Knowledge-guided causal discovery method, device, equipment and medium

Abstract

The present invention provides a knowledge-guided causal discovery method, apparatus, device and medium, the method comprising: randomly extracting n samples from m samples of the enterprise macro monthly indicator data set X to construct a training sample set #imgabs0# using an encoder to convert each training sample x j Embedded into state s j , based on all embedded states, the complete state space #imgabs1# is constructed. The decoder is used to map the state space #imgabs2# to the action space A and output the corresponding variables. After all variables are selected, the initial sorting Π is formed. The idea of ​​direct insertion sorting is used to adjust the initial sorting Π to obtain the adjusted sorting Π'. The optimal sorting Π is obtained by comparing the knowledge reward function. * ; According to the optimal sort π * Compare it with the initial sorting π found by the search, take the changed node relationship as prior knowledge, continuously update the prior knowledge through the iterative process of search, and guide reinforcement learning to find the global optimal solution; transform the global optimal solution to obtain the causal structure.

Description

Technical Field

[0001] The present invention relates to the field of causal discovery, and in particular to a knowledge-guided causal discovery method, apparatus, device and medium. Background Art

[0002] The core of causal discovery lies in identifying causal relationships between data features through fitting methods and measured data. Uncovering causal relationships between variables plays a crucial role in decision-making, guiding interventions to solve problems, and optimizing resources in fields such as biomedicine, public policy evaluation, and genetics. This is particularly true in business scenarios such as customer churn analysis, where qualitative methods can be used to identify the drivers and key factors of customer churn, helping companies stabilize their customer base. For example, by identifying causal relationships among data features within monthly macroeconomic indicators captured by banks' ABM systems, key factors influencing these indicators can be identified, enabling precise regulation. Therefore, constructing an accurate causal structure provides a framework for inferring and predicting causal effects between events, helping to identify and distinguish causal relationships from correlations and identifying key variables influencing outcomes.

[0003] Causal structures are represented by directed acyclic graphs (DAGs), which consist of variables and arrows representing the causal relationships between them. The main goal of causal discovery is to recover the underlying causal structure and the associated conditional probability distribution. Methods for learning causal relationships from passive observational data include constraint-based methods, methods based on causal function models, and methods based on scoring functions and search.

[0004] Constraint-based methods formulate the structure learning problem as a constraint satisfaction problem, constructing the network structure through conditional independence between different variables. Some notable algorithms in this category are the PC algorithm, the grow-shrink (GS) algorithm, and the IAMB algorithm. Constraint-based methods can handle a wider range of data types and distributions, and they are computationally efficient, making them highly interpretable. However, the accuracy of the learning process depends on the number of conditional independence tests performed and the size of the constraint set. They are sensitive to conditional independence tests and data noise, and high-order dependencies are unreliable for large networks and complex data.

[0005] To address the inability of traditional constraint-based methods to identify Markov equivalence classes, some studies have proposed causal function models based on the structural equation model (SEM), drawing on the data distribution characteristics of causal mechanisms. Representative algorithms include the linear non-Gaussian acyclic model (LiNGAM), the post-nonlinear (PNL) method, the additive noise model (ANM) under nonlinear conditions, and corresponding extensions. Causal function models rely heavily on strict assumptions about the causal mechanisms underlying the model and are only applicable to low-dimensional data scenarios.

[0006] Scoring function-based methods use a scoring function to determine the fit between the network structure and the data, and then use a search algorithm to find the optimal structure. Common scoring functions include BIC, BDeu, MDL, and AIC. The search space can be divided into three types: DAG, equivalence class, and node order search space. Most structure learning methods search in the DAG space, but the DAG search space grows exponentially with the number of nodes. Compared to the DAG search space, the equivalence class search space is smaller. The equivalence class search space is partitioned based on Markov equivalence classes, but it is difficult to judge and is complex.

[0007] The node-order search space is partitioned according to the topological order of nodes. This significantly reduces the search scope compared to the DAG space and eliminates the need for acyclicity checks. However, in the node-order search space, the algorithm's accuracy is highly dependent on the node order. Many studies have proposed improving algorithm performance by employing heuristic search strategies, such as genetic algorithms, particle swarm algorithms, ant colony algorithms, and bee colony algorithms. However, due to the heuristic nature of these algorithms, they often become trapped in local optima.

[0008] Currently, scholars are primarily focused on exploring scoring functions and innovating search methods, introducing a large number of strategies and heuristic algorithms to improve search efficiency and escape the constraints of local optimal solutions. However, current scoring search algorithms are still limited by search boundaries and space. Some scoring algorithms are prone to overfitting and underfitting when learning from small samples, resulting in poor efficiency. Furthermore, current scoring functions have scoring equivalence, and the searched network structures cannot overcome the problem of Markov equivalence classes, resulting in a large number of reversed edges. Recently, some studies have used smooth scoring functions and non-cyclic smooth representations to transform combinatorial optimization problems into continuous optimization problems based on gradients, including NOTEARS for linear causal models and subsequent extensions to nonlinear causal relationships.

[0009] Reinforcement learning (RL) has been applied to several combinatorial optimization problems, such as the maximum cut and traveling salesman problems. These efforts aim to learn a policy that serves as a solver for specific types of combinatorial problems. As an alternative to gradient-based optimization methods, some studies leverage the powerful interaction between reinforcement learning and the environment to model causal relationships as Markov decision problems. This approach eliminates the need for smooth scoring functions and can find the optimal network structure within a latent graph space. However, due to the large action space and the significant computational time wasted on acyclicity checks, this approach only achieves good results on problems with small nodes and consumes excessive computational time. To address these issues, some studies have proposed the CORL method, which leverages the reduced space of variable search ranking. This method integrates reinforcement learning into the ranking-based paradigm, achieving high performance and training speed on small-scale problems. However, on large-scale problems, this method suffers from slow training speed and excessively high total iteration counts, wasting significant training resources. Summary of the Invention

[0010] In view of this, the embodiments of the present disclosure provide a knowledge-guided causal discovery method, apparatus, device, and medium, which at least partially solve the problems existing in the prior art.

