A method and device for screening causal features, equipment and storage medium

Abstract

The application discloses a kind of screening method, device and equipment of causal feature and storage medium, the method includes: obtaining the data to be screened including causal feature, and the data to be screened includes transverse federal scene data and longitudinal federal scene data;Condition independence is judged to feature and label variable in the data to be screened;According to condition independence judgment result, feature in the data to be screened is screened in turn, and the final causal feature set is determined.The causal feature screening method provided in the application combines differential privacy and causal feature selection method, ensures that data is not out of the library, realizes the federal feature condition independence detection of multiple participants under the premise, and then completes the constraint-based causal feature selection under the federal learning environment, avoids the coordination party to leak statistical result privacy information, solves the problem that absolute trusted third party cannot be found in engineering implementation, can meet the condition independence judgment demand in different federal scene.

Description

Technical Field

[0001] This invention relates to the field of causal science and technology, and in particular to a method, apparatus, device and storage medium for screening causal features. Background Technology

[0002] Feature selection, as a feature dimensionality reduction technique, is widely used in high-dimensional data analysis scenarios. However, traditional feature selection algorithms generally screen based on the correlation between features and class attributes. Since correlation does not represent causality between features and class attributes, this leads to a lack of interpretability, operability, and robustness in predictive classification models. Causal feature selection discovers the substructures of the Bayesian network for class attributes, namely the Markov blanket (MB). This MB consists of the class attribute's parents (direct causes), children (PCs, direct consequences), and spouses (SPs, other direct causes of the direct consequences), thereby explicitly deriving the causal relationship between local class attributes and features. This allows for the construction of interpretable, operable, and robust predictive classification models.

[0003] In existing technologies, the unseen nature of data in multi-participant federated learning scenarios presents many obstacles to the application of causal feature selection. For example: 1) For data privacy and security reasons, the participants do not share the original data, but only exchange encrypted intermediate data; 2) In current federated learning scenarios, a trusted third party is generally used as the coordinator to process intermediate data or fuse models. However, it is very difficult to find a truly trusted third party in actual production. The coordinator can obtain the results of intermediate interaction data through decryption, which poses a risk of data privacy leakage. Summary of the Invention

[0004] This invention provides a method, apparatus, device, and storage medium for screening causal features, enabling feature selection in multi-participant federated learning scenarios.

[0005] According to one aspect of the present invention, a method for screening causal features is provided, comprising:

[0006] Obtain the data to be filtered that contains the causal features, including horizontal federated scenario data and vertical federated scenario data;

[0007] Perform conditional independence judgment on the features and label variables in the data to be screened;

[0008] Based on the results of the conditional independence judgment, the features in the data to be screened are sequentially screened to determine the final set of causal features.

[0009] Further, obtaining the data to be filtered that includes the causal features includes:

[0010] obtaining original data containing the causal features;

[0011] If in a horizontal federated scenario, performing feature alignment, federated feature engineering and feature value desensitization processing on the original data; if in a vertical federated scenario, performing sample identification alignment, local feature engineering and feature value desensitization processing on the original data;

[0012] The processed original data is used as the data to be screened.

[0013] Further, conditional independence of features and label variables in the data to be screened is judged, including:

[0014] Contingency table statistics are respectively performed on the horizontal federated scenario data and the vertical federated scenario data;

[0015] Conditional independence of features and label variables in the data to be screened is judged according to the contingency table statistics results.

[0016] Further, contingency table statistics are performed on the horizontal federated scenario data, including:

[0017] First sample frequencies of candidate features, label variables and conditional features corresponding to the horizontal federated scenario data are counted, and a local sample frequency contingency table is generated;

[0018] Noise is added to the first sample frequencies to obtain first private sample frequencies;

[0019] Contingency table fusion calculation is performed according to the first private sample frequencies, and a horizontal federated sample frequency contingency table is generated as a contingency table statistics result.

[0020] Further, contingency table statistics are performed on the vertical federated scenario data, including:

[0021] The candidate features, label variables and conditional features corresponding to the vertical federated scenario data are respectively subjected to feature binning processing to obtain respective corresponding sample identification sets;

[0022] According to the feature binning processing results, sample intersections under different feature value combinations are determined as second sample frequencies;

[0023] Noise is added to the second sample frequencies to obtain second private sample frequencies as the contingency table statistics results.

[0024] Further, the noise satisfies a conditional independence noise constraint condition.

[0025] Further, according to the feature binning processing results, sample intersections under different feature value combinations are determined, including:

[0026] According to the binning processing result, hash processing is performed to obtain a hash sample identification set corresponding to each of the candidate feature, the label variable and the conditional feature;

[0027] According to the first random factor corresponding to the candidate feature, the feature binning result is encrypted;

[0028] The encrypted feature binning result is compared using the ciphertext, and the sample intersection is determined according to the intersection of sample identifications under different feature value combinations.

[0029] Further, according to the contingency table statistical result, the conditional independence of the features and the label variable in the to-be-screened data is judged, including:

[0030] According to the contingency table statistical result, the chi-square test value and the degree of freedom are determined;

[0031] According to the chi-square test value and the degree of freedom, the chi-square distribution table is queried to determine whether the features and the label variable satisfy the conditional independence.

[0032] Further, according to the conditional independence judgment result, the features in the to-be-screened data are sequentially screened to determine the final causal feature set, including:

[0033] The features in the to-be-screened data are sequentially screened, and if the features and the label variable satisfy the conditional independence, the corresponding features are added to the candidate Markov blanket;

[0034] The features in the candidate Markov blanket are screened again, and the candidate Markov blanket from which the false positive features are removed is determined as the final causal feature set.

[0035] According to another aspect of the present application, a causal feature screening device is provided, including:

[0036] A to-be-screened data acquisition module is configured to acquire to-be-screened data containing the causal features, wherein the to-be-screened data includes transverse federal scenario data and longitudinal federal scenario data;

[0037] A conditional independence judgment module is configured to judge the conditional independence of the features and the label variable in the to-be-screened data;

[0038] A causal feature set determination module is configured to sequentially screen the features in the to-be-screened data according to the conditional independence judgment result to determine the final causal feature set.

