A data processing method and system
By automatically building search space and adopting multiple search strategies, the existing gain modeling process has been solved, and an efficient gain modeling process has been achieved, which has reduced the professionalism of modeling personnel.
Patent Information
- Application Number
- CN202211642174.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-12-20
- Publication Date
- 2024-11-12
- Estimated Expiration
- 2042-12-20
AI Technical Summary
The existing gain modeling process has high threshold and low efficiency, and modeling personnel need to perform a lot of repetitive work, resulting in high professional requirements and low efficiency in modeling process.
Acquire data sets and intervention columns by responding to user interface configuration operations, perform feature analysis, build a search space, and generate a gain model in the search space according to the search strategy, using a single or fusion search strategy, including random selection, grid search, and evolutionary search , Monte Carlo, Bayesian optimization and meta-learning algorithms automatically build and optimize models.
It realizes automatic construction of gain models, lowers the modeling threshold, improves modeling efficiency, simplifies the gain modeling process, and reduces the professional requirements for modeling personnel.
Smart Images

Figure CN117236426B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of big data analysis, and in particular to a data processing method and system. Background Art
[0002] With the rapid development of artificial intelligence and big data, more and more industries use machine learning trained models for data analysis and prediction. For example, in the marketing field, gain models are used to analyze and predict marketing effects. However, the existing gain model modeling theory is relatively scattered, and modelers need to have strong professionalism for specific scenarios. In the feature screening, algorithm selection, effect evaluation and tuning of the modeling process, modelers need to perform a lot of repetitive work, which leads to high threshold and low efficiency of the gain modeling process. Summary of the invention
[0003] The present embodiment provides a data processing method and system to solve the problem that the existing gain modeling process has a high threshold and low efficiency.
[0004] In order to solve the above technical problems, the present invention provides a data processing method, which comprises:
[0005] In response to a configuration operation of the user interface, a configured data set and an intervention column are obtained;
[0006] Performing feature analysis on the data set and the intervention column to obtain data feature information and intervention information;
[0007] constructing a search space based on the data feature information and the intervention information;
[0008] A search is performed in the search space according to a search strategy to generate a gain model.
[0009] Optionally, in the above method, the data feature information includes quality information and feature distribution information, the intervention information includes intervention column type, and the step of constructing a search space based on the data feature information and the intervention information includes:
[0010] Determine a model stage subspace according to the quality information and / or feature distribution information, and / or intervention column type;
[0011] Based on the determined model stage subspace, a search space is constructed.
[0012] Optionally, in the above method, the model stage subspace includes one or more of a data preprocessing subspace, a data bias processing subspace, a feature engineering subspace, a model algorithm subspace, and a model evaluation subspace;
[0013] The step of determining the model stage subspace according to the quality information and / or feature distribution information and / or intervention column type comprises:
[0014] Determine a data preprocessing subspace according to the quality information and / or the feature distribution information; and / or,
[0015] determining a data bias processing subspace according to the intervention column type; and / or,
[0016] Determine a feature engineering subspace according to the quality information and / or the feature distribution information; and / or,
[0017] A model algorithm subspace and a model evaluation subspace are determined according to the feature distribution information and / or the intervention column type.
[0018] Optionally, in the above method, the step of searching in the search space according to the search strategy to generate a gain model includes:
[0019] Obtain a search strategy based on the configured number of training iterations and / or the constructed search space dimension;
[0020] Searching the search space according to the search strategy to generate a gain model;
[0021] The search strategy includes a single search strategy and a fusion search strategy, and the fusion search strategy includes two or more search algorithms.
[0022] Optionally, in the above method, the step of obtaining the search strategy according to the configured number of training iterations and / or the constructed search space dimension includes:
[0023] When the configured number of training iterations is less than a first threshold, determining that the search strategy is a single search strategy; and / or,
[0024] When the configured number of training iterations is greater than or equal to the first threshold, determining that the search strategy is a fusion search strategy; and / or,
[0025] When the dimension of the constructed search space exceeds a second threshold, determining that the search strategy is a fusion search strategy; and / or,
[0026] When the configured number of training iterations is greater than or equal to the first threshold and the dimension of the constructed search space exceeds the second threshold, the search strategy is determined to be a fusion search strategy.
[0027] Optionally, in the above method, if the search strategy is a fusion search strategy, before the step of searching in the search space according to the search strategy to generate a gain model, the method further includes:
[0028] The switching parameters of the fusion search strategy are determined according to the number of training iterations and / or the search space dimension; wherein the switching parameters include the initial population number n, the switching strategy threshold k and the search algorithm switching mode, and n and k are both positive integers.
[0029] Optionally, in the above method, the search algorithm includes any one of the following:
[0030] Random sampling, grid search, evolutionary search, Monte Carlo, Bayesian optimization, and meta-learning.
[0031] Optionally, in the above method, the step of searching in the search space according to the search strategy to generate a gain model includes:
[0032] Using an initial search algorithm to search in the search space, obtaining n groups of different training parameter combinations;
[0033] Using the current search algorithm to search based on the n sets of training parameter combinations, obtain candidate training parameter combinations, and form candidate model training instances corresponding to the candidate training parameter combinations;
[0034] Based on the data set, model training and scoring are performed on the candidate model training instances corresponding to the candidate training parameter combinations; if the growth change of the scores of the candidate model training instances generated for k consecutive times meets the preset conditions, the current search algorithm is switched to the target search algorithm based on the search algorithm switching mode, and the search is continued until the termination condition is reached, the search process is terminated, and the target search results are obtained;
[0035] A gain model is generated based on the target search results.
[0036] Optionally, in the above method, the step of searching in the search space according to the search strategy to generate a gain model includes:
[0037] Searching the search space based on the search strategy to generate a model training instance;
[0038] Model training is performed according to the data set and the model training instance to generate a gain model.
[0039] Optionally, in the above method, if the model training instances include two or more, the step of performing model training based on the data set and the model training instances to generate a gain model includes:
[0040] Determining the parallelism of the model training instances;
[0041] Based on the parallelism of the model training instance, the data set and the multiple model training instances, parallel model training is performed to generate multiple gain models.
[0042] Optionally, the above method further includes:
[0043] Using the gain model to predict the data set, to obtain a gain value corresponding to the intervention value;
[0044] The actual values of the target column and the intervention column of the data set and the predicted gain values are evaluated based on the evaluation index to obtain a gain evaluation result.
[0045] Optionally, the above method further includes:
[0046] Displays multiple historical model training information sorted by evaluation indicators.
[0047] Optionally, the method further includes at least one of the following:
[0048] If the intervention column type is a multi-value intervention column type, switching different intervention column values based on the user's first selection operation to sort and display the historical model training information;
[0049] Based on the user's second selection operation on multiple historical model training information, the target historical model detailed training information is displayed, and the target historical model detailed training information includes model scoring, scoring curve chart, model process and parameter combination of each stage.
