CN114943607A - Feature discovery method, attribute prediction method and device - Google Patents

Abstract

The embodiment of the specification describes a feature discovery method, an attribute prediction method and a device. According to the feature discovery method of the embodiment, the original features generated in the service operation and used for predicting the attributes contained in the operated service are obtained firstly. And then, after the original characteristics are derived and the quality is evaluated, the attribute prediction characteristics for predicting the attributes contained in the operation service can be obtained. Because the scheme is obtained by deriving the existing characteristics, a large amount of time is not needed to be spent on analyzing and summarizing the original data, and the efficiency of characteristic discovery can be greatly improved. And by carrying out quality evaluation on the derived features, the attribute prediction features with better performance can be obtained by screening according to the importance degree of the derived features to attribute prediction, so that the accuracy of attribute prediction is improved.

Description

Feature discovery method, attribute prediction method and device

technical field

One or more embodiments of this specification relate to the field of computer technology, and in particular, to a feature discovery method, an attribute prediction method, and an apparatus.

Background technique

Feature discovery is an important means of attribute prediction in business scenarios. For example, in the field of risk prevention and control, with the increasing frequency of payment risk control scenarios and the attack and defense of black and gray products, risk strategists need to spend a lot of time analyzing cases and obtaining the entire process of black and gray products committing crimes. After data analysis, the risk characteristics are summarized, and the risk characteristics are abstracted into specific risk characteristics, which are applied to risk prevention and control.

However, the efficiency of feature discovery based on manual work is low.

SUMMARY OF THE INVENTION

One or more embodiments of this specification describe a feature discovery method, an attribute prediction method, and an apparatus, which can improve the efficiency of feature extraction.

According to the first aspect, a feature discovery method is provided, comprising:

Obtain at least one original feature generated in the business operation; wherein, the original feature is used to predict the attributes included in the operated business;

Derivating the at least one original feature to obtain at least one derived feature;

Evaluate the quality of the at least one derived feature to obtain an attribute prediction feature for predicting an attribute included in the operating business; wherein the quality of the derived feature is used to characterize the feature to predict the attribute degree of importance.

In a possible implementation, the at least one original feature includes at least two original features;

The at least one derived feature obtained by deriving the at least one original feature includes:

Screening the at least two original features to obtain at least one screening feature according to the degree of coverage of the attribute by each of the original features;

The at least one screening feature is mapped to a dimension different from the dimension in which the current screening feature is located to obtain the at least one derived feature.

In a possible implementation manner, the at least one screening feature is obtained by screening the at least two original features according to the degree of coverage of the attribute by each of the original features, including:

Among the at least two original features, an original feature whose feature value is greater than a preset first effective feature threshold is determined as the screening feature.

In a possible implementation, the at least one original feature includes at least two original features;

The at least one derived feature obtained by deriving the at least one original feature includes:

For any one of the first original features and any one of the second original features in the at least two original features, perform:

Splitting the first original feature to obtain M first splitting features; and,

Splitting the second original feature to obtain N second splitting features; wherein, M and N are both integers greater than 0;

Combining the M first splitting features and the N second splitting features to obtain M×N primary derived features; wherein, the primary derived features obtained by any combination are composed of one first splitting feature Combined with a second split feature;

The at least one derived feature is obtained by screening the M×N primary derived features according to the degree of coverage of the attribute by each primary derived feature.

In a possible implementation manner, the at least one derived feature is obtained by filtering the M×N primary derived features according to the coverage degree of each primary derived feature to the attribute, including:

Among the M×N primary derivative features, a primary derivative feature whose combined feature value is greater than a preset second effective feature threshold is determined as the derivative feature; wherein, the combined feature value is combined into a corresponding primary derivative feature. The value obtained by the weighted average of the feature values of the first split feature and the second split feature.

In a possible implementation manner, the attribute prediction feature obtained by evaluating the quality of the at least one derived feature includes:

calculating the attribute prediction ability value of each of the derived features;

Among the attribute prediction ability values, a derivative feature corresponding to an attribute prediction ability value greater than a preset evaluation threshold is determined as the attribute prediction feature.

In a possible implementation manner, the attribute includes a first attribute result to be predicted, and a second attribute result opposite to the first attribute result;

The calculating the attribute prediction ability value of each of the derived features includes:

For each of the derived features, execute:

The current derived features are grouped equidistantly to obtain k groups; wherein, the equidistance includes at least one of equal height and equal width; k is an integer greater than 0;

Use the following formula to calculate the attribute prediction ability value corresponding to each group:

Among them, IV _i is used to represent the attribute prediction ability value corresponding to the ith group, y _i is used to represent the number of the ith group containing the first attribute result, y _s is used to represent the current derived feature contains the said th A number of attribute results, x _i is used to represent the number of the i-th packet containing the second attribute result, x _s is used to represent the current derived feature contains the number of the second attribute result;

The attribute prediction ability values obtained by each grouping are summed to obtain the attribute prediction ability value of the current derived feature.

In a possible implementation manner, the attribute prediction feature obtained by evaluating the quality of the at least one derived feature includes:

The attribute prediction feature is determined according to the accuracy of attribute prediction using the original feature and the derived feature respectively.

