Robust causal relationship learning method, system, device and storage medium

Abstract

The application discloses a kind of robust causal relationship learning method, system, equipment and storage medium, large-scale variable set can be effectively handled, and when variable set scale is larger, the problem is decomposed into multiple small-scale problems and is handled, improve the efficiency of processing large-scale variable set;Meanwhile, the application has good scalability, can be used with any causal algorithm, so that the application can adapt to a variety of different application scenarios and needs;And, the application provides a new way to solve complex causal inference problem, the idea of divide and conquer and the concept of causal cut provide a new perspective and tool for subsequent causal inference research;In addition, the directed causal graph finally obtained by the application has high accuracy, and can improve the effect of the application field.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and machine learning, and in particular to a robust causal relationship learning method, system, device, and storage medium. Background Technology

[0002] Causal relationship learning is a crucial component of artificial intelligence and machine learning, particularly valuable in applications such as gene therapy, economic forecasting, and advertising campaign design. However, causal relationship learning often involves handling complex graphical probabilistic models, which is especially challenging when a large number of equivalence classes exist. In such models, it is difficult to understand the data generation process using traditional statistical learning methods, thus hindering the accurate description of the joint distribution of observed variables. This problem is exacerbated when experimental manipulation of samples is not possible. Therefore, researchers have begun to leverage causal asymmetry under the assumption of generative processes to optimize causal relationship learning methods in non-experimental settings. However, these methods still struggle with small sample sizes, posing a significant challenge for many practical applications.

[0003] Chinese invention patent application CN202310743834.0, entitled "A Causal Discovery Method Based on Conditional Independence Testing and Continuous Optimization," employs conditional independence testing to establish a masking matrix, designs constraint terms in the continuous optimization loss function, and then uses a combination of variational autoencoders and structural causal models to establish a variational graph autoencoder model. This method also optimizes the equality-constrained optimization problem using a quadratic penalty function method to derive a weight matrix, further discovering causal relationships. This method performs well in validation on virtual and real biological datasets, helping to overcome the limitations of existing continuous optimization-based causal discovery methods in finding causal relationships in scenarios with limited samples and heterogeneous noise.

[0004] Chinese invention patent application CN202310588420.5, entitled "A Method for Analyzing Medical Risk Factors Based on Causal Discovery of Chronic Diseases in the Elderly," employs a causal discovery algorithm to mine the causal relationships between feature variables in medical data on chronic diseases in the elderly, constructing a causal relationship graph of the features. By identifying feature variables that have direct or indirect causal relationships with adverse events, and using a decision tree model for risk stratification, it achieves aggregated analysis of risk factors for adverse events in chronic diseases in the elderly.

[0005] Chinese invention patent application CN202310518076.2, entitled "A Method for Causal Discovery in Time Series Data Based on Mutual Information Measurement," utilizes the PCSMI algorithm to construct a summary causal graph when confounding factors are identified among time series variables. Conversely, when no confounding factors are identified, the FCISMI algorithm is used to construct a summary causal graph. By defining the summary causal graph, causal summary mutual information, and conditional causal summary mutual information, this method can effectively improve the accuracy of causal discovery in time series data.

[0006] The Chinese invention patent application CN202310540168.0, entitled "Hierarchical Causal Discovery Method, Apparatus, Medium and Equipment for Digital Agricultural Information", constructs a causal discovery framework with a two-layer hierarchical structure including high-level strategies and low-level strategies by acquiring a dataset of digital agricultural information. This reduces the complexity of independence testing and significantly improves the efficiency of causal discovery in the case of large-scale variables.

[0007] While the aforementioned patents each possess their own characteristics and advantages in causal relationship discovery methods, most of them fail to effectively handle small sample problems. Small sample problems primarily refer to the issue of insufficient available data in causal relationship discovery, which can severely impact the results due to noise, leading to poor application performance. In most cases, these methods rely on large sample sizes to return robust results, placing high demands on practical applications. Furthermore, these methods often fail to find true causal relationships when dealing with graphical probabilistic models with numerous equivalence classes. This means that even with massive amounts of data, these methods cannot accurately understand the data generation process, thus failing to accurately describe the joint distribution of observed variables. This problem is even more pronounced when experimental manipulation of samples is not possible. Therefore, how to improve the accuracy and robustness of causal relationship discovery under small sample conditions, thereby enhancing the effectiveness in relevant application fields, is a key research issue. Summary of the Invention

[0008] The purpose of this invention is to provide a robust causal relationship learning method, system, device, and storage medium that not only has theoretical robustness but also demonstrates superior scalability and effectiveness in experiments with synthetic and real data, thereby improving the effectiveness in the relevant application fields.

[0009] The objective of this invention is achieved through the following technical solution:

[0010] A robust causal relationship learning method includes:

[0011] The input data containing the set of variables to be studied, V, is preprocessed to obtain the sample set D; wherein, the variables to be studied are various types of terms in the research field.