[0050] The present invention also provides a data processing system, the system comprising:
[0051] An acquisition module, used for acquiring a configured data set and an intervention column in response to a configuration operation of the user interface;
[0052] A feature analysis module, used to perform feature analysis on the data set and intervention column to obtain data feature information and intervention information;
[0053] A construction module, used to construct a search space based on the data feature information and the intervention information;
[0054] The search module is used to search in the search space according to the search strategy to generate a gain model.
[0055] Optionally, in the above system, the data feature information includes quality information and feature distribution information, the intervention information includes intervention column type, and the construction module includes:
[0056] A determination submodule, configured to determine a model stage subspace according to the quality information and / or feature distribution information, and / or intervention column type;
[0057] The construction submodule is used to construct a search space based on the determined model stage subspace.
[0058] Optionally, in the above system, the model stage subspace includes one or more of a data preprocessing subspace, a data bias processing subspace, a feature engineering subspace, a model algorithm subspace, and a model evaluation subspace;
[0059] The determination submodule is specifically used for:
[0060] Determine a data preprocessing subspace according to the quality information and / or the feature distribution information; and / or,
[0061] determining a data bias processing subspace according to the intervention column type; and / or,
[0062] Determine a feature engineering subspace according to the quality information and / or the feature distribution information; and / or,
[0063] A model algorithm subspace and a model evaluation subspace are determined according to the feature distribution information and / or the intervention column type.
[0064] Optionally, in the above system, the search module includes:
[0065] An acquisition submodule, used to acquire a search strategy according to a configured number of training iterations and / or a constructed search space dimension;
[0066] A first generating submodule, configured to search the search space according to the search strategy to generate a gain model;
[0067] The search strategy includes a single search strategy and a fusion search strategy, and the fusion search strategy includes two or more search algorithms.
[0068] Optionally, in the above system, the acquisition submodule is specifically used for:
[0069] When the configured number of training iterations is less than a first threshold, determining that the search strategy is a single search strategy; and / or,
[0070] When the configured number of training iterations is greater than or equal to the first threshold, determining that the search strategy is a fusion search strategy; and / or,
[0071] When the dimension of the constructed search space exceeds a second threshold, determining that the search strategy is a fusion search strategy; and / or,
[0072] When the configured number of training iterations is greater than or equal to the first threshold and the dimension of the constructed search space exceeds the second threshold, the search strategy is determined to be a fusion search strategy.
[0073] Optionally, in the above system, if the search strategy is a fusion search strategy, the system further includes:
[0074] A determination module is used to determine the switching parameters of the fusion search strategy according to the number of training iterations and / or the search space dimension; wherein the switching parameters include an initial population number n and a switching strategy threshold k, and n and k are both positive integers.
[0075] Optionally, in the above system, the search algorithm includes any one of the following:
[0076] Random sampling, grid search, evolutionary search, Monte Carlo, Bayesian optimization, and meta-learning.
[0077] Optionally, in the above system, the first generating submodule is specifically used for:
[0078] Using an initial search algorithm to search in the search space, obtaining n groups of different training parameter combinations;
[0079] Using the current search algorithm to search on the basis of the n sets of training parameter combinations, a candidate training parameter combination is obtained, and a candidate model training instance corresponding to the candidate training parameter combination is formed; based on the data set, model training and scoring are performed on the candidate model training instance corresponding to the candidate training parameter combination; if the growth change of the score of the candidate model training instance generated k times in succession meets the preset condition, the current search algorithm is switched to the target search algorithm based on the search algorithm switching mode, and the search is continued until the termination condition is reached, the search process is terminated, and the target search result is obtained;
[0080] A gain model is generated based on the target search results.
[0081] Optionally, in the above system, the search module further includes:
[0082] A second generation submodule is used to search the search space based on the search strategy to generate a model training instance;
[0083] The model training submodule is used to perform model training according to the data set and the model training instance to generate a gain model.
[0084] Optionally, in the above system, if the model training instance includes two or more, the model training submodule is specifically used for:
[0085] Determining the parallelism of the model training instances;
[0086] Parallel model training is performed based on the parallelism of the model training instance, the data set and the multiple model training instances to generate multiple gain models.
[0087] Optionally, the above system further includes:
[0088] A prediction module, used to use the gain model to predict the data set to obtain a gain value corresponding to the intervention value;
[0089] An evaluation module is used to evaluate the actual values of the target column and the intervention column of the data set and the predicted gain value based on the evaluation index to obtain a gain evaluation result.
[0090] Optionally, the above system further includes:
[0091] The first display module is used to display multiple historical model training information in order according to evaluation indicators.
[0092] Optionally, the above system further includes:
[0093] A sorting module, for a plurality of pre-column types that are multi-value intervention column types, switching different intervention column values based on a first selection operation of a user to sort and display the historical model training information;
[0094] The second display module is used to display the detailed training information of the target historical model based on the user's second selection operation on multiple historical model training information. The detailed training information of the target historical model includes model scoring, scoring curve chart, model process and parameter combination of each stage.
[0095] An embodiment of the present invention further provides a data processing system, comprising a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the computer program implements the steps of the above-mentioned data processing method when executed by the processor.
[0096] An embodiment of the present invention further provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the data processing method described above are implemented.
[0097] The embodiment of the present invention automatically constructs a search space based on data feature information and intervention information; searches are performed in the automatically constructed search space according to a search strategy to generate a gain model, thereby realizing automatic construction of the gain model, improving the modeling efficiency of the gain model, and lowering the threshold for gain modeling. BRIEF DESCRIPTION OF THE DRAWINGS
[0098] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative labor.
[0099] Figure 1is a flow chart of a data processing method provided by an embodiment of the present invention;
[0100] Figure 2 is a schematic diagram of a configuration interface of a data processing system provided by an embodiment of the present invention;
[0101] Figure 3 is a schematic diagram of a process selection space provided by an embodiment of the present invention;
[0102] Figure 4 is a schematic diagram of an interface for iterative evolution settings provided by an embodiment of the present invention;
[0103] Figure 5 is a flow chart of model training provided by an embodiment of the present invention;
[0104] Figure 6 It is a structural block diagram of a data processing system provided by an embodiment of the present invention. DETAILED DESCRIPTION
[0105] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0106] See also Figure 1 , Figure 1 is a flow chart of a data processing method provided by an embodiment of the present invention, such as Figure 1 As shown, the data processing method includes the following steps:
[0107] Step 101 : In response to a configuration operation of a user interface, a configured data set and an intervention column are obtained.
[0108] When the data processing system builds a gain model, the user interface for model configuration is displayed, and the user can configure the data set and intervention column used to build the model in the configuration interface. Specifically, the data processing system supports multiple access methods for data sets, and is compatible with data sets in various formats, for example, it is compatible with data set formats such as csv, tsv, and txt. Users configure the data set for model training in the user interface of the data processing system, and can choose to access it in the form of file upload, database table, or URL according to the format of the accessed data set or actual needs.
