Method, device and storage medium for predicting business objects

Abstract

The application provides a prediction method and device of a business object and a storage medium, wherein the method comprises: obtaining a feature corresponding to a business object to be analyzed; matching the feature with a rule in a rule learning model, and determining whether the business object to be analyzed meets a requirement and / or a reason for meeting the requirement according to a matching result; and the rule learning model is a rule learning model generated based on column generation technology and a frequent pattern tree. The application can predict whether the business object to be analyzed meets the requirement through the rule learning model, has strong interpretability, can effectively improve user experience, establishes the trust of users in the rule learning model, in addition, the rule learning model generated based on the column generation technology and the frequent pattern tree can also effectively improve the rule learning effect, meet the requirements of different scenes and different dimension data, and improve the overall efficiency of learning and prediction.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer, and particularly relates to a business object prediction method, device and storage medium. BACKGROUND

[0002] With the increasing data scale of various industries, the difficulty of manually analyzing data also increases. Many complex tasks and systems have been automated and intelligentized by adopting and expanding machine learning systems.

[0003] When predicting a business object, a user may not only need a machine learning system to give a prediction result, but also want to know the reason for obtaining the prediction result. However, many machine learning systems are essentially still black box models, for example, a neural network model, which has a complex structure, low processing efficiency and lacks interpretability, resulting in poor user experience. SUMMARY

[0004] The main purpose of the embodiments of the present application is to provide a business object prediction method, device and storage medium to improve prediction efficiency and interpretability, and thus improve user experience.

[0005] In a first aspect, the embodiments of the present application provide a business object prediction method, comprising:

[0006] obtaining a feature corresponding to a business object to be analyzed;

[0007] matching the feature with a rule in a rule learning model, and determining whether the business object to be analyzed meets a requirement and / or a reason for meeting the requirement according to a matching result;

[0008] The rule learning model is a rule learning model generated based on column generation technology and a frequent pattern tree.

[0009] In a second aspect, the embodiments of the present application provide a power load prediction method, comprising:

[0010] obtaining an environmental feature to be analyzed;

[0011] predicting whether a power load corresponding to the environmental feature meets a requirement and / or outputting a reason for determining a power load prediction result according to the environmental feature and a rule learning model;

[0012] The rule learning model is a rule learning model generated based on column generation technology and / or a frequent pattern tree.

[0013] In a third aspect, the embodiments of the present application provide an electronic device, comprising:

[0014] at least one processor; and

[0015] a memory connected with the at least one processor in communication;

[0016] The memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the electronic device to perform the method of the first aspect or the second aspect.

[0017] In a fourth aspect, an embodiment of the present application provides a computer readable storage medium, and the computer readable storage medium stores computer execution instructions, and when a processor executes the computer execution instructions, the method of the first aspect or the second aspect is implemented.

[0018] The business object prediction method, device and storage medium provided by the present application can predict whether the business object to be analyzed meets the requirements through the rule learning model, have strong interpretability, can effectively improve user experience, establish user trust in the rule learning model, in addition, the rule learning model generated based on the column generation technology and the frequent pattern tree can also effectively improve the rule learning effect, meet the needs of different scenes and different dimension data, and improve the overall efficiency of learning and prediction. BRIEF DESCRIPTION OF DRAWINGS

[0019] The accompanying drawings, which are incorporated in and constitute a part of the specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0020] Figure 1 An application scenario schematic diagram provided by an embodiment of the present application;

[0021] Figure 2 A principle schematic diagram of determining an approval result based on a neural network model;

[0022] Figure 3 A principle schematic diagram of determining an approval result based on a rule learning model provided by an embodiment of the present application;

[0023] Figure 4 A reason explanation schematic diagram of an approval failure provided by an embodiment of the present application;

[0024] Figure 5 A flowchart of a business object prediction method provided by an embodiment of the present application;

[0025] Figure 6 A schematic diagram of a frequent pattern tree provided by an embodiment of the present application;

[0026] Figure 7 A schematic diagram of a frequent pattern tree constructed based on a non-empty rule head provided by an embodiment of the present application;

[0027] Figure 8 A set schematic diagram in a recursive process provided for an embodiment of the present application;

[0028] Figure 9 A node linked list schematic diagram provided for an embodiment of the present application;

[0029] Figure 10 A rule learning method flow schematic diagram provided for an embodiment of the present application;

[0030] Figure 11 A power load prediction method flow schematic diagram provided for an embodiment of the present application;

[0031] Figure 12 A power load prediction result schematic diagram provided for an embodiment of the present application;

[0032] Figure 13 A business index abnormality prediction result schematic diagram provided for an embodiment of the present application;

[0033] Figure 14 A recommendation information schematic diagram provided for an embodiment of the present application;

[0034] Figure 15 A schematic diagram of the structure of an electronic device provided for an embodiment of the present application.

[0035] Through the above-mentioned drawings, the specific embodiments of the present application have been shown, and will be described in more detail hereinafter. These drawings and textual descriptions are not intended to limit the scope of the concept of the present application by any means, but to illustrate the concept of the present application to those skilled in the art by referring to specific embodiments. DETAILED DESCRIPTION

[0036] The exemplary embodiments will be described in detail herein with reference to the attached drawings. The following description is made with reference to the accompanying drawings in which like reference numerals represent like elements or similar elements, unless otherwise indicated. The following exemplary embodiments described in the following detailed description are not meant to be limiting of all embodiments that would be consistent with the present application. Instead, they are merely examples of apparatus and methods consistent with some aspects of the present application as detailed in the appended claims.

[0037] First, the terms involved in the present application are explained:

[0038] Rule learning: a machine learning model for learning a set of IF-THEN rules to describe and explain a specified group in the data set.

[0039] BitMap (bitmap): a mapping of bit conversion of an integer.

[0040] FP-Tree (Frequent Pattern Tree): A specific data structure that supports frequent itemset mining.

[0041] Branch and bound: A technique that partitions the search space into small subspaces and prunes the subspaces using the upper and lower bounds of the solutions.

[0042] Column generation: An efficient algorithm for solving large-scale linear optimization problems.

[0043] Pricing problem: An integer programming problem that solves the optimal generating column in the column generation algorithm.

[0044] Check number: The objective function of the pricing problem.

[0045] With the continuous development of big data processing technology, the application of machine learning is also more and more widely. Traditional machine learning can train a model based on existing data and use the trained model for prediction.

[0046] Figure 1 An application scenario diagram is provided for the embodiments of the present application. As shown in Figure 1 The user's characteristics, such as age, income, flow, asset situation, default times, etc., can be input into the model trained by the electronic device, and then the approval result is determined, that is, whether the user's request is passed or not passed.

[0047] Conventional machine learning models, such as neural network models, are often black box models and lack explainability. Although the data can be input into the model to give the result of passing or not passing the approval, the reasons for passing or not passing the approval cannot be given.

[0048] Figure 2 A principle diagram for determining an approval result based on a neural network model is provided. As shown in Figure 2 The neural network model can include multiple layers, such as convolutional layers, pooling layers, fully connected layers, etc. The input data is calculated layer by layer in the neural network model and is passed back, and after the final result is determined, it is impossible to distinguish which reasons in the input data led to the final result, so it performs poorly in explainability.

[0049] Figure 3 A principle diagram for determining an approval result based on a rule learning model is provided for the embodiments of the present application. As shown in Figure 3 The rule learning model contains multiple rules corresponding to not passing the approval, and each rule corresponds to at least one feature such as default times, flow, etc. If the characteristics of the target user match a rule in the rule learning model, it can be determined that the corresponding result is not passing the approval, and the corresponding reason can be fed back to the approval personnel or the user.

[0050] Figure 4 A reason explanation diagram for disapproval of an embodiment of the present application is provided. As shown in Figure 4 the number of defaults, assets, water, and income of user A match a rule in the rule learning model, the approval result corresponding to user A is disapproval, and the corresponding reason is fed back to the approval personnel or user.

[0051] In addition to the above-mentioned financial field, the rule learning model has strong application demand in other fields. For example, in the load prediction scene of the power field, the prediction personnel first need to understand the cause of generating the prediction value, so as to report the prediction result to the dispatching department, thereby providing a reference for subsequent power dispatching, power generation plan, and dynamic pricing of electricity price; in the enterprise operation field, the algorithm can give the reason for the change of the key operation index, which can greatly improve the work efficiency of the analysis and decision personnel, so as to make more scientific and reasonable decisions; in the personalized recommendation scene, the user wants to understand the reason for the recommendation result, so as to make the user have more trust and higher satisfaction with the recommendation system; and the like.

[0052] Therefore, the rule learning model, with its excellent simplicity and interpretability, can realize the auxiliary decision-making of key fields and automatically explore the business value in the data.

[0053] However, the rule learning still faces great challenges at present. First, it is very difficult to simultaneously optimize the prediction ability, rule complexity, and diversity of the model. Most rule learning algorithms are usually divided into two stages, namely the rule generation and rule selection stages. Due to the lack of consideration of the global optimization target, the rule generation stage is usually weakly connected with the final rule set effect, resulting in inaccurate or high overlap rules. It should be noted that reducing the overlap of rules can encourage each rule to cover and describe an independent subspace in the feature space, thereby improving the interpretability of the result. This is particularly important for the interpretability of multi-classification tasks, because if the rules from different categories have high overlap, the class group to which the samples in the overlapping part belong will become ambiguous due to the conflict, resulting in poor prediction performance of the rule learning model. Moreover, there are very limited models that can effectively handle large-scale data. With the increase of the sample size, especially the number of features, the result is usually complex, and the interpretability requirement of the model becomes particularly important. Unfortunately, the rule space that increases exponentially with the feature dimension poses a severe challenge to the development of fast algorithms.

[0054] Therefore, the embodiment of the present application provides a prediction method of a business object, which can generate a rule learning model based on column generation technology and a frequent pattern tree, after obtaining the features corresponding to the business object to be analyzed, the features can be matched with the rules in the rule learning model, and according to the matching result, it can be determined whether the business object meets the requirements, and if it meets the requirements, the reason for meeting the requirements can be further determined. Based on column generation technology, the data processing amount can be effectively reduced, and the efficiency can be improved. Based on the frequent pattern tree, the frequency of occurrence of various possible rules can be intuitively represented, and then efficient representation of data can be realized, and samples meeting certain rules can be quickly accessed, so that the rules can be quickly searched, and the overall efficiency of rule learning and prediction can be further improved. The generated rule learning model can be applied to various business fields, and the prediction results of various business fields can be provided with reason explanation, and the user experience can be improved.

[0055] Some embodiments of the present application will be described in detail below with reference to the accompanying drawings. The embodiments described below and the features in the embodiments can be combined with each other without conflict. In addition, the sequence of steps in each method embodiment described below is only an example, not a strict limitation.

[0056] Figure 5 A flowchart of a prediction method of a business object provided by an embodiment of the present application is shown. The execution subject of the method in the embodiment can be any device with data processing function. As shown in Figure 5 The method can include the following steps.

[0057] Step 501: Obtain the features corresponding to the business object to be analyzed.

[0058] Optionally, the business object can be any possible business object, including but not limited to: user, time, place, event, etc. The features corresponding to the business object can be any feature that can represent the information of the business object, such as user portrait features, changes in environmental information within a certain time, images corresponding to places, and various development stages of events, etc.

[0059] Step 502: Match the features with the rules in the rule learning model, and determine whether the business object to be analyzed meets the requirements and / or the reason for meeting the requirements according to the matching result.

[0060] The rule learning model is a rule learning model generated based on column generation technology and a frequent pattern tree.

[0061] Optionally, the rule learning model can be generated by the rule learning method described in any embodiment of the present application. The characteristics of the business object are matched with the rules in the rule learning model. If matched with at least one rule, it is considered that the business object meets the requirements, otherwise it does not meet the requirements.

[0062] Optionally, it can also be shown whether the business object meets the requirements and / or the reasons for meeting the requirements. Specifically, it can only show whether the requirements are met, only show the reasons, or both. The reasons are determined by the rules matched with the characteristics corresponding to the business object to be analyzed.

[0063] The prediction method provided in the embodiment can be applied to any field. Exemplarily, the rule learning model can be used to predict at least one of the following: whether the power load of the time node to be analyzed meets the requirements, whether the business index corresponding to the time node to be analyzed is abnormal, whether the user to be analyzed will accept the recommended information, whether the user to be analyzed is approved, and whether the user to be analyzed has maintenance needs.

[0064] Correspondingly, for different fields, positive and negative samples that meet the requirements and do not meet the requirements can be found during rule learning, and rule learning is performed based on the positive and negative samples. The learned rules can be used for prediction in the field.

[0065] In summary, the prediction method provided in the embodiment predicts whether the business object to be analyzed meets the requirements through the rule learning model, has strong interpretability, can effectively improve user experience, establish user trust in the rule learning model, and further improve the effect of rule learning based on the column generation technology and the rule learning model generated by the frequent pattern tree, meet the needs of different scenarios and different dimensional data, and improve the overall efficiency of learning and prediction.

[0066] One or more embodiments of the present application can also include a rule learning method, wherein the rule learning method can include: obtaining a training sample set, the training sample set including a positive sample group and a negative sample group, wherein the positive sample includes characteristics corresponding to a business object that meets the requirements, and the negative sample includes characteristics corresponding to a business object that does not meet the requirements; generating a plurality of initial rules based on the training sample set, each rule including at least one characteristic; continuously adding new rules based on column generation technology until a rule learning model is obtained; wherein during column generation, new rules are generated based on a frequent pattern tree.

[0067] The positive sample includes characteristics corresponding to a business object that meets the requirements, and the negative sample includes characteristics corresponding to a business object that does not meet the requirements.

[0068] The positive sample set includes a plurality of positive samples, and the negative sample set includes a plurality of negative samples. The positive sample can be a sample corresponding to a result of interest of a user, and the negative sample can be a sample corresponding to a result of no interest of the user. Each business object can correspond to a sample, and the business object can be any business object that needs to be analyzed. The sample corresponding to each business object can include at least one feature of the business object.

