Causal model off-line evaluation and learning method for personalized decision-making system and instance of causal model off-line evaluation and learning method

Abstract

The invention provides a causal model evaluation and learning method for personalized decision and an example thereof. The method comprises a special causal model evaluation method and a causal model learning method. According to the evaluation method disclosed by the invention, random test data is used, the quasi-decision income of the causal model is obtained through the quasi-random test to evaluate the decision income amount of the causal model, and the learning method disclosed by the invention optimizes the decision output of the causal model by learning the optimal decision grouping label. According to the method, the causal model for personalized decision can be evaluated and learned offline, the evaluation method is weak in hypothesis dependence on data and the to-be-evaluated causal model, does not depend on individual characteristics, is wide in application range, does not have an unknown confusion variable problem, and the evaluation result and the decision income amount are directly linked; the model learned by the learning method is high in individual decision-making benefit, convenient to use and suitable for a data-driven individual decision-making system. The technical problem solved by the disclosure is that a solution for evaluating and learning the causal model in the prior art is too single, and the disclosure provides more choices for evaluating and learning the causal model of a personalized decision system.

Description

Technical Field

[0001] The present invention relates to a causal model for a personalized decision-making system, and in particular to an offline evaluation and learning method and system for a causal model applied to personalized decision-making. Background Art

[0002] With the rapid development of big data and machine learning technologies, personalized decision-making systems have been widely used in many fields, such as personalized credit, personalized recommendation systems, personalized medical care and health care, personalized marketing, etc. These personalized decision-making systems analyze the historical behavior and characteristics of a unit (user or patient) and make personalized decisions for the unit (such as recommending specific content, determining loan amounts, using male or female large models to interact with social software users, etc.) in order to achieve specific goals (such as high interest income, high user click-through rate, no disease progression, stock profits, media news likes, etc.).

[0003] However, traditional personalized decision-making systems mainly rely on correlation prediction models, that is, to explore the linear or nonlinear correlation between the unit's characteristics and the historical decisions of the unit and specific goals based on historical data to predict the unit's target variable outcome. This method has some problems, such as we do not know whether there will be the same or better results if we do not make this decision, that is, we cannot rule out the evaluation of decision benefits due to confounding factors that are not collected or cannot be collected or not thought of.

[0004] For example, when evaluating the effect of a new drug on the recovery of heart patients, it may be observed in the study that patients who receive the new drug in the real world seem to recover more slowly than those who do not receive the treatment. However, if confounding factors such as the patient's age, gender, lifestyle, and dietary habits are not taken into account, it is unknown whether not receiving the new drug will lead to a slower recovery, and the causal effect of the new drug on the speed of recovery may be incorrectly assessed.

[0005] To solve these problems, researchers began to explore the use of causal models to improve personalized decision-making systems. Since the introduction of causal models, people have been paying attention to the improvement of offline evaluation and learning methods of causal models.

[0006] Existing methods mainly include evaluation and learning based on hypothesis-generated synthetic or semi-synthetic data, and evaluation and learning based on predictive performance. However, personalized decision-making systems must rely on the data characteristics of real individuals. This is because, on the one hand, evaluation and learning based on hypothesis-generated data cannot reflect the causal relationship in the real world and it is difficult to create value in real-world applications. On the other hand, even if the generated data can reflect the causal relationship in the real world and represent the characteristics of the real world, it means that a generative model of the real world, that is, a model used to generate data, has been obtained. Since the causal model is to learn some causal knowledge from real-world data, there is no need to use the generated data to learn the causal model and evaluation. Evaluation and learning based on the predictive performance of the target variable cannot exclude the influence of unknown confounding factors and does not consider counterfactual outcomes, thereby misjudging the correlation as a causal relationship and bringing negative impacts to personalized decision-making. Therefore, both are inappropriate.

[0007] In summary, a new causal model evaluation and learning method is urgently needed to replace or supplement the existing technology, especially in today's increasingly popular personalized decision-making services, such as personalized recommendations, personalized medical and health care, personalized credit, personalized marketing and other applications. In fact, in order to more accurately and quantitatively evaluate the quality of causal models in the real world, more and more causal model evaluation and learning works in the world no longer use synthetic data and direct or indirect predictive indicators directly targeting the target variable.

[0008] Randomized trial data analysis that has emerged in recent years is a new method for evaluating and learning causal models. With the availability of randomized trial data, people can no longer tolerate the risks of unknown confounding factors brought about by using observational data for offline evaluation or learning causal models, the application limitations of synthetic data evaluation, and even more so the inability to correctly evaluate the size of the decision benefits brought about by the causal model used. Summary of the invention

[0009] The purpose of the present invention is to avoid the shortcomings of the above-mentioned prior art and to propose an evaluation and learning method and example of a causal model of a personalized decision-making system. Offline random test data is used to provide a learning and evaluation method for the personalized decision-making system with the help of optimal grouping labels, and the method is implemented on different personalized decision-making systems. The method is applicable to various systems and provides a new method for offline evaluation and learning of causal models.

[0010] Specifically, the present disclosure describes systems and methods for evaluating and learning causal models, which are implemented as computer programs for causal model learning and evaluation applied to personalized decision systems on one or more computers in one or more locations.

