Data causal structure learning method and system with group features

Abstract

The invention discloses a data causal structure learning method and system with group features, and the method comprises the steps: obtaining a currently arriving data feature group, and employing a conditional independence test to recognize a feature associated with each feature in the current data feature group, so as to construct a skeleton structure in the current data feature group; performing orientation on the skeleton structure in the current data feature group through V structure identification and using an orientation rule to form a causal skeleton structure in the current data feature group; fusing the causal skeleton structure in the current data feature group with an original causal skeleton structure in the current feature space, and constructing a causal skeleton structure diagram in the current state space; based on the causal skeleton structure diagram, identifying a causal diagram in a current feature space by using an orientation rule; according to the method, the contained causal structure relation can be found from the linear arbitrary distribution data with group characteristics, and the computer resource consumption is low.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data mining and machine learning, and particularly relates to a data causal structure learning method and system with group features. BACKGROUND

[0002] Causal structure learning refers to identifying the causal relationships and causal structures between variables through data analysis and modeling, so as to infer the causal effects between variables, and aims to better understand and predict the relationships between features in the data set by automatically learning the causal relationships between features. Causal structure learning is increasingly widely used in the fields of data science, machine learning and artificial intelligence, and has important significance in decision analysis, medical health, social science, finance and other fields.

[0003] Causal structure learning can be used for decision analysis and interpretation, which helps to better understand the causal relationships and provide support for making correct decisions. By learning the causal structure, the causal relationships between variables can be discovered, and these relationships can be used to build more accurate and reliable prediction models. Therefore, with the continuous development of data science and machine learning technology, causal structure learning has also received extensive attention.

[0004] Existing causal structure learning methods can be mainly divided into three categories: constraint-based causal structure learning methods, score-based causal structure learning methods and hybrid causal structure learning methods. However, when the faithfulness is violated, the score-based causal structure learning method will have a calculation problem, which will result in no results, so almost all the causal structure learning methods that have appeared in recent years are constraint-based causal structure learning methods. However, the existing causal structure learning methods need to obtain the entire feature space in advance, and cannot effectively obtain the causal structure relationship between the group features, and cannot effectively process the features that arrive in groups. In actual applications, group features are very common, such as various physiological indicators in medicine, investment portfolios in finance, or feature groups in image analysis, spam filtering and other application scenarios. Therefore, a new method is needed to effectively process data with group features online and accurately identify the causal relationships within and between feature groups, while further avoiding the reduction of computer resource consumption caused by obtaining the entire feature space.

[0005] In the related art, the patent application document with publication number CN110866162A uses a local causal network structure discovery algorithm based on mutual information, which removes false variables based on regression analysis equations and generates a local causal network structure with direction based on conditional independence test in an undirected graph, and performs causal relationship mining of learning effect; but this scheme needs to obtain the entire feature space of the data set in advance, and cannot process features that arrive in groups. SUMMARY

[0006] The technical problem to be solved by the present application is how to perform causal structure learning on data with group features such as image data or mail data, and reduce computer resource consumption.

[0007] The present application solves the above technical problem by the following technical means:

[0008] The present application provides a method for causal structure learning of data with group features, comprising:

[0009] Obtaining a current data feature group, for each feature in the current data feature group, respectively using conditional independence test to identify features associated with each feature to construct a skeleton structure in the current data feature group;

[0010] Directing the skeleton structure in the current data feature group by V structure identification and using directional rules to form a causal skeleton structure in the current data feature group;

[0011] Fusing the causal skeleton structure in the current data feature group with the original causal skeleton structure under the current feature space to construct a causal skeleton structure graph under the current state space;

[0012] Based on the causal skeleton structure graph, using directional rules to identify a causal graph under the current feature space.

[0013] Further, the data feature group comprises an image feature group or a mail feature group;

[0014] The features in the image feature group comprise a color histogram, a texture feature and a shape descriptor;

[0015] The features in the mail feature group comprise a sender, a subject keyword and a mail content.