[0109] Step 102: Perform feature analysis on the data set and intervention column to obtain data feature information and intervention information.
[0110] The intervention information includes the intervention column name, control value, and intervention column type.
[0111] Specifically, the intervention column is actually also data in the data set. In the embodiment of the present invention, the data processing system automatically analyzes the user-configured data set (the data set includes the intervention column data) and the characteristics of the intervention column. The specific analysis content includes but is not limited to the following options: field column type, field missing rate, intervention column type, target column type, field quantile, field dimension number, target column distribution, intervention column type, etc., so as to obtain data feature information and intervention information. The data feature information includes quality information and feature distribution information, and the intervention information includes the intervention column type. The quality information includes at least one of the following: null value column, missing rate, deduplication, outlier rate, etc.; the feature distribution information includes at least one of the following: maximum value, minimum value and statistical distribution type; the intervention column type includes the following optional items: binary intervention column, multi-value intervention column and continuous intervention column. The intervention column type can be determined based on the intervention value of the intervention column, whether it is a binary intervention column, a multi-value intervention column or a continuous intervention column.
[0112] For example, taking the marketing advertising campaign scenario as an example, the intervention column is whether to distribute discount coupons, and the difference in user purchase intention between those who distribute discount coupons and those who do not distribute discount coupons is analyzed by constructing a gain model. Figure 2 As shown, the user configuration interface of the data processing system is displayed. The user configuration interface of the data processing system displays the data set configured by the user - the coupon promotion data set. It can also display whether the user sets a randomized controlled experiment and the name of the target column field Y. In addition, the intervention information obtained by feature analysis of the intervention column can also be displayed. Specifically, the intervention column field name is set to treatment, and the control value (control) is set to 0.
[0113] In addition, the data set set by the user can be automatically split into a training set and a test set, the training set is used for model training, and the test set is used for model evaluation. Optionally, the split ratio can be configured by the user or defaulted by the data processing system. For example, the input coupon promotion data set is 100,000, of which 70,000 are split into a training set and 30,000 are split into a test set.
[0114] Step 103: construct a search space based on the data feature information and the intervention information.
[0115] Optionally, the data feature information includes quality information and feature distribution information, and the intervention information includes an intervention column type; and the step of constructing a search space based on the data feature information and the intervention information includes:
[0116] According to the quality information and / or feature distribution information, and / or intervention column type, a model stage subspace is determined; and based on the determined model stage subspace, a search space is constructed.
[0117] The model stage subspace includes one or more of a data preprocessing subspace, a data bias processing subspace, a feature engineering subspace, a model algorithm subspace, and a model evaluation subspace. The data preprocessing subspace, the data bias processing subspace, and the feature engineering subspace are optional subspaces, while the model algorithm subspace and the model evaluation subspace are mandatory subspaces.
[0118] Among them, the data preprocessing subspace includes the following options: feature type conversion, missing value filling, outlier processing, hot encoding (OneHot encoding) and normalization, etc.; the data bias processing subspace includes one of the following options: nearest neighbor matching and propensity matching, etc.; the feature selection subspace includes the following options: causal discovery algorithm, statistical filtering algorithm and split gain method, etc.; the algorithm selection subspace includes one of the following options: meta-learning algorithm Meta-Learner, category encoding conversion algorithm Class Transform and direct gain algorithm, etc.; Meta-Learner includes S-Learner, T-Learner, X-Learner, etc. The model evaluation subspace includes at least one of the following: AUUC and Gini index.
[0119] Optionally, the embodiment of the present invention provides the following feasible manner for implementing the step of determining the model stage subspace according to the quality information and / or feature distribution information, and / or intervention column type, including:
[0120] Determine a data preprocessing subspace according to the quality information and / or the feature distribution information; and / or,
[0121] determining a data bias processing subspace according to the intervention column type; and / or,
[0122] Determine a feature engineering subspace according to the quality information and / or the feature distribution information; and / or,
[0123] A model algorithm subspace and a model evaluation subspace are determined according to the feature distribution information and / or the intervention column type.
[0124] The following is a detailed description of the specific implementation of each feasible method for determining the subspace of the model stage. Specifically, the implementation of determining the data preprocessing subspace according to the quality information and / or the feature distribution information includes: determining the missing value filling method of the data preprocessing subspace according to the missing rate in the quality information. Exemplarily, if the field missing rate is high and exceeds a certain threshold, the field deletion method can be used; if the field missing value is medium or low, the median filling method, the mode filling method, etc. can be used; if the numerical missing value is very low (below a certain threshold, such as only a few samples), the sample deletion method can be used. The outlier processing method of the data preprocessing subspace can be determined according to the abnormal rate in the quality information and the feature distribution information. Exemplarily, for a numerical type field, it can be determined whether there are outliers based on the distribution of the feature; for example, according to the ±3σ principle, the number of samples outside ±3σ is counted. If the number of outliers exceeds a certain threshold or ratio, methods such as setting it as the mode, filling in empty values, and not processing can be used. The feature type conversion method can also be determined according to the type of the target column. For example, for a numeric type field, if the different values of the field are low, it can be converted to a categorical type, etc. The encoding method can also be determined according to the type of the target column. For example, for a categorical field, a hot encoding conversion can be performed.
[0125] The data preprocessing subspace belongs to a multiple-choice subspace, such as being able to select mean filling, outlier processing, and hot encoding at the same time. In particular, each selected processing method also has different options, such as missing value filling is selected as mean filling, outlier processing is selected as sample deletion, and the data preprocessing method in the final generated model training instance is a combination of mean filling + outlier deletion + hot encoding. For example, taking a marketing advertising campaign scenario as an example, the features of the data set include fields such as postal code, gender, and whether a discount coupon has been used for purchase, among which a small part of the gender field is empty. The generated model training instance may include mode filling of the gender field, and may include a category conversion method for converting the postal code from a numerical type to a categorical type.
