WO2021139462A1 - Procédé et dispositif de sélection progressive de modèles et support lisible de stockage - Google Patents

Procédé et dispositif de sélection progressive de modèles et support lisible de stockage Download PDF

Info

Publication number
WO2021139462A1
WO2021139462A1 PCT/CN2020/134035 CN2020134035W WO2021139462A1 WO 2021139462 A1 WO2021139462 A1 WO 2021139462A1 CN 2020134035 W CN2020134035 W CN 2020134035W WO 2021139462 A1 WO2021139462 A1 WO 2021139462A1
Authority
WO
WIPO (PCT)
Prior art keywords
model
feature
saliency
feature set
training
Prior art date
Application number
PCT/CN2020/134035
Other languages
English (en)
Chinese (zh)
Inventor
唐兴兴
黄启军
陈瑞钦
林冰垠
李诗琦
Original Assignee
深圳前海微众银行股份有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 深圳前海微众银行股份有限公司 filed Critical 深圳前海微众银行股份有限公司
Publication of WO2021139462A1 publication Critical patent/WO2021139462A1/fr

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning

Definitions

  • This application relates to the field of machine learning technology of Fintech, in particular to a stepwise model selection method, device and readable storage medium.
  • stepwise selection mode is an important model selection strategy. Compared with adding all features into model training, it can effectively prevent the model from overfitting.
  • the current stepwise selection mode usually requires a higher level of modeling personnel.
  • the code development ability can only be implemented on a single machine, that is, the current implementation of the stepwise selection mode has high threshold requirements for modelers, and the modeling time for the stepwise selection mode is long because it can only be achieved on a single machine , The modeling efficiency is low. Therefore, in the prior art, there are technical problems of high modeling threshold and low efficiency of gradually selecting modes.
  • the main purpose of this application is to provide a step-by-step model selection method, device and readable storage medium, aiming to solve the technical problems of high modeling threshold and low efficiency of step-by-step model selection in the prior art.
  • this application provides a stepwise model selection method, which is applied to the server side, and the stepwise model selection method includes:
  • the feature set to be trained includes a first model feature set and a second model feature set
  • the cyclic training model set includes a first cyclic training model set and a second cyclic training model set
  • the step of performing cyclic training on the first initial training model based on each of the first type saliency and each of the second type saliency respectively, and obtaining a set of cyclic training models includes:
  • the first initial training model is cyclically trained and updated until the feature to be removed does not exist in the first model feature set, and the first cyclic training model set is obtained ;
  • the target feature is added to the first model feature set, and the updated first initial training model is cyclically trained based on the first model feature set after the target feature is added, until the target is added If the feature to be eliminated does not exist in the feature set of the first model after the feature and the target feature does not exist in the feature set of the second model, the second cyclic training model set is obtained.
  • the present application also provides a step-by-step model selection device.
  • the step-by-step model selection device is applied to the server, and the step-by-step model selection device includes:
  • the first training module is configured to receive the configuration parameters sent by the client associated with the server and obtain the feature set to be trained, and perform the preset training model based on the feature set to be trained and the configuration parameters Training to obtain the first initial training model;
  • a calculation module for calculating respectively the first type saliency and the second type saliency corresponding to the feature set to be trained
  • a second training module configured to perform cyclic training on the first initial training model based on each of the first type saliency and each of the second type saliency to obtain a cyclic training model set;
  • a selection module for selecting a target training model from the first initial training model and the set of cyclic training models based on the configuration parameters
  • the feedback module is used for generating the visualization data corresponding to the target training model, and feeding back the visualization data to the client.
  • the present application also provides a step-by-step model selection device.
  • the step-by-step model selection device includes a memory, a processor, and a program of the step-by-step model selection method that is stored on the memory and can be run on the processor, and When the program of the stepwise model selection method is executed by the processor, the steps of the stepwise model selection method as described above can be realized.
  • the present application also provides a readable storage medium, the readable storage medium stores a program for implementing the stepwise model selection method, and when the program for the stepwise model selection method is executed by a processor, the stepwise model selection method is implemented as described above. step.
  • This application receives the configuration parameters sent by the client associated with the server and obtains the feature set to be trained, and trains the preset model to be trained based on the feature set to be trained and the configuration parameters to obtain the first initial Training model, respectively calculating the saliency of the first type and the saliency of the second type corresponding to the feature set to be trained, respectively, based on each of the first type of saliency and each of the second type of saliency, for the first type of saliency
  • the initial training model performs cyclic training to obtain a cyclic training model set, and based on the configuration parameters, a target training model is selected from the first initial training model and the cyclic training model set, and a visualization corresponding to the target training model is generated Data, and feed back the visualization data to the client.
  • this application first receives the configuration parameters sent by the client and obtains the feature set to be trained, and trains the preset model to be trained based on the feature set to be trained and the configuration parameters, and obtains the first An initial training model, and then respectively perform calculations of the saliency of the first type and the saliency of the second type corresponding to the feature set to be trained, and then respectively based on each of the first type of saliency and each of the second type of saliency , Performing cyclic training on the first initial training model to obtain a cyclic training model set, and then based on the configuration parameters, selecting a target training model among the first initial training model and the cyclic training model set , And then generate the visualization data corresponding to the target training model, and then feed back the visualization data to the client.
  • this application provides a model selection method of step-by-step selection mode of non-coded distributed modeling and visual modeling.
  • the user only needs to set and send the necessary configuration parameters to the step-by-step model selection server through the client, and the step-by-step model Select the server to feed back the corresponding visualization data and step-by-step model selection results of the corresponding step-by-step model selection process, that is, through the communication connection between the client and the step-by-step model selection server for model modeling, distributed modeling is realized, and the corresponding
  • the modeling efficiency of the step-by-step selection mode is improved, and the step-by-step model selection results corresponding to the obtained modeling parameters are converted into visual data and fed back to the client.
  • FIG. 1 is a schematic flowchart of a first embodiment of a stepwise model selection method according to this application;
  • FIG. 2 is a schematic diagram of a visual interface for configuring the parameters in the step-by-step model selection method of this application;
  • FIG. 3 is a schematic flowchart of a second embodiment of a stepwise model selection method according to this application.
  • FIG. 4 is a schematic diagram of the first cycle process described in the step-by-step model selection method of this application.
  • FIG. 5 is a schematic diagram of a model selection process corresponding to the second cycle process and the first cycle process in the stepwise model selection method of this application;
  • FIG. 6 is a schematic flowchart of a third embodiment of a stepwise model selection method according to this application.
  • FIG. 7 is a schematic diagram of the device structure of the hardware operating environment involved in the solution of the embodiment of the application.
  • the embodiment of the present application provides a stepwise model selection method.
  • the stepwise model selection method is applied to the server.
  • the stepwise model selection method includes:
  • Step S10 Receive configuration parameters sent by the client associated with the server and obtain a feature set to be trained, and train a preset model to be trained based on the feature set to be trained and the configuration parameters to obtain a first initial Training model
  • the client includes a visual interface
  • the user can configure the parameters of the preset model to be trained on the visual interface to obtain the configuration parameters, as shown in Figure 2
  • the parameters such as the maximum iteration coefficient, the minimum convergence error, the stepwise model selection mode, and the category weight are all parameters that need to be set before model training.
  • Step-by-step model selection modes include forward selection mode, backward selection mode, and step-by-step selection mode.
  • the stepwise model selection method is applied to a stepwise model selection server.
  • the feature to be trained includes one or more features, and each feature includes one or more pieces of feature data.
  • the preset model to be trained includes a logistic regression model.
  • the training feature set includes a first model feature set and a second model feature set, wherein the first model feature set is a feature set that has been added to the preset model to be trained for training, and the second model feature set is not The feature set of the preset model to be trained for training is added, and both the first model feature set and the second model feature set include one or more features.