[0016] Further, for each feature in the current data feature group, respectively using conditional independence test to identify features associated with each feature to construct a skeleton structure in the current data feature group, comprising:

[0017] Traversing each feature of the current data feature group, and for any one node feature Y in a parent-child node set PC F of the current traversed feature F, calculating mutual information I(F;Y) between the feature F and the node feature Y;

[0018] For node feature X∈PC F \{F,Y}, calculating I(X;Y) between the node feature X and the node feature Y, and calculating mutual information I(F;X) between the feature F and the node feature X, wherein PC F \{F,Y} represents PCF remove the set of {F, Y};

[0019] based on mutual information I(F; X), I(F; Y) and I(X; Y), determine that the node feature X is a false parent-child node of the feature F, and delete the node feature X;

[0020] In the PC F after all the false parent-child nodes in {F, Y} are deleted, obtain a local skeleton structure around the feature F;

[0021] based on the local skeleton structure around each feature in the current data feature group, construct a skeleton structure within the current data feature group.

[0022] Further, the calculation formula of the mutual information I(F; Y) is:

[0023]

[0024] In the formula, H(F) represents the entropy of the feature F, H(F|Y) represents the entropy of the feature F under the condition of the node feature Y, P(f) represents the probability of the feature F taking the value f, P(y) represents the probability of the node feature Y taking the value y, and P(f, y) represents the probability of the feature F taking the value f and the feature Y taking the value y.

[0025] Further, based on the mutual information I(F; X), I(F; Y) and I(X; Y), when it is determined that the node feature X is a false parent-child node of the feature F, the node feature X is deleted, including:

[0026] determine whether the mutual information I(F; X), I(F; Y) and I(X; Y) satisfy:

[0027] I(F; Y)>I(F; X) && I(X; Y)>I(F; X)

[0028] In the formula, && represents logical and (and);

[0029] If yes, it is determined that the node feature X is a parent-child node of the feature F;

[0030] If no, it is determined that the node feature X is a false parent-child node of the feature F and the node feature X is deleted.

[0031] Further, the skeleton structure within the current data feature group is oriented by V structure recognition and using a directional rule to form a causal skeleton structure within the current data feature group, including:

[0032] all V structures in the skeleton structure within the current data feature group are recognized by V structure recognition, and directional edges are obtained;

[0033] The undirected edges are oriented by using orientation rules, and a causal skeleton in the current data feature set is formed based on all oriented edges, and the orientation rules are specifically:

[0034] For a skeleton structure X-Y-Z, where node feature X and node feature Z are not directly connected, when there is an oriented edge X→Y, the structure Y-Z is oriented as Y→Z;

[0035] When there is a skeleton structure X→Y→Z, if there is a skeleton X-Z, it is oriented as X→Z;

[0036] If there are two skeleton structures X-Z→Y and X-W→Y, and node feature Z and node feature W are not directly connected, then the structure X-Y is oriented as X→Y.

[0037] Further, the V structure in the skeleton structure in the current data feature set is identified by the V structure, and the oriented edge is obtained, including:

[0038] For a skeleton structure F-Y-X, mutual information I(F;X) and I(F;X|Y) are calculated, respectively, F, X, and Y are node features, I(F;X) is the mutual information between node features F and X, and I(F;X|Y) is the conditional mutual information between F and X given variable Y;

[0039] When I(F;X)>0&&I(F;X|Y)>I(F;X) is satisfied, it is determined that the skeleton structure F-Y-X forms a V structure F→Y←X, and the oriented edges F→Y and X→Y are obtained.

[0040] Further, the causal skeleton structure in the current data feature set is fused with the original causal skeleton structure under the current feature space, and a causal skeleton structure diagram under the current state space is constructed, including:

[0041] A node feature G in the current data feature set and a node feature C under the previous feature space are selected, and mutual information I(G;C) between the node feature G and the node feature C is calculated;

[0042] For any node feature X in the parent-child node set PC G of the node feature G, mutual information I(G;X) between the node feature G and the node feature X and mutual information I(C;X) between the node feature C and the node feature X are calculated;

[0043] If I(G;C)>0&&I(G;X)≤I(G;C)&&I(X;C)≤I(G;C) is satisfied, it is determined that the node feature G and the node feature C are directly connected, otherwise the node feature G and the node feature C do not have a directly connected edge;

[0044] After traversing all the features in the current data feature group and the last feature space, a causal skeleton structure diagram in the current state space is constructed.

[0045] Further, the causal graph in the current feature space is identified by using a directional rule based on the causal skeleton structure diagram, including:

[0046] All V structures in the causal skeleton structure diagram are identified by V structures.

