction probability. The tag business constraint matrix is ​​used to characterize the mutual exclusion relationship, hierarchical relationship and accompaniment relationship between tags.

[0077] The tag business constraint matrix is ​​represented as follows: ; In the formula, C represents the tag business constraint matrix. Let K represent the business constraint coefficient between the k-th candidate label and the q-th candidate label, K represent the total number of candidate labels, K×K represent the dimension of matrix C as K rows and K columns, k represent the candidate label number, and q represent the constrained label number.

[0078] When the k-th candidate label and the q-th candidate label are mutually exclusive Take a negative value; when the k-th candidate label and the q-th candidate label have a hierarchical relationship or an accompanying relationship. Take a positive value; when there is no explicit business constraint relationship between the k-th candidate label and the q-th candidate label, Take zero.

[0079] Calculate the constraint correction amount based on the tag business constraint matrix: ; In the formula, This represents the constraint correction amount for the k-th candidate label corresponding to the i-th user. This represents the business constraint coefficient between the k-th candidate label and the q-th candidate label. This represents the initial prediction probability that the i-th user is predicted to have the q-th candidate label. _q represents the average predicted probability of the q-th candidate label in historical samples. This represents the summation of the constraint effects on all candidate labels, where i represents the user index, k represents the candidate label index, and q represents the constrained label index.

[0080] Obtain the calibrated prediction probability: ; In the formula, This represents the calibrated prediction probability that the i-th user is predicted to have the k-th candidate label. Let represent the sigmoid activation function, and let logit represent the inverse function of the sigmoid function. This represents the initial prediction probability that the i-th user is predicted to have the k-th candidate label. Indicates the constraint calibration strength coefficient. This represents the constraint correction amount for the i-th user corresponding to the k-th candidate label, where i represents the user index and k represents the candidate label index.

[0081] S7. Based on the sufficiency of evidence, stability of the label, and aging of the label, calculate the adaptive judgment threshold for each candidate label for each power user.

[0082] First, calculate the sufficiency of evidence: ; In the formula, This represents the sufficiency of evidence for the i-th user corresponding to the k-th candidate label. This represents the evidence weight of the r-th event token of the i-th user relative to the k-th candidate tag. This represents the data reliability of the r-th event token of the i-th user. This represents a weighted sum of the data reliability of all event tokens according to the evidence weights, where i represents the user sequence number, r represents the event token sequence number, and k represents the candidate tag sequence number.

[0083] Secondly, calculate the label stability: ; In the formula, Let H represent the label stability of the i-th user corresponding to the k-th candidate label, and let H represent the number of historical review time windows. This indicates summing over historical review time windows. This represents the predicted probability of the i-th user corresponding to the k-th candidate label within the th time window. This represents the predicted probability of the i-th user corresponding to the k-th candidate label within the (th-1)-th time window. This represents the absolute value of the difference in predicted probabilities between adjacent historical time windows, where i represents the user number, k represents the candidate label number, t represents the current time window number, and h represents the historical review time window number.

[0084] Next, calculate the label aging level: ; In the formula, Let represent the tag aging degree of the k-th candidate tag corresponding to the i-th user, and exp represent an exponential function with the natural constant as the base. Indicates the current time. This represents the last update time of the candidate tag corresponding to the i-th user. Let i represent the aging time constant corresponding to the k-th candidate tag, where i represents the user number and k represents the candidate tag number.

[0085] An adaptive judgment threshold is calculated based on the sufficiency of evidence, label stability, and label aging degree: ; In the formula, This represents the adaptive decision threshold for the k-th candidate label corresponding to the i-th user. This represents the basic decision threshold for the k-th candidate label. This represents the threshold adjustment coefficient corresponding to the sufficiency of evidence. This represents the sufficiency of evidence for the i-th user corresponding to the k-th candidate label. This represents the threshold adjustment coefficient corresponding to label stability. This represents the label stability of the k-th candidate label for the i-th user. This represents the threshold adjustment coefficient corresponding to the degree of label aging. This represents the tag aging degree of the k-th candidate tag corresponding to the i-th user, where i represents the user index and k represents the candidate tag index.