In a possible implementation manner, the determining the attribute prediction feature according to the accuracy of the attribute prediction using the original feature and the derived feature respectively includes:

Using the at least two original features, a first attribute prediction model is obtained by training;

Using at least one of the derived features, training to obtain a second attribute prediction model;

Using the first attribute prediction model and the second attribute prediction model to perform attribute prediction on the same attribute prediction feature to obtain a first prediction result corresponding to the first attribute prediction model and a corresponding second attribute prediction model The second prediction result of ;

Calculate the distance between the first prediction result and the label value of the attribute to be predicted feature to obtain a first similarity value; and,

Calculate the distance between the second prediction result and the label value of the attribute to be predicted feature to obtain a second similarity value;

When the second similarity value is smaller than the first similarity value, the derived feature used for training the second attribute prediction model is determined as the attribute prediction feature.

According to the second aspect, an attribute prediction method is provided, including:

Acquiring at least two attribute prediction features obtained based on any one of the feature discovery methods described in the first aspect for predicting attributes included in the operated business, and quality evaluation results of the at least two attribute prediction features;

dividing the at least two attribute prediction features into at least two levels according to the quality evaluation results of the attribute prediction features; and,

The attribute prediction features of the at least two levels are deployed in different services, so as to predict the attributes in different services; wherein, the higher the importance of the services, the higher the level of the deployed attribute prediction features. high.

According to a third aspect, a feature discovery device is provided, including: a feature acquisition module, a feature derivation module, and a feature evaluation module;

The feature acquisition module is configured to acquire at least one original feature generated in the business operation; wherein, the original feature is used to predict the attributes included in the operated business;

The feature derivation module is configured to derive at least one original feature acquired by the feature acquisition module to obtain at least one derived feature;

The feature evaluation module is configured to evaluate the quality of the at least one derived feature obtained by the feature derivation module, and obtain an attribute prediction feature for predicting attributes included in the operating business; wherein the The quality of a derived feature is used to characterize how important that feature is to predict the attribute.

According to a fourth aspect, an attribute prediction apparatus is provided, including: a prediction data acquisition module, a grade division module, and an attribute prediction module;

The predicted data obtaining module is configured to obtain at least two attribute prediction features obtained based on the feature discovery device described in the third aspect for predicting the attributes included in the operating business, and the at least two attribute prediction features. quality assessment results;

The grade division module is configured to divide the at least two attribute prediction features into at least two grades according to the quality evaluation result of the attribute prediction feature obtained by the prediction data acquisition module;

The attribute prediction module is configured to deploy the attribute prediction features of the at least two grades divided by the grade division module into different services, so as to realize the attribute prediction in different services; wherein, the The higher the importance of the business, the higher the level of the deployed attribute prediction feature.

According to a fifth aspect, a computing device is provided, including: a memory and a processor, where executable code is stored in the memory, and when the processor executes the executable code, the above-mentioned first and second aspects are implemented The method of any of the above.

According to the method and apparatus provided by the embodiments of the present specification, when performing feature discovery, the original feature generated in the business operation for predicting the attributes included in the operated business can be obtained first, and then the original feature can be derived to obtain at least one derived features. Further, the attribute prediction feature included in the operation business for predicting the attribute can be obtained by performing quality evaluation on the obtained derivative feature. It can be seen that this solution is obtained by deriving the existing features, and does not need to spend a lot of time analyzing and summarizing the original data, so the efficiency of feature discovery can be greatly improved.

Description of drawings

In order to more clearly illustrate the technical solutions in the embodiments of the present specification or in the prior art, the following briefly introduces the accompanying drawings that need to be used in the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are For some embodiments of this specification, for those of ordinary skill in the art, other drawings can also be obtained based on these drawings without any creative effort.

1 is a flowchart of a feature discovery method provided by an embodiment of the present specification;

FIG. 2 is a flowchart of a feature derivation method provided by an embodiment of the present specification;

3 is a flowchart of another feature derivation method provided by an embodiment of the present specification;

4 is a flowchart of a feature evaluation method provided by an embodiment of the present specification;

5 is a flowchart of a method for calculating a feature prediction capability value provided by an embodiment of the present specification;

6 is a flowchart of another feature evaluation method provided by an embodiment of the present specification;

FIG. 7 is a flowchart of an attribute prediction method provided by an embodiment of the present specification;

FIG. 8 is a schematic diagram of a feature discovery apparatus provided by an embodiment of the present specification;

FIG. 9 is a schematic diagram of an attribute prediction apparatus provided by an embodiment of the present specification.

Detailed ways

As mentioned above, feature discovery is an important means for attribute prediction in business scenarios. However, attribute prediction needs to analyze and summarize the original data to obtain features that can predict attributes. For example, in the field of risk prevention and control, with the increasing frequency of payment risk control scenarios and attacks and defenses of black and gray products, risk strategists can summarize the risk characteristics by analyzing cases, obtaining the entire process of black and gray products committing crimes, and analyzing data. The risk characteristics are abstracted into specific risk characteristics, and then the risk account can be predicted based on the risk characteristics.

However, the method that relies on operational strategists to mine the features of attribute prediction takes a lot of time and has low output efficiency. Moreover, the generation of features will directly affect the prediction of attributes, as well as the quality of services. If the experience of operation strategists is inconsistent, the features and evaluation criteria for attribute identification will be inconsistent, and it is difficult to standardize the mining of features, resulting in low accuracy of attribute prediction.