[0069] Taking the field of financial examination and approval as an example, meeting the requirement can mean that the examination and approval is not passed, and accordingly, the positive sample can be a sample of examination and approval not passed, the negative sample can be a sample of examination and approval passed, and the business object can be a user initiating the examination and approval. The sample of each user can include at least one feature of the user, and can include but is not limited to at least one of the following: age, income, bank flow, education, occupation, default times, and asset situation.

[0070] Exemplarily, the training sample set can include a plurality of samples as follows.

[0071] The positive sample of examination and approval not passed includes:

[0072] Sample P1: default times 3, income 0, flow 0, and asset no mortgage;

[0073] Sample P2: default times 5, income 1, flow 2, and asset mortgage;

[0074] Sample P3: ……

[0075] The negative sample of examination and approval passed includes:

[0076] Sample N1: default times 0, income 1, flow 1, and asset no mortgage;

[0077] Sample N2: default times 1, income 5, flow 5, and asset mortgage;

[0078] Sample N3: ……

[0079] Optionally, the plurality of rules initially generated can be determined according to actual needs, and the embodiment does not limit this. In the subsequent iteration process, the plurality of rules initially generated can be continuously optimized.

[0080] Each rule can include at least one feature, and the implicit meaning is that if a target business object covers all features included in the rule, the target business object is considered to meet the requirement (such as examination and approval not passed).

[0081] Exemplarily, the initially generated rules and the new rules can adopt the following forms of expression.

[0082] Rule 1: default times is 3;

[0083] Rule 2: breach times is 1, asset is mortgaged;

[0084] Rule 3: ….

[0085] In the field of rule learning, the number of rules will explode with the growth of feature size, so it is difficult to express all the rules explicitly in the model. In the embodiments of the present application, rule learning can be implemented based on column generation technology.

[0086] Column generation technology is a very efficient algorithm for solving large-scale linear optimization problems, which can solve problems by constantly generating new columns. In the embodiments, based on column generation technology, a plurality of rules can be initially generated and put into a rule pool, and new rules can be constantly generated according to the performance of the existing rules in the rule pool, until the newly generated rules cannot make the final result better, and then the model learning is ended.

[0087] Optionally, in the column generation process, new rules can be generated based on the frequent pattern tree. The frequent pattern tree can be used to mine frequent item sets. In rule learning, the frequent pattern tree can intuitively list the number of times each rule covers positive samples, that is, the frequency of the rule, so that the rules that cover more positive samples or cover fewer negative samples can be quickly determined based on the frequent pattern tree, providing a reference for generating new rules.

[0088] Optionally, in the column generation process, based on the frequent pattern tree, the rules that cover more positive samples or cover fewer negative samples can be considered, the determination test numbers corresponding to these rules can be calculated, and new rules can be added when the test numbers meet the requirements, so that the efficiency of generating new rules is improved without traversing all the test numbers.

[0089] Exemplarily, each node in the frequent pattern tree can correspond to a feature, and at least part of the features on each path constitute a rule. If a rule only covers negative samples without covering positive samples, or the number of times a rule covers negative samples is greater than the number of times it covers positive samples, or the difference between the number of times a rule covers positive samples and the number of times it covers negative samples is less than a preset threshold, the rule can not be considered, and the rule with the smallest test number is selected from other rules as a new rule to be added to the rule pool for further iteration.

[0090] In summary, the rule learning method provided in the embodiments of the present application can obtain a training sample set, the training sample set includes a positive sample group and a negative sample group, wherein the positive sample includes a feature corresponding to a business object meeting a requirement, and the negative sample includes a feature corresponding to a business object not meeting the requirement, an initial rule is generated according to the training sample set, each rule includes at least one feature, a new rule is continuously added based on the column generation technology until a rule learning model is obtained, wherein in the column generation process, a new rule is generated based on the frequent pattern tree, so that the data amount processed in each iteration can be reduced based on the column generation technology, and the new rule can be selected to be added to the rule pool based on the frequent degree of various possible rules represented by the frequent pattern tree, the efficiency of rule learning can be improved, the rule learning demand in each business field can be met, the reason for the prediction result is provided, and the user experience is improved.

[0091] In one or more embodiments of the present application, the column generation technology is continuously used to add new rules until the rule learning model is obtained, which can include: constructing a constraint master problem, the constraint master problem includes at least one of the following: accuracy, complexity, and coincidence of the rule learning model; repeatedly performing the following steps until there is no new rule to be added: solving the constraint master problem based on the existing rules to obtain a dual variable; and generating a new rule based on the frequent pattern tree according to the obtained dual variable.

[0092] In the column generation technology, each column can correspond to a rule, after generating a plurality of rules, a constraint master problem can be constructed, the constraint master problem is usually a maximization or minimization problem. After constructing the constraint master problem, the constraint master problem can be solved to obtain a dual variable, and then a new rule can be found based on the dual variable.

[0093] Optionally, in the embodiments of the present application, the constraint master problem can include at least one of the accuracy, complexity, and coincidence of the rule learning model. The accuracy can be used to represent the ability of the model to accurately predict the result, the complexity can be used to represent the complexity of the rules included in the model, and the complexity can be the number of features included in the rule, and the coincidence can be used to represent the diversity of the rules included in the model. If each positive sample is covered by multiple rules, it means that the coincidence is high, the rules are not diversified, and the diversity is poor.

[0094] For example, the constraint master problem can be described as: the highest accuracy, or the lowest complexity, or the lowest weighted sum of the complexity and the coincidence, and the like.

[0095] By mapping various possible rules to columns in the column generation technique, and then constructing a restriction master problem according to the accuracy, complexity, coincidence degree and the like of the rule learning model, and generating new rules based on the restriction master problem, the restriction master problem can be effectively constructed and solved according to the requirements of the rule learning model to complete the generation of the rule learning model, and the accuracy, complexity, coincidence degree and the like of the rule learning model can be optimized.

[0096] In one or more embodiments of the present application, the initial rules can be generated by generating initial rules according to the features contained in the positive sample set and / or the negative sample set, wherein the initial rules are rules containing a single feature. The initial rules can be put into a rule pool.

[0097] Optionally, single features with different values are divided into different rules. For example, "the number of defaults is equal to 3" can be a rule, and "the number of defaults is equal to 2" can also be a rule. Assuming that there are 4 types of features: default times, flow, income, and asset information, and each type of feature has n values, a total of 4n rules containing a single feature can be generated.

[0098] In an optional implementation, the initial rules can be generated based on the positive sample set. Specifically, the initial rules are single features that appear in the positive sample, and rules that do not appear in the positive sample are not considered, thereby reducing the data volume and improving the efficiency.

[0099] In another optional implementation, the initial rules can be generated based on the positive sample set and the negative sample set. Specifically, the single features appearing in the positive sample set and the negative sample set are used as the initial rules, and in the subsequent calculation process, the decision variables are used to distinguish whether an initial rule is adopted or not. This way can more comprehensively cover various rules and improve the rule learning effect.

[0100] In other optional implementations, the initial rules can also be generated based on the negative sample set, for example, among all possible single features, the single features that do not appear in the negative sample or have a low frequency of appearance in the negative sample are selected as the initial rules.

[0101] According to the features contained in the positive sample set and / or the negative sample set, the initial rules are generated, each rule contains a single feature, which can make the initial rules simple and comprehensive, and improve the overall efficiency of rule learning.

[0102] Optionally, according to the optimization target of the rule learning model, the original master problem can be constructed as follows:

[0103]

[0104]

[0105]

[0106] θ∈{0,1} q ,

[0107] μ≥0, v≥0, ω≥0

[0108] Where P represents the positive sample group, N represents the negative sample group, and ω i This represents the number of times a negative sample i is incorrectly covered by the rule describing the positive sample. The vector is used to represent whether sample i is covered by rule j. If the j-th rule covers sample i, then p ij =1, otherwise p ij =0, μ i This indicates whether positive sample i is not covered by any rule describing positive samples, v i c is the number of times a positive sample i is repeatedly covered by multiple rules describing positive samples. T θ represents the sum of the lengths of the rules describing positive samples. θ is a decision variable used to indicate whether each rule is adopted as a rule describing positive samples. α, β1, and β2 are the complexity coefficient, accuracy coefficient, and repeatability coefficient, respectively.

[0109] To improve the efficiency of solving the problem, the original master problem can be replaced by an equivalent substitution, and the rules considered can be restricted to a small rule pool. The decision variables can be relaxed to between 0 and 1, thus obtaining the restricted master problem.

[0110] In one or more embodiments of this application, optionally, constructing the restricted master problem may include: constructing the restricted master problem based on decision variables of existing rules; the decision variable corresponding to each rule is used to indicate whether the rule is adopted.

[0111] The objective function of the constrained main problem may include at least one of the following: the number of negative samples covered by the adopted rule and / or the number of positive samples not covered by the adopted rule, used to characterize the accuracy of the rule learning model; the number of positive samples covered multiple times by the adopted rule, used to characterize the overlap of the rule learning model; and the sum of the lengths of the adopted rules, used to characterize the complexity of the rule learning model.

[0112] The constraints on the main problem may include: for any positive sample, the sum of the uncovered result and the multiple coverage result of the positive sample is greater than or equal to 1; wherein, if the adopted rule does not cover the positive sample, the uncovered result is 1, otherwise it is 0; if the adopted rule covers the positive sample multiple times, the multiple coverage result is the coverage number minus 1; the decision variable takes values ​​between 0 and 1.

[0113] Specifically, the limited master problem can be constructed based on existing rules in a rule pool, and the rules contained in the rule pool are denoted as existing rules. Each rule in the rule pool can be determined whether to be adopted by its corresponding decision variable. After the iteration is completed, the rules contained in the model are all the adopted rules, and when predicting, if the characteristics of the business object to be analyzed match a rule in the model, it is considered to meet the requirements.

[0114] Optionally, the limited master problem can be expressed by the following formula.

[0115]

[0116]

[0117] 0≤θ j ≤1,j∈R′

[0118] μ≥0,v≥0,ω≥0

[0119] wherein θ j is a decision variable of the jth rule, and takes a value of 0 or 1, 1 representing that the rule is adopted, and 0 representing that the rule is not adopted, and the adopted rule corresponds to a rule describing a positive sample in the original master problem. j is the length of the jth sample. The vector represents the case that the ith sample is covered by the existing rules, but the corresponding existing rule is the limited version, which is different from the full version in the original master problem.

[0120] In the limited master problem, the objective function can include four parts, and each part can have its corresponding coefficient, and the limited master problem is constructed by linear weighting. The following will be explained respectively.

[0121] The first part is the number of all negative samples covered by the adopted rules. Assuming that there are 10 adopted rules, and there are 50 negative samples in the training sample set, for each negative sample, if the negative sample contains one or more of the adopted rules, it is considered that the adopted rules cover the negative sample.

[0122] Optionally, the negative sample group can be represented by a two-dimensional matrix, in which the row represents the sample and the column represents the rule. The ith row can be the ith negative sample, and the jth column can be the jth rule. If the jth rule covers the ith sample, the element of the ith row and the jth column in the matrix is 1, otherwise it is 0. The decision variables of the existing multiple rules can be represented as a vector, and the jth element of the vector is the decision variable corresponding to the jth rule. In this way, the negative sample matrix and the decision variable vector are directly multiplied, and all the negative samples covered by the adopted rules are obtained.

[0123] The second part is the number of positive samples not covered by the adopted rules. In practical applications, it is always desirable that the more positive samples are covered by the adopted rules and the less negative samples are not covered by the adopted rules, and therefore, the first part and the second part combined together can be used to represent the accuracy of the rule learning model. In other optional implementations, only the first part or only the second part can be retained.

[0124] Optionally, the positive sample group can also be represented by a two-dimensional matrix, so that the positive sample matrix and the vector of decision variables can be used to more accurately and quickly determine the positive samples covered by the rules.

[0125] The third part is the number of positive samples covered by the adopted rules multiple times, corresponding to the degree of overlap. Specifically, a positive sample can be covered by one or more rules, or can not be covered by any rule. For each positive sample, if it is not covered by any rule, its corresponding second part and third part are 1 and 0 respectively, and if it is covered by multiple rules, the second part is 0 and the third part is the number of coverings minus 1.

[0126] For example, rule 1: number of defaults = 1; rule 2: income = 0. If a positive sample is: number of defaults = 1 and income = 0, this positive sample is covered by the two rules, and the value of the third part should be 2-1 = 1. In practical applications, it can be desirable to have fewer cases covered multiple times to reduce the degree of overlap and improve diversity.

[0127] The fourth part is the sum of the lengths of the adopted rules, corresponding to the complexity. Among them, the length of each rule is the number of features contained in the rule, for example, the length of the rule is 2 if it contains 2 features, and the longer the rule, the higher the complexity.

[0128] In the constraint problem, the first type of constraint condition is: for any positive sample, the sum of the second part and the third part corresponding to the positive sample is greater than 1. The meaning is that for each positive sample, it is either not covered or covered at least once. Specifically, for any positive sample, if it is not covered by the adopted rules, the second part and the third part are 1 and 0 respectively, and if it is covered by at least one adopted rule, the second part is 0 and the third part is an integer greater than or equal to 1, and therefore, whether the positive sample is covered by the adopted rules or not, the sum of the second part and the third part corresponding to the positive sample is greater than or equal to 1.

[0129] The second type of constraint condition is that the range of the decision variable is between 0 and 1. Specifically, the value of the decision variable should be 0 or 1, but limiting the value of the decision variable to 0 or 1 increases the difficulty of solving, so the range of the decision variable can be relaxed to 1 to 0, which can solve the restricted master problem faster, and the value of the decision variable solved finally falls to 1 or 0.

[0130] By constructing the restricted master problem by the above formula, the rule learning and column generation technology can be combined to realize the modeling of the rule generation corresponding to the restricted master problem, and the solution can be solved from the accuracy, overlap, complexity and other dimensions to meet the actual requirements of rule learning.