  • the configuration parameters sent by the client associated with the server are received and the feature set to be trained is obtained, and a preset model to be trained is trained based on the feature set to be trained and the configuration parameters to obtain a first initial training model.
  • the configuration parameters sent by the client associated with the server are received, and the feature set to be trained is extracted from the preset server local database, and the feature set is based on the features of the first model feature set in the feature set to be trained Data, iterative training is performed on the preset model to be trained, and when the iterative training completion judgment condition in the configuration parameters is reached, the training is stopped and the first initial training model is obtained, wherein the iterative training completion judgment
  • the conditions include reaching the maximum number of iterations, reaching the minimum convergence error, etc.
  • Step S20 Calculate the saliency of the first type and the saliency of the second type corresponding to the feature set to be trained respectively;
  • both the first type of significance and the second type of significance can be determined based on the Pearson correlation value, that is, when the Pearson correlation value is less than or Equal to the preset Pearson correlation threshold, it is determined that the feature corresponding to the first type of saliency or the second type of saliency meets the preset saliency requirement, that is, the first type of saliency or the The feature corresponding to the second type of saliency is shown to be significant, and when the Pearson correlation value is greater than the preset Pearson correlation threshold, it is determined that the first type of significance or the second type of significance corresponds to The feature does not meet the preset saliency requirement, that is, the feature corresponding to the saliency of the first type or the saliency of the second type is insignificant, and the feature set to be trained includes the first model feature set and the second model feature set. Model feature set.
  • the feature set to be trained includes a first model feature set and a second model feature set
  • the step of separately calculating the saliency of the first type and the saliency of the second type corresponding to the feature set to be trained includes:
  • Step S21 Calculate the chi-square value wald corresponding to each element in the first model feature set
  • the chi-square value wald corresponding to each element in the first model feature set is calculated, specifically, the chi-square value corresponding to each element in the first model feature set is calculated based on a preset chi-square value wald calculation formula wald, where the preset chi-square value wald calculation formula is as follows:
  • S 1 is the chi-square value wald
  • X is the characteristic data representation matrix corresponding to the feature set to be trained, where X includes n pieces of data, each piece of data includes k values, and X can be represented by a matrix
  • the model parameter obtained by training the preset model to be trained based on X is ⁇ , where ⁇ is a k-dimensional vector ( ⁇ 1 , ⁇ 2 ,..., ⁇ k-1 , ⁇ k ), and
  • the feature set to be trained can be divided into a first model feature set and a second model feature set, the feature data representation matrix corresponding to the first model feature set is X 0 , and the feature data corresponding to the second model feature set represents The matrix is X 1 , where X 0 includes n pieces of data, each piece of data includes (kt) values, and the model parameter obtained by training the preset model to be trained corresponding to X 0 is ⁇ 0 , where ⁇ 0 is (kt)-dimensional vector ( ⁇
  • Step S21 Calculate the first type saliency of each element in the first model feature set based on each chi-square value wald and the degrees of freedom of each element in the first model feature set;
  • the degree of freedom is related to the number of feature data corresponding to the feature. For example, assuming that there are 100 pieces of different data in the feature data, the degree of freedom is 99.
  • the Pearson correlation value of each element in the first model feature set is calculated by a preset Pearson correlation value calculation formula, and further, through each of the Pearson correlation values
  • the significance of the first type of each element in the feature set of the first model for example, assuming that the Pearson correlation values are 0.001, 0.01, and 0.05, respectively, then the corresponding measures to determine the significance of each of the first types Values are 10, 1, and 0.2, wherein the larger the measurement value, the more significant the first significance.
  • Step S23 Calculate the score chi-square value corresponding to each element in the second model feature set.
  • the score chi-square value corresponding to each element in the second model feature set is calculated, specifically, the score chi-square value corresponding to each element in the second model feature set is calculated based on a preset score chi-square value calculation formula
  • the calculation formula for the preset score chi-square value is as follows:
  • S2 is the score chi-square value
  • X is the corresponding feature data representation matrix in the feature set to be trained, where X includes n pieces of data, each piece of data includes k values, and X can be represented by a matrix
  • the model parameter obtained by training the preset model to be trained based on X is ⁇ , where ⁇ is a k-dimensional vector ( ⁇ 1 , ⁇ 2 ,..., ⁇ k-1 , ⁇ k ), and
  • the feature set to be trained can be divided into a first model feature set X and a second model feature set, wherein the feature data representation matrix corresponding to the first model feature set is X 0 , and the feature corresponding to the second model feature set
  • the data representation matrix is X 1 , where X 0 includes n pieces of data, each piece of data includes (kt) numbers, and the model parameter obtained by training the preset model to be trained corresponding to X 0 is ⁇ 0 , where ⁇ 0 is a (k
  • Step S24 Calculate the second-type saliency of each element in the second model feature set based on each of the score chi-square values and the degrees of freedom of each element in the second model feature set.
  • the second-type saliency of each element in the second model feature set is calculated based on each of the score chi-square value and the degree of freedom of each element in the second model feature set, specifically, based on each
  • the score chi-square value and the degree of freedom of each element in the second model feature set are calculated by using a preset Pearson correlation value calculation formula to calculate the Pearson correlation value of each element in the second model feature set, and then through Each of the Pearson correlation values determines the second-type significance of each element in the second model feature set.
  • Step S30 Perform cyclic training on the first initial training model based on each of the first type saliency and each of the second type saliency to obtain a set of cyclic training models;
  • the cyclic training model set includes a first cyclic training model set and a second cyclic training model set, and the first cyclic training model set includes one or more first model elements, The second cyclic training model set includes one or more second model elements.
  • the first initial training model is cyclically trained to obtain a cyclic training model set, specifically, based on each of the first type saliency , Gradually remove features to be removed in the first model feature set that meet the preset significant requirements for removal, and after removing one feature to be removed each time, based on the removed first model feature set,
  • An initial training model is iteratively trained and updated, and one of the first model elements is obtained until the preset first cycle termination condition is reached, and the first cycle training model set is obtained, and then based on the saliency of each second type,
  • target features that meet the preset significance requirements are selected, and the target feature is added to the first model feature set, and after each target feature is added to the first model feature set ,
  • the first initial training model after the iterative update is cyclically trained to obtain one or more second model elements until the preset first cycle is reached at the
  • Step S40 based on the configuration parameters, select a target training model from the first initial training model and the cyclic training model set;
  • the configuration parameters include a model selection strategy.
  • a target training model is selected from the first initial training model and the cyclic training model set. Specifically, based on the model selection strategy, a model that best meets the model selection strategy is selected as the target training model among the elements of the first initial training model and the cyclic training model set.
  • the step of selecting a target training model from the first initial training model and the cyclic training model set based on the configuration parameters includes:
  • Step S41 Obtain a model selection strategy in the parameter configuration, where the model selection strategy includes an AUC value and an AIC value;
  • the AUC value is the criterion for evaluating the training model, and the larger the AUC value is, the better the training model is, where the AUC value is ROC (receiver operating characteristic curve, the area enclosed by the coordinate axis under the receiver operating characteristic curve, and the value of this area will not be greater than 1.
  • the ROC curve is based on a series of different two classification methods (demarcation value or decision Threshold), a curve drawn with true positive rate (sensitivity) as the ordinate and false positive rate (1-specificity) as the abscissa, the AIC value is a value calculated based on the AIC criterion, where the AIC criterion is A standard to measure the goodness of a statistical model.
  • Step S42 if the model selection strategy is the AUC value, compare the AUC values of the elements in the cyclic training model set, and select the element corresponding to the largest AUC value as the target training model ;
  • the model selection strategy is the AUC value
