Multiple Aggregates

[preethiraghavan1]

Description

Please include a summary of the change, the motivation, and any additional context that will help others understand your PR. If it closes one or more open issues, please tag them as described here.

Affected Dependencies

How has this been tested?

• Unit test

Checklist

• [ x] I have followed the Contribution Guidelines and Code of Conduct
• [ x] I have commented my code following the OpenMined Styleguide
• [x ] I have labeled this PR with the relevant Type labels
• [ x] My changes are covered by tests

[dvadym]

Thanks the approach is great! I left comments

[dvadym]

This is the right idea to add numbers to solve problem with duplicating labels!

Nit: f"Aggregation{i}"

Comment about adding numbers to label names: In BeamBackend such function was implemented with UniqueLabelGenerator class. But here it's simple enough, so I think the current approach to add numbers instead of using UniqueLabelGenerator makes sense.

[dvadym]

I've though in more details, in most use cases aggregations will share the same parameters (and also sharing the same parameters will help to optimize performance and utility of queries). Could you please

1. add argument params of type AggregateParams.

2.add argument partition_extractor_fn

Those arguments will be used in each aggregation

[dvadym]

It looks that we can have only one class Aggregate, w/o _Aggregate namely

1. aggregate_value returns self. aggregate_value saves in some member variable information about aggregations.
2. expand works as in _Aggregate

The advantage is that it will be simpler and no need to create multiple instances of _Aggregate. WDYT?

[dvadym]

Nice, this is a correct approach to generate dynamic tuples!

[dvadym]

Since all shared parameters will be provided in the constructor, there are just a few parameters that's needed value_extractor

[dvadym]

I'd suggest to use DPEngine.aggregate instead of PrivatePTransforms Mean, Sum, Count. The benefits are that in future we can optimize performance/utility by computing multiple aggregations with DPEngine.aggregate

[dvadym]

class Aggregate can aggregate only for one partition, but by multiple values.

For each value aggregation, we run DPEngine.aggregate we need to have AggregateParams:

class AggregateParams:
    metrics: Iterable[Metrics]
    max_partitions_contributed: int
    max_contributions_per_partition: int
    budget_weight: float = 1
    min_value: float = None
    max_value: float = None
    public_partitions: Any = None
    noise_kind: NoiseKind = NoiseKind.LAPLACE
    custom_combiners: Iterable['CustomCombiner'] = None

Some of those parameters are common for all values to aggregate (i.e. they needed to be specified in constructor in Aggregate), some of them specific for value (i.e. in aggregate_value arguments).

1. def __init__(self, partition_extractor, common_params:CommonAggregateParams, public_partitions=None)
2.def aggregate_value(self, value_extractor, value_params:ValueParams,output_col:str)

class CommonAggregateParams:  # name AggregateParams is already taken, maybe other name?
  max_partitions_contributed: int
  max_contributions_per_partition: int
  noise_kind: NoiseKind = NoiseKind.LAPLACE

class ValueParams:
    budget_weight: float = 1
    min_value: float = None
    max_value: float = None