[0131] In practical applications, the above four parts may not all be necessary, and one or more parts can be added, reduced or modified as needed. For example, the complexity or overlap part can be deleted. For another example, the support degree of the rule covering the positive sample can be added in the objective function, the larger the better, and the total number of rules adopted is smaller, the better. Or, when searching for rules, the support degree of the new rule covering the positive sample cannot be less than the first threshold value, and the total number of rules adopted cannot exceed the second threshold value in the constraint condition, wherein the first threshold value and the second threshold value can be default values, can be input by the user, or can dynamically change with the total number of samples.

[0132] After constructing the objective function and the constraint condition, the existing column generation technology can be used to solve the restricted master problem to obtain the decision variable and the dual variable. Among them, the constraint condition of each positive sample can correspond to a dual variable, and when the restricted master problem is solved in each iteration, the dual variable corresponding to each constraint is extracted for subsequent calculation of the check number.

[0133] In one or more embodiments of the present application, optionally, each term in the restricted master problem is provided with a coefficient; the method further comprises: obtaining at least one of the accuracy coefficient, the overlap coefficient, and the complexity coefficient input by the user; correspondingly, the restricted master problem is constructed by at least one of the coefficient input by the user, the coefficient determined according to the sample number, and the coefficient set by default.

[0134] Specifically, increasing the coefficient can adjust the importance of each index, and the larger the coefficient, the greater the punishment. These coefficients can be input by the user, or use the default recommended value, or the user can input part of the coefficients and the rest of the coefficients use the default value. By allowing the user to input the coefficient, the learning needs of different users and different training scenarios can be met, which has strong adaptability.

[0135] Optionally, each coefficient can also be dynamically adjusted according to the training sample set. For example, the complexity coefficient can be 1% of the sample size. The more samples, the greater the complexity coefficient, so that the selection of the coefficient can better match the actual training samples and improve the learning effect.

[0136] In one or more embodiments of the present application, optionally, the bitmap information of each feature covering the positive samples and the bitmap information of each feature covering the negative samples can also be counted, wherein the bitmap information is used to achieve at least one of the following: determining the positive samples and / or the negative samples covered by the feature according to the bitmap information corresponding to any feature; calculating the positive samples and / or the negative samples covered by the rules corresponding to at least two features according to the bitmap information corresponding to the at least two features.

[0137] For example, whether each feature covers a sample can be represented by 1 and 0, for example, 1 represents covering the sample, and 0 represents not covering. Therefore, for each feature, its corresponding BitMap can be constructed, and the i-th bit in the BitMap is used to indicate whether the i-th sample is covered by the feature. Considering the feature of the number of defaults = 3, in a positive sample group with a sample size of 100, it is assumed that the first, third, and fifth positive samples contain the number of defaults = 3, and the other positive samples do not contain the feature of the number of defaults = 3. Then, the positive sample BitMap corresponding to the feature can contain 100 bits, the first, third, and fifth bits are 1, and the other bits are 0. In actual application, when it is necessary to view the positive samples covered by the feature, the BitMap can return the identification of the covered positive samples, for example, return 1, 3, and 5, which means that the feature covers the first, third, and fifth positive samples. Similarly, the negative sample BitMap corresponding to each feature can be constructed.

[0138] The BitMap can be used to assist in solving the constrained master problem. Specifically, the positive and negative samples covered by the rules can be used in the constrained master problem, and each rule can contain at least one feature. When a rule contains one feature, the BitMap corresponding to the rule can be directly used to determine which positive and negative samples are covered. When a rule contains multiple features, the BitMap corresponding to each feature can be intersected to determine which positive and negative samples are covered by the rule.

[0139] For example, when it is necessary to determine the positive samples covered by the rule of the feature of the number of defaults = 3 and the income = 0, the positive sample BitMap corresponding to the number of defaults = 3 and the positive sample BitMap corresponding to the income = 0 are intersected. If the i-th bit in the result is 1, it means that the i-th positive sample is covered.

[0140] In addition to being used for calculating the constraint master problem, the BitMap can also be used for subsequent analysis of the frequent pattern tree or calculation of the test number lower bound, and can be used to find the samples covered by the rule. The BitMap can effectively improve the efficiency and accuracy of finding the covered positive and negative samples, and thus improve the overall efficiency and accuracy of rule learning.

[0141] In one or more embodiments of the present application, optionally, based on the obtained dual variable, a new rule is generated based on the frequent pattern tree, including: constructing a corresponding frequent pattern tree through any rule head, calculating the test number lower bound corresponding to the rule head according to the dual variable, to determine whether the rule head needs to be analyzed, and performing recursive analysis on the frequent pattern tree corresponding to the rule head that needs to be analyzed, and finding the rule with the minimum test number as the new rule; wherein the nodes in the frequent pattern tree are used to represent features, and the rule head includes one or more features; the frequent pattern tree corresponding to the rule head is a frequent pattern tree constructed according to the positive samples containing the rule head; and the test number lower bound corresponding to the rule head is used to represent the test number lower bound corresponding to each path rule in the frequent pattern tree corresponding to the rule head.

[0142] Optionally, the dual variable can be obtained by solving the constraint master problem, and then the test number corresponding to each rule can be calculated according to the dual variable. The rule corresponding to the minimum test number can be added to the rule pool as a new rule for the next iteration.

[0143] For example, for the rule c j , its corresponding test number can be calculated by the following formula.

[0144]

[0145] Where λ i is the dual variable corresponding to the ith sample, and p ij = 1 if the jth rule covers the sample i, otherwise p ij = 0, and the meanings of the remaining parameters can be found in the foregoing embodiments.

[0146] In finding the minimum test number, a large amount of computing power is required if each rule is calculated one by one. In this embodiment, a frequent pattern tree can be constructed based on different rule heads, the test number lower bound corresponding to the frequent pattern tree is calculated, and it is determined whether it is necessary to continue analyzing a rule head based on the test number lower bound. Only the rules contained in the frequent pattern tree that need to be analyzed are analyzed, which can effectively reduce the amount of calculation and improve the processing efficiency.

[0147] The construction method of the frequent pattern tree will be explained below. Optionally, the construction of the frequent pattern tree can include the following steps 1a) to 1b).

[0148] Step 1a), count the frequency of single features in the positive sample group, and discard features with a frequency less than a preset frequency threshold.

[0149] Table 1 is a sample example, each sample can include: sample ID, feature, weight, label. Among them, the features contained in the sample, A, B, C can refer to different dimensions of features, for example, default times, income, and flow, a1, b1, c1, etc. can be specific values, for example, A=a1, B=b1, C=c1 respectively refer to: default times=3, income=0, flow=0. The weight can represent the number of times the sample appears, for example, if two samples have exactly the same features, only one record can be recorded, and the weight is 2, indicating that there are two such samples. The label can be used to distinguish between positive and negative samples, where 1 is a positive sample and 0 is a negative sample, for example, it can represent disapproval or approval, respectively.

[0150] Table 1 sample example

[0151] Sample ID Feature Weight Label 1 A = al, B = bl, C = cl 1 1 2 A = al, B = b2, C = cl 2 1 3 A = al, B = b3, C = cl 2 1 4 A = a2, B = bl, C = c2 1 1 5 A = a2, B = bl, C = c3 3 1 6 A = a3, B = bl, C = c3 1 1 7 A = al, B = bl, C = cl 1 0 … … … … 1000 A = a2, B = b2, C = c2 2 1

[0152] Optionally, the positive sample group can be scanned once to count the frequency of single features in all positive samples, for example, the frequency of A=a1, the frequency of A=a2, the frequency of B=b1, and so on. After counting the frequency of single features, features with a frequency less than a threshold can be discarded to further improve processing efficiency. Table 2 is an example of single feature frequency, which shows the frequency of each single feature in the positive sample.

[0153] Table 2 single feature frequency example

[0154] Feature Weight B = bl 6 A = al 5 C = cl 5 A = a2 4 C = c3 4 B = b2 2 B = b3 2

[0155] Step 1b), initialize the root node according to the current rule head, and construct a frequent pattern tree based on the frequency of the single features counted.

[0156] Different rule heads are used to construct different frequent pattern trees, and each rule head corresponds to a frequent pattern tree that is constructed based on all positive samples containing the rule head.

[0157] In an example, the rule head is empty, that is, the corresponding frequent pattern tree is constructed based on all positive samples. Figure 6 A schematic diagram of a frequent pattern tree provided by an embodiment of the present application. As shown in the figure, Figure 6As shown, for each positive sample, the features contained therein can be sorted in descending order according to the single-feature frequencies calculated in the previous step. Specifically, for each feature therein, if a path from the root node to the feature already exists, the frequency corresponding to the feature is added to the corresponding node, and the number on the right of the feature in the figure represents the frequency, and the number in the bracket below represents the sample ID in which the feature appears; if not, a new downstream node is constructed for the feature.

[0158] In another example, the rule head can include one feature, for example, the rule head is B=b1, all positive samples containing the feature B=b1 can be found, and a frequent pattern tree is constructed based on the positive samples, and the root node in the frequent pattern tree can be B=b1.

[0159] In another example, the rule head can include multiple features, for example, the rule head is B=b1 and A=a1, all positive samples containing the features B=b1 and A=a1 can be found, and a frequent pattern tree is constructed based on the positive samples, and the root node in the frequent pattern tree can be B=b1 and A=a1.

[0160] In actual applications, a frequent pattern tree can be constructed based on at least one rule head, and whether each rule in the frequent pattern tree corresponding to the rule head needs to be further analyzed can be determined by calculating the lower bound of the test number corresponding to the rule head, and the analysis is only performed when necessary.

[0161] Optionally, if the lower bound of the test number corresponding to the rule head is greater than or equal to 0, it indicates that the test numbers of each rule in the frequent pattern tree corresponding to the rule head are all greater than or equal to 0, and there is no need for further analysis. Alternatively, if a part of the test numbers corresponding to the rules have been calculated in the process of finding new rules, and the lower bound of the test number corresponding to the rule head is greater than or equal to the minimum test number in the part of the test numbers, it indicates that the test numbers of each rule in the frequent pattern tree corresponding to the rule head will not be smaller than the minimum test number currently calculated, and there is no need for analysis.

[0162] Illustratively, assuming that there are 100 rules in the current rule pool, when a new rule is found as the 101st rule, at least one rule head can be designed, and each rule head is analyzed respectively, if the lower bound of the test number corresponding to the rule head is greater than or equal to 0, the analysis of each rule in the frequent pattern tree corresponding to the rule head is skipped, or, according to the current dual variable, a part of the test numbers corresponding to the rules can be calculated, assuming that the minimum test number in the calculated test numbers is -0.5, if the lower bound of the test number corresponding to a rule head is -0.3, the analysis of each rule in the frequent pattern tree corresponding to the rule head can also be skipped, and further analysis is not necessary.

[0163] After finding the new rule and adding it to the rule pool, the dual variables can be solved again according to the 101 rules in the rule pool, and new rules can be found according to the dual variables and the frequent pattern tree.

[0164] The lower bound of the test number corresponding to the rule head is calculated to determine whether the rule head needs to be analyzed, and the frequent pattern tree corresponding to the rule head that needs to be analyzed is analyzed recursively, and the rule head that does not need to be analyzed is directly skipped. The lower bound of the test number can be used to realize pruning, reduce the amount of calculation, and improve the efficiency of finding new rules.

[0165] In one or more embodiments of the present application, the corresponding frequent pattern tree can be constructed through any rule head, the lower bound of the test number corresponding to the rule head is calculated according to the dual variable to determine whether the rule head needs to be analyzed, and the frequent pattern tree corresponding to the rule head that needs to be analyzed is analyzed recursively, and the rule with the smallest test number found is the new rule, which can include steps 2a) to 2d).

[0166] Step 2a), initialize the optimal test number as 0, and set the initial rule head as null.

[0167] Optionally, the optimal test number is denoted as BestScore, and the initial rule head Cond is Null (empty).

[0168] Step 2b), based on the positive samples satisfying the current rule head, a frequent pattern tree is constructed, and steps 2c) to 2d) are repeatedly executed until the rules included in the positive sample group are analyzed.

[0169] Wherein, when the rule head is empty, all positive samples can be scanned to construct the corresponding frequent pattern tree. Optionally, the construction of the frequent pattern tree can be realized by referring to the aforementioned steps 1a) to 1b).

[0170] Step 2c), according to the frequent pattern tree corresponding to the current rule head, a set of single-feature nodes with an appearance frequency greater than or equal to a preset frequency threshold is determined, for each element in the set, the test number corresponding to the rule formed by the element and the current rule head is determined, and when the corresponding test number is less than the optimal test number, the optimal test number is updated to the corresponding test number.

[0171] All single-feature nodes with an appearance frequency greater than or equal to a preset frequency threshold can be added to the set, regardless of whether they are directly connected to the root node.

[0172] For each element in the set, a rule is formed by the element and the current rule head, and the test number corresponding to the rule is calculated, for example, B = b1 is an element in the set, B = b1 is combined with the current rule head Null to form a rule B = b1, and the test number corresponding to the rule is calculated, and if the test number is better than BestScore, BestScore is updated, and in the embodiment, better can mean smaller.

[0173] Step 2d), for each element in the set, the lower bound of the test number corresponding to the element is calculated, if the lower bound of the test number is greater than or equal to the optimal test number, the element is skipped, otherwise, a new rule head and a corresponding frequent pattern tree are constructed based on the depth-first recursive method.

[0174] The lower bound of the test number corresponding to the element can mean that when the element and the current rule head form a new rule head, the lower bound of the test number of the frequent pattern tree corresponding to the new rule head.

[0175] For example, the current rule head is Null, and an element in the set is e, the lower bound of the test number corresponding to the rule head Cond = Null&e is calculated, and if the lower bound of the test number is greater than or equal to BestScore, the element is skipped.

