Characterizing failures of a machine learning model based on instance features
The present disclosure relates to systems, methods, and computer readable media that evaluate performance of a machine learning system in connection with a test dataset. For example, systems disclosed herein may receive a test dataset and identify label information for the test dataset including feature information and ground truth data. The systems disclosed herein can compare the ground truth data and outputs generated by a machine learning system to evaluate performance of the machine learning system with respect to the test dataset. The systems disclosed herein may further generate feature clusters based on failed outputs and corresponding features and generate a number of performance views that illustrate performance of the machine learning system with respect to clustered groupings of the test dataset.
1. A method, comprising:
receiving a plurality of outputs for a test dataset, the test dataset comprising a plurality of test instances;
identifying label information associated with the test dataset, the label information comprising feature labels associated with respective instances from the plurality of test instances and ground truth data corresponding to expected outputs of a machine learning system with respect to the plurality of test instances;
comparing the label information and the plurality of outputs to identify a set of error labels associated with a subset of outputs from the plurality of outputs that are inaccurate with respect to corresponding label information by identifying combinations of one or more feature labels having a higher correlation to the set of error labels than other combinations of one or more feature labels and grouping the test instances in accordance with the identified combinations of one or more feature labels having the higher correlation to the set of error labels; and
generating a plurality of feature clusters comprising groupings of test instances from the plurality of test instances based on the feature labels and the set of error labels.
2. The method of claim 1 , further comprising generating a plurality of performance views associated with performance of the machine learning system with respect to the plurality of feature clusters.
3. The method of claim 2 , wherein generating the plurality of performance views comprises:
generating a global performance view including an indication of performance of the machine learning system with respect to each feature cluster from the plurality of feature clusters; and
generating a plurality of cluster views including, for each feature cluster of the plurality of feature clusters, a multi-branch view representative of performance of the machine learning system with respect to two or more subsets of test instances from the feature cluster.
4. The method of claim 1 , wherein the machine learning system comprises a first machine learning model trained to generate a first output for the given instance and a second machine learning model trained to generate a second output for the given instance, wherein the output for the given instance is based on a combination of the first output and the second output.
5. The method of claim 4 , wherein comparing the label information and the plurality of outputs further comprises comparing the label information to a first plurality of outputs corresponding to the first machine learning model and a second plurality of outputs corresponding to the second machine learning model, and wherein the method further comprises generating a plurality of quality metrics for the plurality of outputs indicative of performances of the first machine learning model and the second machine learning model.
6. The method of claim 1 , wherein identifying label information comprises applying an augmented feature model trained to identify features from a set of instances to the plurality of test instances to generate the feature labels based on features identified in connection with respective test instances of the plurality of test instances.
7. The method of claim 1 , wherein the output for the given instance comprises a classification of a digital media content item, the classification comprising one or more of a predicted user rating for the digital media content item, an identification of objects within the digital media content item, a count of objects within the digital media content item, a description of content within the digital media content item, or an indication of quality of the digital media content item.
8. The method of claim 1 , wherein identifying the combinations of one or more feature labels having the higher correlation to the set of error labels is based on a minimum cluster constraint in which each grouping of test instances has a number of test instances greater than or equal to a minimum threshold number of test instances.
9. The method of claim 1 , wherein generating the plurality of feature clusters comprises applying a clustering model trained to identify groupings of test instances from a collection of test instances based on groupings of features associated with output failures generated by a given machine learning model.
10. The method of claim 1 , further comprising:
identifying, from the plurality of feature clusters, a feature cluster of interest based on a rate of error labels associated with a corresponding grouping of test instances for the feature cluster of interest; and
refining the machine learning system by selectively collecting additional training data including a plurality of training instances based on associated label information, wherein the associated label information includes one or more feature labels associated with the identified feature cluster of interest.
11. The method of claim 10 , wherein selectively collecting additional training data comprises sampling a subset of training instances from a collection of training instances based on the label information associated with the subset of training instances including the one or more feature labels associated with the identified feature cluster of interest.
12. A system, comprising:
one or more processors;
memory in electronic communication with the one or more processors; and
instructions stored in the memory, the instructions being executable by the one or more processors to:
receive a plurality of outputs for a test dataset, the test dataset comprising a plurality of test instances;
identify label information associated with the test dataset, the label information comprising feature labels associated with respective instances from the plurality of test instances and ground truth data corresponding to expected outputs of a machine learning system with respect to the plurality of test instances;
compare the label information and the plurality of outputs to identify a set of error labels associated with a subset of outputs from the plurality of outputs that are inaccurate with respect to corresponding label information by identifying combinations of one or more feature labels having a higher correlation to the set of error labels than other combinations of one or more feature labels and grouping the test instances in accordance with the identified combinations of one or more feature labels having the higher correlation to the set of error labels; and
generate a plurality of feature clusters comprising groupings of test instances from the plurality of test instances based on the feature labels and the set of error labels.
13. The system of claim 12 , further comprising instructions being executable by the one or more processors to generate a plurality of performance views associated with performance of the machine learning system with respect to the plurality of feature clusters.
14. The system of claim 12 , wherein identifying label information comprises:
receiving label information previously associated with the test dataset; and
augmenting the label information by applying an augmented feature model trained to identify features associated with a given set of instances to the plurality of test instances to identify one or more additional features associated with the plurality of test instances not previously included within the received label information.
15. The system of claim 12 , further comprising instructions being executable by the one or more processors to:
identify, from the plurality of feature clusters, a feature cluster of interest based on a rate of error labels associated with a corresponding grouping of test instances for the feature cluster of interest; and
refine the machine learning system by selectively collecting additional training data including a plurality of training instances based on associated label information, wherein the associated label information includes one or more feature labels associated with the identified feature cluster of interest.
16. A method, comprising:
receiving a plurality of outputs for a test dataset, the test dataset comprising a plurality of test instances;
identifying label information associated with the test dataset, the label information comprising feature labels associated with respective instances from the plurality of test instances and ground truth data corresponding to expected outputs of a machine learning system with respect to the plurality of test instances, wherein identifying label information includes receiving the ground truth data based on information provided by individuals when presented with the plurality of test instances;
comparing the label information and the plurality of outputs to identify a set of error labels associated with a subset of outputs from the plurality of outputs that are inaccurate with respect to corresponding label information; and
generating a plurality of feature clusters comprising groupings of test instances from the plurality of test instances based on the feature labels and the set of error labels.
17. The method of claim 16 , further comprising generating a plurality of performance views associated with performance of the machine learning system with respect to the plurality of feature clusters.
18. The method of claim 16 , wherein the machine learning system comprises a first machine learning model trained to generate a first output for the given instance and a second machine learning model trained to generate a second output for the given instance, wherein the output for the given instance is based on a combination of the first output and the second output.
19. The method of claim 16 , wherein the output for the given instance comprises a classification of a digital media content item, the classification comprising one or more of a predicted user rating for the digital media content item, an identification of objects within the digital media content item, a count of objects within the digital media content item, a description of content within the digital media content item, or an indication of quality of the digital media content item.
20. The method of claim 16 , wherein generating the plurality of feature clusters comprises applying a clustering model trained to identify groupings of test instances from a collection of test instances based on groupings of features associated with output failures generated by a given machine learning model.