[0011] The purpose of evaluating the causal model disclosed in the present invention can be achieved by the following technical solution: Step 101, obtaining the causal model to be evaluated f and the test data set S containing multiple units te , and complete the configuration of the evaluation method; Step 102, for multiple units in the test set, divide the test units with the same proposed decision value G of the causal model f into the same group; Step 103, respectively calculate the difference matrix of potential results when the decision variable values ​​of the units in each group are different; Step 104, calculate the causal benefit of the causal model f for each unit decision according to the decision causal benefit function B(z).

[0012] The purpose of learning causal models in the present disclosure can be achieved by the following technical solutions: Step 201, obtaining a causal model to be learned f and a training data set S containing multiple units te , and complete the configuration of the learning method; step 202, determine the target variable cutoff point Y0 according to the target variable Y of the data unit of multiple units; step 203, given the units in the training data set, group the units according to the target variable cutoff point Y0 and the decision variables A and target variable Y of the data unit of the unit, so as to generate the optimal grouping label of the unit as the optimal decision label G; step 204, learn the causal model according to the feature variables X and the optimal decision label G of the data unit of multiple units.

[0013] In some specific implementations, the test data set S in step 101 and step 201 is te And the training data set S te The decision variable A of unit z in is randomly generated, where a measurement result of unit z is a data unit (X, A, Y), X is the feature variable before the decision variable A is generated, A is the decision variable, and Y is the target variable of unit z after the decision variable A is generated.

[0014] Among them, the characteristic variable types of data units can be table records, natural languages, images, videos, sensors, etc., which can be static data or time series, such as patients' electronic health records and key vital signs monitoring information flows, users' personal information and operation flows.

[0015] Among them, the decision variable of the data unit is the specific decision taken by the personalized decision system for the unit. The decision variable type can be a binary Boolean variable or a multi-valued discrete variable, such as whether to pop up a reminder for the patient, what kind of product to push to the user, the discount amount given to the user, and the price of the product for the user. It can be a one-dimensional variable or a multi-dimensional variable, such as the pushed image, sound, and video.

[0016] Among them, the target variable of the data unit is the specific attribute of the unit that the personalized decision-making system wants to measure in response to the unit's specific decision or behavior. The target variable type can be binary or multi-valued, such as whether the patient dies or whether the user purchases, the user's rating level, the user's total repayment amount, etc. It can be a one-dimensional variable or a multi-dimensional variable, such as whether there is disease progression, multi-dimensional physical indicators, etc.

[0017] In some specific implementations, step 202 includes obtaining a training data set S tr ={z i} i=1 ntr , using an objective boundary function Y0=F({z i} i=1 ntr ), determine the target variable cutoff point Y0=[Y0 1 ,Y0 2 ,…,Y0 DY ], where D Y is the dimension of Y, where the target cutoff point Y0 is determined by the target cutoff function F, the input of which is the target variable Y of all data units used for training, and the output is the target variable cutoff point Y0.

[0018] Function F can be the maximum M equal-division points or hyperplanes of the target variable Y of the sample or the maximum M equal-division points or hyperplanes of the target variable Y interval, where M is the number of possible values ​​of the decision variable A.

[0019] In some specific implementations, step 203 includes marking the units with A=i and Y>Y0 as the i-th group, allocating the units with A=i and Y≤Y0 to other groups that are not group i according to the selected remaining unit allocation method G=R(A,Y), where the total number of groups is M, and the group labels are 0, 1...M-1, respectively; finally, the group label of the unit is used as the predictor variable G, that is, the optimal decision label G for each unit.

[0020] The selected remaining unit allocation method G=R(A,Y) can be to randomly allocate the units with A=i and Y≤Y0 to other groups other than group i with equal probability, or to allocate them to other groups other than group i in turn after sorting them according to the Y value.

[0021] In some specific implementations, step 204 includes, based on the training data set S tr, the unit's characteristic variables and optimal decision labels learn a prediction model f:X→G, where the input X of the prediction model f is the unit's characteristic variables, and the output is the optimal decision label G. When learning the causal model f according to the prediction performance, select the model with the highest precision or the highest recall or the maximum F1-Score or the maximum PR-AUC; when using the causal model f, for the specific unit z, make a one-to-one corresponding decision A according to the output G of the causal model f.

[0022] The causal model f to be learned can be a linear classification model, a Bayesian classification model, a decision tree classification model and its improvements (XGBoost, LightgGBM, Catboost), a fully connected neural network classification model, or a pre-trained classification model based on the Transformer architecture (TabPFM), etc.

[0023] When using a causal model, for a specific unit z, a one-to-one decision A is made based on the output G of the causal model.

[0024] In some specific implementations, step 102 includes, given a causal model f and a test data set S te ={z i} i=1 nte , input the characteristic variable X of the unit in the test data set into the causal model, and obtain the output of the causal model. When the test units with the same value of the proposed decision G of the causal model f are divided into the same group, the group numbers are 0, 1...M-1, where M is the number of possible values ​​of the proposed decision G and the number of possible values ​​of the decision variable A. The input of the causal model f is the characteristic variable X of the unit z, and the output is the proposed decision G for the unit z.