Based on this, this scheme considers the use of existing features for feature derivation, and quickly and efficiently derives features with better performance and higher risk coverage. This can not only improve the efficiency of feature discovery, but also use the derived features to improve the coverage of attributes related to attributes, thereby improving the accuracy of attribute prediction.

As shown in FIG. 1 , an embodiment of the present specification provides a feature discovery method, and the method may include the following steps:

Step 101: Obtain at least one original feature generated in the business operation; wherein, the original feature is used to predict the attributes included in the operated business;

Step 103: Derive at least one original feature to obtain at least one derived feature;

Step 105: Evaluate the quality of at least one derived feature to obtain an attribute prediction feature for predicting an attribute included in the operating business; wherein the quality of the derived feature is used to represent the importance of the feature for predicting the attribute.

In the embodiment of the present specification, when performing feature discovery, the original feature generated in the business operation for predicting the attributes included in the operated business can be obtained first, and then the original feature is derived to obtain at least one derived feature. Further, the attribute prediction feature for predicting the included attributes in the operation business can be obtained by performing quality evaluation on the obtained derivative features. It can be seen that this solution is obtained by deriving existing features, and does not need to spend a lot of time analyzing and summarizing the original data, so the efficiency of feature discovery can be greatly improved.

Moreover, after the derived feature is obtained by deriving the original feature, the quality of the derived feature is further evaluated, so that the attribute prediction feature with better performance can be obtained according to the importance of the derived feature to predict the attribute. The attribute prediction feature can improve the accuracy of attribute prediction when making predictions.

Each step in FIG. 1 will be described below with reference to specific embodiments.

First, in step 101, at least one original feature generated in the business operation is acquired;

In this step, the original feature may be a feature used to predict attributes included in the operating business. For example, business operations can be merchant access risk management, fraud risk management, and online business risk management. The original features generated in business operations can be the features obtained by strategic operators in the daily risk attack and defense process, based on link restoration of risk cases and statistical analysis of historical risk samples. If the attribute prediction is to predict the risk account, the original feature may include the user's age, gender, credit degree, transaction amount, transaction time and other features.

In addition, the original feature may also be a feature obtained by using the feature discovery method of the present solution in a certain period of time before. That is to say, the method provided by this solution can perform feature discovery by using the features derived in the previous time period as the original features in the next time period.

Of course, in a possible implementation manner, the original features may include both the features obtained by the operation strategist through analysis and summary, and the features derived from the features in the previous time period. Such derived features not only improve the efficiency of feature discovery, but also the obtained features contain the results of manual analysis and summary to a certain extent, which can improve the performance of features for attribute prediction.

Then in step 103, at least one original feature is derived to obtain at least one derived feature.

In this step, after the original features are obtained, it is considered to derive the original features by means of spatial transformation and feature combination.

For example, when spatially transforming derived features through a lightweight model, the original features may include at least two. As shown in FIG. 2, when at least one derived feature is obtained by deriving at least one original feature in step 103, the following steps may be performed:

Step 201: Screen at least two original features to obtain at least one screening feature according to the coverage degree of each original feature to the attribute;

Step 203: Map the at least one screening feature to a dimension different from the dimension in which the current screening feature is located to obtain at least one derived feature.

In this embodiment, it may be considered to derive features by means of a lightweight model. Specifically, the original features can be firstly screened according to the coverage degree of the original features to the attributes to obtain the screening features. Then, the obtained screening feature is mapped to a dimension different from the dimension in which the current screening feature is located to obtain a derived feature. The features derived in this way can have good attribute coverage and can be well used to predict attributes.

For example, when performing feature derivation in steps 201 and 203, linear or nonlinear machine learning models, such as algorithms such as Logistic Regression and Decision Tree, can be used to process and refine the input original features to generate features with better performance , to characterize the feature information. Moreover, this method does not require the type of input data, which can be numeric or character, which can be well adapted to application scenarios of various data types.

Step 201 will be described below.

In step 201, when at least two original features are screened to obtain at least one screening feature according to the coverage degree of each original feature to the attribute, at least two original features whose feature value is greater than the preset first effective feature threshold may be selected. The original feature is determined to be the screening feature. In this way, by filtering the original features according to the eigenvalues, it is possible to filter out the data with low information intensity representing attributes, and use the data with high feature representation intensity to generate derived features, which is more conducive to improving the accuracy of attribute prediction.

For example, in practical application scenarios, since many features are not effective features, the feature value they usually reflect is 0. However, if features containing these invalid features are used for feature discovery, it will not only reduce the processing efficiency due to the large amount of data, but also reduce the quality of the derived features. Therefore, when the original features are filtered, the first effective feature threshold may be 0. In this way, all features with eigenvalues of 0 are filtered out, that is, all invalid features are filtered out, so that effective features can be used for feature derivation more efficiently.

Step 203 will be described below.

In step 203, when at least one derived feature is obtained by mapping the at least one screening feature to a dimension different from the dimension in which the current screening feature is located, the at least one derived feature may be obtained by linearly changing the at least one screening feature. By setting an appropriate linear variation coefficient in this way, features containing various levels of information can be derived.