  • the AUC values of the elements in the cyclic training model set are compared, and the element corresponding to the largest AUC value is selected as the The target training model, specifically, if the model selection strategy is the AUC value, compare the AUC values to obtain the maximum AUC value, and use the training model corresponding to the maximum AUC value as the target training A model, wherein the training model includes a first initial training model and each element in the cyclic training model set.
  • Step S43 If the model selection strategy is the AIC value, compare the AIC values of the elements in the cyclic training model set, and select the element corresponding to the smallest AIC value as the target training model .
  • the model selection strategy is the AIC value
  • the AIC value of each element in the cyclic training model set is compared, and the element corresponding to the smallest AIC value is selected as the The target training model, specifically, if the model selection strategy is the AIC value, the AIC values are compared to obtain the minimum AIC value, and the training model corresponding to the minimum AIC value is used as the target training A model, wherein the training model includes a first initial training model and each element in the cyclic training model set.
  • Step S50 Generate visualization data corresponding to the target training model, and feed back the visualization data to the client.
  • the visualization data includes candidate feature visualization data, model selection summary visualization data, and training process visualization data, where the candidate feature is a feature in the feature set to be trained
  • the model selection summary data includes summary data for model selection of the first initial training model and the model elements in the cyclic training model set.
  • Generate visualization data corresponding to the target training model and feed back the visualization data to the client, specifically, generate visualization data corresponding to the acquisition process corresponding to the target training model, wherein the acquisition process includes features Selection process, model training process, model selection process, etc., and then feedback the visualization data to the visualization interface of the client for display to the customer, wherein the feature selection process is the process of selecting features in the feature set to be trained
  • the model training process is a process of training a target model, wherein the target model includes a preset model to be trained, a first initial training model, model elements, etc., and the model selection process is based on a preset model selection strategy The process of selecting the target training model.
  • the client includes a visual interface
  • the step of generating visualization data corresponding to the target training model and feeding back the visualization data to the client includes:
  • Step S51 Obtain candidate feature data, selection summary data, and training process data corresponding to the model selection process of the target training model;
  • the model selection process of the target training model includes a model iterative training process, a feature selection process, a model selection process, etc., wherein the feature selection process is a process of removing the feature to be removed, and the model selection process The process of selecting a target training model based on a preset model selection strategy.
  • Obtain candidate feature data, selection summary data, and training process data corresponding to the model selection process of the target training model specifically, acquire candidate feature data of the feature selection process and selection summary data of the model selection process in real time And training process data of the model iterative training process.
  • Step S52 Generate visualization data corresponding to the candidate feature data, the selection summary data, and the training process data, and feed back the visualization data to the visualization interface in real time.
  • the visualization data includes graphic data, table data, and the like.
  • the selection of the visualization data corresponding to the summary data and the training process data, and the real-time feedback of the visualization data to the visualization interface in real time, wherein the time interval for real-time feedback of the visualization data to the visualization interface The user of the server can be selected by the stepwise model to set it, and the user of the client can query the visualization data in real time on the client.
  • the first model is obtained.
  • the initial training model is to calculate the saliency of the first type and the saliency of the second type corresponding to the feature set to be trained, respectively, based on the saliency of each of the first types and the salience of each of the second types, and the saliency of the first type
  • An initial training model performs cyclic training to obtain a cyclic training model set, and based on the configuration parameters, a target training model is selected from the first initial training model and the cyclic training model set, and the corresponding training model is generated Visualize data, and feed back the visualized data to the client.
  • this embodiment first receives the configuration parameters sent by the client and obtains the feature set to be trained, and trains the preset model to be trained based on the feature set to be trained and the configuration parameters to obtain The first initial training model, and then respectively calculate the saliency of the first type and the saliency of the second type corresponding to the feature set to be trained, and then respectively based on the saliency of each first type and the saliency of each second type Cyclical training of the first initial training model to obtain a cyclic training model set, and then based on the configuration parameters, perform target training model training among the first initial training model and the cyclic training model set Select, and then generate the visualization data corresponding to the target training model, and then feed back the visualization data to the client.
  • this embodiment provides a model selection method of step-by-step selection mode for codeless distributed modeling and visual modeling.
  • the user only needs to set and send the necessary configuration parameters to the step-by-step model selection server through the client.
  • the model selection server can feed back the corresponding visualization data and step-by-step model selection results of the corresponding step-by-step model selection process, that is, through the communication connection between the client and the step-by-step model selection server for model modeling, distributed modeling is realized, and then Compared with the step-by-step selection mode performed by a single machine, the modeling efficiency of the step-by-step selection mode is improved, and the step-by-step model selection results corresponding to the obtained modeling parameters are converted into visual data and fed back to the client.
  • the feature set to be trained includes a first model feature set and a second model feature set
  • the training model set includes the first cyclic training model set and the second cyclic training model set
  • the step of performing cyclic training on the first initial training model based on each of the first type saliency and each of the second type saliency respectively, and obtaining a set of cyclic training models includes:
  • Step S31 based on each of the first type saliency, remove features to be removed from the first model feature set that meet the preset removal saliency requirements;
  • the features to be removed that meet the preset saliency requirements for removal are excluded from the feature set of the first model, specifically, based on the saliency of each first type, Select the feature to be selected with the lowest saliency in the first model feature set, and determine whether the feature to be selected meets the pre-set saliency removal requirement, and if the feature to be selected meets the pre-set saliency removal requirement, all The feature to be selected is used as the feature to be removed. If the feature to be selected does not meet the preset removal significance requirement, then the first loop process corresponding to the first model feature set is jumped out, wherein, the first A schematic diagram of the cycle process is shown in Figure 4.
  • the data is the training data corresponding to each feature in the feature set to be trained, the training model is the preset model to be trained, and the feature added to the model is the first
  • the threshold value is the preset significance threshold for removing.
  • the step of culling features to be removed that meet the preset saliency removal requirement from the first model feature set includes:
  • Step S311 comparing the saliency of each of the first types to select the feature with the lowest saliency in the first model feature set as the feature to be selected;
  • the saliency of each first type is compared, and the feature with the lowest saliency is selected from the feature set of the first model as the feature to be selected. Specifically, the saliency of each first type is selected as the feature to be selected. The Pearson correlation value corresponding to the sex is compared, and the feature with the largest Pearson correlation value is selected as the feature to be selected.
  • Step S312 comparing the to-be-selected saliency of the to-be-selected feature with a preset saliency-removed threshold
  • the preset saliency rejection threshold can be set by the user, and the saliency to be selected is the first type saliency of the feature to be selected.
  • the to-be-selected saliency of the to-be-selected feature is compared with a preset rejection significance threshold, specifically, the Pearson correlation value corresponding to the to-be-selected saliency of the to-be-selected feature is compared with the preset rejection
  • the significance threshold corresponding to the elimination Pearson correlation threshold is compared.
  • Step S313 if the to-be-selected significance is less than the preset rejection significance threshold, determine that the to-be-selected feature meets the preset rejection significance requirement, and use the to-be-selected feature as the to-be-rejected feature feature.
  • the to-be-selected saliency is less than the preset rejection significance threshold, it is determined that the to-be-selected feature meets the preset rejection saliency requirement, and the to-be-selected feature is taken as the The feature to be eliminated, specifically, if the significance to be selected is less than the preset significance threshold for elimination, it indicates that the Pearson correlation value corresponding to the significance to be selected is greater than the elimination Pearson correlation valve Value, the feature to be selected is insignificant, and then it is determined that the feature to be selected meets the preset saliency removal requirement, and the feature to be selected is taken as the feature to be removed.