[0047] When the skeleton structure X-Y-Z satisfies I(X;Z|Y)>I(X;Z), it is determined that the skeleton structure X-Y-Z has a structure Y→Z or a structure Y→X, wherein I(X;Z) is mutual information between the node feature X and the node feature Z, and I(X;Z|Y) is conditional mutual information between X and Z given the variable Y.

[0048] Based on the V structure, the structure among X, Y, and Z is inferred.

[0049] All the skeleton structures in the causal skeleton structure diagram are traversed, and the causal graph in the current feature space is inferred.

[0050] In addition, the present application further provides a data causal structure learning system with group features, the system comprising:

[0051] A construction module is configured to acquire a current arrived data feature group, and for each feature in the current data feature group, a conditional independence test is used to identify features associated with each feature to construct a skeleton structure in the current data feature group.

[0052] A directional module is configured to identify and use a directional rule to direct the skeleton structure in the current data feature group by V structures, so as to form a causal skeleton structure in the current data feature group.

[0053] A fusion module is configured to fuse the causal skeleton structure in the current data feature group with the original causal skeleton structure in the current feature space, so as to construct a causal skeleton structure diagram in the current state space.

[0054] An identification module is configured to identify the causal graph in the current feature space by using a directional rule based on the causal skeleton structure diagram.

[0055] The present application has the following advantages:

[0056] (1) The present application is directed to the learning of the causal structure of a feature group, which can identify the causal structure in a feature group according to the information provided by the current data feature group, without obtaining the entire feature space, and uses the pairwise comparison between mutual information to construct the causal skeleton, which reduces the consumption of computer resources, improves the performance of the operation platform, and identifies the causal structure in the current feature space in real time according to the information provided by all features in the current feature space, and can discover the causal structure relationship contained in linear arbitrarily distributed data with group features.

[0057] Additional aspects and advantages of the present application will be described in the description that follows, will become apparent from the description, or will be learned by practice of the present application. BRIEF DESCRIPTION OF DRAWINGS

[0058] Figure 1 is a flowchart of a data causal structure learning method with group features according to an embodiment of the present application;

[0059] Figure 2 is a principle block diagram of a data causal structure learning method with group features according to an embodiment of the present application;

[0060] Figure 3 is a structure diagram of a data causal structure learning system with group features according to an embodiment of the present application. DETAILED DESCRIPTION

[0061] In order to make the purpose, technical scheme and advantages of the embodiments of the present application more clear, the technical scheme in the embodiments of the present application will be described clearly and completely in combination with the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the protection scope of the present application.

[0062] As shown in Figures 1 to 2 , an embodiment of the present application discloses a data causal structure learning method with group features, which comprises the following steps:

[0063] S10, obtaining the current data feature group, for each feature in the current data feature group, respectively using conditional independence test to identify the features associated with each feature to construct the skeleton structure in the current data feature group;

[0064] S20, orienting the skeleton structure in the current data feature group by V structure identification and using directional rules to form the causal skeleton structure in the current data feature group;

[0065] S30, fuse the causal skeleton structure in the current data feature group with the original causal skeleton structure under the current feature space to construct a causal skeleton structure diagram under the current state space;

[0066] S40, based on the causal skeleton structure diagram, recognize the causal diagram under the current feature space by using a directional rule.

[0067] For a given set of data features, the embodiment first determines the association between each feature through conditional independence test. If two features are conditionally independent, they are not directly connected; otherwise, there is a directly connected skeleton structure. After obtaining the skeleton structure in the feature group, the constructed in-group skeleton structure is oriented by using a directional rule (such as Meek rule). Finally, after determining the causal structure in the feature group, it is combined with the existing causal structure to remove the incorrectly connected edges, and the global causal structure diagram is inferred based on the existing directional edges.

[0068] The data causal structure learning method with group features proposed in the embodiment is suitable for learning the causal structure of data with group features, such as image analysis or spam filtering, etc. The input of the causal learning mechanism is all feature groups, and the output is the causal structure diagram between features.

[0069] In an embodiment, the data feature group includes an image feature group or a mail feature group;

[0070] The features in the image feature group include color histograms, texture features, and shape descriptors;

[0071] The features in the mail feature group include the sender, the subject keyword, and the mail content.