[0025] In some specific implementations, step 103 includes calculating the difference matrix D of potential results when the decision variables of the units in the i-th group take different values i , where the matrix D i There are M-1 rows and M-1 columns, and the element in the jth row and kth column is Y i A=k -Y i A=j , where the potential outcome Y i A=j It means that if a specific unit z in group i executes decision j, then the expected value of the target variable Y measurement result of unit z is recorded as Y i A=j , infer the potential outcome Y of the units in group i i A=j The average value of the target variable Y of the samples whose unit decision variable of group i takes value j is used.

[0026] In some specific implementations, step 104 includes: the decision causal benefit function B=B(z) can be the difference matrix D of the potential results of the group where the unit z is located i The difference matrix D is the potential result of removing the minimum value of the i-th row or taking the group where the unit z is located. i The average of the ith column minus the ith row.

[0027] Furthermore, the decision causal benefit for each unit is calculated according to the decision causal benefit function B=B(z). The total decision causal benefit is the sum of the decision causal benefits of all units, and the average decision causal benefit is the total decision causal benefit of all units divided by the number of units.

[0028] Compared with the prior art, the learning and evaluation method and examples of the causal model for personalized decision-making systems disclosed in the present invention have the following advantages: the evaluation method disclosed in the present invention has weak data assumptions and a wide range of applications. It does not have any conditional independence assumptions, causal directions, or causal graph assumptions about the possible relationships between feature variables, decision variables, target variables, and latent variables. It does not have any assumptions about the form of latent variables or other variables. It does not have any specific strong functional assumptions about the relationship between variables, and it does not have any restrictions or dependencies on the model to be evaluated. The learning method disclosed in the present invention can significantly improve the performance of the causal model by learning the optimal grouping. BRIEF DESCRIPTION OF THE DRAWINGS Figure 1 It is a principle block diagram of the evaluation and learning method of the causal model for personalized decision-making system disclosed in the present invention. Figure 2 is the difference matrix D of the potential outcomes of the group where the decision variables are binary and the decision causal benefit function is the unit z. i Schematic diagram of the principle of the method for evaluating the average return of the causal model when the average value of the i-th column is removed. Figure 3 It is the optimal decision label generation process for all units when the decision variable is binary and the remaining units are assigned by random assignment with equal probability. DETAILED DESCRIPTION

[0029] The present disclosure is further described in detail below with reference to the best embodiments shown in the accompanying drawings: The principle block diagram of the evaluation and learning of the causal model for the personalized decision-making system disclosed in the present invention is as follows Figure 1 In the following specific embodiments, the present disclosure first uses the causal model learning method of step 201, step 202, step 203, and step 204 to learn the causal model, and then uses the causal model evaluation method of step 101, step 102, step 103, and step 104 to evaluate.

[0030] A specific embodiment 1 is to learn and evaluate a causal model for improving the 14-day mortality rate of patients with acute kidney injury using personalized electronic health record alerts for acute kidney injury.

[0031] In this specific embodiment, the data of step 201 and step 101 are from randomized trial data generated by a double-blind, multicenter, parallel, randomized, controlled trial of 6030 adult patients with acute kidney injury defined by the Kidney Disease: ImprovingGlobal Outcomes (KDIGO) creatinine criteria in four teaching hospitals and two non-teaching hospitals (including community hospitals and large nursing centers).

[0032] The target variable Y in this embodiment is the death of the patient within 14 days after the random intervention. The target variable takes the value of 0 in the case of death and takes the value of 1 in the case of non-death.

[0033] Among them, the randomized decision variable or intervention measure A in this embodiment is a pop-up alert window based on electronic health records, which pops up when the doctor opens the patient's medical record. When there is no pop-up alert, the decision variable takes a value of 0, and when there is a pop-up alert, the value takes a value of 1.

[0034] Among them, there are 47 characteristic variables X in this embodiment, including discrete variables and continuous variables: age, age greater than or equal to 90 years old, gender, whether Hispanic, whether African American, hospital (No. 1 city demonstration hospital, No. 2 city demonstration hospital, No. 3 city demonstration hospital, No. 4 suburban demonstration hospital, No. 5 good suburban non-demonstration hospital, No. 6 suburban non-demonstration hospital), whether it is a medical admission; whether there has been hypertension, whether there has been diabetes, whether there has been chronic obstructive pulmonary disease, whether there has been chronic kidney disease, whether there has been congestive heart failure, whether there has been acute kidney disease, whether there has been liver disease, whether there has been a malignant tumor; whether angiotensin converting enzyme inhibitors / angiotensin II receptor antagonists / renin inhibitors have been used for antihypertensive treatment within 24 hours, whether angiotensin converting enzyme inhibitors / angiotensin II receptor antagonists / renin inhibitors have been used for antihypertensive treatment within 72 hours, whether non-steroidal anti-inflammatory drugs have been used within 24 hours, whether non-steroidal anti-inflammatory drugs have been used within 72 hours, whether aminoglycosides have been used within 24 hours, and whether they have been used within 72 hours. whether the patient received proton pump inhibitors within 24 hours, whether the patient had undergone comparative study procedures within 72 hours, whether the patient had undergone CT surgery within 7 days; the number of hours from the onset of acute kidney injury to the randomized intervention, the number of days from the start date of the trial to randomization, and the number of hours from hospital admission to randomization; pulse at randomization, respiratory rate at randomization, systolic blood pressure measurement at randomization, diastolic blood pressure measurement at randomization; baseline creatinine measurement, minimum creatinine measurement within 48 hours, bicarbonate measurement at randomization, urea nitrogen measurement at randomization, chloride measurement at randomization, creatinine measurement at randomization, platelet count measurement at randomization, white blood cells measured at randomization, hemoglobin measurement at randomization, sodium measurement at randomization, potassium measurement at randomization, anion gap measurement at randomization; Elixhauser comorbidity index, simple intensive care index; whether the patient was in the emergency department at the time of randomization, estimated glomerular filtration rate at admission, and whether the patient was in the hospital ward for randomization.