  • the saliency to be selected is If it is greater than or equal to the preset significance threshold for rejection, it indicates that the Pearson correlation value corresponding to the significance to be selected is less than or equal to the Pearson correlation threshold for rejection, and then it is determined that the feature to be selected satisfies all requirements. The pre-determined saliency requirement is eliminated, and the first loop process is jumped out.
  • Step S32 Based on the eliminated first model feature set, the first initial training model is cyclically trained and updated until the feature to be eliminated does not exist in the first model feature set, and the first cycle is obtained. Training model set;
  • the first initial training model is cyclically trained and updated until the feature to be removed does not exist in the first model feature set, and the The first cyclic training model set, specifically, inputting training data corresponding to each feature of the first model feature set after removal into the first initial training model, so as to perform iterative training updates on the first initial training model, Obtain the updated first initial training model, and use the updated first initial training model as one of the first model elements. Further, repeat the first cycle process again, that is, re-execute the first cycle process again.
  • For the first model element repeat the above process to obtain a plurality of the first model elements, until the preset first cycle process termination condition is reached, and the first cycle training model set is obtained.
  • the configuration parameters include the iterative training completion judgment condition
  • the first cyclic training model set includes one or more first model elements
  • the cyclic training update is performed on the first initial training model based on the removed first model feature set, until the feature to be removed does not exist in the first model feature set, and the first cyclic training is obtained
  • the steps of the model set include:
  • Step S321 Perform iterative training update on the first initial training model based on the eliminated first model feature set until the first initial training model satisfies the iterative training completion judgment condition, and obtain each of the first initial training models.
  • the iterative training completion judgment condition includes reaching the maximum number of iterations and reaching the minimum convergence error.
  • Step S322 Recalculate the first type saliency of each element in the feature set of the first model after the removal, so as to repeat the removal of the feature to be removed and the evaluation of the updated first initial training model. Iterative training and updating until the feature to be eliminated does not exist in the first model feature set, and the first cyclic training model set is obtained.
  • a comparison is performed based on the removed first model feature set.
  • Step S33 based on each of the second-type saliences, select target features that meet the preset saliency requirements from the second model feature set;
  • the second type of significance can be determined based on the Pearson correlation value, and when the Pearson correlation value is less than or equal to the preset Pearson correlation threshold, it is determined The feature corresponding to the saliency of the second type meets the preset saliency requirement, that is, the feature corresponding to the saliency of the second type appears to be significant, when the Pearson correlation value is greater than the preset Pearson correlation valve Value, it is determined that the feature corresponding to the second type of saliency does not meet the preset saliency requirement, that is, the feature corresponding to the second saliency is not significant.
  • target features that meet the preset saliency requirements in the second model feature set specifically, based on the saliency of each second type, in the second model feature set
  • Select the most salient feature with the highest salientity and determine whether the most salient feature meets the preset salientity requirements, and if the most salient feature meets the preset salientity requirements, then the most salient feature is taken as the target feature, If the feature to be selected does not meet the preset saliency requirement, the second loop process corresponding to the second model feature set is jumped out. As shown in FIG.
  • the second loop process plus the first A schematic diagram of the model selection process corresponding to the cyclic process, where the data is the training data corresponding to each feature in the feature set to be trained, the training model is the preset model to be trained, and the feature added to the model is the first
  • the threshold value is the preset significance threshold value
  • the significance is the preset significance requirement.
  • the step of selecting a target feature that meets a preset saliency requirement from the second model feature set based on each of the second type saliency includes:
  • Step S331 comparing each of the second-type saliences to select the most significant feature with the highest saliency in the second model feature set;
  • the saliency of each second type is compared to select the most significant feature with the highest saliency in the second model feature set, specifically, the saliency of each second type is corresponding to Compare the Pearson correlation values of, and select the feature with the smallest Pearson correlation value as the most significant feature.
  • Step S332 comparing the target saliency corresponding to the most salient feature with the preset saliency threshold
  • the preset saliency threshold can be set by the user, and the target saliency is the second type saliency of the most salient feature.
  • the target saliency corresponding to the most salient feature is compared with the preset saliency threshold, specifically, the Pearson correlation value corresponding to the target saliency of the most salient feature is compared with the preset saliency
  • the Pearson correlation threshold corresponding to the sexual threshold is compared.
  • Step S333 If the target saliency is greater than or equal to the preset saliency threshold, it is determined that the most significant feature meets the preset saliency requirement, and the most significant feature is taken as the target feature.
  • the target saliency is greater than or equal to the preset saliency threshold, it is determined that the most salient feature meets the preset saliency requirement, and the most salient feature is taken as the The target feature, specifically, if the target significance is greater than or equal to the preset significance threshold, it indicates that the Pearson correlation value corresponding to the target significance is less than or equal to the Pearson correlation threshold, If the feature to be selected is significant, it is determined that the most significant feature to be selected meets the preset saliency requirement, and the most significant feature is taken as the target feature.
  • the target saliency is less than the predicted Set the significance threshold, it indicates that the Pearson correlation value corresponding to the target significance is greater than the Pearson correlation threshold, and then it is determined that the most significant feature does not meet the preset significance requirements, and jumps out The second cycle process.
  • Step S34 Add the target feature to the first model feature set, and perform cyclic training on the updated first initial training model based on the first model feature set after adding the target feature until it is added If the feature to be eliminated does not exist in the first model feature set after the target feature and the target feature does not exist in the second model feature set, the second cyclic training model set is obtained.
  • the target feature is added to the first model feature set, and the first model feature set after adding the target feature is compared with the updated first initial model feature set.
  • the training model performs cyclic training until the feature to be removed does not exist in the first model feature set after adding the target feature and the target feature does not exist in the second model feature set, and the second cyclic training is obtained.
  • the model set specifically, adding the target feature to the first model feature set, and adding the first model feature set after adding the target feature to the first initial training model after the previous iteration training update, Iterative training update is performed on the first initial training model after the last iterative training update to obtain the first initial training model after this update, and the first initial training model after this update is taken as One of the second model elements, further, repeating the first loop process again until the first model feature set after adding the target feature reaches the preset first loop process termination condition, also That is, until the feature to be removed does not exist in the first model feature set after the target feature is added, one or more of the second model
  • the second cyclic training model set includes one or more second model elements
  • the step of obtaining the second cyclic training model set includes:
  • Step S341 Add the target feature to the first model feature set to update the first model feature set and the second model feature set, and obtain the updated first model feature set and the updated model feature set.
  • the second model feature set
  • the target feature is added to the first model feature set to update the first model feature set and the second model feature set to obtain the updated first model feature set and The updated second model feature set.
  • the target feature is added to the first model feature set to update the number and information of features included in the first model feature set, and to update the second model The number of features and information included in the feature set are obtained to obtain the updated first model feature set and the updated second model feature set.
  • the first model feature set includes features X 1 and X 2 , Wherein, the target feature is X 1 , and the second model feature set includes feature X 3 and feature X 4 , then the updated first model feature set includes feature X 2 , and the updated second model The feature set includes feature X 1 , feature X 3 and feature X 4 .