[0011] In a first aspect, the present disclosure provides a knowledge-guided causal discovery method, comprising:

[0012] S101: Obtain a monthly enterprise macroeconomic indicator dataset X, and randomly select n samples from the m samples of the monthly enterprise macroeconomic indicator dataset X to construct a training sample set.

[0013] S102, using the encoder to train the sample set Each training sample x in j Embedded into state s j , forming a complete state space based on all embedded states Among them, the state space The tth element of represents the actual state encountered at the tth decision step when generating the sort;

[0014] S103, using the decoder, by decoding the state, the state space Map to the action space A and output the corresponding variables. In each step, the selected variables are masked out and the remaining unselected variables are output to ensure that the variables are selected once. After all variables are selected, the initial ranking π is formed.

[0015] S104, use the idea of ​​direct insertion sort to adjust the initial sort Π to obtain the adjusted sort Π', and obtain the final optimal sort Π by comparing the knowledge reward function* ;

[0016] S105, according to the optimal sorting π * Compare it with the initial ranking π found by the search, take the changed node relationship as prior knowledge, continuously update the prior knowledge through the iterative process of search, and use the prior knowledge to obtain effective knowledge that reinforcement learning has not obtained in the exploration process to constrain the direction of subsequent reinforcement learning exploration and guide reinforcement learning to find the global optimal solution;

[0017] S106, transforming the global optimal solution to obtain an initial causal structure, pruning the initial causal structure to obtain a final causal structure, and predicting the data to be predicted based on the final causal structure to obtain causal relationships, so as to obtain key factors affecting enterprise indicator data.

[0018] According to a specific implementation of the embodiment of the present disclosure, in step S102, the training sample set d is the total number of variables, n is the size of the training sample; the encoder is represented as is the parameter of the encoder; each state s j Set in the same dimension, i.e. s j ∈R n ; All states embedded by training samples constitute the state space S: = {s1, ..., s d}; Set an initial state s0, represented by s0 to select the first action, then the complete state space

[0019] According to a specific implementation of the embodiment of the present disclosure, the decoder is represented as is the parameter of the encoder, and the action output by the decoder is the variable.

[0020] According to a specific implementation of the embodiment of the present disclosure, in step S104, the knowledge reward function R k Expressed as:

[0021] R k =S BIC (G Π )-S BIC (G Π )

[0022] Among them, S BIC (G Π ) represents the dense reward of ranking π; assuming that the variance of Gaussian noise is equal in each variable, we get:

[0023]

[0024] in represents the residual sum of squares of the i-th variable, is the least squares loss, m is the sample size, and EGΠ is the number of edges in the graph generated by sorting Π.

[0025] According to a specific implementation of the embodiment of the present disclosure, in S105, the prior knowledge P k The expression is:

[0026] P k =P k ∪(X i ,X j ); if I Π *(X i ,X j )≠I Π (X i ,X j )

[0027] in, Represents node X i and X j In the optimal sorting π * The relative index position in ;

[0028] Then S105 specifically includes:

[0029] Use policy gradient to optimize the parameters of the ranking generation model; where the policy gradient is:

[0030]

[0031] Among them, J(φ) represents the maximization strategy, φ={φ e ,φ d}, T is the total number of decision steps, E φ represents the expected value of the policy φ, which means the total return R(τ) of the trajectory τ multiplied by the logarithmic policy gradient under the current policy; τ represents a trajectory, usually refers to the policy π φ A complete path of executing a sequence of actions from the initial state to the termination; Indicates that in strategy π φ In state Next select action a t The probability value of

[0032] The model parameters are trained by the Actor-Critic algorithm, with the current state as input and the output as the probability distribution of the action or the continuous action value. The Critic network then evaluates the quality of the action and adjusts the strategy. Among them, the Actor is the policy network, which consists of an encoder and a decoder; the Critic is the evaluation network, which uses a three-layer feedforward neural network with 512, 256 and 1 units, with the state is input, output given state The predicted value of the current strategy; wherein, during the training process, the penalty weight W o To constrain the optimization direction of the next round of sorting:

[0033]

[0034] where w i Corresponding to the node X in the sort π i The penalty weight of the sorted Π is calculated based on the prior knowledge obtained from the previous round of exploration. i ,X j ) index position conflicts with the parent node X in prior knowledge i Assign penalty weight w i , the remaining nodes have penalty weights w i Set to 0; the penalty weight value of the conflict node is obtained through experimental comparison, that is:

[0035]

[0036] Represented as node X i and X j In prior knowledge p k The relative index position in ;

[0037] According to the penalty weight W of the ranking Π o , set the Actor's loss function Loss Actor And Critic's loss function Loss Critic To constrain the direction of algorithm optimization:

[0038]

[0039] Among them, π * represents the global optimal solution, s t and a t Represents state and action, G t is the experience return, defined as the cumulative reward from time step t to a certain period in the future: G t =R t+1 +γR t+2 +…+γ n-1 Rt+n +γ n Q)s t+n ,a t+n ), γ represents the discount coefficient, N represents the number of training samples, R t+1 represents the reward obtained at time step t+1, Q(s t ,a t ) means in state s t Next select action a t The action value function is the current estimated action value function, V(s t ) means in state s t The state value function under .

[0040] According to a specific implementation method of an embodiment of the present disclosure, in step S106, linear regression is applied to the obtained fully connected initial causal structure, and then the edges with smaller weights are pruned using a threshold to obtain the final causal structure; for nonlinear models, CAM pruning is used to prune the initial causal structure to obtain the final causal structure.

[0041] According to a specific implementation of the embodiment of the present disclosure, the method further includes:

[0042] For each variable X j , for X j A generalized additive model is fitted to the current parents, and then the covariates are tested for significance. If the threshold p-value of the significance test is less than or equal to 0.001, the covariates are significant.

[0043] In a second aspect, the present disclosure provides a knowledge-guided causal discovery device, comprising:

[0044] The training sample construction unit is used to obtain the enterprise macro monthly indicator dataset X, and randomly select n samples from the m samples of the enterprise macro monthly indicator dataset X to construct a training sample set.