[0035] Specifically, 80% of the data set is randomly used as the training data set S tr ={z i} i=1 4824 , and use the other 20% as the test data set S te ={z i} i=4825 6030 .

[0036] In some specific implementations, step 202 includes using a target boundary function Y0=0.5 to determine the target variable boundary point Y0=0.5.

[0037] In some specific implementations, step 203 includes: Figure 3 As shown, patients with acute kidney disease who had no pop-up alerts and had not died within 14 days were marked as Group 0, and patients who had pop-up alerts and had not died within 14 days were marked as Group 1. For the remaining units, patients with acute kidney disease who had pop-up alerts and died within 14 days were marked as Group 0, and patients who had no pop-up alerts and died within 14 days were marked as Group 1.

[0038] In some specific implementations, step 204 includes, based on the training data set S tr ={z i} i=1 4824 , the unit's characteristic variables and the optimal decision label learning prediction model f:X→G. Specifically, the input of the prediction model f is the unit's 47 characteristic variables X, and the output is the optimal decision label G. Among them, the CatBoost tree classification model is used for learning here, and the model is denoted as f1. Among them, the parameter learning of the linear classification model is the default setting of the scikit-learn package in Python, and the parameter learning of the CatBoost tree classification model is the default setting of the catboost package in Python. As a control group, the CatBoost tree classification learning prediction model f0:(X,A)→Y is also used, and T i =argmax A Y(X i ,A) Get the decision variable, that is, select the decision that the model believes will reduce the probability of the patient's death within 14 days.

[0039] In some specific implementations, step 102 includes, given the causal models to be evaluated f1, f0 and the test data set S te ={z i} i=4825 6030 , input the characteristic variable X of the unit in the test data set into the causal model, obtain the quasi-decision G of the causal model, and divide the test units with the same causal model output value G into the same group, and the group numbers are quasi-no-pop-up group 0 and quasi-pop-up group 1.

[0040] In some specific implementations, step 103 includes calculating the difference matrix D of the potential results of the units in group 0 0 =[0, Y 0 A=1 -Y 0 A=0 ; Y 0 A=0 -Y 0 A=1 , 0)] and the difference D between the potential results of the units in group 1 1 =[0, Y 1A=1 -Y 1 A=0 ;Y 1 A=0 -Y 1 A=1 , 0], where the potential outcome Y i A=j It means that if a specific unit z in group i executes decision j, then the expected value of the target variable Y measurement result of unit z is recorded as Y i A=j , infer the potential outcome Y of the units in group i i A=j The average value of the target variable Y of the samples whose unit decision variable of group i takes value j is used.

[0041] In some specific implementations, step 104 includes, for example, Figure 3 As shown, the decision causal benefit for each unit is calculated according to the decision causal benefit function B = B (z), where the decision causal benefit function is the difference matrix D of the potential results of the group where the unit z is located. i The average value of the ith column minus the ith row, that is, for the units in the group 0 that is not going to pop up, the decision causal benefit is Y 0 A=0 -Y 0 A=1 For the units in the pseudo-pop-up group 1, the decision causal benefit is Y 1 A=1 -Y 1 A=0 ; The total causal benefit of all units is the sum of the causal benefits of all units. sum =(Y 0 A=0 -Y 0 A=1 )*n0+(Y 1 A=1 -Y 1 A=0 )*n1, the average causal benefit of decision making is the total causal benefit of decision making divided by the number of units B ave =(Y 0 A=0 -Y 0 A=1 )*n0 / n+(Y 1 A=1 -Y 1 A=0 )*n1 / n, where n1 is the number of units in the test set in the group of proposed pop-ups, n0 is the number of units in the test set in the group of proposed non-pop-ups, and n is the total number of units in the test set.

[0042] The experiment was repeated 1000 times and the average was taken. A two-sample t-test was used for significance test. Finally, it was concluded that the mortality rate of the model f1 trained using the learning method of this embodiment was significantly reduced from 8.92% to 7.37%, and the average causal decision benefit was 0.0155. The mortality rate of the control group model f0 was not significantly reduced from 8.92% to 8.59%, and the average causal decision benefit was 0.0033.

[0043] Specific embodiment 2 is to learn and evaluate a causal model for increasing the number of likes for media news by a large model robot with personalized male or female identity replying to a user's tweet about sports, entertainment or lifestyle.