[0126] Specifically, the feature engineering subspace can be determined according to the feature distribution information. The feature engineering subspace is an optional subspace, and the generated model training instance may or may not include this stage. And the feature engineering subspace is a single-choice subspace, and only one method can be selected from multiple feature selection methods. Among them, the feature engineering subspace includes causal discovery algorithms, statistical filtering algorithms, and split gain methods. The causal discovery algorithm includes but is not limited to PC, Notears, GES and other algorithms. The causal discovery algorithm selects the features that have an impact on the target variable and the intervention variable, and can take the union or intersection of all the features that have an impact on the target variable and the intervention variable as part of the feature engineering subspace. The statistical filtering algorithm uses a variety of statistical filtering algorithms, including but not limited to one or more of the F test, logistic regression likelihood ratio statistics, etc. for feature selection. Taking the F statistic method as an example, first calculate the F statistic of each feature for the target column as the importance score, and sort the scores, and extract the first random number of features as part of the feature engineering subspace. The split gain method includes but is not limited to KL divergence, Euclidean distance, chi-square and other methods. Taking the Euclidean method as an example, first calculate the Euclidean distance gain of each feature after grouping, and use the square of the difference in the number of target columns corresponding to a feature in the intervention group and the control group as the measurement value; before splitting, calculate the Euclidean distance on the overall data as the pre-split value, use the feature values to perform equidistant binning, divide the feature values into different groups of data as subgroups, calculate the Euclidean distance of each subgroup and sum them as the post-split value, subtract the post-split value from the pre-split value as the gain value of the split, sort from large to small by the gain value, and extract the random number of features before extraction as part of the feature engineering subspace. For example, taking the marketing advertising campaign scenario as an example, the data is not a randomized controlled experiment, and the generated model training instance may include a method for feature screening using the F test method.
[0127] Specifically, the data bias processing subspace can be determined according to the intervention column type. If the intervention column type is a binary column, it is determined that the constructed search space contains the data bias processing subspace; if the pre-column type is a multi-value intervention column or a continuous intervention column, it is determined that the constructed search space does not contain the data bias processing subspace. For example, if the intervention column name is treament, the column has only two values 1 and 0, the control value can be set to 0, and the random control data set is set to no, the obtained search space contains the data bias processing subspace. Data bias processing can reconstruct the training data to eliminate the bias of the data. Exemplarily, taking the propensity score matching method as an example, propensity is a common statistical processing method for data bias, and the data is divided into an intervention group (value 1) and a control group (value 0) according to the value of the intervention column. First, the data with an intervention column value of 1 is set as the intervention group, and the data with an intervention column value of 0 is set as the control group. The caliper propensity score matching method based on logistic regression is used to reconstruct the control group data, that is, the data of the intervention value 0 group. The intervention group data and the reconstructed control group data are merged as the subsequent training data set. Data bias is handled as a single-selectable subspace, and only one method of the optional subspace can be selected or not selected. For example, taking a marketing advertising campaign scenario as an example, the data is not a randomized controlled experiment data, and the generated model training instance may include a propensity score matching method.
[0128] Specifically, the model algorithm subspace can be determined according to the feature distribution information and the intervention column type, wherein the model algorithm subspace includes the algorithm type and the algorithm parameters. For each different algorithm, there are also different algorithm parameters. The algorithm parameters include specific algorithms, hyperparameters, etc. Exemplarily, if the intervention column is a binary intervention column type and it is a binary classification problem, then the model algorithm subspace is determined to be a model algorithm that can support the binary intervention column type, such as the ClassTransform algorithm. For example, when the target column is a regression column, it is determined that the model algorithm subspace contains a regression algorithm, such as linear regression; and when the target column is a classification type, it is determined that the model algorithm subspace is a classification algorithm, such as random forest, logistic regression, etc. The algorithm selection stage is a mandatory stage, and this stage must be included in the process of implementing the model training. Exemplarily, taking the marketing advertising activity scenario as an example, the data is not a randomized controlled experiment data, and the ClassTransform method may be used in the generated model training instance, and the parameter combination is XGBClassifier and the corresponding XGBClassifier default parameter value.
[0129] Specifically, the model evaluation subspace may be determined according to the feature distribution information and / or intervention column type. The model evaluation subspace includes but is not limited to Auuc, Gini and other methods.
[0130] Based on the characteristic analysis of the data set and the model parameters, the corresponding subspace is selected from the optional stages such as data preprocessing, data bias processing and feature selection, and the process selection space is formed by combining the required stages such as algorithm stage, model training and model evaluation. The generated process selection space is as follows: Figure 3 As shown. Each model stage has its own optional subspace, that is, different model stages can have multiple optional actual processing methods. For example, data preprocessing is a multiple-choice subspace. When each optional method does not conflict with each other, the generated parameter flow can contain multiple optional data preprocessing methods in the instance. The data bias processing subspace is a single-choice space, and the generated model training instance can only choose one of the methods or not; the feature selection subspace is a single-choice space, and the generated model training instance can only choose one method or not; the model algorithm subspace is a single-choice space, and the generated model training instance can only choose one of the algorithms, and it cannot be unselected.
[0131] Step 104: Search in the search space according to the search strategy to generate a gain model.
[0132] Among them, the search strategy includes search algorithms, which include but are not limited to: random sampling, grid search, evolutionary search, Monte Carlo, Bayesian optimization and meta-learning.
[0133] Specifically, the search space includes all model stage subspaces of the model training process. The search algorithm in the search strategy searches in the high-dimensional search space, and selects different parameter combinations to obtain the corresponding model training instances, thereby obtaining the gain model. If the single selected model training instance is completely consistent with all the subspace parameters contained in the model training instance that has been trained, it is necessary to abandon the model training instance generated this time, and search again to generate a new model training instance. Model training for the training data set includes training according to the process stages of the model training instance. Among them, data preprocessing is first performed, followed by feature preprocessing, bias processing, feature engineering, feature selection, model training, etc.; the gain model is generated through the above process.
[0134] In order to increase the speed at which the search strategy searches for parameter combinations for model training in a high-dimensional search space and improve the efficiency of constructing a gain model, the embodiment of the present invention optionally includes the step of searching in the search space according to the search strategy to generate a gain model, including: obtaining a search strategy according to a configured number of training iterations and / or a constructed search space dimension; searching in the search space according to the search strategy to generate a gain model; wherein the search strategy includes a single search strategy and a fusion search strategy, and the fusion search strategy includes two or more search algorithms.
[0135] In an embodiment of the present invention, if the dimension of the constructed search space is low or the number of training iterations is set to be low, a single search algorithm can be used to quickly search for parameters for constructing a gain model; if the dimension of the constructed search space is high or the number of training iterations is set to be high, a combination of multiple search algorithms can be used to quickly and accurately search for parameters for constructing a gain model.
[0136] Optionally, the present invention provides the following feasible methods for implementing the step of obtaining the search strategy according to the configured number of training iterations and / or the constructed search space dimension, including:
[0137] When the configured number of training iterations is less than a first threshold, determining that the search strategy is a single search strategy; and / or,
[0138] When the configured number of training iterations is greater than or equal to the first threshold, determining that the search strategy is a fusion search strategy; and / or,
[0139] When the dimension of the constructed search space exceeds a second threshold, determining that the search strategy is a fusion search strategy; and / or,
[0140] When the configured number of training iterations is greater than or equal to the first threshold and the dimension of the constructed search space exceeds the second threshold, the search strategy is determined to be a fusion search strategy.