  • Step S342 Perform iterative training update on the first initial training model based on the updated feature set of the first model to obtain one of the second model elements;
  • the first initial training model is iteratively trained and updated to obtain one of the second model elements, specifically, the updated The first model feature set is added to the first initial training model after the last update to perform an iterative training update on the first initial training model until the first initial training model reaches the iterative training completion judgment condition, Obtain the updated first initial training model, that is, obtain one of the second model elements.
  • Step S343 Recalculate the first-type saliency of each element in the updated first model feature set, so as to repeat the elimination of the features to be eliminated and the evaluation of the updated first initial training model. Iterative training and updating to obtain one or more of the second model elements, until the feature to be eliminated does not exist in the first model feature set, then jump out of the first loop process corresponding to the first model feature set;
  • the first loop process is re-executed until the first model feature set does not include the The features to be removed are obtained, and one or more of the second model elements are obtained.
  • Step S344 Recalculate the second-type saliency of each element in the updated second model feature set, so as to repeatedly select the target feature in the second model feature set, and add the target feature to all elements.
  • the first model feature set to repeatedly execute the first loop process to obtain one or more of the second model elements, until the target feature does not exist in the second model feature set, then jump out of the second model feature set.
  • the second cycle process corresponding to the model feature set.
  • the second-type saliency of each element in the updated second model feature set is recalculated to repeatedly select the target feature in the second model feature set, and the target Features are added to the first model feature set to repeat the first loop process to obtain one or more of the second model elements, until the target feature does not exist in the second model feature set, then jump out
  • the second cyclic process corresponding to the second model feature set specifically, recalculating the second type saliency of each element in the second model feature set after the update is based on the recalculated second type Significance, reselect the target feature and add it to the first model feature set to re-execute the first loop process until the feature to be removed does not exist in the first model feature set, and one or more second models are obtained Element, further, continue to reselect target features in the second model feature set to re-execute the first cycle process until the target feature does not exist in the second model feature set, and the second cycle is obtained Training model set.
  • the features to be removed that meet the preset saliency removal requirements are removed from the first model feature set, and then based on the removed first model feature set, all features are removed from the first model feature set.
  • the first initial training model performs cyclic training update until the feature to be removed does not exist in the first model feature set, and the first cyclic training model set is obtained, and then based on the saliency of each second type, In the second model feature set, select target features that meet the preset significance requirements, and then add the target feature to the first model feature set, and update based on the first model feature set after adding the target feature
  • the subsequent first initial training model performs cyclic training until the feature to be eliminated does not exist in the first model feature set after adding the target feature and the target feature does not exist in the second model feature set, Obtain the second cyclic training model set.
  • this embodiment gradually eliminates the features to be eliminated in the first model feature set based on each of the first type salience, and iterates the first initial training model based on the eliminated first model feature set Training update, obtain the first model element, and based on the saliency of each second type, gradually select target features from the second model feature set to add to the first model feature set, so as to train the first initial model Perform iterative training and update to obtain the second model element until the feature to be removed does not exist in the first model feature set and the target feature does not exist in the second model feature, obtain a cyclic training model set, and then
  • the model selection of the stepwise selection mode is realized, which lays the foundation for the model selection of the stepwise selection mode of the realization of codeless distributed modeling and visual modeling. Therefore, it is necessary to solve the high threshold and Inefficient technical problems laid the foundation.
  • the forward model selection method is applied to the client, and the forward model selection method includes:
  • Step A10 Receive a model selection task, and send configuration parameters corresponding to the model selection task to the server associated with the client, so that the server can perform model selection based on the configuration parameters to obtain a target training model , And obtain the visualization data corresponding to the target training model, so as to send the visualization data to the client;
  • the model selection task includes target model requirements
  • the target model requirements are determined by the configuration parameters
  • the configuration parameters include large iteration coefficients, minimum convergence errors, model selection modes, etc. parameter.
  • Receive the model selection task and send the configuration parameters corresponding to the model selection task to the server associated with the client, so that the server can make model selection based on the configuration parameters, obtain the target training model, and obtain
  • the visualization data corresponding to the target training model is sent to the client, specifically, the model selection task is received, and the configuration parameters corresponding to the model selection task are matched in a preset local database or determined by
  • the user sets the configuration parameters by himself based on the model selection task, and further, sends the configuration parameters to the server associated with the client, so that the server can use the configuration parameters and locally acquired waiting parameters.
  • the training model set performs training update on the preset model to be trained to obtain a first initial training model, and then performs a cyclic training update on the first initial training model to obtain a cyclic training model set, and in each of the cyclic training model sets Select a model that meets the preset model selection strategy as the target training model to convert the process data corresponding to the target training model into the visualization data and feed it back to the client, where the visualization data includes candidate feature visualization data , Model selection summary visualization data and model training process visualization data, wherein the candidate feature is each feature in each feature set to be trained, and the model selection summary data includes training the cycle based on a preset model selection strategy
  • the model elements in the model set are summarized data for model selection.
  • Step A20 Receive the visualization data fed back by the server, and display the visualization data on a preset visualization interface.
  • the client can query the visualization data corresponding to the process data of the server in real time on the preset visualization interface, and it can be in the process of model selection or model selection. After the selection is completed, the process data is inquired, and the client is in communication with the server.
  • a model selection task is received, and the configuration parameters corresponding to the model selection task are sent to the server associated with the client, so that the server can perform model selection based on the configuration parameters to obtain target training Model, and obtain the visualization data corresponding to the target training model to send the visualization data to the client, and then receive the visualization data fed back by the server, and set the visualization data in a preset visualization
  • the interface is displayed. That is, this implementation provides a model selection method for codeless distributed modeling and visual modeling.
  • this embodiment implements distributed modeling, improves the modeling efficiency during model selection, and the model selection process does not have any code development capability requirements for users, which reduces the ability threshold requirements for modelers.
  • the server can convert the process data corresponding to the target training model into visualized data and feed it back to the client, it further reduces the ability threshold requirements for modelers, and the visualized data is convenient for modelers to understand and read.
  • the modeling efficiency of modelers can be further improved, and therefore, the technical problems of high threshold and low efficiency of forward selection model modeling in the prior art are solved.
  • FIG. 7 is a schematic diagram of the device structure of the hardware operating environment involved in the solution of the embodiment of the present application.
  • the step-by-step model selection device may include: a processor 1001, such as a CPU, a memory 1005, and a communication bus 1002.
  • the communication bus 1002 is used to implement connection and communication between the processor 1001 and the memory 1005.
  • the memory 1005 may be a high-speed RAM memory, or a non-volatile memory (non-volatile memory), such as a magnetic disk memory.
  • the memory 1005 may also be a storage device independent of the aforementioned processor 1001.
  • the step-by-step model selection device may also include a rectangular user interface, a network interface, a camera, an RF (Radio Frequency) circuit, a sensor, an audio circuit, a WiFi module, and so on.
  • the rectangular user interface may include a display screen (Display) and an input sub-module such as a keyboard (Keyboard), and the optional rectangular user interface may also include a standard wired interface and a wireless interface.
  • the network interface can optionally include a standard wired interface and a wireless interface (such as a WI-FI interface).
