Synthesizing user transactional data for de-identifying sensitive information
As described herein, a system, method, and computer program are provided for synthesizing user transactional data for de-identifying sensitive information. In use, transactional data of a plurality of users is identified. Additionally, the plurality of users are clustered based on the transactional data, to form groups of users having transactional data representing similar transactional behavior. Further, synthesized transactional data is generated for the users in each group by: identifying a subset of the transactional data that corresponds to the users in each group, shuffling the transactional data in the subset across the users in each group, and perturbing portions of the shuffled transactional data.
1. A non-transitory computer-readable media storing computer instructions which when executed by one or more processors of a device cause the device to:
identify transactional data of a plurality of users;
cluster the plurality of users based on the transactional data, to form groups of users having transactional data representing similar transactional behavior;
generate synthesized transactional data for the users in each group by:
identifying a subset of the transactional data that corresponds to the users in each group,
shuffling the transactional data in the subset across the users in each group, wherein the shuffling includes:
constructing a pool of transactions from the subset of the transactional data,
for each user in each group, sampling transactions from the pool based on:
a number of transactions associated with the user in the subset of the transactional data, and
a category of each of the transactions associated with the user in the subset of the transactional data,
wherein at least one of:
a number of transactions sampled from the pool for each user in each group matches the number of transactions in the subset of the transactional data that are associated with the user, or
a number of transactions of a particular category sampled from the pool for each user in each group matches a number of transactions of the particular category in the subset of the transactional data that are associated with the user, and
perturbing portions of the shuffled transactional data.
2. The non-transitory computer-readable media of claim 1 , wherein the transactional data of a plurality of users is transactional records stored for the plurality of users that represent financial transaction associated with the plurality of users.
3. The non-transitory computer-readable media of claim 1 , wherein clustering the plurality of users based on the transactional data includes:
encoding historical transactional data to form a knowledge graph, and
learning a continuous representation for entities in the graph, and
clustering the plurality of users based on the learned entities.
4. The non-transitory computer-readable media of claim 3 , wherein knowledge graph is generated to include:
nodes representing users, merchants, and geo-locations, and
edges representing transaction features.
5. The non-transitory computer-readable media of claim 1 , wherein the shuffling is performed stochastically.
6. The non-transitory computer-readable media of claim 1 , wherein the number of transactions sampled from the pool for each user in each group matches the number of transactions in the subset of the transactional data that are associated with the user.
7. The non-transitory computer-readable media of claim 1 , wherein the number of transactions of a particular category sampled from the pool for each user in each group matches the number of transactions of the particular category in the subset of the transactional data that are associated with the user.
8. The non-transitory computer-readable media of claim 1 , wherein perturbing portions of the shuffled transactional data includes:
perturbing transaction amounts within the shuffled transactional data.
9. The non-transitory computer-readable media of claim 8 , wherein the transaction amounts are perturbed up to a predefined maximum threshold percentage.
10. The non-transitory computer-readable media of claim 1 , wherein perturbing portions of the shuffled transactional data includes:
shuffling elements of transaction descriptions within the shuffled transactional data.
11. The non-transitory computer-readable media of claim 10 , wherein the elements are one or more consecutive words in a transaction description carrying homogeneous related information and are identified using machine learning models.
12. The non-transitory computer-readable media of claim 10 , wherein the elements within each of the transaction descriptions are shuffled randomly.
13. The non-transitory computer-readable media of claim 1 , wherein perturbing portions of the shuffled transactional data includes:
masking Personally Identifiable Information (PII).
14. The non-transitory computer-readable media of claim 13 , wherein the PII includes at least one of:
social security number,
credit card number, or
user name.
15. The non-transitory computer-readable media of claim 13 , wherein the masking is performed using deep learning.
16. The non-transitory computer-readable media of claim 1 , wherein perturbing portions of the shuffled transactional data includes:
adding noise to transaction descriptions.
17. A method, comprising:
at a computer system:
identifying transactional data of a plurality of users;
clustering the plurality of users based on the transactional data, to form groups of users having transactional data representing similar transactional behavior;
generating synthesized transactional data for the users in each group by:
identifying a subset of the transactional data that corresponds to the users in each group,
shuffling the transactional data in the subset across the users in each group, wherein the shuffling includes:
constructing a pool of transactions from the subset of the transactional data,
for each user in each group, sampling transactions from the pool based on:
a number of transactions associated with the user in the subset of the transactional data, and
a category of each of the transactions associated with the user in the subset of the transactional data,
wherein at least one of:
a number of transactions sampled from the pool for each user in each group matches the number of transactions in the subset of the transactional data that are associated with the user, or
a number of transactions of a particular category sampled from the pool for each user in each group matches a number of transactions of the particular category in the subset of the transactional data that are associated with the user, and
perturbing portions of the shuffled transactional data.
18. A system, comprising:
a non-transitory memory storing instructions; and
one or more processors in communication with the non-transitory memory that execute the instructions to:
identify transactional data of a plurality of users;
cluster the plurality of users based on the transactional data, to form groups of users having transactional data representing similar transactional behavior;
generate synthesized transactional data for the users in each group by:
identifying a subset of the transactional data that corresponds to the users in each group,
shuffling the transactional data in the subset across the users in each group, wherein the shuffling includes:
constructing a pool of transactions from the subset of the transactional data,
for each user in each group, sampling transactions from the pool based on:
a number of transactions associated with the user in the subset of the transactional data, and
a category of each of the transactions associated with the user in the subset of the transactional data,
wherein at least one of:
a number of transactions sampled from the pool for each user in each group matches the number of transactions in the subset of the transactional data that are associated with the user, or
a number of transactions of a particular category sampled from the pool for each user in each group matches a number of transactions of the particular category in the subset of the transactional data that are associated with the user, and
perturbing portions of the shuffled transactional data.