[0039] Optionally, the to-be-screened data acquisition module is further configured to:

[0040] Acquire original data containing the causal features;

[0041] If in the horizontal federated scenario, the original data is subjected to feature alignment, federated feature engineering and feature value desensitization processing; if in the vertical federated scenario, the original data is subjected to sample identification alignment, local feature engineering and feature value desensitization processing.

[0042] The processed original data is taken as the data to be screened.

[0043] Optionally, the conditional independence judgment module is further configured to:

[0044] The contingency table statistics is performed on the horizontal federated scenario data and the vertical federated scenario data respectively.

[0045] The conditional independence of the features and the label variables in the data to be screened is judged according to the contingency table statistics result.

[0046] Optionally, the conditional independence judgment module is further configured to:

[0047] The first sample frequency of the candidate features, the label variables and the conditional features corresponding to the horizontal federated scenario data is counted, and a local sample frequency contingency table is generated;

[0048] Noise is added to the first sample frequency to obtain a first private sample frequency;

[0049] Contingency table fusion calculation is performed according to the first private sample frequency, and a horizontal federated sample frequency contingency table is generated as the contingency table statistics result.

[0050] Optionally, the conditional independence judgment module is further configured to:

[0051] The candidate features, the label variables and the conditional features corresponding to the vertical federated scenario data are subjected to feature binning processing respectively, and a sample identification set corresponding to each is obtained;

[0052] According to the feature binning processing result, a sample intersection under different feature value combinations is determined as a second sample frequency;

[0053] Noise is added to the second sample frequency to obtain a second private sample frequency as the contingency table statistics result.

[0054] Optionally, the noise satisfies a conditional independence noise constraint condition.

[0055] Optionally, the conditional independence judgment module is further configured to:

[0056] According to the binning processing result, hash processing is performed to obtain a hash sample identification set corresponding to each of the candidate features, the label variables and the conditional features;

[0057] The feature binning result is encrypted according to a first random factor corresponding to the candidate features.

[0058] The encrypted feature binning results are compared using ciphertext, and the sample intersection is determined according to the intersection of sample identifiers under different feature value combinations.

[0059] Optionally, the conditional independence judgment module is further configured to:

[0060] determine a chi-square test value and a degree of freedom according to the contingency table statistical result;

[0061] determine whether the feature and the label variable satisfy conditional independence according to the chi-square test value and the degree of freedom by querying a chi-square distribution table.

[0062] Optionally, the causal feature set determination module is further configured to:

[0063] sequentially filter features in the to-be-filtered data, and if the features and the label variable satisfy conditional independence, add corresponding features to a candidate Markov blanket;

[0064] perform secondary filtering on the features in the candidate Markov blanket, and determine a candidate Markov blanket from which false positive features are removed as a final causal feature set.

[0065] According to another aspect of the present application, an electronic device is provided, and the electronic device comprises:

[0066] at least one processor; and

[0067] a memory connected in communication with the at least one processor; wherein

[0068] the memory stores a computer program executable by the at least one processor, and the computer program is executed by the at least one processor to enable the at least one processor to perform the causal feature filtering method according to any one of the embodiments of the present application.

[0069] According to another aspect of the present application, a computer readable storage medium is provided, and the computer readable storage medium stores computer instructions for enabling a processor to perform the causal feature filtering method according to any one of the embodiments of the present application when executed by the processor.

[0070] The embodiment of the present application provides a screening method of causal features, first acquires to-be-screened data containing causal features, the to-be-screened data includes horizontal federated scene data and longitudinal federated scene data, then conditional independence of features and label variables in the to-be-screened data is judged, and finally, according to the conditional independence judgment result, the features in the to-be-screened data are screened in sequence to determine a final causal feature set. The screening method of causal features provided by the embodiment of the present application combines differential privacy and causal feature selection methods, realizes federated feature conditional independence detection of multiple participants under the premise that data is not exported from a library, and then completes constraint-based causal feature selection in a federated learning environment, avoids leaking statistical result privacy information by a coordinator, solves the problem that an absolutely trusted third party cannot be found in engineering implementation, and can meet the conditional independence judgment demand in different federated scenes.

[0071] It should be understood that the content described in this part is not intended to identify the key or important features of the embodiments of the present application, nor is it intended to limit the scope of the present application. Other features of the present application will become apparent from the following description. BRIEF DESCRIPTION OF DRAWINGS

[0072] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor.

[0073] Figure 1 is a flow chart of a screening method of causal features according to an embodiment of the present application;

[0074] Figure 2 is a flow chart of a screening method of causal features according to an embodiment of the present application;

[0075] Figure 3 is a causal feature selection framework in a horizontal federated scene according to an embodiment of the present application;

[0076] Figure 4 is a causal feature selection framework in a longitudinal federated scene according to an embodiment of the present application;

[0077] Figure 5 is a structural schematic diagram of a screening device of causal features according to an embodiment of the present application;

[0078] Figure 6 is a structural schematic diagram of an electronic device for implementing the screening method of causal features according to an embodiment of the present application. DETAILED DESCRIPTION

[0079] In order to make the person skilled in the art better understand the present application, the technical solutions in the embodiments of the present application will be described clearly and completely in the following with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should belong to the scope of protection of the present application.

[0080] It should be noted that the terms "first", "second" and the like in the specification and claims of the present application and the above-described drawings are used to distinguish similar objects, and do not necessarily have to describe a specific order or sequence. It should be understood that the data used in this way can be exchanged under appropriate circumstances, so that the embodiments of the application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device that includes a series of steps or units does not have to be limited to those steps or units clearly listed, but can include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0081] Embodiment one

[0082] Figure 1 A flowchart of a cause-effect feature screening method is provided for the first embodiment of the present application. The present embodiment can be applied to the screening of cause-effect features. The method can be performed by a cause-effect feature screening device, which can be realized in the form of hardware and / or software, and can be configured in an electronic device. As shown in the figure, the method comprises: Figure 1

[0083] S110, obtaining cause-effect feature containing to-be-screened data, the to-be-screened data including horizontal federated scenario data and vertical federated scenario data.