[0045] The encoding unit is used to use the encoder to convert the training sample set Each training sample x in j Embedded into state s j , forming a complete state space based on all embedded states Among them, the state space The tth element of represents the actual state encountered at the tth decision step when generating the variable ranking;

[0046] The decoding unit is used to decode the state space by using a decoder. Map to the action space A and output the corresponding variables. In each step, the selected variables are masked out and the remaining unselected variables are output to ensure that the variables are selected once. After all variables are selected, the initial ranking π is formed.

[0047] The sorting adjustment unit is used to adjust the initial sort Π to obtain the adjusted sort Π' by using the idea of ​​direct insertion sorting, and obtain the final optimal sort Π by comparing the knowledge reward function * ;

[0048] Reinforcement learning unit for optimal sorting π * Compare it with the initial ranking π found by the search, take the changed node relationship as prior knowledge, continuously update the prior knowledge through the iterative process of search, and use the prior knowledge to obtain effective knowledge that reinforcement learning has not obtained in the exploration process to constrain the direction of subsequent reinforcement learning exploration and guide reinforcement learning to find the global optimal solution;

[0049] A pruning unit is used to transform the global optimal solution to obtain an initial causal structure, prune the initial causal structure to obtain a final causal structure, and predict the data to be predicted based on the final causal structure to obtain a causal relationship, so as to obtain key factors affecting enterprise indicator data.

[0050] In a third aspect, an embodiment of the present disclosure provides a knowledge-guided causal discovery device, comprising:

[0051] at least one processor; and,

[0052] a memory communicatively connected to the at least one processor; wherein,

[0053] The memory stores instructions that can be executed by the at least one processor. The instructions are executed by the at least one processor to enable the at least one processor to perform the above-mentioned knowledge-guided causal discovery method.

[0054] In a fourth aspect, an embodiment of the present disclosure provides a non-transitory computer-readable storage medium, which stores computer instructions for causing the computer to execute the knowledge-guided causal discovery method as described above.

[0055] In summary, compared with the prior art, this embodiment has at least the following improvements:

[0056] (1) To avoid dealing with acyclic constraints and reduce the size of the search space, this embodiment formulates the combinatorial optimization problem of directed graphs as a variable ranking search problem of a multi-step Markov decision process, and implements the variable ranking generation process in the decoder-encoder architecture.

[0057] (2) In order to reduce the complexity of the interactive environment and the existence of uncertain information, a knowledge acquisition method is designed. The prior knowledge acquired iteratively is used to adaptively assign weights to samples in the training batch, guide the search direction of reinforcement learning in complex environments, and use reinforcement learning to optimize based on the reward function and the weights of the training samples;

[0058] (3) In order to improve the training speed of the algorithm, an asynchronous parallel framework is designed, and a dynamic threshold is set according to the reward scores ranked according to different course stages, so that the method can dynamically adjust the number of iterations and reduce unnecessary iterations.

[0059] This example conducted comparative experiments on synthetic and real datasets for problems of different scales, using metrics including true positive rate, structural Hamming distance, F1 score, and runtime. The experiments demonstrated that this example outperformed other algorithms across all metrics for problems of varying scales. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] In order to more clearly illustrate the technical solutions of the embodiments of the present disclosure, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present disclosure. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0061] Figure 1 A schematic diagram of a causal network diagram consisting of variables and the causal dependencies between them;

[0062] Figure 2 This is an example graph for topological sorting;

[0063] Figure 3 Schematic diagram of the principle of reinforcement learning;

[0064] Figure 4 A schematic diagram of a process flow of a knowledge-guided causal discovery method provided by the first embodiment of the present invention;

[0065] Figure 5 This is the architecture diagram of the Transformer encoder;

[0066] Figure 6 This is the architecture diagram of the LSTM decoder;

[0067] Figure 7 Schematic diagram for sorting adjustment;

[0068] Figure 8 Schematic diagram for prior knowledge acquisition;

[0069] Figure 9 A schematic diagram of the structure of a knowledge-guided causal discovery device provided in a second embodiment of the present invention;

[0070] Figure 10 A schematic diagram of a method or device that can implement an embodiment of the present invention. DETAILED DESCRIPTION

[0071] The embodiments of the present disclosure are described in detail below with reference to the accompanying drawings.

[0072] The following describes the embodiments of the present disclosure through specific examples, and those skilled in the art can easily understand other advantages and effects of the present disclosure from the contents disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present disclosure, rather than all of the embodiments. The present disclosure can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present disclosure. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in the present disclosure, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present disclosure.

[0073] It should be noted that various aspects of the embodiments within the scope of the appended claims are described below. It should be apparent that the aspects described herein can be embodied in a wide variety of forms, and any specific structure and / or function described herein is merely illustrative. Based on this disclosure, it should be understood by those skilled in the art that an aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects described herein can be used to implement an apparatus and / or practice a method. In addition, other structures and / or functionalities other than one or more of the aspects described herein can be used to implement this apparatus and / or practice this method.

[0074] It should also be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present disclosure. The illustrations only show components related to the present disclosure and are not drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component can be changed at will, and the component layout type may also be more complicated.

[0075] Additionally, in the following description, specific details are provided to provide a thorough understanding of the examples. However, one skilled in the art will appreciate that the aspects described can be practiced without these specific details.

[0076] Multiple embodiments of the present invention provide a knowledge-guided causal discovery method, apparatus, device, and medium. These methods utilize the effective knowledge gained during model learning to guide reinforcement learning for automated search in uncertain and complex environments, thereby improving search efficiency and performance. To facilitate understanding of the present invention, the following introduces some of the technologies involved.

[0077] 1. Directed Acyclic Graph (DAG)

[0078] DAG is represented as G = (d, V, E), where d is the number of nodes, V = {v1, v2, ..., v d} represents the node set, E={(v i ,v j )|i,j=1,…,d} is from v i to v j The probability model associated with G is decomposed into Where P(X j |Pa(X j )) is a given X j The superset Pa(X j ):=X k |(v k ,v j )∈E}.