For example, in a possible implementation manner, the filtering features can be mapped to other dimensions by the following calculation formula.

用于表征映射得到的衍生特征，z_v,t用于表征筛选特征，ω^G用于表征对筛选特征进行映射的映射权重，b^G用于表征对筛选特征进行映射的映射修正量。在上述计算式中，ω^G和b^G参数可以是学习参数，可以通过神经网络学习得到。而各个参数的维度大小可以分别为：ω^G∈R^C，b^G∈R^C。in,

It is used to characterize the derived features obtained by mapping, z _{v, t} is used to characterize the screening features, ω ^G is used to characterize the mapping weight for mapping the screening features, and b ^G is used to characterize the mapping correction amount for mapping the screening features. In the above calculation formula, the ω ^G and b ^G parameters can be learning parameters, which can be obtained through neural network learning. The dimensions of each parameter can be respectively: ω ^G ∈ R ^C , b ^G ∈ R ^C .

In another possible implementation manner, when a feature is derived by means of feature combination, the original feature may include at least two. As shown in FIG. 3 , when at least one derived feature is derived from at least one original feature in step 103, the following steps may be performed for any one of the first original features and any one of the second original features among the at least two original features:

Step 301: Split the first original feature to obtain M first split features;

Step 303: Splitting the second original feature to obtain N second splitting features; wherein, M and N are both integers greater than 0;

Step 305: Combine the M first splitting features and the N second splitting features to obtain M×N primary derived features; wherein, the primary derived features obtained by any combination are composed of a first splitting feature and A second split feature combination is obtained;

Step 307 : Screen M×N primary derivative features to obtain at least one derivative feature according to the coverage degree of each primary derivative feature to the attribute.

In this embodiment, it may be considered to generate derived features by combining features. Specifically, the first original feature may be split to obtain M first split features, and then the second original feature may be split to obtain N second split features. The M first split features and the N second split features are further combined to obtain M×N primary derived features. Finally, the derived features can be obtained by feature screening according to the coverage degree of the attributes of each primary derived feature. In this way, by combining the original features in pairs, features with a higher degree of discrimination can be found.

For example, the first original feature is age, and the second original feature is gender. When splitting the first original feature in step 301, age 0-100 can be split into five, such as 0-20, 21-40 , 41-60, 61-80, 81-100; in step 302, when splitting the second original feature, two values of male and female can be taken. In this way, by combining the split age features and gender features in pairs, 2×5=10 features can be obtained. Males aged 0-20, males 21-40, males 41-60, males 61-80, males 81-100, females 0-20, females 21-40, 41 -60 women, 61-80 women, and 81-100 women. In this way, when further screening according to the attribute coverage degree, the combined features can be screened according to specific feature values.

The eigenvalues obtained by the pairwise combination may be obtained by weighted average, that is, the combined eigenvalue of each derived feature obtained after the combination may be the eigenvalue of the first split feature and the second split feature corresponding to the derived feature. Values obtained by weighted average. In step 307, when at least one derivative feature is obtained by screening the M×N primary derivative features according to the coverage degree of each primary derivative feature to the attribute, the combined feature value of the M×N primary derivative features may be greater than the preset value. The primary derived feature of the second valid feature threshold is determined as a derived feature. In this way, after obtaining the primary derived features through combination, further screening according to the feature value can obtain features with higher feature representation strength, that is, features with higher attribute concentration, so that the obtained derived features have better performance in attribute prediction.

Further in step 105, the quality of at least one derived feature is evaluated to obtain an attribute prediction feature for predicting an attribute included in the operating business.

In this step, when evaluating the quality of the derived feature, it mainly evaluates the importance of the derived feature for predicting the attribute, that is, evaluating the predictive ability of the derived feature. For example, in one possible implementation, the evaluation may be performed by calculating a predictive ability value. As shown in Figure 4, the derived features can be evaluated by the following steps:

Step 401: Calculate the attribute prediction ability value of each derived feature;

Step 403 : Determining, among the attribute prediction ability values, the derivative feature corresponding to the attribute prediction ability value greater than the preset evaluation threshold as the attribute prediction feature.

In this embodiment, the attribute prediction ability value of each derivative feature can be calculated first, and then the derivative feature corresponding to the attribute prediction ability value greater than the preset evaluation threshold among the attribute prediction ability values is determined as the attribute prediction feature. In this way, by accurately calculating the attribute prediction ability value of the derived features, the derived features with better performance can be screened out, thereby improving the accuracy of attribute prediction.

In a possible implementation manner, the attribute may include a first attribute result to be predicted, and a second attribute result opposite to the first attribute result. Then, when calculating the attribute prediction ability value in step 401, as shown in FIG. 5, the attribute prediction ability value of each derived feature can be obtained through the following steps:

For each derived feature,

Step 501: The current derived features are grouped equidistantly to obtain k groups; wherein, the equidistance includes at least one of equal height and equal width; k is an integer greater than 0;

Step 503: Calculate the attribute prediction ability value corresponding to each group by using the calculation formula of the attribute prediction ability;

Among them, the calculation formula of attribute prediction ability value is:

Among them, IV _i is used to characterize the attribute prediction ability value corresponding to the ith group, y _i is used to characterize the number of the first attribute result contained in the ith group, and y _s is used to characterize the current derived feature that contains the first attribute result. Quantity, x _i is used to represent the number of results of the second attribute contained in the ith group, and x _s is used to represent the number of results of the second attribute contained in the current derived feature;

Step 505: Summing the attribute prediction ability values obtained by each group to obtain the attribute prediction ability value of the current derived feature.