[0072] For example, in image analysis, the input of the causal learning mechanism can be a set of image features such as color histograms, texture features, shape descriptors, etc. These features can be used to describe different aspects of the image. The output of the causal learning mechanism is a causal structure diagram between features, which can represent the causal relationship between features and further explain which features are more important for image classification or object detection. The causal learning mechanism can analyze and model the input feature group to find the causal relationship between them and generate a causal structure diagram to represent this relationship, which helps to deeply understand the interaction between image features and improve the accuracy and efficiency of image analysis tasks.

[0073] In the email filtering, the input of the causal learning mechanism can be a set of features related to the email, such as the sender, subject keywords, email content, etc., which can be used to determine the category of the email, such as spam or normal email. The output of the causal learning mechanism is a causal structure graph between features, which can represent the causal relationship between features, such as which features have more influence on determining the category of the email. The causal learning mechanism can generate a causal structure graph to describe the relationship between the input feature set by analyzing and modeling the causal relationship between the input feature set, which can help to improve the accuracy and automation of email filtering.

[0074] In an embodiment, the step S10 of identifying the features associated with each feature in the current data feature set to construct the skeleton structure in the current data feature set by using conditional independence test respectively for each feature in the current data feature set comprises the following steps:

[0075] S11, traversing each feature in the current data feature set, and for any node feature Y in the parent-child node set PC F of the current traversed feature F, calculating the mutual information I(F;Y) between the feature F and the node feature Y, wherein PC F represents all features directly connected to the feature F (i.e., the parent-child nodes of the feature F), and the initial state of PC F is the newly arrived data feature set;

[0076] S12, for node feature X e PC F \{F,Y}, calculating the mutual information I(X;Y) between the node feature X and the node feature Y, and calculating the mutual information I(F;X) between the feature F and the node feature X, wherein PC F \{F,Y} represents the set PC F excluding {F,Y};

[0077] S13, based on the mutual information I(F;X), I(F;Y) and I(X;Y), determining that the node feature X is a false parent-child node of the feature F, and deleting the node feature X;

[0078] S14, after all the false parent-child nodes in PC F \{F,Y} are deleted, obtaining the local skeleton structure around the feature F;

[0079] S15, based on the local skeleton structure around each feature in the current data feature set, constructing the skeleton structure in the current data feature set.

[0080] It should be noted that steps S12-S13 are repeated until all the false parent-child nodes are deleted, at which time the local skeleton structure around the feature F is obtained; steps S11-S14 are then repeated to find the local skeleton structure of all features in the current data feature group, and the skeleton structure of the current data feature group is constructed.

[0081] In an embodiment, the formula for calculating the mutual information I(F;Y) is:

[0082]

[0083] In the formula, H(F) represents the entropy of the feature F, H(F|Y) represents the entropy of the feature F under the condition of the node feature Y, P(f) represents the probability of the feature F taking the value f, P(y) represents the probability of the node feature Y taking the value y, and P(f,y) represents the probability of the feature F taking the value f and the feature Y taking the value y. For example, in the image feature and the mail feature, the feature F represents the target feature, the features X and Y represent any two features other than the target feature, and f, x, and y represent possible values of F, X, and Y, respectively.

[0084] It should be noted that the mutual information is an index for measuring the strength of the relationship between two features, and the greater the mutual information, the stronger the correlation between the features.

[0085] It should be understood that the mutual information I(X;Y) between the node feature X and the node feature Y, and the formula for calculating the mutual information I(F;X) between the feature F and the node feature X are similar to the formula for calculating the mutual information I(F;Y), and will not be described here.

[0086] In an embodiment, when the step S13 determines that the node feature X is a false parent-child node of the feature F, the node feature X is deleted, including the following steps:

[0087] S131, it is determined whether the mutual information I(F;X), I(F;Y), and I(X;Y) satisfy I(F;Y)>I(F;X) && I(X;Y)>I(F;X), if yes, step S132 is performed, and if no, step S133 is performed.

[0088] I(F;Y)>I(F;X) && I(X;Y)>I(F;X)

[0089] In the formula, && represents the logical and (and);

[0090] S132, it is determined that the node feature X is a parent-child node of the feature F.

[0091] S133, it is determined that the node feature X is a false parent-child node of the feature F and the node feature X is deleted.

[0092] In an embodiment, the step S20 of orienting the skeleton structure in the current data feature group by V structure identification and using orientation rules to form a causal skeleton structure in the current data feature group comprises the following steps:

[0093] S21. Identify all V structures in the skeleton structure in the current data feature group by V structure identification to obtain oriented edges.

