Transaction feature generation

Patent metadata

Field	Value
Publication number	US-11210673-B2
Application number	US-202016803150-A
Country	US
Kind code	B2
Filing date	Feb 27, 2020
Priority date	May 29, 2019
Publication date	Dec 28, 2021
Grant date	Dec 28, 2021

Abstract

The present specification discloses a method and an apparatus for training a transaction feature generation model, and a method and an apparatus for generating a transaction feature. The method for generating a transaction feature can include the following: obtaining a target dataset, where the target dataset includes some pieces of transaction data; obtaining some original features of the transaction data and determining one or more combination methods for the original features; determining a feature vector of a new feature that is obtained by combining the original features based on each combination method; inputting the feature vector into a trained transaction feature generation model, and outputting a prediction result of the new feature; and selecting some new features whose prediction results meet a specified condition as transaction features generated for the target dataset.

First claim

What is claimed is:

1. A computer-implemented method for evaluating combination transaction features, the method comprising:
obtaining a target dataset, wherein the target dataset comprises transaction data;
obtaining transaction feature vectors having respective values for a plurality of original features represented in the transaction data;
processing the transaction feature vectors to train a transaction feature generation model that generates a predicted feature label for a new feature represented by a feature vector, including, for each new feature of a plurality of new features, performing operations comprising:
determining, for the new feature, a respective combination method for a respective original feature group comprising a plurality of the original features of the transaction data;
generating values of the new feature by combining corresponding values of the plurality of the original features based on the respective combination method;
computing the feature label for the new feature based on a sum of differences between the values of the new feature and corresponding transaction labels;
providing the feature vector for the new feature as a first input to the transaction feature generation model to obtain a prediction result corresponding to the feature label, and
updating parameters of the transaction feature generation model based on a difference between the prediction result and the feature label for the new feature;

receiving another new feature having a second combination method;
computing a second feature vector for the other new feature according to the second combination method;
providing the second feature vector for the other new feature as a second input to the transaction feature generation model to obtain a second prediction result for the other new feature;
evaluating the other new feature relative to one or more additional new features using the second prediction result for the other new feature;
selecting, based on the evaluation, the other new feature as a new transaction feature for training a model to classify transactions; and
using the other new feature to train the model to classify transactions having original feature values including using the second combination method of the other new feature to combine original feature values of the transactions.

2. The method according to claim 1 , further comprising:
obtaining the original features from the transaction data wherein two or more of the original features obtained from the transaction data are of a first data type; and
wherein determining the respective combination method for the respective original feature group comprises:
determining a combination method that matches the first data type as the respective combination method for the respective original feature group.

3. The method according to claim 2 , wherein
when the first data type is a numeric type, the respective combination method comprises one or more of the following:
an arithmetic operation, logarithmic sum calculation, or quadratic sum calculation.

4. The method according to claim 2 , wherein
when the first data type is a string, the respective combination method comprises one or more of the following:
an arithmetic operation of string lengths, a logarithmic sum of string lengths, or a quadratic sum of string lengths.

5. The method according to claim 1 , wherein
the second feature vector is generated based on meta information of the original features and the second combination method.

6. The method according to claim 5 , wherein the meta information comprises one or more of the following:
an average value, a variance, or a number of unique data of the original features.

7. The method according to claim 1 , wherein computing the second feature vector comprises:
computing the second feature vector based on meta information of the original features, the second combination method, and meta features of two or more sample datasets used to train the transaction feature generation model.

8. The method according to claim 7 , wherein the meta features of the two or more sample datasets comprise one or more of the following:
a number of original features, a number of original features of a numeric type, or a ratio of positive and negative samples.

9. The method according to claim 1 , wherein the second combination method combines features of a same type as the respective combination method.

10. The method according to claim 1 , wherein the differences between the values of the new feature and the corresponding transaction labels comprise a mean square error between a first value of the values of the new feature and a first transaction label of the corresponding transaction labels.

11. A non-transitory, computer-readable medium storing one or more instructions executable by a computer system to perform operations comprising:
obtaining a target dataset, wherein the target dataset comprises transaction data;
obtaining transaction feature vectors having respective values for a plurality of original features represented in the transaction data;
processing the transaction feature vectors to train a transaction feature generation model that generates a feature label for a new feature represented by a feature vector, including, for each new feature of a plurality of new features, performing other operations comprising:
determining, for the new feature, a respective combination method for a respective original feature group comprising a plurality of the original features of the transaction data;
generating values of the new feature by combining corresponding values of the plurality of the original features based on the respective combination method;
computing the feature label for the new feature based on a sum of differences between the values of the new feature and corresponding transaction labels;
providing the feature vector for the new feature as a first input to the transaction feature generation model to obtain a prediction result corresponding to the feature label, and
updating parameters of the transaction feature generation model based on a difference between the prediction result and the feature label for the new feature;
receiving another new feature having a second combination method;
computing a second feature vector for the other new feature according to the second combination method;
providing the second feature vector for the other new feature as a second input to the transaction feature generation model to obtain a second prediction result for the other new feature;
evaluating the other new feature relative to one or more additional new features using the second prediction result for the other new feature;
selecting, based on the evaluation, the other new feature as a new transaction feature for training a model to classify transactions; and
using the other new feature to train the model to classify transactions having original feature values including using the second combination method of the other new feature to combine original feature values of the transactions.

12. The non-transitory, computer-readable medium of claim 11 , wherein the operations further comprise :
obtaining the original features from the transaction data wherein two or more of the original features obtained from the transaction data are of a first data type; and
wherein determining the respective combination method for the respective original feature group comprises:
determining a combination method that matches the first data type as the respective combination method for the respective original feature group.

13. The non-transitory, computer-readable medium of claim 11 , wherein

Assignees

• Advanced New Technologies Co Ltd

Inventors

• Li Longfei

Classifications

• G06N20/00 Primary

  Machine learning · CPC title

• G06Q20/4016 Primary

  involving fraud or risk level assessment in transaction processing · CPC title

• G06F18/2178

  based on feedback of a supervisor · CPC title

• G06N3/09

  Supervised learning · CPC title

• G06N3/0499

  Feedforward networks · CPC title