  • step-by-step model selection device shown in FIG. 7 does not constitute a limitation on the step-by-step model selection device, and may include more or less components than shown in the figure, or combine certain components, or different The layout of the components.
  • the memory 1005 which is a computer-readable storage medium, may include an operating system, a network communication module, and a step-by-step model selection program.
  • the operating system is a program that manages and controls the hardware and software resources of the step-by-step model selection device, and supports the step-by-step model selection program and the operation of other software and/or programs.
  • the network communication module is used to realize the communication between the components in the memory 1005 and the communication with other hardware and software in the step-by-step model selection system.
  • the processor 1001 is configured to execute a step-by-step model selection program stored in the memory 1005 to implement the steps of the step-by-step model selection method described in any one of the above.
  • step-by-step model selection device of the present application is basically the same as each embodiment of the step-by-step model selection method described above, and will not be repeated here.
  • the embodiment of the present application also provides a step-by-step model selection device, the step-by-step model selection device is applied to the server, and the step-by-step model selection device includes:
  • the first training module is configured to receive the configuration parameters sent by the client associated with the server and obtain the feature set to be trained, and perform the preset training model based on the feature set to be trained and the configuration parameters Training to obtain the first initial training model;
  • a calculation module for calculating respectively the first type saliency and the second type saliency corresponding to the feature set to be trained
  • a second training module configured to perform cyclic training on the first initial training model based on each of the first type saliency and each of the second type saliency to obtain a cyclic training model set;
  • a selection module for selecting a target training model from the first initial training model and the set of cyclic training models based on the configuration parameters
  • the feedback module is used for generating the visualization data corresponding to the target training model, and feeding back the visualization data to the client.
  • the second training module includes:
  • the first cyclic training sub-module is configured to perform cyclic training and updating of the first initial training model based on the removed first model feature set until the first model feature set does not contain the to-be removed Feature, obtaining the first cyclic training model set;
  • the second cyclic training sub-module is used to add the target feature to the first model feature set, and based on the updated first initial model feature set after adding the target feature to the first model feature set.
  • the training model performs cyclic training until the feature to be eliminated does not exist in the first model feature set after adding the target feature and the target feature does not exist in the second model feature set, and the second cyclic training is obtained Model set.
  • the rejection sub-module includes:
  • the first selection unit is configured to compare the saliency of each of the first types to select the feature with the lowest saliency in the first model feature set as the feature to be selected;
  • the first comparison unit is configured to compare the to-be-selected saliency of the to-be-selected feature with a preset rejection saliency threshold
  • the first determining unit is configured to determine if the to-be-selected significance is less than the preset rejection significance threshold, determine that the to-be-selected feature satisfies the preset rejection significance requirement, and set the to-be-selected significance
  • the feature is used as the feature to be removed.
  • the first cycle training sub-module includes:
  • the first iterative training unit is used for the iterative training update of the first initial training model based on the eliminated first model feature set until the first initial training model satisfies the iterative training completion judgment Condition, to obtain one of the first model elements;
  • the second iterative training unit is used to recalculate the saliency of the first type of each element in the feature set of the first model after the removal, so as to repeatedly remove the feature to be removed and check all the updated features.
  • the iterative training update of the first initial training model until the feature to be eliminated does not exist in the first model feature set, and the first cyclic training model set is obtained.
  • the selection submodule includes:
  • a second selection unit for comparing the saliency of each of the second types to select the most significant feature with the highest saliency in the second model feature set;
  • a second comparison unit configured to compare the target significance corresponding to the most salient feature with the preset significance threshold
  • the second determining unit is configured to determine that if the target significance is greater than or equal to the preset significance threshold, determine that the most significant feature satisfies the preset significance requirement, and compare the most significant feature As the target feature.
  • the cycle training sub-module includes:
  • An update unit for adding the target feature to the first model feature set to update the first model feature set and the second model feature set to obtain the updated first model feature set And the updated feature set of the second model;
  • the third iterative training unit is used to recalculate the first type saliency of each element in the updated first model feature set, so as to repeatedly remove the feature to be removed and check all updated features.
  • the iterative training update of the first initial training model obtains one or more of the second model elements, until the feature to be removed does not exist in the feature set of the first model, then jump out of the feature set corresponding to the first model The first cycle process;
  • Cyclic unit for recalculating the second-type saliency of each element in the second model feature set after the update, so as to repeatedly select the target feature in the second model feature set, and combine the The target feature is added to the first model feature set to repeat the first loop process to obtain one or more of the second model elements, until the target feature does not exist in the second model feature set, then jump out The second cycle process corresponding to the second model feature set.
  • the calculation module includes:
  • the first calculation sub-module is used to calculate the chi-square value wald corresponding to each element in the first model feature set;
  • the second calculation sub-module is configured to calculate the first type saliency of each element in the first model feature set based on each of the chi-square value wald and the degrees of freedom of each element in the first model feature set;
  • the third calculation sub-module is used to calculate the score chi-square value corresponding to each element in the second model feature set
  • the fourth calculation sub-module is used to calculate the second-type saliency of each element in the second model feature set based on each of the score chi-square values and the degrees of freedom of each element in the second model feature set.
  • the selection module includes:
  • An obtaining sub-module for obtaining the model selection strategy in the parameter configuration wherein the model selection strategy includes an AUC value and an AIC value;
  • the first selection sub-module is configured to compare the AUC value of each element in the cyclic training model set if the model selection strategy is the AUC value to select the one corresponding to the largest AUC value Element as the target training model;
  • the second selection sub-module is configured to compare the AIC values of the elements in the cyclic training model set if the model selection strategy is the AIC value to select the smallest AIC value corresponding to the AIC value
  • the element serves as the target training model.
  • the feedback module includes:
  • the feedback sub-module is used to generate the visualization data corresponding to the candidate feature data, the selection summary data, and the training process data, and to feed back the visualization data to the visualization interface in real time.
  • step-by-step model selection device of the present application is basically the same as each embodiment of the above-mentioned step-by-step model selection method, and will not be repeated here.
  • an embodiment of the present application also provides a step-by-step model selection device.
  • the step-by-step model selection device is applied to a client, and the step-by-step model selection device includes:
  • the sending module is configured to receive the model selection task, and send the configuration parameters corresponding to the model selection task to the server associated with the client, so that the server can use the configuration parameters and the acquired configuration parameters.
  • model selection on training features obtaining a target training model, and obtaining visualization data corresponding to the target training model, so as to send the visualization data to the client;
  • the receiving module is configured to receive the visualization data fed back by the server, and display the visualization data on a preset visualization interface.
  • step-by-step model selection device of the present application is basically the same as each embodiment of the above-mentioned step-by-step model selection method, and will not be repeated here.
  • the embodiments of the present application provide a readable storage medium, and the readable storage medium stores one or more programs, and the one or more programs may also be executed by one or more processors for implementation Steps of the stepwise model selection method described in any one of the above.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Software Systems (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Medical Informatics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Physics & Mathematics (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Artificial Intelligence (AREA)
  • Image Analysis (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

