Deep behavioral networks for fraud detection
A transaction processing system includes a transaction processing module configured to receive first information associated with a first proposed transaction, retrieve second information associated with at least one prior transaction that is associated with the first proposed transaction, and calculate a time-decayed algorithm using the second information to generate third information. The transaction processing system also includes a weighting module communicably coupled to the transaction processing module, wherein the weighting module is configured to receive the third information from the neural-based processing module, apply a weighting factor to the third information to generate fourth information, and calculate at least one processing algorithm using the first information and the fourth information to generate an output. The output of the weighting module is used by an additional transaction processing module to determine whether the first proposed transaction is fraudulent.
1. A fraud detection transaction processing system communicably coupled with a client and a user via one or more communication links, the fraud detection transaction processing system comprising:
a neural-based transaction processing module, which is a layer of neural cells with local memory where the memory is time-decayed prior to state updates upon processing of a new sample, the neural-based transaction processing module configured to:
receive, from the client over the one or more communication links, first information associated with a first proposed electronic transaction, the first proposed electronic transaction constituting the new sample for a given iteration,
retrieve, from a communicably coupled storage repository, second information associated with at least one prior electronic transaction that is associated with the first proposed electronic transaction, and
generate third information by calculating, using a controller executing on a hardware processor and associated with, and executable by, the neural-based transaction processing module, an exponential time-decayed algorithm using the second information, wherein calculating the exponential time-decayed algorithm in a current iteration causes the local memory to be time-decayed, exponentially, and comprises
invoking a weighting module to compute a weighting function from a time interval between the at least one prior electronic transaction and the first proposed electronic transaction,
multiplying an output of a previous iteration of calculating the exponential time-decayed algorithm to the weighting function by the weighting module or a controller of the transaction processing system, wherein the output of a previous iteration of calculating the exponential time-decayed algorithm is an exponential time-decayed algorithm calculated in the previous iteration, the exponential time-decayed algorithm calculated in the previous iterations providing long-term retention, for a duration set by a half-life parameter of said exponential time-decayed algorithm, of contributions of prior transactions that are not the most recent transaction, and
adding the multiplication result and the received first information of the first proposed electronic transaction by the weighting module or the controller to update the exponential time-decayed algorithm, thereby generating third information in the current iteration, which comprises a weighted summation of the first proposed electronic transaction relative to the prior transactions; and
wherein the fraud detection transaction processing system further comprises the weighting module, the weighting module communicably coupled to the neural-based transaction processing module, and wherein the weighting module is configured to:
generate fourth information by receiving the third information from the neural-based transaction processing module, and applying a weighting factor to the third information,
calculate at least one processing algorithm using the first information and the fourth information to generate an output including a risk score that is indicative of whether the first proposed electronic transaction is fraudulent; and
provide the output to the client, over the one or more communication links, to flag to the client while the first proposed electronic transaction is in-flight, before the first proposed electronic transaction is completed, whether the first proposed electronic transaction is fraudulent,
wherein a total latency of the fraud detection transaction processing system, as defined by a duration between a time of a receipt of the first information by the neural-based transaction module and a time of a corresponding provision of the output to the client, is a matter of milli-seconds.
2. The fraud detection transaction processing system of claim 1 , wherein the receiving of the third information from the neural-based transaction processing module includes receiving third information from the current iteration and receiving third information from each of previous iterations.
3. The fraud detection transaction processing system of claim 2 , wherein the neural-based transaction processing module and the weighting module are configured to combine to read a vector of relative time intervals between each of the previous iterations and the current iteration.
4. The fraud detection transaction processing system of claim 3 , wherein the weighting module is further configured to compute a vector of relative time encodings based on the read vector of relative time intervals and concatenate the vector of relative time encodings with the received third information from each of the previous iterations.
5. The fraud detection transaction processing system of claim 4 , wherein the weighting module is further configured to produce a key vector based on the concatenation result, produce a query vector based on the received third information in the current iteration, and generate a vector of attention scores based on the key vector and the query vector.
6. The fraud detection transaction processing system of claim 5 , wherein the weighting module is further configured to apply a weighting vector to the vector of attention scores to generate a modified vector of attention scores, and wherein the weighting vector is defined based on the vector of relative time intervals between each of the previous iterations and the current iteration.
7. The fraud detection transaction processing system of claim 6 , wherein the neural-based transaction processing module is further configured to normalize the modified vector of attention scores to produce attention weights, and wherein the normalization is conducted in a softmax layer of the neural-based transaction processing module.
8. The fraud detection transaction processing system of claim 7 , wherein the weighting module is further configured to generate a value vector based on the concatenation result, and produce a context vector by weighting the value vector with the modified vector of attention scores, wherein the output comprises the context vector.