[0086] Based on the calibrated prediction probability and the adaptive decision threshold, the predicted label set is obtained: ; In the formula, Let i represent the set of predicted labels for the i-th user. This represents the k-th candidate label. This represents the calibrated prediction probability that the i-th user is predicted to have the k-th candidate label. This represents the adaptive judgment threshold for the k-th candidate label corresponding to the i-th user, where K represents the total number of candidate labels, i represents the user index, and k represents the candidate label index.

[0087] S8 compares the calibrated prediction probability with the adaptive decision threshold to obtain the predicted label set, and outputs the predicted label set as the profile data of the corresponding user.

[0088] The user profile data includes a unique user identifier, a set of predicted tags, the calibrated prediction probability of each tag, the adaptive judgment threshold of each tag, the main evidence token corresponding to each tag, the tag update time, and the identifier to be reviewed.

[0089] S9 dynamically updates the profile based on label update error and evidence drift.

[0090] When the preset update cycle arrives, the multi-source data for the corresponding user is reacquired, and steps S2 to S8 are re-executed to obtain the updated predicted label set and cross-source evidence vector. The label update error is calculated based on the predicted label sets before and after the update. ; In the formula, This represents the label update error for the i-th user. Let i represent the set of predicted labels after the i-th user is updated. Let $\mathbf{i}$ represent the set of predicted labels before the update for the $i$-th user, $\mathbf{i}$ denotes the symmetric difference operation between sets, $\mathbf{i}$ denotes the union operation between sets, and $\mathbf{i}$ denotes the set of predicted labels before the update for the $i$-th user. | indicates the number of tags that changed before and after the update, | | indicates the total number of related tags before and after the update, and i represents the user's serial number.

[0091] Calculate the evidence drift based on the cross-source evidence vectors before and after the update: ; In the formula, Let K represent the evidence drift for the i-th user, and K represent the total number of candidate labels. This indicates that the summation is performed on all candidate labels. This represents the updated cross-source evidence vector for the i-th user corresponding to the k-th candidate label. This represents the cross-source evidence vector before the update for the i-th user corresponding to the k-th candidate label, where · denotes the vector dot product operation. This represents the norm of the updated cross-source evidence vector. Let represent the norm of the cross-source evidence vector before the update, i represent the user index, and k represent the candidate label index.

[0092] When both the label update error and the evidence drift exceed the corresponding threshold, the profile data of the corresponding user is updated; when the label update error exceeds the corresponding threshold but the evidence drift does not exceed the corresponding threshold, the profile data of the corresponding user is marked as pending review; when neither the label update error nor the evidence drift exceeds the corresponding threshold, the original profile data remains unchanged.

[0093] Example 2: This application provides a power user profile construction system based on multi-label classification, applied to the power user profile construction method based on multi-label classification in Example 1, including:

[0094] The multi-source data acquisition module is used to collect multi-source data from electricity users. The multi-source data includes basic profile data, electricity load data, transaction data, behavioral data, credit data, and auxiliary data.

[0095] The multi-source data processing module is used to preprocess and standardize multi-source data sequentially. The preprocessing includes outlier identification and missing value completion.

[0096] The event token construction module is used to construct a set of power behavior event tokens for each power user within a preset time window based on standardized multi-source data. Power behavior event tokens include load pattern event tokens, transaction fulfillment event tokens, service interaction event tokens, credit risk event tokens, and auxiliary scenario event tokens.

[0097] The cross-source evidence generation module is used to calculate the evidence weight of each event token relative to the candidate label based on the set of power behavior event tokens, and generate the cross-source evidence vector corresponding to each candidate label.

[0098] The multi-label prediction module is used to map standardized multi-source data and cross-source evidence vectors together into user multi-dimensional feature vectors, and to obtain the initial prediction probability of each candidate label using a multi-label classification model.

[0099] The constraint calibration module is used to perform constraint calibration on the initial prediction probability according to the preset tag service constraint matrix to obtain the calibrated prediction probability. The tag service constraint matrix is ​​used to characterize the mutual exclusion relationship, hierarchical relationship and accompaniment relationship between tags.