[0012] If the size of the set of variables to be studied, V, does not meet the set requirements, then a causal cut (C, V1, V2) is found on the set of variables to be studied, and a causal relationship learning algorithm is applied on V1∪C and the sample set D to obtain a directed causal graph G1, and a causal relationship learning algorithm is applied on V2∪C and the sample set D to obtain a directed causal graph G2; wherein V1, V2, and C are three non-overlapping subsets of variables in the set of variables to be studied, and V1∪V2∪C=V, and the subsets V1 and V2 have no edges in the directed graph of the set of variables to be studied, V;

[0013] Merge directed causal graph G1 and directed causal graph G2 to obtain the final directed causal graph G.

[0014] A robust causal relationship learning system includes:

[0015] The preprocessing unit is used to preprocess the input data containing the set of variables to be studied V to obtain the sample set D; wherein, the variables to be studied are various types of terms in the research field.

[0016] The subproblem segmentation and causal relationship learning unit is used to find a causal cut (C, V1, V2) on the set of variables to be studied when the size of the set of variables to be studied does not meet the set requirements. Then, it applies the causal relationship learning algorithm on V1∪C and the sample set D to obtain a directed causal graph G1, and applies the causal relationship learning algorithm on V2∪C and the sample set D to obtain a directed causal graph G2. Here, V1, V2 and C are three non-overlapping variable subsets of the set of variables to be studied V, and V1∪V2∪C=V, and the variable subsets V1 and V2 have no edges in the directed graph of the set of variables to be studied V.

[0017] The result merging unit is used to merge the directed causal graph G1 and the directed causal graph G2 to obtain the final directed causal graph G.

[0018] A processing device includes: one or more processors; and a memory for storing one or more programs;

[0019] When the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned method.

[0020] A readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned method.

[0021] As can be seen from the technical solution provided by the present invention, it can effectively handle large-scale variable sets, and when the variable set is large, it decomposes the problem into multiple smaller problems for processing, thereby improving the efficiency of handling large-scale variable sets. Simultaneously, the present invention has good scalability and can be used in conjunction with any causal algorithm, enabling it to adapt to various application scenarios and needs. Furthermore, the present invention provides a novel approach to solving complex causal inference problems; its divide-and-conquer strategy and the concept of causal cut provide a new perspective and tool for subsequent causal inference research. In addition, the directed causal graph finally obtained by the present invention has high accuracy and can improve the effectiveness in its application domain. Attached Figure Description

[0022] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 A flowchart illustrating a robust causal relationship learning method provided in an embodiment of the present invention;

[0024] Figure 2 A schematic diagram illustrating robust causal relationship learning for sub-problem segmentation and result merging provided in an embodiment of the present invention;

[0025] Figure 3 This is a schematic diagram of a robust causal relationship learning system provided in an embodiment of the present invention;

[0026] Figure 4 This is a schematic diagram of a processing device provided in an embodiment of the present invention. Detailed Implementation

[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0028] First, the following explanations are provided for the terms that may be used in this article:

[0029] The terms “including,” “comprising,” “containing,” “having,” or other similar semantic descriptions should be interpreted as non-exclusive inclusion. For example, “including a technical feature element (such as raw material, component, ingredient, carrier, dosage form, material, size, part, component, mechanism, device, step, process, method, reaction conditions, processing conditions, parameter, algorithm, signal, data, product or article of manufacture, etc.)” should be interpreted as including not only the expressly listed technical feature element, but also other technical feature elements that are not expressly listed and are well-known in the art.

[0030] The term "composed of" excludes any technical features not expressly listed. When used in a claim, it closes the claim to exclude all technical features other than those expressly listed, except for associated conventional impurities. If the term appears only in a clause of a claim, it limits the claim to the elements expressly listed in that clause; elements recited in other clauses are not excluded from the overall claim.

[0031] The following provides a detailed description of a robust causal relationship learning method, system, device, and storage medium provided by this invention. Contents not described in detail in the embodiments of this invention are prior art known to those skilled in the art. Where specific conditions are not specified in the embodiments of this invention, conventional conditions or conditions recommended by the manufacturer in the art shall apply. Reagents or instruments used in the embodiments of this invention, unless otherwise specified by the manufacturer, are all commercially available conventional products.

[0032] Example 1

[0033] This invention provides a robust causal relationship learning method. By performing a carefully designed conditional independence test (i.e., the causal cut finding scheme described later), a complex problem is divided into simpler subproblems, and existing causal algorithms are applied to these subproblems, ultimately improving the efficiency and accuracy of small sample data processing. This invention collects partial results from all subproblems and finally merges them to reveal the complete causal relationship between all variables. This invention not only possesses theoretical robustness, as it consistently returns correct and complete results under optimal settings, but also demonstrates superior scalability and effectiveness in experiments with both synthetic and real-world data. Figure 1 As shown, the method mainly includes the following steps:

[0034] 1. Data preprocessing.

