Adding new auto intent resolving model from the research project

[jzhang-gp]

Description

Adding the latest intent resolving code for additional intent type.

One weird finding on the following test. Both columns are columns with distinct values. First column is from 0 to 99 while the second is from 1 to 100. The first column is treated as Numeric but the second is treated as Droppable. Is this because the ID data usually starts from 1? @christeefy

def test_autointentmapping(step):
    """Test intents automatically mapped for a PreparerStep subclass."""
    import pandas as pd
    import numpy as np

    df = pd.DataFrame(
        [np.arange(i, i + 2) for i in range(100)], columns=["1", "2"]
    )
    step.get_mapping(df)
    assert step.cache_manager["intent", "1"] == "Numeric"
    # TODO this part DOES NOT make sense. Shouldn't both intents be Droppable?
    #  For now I'm temporarily change it to one Numeric and one Droppable to
    #  pass the test. We need to revisit this on the auto intent resolving
    #  side.
    assert step.cache_manager["intent", "2"] == "Droppable"

[christeefy]

Thanks for surfacing this bug, Jing.

I expanded on the test you had and it only fails [1 ... 100] and nothing else. From manual inspection, [1 ... 100] does not seem to appear any more frequently than [0 ... N] in the training data.

Interestingly, I was able to control the resolver's intent prediction simply by changing the "min" metafeature from 1 (Droppable) to not 1 (Numerical). Given this observation, I will include additional training data and retrain the model to tackle this specific issue.

[jzhang-gp]

Ping :)

[jichaoz]

Thanks for surfacing this bug, Jing.

I expanded on the test you had and it only fails [1 ... 100] and nothing else. From manual inspection, [1 ... 100] does not seem to appear any more frequently than [0 ... N] in the training data.

Interestingly, I was able to control the resolver's intent prediction simply by changing the "min" metafeature from 1 (Droppable) to not 1 (Numerical). Given this observation, I will include additional training data and retrain the model to tackle this specific issue.

Is this issue solved?

[christeefy]

Not yet, Ji Chao. I am free to work on it today, along with the dependency upgrade for pandas and scikit-learn

[christeefy]

[Update] Submitted a PR at the automl_research repo to address this issue. We can update this PR once that automl_research PR is approved. :)

[jzhang-gp]

[Update] Submitted a PR at the automl_research repo to address this issue. We can update this PR once that automl_research PR is approved. :)

PR approved with a minor comment :)