[0084] ​Wherein, the causal feature is a feature with a causal relationship. The federated machine learning is a machine learning framework that can effectively help multiple institutions to perform data use and machine learning modeling under the requirements of user privacy protection, data security and government regulations. Further, the horizontal federated learning, also known as sample-partitioned federated learning or example-partitioned federated learning, can be applied to the scenario that the data sets of each participant of the federated learning have the same feature space and different sample spaces. The federated learning composed of participants with the same sample space and different feature spaces on the data set is classified as vertical federated learning (VFL), which can also be understood as feature-partitioned federated learning.

[0085] For example, a bank and an e-commerce cooperate, the bank has a credit label, and the e-commerce has consumption data. After both parties get a batch of the same users, the bank wants to evaluate the data effect of the e-commerce, but the bank does not want to disclose its credit label, and the e-commerce does not want to disclose its consumption data. In this case, the vertical federated learning in the federated learning method can be used. For different subsidiaries of the same bank or different banks, their users are different, and the data features are basically the same. At this time, the business cooperation between different subsidiaries or different banks can use the horizontal federated learning in the federated learning method. The federated learning can make each enterprise jointly train without sharing data through an encrypted distributed machine learning framework, and solve the problem of "data island".

[0086] In the embodiment, before the causal feature screening starts, the data of each participant needs to be preprocessed, and the data after preprocessing is taken as the to-be-screened data. The data preprocessing can be sample alignment, local and federated feature engineering, data desensitization and the like. According to the data type, the data in the horizontal federated scenario and the data in the vertical federated scenario can adopt different preprocessing modes.

[0087] In S120, the conditional independence of the features and the label variable in the to-be-screened data is judged.

[0088] In this step, we mainly care about whether a random variable (a certain feature) and another random variable (a label) are independent of each other. If they are independent, it can be said that the feature does not play a role in determining the label variable, that is, we cannot judge whether the sample belongs to the classification corresponding to the label according to the presence or absence of the feature.

[0089] Optionally, the conditional independence judgment method includes a chi-square test method based on a contingency table, an F test method based on linear correlation, and a conditional independence test method based on mutual information. The embodiments of the present application are not limited to a specific method of conditional independence judgment. Taking the chi-square test as an example, the chi-square test is a commonly used method for testing the conditional independence of two variables in mathematical statistics. The most basic idea is to determine whether the theory is correct by observing the deviation between the actual value and the theoretical value. When performing the method, it is often assumed that the two variables are indeed independent, and then the deviation between the actual value and the theoretical value is observed. If the deviation is small enough, we believe that the error is a natural sample error, which is caused by inaccurate measurement or occurs accidentally, and the two variables are indeed independent. At this time, the original hypothesis is accepted. If the deviation is large enough, such error is unlikely to occur accidentally or due to inaccurate measurement, and we believe that the two variables are not independent, that is, the original hypothesis is denied, and the alternative hypothesis is accepted.

[0090] For other conditional independence judgment methods, such as the mutual information-based method and the F detection method based on correlation, the calculation methods are different, but they are all conditional independence detection methods based on the statistical results of federated data. The difference is that the statistical objects are different.

[0091] S130, according to the conditional independence judgment result, the features in the to-be-screened data are sequentially screened to determine the final causal feature set.

[0092] In this embodiment, based on the conditional independence judgment result, the coordinator can select causal features accordingly.

[0093] Optionally, all to-be-screened features can be sequentially screened. The current feature is taken as a candidate feature. If the candidate feature is not independent of the label variable, the candidate feature is added to the candidate feature set. Otherwise, the feature is discarded, and the next feature is selected as a candidate feature for continuous judgment until all features are screened.

[0094] The embodiments of the present application first obtain to-be-screened data containing causal features, the to-be-screened data including horizontal federated scenario data and vertical federated scenario data, then perform conditional independence judgment on the features and the label variable in the to-be-screened data, and finally screen the features in the to-be-screened data according to the conditional independence judgment result to determine the final causal feature set. The causal feature screening method provided by the embodiments of the present application combines differential privacy and causal feature selection method, realizes conditional independence detection of federated features of multiple participants under the premise that data is not exported, completes constraint-based causal feature selection in a federated learning environment, avoids the leakage of private information of the statistical results by the coordinator, solves the problem that an absolutely trusted third party cannot be found in engineering implementation, and can meet the conditional independence judgment demand in different federated scenarios.

[0095] Embodiment Two

[0096] Figure 2 A flowchart of a screening method of a causal feature provided for Embodiment Two of the present application, which is a refinement of the above-mentioned embodiment. As shown in the figure, the method comprises: Figure 2

[0097] S210, obtaining original data containing a causal feature.

[0098] In this embodiment, the original data for causal feature screening can be obtained according to actual needs, and the original data can be provided by each participant in federated learning.

[0099] In this embodiment, the original data for causal feature screening can be obtained according to actual needs, and the original data can be provided by each participant in federated learning.

[0100] S220, if in a horizontal federated scenario, then performing feature alignment, federated feature engineering and feature value desensitization processing on the original data; if in a vertical federated scenario, then performing sample ID alignment, local feature engineering and feature value desensitization processing on the original data.

[0101] In this embodiment, different preprocessing methods can be used for horizontal federated scenarios and vertical federated scenarios. Assuming that there are two participants A and B, one coordinator C, the feature variable set is F, and the label variable is Y, then in the horizontal federated scenario, each participant has the feature set F and the label variable Y; in the vertical federated scenario, participant A has part of the features FA(FA∈F), and participant B has the remaining part of the features FB(FB∈F) and the label variable Y.

[0102] Specifically, the horizontal federated scenario can perform feature alignment, federated feature engineering and feature value desensitization processing; the vertical federated scenario can perform sample ID alignment, local feature engineering and feature value desensitization processing.

[0103] Feature engineering is the most important part of machine learning modeling. The biggest difference between federated feature engineering and traditional feature engineering is that the feature processing (and possibly the monitoring part) in federated feature engineering relies on encrypted data, and also requires data integration and calculation in the cloud. Feature alignment is to remove non-overlapping features to ensure that the features of the participants in federated learning are the same; sample ID alignment is a private set intersection (PSI) technology, which means that the data holder can calculate the intersection of the data sets of each party without exposing any data set information outside the intersection; feature value desensitization is to encrypt the original data, the purpose of which is to remove or protect the privacy information of the original data.