In this embodiment, when calculating the attribute prediction ability value of each derived feature, firstly, the current derived features can be grouped equidistantly by means of equal height and equal width, and then the corresponding attribute prediction ability value of each grouping can be calculated by using the above attribute prediction ability value calculation. The attribute prediction ability value of . Finally, the attribute prediction ability value of the current derived feature can be obtained by summing the attribute prediction ability values corresponding to each group.

The above attribute prediction ability value measures the amount of information of a variable, and its value determines the degree of influence of the independent variable on the target variable. For example, in an application scenario in which a user performs information feedback, the first attribute result may be the customer who responded, and the second attribute result may be the customer who did not respond. Then the meaning of y _i /y _s is the proportion of the customers who responded in the current i-th group to all the responding customers in the k groups of the entire derived feature, and the meaning of x _i /x _s is the current i-th group. The proportion of customers who did not respond in k groupings of the entire derived feature.

It is worth noting that in any grouping of derived features, there should not be a situation where the number of first attribute results is 0 or the number of second attribute results is 0, when the number of attribute results of a grouping of derived features is 0 , the corresponding ln value is negative infinity, and the IV value is positive infinity. If possible, make this grouping directly into a rule as a precondition or supplement to the model.

In the modeling process, the IV value is mainly used for feature selection. If you want to rank the predictive ability of variables, since the size of the IV value of the feature variable indicates the strength of the variable's predictive ability, you can choose the IV value from high to high. to low filter.

In another possible implementation manner, step 105 may also perform evaluation based on a machine learning algorithm of a distributed gradient decision tree, that is, determine the attribute prediction feature according to the accuracy of attribute prediction using the original feature and the derived feature respectively. For example, as shown in Figure 6, the following steps may be specifically included when determining the attribute prediction feature:

Step 601: Use at least one of the original features to train to obtain a first attribute prediction model;

Step 603: Use at least one of the derived features to train to obtain a second attribute prediction model;

Step 605: Use the first attribute prediction model and the second attribute prediction model to perform attribute prediction on the same attribute prediction feature to obtain a first prediction result corresponding to the first attribute prediction model and a second prediction result corresponding to the second attribute prediction model ;

Step 607: Calculate the distance between the first prediction result and the label value of the attribute to be predicted feature to obtain the first similarity value;

Step 609: Calculate the distance between the second prediction result and the label value of the feature to be predicted for the attribute to obtain a second similarity value;

Step 611: When the second similarity value is smaller than the first similarity value, determine the derived feature used for training the second attribute prediction model as the attribute prediction feature.

In this embodiment, when determining the attribute feature according to the accuracy of the attribute prediction based on the original feature and the derived feature, firstly, the original feature and the derived feature can be used to train the model to obtain the first attribute prediction model and the second attribute prediction model respectively. Then use the first attribute prediction model and the second attribute prediction model to perform attribute prediction on the same attribute to be predicted feature to obtain a first prediction result and a second prediction result. Further, compare the distance between the two prediction results and the actual label value. If the distance between the prediction result obtained by the derived feature and the label value is smaller than the distance between the prediction result obtained by the original feature and the label value, it means that the derived feature has better properties. The effect is predicted, so that the derived feature trained on the second attribute prediction model can be determined as the attribute prediction feature. In this way, the attribute prediction features obtained each time are screened out through the actual prediction, which can have better performance.

In step 601, when using the original features to train the first attribute prediction model, at least one feature may be randomly selected from the original features for training. Similarly, when using the derived features to train the second attribute prediction model in step 603, at least one feature may be randomly selected from the derived features for training. Of course, different numbers of features can be used for model training to further filter attribute features. For example, for the original features A and B, the derived features are C, D, E, and F, where the first attribute prediction model is trained using A and B. For the second attribute prediction model, C and D can be used for training first. If it is obtained through comparison that the performance of the second attribute prediction model is better than the performance of the first attribute prediction model, then C and D can be determined as attribute prediction features. If the second attribute prediction model is obtained by further training with C, D and E, but the performance of the second attribute prediction model has declined at this time, then filter the derived feature C; and if the performance of the second attribute prediction model has improved, then The derived feature C can also be determined as an attribute prediction feature.

Among them, when determining the performance of the model, it can be obtained by predicting the same attribute to be predicted. The closer the distance to the actual label, the better the performance of the model; conversely, the farther the distance from the actual label, the worse the performance of the model .

In yet another possible implementation manner, step 105 may further evaluate the quality of the derived features through a heat map. For example, in step 103, after the features are crossed in pairs by the feature combination method, a table similar to the following Table 1 can be formed for the feature x and the feature y. Then count the data distribution in each unit, and finally give each unit a score, the larger the score, the better the performance of the cross-combined features. Through the following heat map, the performance of each feature can be found more intuitively, so that the derived features can be directly screened to obtain attribute prediction features.

