Granted Patent B2

US 12,373,670 · App. 18/637,142 · Granted Jul 29, 2025

Inventors: Bernardo José Amaral Nunes de Almeida Branco (Lisbon, PT); Pedro Caldeira Abreu (London, GB); Ana Sofia Leal Gomes (Lisbon, PT); Mariana S. C. Almeida (Lisbon, PT); João Tiago Barriga Negra Ascensão (Lisbon, PT); Pedro Gustavo Santos Rodrigues Bizarro (Lisbon, PT)

Assignee: Feedzai—Consultadoria e Inovação Tecnológica, S.A.

G06N3/045G06F7/08G06F16/2379G06F16/27G06N3/042G06N3/044G06N3/063G06N3/08G06Q20/4016G06Q20/409G06Q40/12

Abstract

Claims (86)

1. A method, comprising:

receiving data of a first transaction;

retrieving a first state;

providing the retrieved first state and an input data based on the first transaction to a machine learning trained recurrent neural network model to determine a second state including by:

transforming an input vector into a new vector, and

representing the new vector in a space to feed at least one layer of the recurrent neural network model;

determining a prediction result associated with a fraud threat of the first transaction using at least (i) the first state and (ii) the input data based on the first transaction;

updating a saved recurrent neural network state for an entity associated with the first transaction to be the second state;

receiving data of a second transaction, wherein the second transaction is associated with the entity; and

unlooping a neural network associated with the saved recurrent neural network state including by:

retrieving the second state;

providing the second state and an input data based on the second transaction to the machine learning trained recurrent neural network model to determine a third state;

determining a prediction result associated with a fraud threat of the second transaction using at least (i) the second state and (ii) the input data based on the second transaction, wherein the second transaction is approved based at least in part on a determination that the prediction result associated with the fraud threat of the second transaction is below a threshold; and

updating the saved recurrent neural network state for the entity to be the third state.

2. The method of claim 1 , wherein:

the first transaction and the second transaction are included in an interleaved sequence of transactions; and

the interleaved sequence of transactions includes transactions received in an order in which adjacent transactions are not necessarily associated with a same entity.

3. The method of claim 2 , wherein the interleaved sequence of transactions includes:

an initial transaction received at a first time, the initial transaction being associated with a first entity;

a middle transaction received at a second time after the first time, the middle transaction being associated with a second entity different from the first entity; and

a later transaction received at a third time after the second time, the later transaction being associated with the first entity.

4. The method of claim 2 , further comprising:

receiving data of a third transaction, wherein the third transaction is associated with an entity different from the first transaction, and the third transaction is included in the interleaved sequence of transactions;

retrieving a first state that is a saved recurrent neural network state for the entity associated with the third transaction;

providing the retrieved first state and an input data based on data of the third transaction to a machine learning trained recurrent neural network model to determine a new second state;

determining a prediction result associated with a fraud threat of the third transaction using at least (i) the second state and (ii) the input data based on the third transaction; and

updating the saved recurrent neural network state for the entity associated with the third transaction to be the second state.

5. The method of claim 1 , wherein the saved recurrent neural network state for an entity associated with the first transaction is retrieved from at least one of: gated recurrent unit (GRU) cell or a Long Short-Term Memory (LSTM) cell.

6. The method of claim 1 , wherein data of the first transaction includes at least one of: a numerical field, a categorical field, a timestamp field, or an entity identification field.

7. The method of claim 1 , wherein the first state includes an encoding of transactions associated with the entity associated with the first transaction.

8. The method of claim 1 , further comprising batch processing transactions including by continuously and concurrently reading batches from at least one database and inserting the batch processing transactions into a queue for processing.

9. The method of claim 8 , wherein the batch processing is performed in offline training.

10. The method of claim 8 , further comprising:

building a first database of entities meeting a first criterion;

building a second database of entities meeting a second criterion; and

building a validation dataset using a representative sample of entities from the first database and the second database during a validation period arranged in descending order in number of transactions.