[0104] ​S230, the processed original data is taken as the data to be screened.

[0105] In this embodiment, the original data is processed as described above and then taken as the data to be screened.

[0106] S240, respectively, for the horizontal federated scenario data and the vertical federated scenario data, the contingency table statistics.

[0107] The contingency table is a frequency table listed when the observation data is classified according to two or more attributes (qualitative variables). It is a frequency distribution table of cross classification of two or more variables. Contingency table analysis is a related statistical analysis and inference based on the contingency table. The basic problem of contingency table analysis is to determine whether there is a correlation between the attributes under study, that is, whether they are independent.

[0108] In this embodiment, when the causal feature screening is performed on the data to be screened, the contingency table statistics is first performed. There are different ways of contingency table statistics for the horizontal federated scenario and the vertical federated scenario.

[0109] Optionally, the way of performing the contingency table statistics on the horizontal federated scenario data can be: counting the first sample frequency of the candidate features, the label variable and the condition features corresponding to the horizontal federated scenario data, generating a local sample frequency contingency table; adding noise to the first sample frequency to obtain a first private sample frequency; performing contingency table fusion calculation according to the first private sample frequency to generate a horizontal federated sample frequency contingency table as the contingency table statistics result.

[0110] In this embodiment, the noise satisfies the condition independent noise constraint condition.

[0111] Specifically, in the horizontal federated learning scenario, the contingency table statistics is realized through three steps of local contingency table statistics, local differential privacy method encryption statistics result based on condition independent noise constraint, and federated contingency table fusion calculation. The specific implementation steps are as follows:

[0112] 1) Local contingency table statistics: each participant respectively counts the sample frequency among the candidate features, the label variable and the condition features, and generates a local sample frequency contingency table.

[0113] 2) Privacy computation with conditional independence noise constraint: Each participant encrypts all sample frequencies in the contingency table statistics using a privacy computation method with conditional independence noise constraint. Assuming x represents the sample frequency variable in the contingency table, r(x) represents the noise added to the frequency, differential privacy Laplace mechanism noise or Gaussian mechanism noise can be used, k(x) is a function based on the conditional independence detection method of the contingency table, f(x) is the sample frequency after differential privacy processing. The principle of the differential privacy method based on the conditional independence noise constraint is to ensure that the result of |k(x)-k(x+r(x))| is as small as possible within the maximum change range of the conditional independence detection value under a given significance level (usually 0.05). Assuming the degree of freedom is n, it is known from the chi-square test critical value table that the conditional independence detection value under a significance level of 0.05 is k1, and the conditional independence detection value under a significance level of 0.1 is k2. Therefore, the maximum change value of the conditional independence detection value under a given significance level of 0.05 is: k1-k2, that is: |k(x)-k(x+r(x))| < k1

[0114] k2. Add r(x) that satisfies the conditional independence noise constraint to the sample frequency x to obtain the sample frequency f(x) after differential privacy processing, that is: f(x) = x+r(x).

[0115] 3) Federated contingency table fusion computation: Send the desensitized candidate features, label variables and conditional feature values to the coordination party together with the encrypted sample frequency contingency table statistics in step 2), and perform contingency table fusion computation by the coordination party, that is, add the sample frequencies under the same feature value combination to obtain the federated sample frequency contingency table.

[0116] Optionally, the way to perform contingency table statistics on longitudinal federated scene data can be: performing feature binning on the candidate features, label variables and conditional features corresponding to the longitudinal federated scene data respectively to obtain the sample identification sets corresponding thereto respectively; determining the sample intersection under different feature value combinations as the second sample frequency according to the feature binning results; and adding noise to the second sample frequency to obtain the second private sample frequency as the contingency table statistics result.

[0117] In the embodiment, the above noise satisfies the conditional independence noise constraint condition.

[0118] Specifically, in the longitudinal federated learning scene, the contingency table statistics is realized through three steps of local feature binning, federated feature binning intersection computation and privacy computation method with conditional independence noise constraint. The specific implementation steps are as follows:

[0119] 1) Local feature binning: the participants holding the candidate feature X, the label variable Y and the conditional feature Z respectively perform feature binning on the above candidate feature, label variable and conditional feature locally to obtain all feature values of the candidate feature X and their corresponding sample ID set ID X , all label values of the label variable Y and their corresponding sample ID set ID Y , and all feature values of the conditional feature Z and their corresponding sample ID set ID Z .

[0120] 2) Intersection calculation of federated feature binning: according to the feature binning processing result, the sample intersection under different feature value combinations is determined as the second sample frequency. The label party obtains the encrypted feature binning results of each participant, can calculate the intersection of sample IDs in the feature binning under different feature value combinations through ciphertext comparison, and then obtain the federated contingency table statistical result under the conditional feature.

[0121] 3) The label party sends the encrypted federated contingency table statistical result to the coordinator: the label party encrypts all sample frequencies in the federated contingency table obtained by the differential privacy method based on conditional independence noise constraint, and the details are the same as step 2) in the horizontal federated learning scene, and then sends the encrypted federated contingency table statistical result to the coordinator.

[0122] Further, the way to determine the sample intersection under different feature value combinations according to the feature binning processing result can be: performing hash processing according to the binning processing result to obtain the hash sample identification set corresponding to the candidate feature, the label variable and the conditional feature respectively; encrypting the feature binning result according to the first random factor corresponding to the candidate feature; comparing the encrypted feature binning result using ciphertext, and determining the sample intersection according to the intersection of sample identifications under different feature value combinations.

[0123] Specifically, the intersection calculation of federated feature binning can be implemented according to the following steps:

[0124] 1) Hash calculation of local feature binning result: each participant locally performs Hash calculation on the feature binning result obtained in step 1) in the vertical federated scene to obtain the candidate feature H(X) and its corresponding sample ID set H(ID X ), the label variable H(Y) and its corresponding sample ID set H(ID Y ), and the conditional feature H(Z) and its corresponding sample ID set H(ID Z ).