Table 1

It can be seen from the above Table 1 that if the set threshold is 1, then the filtered features [feature x, feature y, value] are [1, 1, 1.3], [3, 2, 1.5], [4, 2] , 2.3], [4, 3, 1.5], [3, 4, 1.5], [3, 5, 1.4]. The attribute feature that can be obtained in this way is the feature obtained by combining the feature x and the feature y among the six.

As shown in FIG. 7 , an embodiment of the present specification further provides an attribute prediction method, and the method may include the following steps:

Step 701: Acquire at least two attribute prediction features obtained based on the feature discovery method of each of the foregoing embodiments for predicting attributes included in the operating business, and quality evaluation results of the at least two attribute prediction features;

Step 703: Divide the at least two attribute prediction features into at least two levels according to the quality evaluation results of the attribute prediction features; and,

Step 705: Deploy at least two levels of attribute prediction features into different services to implement attribute prediction in different services; wherein, the higher the importance of the service, the higher the level of the deployed attribute prediction features.

In this embodiment, when performing attribute prediction, at least two attribute prediction features obtained by using the feature discovery method provided by each embodiment of this specification for predicting the attributes included in the operating business, and the at least two attribute prediction features may be obtained first. The quality assessment results of each attribute prediction feature. Then, according to the quality evaluation result of the attribute prediction feature, the at least two attribute prediction features are divided into at least two levels, and the attribute prediction features of different levels are deployed in different services, so as to realize the attribute prediction in different services.

For example, in the prediction of the risk account, after the attribute prediction feature is obtained, each attribute feature can be divided into different risk levels according to the feature value. For example, attribute prediction features obtained by cross-combination of user credit and gender. Men with 0-20 credit, men with 21-40 credit, men with 41-60 credit, men with 61-80 credit, men with 81-100 credit, women with 0-20 credit, women with 21-40 credit, Women with 41-60 credit, women with 61-80 credit, and women with 81-100 credit. In this way, different features can be deployed in each scenario, for users with low credit degrees, they can be deployed in business scenarios that they focus on, and users with high credit degrees can be deployed in business scenarios with relatively low degrees of concern.

As shown in FIG. 8 , an embodiment of this specification provides a feature discovery apparatus, including: a feature acquisition module 801, a feature derivation module 802, and a feature evaluation module 803;

The feature acquisition module 801 is configured to acquire at least one original feature generated in the business operation; wherein, the original feature is used to predict the attributes included in the operated business;

The feature derivation module 802 is configured to derive at least one original feature acquired by the feature acquisition module 801 to obtain at least one derived feature;

The feature evaluation module 803 is configured to evaluate the quality of at least one derived feature obtained by the feature derivation module 802, and obtain an attribute prediction feature for predicting the attributes included in the operating business; wherein, the quality of the derived feature is used to characterize the How important the feature is to predict the attribute.

In a possible implementation manner, the at least one original feature obtained by the feature obtaining module 801 includes at least two original features;

When the feature derivation module 802 derives at least one original feature to obtain at least one derived feature, it is configured to perform the following operations:

According to the coverage degree of each original feature to the attribute, at least two original features are screened to obtain at least one screening feature;

The at least one screening feature is mapped to a dimension different from the dimension in which the current screening feature is located to obtain at least one derived feature.

In a possible implementation manner, the feature derivation module 802 is configured to perform the following operations when at least one screening feature is obtained by screening at least two original features according to the coverage degree of each original feature on the attribute:

Among the at least two original features, the original feature whose feature value is greater than the preset first effective feature threshold is determined as the screening feature.

In a possible implementation manner, the at least one original feature obtained by the feature obtaining module 801 includes at least two original features;

When the feature derivation module 802 derives at least one original feature to obtain at least one derived feature, it is configured to perform the following operations:

For any first original feature and any second original feature of the at least two original features, execute:

splitting the first original feature to obtain M first split features; and,

Splitting the second original feature to obtain N second splitting features; wherein, M and N are both integers greater than 0;

Combine M first split features and N second split features to obtain M×N primary derived features; wherein, the primary derived features obtained by any combination are composed of a first split feature and a second split feature. The combination of splitting features is obtained;

According to the coverage degree of each primary derived feature to the attribute, at least one derived feature is obtained by screening M×N primary derived features.

In a possible implementation manner, the feature derivation module 802 is configured to perform the following operations when at least one derived feature is obtained by filtering M×N primary derived features according to the coverage degree of each primary derived feature to the attribute:

Among the M×N primary derivative features, the primary derivative feature whose combined feature value is greater than the preset second effective feature threshold is determined as a derivative feature; wherein, the combined feature value is combined into a first split feature and a corresponding primary derivative feature. The eigenvalues of the second split feature are obtained by weighted averaging.

In a possible implementation manner, the feature evaluation module 803 is configured to perform the following operations when the attribute prediction feature is obtained by evaluating the quality of the at least one derived feature:

Calculate the attribute prediction ability value of each derived feature;

Among the attribute prediction ability values, the derivative feature corresponding to the attribute prediction ability value greater than the preset evaluation threshold is determined as the attribute prediction feature.