9. A method for detecting a fraudulent transaction using a fraud detection transaction processing system communicably coupled with a client and a user via one or more communication links, the method comprising:
receiving, by a neural-based transaction processing module, which is a layer of neural cells with local memory where the memory is time-decayed prior to state updates upon processing of a new sample, the neural-based transaction processing module, first information associated with a first proposed electronic transaction from the client over the one or more communication links, the first proposed electronic transaction constituting the new sample for a given iteration, wherein the first information from the client comprises a currently attempted payment charge, including an account number, an account owner, a day and time of the first proposed electronic transaction, an amount of the first proposed electronic transaction, a location of where the first proposed electronic transaction is occurring, a vendor seeking authorization of the charge, or manner in which the transaction is executed;
retrieving, by the neural-based transaction processing module from a communicably coupled storage repository, second information, wherein the second information comprises a state of an account used to make the first proposed electronic transaction, the state of the account indicative of a duration from at least one prior electronic transaction that is associated with the first proposed electronic transaction;
generating third information by calculating, using a controller executing on a hardware processor and associated with and executable by the neural-based transaction processing module, an exponential time-decayed algorithm using the second information, wherein the calculating of the exponential time-decayed algorithm in a current iteration causes the local memory to be time-decayed, exponentially, and comprises
invoking a weighting module to compute a weighting function from a time interval between the at least one prior electronic transaction and the first proposed electronic transaction,
multiplying, by the weighting module or the controller of the transaction processing system, an output of a previous iteration of calculating the exponential time-decayed algorithm to the weighting function, wherein the output of a previous iteration of calculating the exponential time-decayed algorithm is a time-decayed algorithm calculated in the previous iteration, the exponential time-decayed algorithm calculated in the previous iterations providing long-term retention, for a duration set by a half-life parameter of said exponential time-decayed algorithm, of contributions of prior transactions that are not the most recent transaction, and
adding, by the weighting module or the controller, the multiplication result and the received first information of the first proposed electronic transaction to update the exponential time-decayed algorithm, thereby generating the third information in the current iteration, which comprises a weighted summation of the first proposed electronic transaction relative to the prior transactions;
generate fourth information by
receiving, by the weighting module, the third information from the neural-based transaction processing module,
applying, by the weighting module, a weighting factor to the third information to generate fourth information;
calculating, by the weighting module, at least one processing algorithm using the first information and the fourth information to generate an output including a risk score that is indicative of whether the first proposed electronic transaction is fraudulent; and
providing, by the weighting module or the neural-based transaction processing module, the output to the client, over the one or more communication links, to flag to the client while the first proposed electronic transaction is in-flight, and before the first proposed electronic transaction is completed, whether the first proposed electronic transaction is fraudulent,
wherein a total latency of the fraud detection transaction processing system, as defined by a duration between a time of a receipt of the first information by the neural-based transaction module and a time of a corresponding provision of the output to the client, is a matter of milli-seconds.
10. The method for detecting a fraudulent transaction of claim 9 , wherein the receiving of the third information from the neural-based transaction processing module includes receiving third information from the current iteration and receiving third information from each of previous iterations.
11. The method for detecting a fraudulent transaction of claim 10 , further comprising combining to read, by the neural-based transaction processing module and the weighting module, a vector of relative time intervals between each of the previous iterations and the current iteration.
12. The method for detecting a fraudulent transaction of claim 11 , further comprising:
computing, by the weighting module, a vector of relative time encodings based on the read vector of relative time intervals; and
concatenating the vector of relative time encodings with the received third information from each of the previous iterations.
13. The method for detecting a fraudulent transaction of claim 12 , further comprising:
producing, by the weighting module, a key vector based on the concatenation result;
producing, by the weighting module, a query vector based on the received third information in the current iteration; and
generating, by the weighting module, a vector of attention scores based on the key vector and the query vector.
14. The method for detecting a fraudulent transaction of claim 13 , further comprising applying, by the weighting module, a weighting vector to the vector of attention scores to generate a modified vector of attention scores, wherein the weighting vector is defined based on the vector of relative time intervals between each of the previous iterations and the current iteration.
15. The method for detecting a fraudulent transaction of claim 14 , further comprising normalizing, by the transaction processing module, the modified vector of attention scores to produce attention weights, wherein the normalization is conducted in a softmax layer of the transaction processing module.
16. The method for detecting a fraudulent transaction of claim 15 , further comprising:
generating, by the weighting module, a value vector based on the concatenation result; and
producing, by the weighting module, a context vector by weighting the value vector with the modified vector of attention scores,
wherein the output comprises the context vector.