[0125] 2) The candidate feature party generates a random factor k, multiplies all candidate feature values H(X) and their corresponding sample ID set H(ID X ) to obtain k x H(X) and k x H(IDX Send to the tagger.

[0126] 3) The label generator generates a random factor r, which is then multiplied by all label values ​​H(Y) and their corresponding sample ID set H(ID). Y ), to obtain r×H(Y) and r×H(ID) Y The random factor p is sent to the candidate feature providers. Other non-candidate feature providers generate different random factors p, which are multiplied by the set of feature H(Z) and its corresponding sample IDs H(ID). Z ), to obtain p×H(Z) and p×H(ID) Z Send to the candidate feature provider.

[0127] 4) The candidate feature generator uses a random factor k, multiplied by the label value r×H(Y) sent by the label generator and its corresponding set of sample IDs r×H(ID). Y ), to obtain k×r×H(Y) and k×r×H(ID) Y The candidate feature provider sends the data to the tag provider. Simultaneously, the candidate feature provider multiplies the held feature value p×H(Z) and its corresponding sample ID set p×H(ID) sent by the non-candidate feature provider by a random factor k. Z ), to obtain k×p×H(Z) and k×p×H(ID) Z ), and send it to the corresponding non-candidate feature side.

[0128] 5) The labeling method uses the inverse of the random factor r. -1 Multiply by k×r×H(Y) and k×r×H(ID) Y ), to obtain k×H(Y) and k×H(ID) Y Non-candidate feature squares are calculated by multiplying the inverse p-1 of the random factor p by k×p×H(Z) and k×p×H(ID). Z ), to obtain k×H(Z) and k×H(ID) Z ), and send it to the tagger.

[0129] 6) At this point, the labeler obtains the feature binning results encrypted by all participants using the same random factor k. The intersection of sample IDs in the feature bins under different feature value combinations can be calculated by comparing the ciphertext, thereby obtaining the statistical results of the federated contingency table under the conditional features.

[0130] S250. Based on the statistical results of the contingency table, determine the conditional independence of the features and label variables in the data to be screened.

[0131] In this embodiment, after performing contingency table statistics, the conditional independence of feature and label variables can be determined.

[0132] Optionally, the conditional independence judgment of the features in the data to be screened and the label variable according to the contingency table statistical result can be: determining the chi-square test value and the degree of freedom according to the contingency table statistical result; determining whether the features and the label variable satisfy the conditional independence according to the chi-square test value and the degree of freedom by querying the chi-square distribution table.

[0133] Specifically, based on the federal contingency table statistical result, the coordinator can perform chi-square test calculation, and the chi-square statistic in the conditional independence test is the sum of all chi-square statistics of the contingency table under the conditional feature value combination. The calculation formula is:

[0134]

[0135] wherein χ 2 is the chi-square value, i is the i th feature value of the candidate feature, j is the j th label value of the label variable, k is the k th feature value combination of the conditional feature, is the actual frequency in the cell of the contingency table; is the expected frequency when the candidate feature and the label variable are independent. By calculating the chi-square test and the degree of freedom, the conditional independence of the candidate feature and the label variable can be obtained by querying the chi-square distribution table. The degree of freedom is 2 |z| when the conditional independence is detected, and |z| is the number of variables in the conditional feature set.

[0136] S260, screening the features in the data to be screened in sequence, and if the features and the label variable satisfy the conditional independence, adding the corresponding features to the candidate Markov blanket.

[0137] The Markov blanket (MB) is a minimum feature subset that satisfies the following properties: a feature is conditionally independent of all other features in the feature domain under its Markov blanket. Let the Markov blanket of feature T be MB(T), then the above can be expressed as: P(T|MB(T))=P(T|Y,MB(T)), wherein Y is all non-Markov blanket nodes in the feature domain.

[0138] In this embodiment, based on the conditional independence judgment result, the coordinator can perform causal feature selection accordingly. If the candidate feature and the label variable are not independent under the conditional feature, the candidate feature is added to the candidate MB (Markov blanket); otherwise, the feature is discarded, and the next feature is selected as the candidate feature, until all features are screened.

[0139] S270, performing secondary screening on the features in the candidate Markov blanket, and determining the candidate Markov blanket with false positive features removed as the final causal feature set.

[0140] In the embodiment, the candidate MB obtained after the first screening can contain false positive features, that is, candidate features and label variables are independent of each other, so that the false positive features in the candidate MB (Markov blanket) can be removed through secondary screening to obtain a final causal feature set.

[0141] Figure 3 and Figure 4 respectively provide a causal feature selection framework in a horizontal federated scenario and a vertical federated scenario provided by the embodiment of the application, assuming that there are 2 participants A and B, 1 coordinator C, a feature variable set F, and a label variable Y. In the horizontal federated scenario, each participant has a feature set F and a label variable Y. In the vertical federated scenario, participant A has part of the features FA (FA∈F), participant B has the remaining part of the features FB (FB∈F) and the label variable Y.