[0094] S22. Orient the edges that are not oriented using orientation rules, and form a causal skeleton in the current data feature group based on all oriented edges, wherein the orientation rules are specifically as follows:

[0095] No new V structure is generated: for a skeleton structure X-Y-Z, wherein the node feature X and the node feature Z are not directly connected, when there is a directed edge X→Y, the structure Y-Z is oriented as Y→Z.

[0096] No loop is generated: when there is a skeleton structure X→Y→Z, if there is a skeleton X-Z, it is oriented as X→Z.

[0097] Forced 3-way branching: if there are two skeleton structures X-Z→Y and X-W→Y, and the node feature Z and the node feature W are not directly connected, the structure X-Y is oriented as X→Y.

[0098] Based on the above three auxiliary orientation rules, the skeleton structure constructed in the current data feature group can be oriented to form a causal structure diagram in the feature group.

[0099] In an embodiment, the step S21 of identifying all V structures in the skeleton structure in the current data feature group by V structure identification to obtain oriented edges comprises the following steps:

[0100] For a skeleton structure F-Y-X, mutual information I(F;X) and I(F;X|Y) are calculated, respectively, F, X, and Y are node features, I(F;X) is the mutual information between the node feature F and the node feature X, and I(F;X|Y) is the conditional mutual information between F and X given the variable Y.

[0101] When I(F;X)>0&&I(F;X|Y)>I(F;X) is satisfied, it is determined that the skeleton structure F-Y-X constitutes a V structure F→Y←X, and oriented edges F→Y and X→Y are obtained.

[0102] After all V structures are identified, the remaining edges that are not oriented are oriented using the above three orientation rules in this embodiment.

[0103] In an embodiment, the step S30 of fusing the causal skeleton structure in the current data feature set with the original causal skeleton structure under the current feature space to construct a causal skeleton structure diagram under the current state space comprises the following steps.

[0104] S31, selecting a node feature G in the current data feature set and a node feature C under the last feature space, and calculating mutual information I(G;C) between the node feature G and the node feature C;

[0105] S32, for any node feature X in the parent-child node set PC of the node feature G, G calculating mutual information I(G;X) between the node feature G and the node feature X and mutual information I(C;X) between the node feature C and the node feature X;

[0106] S33, if I(G;C)>0&&I(G;X)≤I(G;C)&&I(X;C)≤I(G;C) is satisfied, it is determined that the node feature G and the node feature C are directly connected, otherwise the node feature G and the node feature C do not exist a directly connected edge;

[0107] S34, after traversing all the features in the current data feature set and under the last feature space, a causal skeleton structure diagram under the current state space is constructed.

[0108] In an embodiment, the step S40 of identifying a causal diagram under the current feature space based on the causal skeleton structure diagram by using a directional rule comprises the following steps.

[0109] S41, all V structures in the causal skeleton structure diagram are identified by V structure.

[0110] It should be noted that, specifically, any one skeleton structure X-Y-Z is selected, and mutual information I(X;Z) and I(X;Z|Y) are calculated respectively, and it is judged by a directional rule whether there is a structure X→Y and a structure Z→Y, and when all the skeleton structures are traversed, all the V structures in the causal skeleton structure diagram are obtained.

[0111] S42, any one skeleton structure X-Y-Z is selected, and when the skeleton structure satisfies I(X;Z|Y)>I(X;Z), it is determined that the skeleton structure X-Y-Z has a structure Y→Z or a structure Y→X, wherein I(X;Z) is mutual information between a node feature X and a node feature Z, and I(X;Z|Y) is conditional mutual information between X and Z given a variable Y;

[0112] S43, based on the V structure, a structure among X, Y, and Z is inferred.

[0113] S44, traverse all skeleton structures in the causal skeleton structure graph, and deduce a causal graph under the current feature space.

[0114] The embodiment is based on the skeleton structure constructed under the current feature space, uses a directional rule to identify part of the newly constructed structure, and deduces the remaining undirectional structure.

[0115] As shown in Figure 2 An embodiment of the present application discloses a data causal structure learning system with group features, which comprises:

[0116] A construction module 10 is configured to acquire a current arrived data feature group, and for each feature in the current data feature group, respectively adopt a conditional independence test to identify features associated with each feature to construct a skeleton structure in the current data feature group.