[0079] Causal models are generally built based on implicit assumptions about functions and / or noise. Classic causal models include Rubin's causal model and Pearl et al.'s causal graph model. Causal graph models are primarily based on Bayesian structural learning algorithms. They detect conditional independence between variables or assume causal Markov assumptions and causal fidelity assumptions based on the joint distribution of variables. These models are suitable for discovering global causal structures in high-dimensional data. However, these methods suffer from the Markov equivalence class problem and are unable to identify conditional independence relationships within the same equivalence class.

[0080] 2. Causal Markov Assumption

[0081] For a set of variables with causal sufficiency, if all variables are conditionally independent from their non-descendant nodes under the condition that the parent node of the variable is known, then this situation is said to satisfy the causal Markov assumption [Causation, prediction and Search]

[0082] 3. Loyalty Hypothesis

[0083] If, given a set of variables V, variable v i and v j Mutually independent or conditionally independent, then in the causal network graph G consisting of variables and their causal dependencies, vi and v j All paths between them are separated by appropriate variables d-separated in the variable set V (schematic diagram as shown in Figure 1 As shown), the joint distribution P of all random variables is said to be causally faithful to the graph G.

[0084] 4. Causal Sufficiency Assumption

[0085] A set of variables V is considered causally sufficient when the direct cause of any two variables in V also exists in V. If the set of observations of the variables is not causally sufficient, then the causal model may contain unobserved common dependent variables.

[0086] 5. Data generation method SEM

[0087] Assume that the data generation process model is as shown in [Hoyer et al. (2009); Peters et al. (2014).], each variable x j is associated with node i in the DAG, and x j The observations of are obtained as a function of their parents in the graph plus independent additive noise ∈, i.e.:

[0088] X j :=f j (Pa(X j ))+∈ j ,j=1,…,d

[0089] Among them, f j Represents X j The functional relationship between Pa(X j ) represents X j The parent node of j Assume joint independence. Assume causality is minimal, which is equivalent to each f j For any X in this structural equation SEM k ∈Pa(X j ) are not constants.

[0090] Causally, without further assumptions on the form of the function and / or noise, the above model can only be identified to a Markov equivalence class under the usual Markov and faithfulness assumptions; in this example, a synthetic dataset generated from a fully identifiable model is considered, so the estimated graph is evaluated as a true DAG.

[0091] 6. Scoring function - BIC

[0092] Given a sample X=[x1,x2,…,x d ]∈R m*d , where xj is a random variable X j The goal is to find the directed acyclic graph G that optimizes the Bayesian Information Criterion (BIC) (or equivalently, the minimum description length) score, defined as follows:

[0093]

[0094] It's X j The kth observation value of θ j The parameter associated with each possibility, |θ j | is the dimension of the parameter.

[0095] 7. Topological sorting

[0096] A topological sort can be obtained for a directed graph only if it is a directed acyclic graph (DAG). A topological sort is a linear sort of all nodes such that for any two nodes in the graph, if there is a directed edge from u to v, then u comes before v in the topological sort. For example, if O = {1, 2, 3, 4, 5}, Figure 2 shown.

[0097] That is, finding a graph (DAG) that satisfies the acyclic constraint is equivalent to finding a topological sort O of the graph G. The same directed graph may have multiple such node sorts, and the set of such sorts can be expressed as:

[0098] Φ(0)={0: fully connected DAG G 0 is a DAG superset of G}

[0099] The superset DAG of G means that all edges in G are contained in the superset. The optimal causal structure G * ∈Φ(O), so the task of this embodiment transforms the problem of finding a graph scoring function that satisfies the acyclicity constraint into the problem of finding the optimal graph topology sorting, which reduces a lot of calculations.

[0100] Reinforcement learning is an intelligent agent learning through trial and error, interacting with the environment to obtain rewards to guide the behavior of the intelligent agent so that the intelligent agent can obtain the maximum reward.

[0101] 8. Reinforcement Learning

[0102] like Figure 3 As shown in Figure 1, standard reinforcement learning is modeled as a Markov process (MDP) on a solvable basis. The state transition model P of the environment is represented by the transition probability T(s'|s,a), which is the probability of taking action a in state s to go to the next state s'. The goal of reinforcement learning is to find an optimal policy π φ(a|s) (where φ is a parameter) maximizes the cumulative reward R obtained by the agent. Reinforcement learning rewards are divided into two types according to different environments: dense reward and episode reward. The former is an immediate reward feedback for each action or state, but in general, the reward is provided after the task is completed, which summarizes the performance of the entire sequence or episode. For episodic tasks, trajectories can be collected by repeatedly executing the strategy. Where T is a finite time horizon.

[0103] Causal structure learning is aimed at the overall structure score, so this example sets the reward to episode deward and calculates the expected cumulative return of the trajectory: where γ∈(0,1) represents the discount rate.

[0104] Based on the above technology, the embodiments of the present invention will be further described below.

[0105] See also Figure 4 The first embodiment of the present invention provides a knowledge-guided causal discovery method, which is used to discover the causal relationship between the macro-monthly indicator data of an enterprise, thereby finding the key factors affecting the enterprise indicator data to help the enterprise perform precise regulation. However, it should be understood that this method can also be used to discover causal relationships in other scenarios. The method can be executed by a knowledge-guided causal discovery device (hereinafter referred to as the discovery device), in particular, by one or more processors within the discovery device to implement the following steps:

[0106] S101, obtain an enterprise macroeconomic monthly indicator dataset X, and randomly select n samples from the m samples of the enterprise macroeconomic monthly indicator dataset X to construct a training sample set.

[0107] In this embodiment, the enterprise macroeconomic monthly indicator dataset X mainly includes financial indicators such as enterprise scale, asset load ratio, net profit margin of total assets, proportion of intangible assets, cash flow ratio, comprehensive tax rate, financial leverage, enterprise investment level, total asset growth rate, as well as corporate governance indicators such as board size, management gender ratio, management compensation, and independent director ratio.