[0141] Optionally, if the search strategy is a fusion search strategy, before the step of searching in the search space according to the search strategy to generate a gain model, the method further includes:
[0142] The switching parameters of the fusion search strategy are determined according to the number of training iterations and / or the search space dimension; wherein the switching parameters include the initial population number n, the switching strategy threshold k and the search algorithm switching mode, and n and k are both positive integers.
[0143] The search algorithm switching method includes random switching or polling switching.
[0144] In a feasible embodiment, the above-mentioned automatic determination of the switching parameters of the fusion search strategy according to the number of training iterations and / or the search space dimension specifically includes: when the number of training iterations reaches a certain value and / or the search space dimension reaches a certain value, the system determines the switching parameters of the fusion search strategy according to an empirical algorithm. For example, when the number of training iterations or the search space dimension reaches a certain value, the system determines the initial population number n in the switching parameters of the fusion search strategy according to the empirical algorithm to be one-half of the search space dimension, and the switching strategy threshold k is one-fifth of the number of training iterations; for another example, when the number of training iterations or the search space dimension reaches a first limit value, the initial population number n is determined to be a first default value, and the switching strategy threshold k is a second default value; when the number of training iterations or the search space dimension reaches a second limit value, the initial population number n is determined to be a third default value, and the switching strategy threshold k is a fourth default value, wherein, when the number of training iterations is larger or the search space dimension is higher, the initial population number n and the switching strategy threshold k can be set to be larger.
[0145] Furthermore, the switching parameters in the fusion search strategy can also be set by the user. The data processing system displays an iterative evolution setting interface for setting model training iterations. In response to the user's configuration operation on the iterative evolution setting interface, the initial population number n, the switching strategy threshold k, and the search algorithm switching mode are obtained. Exemplarily, the user can set the switching parameters of the fusion search strategy on the iterative evolution setting interface of the data processing system. Figure 4 As shown, the user can set the number of initialization populations to 30, the number of switching strategy thresholds to 20, the total number of training times to 200, and the search algorithm switching mode to random switching in the iterative evolution setting interface.
[0146] Optionally, in the embodiment of the present invention, if the search strategy is a fusion search strategy, the step of searching in the search space according to the search strategy to generate a gain model includes:
[0147] Using an initial search algorithm to search in the search space, obtaining n groups of different parameter combinations to form n different initial parameter combinations;
[0148] Using the current search algorithm to search based on the n sets of training parameter combinations, obtain candidate training parameter combinations, and form candidate model training instances corresponding to the candidate training parameter combinations;
[0149] Based on the data set, model training and scoring are performed on the candidate model training instances corresponding to the candidate training parameter combinations; if the growth change of the scores of the candidate model training instances generated for k consecutive times meets the preset conditions, the current search algorithm is switched to the target search algorithm based on the search algorithm switching method, and the search is continued until the termination condition is reached, the search process is terminated, and the target search results are obtained; based on the target search results, a gain model is generated.
[0150] The above-mentioned preset condition may be that the growth change rate is less than a preset value, or that the growth value between the first score and the kth score is less than a preset value. The above-mentioned termination condition may be that the number of training iterations is reached, or that there is no new parameter combination, etc. The above-mentioned description of the preset conditions and termination conditions is only a feasible method provided by the present invention, and the present invention does not specifically limit the setting of the preset conditions and termination conditions.
[0151] Based on the characteristics of automatic modeling, the embodiment of the present invention supports the iterative evolution of the model, such as Figure 5 As shown in the figure, a fusion search strategy is used to continuously generate parameter combination process instances in the search space through an iterative evolution method to train the model. The iterative process is terminated when the training termination condition is reached, thereby completing the entire automatic modeling training. The specific process is as follows:
[0152] a): Get the number of initialization populations n, where n is a positive integer; the fusion search strategy first uses random extraction to extract n groups of different parameter combinations in the search space to form n different parameter combination process instances. Perform model training on each process instance based on the training data set. Model training can be distributed or stand-alone, and use scoring indicators to score each trained process instance to obtain the initial parameter combination scoring history, and generate the initial parameter combination scoring history;
[0153] b): After n times, based on the search algorithm switching method, such as random switching, the random prime search strategy is switched to a certain search strategy for searching, such as switching to Monte Carlo search. Based on the existing parameter combination scoring history, Monte Carlo search is used to select m groups of different parameter combinations in the search space to form candidate parameter combination process instances, and then form candidate parameter combination process instances. The candidate parameter combination process instances are trained and scored using the training data set, and the scoring results are added to the parameter combination scoring history.
[0154] c): If the current search strategy performs poorly and the scores of the process instances generated for k consecutive times (k is the number of times the strategy is switched to a threshold) do not increase, then switch to another search strategy to continue searching in the search space, such as switching to an evolutionary algorithm for search;
[0155] d): Repeat the search process of c, generate different parameter combination process instances in the high-dimensional optional process space by integrating multiple search strategies, and complete the entire automatic modeling training until the termination condition is reached. The termination condition includes that the number of model training times reaches a threshold or that no new parameter combination can be found in the search space; the threshold is the total number of model training times.
[0156] To prevent duplicate parameter combination process instances, the parameter combination process instances generated in any of the above strategies will check the existing parameter combination scoring history. If a duplicate parameter combination process instance appears, a new parameter combination process instance will be regenerated using the current strategy. For example, if the current strategy is a random search, when a duplicate parameter combination process is encountered, a new parameter combination process instance will be randomly generated.
[0157] Switching between multiple search algorithms supports two modes: random switching or round-robin switching. Random switching means randomly switching to a search algorithm in the above steps b, c, and d. For example, in step b, it randomly switches to evolutionary search, and in step c, it randomly switches to meta-learning algorithm. Round-robin switching means setting a fixed order for all search algorithms, and switching to the next search algorithm in steps b and c in a fixed order. Before reaching the termination condition of the search (such as the total number of training times), it supports the same search algorithm to be switched multiple times.
[0158] For example, take the marketing advertising campaign scenario as an example. Figure 4 As shown, the parameters for iterative evolution setting are as follows: the number of population initializations is 30, the number of strategy switching thresholds is 20, the total number of training times is 200, and the switching strategy is random switching. The data processing system first uses a random search algorithm to initialize and randomly generate 30 parameter combination process instances in the selection space, performs model training and scoring on these 30 parameter combination process instances, and obtains different parameter combinations and scoring values; then randomly switches to Monte Carlo search, generates parameter combination process instances in the selection space, and continues to perform model training and scoring on the generated parameter combination process instances. The scores are improved in the 45th, 62nd, and 80th times, respectively, until the 100th process instance, when the scores for the last 20 consecutive times have not improved; in this case, randomly switches to the evolutionary algorithm, generates new parameter combination process instances in the selection space for model training and scoring, and the scores are improved in the 115th, 129th, and 140th times, respectively, until the 160th process instance, when the scores for the last 20 times have not improved; in this case, switches to the meta-learning algorithm, generates parameter combination process instances in the selection space for training and scoring, and the scores are improved in the 165th, 180th, and 179th times, respectively, until the 200th time, reaching the total number of training times, and completing the model iteration evolution stage.