[0176] If the lower bound of the test number is less than BestScore, it means that the current element can be further analyzed, and a new rule head and a corresponding frequent pattern tree can be further constructed based on the element.

[0177] Optionally, when the feature is added, the added feature can be determined based on the depth-first recursive method. The depth-first recursive method can mean that in the search process, the longest path is always analyzed first, and then the search is restarted from the last branch.

[0178] After a new rule head is constructed each time, steps 2c) and 2d) are executed again until all rules contained in the positive sample group are analyzed, and here analyzed can mean that the test number is calculated or skipped. In all possible rules, part of the rules have been calculated for the test number, and the remaining part is skipped, and then it can be considered that the process of finding a new rule is completed, and the new rule is the rule with the smallest test number.

[0179] By initializing the optimal test number and the rule head and constructing the corresponding frequent pattern tree, and then optimizing the minimum test number based on the single feature of the current frequent pattern tree as the comparison benchmark of the calculated lower bound of the test number, compared with comparing the lower bound of the test number with 0, the traversal of all rules can be more quickly realized, and the smallest test number can be more quickly found, and the processing efficiency is further improved.

[0180] In one or more embodiments of this application, optionally, step 2d) may include: for each element in the set, if there are still nodes to be analyzed along the depth direction starting from the element on the path where the rule formed by the element and the current rule head is located in the initial frequent pattern tree, then calculate the lower bound of the test number corresponding to the element; if the lower bound of the test number is greater than or equal to the optimal test number, then skip the analysis of the element; if the lower bound of the test number is less than the optimal test number, then construct a frequent pattern tree based on the new rule head formed by the current rule head and the element; if there are no nodes to be analyzed along the depth direction, then in the initial frequent pattern tree, backtrack from the element to find the next node to be analyzed, and construct a frequent pattern tree based on the rule head corresponding to the next node.

[0181] In practical applications, frequent pattern trees constructed using rule header nulls are as follows: Figure 6 As shown, this frequent pattern tree covers all features that have appeared in the positive sample group. The possible rules include not only rules corresponding to paths from the root node to leaf nodes (e.g., rules consisting of the features B=b1, A=a1, and C=c1), but also rules corresponding to paths from the root node to intermediate nodes (e.g., rules consisting of the features B=b1 and A=a1), and even rules corresponding to combinations of non-adjacent nodes (e.g., rules consisting of the features B=b1 and C=c1). Analyzing only the rules corresponding to each path in the current frequent pattern tree may not cover all rules. Therefore, each node in the current frequent pattern tree can be analyzed, and new rule heads can be constructed based on each node for further analysis until all possible rules have been analyzed.

[0182] Optionally, after constructing the frequent pattern tree with a null rule header, the leaf nodes in the current frequent pattern tree can be analyzed first, gradually progressing from leaf nodes to intermediate nodes to the root node. For each element in the set, if there are nodes to be analyzed along the depth direction, it indicates that further analysis is needed. The lower bound of the test number corresponding to the element can be calculated. If the lower bound of the test number is greater than the optimal test number, the element is skipped from analysis. If there are no nodes to be analyzed along the depth direction for the element, there is no need to further analyze it, and therefore no need to calculate the corresponding lower bound of the test number.

[0183] The following describes a specific implementation scheme of rule search according to the frequent pattern number, taking the depth direction from the leaf node to the root node as an example. In combination with steps 2a) to 2d), first, the initial BestScore is set to 0, and the initial rule head is Null. After the frequent pattern tree is constructed with the rule head Null, in the set E1 composed of single-feature nodes with the occurrence frequency greater than or equal to the preset frequency threshold, the test number corresponding to the rule composed of each element e and the rule head Null, i.e., the test number corresponding to the rule e, is calculated, and when the test number corresponding to the element e is less than the current BestScore, the BestScore is updated, so that the BestScore is always less than 0 and consistent with the minimum value in the test numbers that have been calculated.

[0184] Further, each element e in the set E1 is analyzed. Optionally, since the depth direction is from the leaf node to the root node, the analysis can be started from the leaf node, for example, B = b3.

[0185] Since B = b3 has other nodes C = c1 and A = a1 towards the root node, i.e., the rule composed of the element B = b3 and the current rule head Null, i.e., B = b3, there are nodes to be analyzed in the direction from the element B = b3 towards the root node on the path in the initial frequent pattern tree, so the test number lower bound corresponding to the element B = b3 can be calculated. If the test number lower bound is greater than the optimal test number, the element is skipped and there is no need to build a new frequent pattern tree with the element B = b3 as the rule head. If the test number lower bound is less than the BestScore, the frequent pattern tree can be built based on B = b3 as the rule head for analysis.

[0186] Figure 7 A schematic diagram of a frequent pattern tree constructed based on a non-empty rule head provided by an embodiment of the present application is shown in FIG. 1. Figure 7 As shown, when B = b3 is taken as the rule head, the frequent pattern tree is constructed according to all positive samples containing the feature B = b3, and the frequent pattern tree contains the path of B = b3, C = c1, and A = a1.

[0187] For each element in the set E2 composed of single-feature nodes C = c1 and A = a1 in the current frequent pattern tree, the corresponding test number is calculated, and the BestScore is updated when the test number is better than the BestScore.

[0188] For each element in the set E2, if the rule composed of the element and the current rule head has nodes to be analyzed in the depth direction on the path in the initial frequent pattern tree, the test number lower bound corresponding to the element is calculated.

[0189] Taking the element C=c1 in the set E2 as an example, there is a node A=a1 to be analyzed along the depth direction (towards the root node) from the element C=c1 on the path where the element C=c1 and the current rule head B=b3 are located in the initial frequent pattern tree, and therefore the element C=c1 needs to be further analyzed: a lower bound of the test number corresponding to the element C=c1 is calculated, which is used to represent the lower bound of the test number of the frequent pattern tree corresponding to the new rule head composed of the element C=c1 and the current rule head B=b3, i.e., the rule head <C=c1 and B=b3>, and if the lower bound of the test number is greater than or equal to BestScore, it is not necessary to further analyze the frequent pattern tree constructed with the rule head <C=c1 and B=b3>, and if the lower bound of the test number is less than BestScore, the frequent pattern tree can be constructed with the rule head <C=c1 and B=b3>.

[0190] Since the positive sample containing B=b3 and the positive sample containing C=c1 and B=b3 are completely identical in the training sample set, the frequent pattern tree constructed with the rule head <C=c1 and B=b3> is also identical to the frequent pattern tree constructed with the rule head <B=b3>. Figure 7

[0191] It should be noted that the new frequent pattern tree constructed based on a certain node can not be completely identical to the path containing the node in the initial frequent pattern, for example, when constructing the new frequent pattern tree, some features on the path where the node is located in the initial frequent tree can be filtered out because the occurrence frequency of the features does not meet the requirement, and therefore the features can be skipped when analyzing the new frequent pattern tree. In addition, when a feature corresponds to multiple leaf nodes in the initial frequent pattern tree, the new frequent pattern tree constructed can contain multiple rule paths in the initial frequent pattern tree, and all the multiple rule paths need to be analyzed.

[0192] In this frequent pattern tree, C=c1 and B=b3 are the rule head, and the single-feature node only has A=a1, and therefore the set E3 of the single-feature nodes corresponding to the current frequent pattern tree only contains the element A=a1, and the test number corresponding to the element A=a1 is calculated, and if the test number is better than BestScore, BestScore is updated.

[0193] Then, for the element A=a1, it is judged whether there is a node to be analyzed along the depth direction from the element. As can be seen from the initial frequent pattern tree, there is no other node to be analyzed along the direction towards the root node from the element A=a1 on the path where the element A=a1 and the current rule head <C=c1 and B=b3> are located, and therefore the element A=a1 does not need to calculate the lower bound of the test number corresponding to the element, and the frequent pattern tree constructed with the current rule head <C=c1 and B=b3> is analyzed. Therefore, in the following description, the frequent pattern tree constructed with the rule head <C=c1 and B=b3> is referred to as the current frequent pattern tree.​Figure 7 In the embodiment, the analysis direction is gradually deepening from the leaf node to the intermediate node and then to the root node, as indicated by the arrow.

[0194] When there is no node to be analyzed in the depth direction, the next node to be analyzed can be found from the element in the initial frequent pattern tree, and the frequent pattern tree can be constructed based on the rule head corresponding to the next node. Assuming that in the initial frequent pattern tree, the node C=c1 is connected to the root node through the node A=a1 in addition to other nodes, the node C=c1 can be returned to, and the other nodes in the depth direction can be analyzed until all the nodes in the depth direction of the node C=c1 are traversed.

[0195] From the perspective of the set, after the elements in the set E3 are analyzed, the elements in the set E2 that have not been analyzed are further analyzed. Specifically, in the set E2, the element C=c1 has been analyzed, and for the element A=a1, it can be determined whether it needs to be analyzed. In the path containing A=a1 and the rule head B=b3, the possible rules have been analyzed, and the inspection number lower bound of A=a1 does not need to be calculated. Assuming that there are other nodes between A=a1 and the root node, the inspection number lower bound of A=a1 can be further calculated. Figure 7

[0196] After the set E2 is analyzed, the remaining elements in the set E1 can be analyzed again. The specific implementation is similar to the implementation of analyzing B=b3, which will not be described here.

[0197] Through the above scheme, various possible rules can be traversed according to the depth-first recursive method, and all unanalyzed rules can be quickly and accurately found for analysis and processing by constantly exploring the information in the depth direction, and the rules that have been analyzed can be skipped, thereby further improving the efficiency of finding the minimum inspection number.

[0198] In some other optional implementations, after the set of single-feature nodes with an occurrence frequency greater than or equal to the preset frequency threshold is determined according to the frequent pattern tree corresponding to the current rule head, the nodes that have been analyzed in the set can be removed.

[0199] ​For example, in the initial frequent pattern tree, the corresponding set E1 includes at least two elements B=b3 and C=c1. The leaf node B=b3 is analyzed first as the rule header corresponding frequent pattern tree. During the analysis, the test number corresponding to the rule composed of B=b3 and C=c1 is calculated. After the analysis of B=b3 is completed, C=c1 in set E1 is analyzed as the rule header corresponding frequent pattern tree. The node B=b3 is included in the frequent pattern tree corresponding to C=c1 as the rule header. At this time, the test number corresponding to the rule composed of B=b3 and C=c1 has been calculated, and thus it is not necessary to calculate again. That is, in the initial frequent pattern tree, each element in set E1 is analyzed along the direction from the element to the root node. The node away from the root node direction of the element can be deleted from the single feature node set of the frequent pattern tree corresponding to the element as the rule header. Thus, the recursive processing can be performed in a certain order, the number of repeated calculations is reduced, the calculation amount is reduced, and the efficiency is further improved.

[0200] Figure 8 A set schematic diagram in a recursive process is provided for the embodiments of the present application. As shown in Figure 8 the corresponding single feature node set E1 can be {B=b3, B=b2, …} after the initial frequent pattern tree is constructed. The corresponding single feature node set E2 can be {C=c1, A=a1} after the frequent pattern tree is constructed with B=b3 as the rule header. Further, the corresponding single feature node set E3 can be {A=a1} after the frequent pattern tree is constructed with the new rule header composed of C=c1 and the current rule header B=b3. Each level set can be analyzed in turn. After the element B=b3 in E1 is analyzed, the element B=b2 can be analyzed, and the analysis is continued until all possible rules included in the positive sample set are analyzed.

[0201] In some optional implementation manners, the single feature node set corresponding to the initial frequent pattern tree can also not be analyzed in the depth direction from the leaf node to the root node, but can be analyzed in the depth direction from the root node to the leaf node. For example, B=b1 is analyzed first, and then A=a1 is analyzed, and the analysis is continued in turn. Alternatively, the analysis can be started from the middle node, as long as the traversal of all possible rules is completed. The present application does not limit the specific traversal order.

[0202] In one or more embodiments of the present application, after the initial frequent pattern tree is constructed, a node chain table can be further constructed according to all nodes in the initial frequent pattern tree, the node chain table being used to indicate nodes on the initial frequent pattern tree that belong to the same feature but are distributed on different paths; and accordingly, the frequent pattern tree is constructed according to the new rule head formed by the current rule head and the element, including: constructing the frequent pattern tree corresponding to the new rule head formed by the current rule head and the element according to the node chain table.

[0203] Figure 9 A schematic diagram of a node chain table is provided for the embodiments of the present application. As shown in the figure, Figure 9 After the frequent feature tree is constructed, a node chain table can be constructed for all nodes on the tree, and the nodes on the tree that belong to the same feature but are distributed on different rule paths are linked through the chain table (shown by the dashed line in the figure).

[0204] For example, node B=b1 is distributed on one rule path, node A=a1 is distributed on two rule paths,..., and node B=b3 is distributed on one rule path.

[0205] When a new frequent pattern tree needs to be constructed based on a certain node, the construction can be implemented according to the node chain table. At least one rule path corresponding to the node in the chain table constitutes the frequent pattern tree corresponding to the node. For example, when the frequent pattern tree corresponding to node B=b3 as the rule head needs to be constructed, only the rule path corresponding to node B=b3 in the chain table needs to be found. In this way, the construction process can be effectively simplified, and the efficiency of constructing the frequent pattern tree can be improved through the node chain table.

[0206] Optionally, the frequent pattern tree corresponding to the new rule head formed by the current rule head and the element is constructed according to the node chain table, including: determining the rule path corresponding to the new rule head according to the node chain table, and constructing the frequent pattern tree according to the nodes on the rule path from the element to the root node.

[0207] For example, when the frequent pattern tree with element B=b3 in set E1 as the rule head is constructed, the rule path where B=b3 is located can be found according to the chain table, and the frequent pattern tree with B=b3 as the rule head can be constructed by going up along the rule path to the root node. When the frequent pattern tree with B=b3 and C=c1 as the rule head is constructed, the frequent pattern tree with B=b3 and C=c1 as the rule head can be constructed by going up along the rule path from C=c1 to the root node on the basis of the aforementioned rule path.