[0108] In this embodiment, specifically, small batch training is adopted to randomly select n samples from the m samples of the enterprise macro monthly indicator dataset X to construct a training sample set. Where d is the dimension of the sample.

[0109] S102, using an encoder to embed each training sample in the training sample set into a state, and construct a complete state space based on all embedded states; wherein the t-th element of the state space represents the actual state encountered in the t-th decision step when generating the variable ranking.

[0110] In this embodiment, if Figure 5 As shown, considering that the transformer encoder can better capture the causal relationship between variables, the encoder is used For each training sample x j Embedded into state s j , all embedded states constitute the state space S: = {s1, ..., s d Then, we set an initial state, represented by s0, to select the first action, and the complete state space will be

[0111] Among them, each state s j Set in the same dimension, i.e. s j ∈R n , here we will use to represent the actual state encountered at the tth decision step when generating the variable ranking. i For the encoder variable x i The state value obtained after encoding. If the t-th decision step selects variable x i , then the state is transferred to s i , then the actual state after the transfer is Right now

[0112] In this embodiment, the initial state s0 is set to

[0113] S103: Use the decoder to decode the state, map the state space to the action space, and output the corresponding variables. In each step, the selected variables are masked out and the remaining unselected variables are output to ensure that the variables are selected once. After all variables are selected, the initial ranking π is formed.

[0114] In this embodiment, there is a potential causal relationship between each state. In the sequence problem, traditional neural network models such as RNN are difficult to effectively model long-term dependencies. Therefore, this embodiment adopts the LSTM model, using the decoder By decoding the state, the state space Mapped to the action space A, the decoder structure is as follows Figure 6 shown.

[0115] The state s is converted by the decoder jMapped to action a (action is variable X j ), according to the LSTM input state s j , output the corresponding variables. At each step, the selected variables are masked and the remaining unselected variables are output to ensure that the variables are selected once. After all variables are selected, the variable ranking is formed. Among them, the order in which each variable is selected is different.

[0116] Among them, an action (i.e., variable) is selected from the action space A composed of all variables in each decision step. The size of the action space is |A|=d. Compared with the DAG space size O{2 d×d}, the action space is smaller.

[0117] S104, using the idea of ​​direct insertion sort to adjust the initial sort to obtain the adjusted sort, and obtaining the final optimal sort by comparing the knowledge reward function.

[0118] In this embodiment, if Figure 7 As shown, since the initial sorting Π is the initial optimal sorting explored by reinforcement learning, the idea of ​​direct insertion sorting is used to adjust the initial sorting Π to obtain the adjusted sorting Π'. By comparing the knowledge reward function R k To obtain the final optimal sort π * :

[0119] R k =S BIC (G Π )-S BIC (G Π′ )

[0120] Among them, S BIC (G Π ) represents the dense reward of ranking π; assuming that the variance of Gaussian noise is equal in each variable, we get:

[0121]

[0122] in represents the residual sum of squares of the i-th variable, is the least squares loss, m is the sample size, The number of edges in the graph generated for sorting π.

[0123] S105, based on the comparison between the optimal order and the initial order found, the changed node relationship is used as prior knowledge, and the prior knowledge is continuously updated through the iterative process of the search. The prior knowledge is used to obtain effective knowledge that the reinforcement learning did not obtain during the exploration process to constrain the direction of subsequent reinforcement learning exploration and guide the reinforcement learning to find the global optimal solution.

[0124] In this embodiment, according to the optimal sorting π* Compare it with the initial sort π found, and change the relationship between the nodes (X i ,X j ) as prior knowledge P k :

[0125] P k =P k ∪(X i ,X j ); if I Π *(X i ,X j )≠I Π (X i ,X j )

[0126] in, Represents node X i and X j In the optimal sorting π * The relative index position in the search. The prior knowledge is continuously updated through the iterative process of search. Prior knowledge is used to obtain effective knowledge that reinforcement learning has not obtained during the exploration process to constrain the direction of subsequent reinforcement learning exploration and guide reinforcement learning to find the global optimal solution faster. The specific details of prior knowledge acquisition are as follows Figure 8 shown.

[0127] Traditional reinforcement learning often faces the dilemma of insufficient exploration or over-exploitation, requiring a balance between exploration and exploitation. This embodiment constrains the exploration and optimization direction of reinforcement learning based on acquired prior knowledge, allowing reinforcement learning to rationally utilize the explored effective ranking knowledge without falling into local optimal solutions.

[0128] The goal of the optimization process is to maximize the policy J(φ), where φ = {φ e ,φ d}, where φ e and φ d is with the encoder f enc and decoder f dec Based on the above definition, the associated parameters are optimized using policy gradients to generate the sorting model parameters. The policy gradient is:

[0129]

[0130] T is the total number of decision steps, E φ represents the expected value of the policy φ, which means the total return R(τ) of the trajectory τ multiplied by the logarithmic policy gradient under the current policy; τ represents a trajectory, usually refers to the policy π φ A complete path of executing a sequence of actions from the initial state to the termination; Indicates that in strategy π φ In state Next select action a t The probability value of .

[0131] The model parameters are trained using the Actor-Critic algorithm. The Critic is a judgment network. When the input is the environment state, it can evaluate the value of the current state. When the input is the environment state and the action taken, it can evaluate the value of taking the action under the current state. The Actor is a policy network. It takes the current state as input and outputs the probability distribution of the action or the continuous action value. The Critic network then evaluates the quality of the action and adjusts the policy. The learning rate of the Actor and Critic is set to lr = 1e -4 and lr = 1e -4 The Adam optimizer is used. The discount factor γ = 0.98. The actor consists of an encoder and a decoder, which are selected as described above. The critic uses a three-layer feedforward neural network with 512, 256, and 1 units, which is based on the state is input, output given state The predicted value of the current policy.