[0142] As shown in Figure 3 in the horizontal federated scenario, after feature alignment, federated feature engineering and feature value desensitization processing, the coordinator initializes Then the coordinator randomly selects a feature X from the feature set F as a candidate feature, that is, The candidate feature X and the conditional feature set Z (Z=CMB(Y)) are sent to each participant. Each participant locally counts the sample frequency contingency table between the candidate feature variable X, the label variable Y and the conditional variable Z (Z=CMB(Y)). Each participant uses a differential privacy algorithm based on conditional independence noise to encrypt the sample frequency in the local contingency table statistics result. The coordinator receives the encrypted sample frequency statistics result sent by each participant and performs fusion calculation. The coordinator performs conditional independence judgment calculation based on the fused sample frequency contingency table result. The coordinator updates the candidate MB set CMB(Y) and the feature set F according to the conditional independence judgment result, specifically: if the feature X is conditionally independent of the label variable Y, the feature X is added to the candidate MB set CMB(Y), and X is deleted from the feature set F, that is: CMB(Y)=CMB(Y)∪X, F=F\X. The coordinator continues to randomly select a feature from the feature set F, and repeatedly executes the steps of selecting a candidate feature until the candidate MB set CMB(Y) and the feature set F are updated according to the conditional independence judgment result until there is no feature in the feature set F. The coordinator randomly selects a feature X from the current CMB(Y), Let Z be CMB(Y) without X, i.e., Z = CMB(Y)\X. Send features X and Z to each participant and perform the contingency table statistics to conditional independence judgment steps. Based on the conditional independence judgment results, update the candidate MB set CMB(Y). Specifically, if a feature is conditionally independent of the label variable Y, then remove feature X from the candidate MB set CMB(Y), i.e., CMB(Y) = CMB(Y)\X, until no feature can be removed from CMB(Y). The final output CMB(Y) is the final causal feature set in this embodiment. Taking federated learning of a disease prediction model as an example, suppose that medical institutions A and B in a certain region each have insufficient data, but both have the same data features and can cooperate on data. Medical institutions A and B, as two participating parties, locally compile contingency tables of sample frequencies among candidate feature variables X, label variables Y, and condition variables Z. They then encrypt the sample frequencies in their local contingency table statistics using a differential privacy algorithm based on conditional independence noise, and send the encrypted sample frequency statistics to a third-party coordinator. The coordinator performs conditional independence checks on the sample frequency contingency table results after fusion calculations and updates the candidate MB set based on the results. The candidate MB set is then processed according to the above steps to finally obtain the causal feature set. By jointly using the data from both institutions A and B, a disease prediction model is established, thereby improving the diagnosis rate of the disease in actual medical services.

[0143] like Figure 4 As shown, in a vertical federated scenario, after sample ID alignment, local feature engineering, and feature value anonymization, all participants locally bin all feature variables and label variables, and count the sample ID set corresponding to each feature bin. The coordinator initializes CMB(Y) = φ, and the coordinator randomly selects X from the feature set F as a candidate feature, i.e. The candidate feature X and the conditional feature set Z (Z=CMB(Y)) are sent to the participant holding the feature. After the federal contingency table statistics, the label party encrypts all sample frequency contingency tables obtained by statistical methods based on conditional independence noise constraints, and then sends the encrypted contingency table statistics to the coordinator. After receiving the sample frequency contingency table results, the coordinator performs conditional independence judgment, that is, calculates the conditional independence of X and Y under the given condition Z. According to the conditional independence detection result of X and Y, the coordinator updates the candidate MB set CMB(Y) and the feature set F, specifically: if the feature X is not conditionally independent of the label variable Y, the feature X is added to the candidate MB set CMB(Y), and X is deleted from the feature set F, that is: CMB(Y)=CMB(Y)∪X, F=F\X. The subsequent steps are similar to those in the horizontal federal scenario, and are not described here. Taking the federal learning of group rental house intelligent analysis as an example: the active party is a certain power department, which provides electricity data (including the label of whether it is a group rental house) and defines model parameters and other information as the initiator of the federal learning task; the participant is the government water-related department, which provides water consumption data and participates in federal modeling (without labels); the coordinator is deployed in the relevant group rental house management department, as a safe and supervised third party, responsible for providing computing power and distributing secret keys. In this example, after sample ID alignment, local feature engineering and feature value desensitization processing, the certain power department and the government water-related department as participants perform feature binning on all feature variables and label variables locally, and count the sample ID sets corresponding to each feature bin. The group rental house management department as the coordinator takes out the candidate features from the feature set and sends the candidate features and the conditional feature set to the corresponding participant holding the feature. After federal contingency table statistics, the power department as the label party encrypts all sample frequency contingency tables obtained by statistical methods based on conditional independence noise constraints, and then sends the encrypted contingency table statistics to the group rental house management department. After receiving the sample frequency contingency table results, the group rental house management department performs conditional independence judgment, and finally outputs the causal feature set according to the conditional independence detection result.

[0144] The embodiment of the application first acquires original data containing causal features, if in a horizontal federated scenario, then performs feature alignment, federated feature engineering and feature value desensitization processing on the original data, if in a vertical federated scenario, then performs sample identification alignment, local feature engineering and feature value desensitization processing on the original data, then takes the processed original data as screening data, then performs contingency table statistics on the horizontal federated scenario data and the vertical federated scenario data respectively, then performs conditional independence judgment on the features and label variables in the screening data according to the contingency table statistics result, then sequentially screens the features in the screening data, if the features and the label variables satisfy the conditional independence, then adds the corresponding features to a candidate Markov blanket, and finally performs secondary screening on the features in the candidate Markov blanket, and determines the candidate Markov blanket from which false positive features are removed as the final causal feature set. The causal feature screening method provided by the embodiment of the application combines differential privacy and causal feature selection method, realizes conditional independence detection of federated features of multiple participants under the premise that data is not exported, and then completes constraint-based causal feature selection in a federated learning environment, avoids leaking statistical result privacy information by a coordinator, solves the problem that an absolutely trusted third party cannot be found in engineering implementation, and can meet the conditional independence judgment demand in different federated scenarios.

[0145] Embodiment three

[0146] Figure 5 A structural schematic diagram of a causal feature screening device provided by the third embodiment of the application is shown in FIG. 3. Figure 5 As shown in the figure, the device comprises a screening data acquisition module 310, a conditional independence judgment module 320 and a causal feature set determination module 330.

[0147] The screening data acquisition module 310 is configured to acquire screening data containing the causal features, and the screening data comprises horizontal federated scenario data and vertical federated scenario data.

[0148] The conditional independence judgment module 320 is configured to perform conditional independence judgment on the features and label variables in the screening data.

[0149] The causal feature set determination module 330 is configured to sequentially screen the features in the screening data according to the conditional independence judgment result, and determine the final causal feature set.

[0150] Optionally, the screening data acquisition module 310 is further configured to:

[0151] Obtain the original data containing causal features; in a horizontal federated scenario, perform feature alignment, federated feature engineering, and feature value desensitization on the original data; in a vertical federated scenario, perform sample label alignment, local feature engineering, and feature value desensitization on the original data; use the processed original data as the data to be screened.

[0152] Optionally, the conditional independence judgment module 320 is also used for:

[0153] Contingency table statistics were performed on the horizontal and vertical federated scenario data respectively; based on the contingency table statistics results, conditional independence was determined for the features and label variables in the data to be screened.