[0044] In this specific embodiment, the data of step 201 and step 101 come from a two-week experimental data of 28,457 users on a social platform using 28 male or female identities to automatically reply to the GPT-2 large model robot to increase users' exposure to and participation in media news.

[0045] The target variable Y in this embodiment is the increment of users' likes for media content news.

[0046] Among them, the randomized decision variable or intervention measure A of this embodiment is which large model robot is used to reply and interact with user posts. The value of using a feminine large model robot for interaction is recorded as 0, and the value of using a masculine large model robot for interaction is recorded as 1.

[0047] Among them, the characteristic variable X of this embodiment has 11 items: the number of lists of account functions, the number of likes from the account, the number of posts from the account, the number of accounts followed by the user, the number of accounts following the user, the category of tweets most frequently sent by the user, the number of likes of media news before interaction, the number of media followed before interaction, the number of forwarded media news before interaction, and the number of likes of media news before interaction.

[0048] In some specific implementations, step 202 includes using a target boundary function Y0=Average(Y tr ), where the Average function takes the mean of the target variable Y of the units in the training data set, thereby determining the target variable cutoff point Y0=0.51.

[0049] In some specific implementations, step 203 includes marking the users who interact with the feminine large model robot and whose number of likes for media news after interaction exceeds Y0 as group 0, marking the users who interact with the masculine large model robot and whose number of likes for media news after interaction exceeds Y0 as group 1, and for the remaining units, using an equal probability random allocation strategy, marking the users who interact with the feminine large model robot and whose number of likes for media news after interaction does not exceed Y0 as group 0, and marking the users who interact with the masculine large model robot and whose number of likes for media news after interaction does not exceed Y0 as group 1. The grouping of units is used as a new predictor variable G, that is, the optimal decision label of each unit, which corresponds one-to-one to the value of the decision variable A.

[0050] In some specific implementations, the step 204 of causal model learning includes, based on the training data set S tr ={z i} i=1 22765 , the unit's characteristic variables and optimal decision labels learn the prediction model f:X→G. Specifically, the input of the prediction model f is the unit's 11 characteristic variables X, and the output is the optimal decision label G. Here, the CatBoost tree regression model is used for learning, and the model is denoted as f1. The parameter learning of the CatBoost tree regression model is the default setting of the catboost package in Python. As a control group, the CatBoost tree regression model is also used to learn the prediction model f0:(X,A)→Y, and T i =argmax A Y(X i ,A) Get the decision variable, that is, the decision that the selection model believes will lead to a higher increment of users’ likes for media news after interaction.

[0051] In some specific implementations, step 102 includes, given the causal models to be evaluated f1, f0 and the test data set S te ={z i} i=22765 28457 , input the characteristic variable X of the unit in the test data set into the causal model to obtain the proposed decision G of the causal model, and divide the test units with the same output value G of the causal model into the same group, and the group numbers are group 0 for interaction with the feminine large model and group 1 for interaction with the masculine large model.

[0052] In some specific implementations, step 103 includes calculating the difference matrix D of the potential results of the units in group 0 0 =[0, Y 0A=1 -Y 0 A=0 ; Y 0 A=0 -Y 0 A=1 , 0] and the moment of difference D between the potential results of the units in group 1 1 =[0, Y 1 A=1 -Y 1 A=0 ; Y 1 A=0 -Y 1 A=1 , 0], where the potential outcome Y i A=j It means that if a specific unit z in group i executes decision j, then the expected value of the target variable Y measurement result of unit z is recorded as Y i A=j , infer the potential outcome Y of the units in group i i A=j The average value of the target variable Y of the samples whose unit decision variable of group i takes value j is used.

[0053] In some specific implementations, step 104 includes calculating the decision causal benefit for each unit according to the decision causal benefit function B=B(z), wherein the decision causal benefit function is the average value of the ith column minus the ith row of the difference matrix D_i of the potential results of the group in which unit z is located, that is, for the unit in group 0 to be interacted with the feminized large model, the decision causal benefit is Y 0 A=0 -Y 0 A=1 , for units in group 1 that are intended to interact with the masculinity model, the decision causal payoff is Y 1 A=1 -Y 1 A=0 ; The total causal benefit of all units is the sum of the causal benefits of all units. sum =(Y 0 A=0 -Y 0 A=1 )*n0+(Y 1 A=1 -Y 1 A=0 )*n1, the average causal benefit of decision making is the total causal benefit of decision making divided by the number of units B ave =(Y 0 A=0 -Y 0 A=1 )*n0 / n+(Y1 A=1 -Y 1 A=0 )*n1 / n.

[0054] The experiment was repeated 100 times and the average was taken. Finally, it was concluded that the increase in media news likes for the model f_1 trained using the learning method of this embodiment increased from 0.74 to 0.91, and the average benefit of causal decision was 0.27 likes. The increase in media news likes for the control group model f_0 decreased from 0.74 to 0.63, and the average benefit of causal decision was -0.11 likes.