[0159] The present invention implements and supports iterative evolution of the gain model, improves the intelligence level of the gain model training, and improves the accuracy of the gain model training.
[0160] The step 104 of searching the search space according to the search strategy to generate a gain model includes:
[0161] Searching the search space based on the search strategy to generate a model training instance;
[0162] Model training is performed according to the data set and the model training instance to generate a gain model.
[0163] Optionally, when the model training instance includes two or more, the step of performing model training based on the data set and the model training instance to generate a gain model includes:
[0164] Determining the parallelism of the model training instances;
[0165] Based on the parallelism of the model training instance, the data set and the multiple model training instances, parallel model training is performed to generate multiple gain models.
[0166] Specifically, the data processing system supports parallel training or single-machine training. Parallel training is to generate multiple different model training parallel instances from the optional process space for training at the same time. In addition, the training parallelism can be set according to the resources and task volume of the data processing system, or the data processing system has a default training parallelism.
[0167] Optionally, the method further includes:
[0168] Using the gain model to predict the data set, to obtain a gain value corresponding to the intervention value;
[0169] The actual values of the target column and the intervention column of the data set and the predicted gain values are evaluated based on the evaluation index to obtain a gain evaluation result.
[0170] Specifically, the gain model obtained through training is used to make predictions on the test set to obtain the gain value corresponding to the specified intervention value, or the gain value of all intervention values. It should be noted that model prediction includes data preprocessing, feature selection and other stages of parameter combination instances of model training. The gain model obtained through model training is applied to the test set to ensure consistency in the data processing process. The actual values of the target column and intervention column of the test data set are used to score the predicted gain values to form a score value result.
[0171] It should be noted that it is necessary to distinguish the differences in the scoring methods for binary intervention columns, multi-value intervention columns, and continuous intervention columns. For binary intervention column types, the gain value between the non-control value and the control value is predicted, and the scoring method is to score the obtained gain value. For multi-value intervention column types, the gain value of each non-control value relative to the control value is predicted separately, and the gain value of each non-control value relative to the control value is scored. For continuous intervention columns, the gain value between the specified intervention value entered by the user and the control value and the corresponding scoring index are predicted.
[0172] The embodiment of the present invention can also provide a visual display of the training results of each model implemented above, so as to facilitate the user to understand the information of the model training process. The specific visual display is as follows:
[0173] Optionally, the method further includes:
[0174] Displays multiple historical model training information sorted by evaluation indicators.
[0175] The method further comprises:
[0176] If the intervention column type is a multi-value intervention column type, switching different intervention column values based on the user's first selection operation to sort and display the historical model training information;
[0177] Based on the user's second selection operation on multiple historical model training information, the target historical model detailed training information is displayed, and the target historical model detailed training information includes model scoring, scoring curve chart, model process and parameter combination of each stage.
[0178] Specifically, a training history list of all training sub-processes can be provided on the model training result page of the data processing system, and the training history list is sorted and displayed by default according to the specified scoring index. If it is a multi-value intervention column type, the historical model training information can be sorted and displayed by switching different intervention column values based on the user's first selection operation. When the user's second selection operation for multiple historical model training information is received, the detailed training information of the target historical model in the historical model list selected by the user is displayed, including model scores, scoring curves, model processes, and parameter combinations for each stage.
[0179] For example, taking the marketing advertising campaign scenario as an example, the model training result list shows the training history of 200 process instances, sorted from high to low according to the AUUC score. After clicking, the user can see the basic information of the model training, including that the data is not randomized controlled trial data, the intervention column field name is treatment, the control value is 0, the target column field name is Y, and the score wants to see AUUC. The user can see that the missing values of the gender field of the current process instance are filled with the mode, the postal code value is converted from numeric to categorical; the data bias processing method of propensity matching is adopted; the F test method is used for feature screening, the model training is carried out through the Class Transform method, and the XGB Classifier method and the default parameter value are used as the specific algorithm. The user can view the final score value AUUC and the corresponding curve. The AUUC curve chart shows that the user judgment model algorithm has been improved to a certain extent, which can better avoid the loss of some potential customers than randomly issuing coupons. For example, some customers did not buy the product, but once they were issued coupons, they had a higher conversion rate, or some customers originally had a high willingness to buy, but after being issued coupons, they did not want to buy. Then, the model can be used to predict the gain of customers' future purchases of products, specify more accurate coupon issuance strategies, and increase marketing benefits. Based on the above method, marketing business personnel can also model based on data and meet certain professional modeling scoring standards.
[0180] The embodiment of the present invention improves the modeling efficiency of the gain model and lowers the threshold of gain modeling by automatically constructing the gain model.
[0181] Based on the data processing method provided in the above embodiment, the embodiment of the present invention also provides an implementation of the above data processing system, please refer to Figure 6 The data processing system 600 provided in the embodiment of the present invention includes:
[0182] An acquisition module 601 is used to acquire a configured data set and an intervention column in response to a configuration operation of the user interface;
[0183] A feature analysis module 602 is used to perform feature analysis on the data set and the intervention column to obtain data feature information and intervention information;
[0184] A construction module 603, configured to construct a search space based on the data feature information and the intervention information;
[0185] The search module 604 is used to search in the search space according to the search strategy to generate a gain model.
[0186] Optionally, the data feature information includes quality information and feature distribution information, the intervention information includes intervention column type, and the construction module 603 includes:
[0187] A determination submodule, configured to determine a model stage subspace according to the quality information and / or feature distribution information, and / or intervention column type;
[0188] The construction submodule is used to construct a search space based on the determined model stage subspace.
[0189] Optionally, the model stage subspace includes one or more of a data preprocessing subspace, a data bias processing subspace, a feature engineering subspace, a model algorithm subspace, and a model evaluation subspace;
[0190] The determination submodule is specifically used for:
[0191] Determine a data preprocessing subspace according to the quality information and / or the feature distribution information; and / or,
[0192] determining a data bias processing subspace according to the intervention column type; and / or,
[0193] Determine a feature engineering subspace according to the quality information and / or the feature distribution information; and / or,
[0194] A model algorithm subspace and a model evaluation subspace are determined according to the feature distribution information and / or the intervention column type.
[0195] Optionally, the search module 604 includes:
[0196] An acquisition submodule, used to acquire a search strategy according to a configured number of training iterations and / or a constructed search space dimension;
[0197] A first generating submodule, configured to search the search space according to the search strategy to generate a gain model;
[0198] The search strategy includes a single search strategy and a fusion search strategy, and the fusion search strategy includes two or more search algorithms.