[0035] In this embodiment of the invention, the input data containing the set of variables to be studied, V, is preprocessed to obtain the sample set D.

[0036] In this embodiment of the invention, the preprocessing of the input data containing the set of variables to be studied V includes: (1) cleaning the input data, such as removing outliers and filling in missing values; and (2) standardizing or normalizing the data to make different data on the same scale.

[0037] In this embodiment of the invention, the variables to be studied are terms of various types in the research field; the input data package contains multiple samples corresponding to the set of variables to be studied V, and each sample contains information content corresponding to the set of variables to be studied V.

[0038] For example, research areas may include gene therapy, economic forecasting, and advertising campaign design. Taking advertising campaign design as an example, the variables in the set V to be studied may include: advertising budget, advertising type, advertising frequency, target audience, season, advertising platform, click-through rate, conversion rate, user satisfaction, sales revenue, brand awareness, and market share, etc.

[0039] For example, the input data can be in the form of a list, with each row representing a set of samples and each column representing the value of each variable under study in the corresponding sample.

[0040] 2. Causal relationship learning based on sub-problem segmentation.

[0041] In this embodiment of the invention, a variable threshold θ is introduced. If the size (i.e. the number of variables) of the variable set V to be studied is not greater than the variable threshold θ, then the causal algorithm A can be run on the variable set V to be studied and the sample set D to obtain a directed causal graph.

[0042] If the size of the variable set V is greater than the variable threshold θ (i.e., the size of the variable set V does not meet the set requirements), then a causal cut (C, V1, V2) needs to be found on the variable set V. A causal relationship learning algorithm is then applied to V1∪C and the sample set D to obtain a directed causal graph G1. Similarly, a causal relationship learning algorithm is applied to V2∪C and the sample set D to obtain a directed causal graph G2. This part divides a complex problem into simpler subproblems. Specifically, the complex problem is to discover causality on the variable set V, and the subproblems are to discover causality on the variable sets V1∪C and V2∪C respectively.

[0043] The definition of causal cut is as follows: In a directed graph G = (V, E) on a set of variables V to be studied, if there exists a set of variables... It can derive three non-overlapping variable subsets V1, V2, and C of V, satisfying the following two conditions: (1) V1∪V2∪C=V; (2) There is no edge between V1 and V2 in E. Then, this variable set C is said to form a causal cut on graph G.

[0044] The steps for finding a causal cut (C,V1,V2) on the set of variables V to be studied include:

[0045] (1) Initialize variables. For k iterations, two variables u and v are randomly selected in each iteration. They are conditionally independent in the set of variables V to be studied, excluding u and v.

[0046] (2) Find the smallest subset: Find the smallest subset V^, which is a subset of V–{u,v} (i.e., the set of variables remaining after removing u and v from V), and makes u and v independent under the condition of subset V^.

[0047] (3) Initialize the variable subsets V1, V2 and C: initialize V1 = {u}, V2 = {v}, and C = V^.

[0048] (4) Remove variables: Remove variables from V1, V2 and C from the set of variables to be studied V.

[0049] (5) Variable classification: For each variable w remaining in the set of variables to be studied V, if there exists a subset for all u∈V1. If w and u are independent under condition C1, then add w to V2; if there exists a subset for all v∈V2. If w and v are independent under condition C2, then add w to V1; if variable w is not added to V1 or V2, then add w to C.

[0050] (6) Adjusting variables in the variable subset C: For each variable s in C, if for all u∈V1, there exists a subset... If s and u are independent under condition C3, then move s from C to V2; if a subset exists for all v∈V2. If s and v are independent under condition C4, then move s from C to V1.

[0051] (7) Record the results of each iteration: After each iteration, record the current result using Φj=(C,V1,V2).

[0052] (8) Output: Among all Φj, return the Φj that maximizes min{|V1|,|V2|} as the final causal cut.

[0053] 3. Merge the results of the cause-effect graph.

[0054] In this embodiment of the invention, the directed causal graph G1 and the directed causal graph G2 are merged to obtain the final directed causal graph G. The main steps are as follows:

[0055] (1) Merge G1 and G2 to obtain a directed causal graph G'.

[0056] (2) Sorting the edges: Sort the edges in the directed causal graph G' in descending order of significance.

[0057] (3) Mark variable pairs: Mark all variable pairs as unreachable.

[0058] (4) Process each edge: For each edge (v1→v2) in the directed causal graph G', if the variable pair (v1,v2) is unreachable, mark it as reachable; otherwise, remove the edge (v1→v2) from the directed causal graph G'.