[0055] Although the above two embodiments describe the use of certain terms to refer to the features of the system or the actions performed by the system. It should be understood that these terms are not the only terms that can be used to describe the operation of the system. Some examples of alternative terms are as follows. For example, a unit may alternatively be referred to as a unit, an instance, an individual, or an element. As another example, a feature variable may alternatively be referred to as a covariate or an independent variable. As another example, a decision variable may alternatively be referred to as an intervention variable or a treatment variable or an intervention measure. As another example, a target variable may alternatively be referred to as an outcome variable or an outcome variable or a variable of interest. As another example, an environment may alternatively be referred to as a context or a contextual state or a system. As another example, a randomized trial may alternatively be referred to as an A / B test.

[0056] Although the decision variables, feature variables, and target variables used in the above two embodiments have specific dimensions, it should be understood that for variables of other different dimensions, a simple generalization of the evaluation and learning methods provided in the present disclosure also belongs to the claims of the present invention. For example, the "target variable dividing point" can be simply generalized to the "target variable dividing hyperplane", and the number of values ​​of a high-dimensional decision variable can be obtained by traversing the grid points of the high-dimensional space. The increase in the dimension of the feature variable can be achieved by simply acquiring more diversified features of the unit.

[0057] In this disclosure, the term "repeatedly", i.e., in the context of repeatedly performing an operation, is generally used to mean that the operation occurs multiple times, with or without a particular sequence. For example, a process may constantly or iteratively follow a set of steps in a specified order, or may follow these steps randomly or non-sequentially. In addition, the steps may not all be performed with the same frequency, for example, the frequency of performing causal model evaluations may be higher than the frequency of updating the causal model, and the frequency of the latter may change over time, for example, as the utilization phase becomes dominant and / or as computing power / speed requirements change over time.

[0058] In this disclosure, the term "randomization" means that a random number is generated through a random process to assign a value to a specific variable or perform a certain behavior and intervention. For example, whether to pop up a window is determined by the result of a coin toss, whether to issue a coupon is determined by computer thermal noise, and whether to invest is determined by a true random number generated by a quantum device. Among them, the random process is random, unpredictable and non-reproducible. If necessary, pseudo-random numbers can be used as an alternative. For example, the random number generated by the pseudo-random number generated by a computer random number table determines the push content.

[0059] In this disclosure, the term "configuration" is used in conjunction with a computer program and an environment. For an environment of one or more computers to be configured to perform a specific operation or action, it means that digital electronic circuits, software, firmware, hardware (including the structures disclosed in this specification and their structural equivalents), or a combination thereof, have been installed in the environment, and these software, firmware, hardware, or a combination thereof cause the system to perform these operations or actions in operation. For one or more computer programs to be configured to perform a specific operation or action, it means that one or more programs include instructions that, when executed by a data processing device, cause the device to perform these operations or actions.

[0060] The embodiments of the subject matter and functional operations described in the present disclosure may be implemented in digital electronic circuits, in tangibly embodied computer software or firmware, in computer hardware (including the structures disclosed in this specification and their structural equivalents), or in a combination of one or more of them. The embodiments of the subject matter described in this specification may be implemented as one or more computer programs. One or more computer programs may include one or more modules of computer program instructions, which are encoded on a tangible non-transitory storage medium for data processing equipment to execute or control the operation of data processing equipment. The computer storage medium may be a machine-readable storage device, a machine-readable storage substrate, a random or serial access memory device, or a combination of one or more of them. Alternatively or in addition, program instructions may be encoded on an artificially generated propagation signal (e.g., a machine-generated electrical signal, an optical signal, or an electromagnetic signal) to be transmitted to a suitable receiver device for execution by a data processing device, and the propagation signal is generated to encode information.

[0061] The term "data processing equipment" refers to data processing hardware, and encompasses all types of equipment, devices, and machines for processing data, including (by way of example) a programmable processor, a computer, or a plurality of processors or computers. The equipment may also be or also include special-purpose logic circuits, such as an FPGA (field programmable gate array) or an ASIC (application-specific integrated circuit). In addition to hardware, the equipment may optionally include code that creates an execution environment for a computer program, for example, code constituting processor firmware, a protocol stack, a database management system, an operating system, or a combination of one or more of them.

[0062] The term "computer program" (which may also be referred to or described as a program, software, software application, app, module, software module, script, or code) may be written in any form of programming language, including compiled or interpreted languages, or declarative or procedural languages; and it may be deployed in any form, including as a stand-alone program or as a module, component, subroutine, or other unit suitable for a computing environment. A computer program may (but need not) correspond to a file in a file system. A program may be stored in a portion of a file that holds other programs or data (e.g., one or more scripts stored in a markup language document), in a single file dedicated to the program in question, or in multiple cooperating files (e.g., files storing one or more modules, subroutines, or portions of code). A computer program may be deployed to execute on one computer, or on multiple computers located at one site or distributed over multiple sites and interconnected by a data communications network.

[0063] In this disclosure, the term "database" is used broadly to refer to any collection of data: the data need not be structured in any particular way, or at all, and it may be stored on a storage device in one or more locations. Thus, for example, an index database may include multiple data collections, each of which may be organized and accessed in a different manner.

