Significance Analysis

This package is used to analyse datasets of different HPO-algorithms performing on multiple benchmarks, using a Linear Mixed-Effects Model-based approach.

Note

As indicated with the v0.x.x version number, Significance Analysis is early stage code and APIs might change in the future.

Documentation

For an interactive overview, please have a look at our example.

Every dataset should be a pandas dataframe of the following format:

As it is used to train a model, there can not be missing values, but duplicates are allowed. Our function dataset_validator checks for this format.

Installation

Using R, >=4.0.0 install packages: Matrix, emmeans, lmerTest and lme4

Using pip

pip install significance-analysis

Usage for significance testing

1. Generate data from HPO-algorithms on benchmarks, saving data according to our format.
2. Build a model with all interesting factors
3. Do post-hoc testing
4. Plot the results as CD-diagram

In code, the usage pattern can look like this:

import pandas as pd
from significance_analysis import dataframe_validator, model, cd_diagram


# 1. Generate/import dataset
data = dataframe_validator(pd.read_parquet("datasets/priorband_data.parquet"))

# 2. Build the model
mod = model("value ~ algorithm + (1|benchmark) + prior", data)

# 3. Conduct the post-hoc analysis
post_hoc_results = mod.post_hoc("algorithm")

# 4. Plot the results
cd_diagram(post_hoc_results)

Usage for hypothesis testing

Use the GLRT implementation or our prepared sanity checks to conduct LMEM-based hypothesis testing.

In code:

from significance_analysis import (
    dataframe_validator,
    glrt,
    model,
    seed_dependency_check,
    benchmark_information_check,
    fidelity_check,
)

# 1. Generate/import dataset
data = dataframe_validator(pd.read_parquet("datasets/priorband_data.parquet"))

# 2. Run the preconfigured sanity checks
seed_dependency_check(data)
benchmark_information_check(data)
fidelity_check(data)

# 3. Run a custom hypothesis test, comparing model_1 and model_2
model_1 = model("value ~ algorithm", data)
model_2 = model("value ~ 1", data)
glrt(model_1, model_2)

Usage for metafeature impact analysis

Analyzing the influence, a metafeature has on two algorithms performances.

In code:

from significance_analysis import dataframe_validator, metafeature_analysis

# 1. Generate/import dataset
data = dataframe_validator(pd.read_parquet("datasets/priorband_data.parquet"))

# 2. Run the metafeature analysis
scores = metafeature_analysis(data, ("HB", "PB"), "prior")

For more details and features please have a look at our example.

Contributing

We welcome contributions from everyone, feel free to raise issues or submit pull requests.

To cite the paper or code

@misc{geburek2024lmemsposthocanalysishpo,
title={LMEMs for post-hoc analysis of HPO Benchmarking},
author={Anton Geburek and Neeratyoy Mallik and Danny Stoll and Xavier Bouthillier and Frank Hutter},
year={2024},
eprint={2408.02533},
archivePrefix={arXiv},
primaryClass={cs.LG}
}