[0059] (5) Check each edge: For each edge (v1→v2) in the directed causal graph G', if the directed causal graph G' contains the path v1→···→v2, then let Path(v1→v2) contain all variables involved in v1→···→v2 except v1 and v2, where the path v1→···→v2 contains variables v1 and v2 as well as other variables; if there exists a set of variables... If v1⊥v2|V′, then remove the edge (v1→v2) from G. v1⊥v2|V′ means that, given the values ​​of all variables in the set of variables V′, the values ​​of v1 and v2 are still independent.

[0060] (6) Return result: After the above processing, the final directed causal graph G is obtained.

[0061] like Figure 2 As shown, an example of a robust causal learning process involving subproblem segmentation and result merging is presented.

[0062] Taking the research fields of gene therapy, economic forecasting, and advertising campaign design introduced earlier as examples: (1) In gene therapy, various genes are research variables. A specific gene (called A) controls the expression of certain other genes (such as B), thereby controlling a specific physiological process. If the causal graph between genes cannot be obtained, from the experimental phenomenon alone, regulating the physiological process can adjust both gene A and gene B. However, in reality, choosing to adjust gene A may not be the optimal choice, because it may affect other physiological processes. By learning the specific causal graph, non-optimal decision-making can be avoided. (2) In economic forecasting, various factors are research variables. In a certain scenario, two phenomena (such as increased electricity consumption and increased ice cream sales) may be caused by hot weather at the same time. If the correct causal graph is not found, there may be conclusions that "increased electricity consumption" leads to "increased ice cream sales", or "increased ice cream sales" leads to "increased electricity consumption", thus misleading economic forecasting. (3) Advertising campaign design is similar. The combined effect of different designs is reflected in sales, and these designs have a positive synergistic effect or the opposite result on consumers. By accurately drawing a cause-and-effect diagram, we can correctly understand the interaction between different design elements, which is conducive to designing effective advertisements.

[0063] To facilitate the explanation of the solution process, a specific example is provided below. This example provides a detailed illustration of a robust causal relationship learning framework that involves sub-problem segmentation and result merging. It should be noted that the specific variables and their quantities involved in the following example are illustrative and not intended to constitute a limitation. In practical applications, users can select the necessary variables from the required domain based on their specific circumstances.

[0064] I. Data Preprocessing.

[0065] Step a1: Data cleaning.

[0066] First, let V be the set of variables to be studied. This set of variables includes 12 key nodes in advertising campaign design, such as advertising budget, ad type, ad frequency, target audience, season, advertising platform, click-through rate, conversion rate, user satisfaction, sales revenue, brand awareness, and market share. For the input data of these variables, necessary data cleaning is performed first, such as removing outliers and imputing missing values. For example, if the advertising budget data contains values ​​that significantly deviate from the normal range, such as negative or excessively large values, they are considered outliers and removed. For missing data, such as missing ad frequency data for certain advertising campaigns, interpolation or estimation methods can be used to impute the missing data.

[0067] Step a2: Data standardization or normalization.

[0068] Next, the input data for the variable set V is standardized or normalized. This is to ensure that all variables are on the same scale, facilitating subsequent calculations and comparisons. For example, advertising budgets might be measured in thousands or tens of thousands, while click-through rate (CTR) is a decimal between 0 and 1; the scale difference between the two is significant. Therefore, standardization or normalization can transform all variables to the same scale. Specific methods could include subtracting the mean and dividing by the standard deviation (standardization) or scaling the values ​​to the 0-1 range (normalization).

[0069] Step a3: Data sample representation.

[0070] Finally, let D denote the processed data sample. This is the dataset that will be used in subsequent steps.

[0071] II. Finding the cause-and-effect divide.

[0072] In this section, we will find the causal cut (C, V1, V2) based on the sample set D and the variable set V. According to the definition of a causal cut, we need to find a subset C such that V can be partitioned into three non-overlapping subsets V1, V2, and C, satisfying V1∪V2∪C=V, and there is no edge between V1 and V2.

[0073] Taking advertising campaign design as an example, let V be a set of 12 key nodes, which include advertising budget, advertising type, advertising frequency, target audience, season, advertising platform, click-through rate, conversion rate, user satisfaction, sales, brand awareness, and market share.

[0074] Step b1: Initialize variables.

[0075] For k iterations, two variables u and v are randomly selected in each iteration, such as 'ad budget' and 'click rate', which are conditionally independent in the set V other than 'ad budget' and 'click rate'.

[0076] Step b2: Find the smallest subset.

[0077] We need to find the smallest subset V^, which is a subset of V - 'ad budget' and 'click-through rate', such that 'ad budget' and 'click-through rate' are independent under the condition of V^. Let's assume that the smallest subset V^ is {'ad type', 'ad platform'}.

[0078] Step b3: Initialize V1, V2, and C.

