Method for training model and information recommendation system

1. A method for training a model, being applicable to an information recommendation system comprising a target model, the method comprising:

obtaining a plurality of types of features of the target model;

generating a feature group sequence of the plurality of types of features, the feature group sequence comprising a plurality of feature groups and a sequence relation between the plurality of feature groups, each feature group containing at least one type of features among the plurality of types of features;

classifying the target model into a multi-level model according to the feature group sequence, wherein feature weight values corresponding to each level of a multi-level model match with features in a feature group at a corresponding position of the feature group sequence; and

obtaining a trained target model by executing a preset training operation on the feature weight values corresponding to each level of the multi-level model, one or more levels of the trained target model being configured to estimate a recommendation result of a target recommendation task; wherein

generating the feature group sequence of the plurality of types of features comprises:

determining stability of each type of features of the plurality of types of features; and

generating the feature group sequence according to the stability of each type of features of the plurality of types of features.

2. The method of claim 1 , wherein the target model is configured to execute a recommendation task for items to be recommended, and types of features of the target model comprise at least one of: features of attribute of the items to be recommended, features of a target user, and features of a scenario in which the target user is located.

3. The method of claim 1 , wherein determining the stability of each type of features of the plurality of types of features comprises:

obtaining a first feature weight value of each type of features by training the target model via first sample data;

obtaining a second feature weight value of each type of features by training the target model via second sample data;

determining a difference in feature weight values of each type of features according to the first feature weight value and the second feature weight value;

determining a ratio of sample data with partially missing features, from among the first sample data and the second sample data; and

determining the stability of each type of features according to the ratio and the difference.

4. The method of claim 1 , wherein the feature weight values corresponding to each level of the multi-level model match with features in a feature group at a corresponding position of the feature group sequence when feature groups in the feature group sequence are sorted in a descending order of stability.

5. The method of claim 1 , wherein the preset training operation comprises:

keeping feature weight values corresponding to levels other than a current level in the multi-level model unchanged; and

training feature weight values corresponding to the current level according to preset sample data.

6. The method of claim 1 , wherein the method further comprises:

after obtaining the trained target model by executing the preset training operation on the feature weight values corresponding to each level of the multi-level model, executing the target recommendation task.

7. A system for information recommendation, comprising:

a processor; and

a memory configured to store executable program codes which, when executed, cause the processor to:

obtain a plurality of types of features of a target model;

generate a feature group sequence of the plurality of types of features, the feature group sequence comprising a plurality of feature groups and a sequence relation between the plurality of feature groups, each feature group containing at least one type of features among the plurality of types of features;

classify the target model into a multi-level model according to the feature group sequence, wherein feature weight values corresponding to each level of a multi-level model match with features in a feature group at a corresponding position of the feature group sequence; and

obtain a trained target model by executing a preset training operation on the feature weight values corresponding to each level of the multi-level model, one or more levels of the trained target model being configured to estimate a recommendation result of a target recommendation task; wherein

the executable program codes causing the processor to generate the feature group sequence of the plurality of types of features cause the processor to:

determine stability of each type of features of the plurality of types of features; and

generate the feature group sequence according to the stability of each type of features of the plurality of types of features.

8. The system of claim 7 , wherein the target model is configured to execute a recommendation task for items to be recommended, and types of features of the target model comprise at least one of: features of attribute of the items to be recommended, features of a target user, and features of a scenario in which the target user is located.

9. The system of claim 7 , wherein the executable program codes causing the processor to determine the stability of each type of features of the plurality of types of features cause the processor to:

obtain a first feature weight value of each type of features by training the target model via first sample data;

obtain a second feature weight value of each type of features by training the target model via second sample data;

determine a difference in feature weight values of each type of features according to the first feature weight value and the second feature weight value;

determine a ratio of sample data with partially missing features, from among the first sample data and the second sample data; and

determine the stability of each type of features according to the ratio and the difference.

10. The system of claim 7 , wherein the preset training operation comprises:

keeping feature weight values corresponding to levels other than a current level in the multi-level model unchanged; and

training feature weight values corresponding to the current level according to preset sample data.

11. The system of claim 7 , wherein the executable program codes further cause the processor to:

execute the target recommendation task.

12. The system of claim 7 , wherein the feature weight values corresponding to each level of the multi-level model match with features in a feature group at a corresponding position of the feature group sequence when feature groups in the feature group sequence are sorted in a descending order of stability.

13. A method for training a model, being applicable to an information recommendation system comprising a first system, the first system comprising a target model, the method comprising:

obtaining, by the first system, a plurality of types of features of the target model;

generating, by the first system, a feature group sequence of the plurality of types of features, the feature group sequence comprising a plurality of feature groups and a sequence relation between the plurality of feature groups, each feature group containing at least one type of features among the plurality of types of features;

classifying, by the first system, the target model into a multi-level model according to the feature group sequence, wherein feature weight values corresponding to each level of a multi-level model match with features in a feature group at a corresponding position of the feature group sequence; and

obtaining, by the first system, a trained target model by executing a preset training operation on the feature weight values corresponding to each level of the multi-level model, one or more levels of the trained target model being configured to estimate a recommendation result of a target recommendation task; wherein

generating, by the first system, the feature group sequence of the plurality of types of features comprises:

determining, by the first system, stability of each type of features of the plurality of types of features; and

generating, by the first system, the feature group sequence according to the stability of each type of features of the plurality of types of features.

14. The method of claim 13 , wherein determining, by the first system, the stability of each type of features of the plurality of types of features comprises:

obtaining, by the first system, a first feature weight value of each type of features by training the target model via first sample data;

obtaining, by the first system, a second feature weight value of each type of features by training the target model via second sample data;

determining, by the first system, a difference in feature weight values of each type of features according to the first feature weight value and the second feature weight value;

determining, by the first system, a ratio of sample data with partially missing features, from among the first sample data and the second sample data; and

determining, by the first system, the stability of each type of features according to the ratio and the difference.

15. The method of claim 13 , wherein the preset training operation comprises:

keeping feature weight values corresponding to levels other than a current level in the multi-level model unchanged; and

training feature weight values corresponding to the current level according to preset sample data.

16. The method of claim 13 , the information recommendation system further comprising a second system, wherein the method further comprises:

after the obtaining, by the first system, the trained target model by executing the preset training operation on the feature weight values corresponding to each level of the multi-level model, sending, by the first system, the trained target model to the second system; the trained target model being configured for the second system to execute the target recommendation task, and the target recommendation task being a recommendation task for one or more target items to be recommended.

17. The method of claim 13 , wherein the feature weight values corresponding to each level of the multi-level model match with features in a feature group at a corresponding position of the feature group sequence when feature groups in the feature group sequence are sorted in a descending order of stability.