[0154] Optionally, the conditional independence judgment module 320 is also used for:

[0155] The frequency of the first sample corresponding to the candidate features, label variables and conditional features in the horizontal federated scenario data is counted to generate a local sample frequency contingency table; noise is added to the first sample frequency to obtain the first privacy sample frequency; contingency table fusion calculation is performed based on the first privacy sample frequency to generate a horizontal federated sample frequency contingency table as the contingency table statistical result.

[0156] Optionally, the conditional independence judgment module 320 is also used for:

[0157] The candidate features, label variables, and conditional features corresponding to the longitudinal federated scenario data are binned to obtain their respective sample identifier sets. Based on the binning results, the intersection of samples under different feature value combinations is determined as the second sample frequency. Noise is added to the second sample frequency to obtain the second privacy sample frequency as the contingency table statistical result.

[0158] Optionally, the noise must satisfy the noise independence constraint condition.

[0159] Optionally, the conditional independence judgment module 320 is also used for:

[0160] The binning results are hashed to obtain the hashed sample identifier sets corresponding to the candidate features, label variables, and conditional features. The feature binning results are encrypted according to the first random factor corresponding to the candidate features. The encrypted feature binning results are compared using the ciphertext, and the sample intersection is determined based on the intersection of the sample identifiers under different feature value combinations.

[0161] Optionally, the conditional independence judgment module 320 is also used for:

[0162] Determine the chi-square test value and degrees of freedom based on the statistical results of the contingency table; consult the chi-square distribution table based on the chi-square test value and degrees of freedom to determine whether the feature and label variables meet the condition of independence.

[0163] Optionally, the cause-effect feature set determination module 330 is further configured to:

[0164] The features in the data to be screened are sequentially screened, if the features and the label variable satisfy the conditional independence, the corresponding features are added to the candidate Markov blanket, and the features in the candidate Markov blanket are secondarily screened, the candidate Markov blanket from which the false positive features are removed is determined as the final cause-effect feature set.

[0165] The cause-effect feature screening device provided in the embodiments of the present application can execute the cause-effect feature screening method provided in any of the embodiments of the present application, and has the corresponding function modules and beneficial effects of the execution method.

[0166] Embodiment four

[0167] Figure 6 A structural schematic diagram of an electronic device 10 that can be used to implement embodiments of the present application is shown. The electronic device is intended to represent various forms of digital computers, such as laptops, desktops, tablets, personal digital assistants, servers, blade servers, mainframes, and other appropriate computers. The electronic device can also represent various forms of mobile devices, such as personal digital processors, cellular telephones, smart phones, wearable devices (e.g., headsets, glasses, watches, etc.), and other similar computing devices. The components shown herein, their connections and relationships, and their functions, are meant to be examples only, and are not intended to limit the implementations of the present application described and / or claimed in this document.

[0168] As shown in Figure 6 The electronic device 10 includes at least one processor 11, and a memory, such as a read-only memory (ROM) 12, a random access memory (RAM) 13, etc., which are communicatively connected to the at least one processor 11, wherein the memory stores a computer program that can be executed by the at least one processor, and the processor 11 can perform various appropriate actions and processes according to the computer program stored in the read-only memory (ROM) 12 or the computer program loaded from the storage unit 18 into the random access memory (RAM) 13. In the RAM 13, various programs and data required for the operation of the electronic device 10 can also be stored. The processor 11, the ROM 12, and the RAM 13 are connected to each other through a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14.

[0169] A plurality of components in the electronic device 10 are connected to the I / O interface 15, including: an input unit 16, such as a keyboard, a mouse, etc.; an output unit 17, such as various types of displays, speakers, etc.; a storage unit 18, such as a magnetic disk, an optical disk, etc.; and a communication unit 19, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 19 allows the electronic device 10 to exchange information / data with other devices through a computer network, such as the Internet, and / or various telecommunication networks.

[0170] The processor 11 can be various general and / or special purpose processing components with processing and computing capabilities. Some examples of the processor 11 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, a digital signal processor (DSP), and any appropriate processor, controller, microcontroller, etc. The processor 11 performs various methods and processes described above, such as the causal feature screening method.

[0171] In some embodiments, the causal feature screening method can be implemented as a computer program tangibly embodied in a computer readable storage medium, such as the storage unit 18. In some embodiments, part or all of the computer program can be loaded and / or installed onto the electronic device 10 via the ROM 12 and / or the communication unit 19. When the computer program is loaded onto the RAM 13 and executed by the processor 11, one or more steps of the causal feature screening described above can be performed. Alternatively, in other embodiments, the processor 11 can be configured to perform the causal feature screening method by any other appropriate means, such as by means of firmware.

[0172] Various implementations of the systems and techniques described above can be realized in digital electronic circuitry, integrated circuitry, a field programmable gate array (FPGA), an application specific integrated circuit (ASIC), a system on a chip (SOC), a programmable logic device (PLD), a computer hardware, firmware, software, and / or combinations thereof. These various implementations can include implementation in one or more computer programs that are executable and / or interpretable on a programmable system including at least one programmable processor, which can be special or general purpose, coupled to receive data and instructions from, and to transmit data and instructions to, a storage system, at least one input device, and at least one output device.

[0173] Computer programs for implementing the methods of the present application can be written in any combination of one or more programming languages. These computer programs can be provided to a processor of a general purpose computer, special purpose computer, or other programmable data processing apparatus, such that the computer program, when executed, can cause instructions defined in the flow charts and / or block diagrams to be implemented. The computer program can be executed entirely on a machine, partially on a machine, partially on a machine as a stand-alone software package, and partially on a remote machine or entirely on a remote machine or server.

[0174] In the context of the present application, a computer-readable storage medium can be a tangible medium that can contain or store a computer program for use by or in connection with an instruction execution system, apparatus, or device. A computer-readable storage medium can include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the foregoing. Alternatively, a computer-readable storage medium can be a machine-readable signal medium. More specific examples of a machine-readable storage medium will include one or more lines of electrical connections, a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0175] To provide for interaction with a user, the systems and techniques described here can be implemented on an electronic device having a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user and a keyboard and a pointing device (e.g., a mouse or a trackball) by which the user can provide input to the electronic device. Other kinds of devices can be used to provide for interaction with a user as well; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form, including acoustic, speech, or tactile input.