[0079] Next, initialize V1 = {'Ad Budget'}, V2 = {'Click-Through Rate'}, and C = {'Ad Type', 'Ad Platform'}.

[0080] Step b4: Remove the variable.

[0081] Remove the variables 'ad budget', 'click rate', 'ad type', and 'ad platform' from set V in V, specifically from V1, V2, and C.

[0082] Step b: Variable classification.

[0083] For each remaining variable w in V, it needs to be categorized. For example, for the variable 'target population', if it is independent of all variables in V1 under the condition of a subset of C, then add it to V2. If it is independent of all variables in V2 under the condition of a subset of C, then add it to V1. If neither condition is met, then add it to C.

[0084] Step b6: Adjust the variables in C.

[0085] For each variable s in C, adjustments are needed. For example, for the variable 'ad type', if it is independent of all variables in V1 under the condition of a subset of C-{'ad type'}, then move it from C to V2. If it is independent of all variables in V2 under the condition of a subset of C-{'ad type'}, then move it from C to V1.

[0086] Step b7: Record the results of each iteration.

[0087] After each iteration, the current result is recorded using Φj=(C,V1,V2).

[0088] Step b8: Output the results.

[0089] Of all Φj, return the one that maximizes min{|V1|,|V2|} as the final causal cut. This is the optimal causal cut found in the advertising campaign design.

[0090] III. Learning the causal relationship between sub-problem segmentation and result merging.

[0091] This section will use an example of advertising campaign design to demonstrate causal relationship learning. The input includes a sample set D, a variable set V, variable thresholds θ, and a causal algorithm A. The goal of this process is to output a directed causal graph G.

[0092] Step c1: Determine the size of the variable set.

[0093] First, we need to determine the size of the variable set V. In the example of advertising campaign design, suppose the threshold θ is set to 5, and the variable set V contains 12 key nodes. Therefore, the size of the variable set V is greater than the threshold θ.

[0094] Step c2: Find the causal cut.

[0095] Since the size of the variable set V is greater than the threshold θ, a causal cut (C, V1, V2) needs to be found on V. The method for finding the causal cut was introduced in the previous section. The found causal cuts are V1 = {'Ad budget', 'Target audience'}, V2 = {'Click-through rate', 'Sales revenue'} and C = {'Ad type', 'Ad platform'}.

[0096] Step c3: Recursively apply the causal relationship learning algorithm.

[0097] Next, we need to apply the causal relationship learning algorithm to V1∪C and V2∪C respectively to obtain directed causal graphs G1 and G2.

[0098] Applying a causal learning algorithm to V1∪C={'Ad Budget', 'Target Audience', 'Ad Type', 'Ad Platform'} yields a directed causal graph G1. Similarly, applying the causal learning algorithm to V2∪C={'Click-Through Rate', 'Sales Revenue', 'Ad Type', 'Ad Platform'} yields a directed causal graph G2.

[0099] Step c4: Merge the graphs.

[0100] The final directed causal graph G needs to be obtained by merging G1 and G2. This merging process will be explained in detail in the next section.

[0101] By following the steps above, a set of causal cuts can be found in the variable set V, and causal relationships can be learned based on these cuts, ultimately resulting in a directed causal graph G, which can help to better understand and design advertising campaigns.

[0102] IV. Merging the results of the cause-effect graph.

[0103] This section explains how to combine the results of two cause-effect graphs in an example of advertising campaign design to obtain a complete cause-effect graph. The inputs are G1 and G2, which are solutions to V1∪C and V2∪C respectively, and the output is G, which is the solution to V1∪V2∪C.

[0104] Step d1: Input and output definition.

[0105] First, we have obtained two causal graphs, G1 and G2, which are the solutions to V1∪C={'Ad Budget','Target Audience','Ad Type','Ad Platform'} and V2∪C={'Click-Through Rate','Sales Revenue','Ad Type','Ad Platform'}, respectively.

[0106] Step d2: Merge solutions.

[0107] Next, G1 and G2 need to be merged to obtain a causal graph G' containing V1∪V2∪C={'ad budget','target audience','click rate','sales revenue','ad type','ad platform'}.

[0108] Step d3: Sort the edges.

[0109] The edges in G need to be sorted in descending order of significance. Significance can be obtained through statistical tests or other methods.

[0110] Step d4: Mark the variable pairs.

[0111] After sorting, all variable pairs are marked as unreachable. This indicates that a causal relationship between any two variables has not yet been determined.

[0112] Step d5: Process each edge.

[0113] For each edge (v1→v2) in G', if (v1,v2) is reachable, then remove this edge from G. Otherwise, mark (v1,v2) as reachable.

[0114] Step d6: Inspect each edge.

[0115] For each edge v1→v2 in G', if there exists a path v1→···→v2 in G, let Path(v1→v2) contain all the variables involved in v1→···→v2. If v1⊥v2|V′, then remove this edge from G. This step ensures that every edge in G represents a strong causal relationship, meaning that the two endpoints of the edge are not independent given any other variables.