In a possible implementation manner, the attribute includes a first attribute result to be predicted, and a second attribute result opposite to the first attribute result;

When the feature evaluation module 803 calculates the attribute prediction ability value of each derived feature, it is configured to perform the following operations:

For each derived feature, execute:

The current derived features are grouped at equal distances to obtain k groups; wherein, the equal distance includes at least one of equal height and equal width; k is an integer greater than 0;

Use the following formula to calculate the attribute prediction ability value corresponding to each group:

Among them, IV _i is used to characterize the attribute prediction ability value corresponding to the ith group, y _i is used to characterize the number of the first attribute result contained in the ith group, and y _s is used to characterize the current derived feature that contains the first attribute result. Quantity, x _i is used to represent the number of results of the second attribute contained in the ith group, and x _s is used to represent the number of results of the second attribute contained in the current derived feature;

The attribute prediction ability values obtained by each group are summed to obtain the attribute prediction ability value of the current derived feature.

In a possible implementation manner, the feature evaluation module 803 is configured to perform the following operations when the attribute prediction feature is obtained by evaluating the quality of the at least one derived feature:

Attribute prediction features are determined according to the accuracy of attribute prediction using original features and derived features respectively.

In a possible implementation manner, the feature evaluation module 803 is configured to perform the following operations when determining the attribute prediction feature according to the accuracy of the attribute prediction using the original feature and the derived feature respectively:

Using at least one of the original features, training to obtain a first attribute prediction model;

Using at least one of the derived features, training to obtain a second attribute prediction model;

Using the first attribute prediction model and the second attribute prediction model to perform attribute prediction on the same attribute prediction feature to obtain a first prediction result corresponding to the first attribute prediction model and a second prediction result corresponding to the second attribute prediction model;

Calculate the distance between the first prediction result and the label value of the feature to be predicted by the attribute to obtain the first similarity value; and,

Calculate the distance between the second prediction result and the label value of the attribute to be predicted feature to obtain the second similarity value;

When the second similarity value is smaller than the first similarity value, the derived feature used for training the second attribute prediction model is determined as the attribute prediction feature.

As shown in FIG. 9 , the embodiment of the present specification further provides an attribute prediction apparatus, including: a prediction data acquisition module 901 , a grade division module 902 and an attribute prediction module 903 ;

The prediction data acquisition module 901 is configured to acquire at least two attribute prediction features obtained by the feature discovery device of each of the above embodiments for predicting the attributes included in the operating business, and the quality evaluation results of the at least two attribute prediction features ;

The grade division module 902 is configured to divide the at least two attribute prediction features into at least two grades according to the quality evaluation result of the attribute prediction feature obtained by the prediction data acquisition module 901;

The attribute prediction module 903 is configured to deploy the attribute prediction features of at least two grades divided by the grade division module 902 into different services, so as to predict attributes in different services; The attribute of predicts the higher the rank of the feature.

The present specification also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed in the computer, the computer is made to execute the method in any one of the embodiments in the specification.

The present specification also provides a computing device including a memory and a processor, where executable codes are stored in the memory, and when the processor executes the executable codes, the method in any one of the embodiments in the specification is implemented.

It can be understood that the structures illustrated in the embodiments of the present specification do not constitute specific limitations on the feature discovery apparatus and the attribute prediction apparatus. In other embodiments of the specification, the feature discovery apparatus and the attribute prediction apparatus may include more or less components than shown, or combine some components, or separate some components, or different component arrangements. The illustrated components may be implemented in hardware, software, or a combination of software and hardware.

Since the information exchange and execution process among the units in the above apparatus are based on the same concept as the method embodiments in this specification, the specific content can be found in the descriptions in the method embodiments in this specification, which will not be repeated here.

Those skilled in the art should realize that, in one or more of the above examples, the functions described in this specification can be implemented by hardware, software, accessories or any combination thereof. When implemented in software, the functions may be stored on or transmitted over as one or more instructions or code on a computer-readable medium.

The specific embodiments described above further describe the purposes, technical solutions and beneficial effects described in this specification. It should be understood that the above descriptions are only specific embodiments of the present invention, and are not intended to limit the present invention. Any modification, equivalent replacement, improvement, etc. made on the basis of the technical solution of the present invention shall be included within the protection scope of the present invention.

Claims (13)

