System for constructing effective machine-learning pipelines with optimized outcomes
A system, apparatus and a method for constructing pipelines, including enumerate a subspace of valid integrated pipelines, collecting a set of metrics for each of the plurality of pipelines, reducing the set of metrics to a single metric, selecting a final integrated pipeline from among the valid integrate pipelines based on reduced metric.
1 . A method comprising:
enumerating a subspace of valid integrated machine learning pipelines, wherein each valid integrated machine learning pipeline integrates at least one distinct value-add via a respective machine learning pipeline transformation, wherein each valid integrated machine learning pipeline is a transformed version of same initial machine learning pipeline;
collecting a set of objective and side-effect metrics for each of the valid integrated machine learning pipelines;
reducing the set of objective and side-effect metrics to a single metric for each of the valid integrated machine learning pipelines;
ordering the valid integrated machine learning pipelines based on the single metric;
and selecting a final integrated machine learning pipeline from the valid integrated machine learning pipelines based on the final integrated machine learning pipeline having a top ranked single metric.
2 . The method according to claim 1 , wherein during the enumeration step, dynamically pruning machine learning pipelines from the subspace using preconditions, where pruning using preconditions includes removing any machine learning pipelines from the subspace having a metric that violates a metric range threshold.
3 . The method according to claim 1 , wherein during the enumerating, pruning machine learning pipelines from the subspace using a statically produced partial ordering of value-adds.
4 . The method according to claim 1 , wherein during the enumerating, pruning machine learning pipelines from the subspace based on a time threshold.
5 . The method according to claim 1 , wherein input data used to train machine learning pipelines considered for the subspace is multi-modal data from a plurality of sources, and
wherein machine learning pipelines are added to the subspace by relaxing upper and lower bounds of value-add metrics until a threshold number of machine learning pipelines is met.
6 . The method according to claim 1 , wherein the final integrated machine learning pipeline is selected based on the top ranked single metric being on a frontier of a Pareto frontier.
7 . The method according to claim 1 being cloud implemented.
8 . The method of claim 1 , wherein the subspace of valid integrated machine learning pipelines includes at least one valid integrated machine learning pipeline that is a direct transformation of the initial machine learning pipeline and at least one valid integrated machine learning pipeline that is a transformation of the direct transformation of the initial machine learning pipeline.
9 . The method of claim 1 , wherein machine learning pipeline transformations applied to valid integrated machine learning pipelines within the subspace include de-biasing, compression, and hyperparameter optimization.
10 . The method of claim 1 , wherein a first valid integrated machine learning pipeline is a first transformed version of the initial machine learning pipeline by a de-biasing transformation, wherein a second valid integrated machine learning pipeline is a second transformed version of the initial machine learning pipeline by a compression transformation, wherein a third valid integrated machine learning pipeline is a third transformed version of the initial machine learning pipeline by a compression transformation.
11 . A system comprising:
one or more processors; and
one or more computer-readable storage media collectively storing program instructions which, when executed by the one or more processors, are configured to cause the one or more processors to perform a method comprising:
enumerating a subspace of valid integrated machine learning pipelines, wherein each valid integrated machine learning pipeline integrates at least one distinct value-add via a respective machine learning pipeline transformation, wherein each valid integrated machine learning pipeline is a transformed version of same initial machine learning pipeline;
collecting a set of objective and side-effect metrics for each of the valid integrated machine learning pipelines;
reducing the set of objective and side-effect metrics to a single metric for each of the valid integrated machine learning pipelines;
ordering the valid integrated machine learning pipelines based on the single metric;
and selecting a final integrated machine learning pipeline from the valid integrated machine learning pipelines based on the final integrated machine learning pipeline having a top ranked single metric.
12 . The system according to claim 11 , wherein during the enumeration step, dynamically pruning machine learning pipelines from the subspace using preconditions, where pruning using preconditions includes removing any machine learning pipelines from the subspace having a metric that violates a metric range threshold.
13 . The system according to claim 11 , wherein during the enumerating, pruning machine learning pipelines from the subspace using a statically produced partial ordering of value-adds.
14 . The system according to claim 11 , wherein input data used to train machine learning pipelines considered for the subspace is multi-modal data from a plurality of sources, and
wherein machine learning pipelines are added to the subspace by relaxing upper and lower bounds of value-add metrics until a threshold number of machine learning pipelines is met.
15 . The system according to claim 11 , wherein the set of objective and side-effect metrics for each valid integrated machine learning pipeline are normalized yielding a tuple of metric values for each valid integrated machine learning pipeline, wherein the set of objective and side-effect metrics are reduced to the single metric by computing a geometric mean over the tuple of metric values for each valid integrated machine learning pipeline.
16 . A computer program product comprising one or more computer readable storage media, and program instructions collectively stored on the one or more computer readable storage media, the program instructions comprising instructions configured to cause one or more processors to perform a method comprising:
enumerating a subspace of valid integrated machine learning pipelines, wherein each valid integrated machine learning pipeline integrates at least one distinct value-add via a respective machine learning pipeline transformation, wherein each valid integrated machine learning pipeline is a transformed version of same initial machine learning pipeline;
collecting a set of objective and side-effect metrics for each of the valid integrated machine learning pipelines;
reducing the set of objective and side-effect metrics to a single metric for each of the valid integrated machine learning pipelines;
ordering the valid integrated machine learning pipelines based on the single metric;
and selecting a final integrated machine learning pipeline from the valid integrated machine learning pipelines based on the final integrated machine learning pipeline having a top ranked single metric.
17 . The computer program product according to claim 16 , wherein during the enumeration step, dynamically pruning machine learning pipelines from the subspace using preconditions, where pruning using preconditions includes removing any machine learning pipelines from the subspace having a metric that violates a metric range threshold.
18 . The computer program product according to claim 16 , wherein during the enumerating, pruning machine learning pipelines from the subspace using a statically produced partial ordering of value-adds.
19 . The computer program product according to claim 16 , wherein input data used to train machine learning pipelines considered for the subspace is multi-modal data from a plurality of sources, and
wherein machine learning pipelines are added to the subspace by relaxing upper and lower bounds of value-add metrics until a threshold number of machine learning pipelines is met.
20 . The computer program product according to claim 16 , wherein the set of objective and side-effect metrics for each valid integrated machine learning pipeline are normalized yielding a tuple of metric values for each valid integrated machine learning pipeline, wherein the set of objective and side-effect metrics are reduced to the single metric by computing a weighted average over the tuple of metric values for each valid integrated machine learning pipeline.