[0116] Step d7: Return the result.

[0117] Finally, we return to the final directed causal graph G. This directed causal graph G represents the causal relationships found in the advertising campaign design, including the relationships between key factors such as 'ad budget', 'target audience', 'click-through rate', 'sales', 'ad type', and 'ad platform'.

[0118] For example, sample data might show that both increasing ad budgets and changing ad types increase sales, making changing ad types a viable decision. However, a causal graph might indicate that ad type does not affect sales. The actual causal relationship might be that increasing ad budgets, while increasing sales, makes the cost of streaming ads acceptable; therefore, streaming ads are often used with high ad budgets, and high budgets increase sales. Thus, an accurate causal graph can help users make the right decisions.

[0119] Through the above description of the embodiments, those skilled in the art can clearly understand that the above embodiments can be implemented by software, or by using software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solutions of the above embodiments can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, mobile hard drive, etc.), including several instructions to cause a computer device (such as a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0120] Example 2

[0121] This invention also provides a robust causal relationship learning system, which is mainly used to implement the methods provided in the foregoing embodiments, such as... Figure 3 As shown, the system mainly includes:

[0122] The preprocessing unit is used to preprocess the input data containing the set of variables to be studied V to obtain the sample set D; wherein, the variables to be studied are various types of terms in the research field.

[0123] The subproblem segmentation and causal relationship learning unit is used to find a causal cut (C, V1, V2) on the set of variables to be studied when the size of the set of variables to be studied does not meet the set requirements. Then, it applies the causal relationship learning algorithm on V1∪C and the sample set D to obtain a directed causal graph G1, and applies the causal relationship learning algorithm on V2∪C and the sample set D to obtain a directed causal graph G2. Here, V1, V2 and C are three non-overlapping variable subsets of the set of variables to be studied V, and V1∪V2∪C=V, and the variable subsets V1 and V2 have no edges in the directed graph of the set of variables to be studied V.

[0124] The result merging unit is used to merge the directed causal graph G1 and the directed causal graph G2 to obtain the final directed causal graph G.

[0125] In this embodiment of the invention, the preprocessing of the input data containing the set of variables to be studied V includes: cleaning the input data and then standardizing or normalizing it.

[0126] The input data packet contains multiple samples corresponding to the set of variables to be studied, V, and each sample contains information content corresponding to the set of variables to be studied, V.

[0127] In this embodiment of the invention, the step of finding a causal cut (C, V1, V2) on the set of variables to be studied V includes:

[0128] Initialize variables: For k iterations, in each iteration, two variables u and v are randomly selected, which are conditionally independent in the set of variables V to be studied, excluding u and v.

[0129] Find the smallest subset: Find the smallest subset V^, which is a subset of V–u,v, such that u and v are independent under the condition of subset V^;

[0130] Initialize the variable subsets V1, V2 and C: initialize V1 = {u}, V2 = {v}, C = V^;

[0131] Remove variables: Remove variables V1, V2, and C from the set of variables to be studied, V;

[0132] Variable classification: For each variable w remaining in the set of variables to be studied V, if for all u∈V1, there exists a subset... If w and u are independent under condition C1, then add w to V2; if there exists a subset for all v∈V2. If w and v are independent under condition C2, then add w to V1; if variable w is not added to V1 or V2, then add w to C.

[0133] Adjusting variables in a subset C: For each variable s in C, if for all u∈V1, there exists a subset... If s and u are independent under condition C3, then move s from C to V2; if a subset exists for all v∈V2. If s and v are independent under condition C4, then move s from C to V1;

[0134] Record the result of each iteration: After each iteration, record the current result using Φj = (C, V1, V2);

[0135] Output: Among all Φj, return the Φj that maximizes min{|V1|,|V2|} as the final causal cut.

[0136] In this embodiment of the invention, merging directed causal graph G1 and directed causal graph G2 to obtain the final directed causal graph G includes:

[0137] Merging G1 and G2 yields a directed causal graph G';

[0138] Edge sorting: Sort the edges in the directed causal graph G' in descending order of significance;

[0139] Mark variable pairs: Mark all variable pairs as unreachable;

[0140] Process each edge: For each edge (v1→v2) in the directed causal graph G', if the variable pair (v1,v2) is unreachable, mark it as reachable; otherwise, remove the edge (v1→v2) from the directed causal graph G'.

[0141] Examine each edge: For each edge (v1→v2) in the directed causal graph G', if the directed causal graph G' contains the path v1→···→v2, then let Path(v1→v2) contain all variables involved in v1→···→v2 except v1 and v2, where the path v1→···→v2 contains variables v1 and v2 as well as other variables; if a set of variables exists... If v1⊥v2|V′, then remove the edge (v1→v2) from G. v1⊥v2|V′ means that, given the values ​​of all variables in the set of variables V′, the values ​​of v1 and v2 are still independent.