[0100] The adaptive threshold determination module is used to calculate the adaptive determination threshold for each candidate label for each power user based on the evidence sufficiency, label stability and label aging degree of each candidate label, and generate a predicted label set.

[0101] The profile output module is used to compare the calibrated prediction probability with the adaptive judgment threshold to obtain the predicted label set, and output the predicted label set as the profile data of the corresponding user.

[0102] Example 3: This application provides an electronic device, including at least one processor, at least one memory, and a data bus;

[0103] In this embodiment, the processor and the memory communicate with each other through a data bus; the memory stores program instructions that can be executed by the processor, and the processor calls the program instructions to execute the method for constructing a power user profile based on multi-label classification in Embodiment 1.

[0104] Example 4: This application provides a non-transitory computer-readable storage medium that stores computer instructions, which cause the computer to execute a method for constructing a power user profile based on multi-label classification as described in Example 1.

[0105] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0106] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0107] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0108] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0109] Those skilled in the art will understand that all or part of the steps in the above facts and methods can be implemented by a program instructing related hardware. The program or the program described therein can be stored in a computer-readable storage medium. When the program is executed, it includes the following steps: at this time, the corresponding method steps are introduced. The storage medium can be ROM / RAM, magnetic disk, optical disk, etc.

[0110] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for constructing power user profiles based on multi-label classification, characterized in that, The specific steps include the following: Collect multi-source data from electricity users, including basic profile data, electricity load data, transaction data, behavioral data, credit data, and auxiliary data; The multi-source data are preprocessed and standardized sequentially. The preprocessing includes outlier identification and missing value completion. Based on the standardized multi-source data, a set of power behavior event tokens for each power user within a preset time window is constructed. The power behavior event tokens include load pattern event tokens, transaction fulfillment event tokens, service interaction event tokens, credit risk event tokens, and auxiliary scenario event tokens. Based on the set of power behavior event tokens, calculate the evidence weight of each event token relative to the candidate tag, and generate a cross-source evidence vector corresponding to each candidate tag based on the evidence weight; The standardized multi-source data and the cross-source evidence vector are jointly mapped into a user multi-dimensional feature vector, and a multi-label classification model is used to obtain the initial predicted probability of each candidate label. Based on the preset tag service constraint matrix, the initial prediction probability is constrained and calibrated to obtain the calibrated prediction probability. The tag service constraint matrix is ​​used to characterize the mutual exclusion relationship, hierarchical relationship and accompaniment relationship between tags. Based on the sufficiency of evidence, stability of the label, and aging of the label, the adaptive judgment threshold for each candidate label for each power user is calculated. The calibrated predicted probability is compared with the adaptive decision threshold to obtain a predicted label set, and the predicted label set is used as the profile data of the corresponding user and output.

2. The method for constructing a power user profile based on multi-label classification according to claim 1, characterized in that, The step of constructing a set of power behavior event tokens for each power user within a preset time window based on the standardized multi-source data specifically includes: Extract load change rate, peak and valley load, duration and periodic repetition features from electricity load data to generate load pattern event tokens; Extract payment time, arrears duration, frequency of supplementary payments, and electricity price execution information from transaction data to generate transaction fulfillment event tokens; Extract query frequency, business processing type, fault reporting and electricity consultation content from behavioral data, and generate service interaction event tokens; Extract records of defaulted electricity use, electricity theft, overdue payments, and compliant electricity use from credit data to generate credit risk event tokens; Extract regional, seasonal, meteorological, holiday, and industry production cycle data from auxiliary data to generate auxiliary scenario event tokens.