[0117] A directional module 20 is configured to identify and direct the skeleton structure in the current data feature group by a V structure identification and a directional rule to form a causal skeleton structure in the current data feature group.

[0118] A fusion module 30 is configured to fuse the causal skeleton structure in the current data feature group with a previously existing causal skeleton structure under a current feature space to construct a causal skeleton structure graph under a current state space.

[0119] An identification module 40 is configured to identify a causal graph under the current feature space based on the causal skeleton structure graph and by using a directional rule.

[0120] The embodiment is based on the conditional independence test to identify the association between features for a newly arrived feature group, and then obtain a skeleton structure in the feature group; then the V structure identification and the Meek rule are used to direct the skeleton structure to obtain a causal structure in the feature group; finally, the skeleton structure constructed by the newly arrived feature group and the previously existing skeleton structure under the current feature space are merged, the edges between the features are removed, and the global causal structure graph is deduced based on the existing directional edges. The problem that the traditional causal structure learning method needs to obtain the entire feature space in advance and cannot effectively obtain the causal structure relationship between the feature groups, and then cannot be applied to the scene where the feature groups arrive is solved.

[0121] In an embodiment, the data feature group comprises an image feature group or a mail feature group.

[0122] The features in the image feature group comprise a color histogram, a texture feature and a shape descriptor.

[0123] The features in the mail feature group comprise a sender, a subject keyword and a mail content.

[0124] In an embodiment, the constructing module 10 comprises:

[0125] a first mutual information calculating unit, configured to traverse each feature of a current data feature set, and calculate mutual information I(F; Y) between a feature F currently traversed and any node feature Y in a parent-child node set PC F of the feature F;

[0126] a second mutual information calculating unit, configured to calculate mutual information I(X; Y) between a node feature X and the node feature Y, and calculate mutual information I(F; X) between the feature F and the node feature X, for the node feature X∈PC F \{F,Y}; F wherein PC F \{F,Y} represents a set of PC

[0127] a node deleting unit, configured to delete the node feature X when the node feature X is determined to be a false parent-child node of the feature F based on the mutual information I(F; X), I(F; Y) and I(X; Y);

[0128] a local skeleton structure constructing unit, configured to obtain a local skeleton structure around the feature F after all false parent-child nodes in PC F \{F,Y} are deleted;

[0129] a skeleton constructing unit, configured to construct a skeleton structure in the current data feature set based on the local skeleton structure around each feature in the current data feature set.

[0130] In an embodiment, the node deleting unit is specifically configured to:

[0131] determine whether the mutual information I(F; X), I(F; Y) and I(X; Y) satisfy:

[0132] I(F; Y)>I(F; X) && I(X; Y)>I(F; X)

[0133] wherein && represents logical and (and);

[0134] if yes, it is determined that the node feature X is a parent-child node of the feature F;

[0135] if no, it is determined that the node feature X is a false parent-child node of the feature F and the node feature X is deleted.

[0136] In an embodiment, the directing module 20 comprises:

[0137] a first V structure identifying unit, configured to identify all V structures in a skeleton structure in the current data feature set through V structure, to obtain a directed edge;

[0138] The orientation unit is used to orient unoriented edges using orientation rules, and to form a causal skeleton within the current data feature group based on all oriented edges. The orientation rules are specifically as follows:

[0139] For the skeleton structure XYZ, where node features X and node features Z are not directly connected, when there is an oriented edge X→Y, the structure YZ is oriented as Y→Z;

[0140] When there is a skeleton structure X→Y→Z, if there is a skeleton XZ, then the orientation is X→Z;

[0141] If there are two skeleton structures XZ→Y and XW→Y, and node features Z and W are not directly connected, then the structure XY is oriented as X→Y.

[0142] In one embodiment, the V-structure identification unit is used for:

[0143] For the skeleton structure F–Y–X, calculate the mutual information I(F;X) and I(F;X|Y) respectively, where F, X, and Y are the node features, I(F;X) is the mutual information between node feature F and node feature X, and I(F;X|Y) is the conditional mutual information between F and X given variable Y.

[0144] When I(F;X)>0 &&I(F;X|Y)>I(F;X), the skeleton structure F–Y–X is determined to form the V structure F→Y←X, and the oriented edges F→Y and X→Y are obtained.