[0142] Return result: After the above processing, the final directed causal graph G is obtained.

[0143] Since the specific technical details involved in this system have been described in detail in the previous embodiment 1, they will not be repeated here.

[0144] Those skilled in the art will understand that, for the sake of convenience and brevity, the above-described division of functional modules is used as an example. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the system can be divided into different functional modules to complete all or part of the functions described above.

[0145] Example 3

[0146] The present invention also provides a processing device, such as Figure 4 As shown, it mainly includes: one or more processors; a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the method provided in the foregoing embodiments.

[0147] Furthermore, the processing device also includes at least one input device and at least one output device; in the processing device, the processor, memory, input device, and output device are connected via a bus.

[0148] In this embodiment of the invention, the specific types of the memory, input device, and output device are not limited; for example:

[0149] Input devices can be touchscreens, image acquisition devices, physical buttons, or mice, etc.

[0150] The output device can be a display terminal;

[0151] The memory can be random access memory (RAM) or non-volatile memory, such as disk storage.

[0152] Example 4

[0153] The present invention also provides a readable storage medium storing a computer program that, when executed by a processor, implements the method provided in the foregoing embodiments.

[0154] In this embodiment of the invention, the readable storage medium is a computer-readable storage medium and can be disposed in the aforementioned processing device, for example, as a memory in the processing device. Furthermore, the readable storage medium can also be any medium capable of storing program code, such as a USB flash drive, portable hard drive, read-only memory (ROM), magnetic disk, or optical disk.

[0155] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A robust causal relationship learning method, characterized in that, include: The input data containing the set of variables to be studied, V, is preprocessed to obtain the sample set D; wherein, the variables to be studied are various types of terms in the research field. If the size of the set of variables to be studied, V, does not meet the set requirements, then a causal cut (C, V1, V2) is found on the set of variables to be studied, and a causal relationship learning algorithm is applied to V1 ∪ C and the sample set D to obtain a directed causal graph G1, and a causal relationship learning algorithm is applied to V2 ∪ C and the sample set D to obtain a directed causal graph G2; wherein V1, V2, and C are three non-overlapping subsets of variables in the set of variables to be studied, and V1 ∪ V2 ∪ C = V, and the subsets V1 and V2 have no edges in the directed graph of the set of variables to be studied, V; The directed causal graphs G1 and G2 are merged to obtain the final directed causal graph G. When the research field is advertising campaign design, the variables in the set of variables to be studied V include: advertising budget, advertising type, advertising frequency, target audience, season, advertising platform, click-through rate, conversion rate, user satisfaction, sales, brand awareness and market share. The directed causal graph G represents the causal relationship found in advertising campaign design. The steps for finding the causal cut (C, V1, V2) on the set of variables to be studied V include: Initialize variables: For k iterations, in each iteration, two variables u and v are randomly selected, which are conditionally independent in the set of variables V to be studied, excluding u and v. Find the smallest subset: Find the smallest subset Vˆ, which is a subset of V - u, v, such that u and v are independent under the condition of subset Vˆ; Initialize the variable subsets V1, V2 and C: initialize V1 = {u}, V2 = {v}, and C = Vˆ; Remove variables: Remove variables V1, V2, and C from the set of variables to be studied, V; Variable classification: For each variable w remaining in the set of variables to be studied V, if for all u ∈ V1, there exists a subset C′⊆C such that w and u are independent under the condition of C′, then add w to V2; if for all v ∈ V2, there exists a subset C′⊆C such that w and v are independent under the condition of C′, then add w to V1; otherwise, add w to C. Adjusting variables in a subset C: For each variable s in C, if for all u ∈ V1, there exists a subset C′⊆C - {s} such that s and u are independent under the condition of C′, then move s from C to V2; if for all v ∈ V2, there exists a subset C′⊆C - {s} such that s and v are independent under the condition of C′, then move s from C to V1. Record the result of each iteration: After each iteration, record the current result using Φj = (C, V1, V2); Output: Among all Φj, return the Φj that maximizes min{|V1|, |V2|} as the final causal cut; The step of merging directed causal graph G1 and directed causal graph G2 to obtain the final directed causal graph G includes: Merging G1 and G2 yields a directed causal graph G'; Edge sorting: Sort the edges in the directed causal graph G' in descending order of significance; Mark variable pairs: Mark all variable pairs as unreachable; Process each edge: For each edge (v1→v2) in the directed causal graph G', if the variable pair (v1, v2) is reachable, remove the edge (v1→v2) from the directed causal graph G'; otherwise, mark the variable pair (v1, v2) as reachable. Examine each edge: For each edge (v1→v2) in the directed causal graph G', if v1 → · · · → v2 is in the directed causal graph G', then let Path(v1 → v2) contain all the variables involved in v1 → · · · → v2; if there exists V′⊂ Path(v1 → v2) such that v1⊥v2|V′, then remove the edge (v1→v2) from G; Return result: After the above processing, the final directed causal graph G is obtained.