[0199] Optionally, the acquisition submodule is specifically used for:
[0200] When the configured number of training iterations is less than a first threshold, determining that the search strategy is a single search strategy; and / or,
[0201] When the configured number of training iterations is greater than or equal to the first threshold, determining that the search strategy is a fusion search strategy; and / or,
[0202] When the dimension of the constructed search space exceeds a second threshold, determining that the search strategy is a fusion search strategy; and / or,
[0203] When the configured number of training iterations is greater than or equal to the first threshold and the dimension of the constructed search space exceeds the second threshold, the search strategy is determined to be a fusion search strategy.
[0204] Optionally, if the search strategy is a fusion search strategy, the system further includes:
[0205] A determination module is used to determine the switching parameters of the fusion search strategy based on the number of training iterations and / or the search space dimension; wherein the switching parameters include the initial population number n, the switching strategy threshold k and the search algorithm switching mode, and n and k are both positive integers.
[0206] Optionally, the search algorithm includes any of the following:
[0207] Random sampling, grid search, evolutionary search, Monte Carlo, Bayesian optimization, and meta-learning.
[0208] Optionally, the first generating submodule is specifically used for:
[0209] Using an initial search algorithm to search in the search space, obtaining n groups of different training parameter combinations;
[0210] Performing model training on each of the process instances based on the data set, and scoring each of the trained process instances to obtain an initial parameter combination scoring history;
[0211] Using the current search algorithm to search on the basis of the n sets of training parameter combinations, a candidate training parameter combination is obtained, and a candidate model training instance corresponding to the candidate training parameter combination is formed; based on the data set, model training and scoring are performed on the candidate model training instance corresponding to the candidate training parameter combination; if the growth change of the score of the candidate model training instance generated k times in succession meets the preset condition, the current search algorithm is switched to the target search algorithm based on the search algorithm switching mode, and the search is continued until the termination condition is reached, the search process is terminated, and the target search result is obtained;
[0212] A gain model is generated based on the target search results.
[0213] Optionally, the search module 604 further includes:
[0214] A second generation submodule is used to search the search space based on the search strategy to generate a model training instance;
[0215] The model training submodule is used to perform model training according to the data set and the model training instance to generate a gain model.
[0216] Optionally, if the model training instance includes two or more, the model training submodule is specifically used for:
[0217] Determining the parallelism of the model training instances;
[0218] Based on the parallelism of the model training instance, the data set and the multiple model training instances, parallel model training is performed to generate multiple gain models.
[0219] Optionally, the system further includes:
[0220] A prediction module, used to use the gain model to predict the data set to obtain a gain value corresponding to the intervention value;
[0221] An evaluation module is used to evaluate the actual values of the target column and the intervention column of the data set and the predicted gain value based on the evaluation index to obtain a gain evaluation result.
[0222] Optionally, the system further includes:
[0223] The first display module is used to display multiple historical model training information in order according to evaluation indicators.
[0224] Optionally, the system further includes:
[0225] A sorting module, for a plurality of pre-column types that are multi-value intervention column types, switching different intervention column values based on a first selection operation of a user to sort and display the historical model training information;
[0226] The second display module is used to display the detailed training information of the target historical model based on the user's second selection operation on multiple historical model training information. The detailed training information of the target historical model includes model scoring, scoring curve chart, model process and parameter combination of each stage.
[0227] An embodiment of the present invention provides a data processing system, including a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the computer program implements the steps of the data processing method described in the above embodiment when executed by the processor.
[0228] An embodiment of the present invention provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the data processing method described in the above embodiment are implemented.
[0229] The embodiment of the present invention further provides a readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, each process of the above-mentioned data processing method embodiment is implemented, and the same technical effect can be achieved. To avoid repetition, it is not repeated here. The computer readable storage medium is, for example, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.
[0230] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or device including the element.
[0231] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, a magnetic disk, or an optical disk), and includes a number of instructions for enabling a terminal (which can be a mobile phone, a computer, a server, an air conditioner, or a network device, etc.) to execute the methods described in each embodiment of the present invention.
[0232] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.
Claims
1. A data processing method, characterized in that: The method comprises: In response to a configuration operation of the user interface, a configured data set and an intervention column are obtained; Performing feature analysis on the data set and the intervention column to obtain data feature information and intervention information; constructing a search space based on the data feature information and / or the intervention information; Searching the search space according to the search strategy to generate a gain model; The data feature information includes quality information and feature distribution information, the intervention information includes intervention column type, and the step of constructing a search space based on the data feature information and / or the intervention information includes: Determine a model stage subspace according to at least one of the quality information, the feature distribution information, and the intervention column type; Constructing a search space based on the determined model stage subspace; The model stage subspace includes one or more of a data preprocessing subspace, a data bias processing subspace, a feature engineering subspace, a model algorithm subspace, and a model evaluation subspace; The step of determining the model stage subspace according to at least one of the quality information, feature distribution information, and intervention column type comprises at least one of the following: Determine a data preprocessing subspace according to the quality information and / or the feature distribution information; Determining a data bias processing subspace according to the intervention column type; Determine a feature engineering subspace according to the quality information and / or the feature distribution information; A model algorithm subspace and a model evaluation subspace are determined according to the feature distribution information and / or the intervention column type.
2. The data processing method according to claim 1, characterized in that: Before searching the search space according to the search strategy to generate the gain model, the method further includes: Obtain a search strategy based on the configured number of training iterations and / or the constructed search space dimension; The search strategy includes a single search strategy and a fusion search strategy, and the fusion search strategy includes two or more search algorithms.
3. The data processing method according to claim 2, characterized in that: The step of obtaining the search strategy according to the configured number of training iterations and / or the constructed search space dimension includes: According to the configured number of training iterations being less than the first threshold, determining that the search strategy is a single search strategy; or, According to the configured number of training iterations being greater than or equal to the first threshold, it is determined that the search strategy is a fusion search strategy.
4. The data processing method according to claim 2, characterized in that: The step of obtaining the search strategy according to the configured number of training iterations and / or the constructed search space dimension includes: According to the constructed search space dimension exceeding the second threshold, determining that the search strategy is a fusion search strategy; or, According to the configured number of training iterations being greater than or equal to the first threshold and the constructed search space dimension exceeding the second threshold, it is determined that the search strategy is a fusion search strategy.
5. The data processing method according to any one of claims 2 to 4, characterized in that: If the search strategy is a fusion search strategy, before the step of searching the search space according to the search strategy to generate a gain model, the method further includes: The switching parameters of the fusion search strategy are determined according to the number of training iterations and / or the search space dimension; wherein the switching parameters include the initial population number n, the switching strategy threshold k and the search algorithm switching mode, and n and k are both positive integers.
6. The data processing method according to claim 5, characterized in that: The search algorithm includes any of the following: Random sampling, grid search, evolutionary search, Monte Carlo, Bayesian optimization, and meta-learning.