L'invention concerne un procédé et un dispositif de sélection progressive de modèles et un support lisible de stockage. Le procédé consiste : à recevoir un paramètre de configuration envoyé par un client associé à un serveur, à acquérir des ensembles de caractéristiques à instruire et, en fonction des ensembles de caractéristiques à instruire et du paramètre de configuration, à instruire un modèle prédéfini à instruire, de façon à obtenir un premier modèle initial d'instruction (S10) ; à calculer respectivement une importance de premier type et une importance de second type correspondant aux ensembles de caractéristiques à instruire (S20) ; en fonction de chaque importance de premier type et de chaque importance de second type respectivement, à effectuer une instruction cyclique sur le premier modèle initial d'instruction pour obtenir un ensemble de modèles d'instruction cyclique (S30) ; en fonction du paramètre de configuration, à sélectionner un modèle cible d'instruction à partir du premier modèle initial d'instruction et de l'ensemble de modèles d'instruction cyclique (S40) ; et à générer des données de visualisation correspondant au modèle cible d'instruction et à renvoyer les données de visualisation au client (S50).
PCT/CN2020/134035 2020-01-09 2020-12-04 Procédé et dispositif de sélection progressive de modèles et support lisible de stockage WO2021139462A1 (fr)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202010024336.7 2020-01-09
CN202010024336.7A CN111241745B (zh) 2020-01-09 2020-01-09 逐步模型选择方法、设备及可读存储介质