1. Feature discovery methods, including: Obtain at least one original feature generated in the business operation; wherein, the original feature is used to predict the attributes included in the operated business; Derivating the at least one original feature to obtain at least one derived feature; Evaluate the quality of the at least one derived feature to obtain an attribute prediction feature for predicting an attribute included in the operating business; wherein the quality of the derived feature is used to characterize the feature to predict the attribute degree of importance. 2. The method of claim 1, wherein the at least one original feature comprises at least two original features; The at least one derived feature obtained by deriving the at least one original feature includes: Screening the at least two original features to obtain at least one screening feature according to the degree of coverage of the attribute by each of the original features; The at least one screening feature is mapped to a dimension different from the dimension in which the current screening feature is located to obtain the at least one derived feature. 3. The method according to claim 2, wherein the at least one screening feature is obtained by screening the at least two original features according to the coverage degree of each of the original features on the attribute, comprising: Among the at least two original features, an original feature whose feature value is greater than a preset first effective feature threshold is determined as the screening feature. 4. The method of claim 1, wherein the at least one original feature comprises at least two original features; The at least one derived feature obtained by deriving the at least one original feature includes: For any one of the first original features and any one of the second original features in the at least two original features, perform: Splitting the first original feature to obtain M first splitting features; and, Splitting the second original feature to obtain N second splitting features; wherein, M and N are both integers greater than 0; Combining the M first splitting features and the N second splitting features to obtain M×N primary derived features; wherein, the primary derived features obtained by any combination are composed of one first splitting feature Combined with a second split feature; The at least one derived feature is obtained by screening the M×N primary derived features according to the degree of coverage of the attribute by each primary derived feature. 5 . The method according to claim 4 , wherein the at least one derived feature is obtained by filtering the M×N primary derived features according to the coverage degree of each primary derived feature on the attribute, comprising: 6 . Among the M×N primary derivative features, a primary derivative feature whose combined feature value is greater than a preset second effective feature threshold is determined as the derivative feature; wherein, the combined feature value is combined into a corresponding primary derivative feature. The value obtained by the weighted average of the feature values of the first split feature and the second split feature. 6. The method according to claim 1, wherein the property prediction feature obtained by evaluating the quality of the at least one derived feature comprises: calculating the attribute prediction ability value of each of the derived features; Among the attribute prediction ability values, a derivative feature corresponding to an attribute prediction ability value greater than a preset evaluation threshold is determined as the attribute prediction feature. 7. The method of claim 6, wherein the attribute comprises a first attribute result to be predicted, and a second attribute result opposite to the first attribute result; The calculating the attribute prediction ability value of each of the derived features includes: For each of the derived features, execute: The current derived features are grouped equidistantly to obtain k groups; wherein, the equidistance includes at least one of equal height and equal width; k is an integer greater than 0; Use the following formula to calculate the attribute prediction ability value corresponding to each group:

Among them, IV _i is used to represent the attribute prediction ability value corresponding to the ith group, y _i is used to represent the number of the ith group containing the first attribute result, y _s is used to represent the current derived feature contains the said th A number of attribute results, x _i is used to represent the number of the i-th packet containing the second attribute result, and x _s is used to represent the current derived feature contains the number of the second attribute result; The attribute prediction ability values obtained by each grouping are summed to obtain the attribute prediction ability value of the current derived feature. 8. The method according to claim 1, wherein the property prediction feature obtained by evaluating the quality of the at least one derived feature comprises: The attribute prediction feature is determined according to the accuracy of attribute prediction using the original feature and the derived feature respectively. 9. The method according to claim 8, wherein the determining the attribute prediction feature according to the accuracy of the attribute prediction using the original feature and the derived feature respectively comprises: Using at least one of the original features, training to obtain a first attribute prediction model; Using at least one of the derived features, training to obtain a second attribute prediction model; Using the first attribute prediction model and the second attribute prediction model to perform attribute prediction on the same attribute prediction feature to obtain a first prediction result corresponding to the first attribute prediction model and a corresponding second attribute prediction model The second prediction result of ; Calculate the distance between the first prediction result and the label value of the attribute to be predicted feature to obtain the first similarity value; and, Calculate the distance between the second prediction result and the label value of the attribute to be predicted feature to obtain a second similarity value; When the second similarity value is smaller than the first similarity value, the derived feature used for training the second attribute prediction model is determined as the attribute prediction feature. 10. Attribute prediction methods, including: Acquiring at least two attribute prediction features obtained based on the feature discovery method according to any one of claims 1 to 9 for predicting the attributes included in the operated business, and the quality evaluation results of the at least two attribute prediction features; dividing the at least two attribute prediction features into at least two levels according to the quality evaluation results of the attribute prediction features; and, Deploy the attribute prediction features of the at least two levels into different services, so as to predict the attributes in different services; wherein, the higher the importance of the services, the higher the level of the deployed attribute prediction features. high. 11. A feature discovery device, comprising: a feature acquisition module, a feature derivation module, and a feature evaluation module; The feature acquisition module is configured to acquire at least one original feature generated in the business operation; wherein, the original feature is used to predict the attributes included in the operated business; The feature derivation module is configured to derive at least one original feature acquired by the feature acquisition module to obtain at least one derived feature; The feature evaluation module is configured to evaluate the quality of the at least one derived feature obtained by the feature derivation module, and obtain an attribute prediction feature for predicting the attributes included in the operating business; wherein the The quality of a derived feature is used to characterize how important that feature is to predict the attribute. 12. An attribute prediction device, comprising: a prediction data acquisition module, a grade division module and an attribute prediction module; The predicted data acquisition module is configured to acquire at least two attribute prediction features obtained based on the feature discovery device according to claim 11 for predicting the attributes included in the operating business, and the at least two attribute prediction features. quality assessment results; The grade division module is configured to divide the at least two attribute prediction features into at least two grades according to the quality evaluation result of the attribute prediction feature obtained by the prediction data acquisition module; The attribute prediction module is configured to deploy the attribute prediction features of the at least two grades divided by the grade division module into different services, so as to predict the attributes in different services; wherein, the The higher the importance of the business, the higher the level of the deployed attribute prediction feature. 13. A computing device comprising a memory and a processor, wherein executable code is stored in the memory, and when the processor executes the executable code, the method of any one of claims 1-10 is implemented.

Images