[0208] When a frequent pattern tree with the element C=c1 in the set E1 as the rule head needs to be constructed, the rule path in which the element C=c1 is located can be determined according to the node chain table, and then the frequent pattern tree with the element C=c1 as the rule head is constructed in the direction towards the root node, and can not contain the element B=b3, so that, in combination with the aforementioned sequence of analysis in the depth direction from the leaf node to the root node, repeated analysis can be effectively reduced, thereby realizing faster rule search based on the node chain table.

[0209] In one or more embodiments of the present application, optionally, when calculating the lower bound of the test number corresponding to the element, the lower bound of the test number can be calculated according to at least one of the following costs: the number of features contained in the current rule; the negative of the sum of the dual variables corresponding to the positive samples covered by the current rule; the support degree of the path rule from the root node to the leaf node in the frequent pattern tree constructed with the current rule as the rule head in the positive sample group, which covers the least positive samples; the support degree of the path rule from the root node to the leaf node in the frequent pattern tree constructed with the current rule as the rule head in the negative sample group, which covers the least negative samples; wherein, when calculating the lower bound of the test number corresponding to the element, the current rule is the rule formed by the element and the current rule head.

[0210] For example, assuming that the current rule head is Null and the element is B=b1, when calculating the lower bound of the test number corresponding to the element, the current rule is set as the rule formed by Null and B=b1, i.e. the rule B=b1, and at least one of the above costs is considered.

[0211] The first cost is the number of features contained in the current rule. Since the current rule B=b1 only contains one feature, the cost thereof can be 1, which is used to represent the complexity. In the case where other costs are fixed, the higher the complexity, the larger the lower bound of the test number, and therefore, the smaller the length of the current rule is better.

[0212] The second cost is the negative of the sum of the dual variables corresponding to the positive samples covered by the current rule. Assuming that the current rule B=b1 covers 40 positive samples, the sum of the dual variables corresponding to the 40 positive samples is taken as the negative to obtain the second cost. For any rule, due to the existence of the dual variable, the test number decreases with each additional positive sample covered by the rule, and the more positive samples covered, the smaller the test number, which is equivalent to a smaller cost.

[0213] The third cost is the support degree of the path rule from the root node to the leaf node in the frequent pattern tree constructed by taking the current rule as the rule head, which covers the least positive samples in the positive sample group. The support degree corresponding to the path rule can refer to the number of positive samples covered by the path rule, or the ratio of the number of positive samples covered by the path rule to the total number of positive samples. The third part can reflect the coincidence degree to a certain extent. The more positive samples covered, the greater the coincidence degree.

[0214] The fourth cost is the support degree of the path rule from the root node to the leaf node in the frequent pattern tree constructed by taking the current rule as the rule head, which covers the least negative samples in the negative sample group. Generally, it is considered that the less negative samples covered, the better. Therefore, the more negative samples covered, the worse the accuracy, and the higher the cost.

[0215] If there are multiple rule paths in the frequent pattern tree, the path rule covering the least positive samples and the path rule covering the least negative samples are found in the third cost and the fourth cost. In the frequent pattern tree constructed by taking B=b1 as the rule head, the three path rules on the left side of the tree containing Figure 6 Assuming that among the three rule paths, the leftmost rule path (C=c1, A=a1, B=b1) covers the least positive samples, and the middle rule path (C=c3, A=a2, B=b1) covers the least negative samples, the third cost can be determined by the leftmost path rule, and the fourth cost can be determined by the middle path rule.

[0216] The advantages of the lower bound of the test number provided by the embodiments of the present application are explained and described below.

[0217] The test number lower bound is used to determine the lower bound of the test number corresponding to each path rule in the frequent pattern tree constructed with the current rule as the rule head. In the test number lower bound provided in the embodiments of the present application, the first term cost is determined by the length of the rule head, and each path rule in the frequent pattern tree will not be smaller than the length of the rule head, so the first term cost in the test number corresponding to each path rule will be greater than the first term cost in the test number lower bound. The second term cost is determined by the sum of the dual variables of the covered positive samples, and the rule head can always cover more or equal positive samples relative to each path rule in the tree, so the sum of the dual variables will be greater, and after taking the negative, a smaller cost will be obtained, thus the second term cost in the test number corresponding to each path rule will not be smaller than the second term cost in the test number lower bound. The third term cost and the fourth term cost are determined by the path rule covering the least positive samples and negative samples, thus these two costs consider the best case, for each rule path, the third term cost of the test number corresponding to the rule path will not be smaller than the third term cost obtained by covering the least positive samples of the rule path, and the fourth term cost of the test number corresponding to the rule path will not be smaller than the fourth term cost obtained by covering the least negative samples of the rule path.

[0218] In summary, the test number lower bound constructed by the above four costs will not be greater than the test number of each path rule in the actual tree, thus on the basis of measuring the complexity, coincidence degree and accuracy, an accurate test number lower bound is constructed based on the dual variable, and the efficiency and accuracy of pruning are improved.

[0219] Optionally, each cost has a corresponding penalty coefficient, and the penalty coefficient matches the coefficient in the objective function of the main problem in the column generation technique. Exemplarily, the test number lower bound can be calculated by the following formula.

[0220]

[0221] wherein S' represents all possible path rules in the frequent pattern tree constructed with the current rule as the rule head, cost S' represents the cost of the path rule S', which is equivalent to the test number, S is the rule head, λ is the dual variable, and pbits represents the BitMap of the rule in the positive samples (P), and the sum of λ in the formula represents the sum of the dual variables corresponding to the positive samples covered by the rule head, and supp(a, P / N) represents the support of the rule a in the positive samples P (or the negative samples N), and here S* + and S* - respectively represent the path from the root node to the leaf node covering the least positive samples and negative samples in the FP-tree with S as the rule head.

[0222] The lower bound of the test number constructed by the above formula, each cost has a corresponding penalty coefficient, and the penalty coefficient matches the coefficient in the objective function, and the lower bound of the test number that can be calculated is more accurate. Using the above formula can effectively determine whether a tree is worth continuing to analyze.

[0223] In other optional implementations, the lower bound of the test number can also be adjusted on the basis of the above formula, for example, the cost corresponding to the complexity or the coincidence degree is deleted, or the cost corresponding to the complexity can be the length of the rule head plus 1, so as to construct a more strict lower bound of the test number, and further improve the pruning effect.

[0224] Figure 10 A flowchart of a rule learning method provided by an embodiment of the present application is shown in FIG. 1. As shown in FIG. 1, the rule learning method can include the following three processes: data preprocessing, rule learning, and post-processing. The following will be described respectively. Figure 10

[0225] Step 1001, data preprocessing.

[0226] Optionally, the input data can be divided into two groups: a positive sample group (P) and a negative sample group (N). By default, the rules of the samples in the positive sample group are of interest. The continuous features in the positive and negative sample groups are discretized, for example, the income is a continuous feature, and the income is divided into multiple intervals, each interval corresponding to a feature. A BitMap of the covered positive and negative samples is generated for each discretized feature.

[0227] Step 1002, rule learning.

[0228] Optionally, rule learning can be achieved by repeatedly executing the following two steps: modeling and solving the restricted master problem; and performing rule search of the minimum test number.

[0229] Optionally, in the first step, a single feature is used in the first iteration to construct the restricted master problem, and in the subsequent iterations, the rule with the minimum test number (i.e., the new column) is generated and added to the restricted master problem. The construction and solution of the restricted master problem can refer to the foregoing embodiments, which will not be described here.

[0230] In each iteration, when the solution of the restricted master problem is completed, the dual variable is extracted and used to calculate the minimum test number.

[0231] In the second step, rule search can be implemented based on the BitMap FP-tree and branch and bound to find the rule corresponding to the minimum test number. Specifically, it can include the following 3a) to 3b).

[0232] ​Step 3a), initialize a variable BestScore as 0, determine a preset frequency threshold to be satisfied by the rule, and set a rule head Cond of the FP-tree as Null.

[0233] Step 3b), using the data satisfying Cond, construct a BitMap FP-tree, and determine a single-feature node set satisfying the preset frequency threshold, denoted as E. For each element e in E, evaluate a test number corresponding to a new rule Cond&e, and if the test number is better than BestScore, update BestScore.

[0234] Step 3c), for each element e in E, construct a test number lower bound LB of a new FP-tree corresponding to the element e. If the LB is greater than BestScore, skip analyzing the element e, otherwise, use depth-first recursion to construct an FP-tree with a new rule head Cond&e. Then repeat steps 3b) and 3c).

[0235] Optionally, the BitMap FP-tree can refer to an FP-tree containing BitMap information. In step 3b), the BitMap FP-tree can be constructed according to the method provided in the foregoing embodiments, and BitMap corresponding to the rule covering positive and negative samples from the root node to the current node is stored, and a corresponding node linked list is stored for convenience of calculation.

[0236] The modeling and solving of the restricted main problem, and the rule search of the minimum test number are repeatedly performed until no rule with a test number less than 0 can be found, and a rule learning model can be obtained.

[0237] Step 1003, post-processing.

[0238] In the obtained rule learning model, a rule can contain multiple features. In order to improve the interpretability of the rule, post-processing operation can be further performed. For example, a user can want to know which features in a rule are the most important, and the post-processing operation can sort the features in the rule and count the most important features.

[0239] In one or more embodiments of the present application, optionally, after obtaining the rule learning model, post-processing operation can be performed on any rule containing multiple features in the rule learning model to determine the importance of at least one feature in the rule; wherein the importance is used to show the user the core reason why the business object to be analyzed meets the requirements.

[0240] For example, if a business object to be analyzed corresponds to the features including: number of defaults = 10, income = 1, flow = 1, and asset without mortgage, the prediction result obtained by inputting the features into the trained rule learning model is that the approval is not passed. On this basis, the user can want to know the core reason for the approval not being passed. Therefore, the importance of each feature included in the rule can be counted, and one or more most important features can be the core reason. For example, the core reason for the approval not being passed is that the number of defaults = 10, and the features of income = 1, flow = 1, and asset without mortgage have little to do with the approval not being passed. Therefore, the user can be shown that the core reason for the different approval is that the number of defaults is too many, thereby improving the interpretability of the rule learning model.

[0241] In one or more embodiments of the present application, optionally, determining the importance of at least one feature in the rule comprises: for each feature in the rule, calculating a first difference value between the proportion of positive samples covered by the feature and the proportion of negative samples covered by the feature; calculating a second difference value between the proportion of positive samples covered by the rule and the proportion of negative samples covered by the rule; calculating the importance of each feature according to the second difference value corresponding to the rule and the first difference value corresponding to each feature in the rule; wherein the feature with an importance satisfying a preset condition is the core reason for the business object meeting the requirements of the rule.

[0242] Suppose the rule includes multiple features a1, b1, and c1. For feature a1, the proportion of positive samples covered by the feature a1 and the proportion of negative samples covered by the feature a1 are calculated, wherein the proportion of positive samples covered by the feature a1 is the ratio of the number of positive samples covered by the feature a1 to the total number of positive samples, and similarly, the proportion of negative samples covered by the feature a1 is the ratio of the number of negative samples covered by the feature a1 to the total number of negative samples. The difference between the two ratios is calculated to obtain the first difference value corresponding to the feature a1.

[0243] Similarly, the first difference value corresponding to feature b1 and the first difference value corresponding to feature c1 can be obtained, and the second difference value corresponding to the rule can also be calculated. According to the first difference values corresponding to a1, b2, and c1 and the second difference value corresponding to the rule, the importance of each feature can be calculated.

[0244] Optionally, for each feature, if the first difference value corresponding to the feature is greater than the product of the second difference value and a proportion coefficient, the feature is an important feature, otherwise it is an unimportant feature, and the core reason can be the important feature. Optionally, the proportion coefficient can be 0.5.

[0245] Exemplarily, in the various features included in the construction rule, the proportion of positive samples covered by a certain feature is 8% higher than the proportion of negative samples covered, and the proportion of positive samples covered by the entire rule is only 10% higher than the proportion of negative samples covered, and it can be considered that the feature is sufficient to make a judgment, and when the reasons corresponding to the prediction result are displayed, only the feature or the feature can be displayed, so that the user can quickly understand the core reasons corresponding to the prediction result, and the user experience is improved.

[0246] Optionally, when using the rule learning model to predict the business object, the feature is matched with the rule in the rule learning model, and whether the business object to be analyzed meets the requirement and / or the reason for meeting the requirement is determined according to the matching result, which can include: if there is a rule in the rule learning model that matches the feature corresponding to the business object to be analyzed, it is determined that the business object to be analyzed meets the requirement, and the reason why the business object meets the requirement is the matched rule; according to the proportion of positive and negative samples covered by each feature in the matched rule and the proportion of positive and negative samples covered by the rule, the core reason why the business object to be analyzed meets the requirement is determined; and at least one of the following is displayed: whether the business object to be analyzed meets the requirement, the reason why the requirement is met, and the core reason why the requirement is met.

[0247] Exemplarily, any one of whether the requirement is met, the reason why the requirement is met, and the core reason why the requirement is met can be displayed, and multiple items can also be displayed.

[0248] When predicting, the core reason can be determined in a manner similar to the aforementioned post-processing. By displaying the core reason, the interpretability of the model can be increased, the user can focus on the core reason for meeting the requirement, and the user experience can be further improved.

[0249] The embodiments of the present application also provide a rule learning method, which includes: obtaining a training sample set, the training sample set including a positive sample group and a negative sample group, wherein the positive sample includes a feature corresponding to a business object meeting a requirement, and the negative sample includes a feature corresponding to a business object not meeting the requirement; generating a rule learning model based on column generation technology and / or a frequent pattern tree according to the training sample set; and the generated rule learning model is used to predict whether a business object to be analyzed meets the requirement.