[0132] In order to better constrain the exploration of reinforcement learning, this embodiment introduces the penalty weight W o To constrain the optimization direction of the next round of sorting:

[0133]

[0134] where w i Corresponding to the node X in the sort π i The penalty weight of the sorted Π is calculated based on the prior knowledge obtained from the previous round of exploration. i ,X j ) index position conflicts with the parent node X in prior knowledge i Assign penalty weight w i , the remaining nodes have penalty weights w i Set to 0. The penalty weight value of the conflicting node is obtained through experimental comparison, namely:

[0135]

[0136] Represented as node X i and X j In prior knowledge p k The relative index position in ;

[0137] According to the penalty weight W of the ranking Π o , set the loss function of Actor and Critic to constrain the direction of algorithm optimization:

[0138]

[0139] Among them, π * represents the global optimal solution, s t and a t Represents state and action, G t is the experience return, defined as the cumulative reward from time step t to a certain period in the future: G t =R t+1 +γR t+2 +…+γ n-1 R t+n +γ n Q(s t+n ,a t+n ), γ represents the discount coefficient, N represents the number of training samples, R t+1 represents the reward obtained at time step t+1, Q(s t ,a t ) means in state s t Next select action a t The action value function is the current estimated action value function, V(s t ) means in state s t The state value function under .

[0140] S106, transforming the global optimal solution to obtain an initial causal structure, pruning the initial causal structure to obtain a final causal structure, and predicting the data to be predicted based on the final causal structure to obtain causal relationships, so as to obtain key factors affecting enterprise indicator data.

[0141] In this embodiment, a causal graph can be obtained from the global optimal solution by performing variable selection methods, such as sparse candidate and covariate significance tests. In this work, for linear data models, linear regression is applied to the obtained fully connected initial causal structure, and then a threshold is used to prune edges with small weights to obtain the final causal structure. For nonlinear models, CAM is used to prune the initial causal structure to obtain the final causal structure.

[0142] Among them, the final causal structure can be used to predict the data to obtain causal relationships, and then obtain key factors affecting corporate indicator data, such as the proportion of independent directors, debt-to-asset ratio, corporate scale, total asset net profit margin and other key factors.

[0143] In addition, for each variable X j , can be targeted at X j A generalized additive model was fitted to the current parents, and then the covariates were tested for significance, which was considered significant if the threshold p value was less than or equal to 0.001.

[0144] In summary, compared with the prior art, this embodiment has at least the following improvements:

[0145] (1) To avoid dealing with acyclic constraints and reduce the size of the search space, this embodiment formulates the combinatorial optimization problem of directed graphs as a variable ranking search problem of a multi-step Markov decision process, and implements the variable ranking generation process in the decoder-encoder architecture.

[0146] (2) In order to reduce the complexity of the interactive environment and the existence of uncertain information, a knowledge acquisition method is designed. The prior knowledge acquired iteratively is used to adaptively assign weights to samples in the training batch, guide the search direction of reinforcement learning in complex environments, and use reinforcement learning to optimize based on the reward function and the weights of the training samples;

[0147] (3) In order to improve the training speed of the algorithm, an asynchronous parallel framework is designed, and a dynamic threshold is set according to the reward scores ranked according to different course stages, so that the method can dynamically adjust the number of iterations and reduce unnecessary iterations.

[0148] This example conducted comparative experiments on synthetic and real datasets for problems of different scales, using metrics including true positive rate, structural Hamming distance, F1 score, and runtime. The experiments demonstrated that this example outperformed other algorithms across all metrics for problems of varying scales.

[0149] See also Figure 9 The second embodiment of the present invention further provides a knowledge-guided causal discovery device, comprising:

[0150] The training sample construction unit 210 is used to obtain the enterprise macro monthly indicator dataset X, and randomly select n samples from the m samples of the enterprise macro monthly indicator dataset X to construct a training sample set.

[0151] The encoding unit 220 is used to use an encoder to encode the training sample set Each training sample x in j Embedded into state s j , forming a complete state space based on all embedded states Among them, the state space The tth element of represents the actual state encountered at the tth decision step when generating the variable ranking;

[0152] The decoding unit 230 is used to decode the state space by using a decoder. Map to the action space A and output the corresponding variables. In each step, the selected variables are masked out and the remaining unselected variables are output to ensure that the variables are selected once. After all variables are selected, the initial ranking π is formed.

[0153] The sorting adjustment unit 240 is used to adjust the initial sort Π to obtain the adjusted sort Π' by using the idea of ​​direct insertion sorting, and obtain the final optimal sort Π by comparing the knowledge reward function. * ;

[0154] Reinforcement learning unit 250, for * Compare it with the initial ranking π found by the search, take the changed node relationship as prior knowledge, continuously update the prior knowledge through the iterative process of search, and use the prior knowledge to obtain effective knowledge that reinforcement learning has not obtained in the exploration process to constrain the direction of subsequent reinforcement learning exploration and guide reinforcement learning to find the global optimal solution;

[0155] The pruning unit 260 is used to transform the global optimal solution to obtain an initial causal structure, prune the initial causal structure to obtain a final causal structure, and predict the data to be predicted based on the final causal structure to obtain a causal relationship, so as to obtain key factors affecting enterprise indicator data.

[0156] A third embodiment of the present invention further provides a knowledge-guided causal discovery device, comprising:

[0157] at least one processor; and,

[0158] a memory communicatively connected to the at least one processor; wherein,

[0159] The memory stores instructions that can be executed by the at least one processor. The instructions are executed by the at least one processor to enable the at least one processor to perform the above-mentioned knowledge-guided causal discovery method.

[0160] The third embodiment of the present invention further provides a non-transitory computer-readable storage medium, which stores computer instructions for causing the computer to execute the knowledge-guided causal discovery method as described above.

[0161] The fifth embodiment of the present invention also provides a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium, and the computer program includes program instructions. When the program instructions are executed by a computer, the computer executes the knowledge-guided causal discovery method of any of the aforementioned embodiments.

[0162] Figure 10A schematic diagram of a method or device 1000 that can implement an embodiment of the present invention is shown. In some embodiments, the method or device 1000 may include more or fewer devices than shown. In some embodiments, the method or device 1000 may be implemented using a single device or multiple devices. In some embodiments, the method or device 1000 may be implemented using cloud-based or distributed devices.