[0064] The embodiments of the themes described in this disclosure may be implemented as one or more computer programs. One or more computer programs may include one or more modules of computer program instructions, which are encoded on a tangible non-transient storage medium for data processing equipment to perform or control the operation of data processing equipment. Computer storage media may be a machine-readable storage device, a machine-readable storage substrate, a random or serial access memory device, or a combination of one or more of them. Alternatively or in addition, program instructions may be encoded on an artificially generated propagation signal (e.g., a machine-generated electrical signal, an optical signal, or an electromagnetic signal) to be transmitted to a suitable receiver device for data processing equipment to perform, and the propagation signal is generated to encode information.

[0065] The processes and logic flows described in the present disclosure may be performed by one or more programmable computers executing one or more computer programs to perform functions by operating on input data and generating output. The processes and logic flows may also be performed by a special purpose logic circuit (e.g., an FPGA or ASIC) or by a combination of a special purpose logic circuit and one or more programmed computers.

[0066] Computers suitable for executing computer programs may be based on general or special microprocessors or both, or any other type of central processing unit. Generally speaking, the central processing unit will receive instructions and data from a read-only memory or a random access memory or both. The basic elements of a computer are a central processing unit for implementing or executing instructions and one or more memory devices for storing instructions and data. The central processing unit and the memory may be supplemented by a dedicated logic circuit, or may be incorporated into a dedicated logic circuit. Generally speaking, the computer will also include one or more mass storage devices (e.g., magnetic disks, magneto-optical disks, or optical disks) for storing data, or will be operably connected to receive data from a mass storage device, or to transfer data to a mass storage device, or both. However, the computer does not have to have these devices. In addition, the computer may be embedded in another device, such as a mobile phone, a personal digital assistant (PDA), a mobile audio or video player, a game console, a global positioning system (GPS) receiver, or a portable storage device (e.g., a universal serial bus (USB) flash drive), to name a few.

[0067] Computer-readable media suitable for storing computer program instructions and data include all forms of nonvolatile memory, media, and memory devices, including (by way of example) semiconductor memory devices (e.g., EPROM, EEPROM, and flash memory devices); magnetic disks (e.g., internal hard disks or removable disks); magneto-optical disks; and CD-ROM and DVD-ROM disks.

[0068] To provide for interaction with a user, embodiments of the subject matter described in the present disclosure may be implemented on a computer having a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) for displaying information to a user), a keyboard, and a pointing device (e.g., a mouse or trackball, through which a user can provide input to the computer). Other types of devices may also be used to provide for interaction with a user; for example, feedback provided to a user may be any form of sensory feedback, such as visual feedback, audio feedback, or tactile feedback; and input from a user may be received in any form, including audio, speech, or tactile input. In addition, a computer may interact with a user by sending files to or receiving files from a device used by a user; for example, by sending a web page to a web browser on a user's device in response to a request received from the web browser. In addition, a computer may interact with a user by sending text messages or other forms of messages to a personal device (e.g., a smart phone running a messaging application) and receiving a response message from the user in return.

[0069] Embodiments of the subject matter described in this disclosure may be implemented in a computing system that includes a back-end component (e.g., as a data server), or includes a middleware component (e.g., an application server), or includes a front-end component (e.g., a client computer with a graphical user interface, a web browser, or an app through which a user can interact with a specific implementation of the subject matter described in this specification), or any combination of one or more such back-end, middleware, or front-end components. The components of the system may be interconnected by any form or medium of digital data communication, such as a communication network. Examples of communication networks include local area networks (LANs) and wide area networks (WANs), such as the Internet.

[0070] A computing system may include a client and a server. The client and the server are usually far apart from each other and generally interact through a communication network. The reason for the client and server relationship is that the computer programs that run on the respective computers and have a client-server relationship with each other. In some embodiments, the server transmits data (e.g., an HTML web page) to a user device, for example, for the purpose of displaying data to a user interacting with a device used as a client or receiving user input from the user. Data generated at the device, for example, due to user interaction, can be received from the user device at the server.

[0071] Although the present disclosure includes many specific implementation details, these specific implementation details should not be understood as limitations on the scope of any invention or the scope of the claimed content, but should be understood as descriptions of features that may be specific to specific embodiments of a particular invention. Certain features described in the present disclosure in the context of separate embodiments may also be implemented in combination in a single embodiment. Conversely, the individual features described in the context of a single embodiment may also be implemented individually in multiple embodiments or in any suitable sub-combination. In addition, although features may be described above as functioning in certain combinations and thus even initially protected by claims, one or more features derived from a claimed combination may be separated from the combination in some cases, and a claimed combination may involve sub-combinations or variations of sub-combinations.

[0072] Similarly, although the operations shown in the drawings and described in the claims have a specific order, this should not be understood as requiring that these operations must be performed in the specific order shown or in a sequential order, or that all of the operations shown must be performed to achieve the desired results. In some cases, multitasking and parallel processing may be advantageous. In addition, the separation of the various system modules and components in the above-described embodiments should not be understood as requiring such separation in all embodiments, and should be understood that the program components and systems can generally be integrated together in a single software product or can be packaged into multiple software products.

[0073] Specific embodiments of the subject matter of the present invention have been described. Other embodiments are within the scope of the following claims. For example, the actions described in the claims can be performed in a different order and still achieve the desired results. For example, the processes depicted in the accompanying drawings do not necessarily require the specific order or sequential order shown to achieve the desired results. In some cases, multitasking and parallel processing may be advantageous.

