Contextual impact adjustment for machine learning models
Methods, systems, and computer program products for detecting contextual bias in text are provided herein. A computer-implemented method includes identifying, by a machine learning network, a protected attribute in one or more data samples; processing the identified data samples using a first sub-network of the machine learning network, wherein the first sub-network is configured to determine a plurality of contexts of the protected attribute across the identified data samples; determining an impact of each of the plurality of contexts on a second sub-network of the machine learning network, wherein the second sub-network of the machine learning network is configured to classify a given data sample into one of a plurality of classes; and adjusting the second sub-network of the machine learning to account for the impact of at least one of the plurality of contexts on the second sub-network.
1. A computer-implemented method, the method comprising:
identifying, by a machine learning network, a plurality of data samples comprising a protected attribute;
processing the identified data samples using a first sub-network of the machine learning network, wherein the first sub-network is configured to determine information corresponding to a plurality of contexts of the protected attribute across the identified data samples, wherein each of the plurality of contexts corresponds to a different meaning associated with the protected attribute;
determining respective impacts of the plurality of contexts on a second sub-network of the machine learning network, wherein the second sub-network of the machine learning network is configured to classify a given data sample into one of a plurality of classes; and
adjusting the second sub-network of the machine learning network to account for the impact corresponding to at least one of the plurality of contexts on the second sub-network;
wherein the method is carried out by at least one computing device.
2. The computer-implemented method of claim 1 , wherein the first sub-network is implemented as a residual connection in the machine learning network.
3. The computer-implemented method of claim 2 , wherein the first sub-network comprises a multi-head attention residual network.
4. The computer-implemented method of claim 1 , wherein the second sub-network comprises a feature learning neural network.
5. The computer-implemented method of claim 1 , comprising:
generating and outputting an explanation of the impact corresponding to the at least one of the plurality of contexts on the second sub-network.
6. The computer-implemented method of claim 1 , wherein each of the data samples comprises a sequence of text, and wherein the protected attribute corresponds to a portion of the sequence of text.
7. The computer-implemented method of claim 1 , wherein at least one of:
the relevance of different portions of the given one of the identified data samples to the portion of the given identified data sample is based on an attention computed for the given one of the identified data samples; and
the information comprises weights assigned by the first sub-network to the respective identified samples.
8. The computer-implemented method of claim 7 , wherein said processing comprises:
aggregating the information for the identified data samples based on at least one of: frequencies of types of associations and the assigned weights.
9. The computer-implemented method of claim 1 , wherein the machine learning network is trained on a first set of training data, and wherein said adjusting comprises:
retraining the machine learning network using a second set of training data that reduces the impact of at least one of the plurality of contexts on the second sub-network relative to the first set of training data.
10. The computer-implemented method of claim 1 , wherein software is provided as a service in a cloud environment.
11. A computer program product comprising a computer readable storage medium having program instructions embodied therewith, the program instructions executable by a computing device to cause the computing device to:
identify, by a machine learning network, a plurality of data samples comprising a protected attribute;
process the identified data samples using a first sub-network of the machine learning network, wherein the first sub-network is configured to determine information corresponding to a plurality of contexts of the protected attribute across the identified data samples, wherein each of the plurality of contexts corresponds to a different meaning associated with the protected attribute;
determine respective impacts of the plurality of contexts on a second sub-network of the machine learning network, wherein the second sub-network of the machine learning network is configured to classify a given data sample into one of a plurality of classes; and
adjust the second sub-network of the machine learning network to account for the impact corresponding to at least one of the plurality of contexts on the second sub-network.
12. The computer program product of claim 11 , wherein the first sub-network is implemented as a residual connection in the machine learning network.
13. The computer program product of claim 12 , wherein the first sub-network comprises a multi-head attention residual network.
14. The computer program product of claim 11 , wherein the second sub-network comprises a feature learning neural network.
15. The computer program product of claim 11 , wherein the program instructions executable by the computing device cause the computing device to:
generate and output an explanation corresponding to the impact of the at least one of the plurality of contexts on the second sub-network.
16. The computer program product of claim 11 , wherein each of the data samples comprises a sequence of text, and wherein the protected attribute corresponds to a portion of the sequence of text.
17. The computer program product of claim 11 , wherein at least one of:
the relevance of different portions of the given one of the identified data samples to the portion of the given identified data sample is based on an attention computed for the given one of the identified data samples; and
the information comprises weights assigned by the first sub-network to the respective identified samples.
18. The computer program product of claim 17 , wherein said processing comprises:
aggregating the information for the identified data samples based on at least one of: frequencies of types of associations and the assigned weights.
19. The computer program product of claim 11 , wherein the machine learning network is trained on a first set of training data, and wherein said adjusting comprises:
retraining the machine learning network using a second set of training data that reduces the impact of at least one of the plurality of contexts on the second sub-network relative to the first set of training data.
20. A system comprising:
a memory configured to store program instructions;
a processor operatively coupled to the memory to execute the program instructions to:
identify, by a machine learning network, a plurality of data samples comprising a protected attribute;
process the identified data samples using a first sub-network of the machine learning network, wherein the first sub-network is configured to determine information corresponding to a plurality of contexts of the protected attribute across the identified data samples, wherein each of the plurality of contexts corresponds to a different meaning associated with the protected attribute;
determine respective impacts of the plurality of contexts on a second sub-network of the machine learning network, wherein the second sub-network of the machine learning network is configured to classify a given data sample into one of a plurality of classes; and
adjust the second sub-network of the machine learning network to account for the impact corresponding to at least one of the plurality of contexts on the second sub-network.