[0176] The systems and techniques described herein can be implemented in a computing system that includes a back end component, e.g., as a data server, or that includes a middleware component, e.g., an application server, or that includes a front end component, e.g., a user computer having a graphical user interface or a Web browser through which a user can interact with an implementation of the systems and techniques described herein, or any combination of such back end, middleware, or front end components. The components of the system can be interconnected by any form or medium of digital data communication, e.g., a communication network. Examples of communication networks include a local area network (LAN), a wide area network (WAN), a blockchain network, and the Internet.

[0177] The computing system can include clients and servers. A client and server are generally remote from each other and typically interact through a communication network. The relationship of client and server arises by virtue of computer programs running on the respective computers and having a client-server relationship to each other. A server can be a cloud server, also known as a cloud computing server or cloud host, which is a host product in the cloud computing service system, to solve the defects of large management difficulty and weak business scalability in traditional physical host and VPS service.

[0178] It should be understood that the various forms of flow shown above can be re-ordered, added to, or deleted from without departing from the scope of the present disclosure. For example, the steps recited in the present disclosure can be executed in parallel, executed in series, or executed in different orders, as long as the desired results of the technical solutions of the present disclosure can be achieved, and the present disclosure is not limited herein.

[0179] The above detailed description does not constitute a limitation on the protection scope of the present application. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent replacements, and improvements made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for screening causal features, characterized in that, The method comprises the following steps: obtaining to-be-screened data containing the causal feature, wherein the to-be-screened data comprises transverse federated scene data and longitudinal federated scene data; statistically obtaining first sample frequencies of candidate features, label variables and conditional features corresponding to the transverse federated scene data, generating a local sample frequency contingency table, adding noise to the first sample frequencies to obtain first privacy sample frequencies, performing contingency table fusion calculation according to the first privacy sample frequencies, and generating a transverse federated sample frequency contingency table as a contingency table statistical result; performing feature binning processing on the candidate features, label variables and conditional features corresponding to the longitudinal federated scene data respectively to obtain respective corresponding sample identifier sets; determining sample intersections under different feature value combinations as second sample frequencies according to the feature binning processing result; adding noise to the second sample frequencies to obtain second privacy sample frequencies as the contingency table statistical result; wherein the noise added to the first sample frequencies and the second sample frequencies satisfies a conditionally independent noise constraint condition; and performing conditionally independent judgment on the features and the label variables in the to-be-screened data according to the contingency table statistical result. screening the features in the to-be-screened data in sequence according to the conditionally independent judgment result, and determining a final causal feature set.

2. The method of claim 1, wherein, The method for obtaining to-be-screened data containing the causal feature comprises the following steps: obtaining original data containing the causal feature; if in a transverse federated scene, performing feature alignment, federated feature engineering and feature value desensitization processing on the original data; if in a longitudinal federated scene, performing sample identifier alignment, local feature engineering and feature value desensitization processing on the original data; taking the processed original data as the to-be-screened data.

3. The method of claim 1, wherein, The method for determining sample intersections under different feature value combinations according to the feature binning processing result comprises the following steps: performing hash processing according to the binning processing result to obtain hash sample identifier sets corresponding to the candidate features, label variables and conditional features respectively; encrypting the feature binning result according to a first random factor corresponding to the candidate feature; comparing the encrypted feature binning result by using ciphertext, and determining the sample intersections according to the intersections of sample identifiers under different feature value combinations.

4. The method of claim 1, wherein, The method for performing conditionally independent judgment on the features and the label variables in the to-be-screened data according to the contingency table statistical result comprises the following steps: determining a chi-square test value and a degree of freedom according to the contingency table statistical result; querying a chi-square distribution table according to the chi-square test value and the degree of freedom to determine whether the features and the label variables satisfy conditionally independent.

5. The method of claim 1, wherein, The method for screening the features in the to-be-screened data in sequence according to the conditionally independent judgment result, and determining a final causal feature set comprises the following steps: screening the features in the to-be-screened data in sequence, and adding the corresponding features to a candidate Markov blanket if the features and the label variables satisfy conditionally independent; performing secondary screening on the features in the candidate Markov blanket, and determining a candidate Markov blanket from which false positive features are removed as the final causal feature set.

6. A device for screening for causal features, characterized in that The method comprises the following steps: The to-be-screened data acquisition module is configured to acquire to-be-screened data containing the causal features, wherein the to-be-screened data includes transverse federated scenario data and longitudinal federated scenario data. The conditional independence judgment module is configured to: count first sample frequencies of candidate features, label variables, and conditional features corresponding to the transverse federated scenario data, and generate a local sample frequency contingency table; add noise to the first sample frequencies to obtain first private sample frequencies; perform contingency table fusion calculation based on the first private sample frequencies, and generate a transverse federated sample frequency contingency table as a contingency table statistical result; perform feature binning processing on the candidate features, the label variables, and the conditional features corresponding to the longitudinal federated scenario data respectively, and obtain respective corresponding sample identification sets; determine sample intersections under different feature value combinations as second sample frequencies based on the feature binning processing results; add noise to the second sample frequencies to obtain second private sample frequencies as the contingency table statistical result; wherein the noise added to the first sample frequencies and the second sample frequencies satisfies a conditional independence noise constraint condition; and perform conditional independence judgment on features and label variables in the to-be-screened data based on the contingency table statistical result. The causal feature set determination module is configured to: perform screening on the features in the to-be-screened data in sequence based on the conditional independence judgment result, and determine a final causal feature set.

7. An electronic device, comprising: The electronic device includes: at least one processor; and a memory connected to the at least one processor in communication; wherein the memory stores a computer program that can be executed by the at least one processor, and the computer program is executed by the at least one processor to enable the at least one processor to execute the causal feature screening method of any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, The computer readable storage medium stores computer instructions for causing the processor to execute when the processor executes to implement the causal feature screening method of any one of claims 1-5. The computer readable storage medium stores computer instructions for causing the processor to execute when the processor executes to implement the causal feature screening method of any one of claims 1-5.

Images