[0250] In the first optional implementation, the rule learning model can be generated based on column generation technology. Specifically, the initial rules can be generated, the restricted master problem can be constructed and solved, and then the dual variables obtained by solving can be used to find new rules. Even if the frequent pattern tree is not used, all rules can be traversed to find the rule with the smallest test number as the new rule added to the rule pool for further iteration until no rule with a test number less than 0 can be found.

[0251] Optionally, in searching for new rules, various possible rules can be screened first, for example, features with an occurrence frequency less than a preset frequency threshold are removed, which can effectively reduce the amount of calculation.

[0252] Optionally, the entire problem can also be formalized as an integer programming, and a rule learning model is generated by directly solving the column generation technique of the pricing problem.

[0253] In the second optional implementation, a rule learning model can be generated based on a frequent pattern tree. Optionally, a frequent pattern tree can be directly constructed, frequent item sets are found from the frequent pattern tree, rules with a frequency exceeding a certain requirement are generated, and then the number of positive samples covered by each rule and the number of negative samples covered by each rule are determined to determine whether to adopt the rule, for example, if the number of positive samples covered exceeds the number of negative samples covered, the rule can be adopted, otherwise it is not adopted.

[0254] In the third optional implementation, generating a rule learning model based on the column generation technique and / or the frequent pattern tree according to the training sample set can include: generating a plurality of initial rules based on the training sample set, each rule including at least one feature; continuously adding new rules based on the column generation technique until a rule learning model is obtained; and wherein, in the column generation process, new rules are generated based on the frequent pattern tree. This implementation is consistent with the foregoing embodiments, and specific reference can be made to the foregoing embodiments.

[0255] The third implementation has the following advantages over the first two implementations which only use the column generation technique or only use the frequent pattern tree to generate a rule learning model: compared with the first implementation, the present scheme combines the frequent pattern tree for analysis and processing, improves the efficiency, and can consider the rule overlap degree, etc., and can output rules with less overlap, which is more advantageous in further improving the explainability. At the same time, the present scheme does not directly use an integer programming solver to process the pricing problem, and thus has stronger ability to process big data; compared with the second implementation, the present scheme uses the column generation technique to search for rules that only have potential help for the objective function according to a unified optimization target, and combines the pruning lower bound derived for the test number to effectively process the dimension explosion of the data amount and thus improve the ability to process big data. Moreover, the present scheme can only analyze rules for describing and distinguishing positive samples, and does not learn rules for describing negative samples, thus having obvious efficiency improvement when the positive and negative samples are imbalanced. At the same time, by using an FP-tree containing rule BitMap information, the rule intersection / union processing and data filtering are very efficient.

[0256] In addition, the present scheme also has strong advantages over other rule learning methods.

[0257] An alternative rule learning method is a two-stage rule generation and rule selection scheme. The rule generation mainly relies on the mature association rule mining techniques or by training a tree-based classifier and extracting the root-to-leaf paths as rules. The rule selection mainly depends on optimizing specific criteria, such as balancing accuracy and sparsity. First, this method lacks consideration of global optimization objectives, and the rule generation stage is usually weakly connected to the final rule set, resulting in inaccurate or highly overlapping rules. For example, some important rules can be incorrectly filtered out in the first stage. Second, even if only one well-designed optimization objective is used to select rules, the optimization result is still poor when the generated rules are numerous. The present scheme has a powerful and efficient data processing capability compared with the above method.

[0258] Another alternative rule learning method is a diversity rule learning algorithm based on random sampling and approximation. This kind of algorithm mainly uses random sampling and rule extraction for learning, but due to its reliance on greedy algorithms for approximation and the uncertainty of random algorithms, its actual effect and ability to process large-scale data are very limited. Compared with the above method, the present scheme does not use a random approach to output heuristic rules, but relies on branch-and-bound exact solution, and thus has more advantages in result accuracy.

[0259] Yet another alternative rule learning method is a rule list and optimal decision tree model. This kind of algorithm naturally has the advantage of low overlap between rules, but due to the need to consider the order of rules, the optimization difficulty of this kind of algorithm is very large, and the data size that can be processed is very limited. And when understanding in the form of list and tree, all previous logical rule conditions need to be considered, and the interpretability is poor. Compared with the above method, the rules output by the present scheme do not have a tree structure or an order, and thus are more convenient for users to understand.

[0260] In summary, the present scheme is the first rule generation system that can efficiently process large-scale data, output accurate, diverse and concise rules. Combining column generation technology and rule generation algorithm based on FP-tree data structure, the accuracy, complexity and overlap of the learned rules can be balanced at the same time. On the premise of ensuring high accuracy, the interpretability of the results is further improved. Moreover, the branch-and-bound algorithm developed for FP-tree innovatively theoretically proposes to bound the rules contained in the FP-tree, thereby accelerating the pruning efficiency and improving the ability to process large-scale data.

[0261] The following describes the application of the rule learning method provided by the embodiments of the present application in actual application by means of several examples.

[0262] Figure 11A flowchart of a power load prediction method provided by an embodiment of the present application is shown in FIG. 1. Figure 11 As shown in FIG. 1, the method can include:

[0263] Step 1101: obtaining an environment feature to be analyzed.

[0264] The environment feature to be analyzed can refer to an environment feature corresponding to a time node and / or a region to be analyzed.

[0265] Step 1102: predicting, according to the environment feature and a rule learning model, whether a power load corresponding to the environment feature meets a requirement, and / or outputting a reason for determining a power load prediction result.

[0266] The rule learning model is a rule learning model generated based on a column generation technique and / or a frequent pattern tree.

[0267] Optionally, generating the rule learning model based on the column generation technique and / or the frequent pattern tree can include: obtaining a training sample set, the training sample set including a positive sample group and a negative sample group, wherein the positive sample includes an environment feature corresponding to a power load meeting a requirement, and the negative sample includes an environment feature corresponding to a power load not meeting a requirement; generating the rule learning model based on the column generation technique and / or the frequent pattern tree according to the training sample set; and the generated rule learning model is used to predict whether a power load corresponding to an environment feature to be analyzed meets a requirement.

[0268] Based on the technical solutions provided in the foregoing embodiments, the business object can be set as a time node or a region to be analyzed, the corresponding feature can be an environment feature, and meeting the requirement can specifically refer to a power load meeting a requirement, thereby generating a rule learning model capable of predicting whether a power load meets a requirement. Not only can the prediction result be given, but also the corresponding reason can be given, thereby improving user experience and improving the overall efficiency and accuracy of power load prediction based on the column generation technique and the frequent pattern tree.

[0269] Optionally, in a power industry busbar load prediction scenario, the environment feature can include at least one of the following: temperature, humidity, apparent temperature, abnormal information, batch modification information, and load transfer information. The corresponding power load meeting the requirement can be that the corresponding power load is greater than a load threshold value, and / or the gap between the corresponding power load and a historical same-period power load is greater than a gap threshold value.

[0270] The time node can be an hour, a day, a month, a year, etc. The abnormal information can refer to environmental protection factors causing a change in power load, and the batch modification information can refer to batch control or modification of the power load by a load prediction personnel.

[0271] In an optional implementation, the environmental feature to be analyzed is an environmental feature corresponding to a time node to be analyzed.

[0272] Correspondingly, in model learning, the positive sample includes an environmental feature corresponding to a time node in which the power load meets the requirement, and the negative sample includes an environmental feature corresponding to a time node in which the power load does not meet the requirement; and the generated rule learning model is used to predict whether the power load of the time node to be analyzed meets the requirement. The time node in which the power load meets the requirement can be a time node in which the power load is greater than a load threshold and / or a time node in which a gap of the power load relative to a historical same period power load is greater than a gap threshold.

[0273] In an example, the time node in which the power load meets the requirement is a time node in which the power load is greater than the load threshold, and the time node in which the power load does not meet the requirement can be a time node in which the power load is less than or equal to the load threshold.

[0274] For example, the power load and the environmental feature of each hour can be collected, and divided into positive and negative sample groups according to the load threshold, and the rule learning generation model can be used to predict whether the power load is greater than the load threshold according to the environmental feature of each subsequent hour.

[0275] In another example, the time node in which the power load meets the requirement is a time node in which a gap of the power load relative to a historical same period power load is greater than a gap threshold.

[0276] For example, the power load gap of each day relative to the same day of the previous month can be collected, and divided into positive and negative sample groups according to the gap threshold, and the rule learning generation model can be used to predict whether the power load of a certain day relative to the same day of the previous month is greater than the gap threshold.

[0277] In yet another example, the time node in which the power load meets the requirement can be a time node in which the power load is greater than a load threshold and a gap of the power load relative to a historical same period power load is greater than a gap threshold.

[0278] For example, only the date in which the load of the same day is large and the gap relative to the historical same period is also large is divided into the positive sample, and the model generated by rule learning can be used to predict whether the condition is met.

[0279] Through at least one of the above environmental information, it can be predicted whether the power load of a certain time node exceeds a threshold or whether a fluctuation of the power load relative to a historical same period power load exceeds a threshold, and a specific reason is given, thereby helping a load prediction personnel to establish trust in a predicted value and reporting to a dispatching department, thereby providing a reference for power generation planning, power dispatching, and electricity pricing.

[0280] Optionally, in predicting whether the power load will have a large fluctuation compared with the historical same period, the training sample set can be obtained, which can include: obtaining data corresponding to a plurality of dates, the data corresponding to the dates including the environmental characteristics corresponding to the dates, the power load and the environmental characteristics of the historical same period, the power load; dividing a positive sample group and a negative sample group according to the data corresponding to the plurality of dates, wherein if the difference between the power load of a certain date and the historical same period is greater than a difference threshold, the difference between the environmental characteristics of the certain date and the historical same period is a positive sample, otherwise the difference between the environmental characteristics of the certain date and the historical same period is a negative sample; wherein the generated rule learning model is used to predict whether the power load of the date to be analyzed will change more than the difference threshold compared with the historical same period according to the difference between the environmental characteristics of the date to be analyzed and the historical same period. Optionally, the difference can be a difference value or a ratio. Wherein the historical same period of different dates can be different. Optionally, the historical same period of a certain date can be N days before the date, and N is a positive integer. For example, for July 1, the historical same period is June 1. For July 15, the historical same period can be June 15.

[0281] In actual application, the environmental characteristics and the power load of each day and the historical same period can be obtained, and the difference value of the environmental characteristics and the difference value of the power load are calculated, and the positive and negative samples are divided according to the difference value of the power load, and the environmental characteristics are put into the sample group.

[0282] Specifically, the present scheme calculates the difference between the predicted value and the comparison value of the historical same period under all dimension combinations. In order to help the load prediction personnel understand which dimension combination the difference between the predicted value and the comparison value is significant, the present scheme first groups the data after the difference into positive and negative samples, and defines the group with significant difference as a positive sample.

[0283] Exemplarily, the positive sample group can include: positive sample 1: temperature difference = 10, humidity difference = 20; positive sample 2: temperature difference = 15, humidity difference = 15; …… According to the generated rule learning model, whether the power load of the date to be analyzed will have a large difference compared with the historical same period can be predicted.

[0284] Next, the rule learning model can accept user configuration parameters to balance the accuracy of the output rule set covering result, the complexity and coincidence of the rules, so as to automatically analyze and output a rule set for explaining the positive sample (the group with significant difference). According to the returned result, the prediction personnel can quickly perceive the dimension combination with significant difference, so as to check and adjust the prediction model and the prediction result.

[0285] In another optional implementation manner, the environmental characteristics to be analyzed can be the environmental characteristics corresponding to the region to be analyzed.

[0286] Correspondingly, data corresponding to a plurality of regions can be acquired, the data corresponding to the region including environmental characteristics of the region at a first time combination and power load corresponding to the first time combination; wherein the first time combination includes at least two time nodes; according to the data corresponding to the plurality of regions, a positive sample group and a negative sample group are divided, wherein if a difference between the power load corresponding to the at least two time nodes is greater than a difference threshold, then a difference between the environmental characteristics of the region corresponding to the at least two time nodes is a positive sample, otherwise, the difference between the environmental characteristics of the region corresponding to the at least two time nodes is a negative sample; according to the positive sample group and the negative sample group, a rule learning model is generated based on column generation technology and / or a frequent pattern tree.

[0287] The generated rule learning model is used to predict whether the power load of the region to be analyzed at a second time combination exceeds the difference threshold according to a difference between environmental characteristics of the region to be analyzed at the second time combination; the first time combination and the second time combination can have at least one different time node.

[0288] Taking the first time combination including two time nodes as an example, assuming that the first time combination includes July 1 and June 1, the data corresponding to each region includes environmental characteristics and power load of the region on July 1 and environmental characteristics and power load of the region on June 1.

[0289] Through the data corresponding to a plurality of regions, a positive sample group and a negative sample group can be constructed, if the difference between the power load of a region on July 1 and the power load of the region on June 1 is greater than the difference threshold, then the requirement is met, and the difference between the environmental characteristics of the region on July 1 and the environmental characteristics of the region on June 1 can be used as a positive sample, otherwise, the requirement is not met, and it is used as a negative sample.

[0290] According to the constructed positive sample group and negative sample group, a rule learning model can be generated based on column generation technology and / or a frequent pattern tree. When predicting a region, the model can be used to predict whether the power load on August 1 exceeds the difference threshold relative to the power load on July 1 according to the difference between the environmental characteristics of the region on August 1 and July 1.

[0291] If the first time combination includes three or more time nodes, the difference between two nodes can be calculated to obtain a plurality of differences, and correspondingly, the difference greater than the difference threshold is divided into a positive sample, which can specifically mean that any one of the differences is greater than the difference threshold, or all the differences are greater than the difference threshold, and then the positive sample is divided.

[0292] By using regions as the finest granularity for model learning and prediction of power load, data from various regions can be integrated to guide the prediction of power load in the region under analysis, thereby increasing the breadth of the sample and meeting the prediction needs of power load in different regions.

[0293] Figure 12 This is a schematic diagram illustrating a power load forecasting result provided in an embodiment of this application. Figure 12 As shown, the display management module allows users to customize the historical dates to be compared. Users can select one or more historical dates, and the module displays the power load changes of the date to be analyzed relative to each historical date. The middle line chart shows the power load changes over time. Below, the specific power load changes on August 1st compared to the historical same period on July 1st can be displayed, along with explanations of the differences in predicted values.