7. The data processing method according to claim 5, characterized in that: The step of searching in the search space according to the search strategy to generate a gain model comprises: Using an initial search algorithm to search in the search space, obtaining n groups of different training parameter combinations; Using the current search algorithm to search based on the n sets of training parameter combinations, obtain candidate training parameter combinations, and form candidate model training instances corresponding to the candidate training parameter combinations; Performing model training and scoring on the candidate model training instances corresponding to the candidate training parameter combinations based on the data set; If the growth change of the score of the candidate model training instance generated k times in succession meets the preset condition, the current search algorithm is switched to the target search algorithm based on the search algorithm switching mode, and the search is continued until the termination condition is reached, the search process is terminated, and the target search result is obtained; A gain model is generated based on the target search results.
8. The data processing method according to claim 1 or 2, characterized in that: The step of searching in the search space according to the search strategy to generate a gain model comprises: Searching the search space based on the search strategy to generate a model training instance; Model training is performed according to the data set and the model training instance to generate a gain model.
9. The data processing method according to claim 8, characterized in that: If the model training instances include two or more, the step of performing model training based on the data set and the model training instances to generate a gain model includes: Determining the parallelism of the model training instances; Parallel model training is performed based on the parallelism of the model training instance, the data set and the multiple model training instances to generate multiple gain models.
10. The data processing method according to claim 1, characterized in that: The method further comprises: Using the gain model to predict the data set, to obtain a gain value corresponding to a target intervention value; The actual values of the target column and the intervention column of the data set and the gain value are evaluated based on the evaluation index to obtain a gain evaluation result.
11. The data processing method according to claim 1 or 2, characterized in that: The method further comprises: The multiple historical model training information are sorted according to the evaluation index, and the sorted multiple historical model training information are displayed.
12. The data processing method according to claim 11, characterized in that: The method further comprises: Based on the user's selection operation on the target historical model, the training information corresponding to the target historical model is displayed, and the training information includes model scoring, scoring curve chart, model process and parameter combination of each stage.
13. A data processing system, characterized in that: The system comprises: A first acquisition module, configured to acquire a configured data set and an intervention column in response to a configuration operation of the user interface; A feature analysis module, used to perform feature analysis on the data set and intervention column to obtain data feature information and intervention information; A construction module, configured to construct a search space based on the data feature information and / or the intervention information; A search module, used to search in the search space according to a search strategy to generate a gain model; The data feature information includes quality information and feature distribution information, the intervention information includes intervention column type, and the construction module includes: A determination submodule, configured to determine a model stage subspace according to at least one of the quality information, the feature distribution information, and the intervention column type; A construction submodule, used for constructing a search space based on the determined model stage subspace; The model stage subspace includes one or more of a data preprocessing subspace, a data bias processing subspace, a feature engineering subspace, a model algorithm subspace, and a model evaluation subspace; The determination submodule is specifically used for at least one of the following: Determine a data preprocessing subspace according to the quality information and / or the feature distribution information; determine a data bias processing subspace according to the intervention column type; determine a feature engineering subspace according to the quality information and / or the feature distribution information; determine a model algorithm subspace and a model evaluation subspace according to the feature distribution information and / or the intervention column type.
14. The data processing system according to claim 13, characterized in that: The system further comprises: An acquisition submodule, used to acquire a search strategy according to a configured number of training iterations and / or a constructed search space dimension; A first generating submodule, configured to search the search space according to the search strategy to generate a gain model; The search strategy includes a single search strategy and a fusion search strategy, and the fusion search strategy includes two or more search algorithms.
15. The data processing system according to claim 14, characterized in that: The acquisition submodule is specifically used for: According to the configured number of training iterations being less than the first threshold, determining that the search strategy is a single search strategy; or, According to the configured number of training iterations being greater than or equal to the first threshold, it is determined that the search strategy is a fusion search strategy.
16. The data processing system according to claim 14, characterized in that: The acquisition submodule is specifically used for: According to the constructed search space dimension exceeding the second threshold, determining that the search strategy is a fusion search strategy; or, According to the configured number of training iterations being greater than or equal to the first threshold and the constructed search space dimension exceeding the second threshold, it is determined that the search strategy is a fusion search strategy.
17. The data processing system according to claim 14, characterized in that: If the search strategy is a fusion search strategy, the system further includes: A determination module is used to determine the switching parameters of the fusion search strategy according to the number of training iterations and / or the search space dimension; wherein the switching parameters include the initial population number n, the switching strategy threshold k and the search algorithm switching mode, and n and k are both positive integers.
18. The data processing system according to claim 14, characterized in that: The search algorithm includes any of the following: Random sampling, grid search, evolutionary search, Monte Carlo, Bayesian optimization, and meta-learning.
19. The data processing system according to claim 17, characterized in that: The first generation submodule is specifically used for: Using an initial search algorithm to search in the search space, obtaining n groups of different training parameter combinations; Using the current search algorithm to search based on the n sets of training parameter combinations, obtain candidate training parameter combinations, and form candidate model training instances corresponding to the candidate training parameter combinations; Performing model training and scoring on the candidate model training instances corresponding to the candidate training parameter combinations based on the data set; If the growth change of the score of the candidate model training instance generated k times in succession meets the preset condition, the current search algorithm is switched to the target search algorithm based on the search algorithm switching mode, and the search is continued until the termination condition is reached, the search process is terminated, and the target search result is obtained; A gain model is generated based on the target search results.
20. The data processing system according to claim 13, characterized in that: The search module also includes: A second generation submodule is used to search the search space based on the search strategy to generate a model training instance; The model training submodule is used to perform model training according to the data set and the model training instance to generate a gain model.
21. The data processing system according to claim 20, characterized in that If the model training instance includes two or more than two, the model training submodule is specifically used for: Determining the parallelism of the model training instances; Parallel model training is performed based on the parallelism of the model training instance, the data set and the multiple model training instances to generate multiple gain models.
22. The data processing system according to claim 13, characterized in that: The system further comprises: A prediction module, used to use the gain model to predict the data set to obtain a gain value corresponding to a target intervention value; An evaluation module is used to evaluate the actual values of the target column and the intervention column of the data set and the gain value based on the evaluation index to obtain a gain evaluation result.
23. The data processing system according to any one of claims 13 to 22, characterized in that: The system further comprises: The first display module is used to sort the multiple historical model training information according to the evaluation index and display the sorted multiple historical model training information.
24. The data processing system according to claim 23, characterized in that: The system further comprises: The second display module is used to display the training information corresponding to the target historical model based on the user's selection operation on the target historical model, and the training information includes model scoring, scoring curve chart, model process and parameter combination of each stage.
25. A data processing system, characterized in that: The method comprises a processor, a memory and a computer program stored in the memory and executable on the processor, wherein the computer program implements the steps of the data processing method according to any one of claims 1 to 12 when executed by the processor.
26. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the data processing method according to any one of claims 1 to 12 are implemented.
