Temporal explanations of machine learning model outcomes
In transactional systems where past transactions can have impact on the current score of a machine learning based decision model, the transactions that are most responsible for the score and the associated reasons are determined by the transactional system. A system and method identifies such past transactions that maximally impact the current score and allow for a more effective understanding of the scores generated by a model in a transactional system and explanation of specific transactions for automated decisioning, to explain the scores in terms of past transactions. Further an existing instance-based explanation system is used to identify the reasons for the score, and how the identified transactions influence these reasons. A combination of impact on score and impact on reasons determines the most impactful past transaction with respect to the most recent score being explained.
1 . A computer-implemented machine learning system for managing abnormal transactions in a communications network, the system comprising:
one or more data storage media devices communicatively coupled to the communications network, the one or more data storage media devices storing entity data communicated over the communications network;
a first computing system communicatively coupled to the communications network, the first computing system collecting the entity data and determining an input vector based on a plurality of transactions associated with the entity data, wherein the plurality of transactions include one or more abnormal transactions involving potentially fraudulent or malicious activity;
a first analytic model for recursively generating a first value based on the input vector, the first value being updated by the first analytic model as one or more transactions are omitted from the plurality of transactions; and
a second analytic model generating a weighted reason vector and associated plurality of ranked reasons that provide an understanding of the first value based on a set of top contributor variables in the input vector and latent features of the first analytic model associated with the first value;
wherein one or more transactions involving potentially fraudulent or malicious activity, are identified and omitted from the plurality of transactions associated with the entity data, to enable an early elimination of transactions having fraudulent or malicious nature, responsive to a quantification of a maximal effect on the first value recursively generated based on the input vector.
2 . The system of claim 1 , wherein the entity data represents the one or more transactions and an output instance comprising a set of one or more values is generated to identify one or more temporal events or transactions most responsible for generating the first value based on the input vector.
3 . The system of claim 2 , wherein the one or more transactions are selectively and recursively omitted from the plurality of transactions.
4 . The system of claim 3 , wherein the weighted reason vector provides a set of top contributor variables in the input vector and latent features of the first analytic model that explain the first value.
5 . The system of claim 3 , wherein responsive to the one or more transactions being selectively and recursively omitted from the plurality of transactions, the maximal effect of the one or more transactions on the first value, the weighted reason vector and the plurality of ranked reasons is determined based on a quantum of change in the first value.
6 . The system of claim 5 , wherein an importance measure is generated based on the recursive omission.
7 . The system of claim 6 , wherein the importance measure is a function of a change in the first value and a change in the weighted reason vector.
8 . The system of claim 6 , wherein the importance measure is used to determine the one or more transactions that have the maximal effect.
9 . The system of claim 6 , wherein an importance measure is determined based on the recursive omission, the importance measure quantifying an influence of at least one omitted transaction.
10 . The system of claim 9 , wherein the importance measure is a function of change in the first value and a change in energy associated with at least one of: the weighted reason vector, the plurality of ranked reasons, and a change in a rank associated with the weighted reason vector or the plurality of ranked reasons.
11 . The system of claim 1 , wherein at least one of the one or more transactions contributes to the one or more temporal events or transactions most responsible for the first value.
12 . The system of claim 5 , wherein the first analytic model generates a revised first value, and wherein the second analytic model generates the reason vector, and the associated plurality of ranked reasons based on the maximal effect of the omission of at least one of the one or more transactions.
13 . The system of claim 6 , wherein the importance measure is used for identifying and generating a set of temporal events of past transactions with a maximal importance measure.
14 . The system of claim 13 , further comprising a translation module for translating at least one set of temporal explanations or reason codes into a human-readable explanation of a reason for the first value.
15 . The system of claim 14 , wherein the reason codes are associated with at least a transaction or event that had a maximal contribution to at least one of: the first value, the weighted reason vector, and a ranked reason state.
16 . The system of claim 14 , wherein the output instance provides an output file that includes at least a transaction or event that had a maximal contribution to at least one of: the first value, the weighted reason vector, and a ranked reason state.
17 . A computer-implemented method for managing abnormal transactions in a communications network, the method comprising:
collecting, by a first computing system communicatively coupled to the communications network, entity data and determining an input vector based on a plurality of transactions associated with the entity data, wherein the plurality of transactions include one or more abnormal transactions involving potentially fraudulent or malicious activity, and
wherein the one or more data storage media devices are communicatively coupled to the communications network, the one or more data storage media devices storing the entity data communicated over the communications network,
wherein a first analytic model recursively generates a first value based on the input vector, the first value being updated by the first analytic model as one or more transactions are omitted from the plurality of transactions,
wherein a second analytic model generates a weighted reason vector and associated plurality of ranked reasons that provide an understanding of the first value based on a set of top contributor variables in the input vector and latent features of the first analytic model associated with the first value, and
wherein one or more transactions involving potentially fraudulent or malicious activity are identified and omitted from the plurality of transactions associated with the entity data to enable an early elimination of transactions having fraudulent or malicious nature, responsive to a quantification of a maximal effect on the first value recursively generated based on the input vector.
18 . The method of claim 17 , wherein the entity data represents the one or more transactions and an output instance comprising a set of one or more values is generated to identify one or more temporal events or transactions most responsible for generating the first value based on the input vector and wherein the one or more transactions are selectively and recursively omitted from the plurality of transactions.
19 . The method of claim 18 , wherein the weighted reason vector provides a set of top contributor variables in the input vector and latent features of the first analytic model that explain the first value.
20 . The method of claim 18 , wherein responsive to the one or more transactions being selectively and recursively omitted from the plurality of transactions, the maximal effect of the one or more transactions on the first value, the weighted reason vector and the plurality of ranked reasons is determined based on a quantum of change in the first value.
21 . A computer program product comprising a non-transitory machine-readable medium storing instructions that, when executed by at least one programmable processor, cause the at least one programmable processor to perform operations comprising:
collecting, by a first computing system communicatively coupled to the communications network, entity data and determining an input vector based on a plurality of transactions associated with the entity data, wherein the plurality of transactions include one or more abnormal transactions involving potentially fraudulent or malicious activity, and
wherein the one or more data storage media devices are communicatively coupled to the communications network, the one or more data storage media devices storing the entity data communicated over the communications network,
wherein a first analytic model recursively generates a first value based on the input vector, the first value being updated by the first analytic model as one or more transactions are omitted from the plurality of transactions,
wherein a second analytic model generates a weighted reason vector and associated plurality of ranked reasons that provide an understanding of the first value based on a set of top contributor variables in the input vector and latent features of the first analytic model associated with the first value, and
wherein one or more transactions involving potentially fraudulent or malicious activity are identified and omitted from the plurality of transactions associated with the entity data to enable an early elimination of transactions having fraudulent or malicious nature, responsive to a quantification of a maximal effect on the first value recursively generated based on the input vector.