3. The method for constructing a power user profile based on multi-label classification according to claim 2, characterized in that, The load pattern event token is obtained as follows: Let the normalized load sequence of the i-th user in the t-th time window be: ; In the formula, This represents the standardized load sequence of the i-th user within the t-th time window. Let M represent the standardized load value of the i-th user at the m-th sampling point within the t-th time window, where M represents the total number of sampling points within the time window, i represents the user number, t represents the time window number, and m represents the sampling point number. Calculate the load change rate between adjacent sampling points: ; In the formula, This represents the rate of change of load for the i-th user at the m-th sampling point within the t-th time window relative to the previous sampling point. This represents the standardized load value of the i-th user at the m-th sampling point within the t-th time window. This represents the standardized load value of the i-th user at the (m-1)-th sampling point within the t-th time window, where i represents the user number, t represents the time window number, and m represents the sampling point number. If q consecutive sampling points satisfy: ; In the formula, This represents the absolute value of the load change rate for user i at the m-th sampling point within the t-th time window. The threshold for determining a step change is represented by i, where i represents the user number, t represents the time window number, and m represents the sampling point number. Then, a step change event token is generated.

4. The method for constructing a power user profile based on multi-label classification according to claim 1, characterized in that, The step of calculating the evidence weight of each event token relative to the candidate label based on the set of power behavior event tokens is specifically as follows: ； In the formula, This represents the evidence weight of the r-th event token of the i-th user relative to the k-th candidate tag. Let represent the evidence matching score between the r-th event token of the i-th user and the k-th candidate tag, and exp represent an exponential function with the natural constant as the base. This represents the exponential summation of the evidence matching scores for all event tokens of the i-th user, where i represents the user index, r represents the event token index, and k represents the candidate tag index.

5. The method for constructing a power user profile based on multi-label classification according to claim 1, characterized in that, Based on the aforementioned evidence weights, cross-source evidence vectors corresponding to each candidate label are generated, specifically as follows: ； In the formula, This represents the cross-source evidence vector for the i-th user corresponding to the k-th candidate label. This represents the evidence weight of the r-th event token of the i-th user relative to the k-th candidate tag. The vector representation of the r-th event token of the i-th user. This represents a weighted sum of all event tokens for the i-th user according to the evidence weights, where i represents the user sequence number, r represents the event token sequence number, and k represents the candidate tag sequence number.

6. The method for constructing a power user profile based on multi-label classification according to claim 1, characterized in that, The initial predicted probability of each candidate label is obtained using a multi-label classification model, specifically as follows: ； In the formula, This represents the initial prediction probability that the i-th user is predicted to have the k-th candidate label. This represents the Sigmoid activation function. Let T represent the weight vector corresponding to the k-th candidate label. This represents the multidimensional feature vector of the i-th user. This represents the cross-source evidence vector for the i-th user corresponding to the k-th candidate label. ; ] indicates that and Perform vector concatenation. This represents the bias term corresponding to the k-th candidate label, where i represents the user index and k represents the candidate label index.

7. The method for constructing a power user profile based on multi-label classification according to claim 1, characterized in that, The step of constraining and calibrating the initial prediction probability according to the preset label service constraint matrix specifically involves: ； In the formula, This represents the constraint correction amount for the k-th candidate label corresponding to the i-th user. This represents the business constraint coefficient between the k-th candidate label and the q-th candidate label. This represents the initial prediction probability that the i-th user is predicted to have the q-th candidate label. This represents the average predicted probability of the q-th candidate label in historical samples. This represents the summation of the constraint effects on all candidate labels, where i represents the user index, k represents the candidate label index, and q represents the constrained label index. ； In the formula, This represents the calibrated prediction probability that the i-th user is predicted to have the k-th candidate label. Let represent the sigmoid activation function, and let logit represent the inverse function of the sigmoid function. This represents the initial prediction probability that the i-th user is predicted to have the k-th candidate label. Indicates the constraint calibration strength coefficient. This represents the constraint correction amount for the i-th user corresponding to the k-th candidate label, where i represents the user index and k represents the candidate label index.