[0145] In one embodiment, the fusion module 30 includes:

[0146] The third mutual information calculation unit is used to select a node feature G in the current data feature group and a node feature C in the previous feature space, and calculate the mutual information I(G;C) between node feature G and node feature C.

[0147] The fourth mutual information calculation unit is used for the set of parent and child nodes PC of node feature G. G For any node feature X in the dataset, calculate the mutual information I(G;X) between node feature G and node feature X, and the mutual information I(C;X) between node feature C and node feature X.

[0148] An edge-determining unit is used to determine if node feature G and node feature C are directly connected if I(G;C)>0&&I(G;X)≤I(G;C)&&I(X;C)≤I(G;C), otherwise node feature G and node feature C do not have a directly connected edge.

[0149] The causal skeleton structure graph construction unit is used to traverse all features within the current data feature group and the previous feature space to construct the causal skeleton structure graph in the current state space.

[0150] In one embodiment, the identification module 40 includes:

[0151] The second V-structure identification unit is used to identify all V-structures in the causal skeleton structure diagram through V-structures;

[0152] The skeleton structure determination unit is used to select any skeleton structure X–Y–Z. When the skeleton structure satisfies I(X;Z|Y)>I(X;Z), it is determined that the skeleton structure X–Y–Z has a structure Y→Z or a structure Y→X. Here, I(X;Z) is the mutual information between node features X and node features Z, and I(X;Z|Y) is the conditional mutual information between given variables Y, X, and Z.

[0153] The inference unit is used to infer the structure between X, Y, and Z based on the V structure;

[0154] The identification unit is used to traverse all skeleton structures in the causal skeleton structure graph and infer the causal graph in the current feature space.

[0155] It should be noted that other embodiments or implementation methods of the data causal structure learning system with group features described in this invention can refer to the above-described method embodiments, and will not be repeated here.

[0156] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0157] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0158] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A data causal structure learning method with group features, characterized in that, The method includes: Obtain the currently arrived data feature group, and for each feature in the current data feature group, use the conditional independence test to identify the features associated with each feature in order to construct the skeleton structure of the current data feature group; By identifying V-structures and using orientation rules, the skeleton structure within the current data feature group is oriented to form a causal skeleton structure within the current data feature group. The causal skeleton structure within the current data feature group is merged with the original causal skeleton structure in the current feature space to construct a causal skeleton structure diagram in the current state space. Based on the causal skeleton structure graph, the causal graph in the current feature space is identified using orientation rules.

2. The data causal structure learning method with group features as described in claim 1, characterized in that, The data feature group includes an image feature group or an email feature group; The features in the image feature group include color histogram, texture features, and shape descriptors; The features in the email feature group include the sender, subject keywords, and email content.

3. The data causal structure learning method with group features as described in claim 1, characterized in that, The step of using a conditional independence test to identify associated features for each feature within the current data feature group to construct the skeleton structure within the current data feature group includes: Iterate through each feature in the current data feature group, and for the currently traversed feature F, find the set of parent and child nodes PC. F For any node feature Y in the dataset, calculate the mutual information I(F;Y) between feature F and node feature Y; For node features X∈PC F Let {F,Y} be the region of interest. Calculate the mutual information I(X;Y) between node features X and Y, and calculate the mutual information I(F;X) between feature F and node feature X, where PC... F {F,Y} represents PC F The set excluding {F,Y}; Based on mutual information I(F;X), I(F;Y) and I(X;Y), when it is determined that node feature X is a false parent-child node of feature F, the node feature X is deleted; On PC F After deleting all pseudo-parent-child nodes in \{F,Y}, the local skeleton structure around feature F is obtained; Based on the local skeleton structure around each feature in the current data feature group, construct the skeleton structure within the current data feature group.

4. The data causal structure learning method with group features as described in claim 3, characterized in that, The formula for calculating the mutual information I(F;Y) is as follows: In the formula, H(F) represents the entropy of feature F, H(F|Y) represents the entropy of feature F under the condition of node feature Y, P(f) represents the probability that feature F takes the value f, P(y) represents the probability that node feature Y takes the value y, and P(f,y) represents the probability that feature F takes the value f and feature Y takes the value y.