[0294] Specifically, load forecasters can compare the predicted curves given by the algorithm with historical curves of similar days. Through the rule learning model, the differences between the two curves can be automatically analyzed and explained from the dimensions of temperature, humidity, and perceived temperature.

[0295] By classifying positive and negative samples based on whether the difference between the power load on a given date and the historical same period is greater than a threshold, and then performing rule learning, we can more accurately use environmental differences to predict power load differences and improve the accuracy of predictions.

[0296] Optionally, the positive samples covered by the rules can be determined based on the difference between the environmental characteristics of the region to be analyzed on the date of analysis and the same period in history. Based on the positive samples covered by the rules, the power load of the region to be analyzed on the date of analysis, and / or the difference between the power load of the region to be analyzed on the date of analysis and the same period in history.

[0297] For example, assuming the matching rule covers 10 positive samples, the average difference between these 10 positive samples and the historical power load for the same period can be calculated, and the average of the groups can be used to reflect the power load change corresponding to this rule. After obtaining the average, it can be displayed as the difference between the power load of the date to be analyzed and the historical power load for the same period, or the average can be added to the historical power load for the date to be analyzed to obtain the predicted power load value for the date to be analyzed and then displayed.

[0298] By using the mean of the power load gap of the positive sample corresponding to the rule in the matching to reflect the corresponding power load change, the power load of the region to be analyzed on the date to be analyzed can be effectively predicted according to the similar historical data, a quantitative analysis result is provided for the power load prediction, and the user can understand the specific degree of the power load change, thereby providing a more accurate reference for power dispatching.

[0299] Optionally, the following at least one of the load prediction personnel input can also be obtained: a time node to be analyzed, a reference time node compared with the time to be analyzed, a region to be analyzed; and correspondingly, whether the power load corresponding to the environmental feature meets the requirement can be predicted according to the environmental feature of the region to be analyzed at the time node to be analyzed and the reference time node, which can include: determining whether the power load change of the time node to be analyzed relative to the reference time node meets the requirement according to the gap of the environmental feature of the region to be analyzed at the time node to be analyzed and the reference time node.

[0300] Optionally, the time node can be a date, and whether the power load change of the date to be analyzed relative to one or more reference dates meets the requirement can be determined by the rule learning model.

[0301] Exemplarily, the date to be analyzed can be August 1, the user can input the reference dates July 1 and July 15, and whether the power load change of August 1 relative to July 1 meets the requirement can be calculated by the rule learning model, and whether the power load change of August 1 relative to July 15 meets the requirement can also be calculated.

[0302] One or more reference dates can be input by the load prediction personnel, the date to be analyzed is the current day by default, or the load prediction personnel can input the date to be analyzed, the reference dates are the historical same period by default, or the date to be analyzed and the reference dates can be input by the load prediction personnel. In addition, the region to be analyzed can also be input by the user, so that the date and the region can be selected according to the demand of the load prediction personnel, and the power load prediction demand in different scenes can be met.

[0303] The embodiment of the application provides a rule learning method applied to business index abnormality prediction. Optionally, the method comprises: obtaining a training sample set, the training sample set comprising a positive sample group and a negative sample group, wherein the positive sample comprises an operating feature when a business index is abnormal, and the negative sample comprises an operating feature when the business index is normal; generating a rule learning model based on column generation technology and / or a frequent pattern tree according to the training sample set; and the generated rule learning model is used to predict whether a business index corresponding to an operating feature to be analyzed is abnormal.

[0304] Optionally, the business feature can be a business feature corresponding to a time node or a region. Specifically, the business feature when the business index is abnormal can be a business feature corresponding to a time node or a region when the business index is abnormal, and the business feature to be analyzed can be a business feature of a time node or a region to be analyzed.

[0305] On the basis of the technical solutions provided in the foregoing embodiments, the business object can be a time node or a region to be analyzed, the corresponding feature can be a business feature, and the requirement can be that the business index is abnormal, so that a rule learning model capable of predicting whether the business index is abnormal is output. Not only can the prediction result be given, but also the corresponding reason can be given, the user experience is improved, and the overall efficiency and accuracy of the business index abnormality prediction are also improved based on the column generation technique and the frequent pattern tree.

[0306] Optionally, the business feature includes at least one of the following: an industry, a target user group, a price, and commodity inventory information; the business index is a total sales transaction amount (GMV); and the business index abnormality is that the total sales transaction amount is less than a total amount threshold value and / or a difference between the total sales transaction amount and a total sales transaction amount of a historical same period is less than a difference threshold value.

[0307] In prediction of whether the difference is less than the difference threshold value, the rule learning method can refer to a rule learning method applied to power load prediction, and data related to power load prediction in the rule learning method is replaced by data related to business index abnormality prediction.

[0308] Figure 13 A schematic diagram of a business index abnormality prediction result provided by an embodiment of the present application is shown in FIG. 1. Figure 13 As shown in FIG. 1, a curve diagram shows the GMV change over time, and an abnormal score and root cause drilling can also be given.

[0309] Optionally, the abnormal score and the root cause drilling can be determined by a positive sample corresponding to a rule matched by the business feature to be analyzed in the rule learning model. The abnormal score can be used to represent the abnormality degree, and can be determined according to the abnormality degree of the matched positive sample. The root cause drilling can be used to represent the proportion of the reason for the abnormality, and can be determined by calculating the importance of each feature according to the matched positive sample, so as to reflect the core factor causing the index abnormality.

[0310] In the business operation process, the abnormal change of the key indicators such as GMV has an important influence on the health of the business operation, so the business usually monitors the key indicators, such as the abnormal change, and hopes to locate the cause as soon as possible and take action to resolve the risk. Through the above method, whether the key indicators of the enterprise are less than expected or whether there will be a larger fluctuation can be predicted according to the operating characteristics of the time node or region to be analyzed. When the key indicators of the enterprise appear abnormal, the key factors causing the abnormal indicators can be automatically found through the rule learning model, and the analysis and decision-making personnel are given, so as to improve the efficiency of analysis and decision-making, and make more scientific and reasonable decisions.

[0311] The embodiment of the present application provides a business indicator abnormality prediction method. The method can include: obtaining an operating characteristic to be analyzed; predicting whether a business indicator corresponding to the operating characteristic to be analyzed is abnormal and / or determining the cause of the abnormal or normal business indicator according to the operating characteristic and a rule learning model, wherein the rule learning model is generated by the rule learning method of any of the preceding embodiments. The specific implementation principle and effects of the business indicator abnormality prediction method can be referred to the preceding embodiments, which will not be repeated here.

[0312] The embodiment of the present application provides a rule learning method applied to recommendation information acceptance prediction. The method can include: obtaining a training sample set, the training sample set including a positive sample group and a negative sample group, wherein the positive sample includes the features corresponding to the users who accept the recommendation information, and the negative sample includes the features corresponding to the users who do not accept the recommendation information; generating a rule learning model based on column generation technology and / or frequent pattern tree according to the training sample set; and the generated rule learning model is used to predict whether the user to be analyzed will accept the recommendation information or the acceptance degree of the user to be analyzed to the recommendation information.

[0313] The acceptance degree can include acceptance and non-acceptance, or the acceptance degree can include a percentage of acceptance, such as an acceptance degree of 50%.

[0314] On the basis of the technical solutions provided in the preceding embodiments, the business object can be set as the user to be analyzed, which meets the requirement of accepting the recommendation information, so as to output the rule learning model capable of predicting whether the user will accept the recommendation information. Not only the prediction result can be given, but also the corresponding cause can be given, which improves the user experience and improves the overall efficiency and accuracy of the recommendation information acceptance prediction based on the column generation technology and the frequent pattern tree.

[0315] Optionally, the features of the user include at least one of the following: attribute information of the user such as age, occupation, income, etc., historical browsing records, historical purchase records, preference information, the number of times that the user historically accepted or did not accept recommended information, reasons for the user historically feeding back acceptance or non-acceptance of recommended information; and the recommended information is commodity information and / or advertisement information recommended to the user, so as to predict whether the user will accept the recommended commodity information and advertisement information.

[0316] Figure 14 A schematic diagram of recommended information is provided for an embodiment of the present application. As shown in Figure 14 recommended commodity information can be shown to the user, and if the user does not accept the recommended commodity information, reasons for non-acceptance can be provided for the user to select. In the rule learning process, the number of times that the user historically accepted or did not accept recommended information, reasons for the user historically feeding back acceptance or non-acceptance of recommended information, etc. can be learned as features.

[0317] Through the above method, commodity push and advertisements can be targetedly pushed to users who have acceptance intention, and the needs of different users can be met. In actual application, a rule learning model can be introduced into a personalized recommendation module of various e-commerce and content APPs, a recommended reason function can be constructed, and therefore the recommended reason of each recommended result can be more finely output, so as to solve the confusion of users about the recommended result and the distrust of users for the recommendation system, and thus the trust degree and the satisfaction degree of the recommendation system can be improved.

[0318] An embodiment of the present application provides a recommended information acceptance prediction method. The method can include: obtaining features of a user to be analyzed; predicting whether the user to be analyzed accepts recommended information and / or determining reasons for the user to be analyzed to accept or not accept recommended information according to the features of the user to be analyzed and a rule learning model; and the rule learning model is generated by the method of any of the preceding embodiments. The specific implementation principle and effect of the recommended information acceptance prediction method can be referred to the preceding embodiments, and will not be described here.

[0319] An embodiment of the present application provides a rule learning method applied to approval prediction. The method can include: obtaining a training sample set, the training sample set including a positive sample group and a negative sample group, wherein the positive sample includes features corresponding to a user whose approval is not passed, and the negative sample includes features corresponding to a user whose approval is passed; generating a rule learning model based on column generation technology and / or a frequent pattern tree according to the training sample set; and the generated rule learning model is used to predict whether a user to be analyzed passes the approval.

[0320] Optionally, the features of the user include at least one of the following: the number of defaults in a preset period, asset information, income information, and flow information.

[0321] In the field of financial anti-fraud, in a financial credit background, a credit personnel can check information of a loan applicant and a credit automatic approval result. After the result, there is an explanation of the approval reason, which can find the reason of the approval result from various information of the application client, and automatically generate a human understandable reason output to the system. For details, see Figure 4 .

[0322] The embodiment of the present application provides an approval prediction method. The method can include: obtaining the characteristics of the user to be analyzed; predicting whether the user to be analyzed is approved according to the characteristics of the user to be analyzed and the rule learning model, and / or determining the reason why the user to be analyzed is approved or not approved; wherein the rule learning model is generated by the method of any of the preceding embodiments. The specific implementation principle and effect of the approval prediction method can be referred to the preceding embodiments, which will not be repeated here.

[0323] The embodiment of the present application provides a rule learning method applied to maintenance demand prediction. The method can include: obtaining a training sample set, the training sample set including a positive sample group and a negative sample group; wherein the positive sample includes the characteristics corresponding to the user with maintenance demand, and the negative sample includes the characteristics corresponding to the user without maintenance demand; generating a rule learning model based on column generation technology and / or frequent pattern tree according to the training sample set; and the generated rule learning model is used to predict whether the user to be analyzed has maintenance demand, and the maintenance demand is used for the maintenance personnel to determine whether to maintain the user.

[0324] On the basis of the technical solutions provided in the preceding embodiments, the business object can be set as the user to be analyzed, which meets the requirement of having maintenance demand, thereby outputting the rule learning model capable of predicting whether the user has maintenance demand. Not only the prediction result can be given, but also the corresponding reason can be given, which improves the user experience, and based on the column generation technology and the frequent pattern tree, the overall efficiency and accuracy of the customer maintenance demand prediction are also improved.

[0325] Optionally, the embodiment can be applied to customer acquisition and retention analysis. In one example, the positive sample includes the characteristics corresponding to the converted user, the negative sample includes the characteristics corresponding to the unconverted user, and the generated rule learning model is used to predict whether the user to be analyzed will be converted. In another example, the positive sample includes the characteristics corresponding to the retained user, the negative sample includes the characteristics corresponding to the lost user, and the generated rule learning model is used to predict whether the user to be analyzed will be retained. Wherein, the converted user or the retained user is the user with maintenance demand.

[0326] Specifically, rule learning can be performed on a target customer group (e.g., clicking a marketing advertisement and successfully converting) and a reference group (e.g., not converting) to find rules that can accurately describe the target group, facilitate business personnel to understand the attributes of this type of customer, and thus provide a basis for better marketing reach. Similarly, rule learning can be performed on retained customers and lost customers to locate the significant reasons for easy loss of customers, thus providing insight analysis for retaining customers and improving loyalty.

[0327] The embodiment of the present application provides a maintenance demand prediction method. The method can include: obtaining the characteristics of a user to be analyzed; predicting whether the user to be analyzed has a maintenance demand according to the characteristics of the user to be analyzed and a rule learning model, and / or determining the reason why the user to be analyzed has or does not have a maintenance demand; wherein the rule learning model is generated by the method described in any of the preceding embodiments. The specific implementation principle and effect of the maintenance demand prediction method can be referred to the preceding embodiments, which will not be described here again.

[0328] Through the above method, the rule learning model can be applied in various aspects of actual production and life. It can not only identify high-risk groups from multi-dimensional data and automatically mine the clustering groups in the high-risk groups for identification and attack, but also can perform drilling analysis on the changes of aggregated KPI indicators to find the most explanatory reason combination of the changes, and can also realize user retention analysis, mine the main reasons for user loss, and portrait user groups, which has good performance in various industries.