2. The robust causal relationship learning method according to claim 1, characterized in that, The preprocessing of input data containing the set of variables to be studied V includes: cleaning the input data and then standardizing or normalizing it. The input data packet contains multiple samples corresponding to the set of variables to be studied, V, and each sample contains information content corresponding to the set of variables to be studied, V.

3. A robust causal relationship learning system, characterized in that, include: The preprocessing unit is used to preprocess the input data containing the set of variables to be studied V to obtain the sample set D; wherein, the variables to be studied are various types of terms in the research field. The subproblem segmentation and causal relationship learning unit is used to find causal cuts (C, V1, V2) on the set of variables to be studied when the size of the set of variables to be studied does not meet the set requirements. Then, it applies the causal relationship learning algorithm on V1 ∪ C and the sample set D to obtain a directed causal graph G1, and applies the causal relationship learning algorithm on V2 ∪ C and the sample set D to obtain a directed causal graph G2. Here, V1, V2 and C are three non-overlapping subsets of variables in the set of variables to be studied V, and V1 ∪ V2 ∪ C = V, and the subsets V1 and V2 have no edges in the directed graph of the set of variables to be studied V. The result merging unit is used to merge the directed causal graph G1 and the directed causal graph G2 to obtain the final directed causal graph G. When the research field is advertising campaign design, the variables in the set of variables to be studied V include: advertising budget, advertising type, advertising frequency, target audience, season, advertising platform, click-through rate, conversion rate, user satisfaction, sales, brand awareness and market share. The directed causal graph G represents the causal relationship found in the advertising campaign design. The steps for finding the causal cut (C, V1, V2) on the set of variables to be studied V include: Initialize variables: For k iterations, in each iteration, two variables u and v are randomly selected, which are conditionally independent in the set of variables V to be studied, excluding u and v. Find the smallest subset: Find the smallest subset Vˆ, which is a subset of V - u, v, such that u and v are independent under the condition of subset Vˆ; Initialize the variable subsets V1, V2 and C: initialize V1 = {u}, V2 = {v}, and C = Vˆ; Remove variables: Remove variables V1, V2, and C from the set of variables to be studied, V; Variable classification: For each variable w remaining in the set of variables to be studied V, if for all u ∈ V1, there exists a subset C′⊆C such that w and u are independent under the condition of C′, then add w to V2; if for all v ∈ V2, there exists a subset C′⊆C such that w and v are independent under the condition of C′, then add w to V1; otherwise, add w to C. Adjusting variables in a subset C: For each variable s in C, if for all u ∈ V1, there exists a subset C′⊆C - {s} such that s and u are independent under the condition of C′, then move s from C to V2; if for all v ∈ V2, there exists a subset C′⊆C - {s} such that s and v are independent under the condition of C′, then move s from C to V1. Record the result of each iteration: After each iteration, record the current result using Φj = (C, V1, V2); Output: Among all Φj, return the Φj that maximizes min{|V1|, |V2|} as the final causal cut; The step of merging directed causal graph G1 and directed causal graph G2 to obtain the final directed causal graph G includes: Merging G1 and G2 yields a directed causal graph G'; Edge sorting: Sort the edges in the directed causal graph G' in descending order of significance; Mark variable pairs: Mark all variable pairs as unreachable; Process each edge: For each edge (v1→v2) in the directed causal graph G', if the variable pair (v1, v2) is reachable, remove the edge (v1→v2) from the directed causal graph G'; otherwise, mark the variable pair (v1, v2) as reachable. Examine each edge: For each edge (v1→v2) in the directed causal graph G', if v1 → · · · → v2 is in the directed causal graph G', then let Path(v1 → v2) contain all the variables involved in v1 → · · · → v2; if there exists V′⊂ Path(v1 → v2) such that v1⊥v2|V′, then remove the edge (v1→v2) from G; Return result: After the above processing, the final directed causal graph G is obtained.

4. A robust causal relationship learning system according to claim 3, characterized in that, The preprocessing of input data containing the set of variables to be studied V includes: cleaning the input data and then standardizing or normalizing it. The input data packet contains multiple samples corresponding to the set of variables to be studied, V, and each sample contains information content corresponding to the set of variables to be studied, V.

5. A processing device, characterized in that, include: One or more processors; Memory, used to store one or more programs; Wherein, when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the method as described in any one of claims 1 to 2.

6. A readable storage medium storing a computer program, characterized in that, When a computer program is executed by a processor, it implements the method as described in any one of claims 1 to 2.

Images