Claims

1. A causal model evaluation method for personalized decision making, characterized in that: The process includes repeatedly performing the following steps: Step 101, obtaining a causal model f with evaluation and a test data set S containing multiple units te , and complete the parameter configuration of the evaluation method; Step 102, for multiple units in the test set, divide the test units with the same proposed decision value G of the causal model f into the same group; Step 103, respectively calculate the difference matrix of the potential results when the decision variable A of the units in each group has different values; Step 104, calculating the causal benefit of the causal model for each unit decision according to the decision benefit function B(z).

2. The causal model evaluation method for personalized decision-making according to claim 1, characterized in that: In step 101, the test data set S te The decision variable A of unit z in is randomly generated, where a measurement result of unit z is a data unit (X, A, Y), X is the feature variable before the decision variable A is generated, A is the decision variable, and Y is the target variable of unit z after the decision variable A is generated.

3. The causal model evaluation method for personalized decision-making according to claim 1, characterized in that: The input of the causal model f is the characteristic variable X of the unit z, and the output is the proposed decision G for the unit z. In the step 102, when the test units with the same value of the proposed decision G of the causal model f are divided into the same group, the group numbers are 0, 1...M-1, where M is the number of possible values ​​of the proposed decision G, which is also the number of possible values ​​of the decision variable A.

4. The causal model evaluation method for personalized decision-making according to claim 1, characterized in that: In step 103, the difference matrix D of the potential results when the decision variables of the units in the i-th group have different values ​​is calculated. i , where the matrix D i There are M-1 rows and M-1 columns, and the element in the jth row and kth column is Y i A=k -Y i A=j , potential outcome Y i A=j It means that if a specific unit z in group i executes decision j, then the expected value of the target variable Y measurement result of unit z is recorded as Y i A=j , infer the potential outcome Y of the units in group i i A=j The average value of the target variable Y of the samples whose unit decision variable of group i takes value j is used.

5. The causal model evaluation method for personalized decision-making according to claim 1, characterized in that: In step 104, the decision causal benefit function B=B(z) can be the difference matrix D of the potential results of the group where the unit z is located i The difference matrix D is the potential result of removing the minimum value of the i-th row or taking the group where the unit z is located. i The average value of the ith column minus the ith row. Further, the decision causal benefit for each unit is calculated according to the decision causal benefit function B=B(z). The total decision causal benefit is the sum of the decision causal benefits of all units, and the average decision causal benefit is the total decision causal benefit of all units divided by the number of units.

6. A causal model learning method for personalized decision making, characterized in that: The method comprises repeatedly performing the following steps: Step 201, obtaining a causal model to be learned f and a training data set S containing multiple units te , and complete the configuration of the learning method; Step 202, determine the target variable cutoff point Y0 according to the target variable Y of the data units of multiple units; Step 203, given a unit in the training data set, the unit is grouped according to the target variable cutoff point Y0 and the decision variable A and target variable Y of the data unit of the unit, so as to generate the optimal grouping label of the unit as the optimal decision label G; Step 204 , learning a causal model based on the feature variables X and the optimal decision labels G of the data units of the plurality of units.

7. The causal model learning method for personalized decision-making according to claim 6, characterized in that: In step 201, the training data set S te The decision variable A of unit z in is randomly generated, where a measurement result of unit z is a data unit (X, A, Y), X is the feature variable before the decision variable A is generated, A is the decision variable, and Y is the target variable of unit z after the decision variable A is generated.

8. The causal model learning method for personalized decision-making according to claim 6, characterized in that: In step 202, the target dividing point Y0 is determined by the target dividing function F, the input of the function F is the target variable Y of all the data units used for training, and the output is the target variable dividing point Y0. The function F can be the maximum M equal division points of the target variable Y of the sample or the maximum M equal division points of the target variable Y interval, where M is the number of possible values ​​of the decision variable A.

9. The causal model learning method for personalized decision-making according to claim 6, characterized in that: In step 203, the units with A=i and Y>Y0 are marked as the i-th group, and the units with A=i and Y≤Y0 are allocated to other groups that are not group i according to the selected remaining unit allocation method G=R(A,Y), where the total number of groups is M, and the group labels are 0, 1...M-1 respectively. Finally, the group label of the unit is used as the prediction variable G, that is, the optimal decision label G for each unit; the selected remaining unit allocation method G=R(A,Y) can be to randomly allocate the units with A=i and Y≤Y0 to other groups that are not group i with equal probability, or to allocate them to other groups that are not group i in turn after being sorted by Y value.

10. The causal model learning method for personalized decision-making according to claim 6, characterized in that: In step 204, according to the training data set S tr , the unit's characteristic variables and optimal decision labels learn a prediction model f:X→G, where the input X of the prediction model f is the unit's characteristic variables, and the output is the optimal decision label G. When learning the causal model f according to the prediction performance, select the model with the highest precision or the highest recall or the maximum F1-Score or the maximum PR-AUC; when using the causal model f, for the specific unit z, make a one-to-one corresponding decision A according to the output G of the causal model f.

Images