[0163] like Figure 10 As shown, device 1000 includes a processor 1001, which can perform various appropriate operations and processes according to the programs and / or data stored in a read-only memory (ROM) 1002 or the programs and / or data loaded from a storage portion 1008 into a random access memory (RAM) 1003. Processor 1001 can be a multi-core processor or can include multiple processors. In some embodiments, processor 1001 can include a general-purpose main processor and one or more special coprocessors, such as a central processing unit (CPU), a graphics processing unit (GPU), a neural network processor (NPU), a digital signal processor (DSP), etc. Various programs and data required for the operation of device 1000 are also stored in RAM 1003. Processor 1001, ROM 1002, and RAM 1003 are connected to each other via a bus 1004. An input / output (I / O) interface 1005 is also connected to bus 1004.

[0164] The processor and memory are used together to execute the program stored in the memory. When the program is executed by the computer, the methods, steps or functions described in the above embodiments can be implemented.

[0165] The following components are connected to the I / O interface 1005: an input section 1006 including a keyboard, a mouse, a touch screen, and the like; an output section 1007 including devices such as a cathode ray tube (CRT), a liquid crystal display (LCD), and a speaker; a storage section 1008 including a hard disk and the like; and a communication section 1009 including a network interface card such as a LAN card or a modem. The communication section 1009 performs communication processing via a network such as the Internet. A drive 1010 is also connected to the I / O interface 1005 as needed. A removable medium 1011, such as a magnetic disk, an optical disk, a magneto-optical disk, or a semiconductor memory, is installed in the drive 1010 as needed, so that computer programs read therefrom can be installed into the storage section 1008 as needed. Figure 10 Only some components are shown schematically, which does not mean that the device 1000 only includes Figure 10 Components shown.

[0166] The systems, devices, modules, or units described in the above embodiments may be implemented by a computer or its associated components. The computer may be, for example, a mobile terminal, a smartphone, a personal computer, a laptop computer, an in-vehicle human-computer interaction device, a personal digital assistant, a media player, a navigation device, a game console, a tablet computer, a wearable device, a smart TV, an Internet of Things system, a smart home, an industrial computer, a server, or a combination thereof.

[0167] Although not shown, in an embodiment of the present invention, a computer-readable storage medium is provided, on which a computer program / instruction is stored. When the computer program / instruction is executed by a processor, the method described in Embodiment 1 is implemented.

[0168] Storage media in embodiments of the present invention include permanent and non-permanent, removable and non-removable items that can be used to store information using any method or technology. Examples of storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transmission medium that can be used to store information that can be accessed by a computing device.

[0169] Although not shown, an embodiment of the present invention further provides a computer program product, including: a computer program / instruction, which implements the method described in Example 1 when executed by a processor.

[0170] The methods, programs, systems, and apparatuses of the embodiments of the present invention may be executed or implemented in a single or multiple networked computers, or may be practiced in a distributed computing environment. In the embodiments of this specification, in these distributed computing environments, tasks may be performed by remote processing devices connected via a communication network.

[0171] Those skilled in the art will appreciate that the embodiments of this specification may be provided as methods, systems, or computer program products. Therefore, those skilled in the art will appreciate that the functional modules / units or controllers and related method steps described in the above embodiments may be implemented using software, hardware, or a combination of software / hardware.

[0172] Unless explicitly stated, the actions or steps of the methods, procedures, and methods described in accordance with the embodiments of the present invention do not have to be performed in a specific order and can still achieve the desired results. In some implementations, multitasking and parallel processing are also possible or may be advantageous.

[0173] In this document, multiple embodiments of the present invention are described, but for the sake of brevity, the description of each embodiment is not exhaustive, and the same or similar features or parts between the embodiments may be omitted. In this document, "one embodiment", "some embodiments", "example", "specific example", or "some examples" are intended to apply to at least one embodiment or example according to the present invention, but not all embodiments. The above terms do not necessarily mean to refer to the same embodiment or example. Those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, unless they are mutually contradictory.

[0174] While the exemplary systems and methods of the present invention have been specifically shown and described with reference to the foregoing embodiments, these are merely examples of the best modes for implementing the present systems and methods. Those skilled in the art will appreciate that various changes may be made to the embodiments of the systems and methods described herein when implementing the present systems and / or methods without departing from the spirit and scope of the present invention as defined in the appended claims.

Claims

1. A knowledge-guided causal discovery method for discovering the causal relationship between enterprise macroeconomic monthly indicator data to find the key factors affecting enterprise indicator data, characterized by: include: S101: Obtain a monthly enterprise macroeconomic indicator dataset X, which includes financial indicators and corporate governance indicators. Randomly select n samples from the m samples in the monthly enterprise macroeconomic indicator dataset X to construct a training sample set. S102, using the encoder to train the sample set Each training sample x in j Embedded into state s j , forming a complete state space based on all embedded states Among them, the state space The tth element of represents the actual state encountered at the tth decision step when generating the sort; S103, using the decoder, by decoding the state, the state space Map to the action space A and output the corresponding variables. In each step, the selected variables are masked out and the remaining unselected variables are output to ensure that the variables are selected once. After all variables are selected, the initial ranking π is formed. S104, use the idea of ​​direct insertion sort to adjust the initial sort Π to obtain the adjusted sort Π', and obtain the final optimal sort Π by comparing the knowledge reward function * ; S105, according to the optimal sorting π * Compare it with the initial ranking π found by the search, take the changed node relationship as prior knowledge, continuously update the prior knowledge through the iterative process of search, and use the prior knowledge to obtain effective knowledge that reinforcement learning has not obtained in the exploration process to constrain the direction of subsequent reinforcement learning exploration and guide reinforcement learning to find the global optimal solution; S106, transforming the global optimal solution to obtain an initial causal structure, pruning the initial causal structure to obtain a final causal structure, and predicting the data to be predicted based on the final causal structure to obtain causal relationships, so as to obtain key factors affecting enterprise indicator data.