8. The method for constructing a power user profile based on multi-label classification according to claim 1, characterized in that, The adaptive judgment threshold for each candidate tag for each electricity user is calculated based on the sufficiency of evidence, tag stability, and tag aging degree of each candidate tag, specifically as follows: ； In the formula, This represents the adaptive decision threshold for the k-th candidate label corresponding to the i-th user. This represents the basic decision threshold for the k-th candidate label. This represents the threshold adjustment coefficient corresponding to the sufficiency of evidence. This represents the sufficiency of evidence for the i-th user corresponding to the k-th candidate label. This represents the threshold adjustment coefficient corresponding to label stability. This represents the label stability of the k-th candidate label for the i-th user. This represents the threshold adjustment coefficient corresponding to the degree of label aging. This indicates the tag aging degree of the k-th candidate tag corresponding to the i-th user, where i represents the user index and k represents the candidate tag index; The predicted label set is obtained through the following method: ； In the formula, Let i represent the set of predicted labels for the i-th user. This represents the k-th candidate label. This represents the calibrated prediction probability that the i-th user is predicted to have the k-th candidate label. This represents the adaptive judgment threshold for the k-th candidate label corresponding to the i-th user, where K represents the total number of candidate labels, i represents the user index, and k represents the candidate label index.

9. The method for constructing a power user profile based on multi-label classification according to claim 1, characterized in that, The method further includes: Based on a preset update cycle, reacquire multi-source data for the corresponding user and regenerate the predicted label set; Based on the predicted label sets before and after the update and the cross-source evidence vectors, calculate the label update error and evidence drift. When both the label update error and the evidence drift exceed the corresponding threshold, update the profile data of the corresponding user. When the label update error exceeds the corresponding threshold but the evidence drift does not exceed the corresponding threshold, the profile data of the corresponding user will be marked as pending review. The label update error is obtained in the following way: ； In the formula, This represents the label update error for the i-th user. Let i represent the set of predicted labels after the i-th user is updated. Let $\mathbf{i}$ represent the set of predicted labels before the update for the $i$-th user, $\mathbf{i}$ denotes the symmetric difference operation between sets, $\mathbf{i}$ denotes the union operation between sets, and $\mathbf{i}$ denotes the set of predicted labels before the update for the $i$-th user. | indicates the number of tags that changed before and after the update, | | indicates the total number of all relevant tags before and after the update, and i represents the user's serial number; The evidence drift is obtained in the following way: ； In the formula, Let K represent the evidence drift for the i-th user, and K represent the total number of candidate labels. This indicates that the summation is performed on all candidate labels. This represents the updated cross-source evidence vector for the i-th user corresponding to the k-th candidate label. This represents the cross-source evidence vector before the update for the i-th user and the k-th candidate label, where · denotes the vector dot product operation. This represents the norm of the updated cross-source evidence vector. Let represent the norm of the cross-source evidence vector before the update, i represent the user index, and k represent the candidate label index.

10. A power user profile construction system based on multi-label classification, characterized in that, include: The multi-source data acquisition module is used to collect multi-source data from electricity users, including basic profile data, electricity load data, transaction data, behavioral data, credit data, and auxiliary data. A multi-source data processing module is used to perform preprocessing and standardization processing on the multi-source data in sequence. The preprocessing includes outlier identification processing and missing value completion processing. The event token construction module is used to construct a set of power behavior event tokens for each power user within a preset time window based on the standardized multi-source data. The power behavior event tokens include load mode event tokens, transaction fulfillment event tokens, service interaction event tokens, credit risk event tokens, and auxiliary scenario event tokens. The cross-source evidence generation module is used to calculate the evidence weight of each event token relative to the candidate label based on the set of power behavior event tokens, and generate the cross-source evidence vector corresponding to each candidate label. The multi-label prediction module is used to map the standardized multi-source data and the cross-source evidence vector together into a user multi-dimensional feature vector, and to use a multi-label classification model to obtain the initial prediction probability of each candidate label. The constraint calibration module is used to perform constraint calibration on the initial prediction probability according to the preset tag service constraint matrix to obtain the calibrated prediction probability. The tag service constraint matrix is ​​used to characterize the mutual exclusion relationship, hierarchical relationship and accompaniment relationship between tags. The adaptive threshold determination module is used to calculate the adaptive determination threshold for each candidate label for each power user based on the evidence sufficiency, label stability and label aging degree of each candidate label, and generate a predicted label set. The profile output module is used to compare the calibrated prediction probability with the adaptive judgment threshold to obtain a set of predicted labels, and output the set of predicted labels as the profile data of the corresponding user.

Images