5. The data causal structure learning method with group features as described in claim 3, characterized in that, When determining that node feature X is a false parent-child node of feature F based on mutual information I(F;X), I(F;Y), and I(X;Y), deleting node feature X includes: Determine whether the mutual information I(F;X), I(F;Y) and I(X;Y) satisfy the following: I(F;Y)>I(F;X)&&I(X;Y)>I(F;X) In the formula, && represents logical AND; If the condition is met, then the node feature X is determined to be the parent-child node of feature F; If the condition is not met, then the node feature X is determined to be a false parent-child node of feature F and the node feature X is deleted.

6. The data causal structure learning method with group features as described in claim 1, characterized in that, The step of orienting the skeleton structure within the current data feature group by identifying the V-structure and using orientation rules to form a causal skeleton structure within the current data feature group includes: By identifying all V-structures in the skeleton structure within the current data feature group through V-structure identification, directional edges are obtained; Orientation rules are used to orient unoriented edges, and a causal skeleton within the current data feature group is formed based on all oriented edges. Specifically, the orientation rules are as follows: For the skeleton structure XYZ, where node features X and node features Z are not directly connected, when there is an oriented edge X→Y, the structure YZ is oriented as Y→Z; When there is a skeleton structure X→Y→Z, if there is a skeleton XZ, then the orientation is X→Z; If there are two skeleton structures XZ→Y and XW→Y, and node features Z and W are not directly connected, then the structure XY is oriented as X→Y.

7. The data causal structure learning method with group features as described in claim 6, characterized in that, The step of identifying all V-structures in the skeleton structure within the current data feature group through V-structure identification to obtain directional edges includes: For the skeleton structure F–Y–X, calculate the mutual information I(F;X) and I(F;X|Y) respectively, where F, X, and Y are the node features, I(F;X) is the mutual information between node feature F and node feature X, and I(F;X|Y) is the conditional mutual information between F and X given variable Y. When I(F;X)>0 &&I(F;X|Y)>I(F;X), the skeleton structure F–Y–X is determined to form the V structure F→Y←X, and the oriented edges F→Y and X→Y are obtained.

8. The data causal structure learning method with group features as described in claim 1, characterized in that, The step of fusing the causal skeleton structure within the current data feature group with the original causal skeleton structure in the current feature space to construct a causal skeleton structure diagram in the current state space includes: Select a node feature G within the current data feature group and a node feature C in the previous feature space, and calculate the mutual information I(G;C) between node feature G and node feature C; For the set of parent and child nodes PC of node feature G G For any node feature X in the dataset, calculate the mutual information I(G;X) between node feature G and node feature X, and the mutual information I(C;X) between node feature C and node feature X. If I(G;C)>0&&I(G;X)≤I(G;C)&&I(X;C)≤I(G;C), then node feature G and node feature C are directly connected; otherwise, node feature G and node feature C do not have a directly connected edge. After traversing all features within the current data feature group and the previous feature space, construct a causal skeleton structure diagram in the current state space.

9. The data causal structure learning method with group features as described in claim 1, characterized in that, The step of identifying the causal graph in the current feature space based on the causal skeleton structure graph and using orientation rules includes: Identify all V-structures in the causal skeleton structure diagram using V-structures; Choose any skeleton structure X–Y–Z. When the skeleton structure satisfies I(X;Z|Y)>I(X;Z), determine that the skeleton structure X–Y–Z has a structure Y→Z or a structure Y→X. Here, I(X;Z) is the mutual information between node features X and node features Z, and I(X;Z|Y) is the conditional mutual information between given variables Y, X and Z. Based on the V structure, the structure between X, Y, and Z can be inferred; By traversing all the skeleton structures in the causal skeleton structure graph, the causal graph in the current feature space is inferred.

10. A data causal structure learning system with group features, characterized in that, The system includes: The construction module is used to obtain the currently arrived data feature group, and for each feature in the current data feature group, the conditional independence test is used to identify the features associated with each feature in order to construct the skeleton structure of the current data feature group; The orientation module is used to orient the skeleton structure within the current data feature group by identifying the V-structure and using orientation rules to form a causal skeleton structure within the current data feature group. The fusion module is used to fuse the causal skeleton structure in the current data feature group with the original causal skeleton structure in the current feature space to construct a causal skeleton structure diagram in the current state space. The identification module is used to identify the causal graph in the current feature space based on the causal skeleton structure graph and using orientation rules.

Images