11. The method of claim 10 , wherein the first criterion includes containing at least one fraudulent transaction and the second criterion is not containing any fraudulent transactions.

12. The method of claim 10 , wherein building the validation dataset includes using a parameter defining a proportion of information to use from the first database and the second database in response to a determination that there is class imbalance between the first database and the second database.

13. The method of claim 10 , wherein the representative sample of entities includes a random sample.

14. The method of claim 1 , further comprising:

in response to a transaction being a scorable instance, performing a forward pass and a backward pass; and

in response to a transaction being a non-scorable instance, performing a forward pass without a backward pass.

15. The method of claim 1 , further comprising performing a batch inference process including by:

obtaining a current state for a given card identifier; and

obtaining a new state for the card identifier and a score for a respective transaction.

16. The method of claim 1 , wherein further comprising, prior to retrieving a first state that is a saved recurrent neural network state for an entity associated with the first transaction, pre-processing the input data based on data of the first transaction.

17. The method of claim 16 , wherein the pre-processing includes at least one of:

creating at least one of a time feature or an entity-related feature;

fitting transformer objects for normalization, bucketing, and indexing;

applying transformers to training, validation, and test datasets;

creating a list of chronologically-sorted sequences of transactions based at least in part on the training, validation, and test datasets;

removing those sequences of transactions that do not contain any transaction in a period of interest; and

storing the sequences of transactions.

18. A system, comprising:

a processor configured to:

receive data of a first transaction;

retrieve a first state;

provide the retrieved first state and an input data based on the first transaction to a machine learning trained recurrent neural network model to determine a second state including by:

transforming an input vector into a new vector, and

representing the new vector in a space to feed at least one layer of the recurrent neural network model;

determine a prediction result associated with a fraud threat of the first transaction using at least (i) the first state and (ii) the input data based on the first transaction;

update a saved recurrent neural network state for an entity associated with the first transaction to be the second state;

receive data of a second transaction, wherein the second transaction is associated with the entity; and

unloop a neural network associated with the saved recurrent neural network state including by:

retrieving the second state;

providing the second state and an input data based on the second transaction to the machine learning trained recurrent neural network model to determine a third state;

determining a prediction result associated with a fraud threat of the second transaction using at least (i) the second state and (ii) the input data based on the second transaction, wherein the second transaction is approved based at least in part on a determination that the prediction result associated with the fraud threat of the second transaction is below a threshold; and

updating the saved recurrent neural network state for the entity to be the third state.

19. The system of claim 18 , wherein:

each subset is stored in a separate database;

a test dataset is used to pick a model; and

the test dataset includes a representative sample of entities from a first database and a second database arranged in descending order in number of transactions.

20. A computer program product embodied in a non-transitory computer readable medium and comprising computer instructions for:

receiving data of a first transaction;

retrieving a first state;

providing the retrieved first state and an input data based on the first transaction to a machine learning trained recurrent neural network model to determine a second state including by:

transforming an input vector into a new vector, and

representing the new vector in a space to feed at least one layer of the recurrent neural network model;

determining a prediction result associated with a fraud threat of the first transaction using at least (i) the first state and (ii) the input data based on the first transaction;

updating a saved recurrent neural network state for an entity associated with the first transaction to be the second state;

receiving data of a second transaction, wherein the second transaction is associated with the entity; and

unlooping a neural network associated with the saved recurrent neural network state including by:

retrieving the second state;

providing the second state and an input data based on the second transaction to the machine learning trained recurrent neural network model to determine a third state;

determining a prediction result associated with a fraud threat of the second transaction using at least (i) the second state and (ii) the input data based on the second transaction, wherein the second transaction is approved based at least in part on a determination that the prediction result associated with the fraud threat of the second transaction is below a threshold; and

updating the saved recurrent neural network state for the entity to be the third state.

Continuity (3)

References Cited (61)