Publications (1)

Publication Number Publication Date
WO2021139462A1 true WO2021139462A1 (fr) 2021-07-15

Family

ID=70864097

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/134035 WO2021139462A1 (fr) 2020-01-09 2020-12-04 Procédé et dispositif de sélection progressive de modèles et support lisible de stockage

Country Status (2)

Country Link
CN (1) CN111241745B (fr)
WO (1) WO2021139462A1 (fr)

Families Citing this family (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111241746B (zh) * 2020-01-09 2024-01-26 深圳前海微众银行股份有限公司 向前模型选择方法、设备和可读存储介质
CN111241745B (zh) * 2020-01-09 2024-05-24 深圳前海微众银行股份有限公司 逐步模型选择方法、设备及可读存储介质
CN113407680B (zh) * 2021-06-30 2023-06-02 竹间智能科技（上海）有限公司 异质集成模型筛选方法和电子设备
CN113746899B (zh) * 2021-07-29 2023-04-07 济南浪潮数据技术有限公司 一种云平台访问方法及装置
CN113780582B (zh) * 2021-09-15 2023-04-07 杭银消费金融股份有限公司 基于机器学习模型的风控特征筛选方法及系统
CN116704427B (zh) * 2023-04-19 2024-01-26 广东建设职业技术学院 一种基于3d cnn循环施工过程监测方法

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105550746A (zh) * 2015-12-08 2016-05-04 北京旷视科技有限公司 机器学习模型的训练方法和训练装置
WO2017219991A1 (fr) * 2016-06-23 2017-12-28 华为技术有限公司 Procédé et appareil d'optimisation adaptés à un modèle de reconnaissance des formes, et dispositif terminal
CN109409528A (zh) * 2018-09-10 2019-03-01 平安科技（深圳）有限公司 模型生成方法、装置、计算机设备及存储介质
CN110543946A (zh) * 2018-05-29 2019-12-06 百度在线网络技术（北京）有限公司 用于训练模型的方法和装置
CN111241745A (zh) * 2020-01-09 2020-06-05 深圳前海微众银行股份有限公司 逐步模型选择方法、设备及可读存储介质

Family Cites Families (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20180314971A1 (en) * 2017-04-26 2018-11-01 Midea Group Co., Ltd. Training Machine Learning Models On A Large-Scale Distributed System Using A Job Server
CN108875289B (zh) * 2017-05-08 2021-12-14 腾讯科技（深圳）有限公司 一种算法调试方法、客户端、后台服务器及系统
CN110298389A (zh) * 2019-06-11 2019-10-01 上海冰鉴信息科技有限公司 训练模型时的多轮循环特征选择方法和装置
CN110400215B (zh) * 2019-07-31 2023-11-03 浪潮软件集团有限公司 面向企业家族的小微企业信用评估模型构建方法及系统

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105550746A (zh) * 2015-12-08 2016-05-04 北京旷视科技有限公司 机器学习模型的训练方法和训练装置
WO2017219991A1 (fr) * 2016-06-23 2017-12-28 华为技术有限公司 Procédé et appareil d'optimisation adaptés à un modèle de reconnaissance des formes, et dispositif terminal
CN110543946A (zh) * 2018-05-29 2019-12-06 百度在线网络技术（北京）有限公司 用于训练模型的方法和装置
CN109409528A (zh) * 2018-09-10 2019-03-01 平安科技（深圳）有限公司 模型生成方法、装置、计算机设备及存储介质
CN111241745A (zh) * 2020-01-09 2020-06-05 深圳前海微众银行股份有限公司 逐步模型选择方法、设备及可读存储介质

Also Published As

Publication number Publication date
CN111241745B (zh) 2024-05-24
CN111241745A (zh) 2020-06-05

Similar Documents

Publication Publication Date Title
WO2021139462A1 (fr) Procédé et dispositif de sélection progressive de modèles et support lisible de stockage
WO2021139465A1 (fr) Dispositif et procédé de sélection de modèle arrière, et support d'enregistrement lisible
CN113095512A (zh) 联邦学习建模优化方法、设备、介质及计算机程序产品
US20220374776A1 (en) Method and system for federated learning, electronic device, and computer readable medium
CN107770783B (zh) 一种基站扩容改造方案设计方法及相关设备
EP4239491A1 (fr) Procédé et système de traitement de tables de données et d'entraînement automatique de modèle d'apprentissage machine
CN110287219B (zh) 一种数据处理方法及系统
WO2021139483A1 (fr) Procédé et dispositif de sélection de modèles de transfert et support lisible de stockage
CN113326852A (zh) 模型训练方法、装置、设备、存储介质及程序产品
CN111428883A (zh) 基于向后法的联邦建模方法、设备和可读存储介质
Hvarfner et al. Joint entropy search for maximally-informed Bayesian optimization
CN114881223B (zh) 深度学习模型的转换方法、装置、电子设备和存储介质
CN113902956B (zh) 融合模型的训练方法、图像融合方法、装置、设备及介质
WO2023024349A1 (fr) Procédé d'optimisation d'une prédiction fédérée verticale, dispositif, support et produit programme d'ordinateur
Bergami et al. Probabilistic trace alignment
CN112785144A (zh) 基于联邦学习的模型构建方法、设备及存储介质
CN111428884A (zh) 基于向前法的联邦建模方法、设备和可读存储介质
US11847599B1 (en) Computing system for automated evaluation of process workflows
Wang et al. Inverse reinforcement learning with graph neural networks for iot resource allocation
WO2020191722A1 (fr) Procédé et système pour déterminer une relation causale et produit-programme informatique
CN114926322B (zh) 图像生成方法、装置、电子设备和存储介质
CN112070162A (zh) 多类别处理任务训练样本构建方法、设备及介质
CN111274480A (zh) 用于内容推荐的特征组合方法及装置
US12020214B2 (en) System for applying an artificial intelligence engine in real-time to affect course corrections and influence outcomes
US20220245469A1 (en) Decision Making Using Integrated Machine Learning Models and Knowledge Graphs

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20911449

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20911449

Country of ref document: EP

Kind code of ref document: A1