Convolutional neural networks for variable prediction using raw data
While artificial neural networks can be used to predict particular values in certain contexts, convolutional neural networks are not typically used in these contexts—instead they may be employed for image recognition. However, raw transactional data may be structured to take advantage of convolutional neural network (CNN) techniques by arranging the data such that correlations are increased between nearby other data. In arranging data in this manner, the structured CNN (SCNN) can operate efficiently without having to make use of engineered data features, the generation and maintenance of which can be a time-consuming process.
1. A method for training a structured convolutional neural network (SCNN) usable with raw transaction data, comprising:
creating the SCNN by a computer system, including creating an input layer, a first convolutional module directly connected to the input layer, a second convolutional module directly connected to the first convolutional module, and a fully connected module directly connected to the second convolutional module,
wherein the first convolutional module comprises a first convolutional layer, a second convolutional layer, and a first average pooling layer,
wherein the second convolutional module comprises a third convolutional layer, a fourth convolutional layer, and a second average pooling layer,
wherein the fully connected module comprises a first fully connected layer and a second fully connected layer;
accessing, by the computer system, raw transaction data comprising a plurality of records each having a plurality of fields containing a respective feature value from a set of features;
arranging the raw transaction data such that at least one of the plurality of fields is re-ordered to increase a similarity of raw data between adjacent fields of the plurality of fields;
training the SCNN, comprising:
repeatedly feeding the raw transaction data into the SCNN, comparing outputs of the SCNN to known results associated with the raw transaction data, and making adjustments to a plurality of weighting parameters for the SCNN based on results of the comparing; and
assessing a risk for an electronic payment transaction requested by a user based on an output of the SCNN for data from the user.
2. The method of claim 1 , further comprising predicting, for a plurality of users, a customer value for a particular future time period using the SCNN.
3. The method of claim 1 , wherein arranging the raw transaction data is performed manually.
4. The method of claim 1 , wherein arranging the raw transaction data is performed according to automatically created specifications based on statistical observations of correlation between different feature values in the plurality of records.
5. The method of claim 1 , wherein the first, second, third, and fourth convolutional layers each use at least three weighting parameters to optimize results.
6. The method of claim 1 , wherein the raw transaction data includes fields for a transaction amount, a currency type, and a network address associated with a user.
7. The method of claim 1 , wherein the first average pooling layer is configured to use average values from a plurality of neurons at the second convolutional layer.
8. The method of claim 1 , wherein the SCNN is trained to optimize predicted customer values for customers of an electronic transaction payment service provider.
9. The method of claim 1 , further comprising updating the SCNN based on a newly acquired plurality of records that are more recent than the plurality of records.
10. The method of claim 1 , wherein the computer system comprises one or more processors and one or more memories connected via a network to a database associated with an electronic service provider at which the plurality of records are stored.
11. A computer system, comprising:
a processor; and
a computer-readable medium having stored thereon instructions that are executable to cause the computer system to perform operations comprising:
accessing a structured convolutional neural network (SCNN) that includes an input layer, a first convolutional module connected to the input layer, a second convolutional module connected to the first convolutional module, and a fully connected module directly connected to the second convolutional module,
wherein the first convolutional module comprises a first convolutional layer, a second convolutional layer, and a first average pooling layer,
wherein the second convolutional module comprises a third convolutional layer, a fourth convolutional layer, and a second average pooling layer,
wherein the fully connected module comprises a first fully connected layer and a second fully connected layer;
accessing raw transaction data comprising a plurality of records each having a plurality of fields containing a respective feature value from a set of features;
arranging the raw transaction data such that at least one of the plurality of fields is re-ordered to increase a similarity of raw data between adjacent fields of the plurality of fields;
training the SCNN, comprising:
repeatedly feeding the raw transaction data into the SCNN, comparing outputs of the SCNN to known results associated with the raw transaction data, and making adjustments to a plurality of weighting parameters for the SCNN based on results of the comparing; and
assessing a risk for an electronic payment transaction requested by a user based on an output of the SCNN from data for the user.
12. The computer system of claim 11 , wherein arranging the raw transaction data is performed according to automatically created specifications based on statistical observations of correlations between different feature values in the plurality of records.
13. The computer system of claim 11 , wherein the first, second, third, and fourth convolutional layers each use at least three weighting parameters to optimize results.
14. The computer system of claim 11 , wherein the raw transaction data includes fields for a home address country, a home address postal code, and a network address associated with a user.
15. The computer system of claim 11 , wherein the first average pooling layer is configured to use average values from a plurality of neurons at the second convolutional layer.
16. A non-transitory computer-readable medium having stored thereon instructions that are executable by a computer system to cause the computer system to perform operations comprising:
accessing a structured convolutional neural network (SCNN) that includes an input layer, a first convolutional module connected to the input layer, a second convolutional module directly connected to the first convolutional module, and a fully connected module connected to the second convolutional module,
wherein the first convolutional module comprises a first convolutional layer, a second convolutional layer, and a first average pooling layer,
wherein the second convolutional module comprises a third convolutional layer, a fourth convolutional layer, and a second average pooling layer,
wherein the fully connected module comprises a first fully connected layer and a second fully connected layer;
accessing raw transaction data comprising a plurality of records each having a plurality of fields containing a respective feature value from a set of features;
arranging the raw transaction data such that at least one of the plurality of fields is re-ordered to increase a similarity of raw data between adjacent fields of the plurality of fields;
training the SCNN, comprising:
repeatedly feeding the raw transaction data into the SCNN, comparing outputs of the SCNN to known results associated with the raw transaction data, and making adjustments to a plurality of weighting parameters for the SCNN based on results of the comparing; and
assessing a risk for an electronic payment transaction requested by a user based on an output of the SCNN from data for the user.
17. The non-transitory computer-readable medium of claim 16 , wherein arranging the raw transaction data is performed according to automatically created specifications based on statistical observations of correlation between different feature values in the plurality of records.
18. The non-transitory computer-readable medium of claim 16 , wherein the first, second, third, and fourth convolutional layers each use at least three weighting parameters to optimize results.
19. The non-transitory computer-readable medium of claim 16 , wherein the second average pooling layer is configured to use average values from a plurality of neurons at the fourth convolutional layer.
20. The non-transitory computer-readable medium of claim 16 , wherein the raw transaction data includes fields for a home address country, a home address postal code, and a network address associated with a user.