2. The knowledge-guided causal discovery method according to claim 1, characterized in that: In step S102, the training sample set d is the total number of variables, n is the size of the training sample; the encoder is represented as is the parameter of the encoder; each state s j Set in the same dimension, i.e. s j ∈R n ; All states embedded by training samples constitute the state space S: = {s1, ..., s d }; Set an initial state s0, represented by s0 to select the first action, then the complete state space 3. The knowledge-guided causal discovery method according to claim 2, characterized in that: In step S103, the decoder is represented as is the parameter of the encoder, and the action output by the decoder is the variable.

4. The knowledge-guided causal discovery method according to claim 3, characterized in that: In step S104, the knowledge reward function R k Expressed as: R k =S BIC (G ∏ )-S BIC (G ∏′ ) Among them, S BIC (G ∏ ) represents the dense reward of sorting Ώ; assuming that the variance of Gaussian noise is equal in each variable, we get: in represents the residual sum of squares of the i-th variable, is the least squares loss, m is the sample size, The number of edges in the graph generated for sorting Π.

5. The knowledge-guided causal discovery method according to claim 4, characterized in that: In S105, the prior knowledge P k The expression is: in, Represents node X i and X j In the optimal sorting * The relative index position in ; Then S105 specifically includes: Use policy gradient to optimize the parameters of the ranking generation model; where the policy gradient is: Among them, J(φ) represents the maximization strategy, φ={φ e ,φ d }, T is the total number of decision steps, E φ represents the expected value of the policy φ, which means the total return R(τ) of the trajectory τ multiplied by the logarithmic policy gradient under the current policy; τ represents a trajectory, usually refers to the policy π φ A complete path of executing a sequence of actions from the initial state to the termination; Indicates that in strategy π φ In state Next select action a t The probability value of The model parameters are trained by the Actor-Critic algorithm, with the current state as input and the output as the probability distribution of the action or the continuous action value. The Critic network then evaluates the quality of the action and adjusts the strategy. Among them, the Actor is the policy network, which consists of an encoder and a decoder; the Critic is the evaluation network, which uses a three-layer feedforward neural network with 512, 256 and 1 units, with the state is input, output given state The predicted value of the current strategy; wherein, during the training process, the penalty weight W o To constrain the optimization direction of the next round of sorting: where w i Corresponding to the node X in the sort ∏ i The penalty weight of the sorting ∏ (X i , X j ) index position conflicts with the parent node X in prior knowledge i Assign penalty weight w i , the remaining nodes have penalty weights w i Set to 0; the penalty weight value of the conflict node is obtained through experimental comparison, that is: Represented as node X i and X j In prior knowledge p k The relative index position in ; According to the penalty weight W of the sort Ώ o , set the Actor's loss function Loss Actor And Critic's loss function Loss Critic To constrain the direction of algorithm optimization: Among them, π * represents the global optimal solution, s t and a t Represents state and action, G t is the experience return, defined as the cumulative reward from time step t to a certain period in the future: G t =R t+1 +γR t+2 +…+γ n-1 R t+n +γ n Q(s t+n , a t+n ), γ represents the discount coefficient, N represents the number of training samples, R t+1 represents the reward obtained at time step t+1, Q(s t , a t ) means in state s t Next select action a t The action value function is the current estimated action value function, V(s t ) means in state s t The state value function under .

6. The knowledge-guided causal discovery method according to claim 1, characterized in that: In step S106, linear regression is applied to the obtained fully connected initial causal structure, and then the edges with smaller weights are pruned using a threshold to obtain the final causal structure; for nonlinear models, CAM pruning is used to prune the initial causal structure to obtain the final causal structure.

7. The knowledge-guided causal discovery method according to claim 1, characterized in that: Also includes: For each variable X j , for X j A generalized additive model is fitted to the current parents, and then the covariates are tested for significance. If the threshold p-value of the significance test is less than or equal to 0.001, the covariates are significant.

8. A knowledge-guided causal discovery device for discovering the causal relationship between enterprise macroeconomic monthly indicator data to find the key factors affecting enterprise indicator data, characterized in that: include: The training sample construction unit is used to obtain a monthly macroeconomic indicator dataset X of an enterprise, wherein the monthly macroeconomic indicator dataset X includes financial indicators and corporate governance indicators, and randomly select n samples from the m samples of the monthly macroeconomic indicator dataset X to construct a training sample set. The encoding unit is used to use the encoder to convert the training sample set Each training sample x in j Embedded into state s j , forming a complete state space based on all embedded states Among them, the state space The tth element of represents the actual state encountered at the tth decision step when generating the variable ranking; The decoding unit is used to decode the state space by using a decoder. Map to the action space A and output the corresponding variables. In each step, the selected variables are masked out and the remaining unselected variables are output to ensure that the variables are selected once. After all variables are selected, the initial ranking π is formed. The sorting adjustment unit is used to adjust the initial sort Π to obtain the adjusted sort Π' by using the idea of ​​direct insertion sorting, and obtain the final optimal sort Π by comparing the knowledge reward function * ; Reinforcement learning unit for optimal sorting π * Compare it with the initial ranking π found by the search, take the changed node relationship as prior knowledge, continuously update the prior knowledge through the iterative process of search, and use the prior knowledge to obtain effective knowledge that reinforcement learning has not obtained in the exploration process to constrain the direction of subsequent reinforcement learning exploration and guide reinforcement learning to find the global optimal solution; A pruning unit is used to transform the global optimal solution to obtain an initial causal structure, prune the initial causal structure to obtain a final causal structure, and predict the data to be predicted based on the final causal structure to obtain a causal relationship, so as to obtain key factors affecting enterprise indicator data.

9. A knowledge-guided causal discovery device, characterized in that: The device includes: at least one processor; and, a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the knowledge-guided causal discovery method described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium, characterized in that The non-transitory computer-readable storage medium stores computer instructions, which are used to enable the computer to execute the knowledge-guided causal discovery method according to any one of claims 1 to 7.

Images