[0329] In actual application, when using the scheme provided by the embodiment of the present application, the user can customize the related parameters of the rule learning model to output results meeting the user's individual needs. Specifically, the present application aims to output a set of accurate, concise and diverse rules about events that the user is interested in. The user can control the balance between the three competing targets according to his own preferences. The custom parameters involved can include the support of the rule coverage group, the complexity cost of the rule, the overlap cost of the rule, the number of rules, and the penalty for misclassification of positive and negative samples. For example, by increasing the support threshold of the rule coverage group, the size of the group involved in the rules received by the user will increase, and the time of rule search will decrease. For another example, by increasing the cost of rule overlap, the diversity of the interpretable rules output will increase, and the degree of overlap between rules will decrease. In addition, the set of efficient search strategies proposed by the present application will also change in running time according to the parameters input by the user. For example, by increasing the support of the rule coverage group and the pruning intensity parameter, the number of search trees with weak relevance is reduced to speed up the output of the results.

[0330] The embodiment of the present application provides a business object prediction device, which comprises:

[0331] The first obtaining module is configured to obtain a feature corresponding to a service object to be analyzed.

[0332] The first processing module is configured to match the feature with rules in a rule learning model, and determine whether the service object to be analyzed meets a requirement and / or a reason for meeting the requirement according to a matching result, wherein the rule learning model is a rule learning model generated based on column generation technology and a frequent pattern tree.

[0333] The embodiment of the present application further provides an electric power load prediction device, which comprises:

[0334] The second obtaining module is configured to obtain an environmental feature to be analyzed.

[0335] The second processing module is configured to predict whether an electric power load corresponding to the environmental feature meets a requirement and / or output a reason for determining an electric power load prediction result according to the environmental feature and a rule learning model, wherein the rule learning model is a rule learning model generated based on column generation technology and / or a frequent pattern tree.

[0336] The device provided by the embodiment can be used to execute the method provided by any one of the preceding embodiments. Figures 1 to 14 The technical solutions provided by the embodiments shown in the drawings have similar implementation principles and technical effects, and thus will not be described here.

[0337] Figure 15 A structural schematic diagram of an electronic device provided by the embodiment of the present application is shown in FIG. 1. Figure 15 The electronic device of the embodiment can comprise:

[0338] at least one processor 1501; and

[0339] a memory 1502 in communication connection with the at least one processor;

[0340] The memory 1502 stores instructions executable by the at least one processor 1501, and the instructions are executed by the at least one processor 1501 to enable the electronic device to execute the method described in any one of the preceding embodiments.

[0341] Optionally, the memory 1502 can be independent or integrated with the processor 1501.

[0342] The implementation principles and technical effects of the electronic device provided by the embodiment can be referred to the preceding embodiments, and thus will not be described here.

[0343] The embodiment of the present application further provides a computer readable storage medium, which stores computer execution instructions, and when a processor executes the computer execution instructions, the method described in any one of the preceding embodiments is implemented.

[0344] The embodiment of the present application further provides a computer program product comprising a computer program which, when executed by a processor, implements the method of any of the preceding embodiments.

[0345] In several embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other manners. For example, the above-described device embodiments are merely illustrative, and the division of the modules is merely a logical function division, and there can be another division manner in actual implementation. For example, a plurality of modules or features can be combined or integrated into another system, or some features can be ignored or not executed.

[0346] The integrated modules implemented in the form of software function modules can be stored in a computer readable storage medium. The software function modules stored in the storage medium can include a number of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) or a processor to perform some of the steps of the methods described in the various embodiments of the present application.

[0347] It should be understood that the processor can be a central processing unit (CPU), or can also be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), or the like. The general-purpose processor can be a microprocessor, or the processor can also be any conventional processor. The steps of the methods disclosed in the present application can be directly embodied as hardware processor execution, or a combination of hardware and software modules in the processor to execute. The memory can include a high-speed RAM memory, and can also include a non-volatile storage NVM, such as at least one disk memory, and can also be a U disk, a mobile hard disk, a read-only memory, a magnetic disk or an optical disk, etc.

[0348] The storage medium can be implemented by any type of volatile or non-volatile storage devices or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk or optical disk. The storage medium can be any available medium that can be accessed by a general-purpose or special-purpose computer.

[0349] An exemplary storage medium is coupled to the processor such that the processor can read information from, and can write information to, the storage medium. Of course, the storage medium can be a part of the processor. Consistent with the teachings provided herein, the processor can execute instructions embodied by computer code stored on the storage medium, and the storage medium can store computer code which, when executed by the processor, causes the processor to perform methods as described herein. The processor and the storage medium can be supplemented by, or incorporated in, ASICs (application- specific integrated circuits).

[0350] It should be noted that, as used in this document, the terms "includes" and / or "containing" or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements is not limited to those elements, but can include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element preceded by "comprises a" does not, without more constraints, foreclose the existence of additional identical elements in the process, method, article, or apparatus that includes the element.

[0351] The above-mentioned sequence numbers of the embodiments of the present application are only for description, and do not represent the advantages and disadvantages of the embodiments.

[0352] From the above description of the embodiments, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be realized by means of software and necessary general hardware platforms, and of course can also be realized by hardware, but in many cases the former is a better embodiment. Based on such understanding, the technical solutions of the present application can be embodied in the form of a software product, which is stored in a storage medium (such as a ROM / RAM, a magnetic disk, or an optical disk), and includes a plurality of instructions for causing a terminal device (which can be a mobile phone, a computer, a server, an air conditioner, or a network device) to execute the methods described in the various embodiments of the present application.

[0353] The above is only the preferred embodiment of the present application, and does not limit the patent scope of the present application, and any equivalent structure or equivalent process transformation, or direct or indirect application in other related technical fields, is also included in the patent protection scope of the present application.

Claims

1. A method for predicting a business object, applied to the field of power load prediction, characterized in that, The method comprises: obtaining features corresponding to a business object to be analyzed; matching the features with rules in a rule learning model, and determining whether the business object to be analyzed meets the requirements and / or the reason for meeting the requirements according to a matching result; wherein the rule learning model is a rule learning model generated based on column generation technology and a frequent pattern tree; The method further comprises: obtaining a training sample set, the training sample set comprising a positive sample group and a negative sample group, wherein the positive samples comprise features corresponding to business objects meeting the requirements, and the negative samples comprise features corresponding to business objects not meeting the requirements; generating a plurality of initial rules according to the training sample set, each rule comprising at least one feature; continuously adding new rules based on column generation technology until a rule learning model is obtained; wherein during column generation, new rules are generated based on a frequent pattern tree; continuously adding new rules based on column generation technology until a rule learning model is obtained, comprising: constructing a constraint master problem, the constraint master problem comprising at least one of the following: accuracy, complexity, and coincidence of the rule learning model; repeating the following steps until no new rule can be added: solving the constraint master problem based on existing rules to obtain dual variables; and generating new rules based on the frequent pattern tree according to the obtained dual variables, the new rules being rules with minimum test numbers.

2. The method of claim 1, wherein, constructing a constraint master problem, comprising: constructing a constraint master problem according to decision variables of existing rules; a decision variable corresponding to each rule is used to represent whether the rule is adopted; wherein the objective function of the constraint master problem comprises at least one of the following: a number of negative samples covered by adopted rules and / or a number of positive samples not covered by adopted rules, used to represent the accuracy of the rule learning model; a number of positive samples covered multiple times by adopted rules, used to represent the coincidence of the rule learning model; a sum of lengths of adopted rules, used to represent the complexity of the rule learning model.

3. The method of claim 2, wherein, Each item in the constraint master problem is provided with a coefficient; the method further comprises: obtaining at least one of the following: an accuracy coefficient, a coincidence coefficient, and a complexity coefficient input by a user; correspondingly, the constraint master problem is constructed by at least one of the following: a coefficient input by a user, a coefficient determined according to a sample number, and a coefficient set by default.

4. The method of claim 1, wherein, generating new rules based on the frequent pattern tree according to the obtained dual variables, comprising: constructing a corresponding frequent pattern tree through any rule head, calculating a lower bound of a test number corresponding to the rule head according to the dual variables to determine whether the rule head needs to be analyzed, and recursively analyzing the frequent pattern tree corresponding to the rule head that needs to be analyzed to find the rule with the minimum test number as the new rule; wherein in the frequent pattern tree, a node is used to represent a feature, and a rule head comprises one or more features; the frequent pattern tree corresponding to the rule head is a frequent pattern tree constructed according to positive samples containing the rule head; and the lower bound of the test number corresponding to the rule head is used to represent lower bounds of test numbers corresponding to various path rules in the frequent pattern tree corresponding to the rule head.

5. The method of claim 4, wherein, The corresponding frequent pattern tree is constructed through any rule head, the test number lower bound corresponding to the rule head is calculated according to the dual variable, whether the rule head needs to be analyzed is determined, the frequent pattern tree corresponding to the rule head which needs to be analyzed is analyzed recursively, and the rule with the minimum test number found is the new rule, comprising: Initializing the optimal test number as 0, and setting the initial rule head as empty; An initial frequent pattern tree is constructed based on the positive samples satisfying the current rule head, and the following steps are repeatedly executed until the rules included in the positive sample group are analyzed: According to the frequent pattern tree corresponding to the current rule head, a set of single-feature nodes with an occurrence frequency greater than or equal to a preset frequency threshold is determined, for each element in the set, the test number corresponding to the rule formed by the element and the current rule head is determined, and when the corresponding test number is less than the optimal test number, the optimal test number is updated as the corresponding test number; For each element in the set, the test number lower bound corresponding to the element is calculated, if the test number lower bound is greater than or equal to the optimal test number, the element is skipped, otherwise, a new rule head and the corresponding frequent pattern tree are constructed based on the depth-first recursive method.

6. The method of claim 5, wherein, The test number lower bound corresponding to the element is calculated, if the test number lower bound is greater than or equal to the optimal test number, the element is skipped, otherwise, a new rule head and the corresponding frequent pattern tree are constructed based on the depth-first recursive method, comprising: If the rule formed by the element and the current rule head exists on the path of the initial frequent pattern tree, and there are nodes to be analyzed along the depth direction from the element, the test number lower bound corresponding to the element is calculated, if the test number lower bound is greater than or equal to the optimal test number, the element is skipped; If the test number lower bound is less than the optimal test number, a frequent pattern tree is constructed according to the new rule head formed by the current rule head and the element; If there is no node to be analyzed along the depth direction, the next node to be analyzed is found by returning from the element in the initial frequent pattern tree, and a frequent pattern tree is constructed based on the rule head corresponding to the next node.

7. The method of claim 5, wherein, The test number lower bound corresponding to the element is calculated, comprising: The test number lower bound is calculated according to at least one of the following costs: The number of features included in the current rule; The negative number of the sum of the dual variables corresponding to the positive samples covered by the current rule; The support degree of the path rule from the root node to the leaf node in the frequent pattern tree constructed by taking the current rule as the rule head, which covers the least positive samples, in the positive sample group; The support degree of the path rule from the root node to the leaf node in the frequent pattern tree constructed by taking the current rule as the rule head, which covers the least negative samples, in the negative sample group; Wherein, when the test number lower bound corresponding to the element is calculated, the current rule is the rule formed by the element and the current rule head.

8. The method according to any one of claims 1 to 7, characterized in that, The feature is matched with the rules in the rule learning model, whether the business object to be analyzed meets the requirements and / or the reason is determined according to the matching result, comprising: If there is a rule in the rule learning model that matches the feature corresponding to the business object to be analyzed, it is determined that the business object to be analyzed meets the requirement, and the reason why the business object meets the requirement is that the matched rule meets the requirement. According to the proportion of positive and negative samples covered by each feature in the matched rule and the proportion of positive and negative samples covered by the rule, the core reason why the business object to be analyzed meets the requirement is determined. At least one of the following is displayed: whether the business object to be analyzed meets the requirement, the reason why it meets the requirement, and the core reason why it meets the requirement.

9. A power load forecasting method characterized by, It includes: Obtaining environmental features to be analyzed; According to the environmental features and the rule learning model, it is predicted whether the power load corresponding to the environmental features meets the requirement, and / or the reason why the power load prediction result is determined is output; The rule learning model is generated based on column generation technology and / or frequent pattern tree; It also includes: Obtaining a training sample set, the training sample set including a positive sample group and a negative sample group, wherein the positive sample includes an environmental feature that meets the requirement, and the negative sample includes an environmental feature that does not meet the requirement; According to the training sample set, an initial rule is generated, and each rule includes at least one feature; Based on column generation technology, new rules are continuously added until a rule learning model is obtained; during column generation, new rules are generated based on a frequent pattern tree; Based on column generation technology, new rules are continuously added until a rule learning model is obtained, including: Constructing a constraint master problem, the constraint master problem including at least one of the following: accuracy, complexity, and coincidence of the rule learning model; Repeat the following steps until no new rule can be added: solve the constraint master problem based on the existing rule to obtain a dual variable; generate a new rule based on the frequent pattern tree according to the obtained dual variable, and the new rule is the rule with the smallest test number.

10. The method of claim 9, wherein, It also includes: Obtaining data corresponding to multiple regions, the data corresponding to the region including environmental features corresponding to a first time combination and power load corresponding to the first time combination; wherein the first time combination includes at least two time nodes; According to the data corresponding to the multiple regions, divide the positive sample group and the negative sample group, wherein if the difference between the power loads corresponding to the at least two time nodes is greater than a difference threshold, the difference between the environmental features corresponding to the at least two time nodes is a positive sample, otherwise the difference between the environmental features corresponding to the at least two time nodes is a negative sample; According to the positive sample group and the negative sample group, generate a rule learning model based on column generation technology and / or a frequent pattern tree; The generated rule learning model is used to predict whether the variation of the power load of the region to be analyzed in a second time combination exceeds the difference threshold according to the difference between the environmental features of the region to be analyzed in the second time combination.

11. An electronic device, comprising: It includes: At least one processor; And The memory is connected in communication with the at least one processor; The memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the electronic device to perform the method of any one of claims 1-10.

12. A computer-readable storage medium, characterized in that, The computer readable storage medium stores computer execution instructions, and when the processor executes the computer execution instructions, the method of any one of claims 1-10 is implemented.

Images