Skip to main content

Meta-feature Extractor

Project description

Build Status codecov Documentation Status PythonVersion Pypi Black

pymfe: Python Meta-Feature Extractor

The pymfe (python meta-feature extractor) provides a comprehensive set of meta-features implemented in python. The package brings cutting edge meta-features, following recent literature propose. The pymfe architecture was thought to systematically make the extraction, which can produce a robust set of meta-features. Moreover, pymfe follows recent meta-feature formalization aiming to make MtL reproducible.

Here, you can use different measures and summary functions, setting their hyperparameters, and also measuring automatically the elapsed time. Moreover, you can extract meta-features from specific models, or even extract meta-features with confidence intervals using bootstrap. There are a lot of other interesting features and you can see more about it looking at the documentation.

Meta-feature

In the Meta-learning (MtL) literature, meta-features are measures used to characterize data sets and/or their relations with algorithm bias.

"Meta-learning is the study of principled methods that exploit meta-knowledge to obtain efficient models and solutions by adapting the machine learning and data mining process." - (Brazdil et al. (2008))

Meta-features are used in MtL and AutoML tasks in general, to represent/understand a dataset, to understanding a learning bias, to create machine learning (or data mining) recommendations systems, and to create surrogates models, to name a few.

Pinto et al. (2016) and Rivolli et al. (2018) defined a meta-feature as follows. Let $D \in \mathcal{D}$ be a dataset, $m\colon \mathcal{D} \to \mathbb{R}^{k'}$ be a characterization measure, and $\sigma\colon \mathbb{R}^{k'} \to \mathbb{R}^{k}$ be a summarization function. Both $m$ and $\sigma$ have also hyperparameters associated, $h_m$ and $h_\sigma$ respectively. Thus, a meta-feature $f\colon \mathcal{D} \to \mathbb{R}^{k}$ for a given dataset $D$ is

$$ f\big(D\big) = \sigma\big(m(D,h_m), h_\sigma\big). $$

The measure $m$ can extract more than one value from each data set, i.e., $k'$ can vary according to $D$, which can be mapped to a vector of fixed length $k$ using a summarization function $\sigma$.

In this package, We provided the following meta-features groups:

  • General: General information related to the dataset, also known as simple measures, such as the number of instances, attributes and classes;
  • Statistical: Standard statistical measures to describe the numerical properties of data distribution;
  • Information-theoretic: Particularly appropriate to describe discrete (categorical) attributes and their relationship with the classes;
  • Model-based: Measures designed to extract characteristics from simple machine learning models;
  • Landmarking: Performance of simple and efficient learning algorithms.
    • Relative Landmarking: Relative performance of simple and efficient learning algorithms;
    • Subsampling Landmarking: Performance of simple and efficient learning algorithms from a subsample of the dataset;
  • Clustering: Clustering measures extract information about dataset based on external validation indexes;
  • Concept: Estimate the variability of class labels among examples and the examples density;
  • Itemset: Compute the correlation between binary attributes; and
  • Complexity: Estimate the difficulty in separating the data points into their expected classes.

In the pymfe package, you can use different measures and summary functions, setting their hyperparameters, and automatically measure the elapsed time. Moreover, you can extract meta-features from specific models, or even obtain meta-features with confidence intervals using bootstrap. There are many other exciting features. You can see more about it looking at the documentation.

Dependencies

The main pymfe requirement is:

  • Python (>= 3.6)

Installation

The installation process is similar to other packages available on pip:

pip install -U pymfe

It is possible to install the development version using:

pip install -U git+https://github.com/ealcobaca/pymfe

or

git clone https://github.com/ealcobaca/pymfe.git
cd pymfe
python setup.py install

Example of use

The simplest way to extract meta-features is by instantiating the MFE class. It computes five meta-features groups by default using mean and standard deviation as summary functions: General, Statistical, Information-theoretic, Model-based, and Landmarking. The fit method can be called by passing the X and y. Then the extract method is used to extract the related measures. A simple example using pymfe for supervised tasks is given next:

# Load a dataset
from sklearn.datasets import load_iris
from pymfe.mfe import MFE

data = load_iris()
y = data.target
X = data.data

# Extract default measures
mfe = MFE()
mfe.fit(X, y)
ft = mfe.extract()
print(ft)

# Extract general, statistical and information-theoretic measures
mfe = MFE(groups=["general", "statistical", "info-theory"])
mfe.fit(X, y)
ft = mfe.extract()
print(ft)

# Extract all available measures
mfe = MFE(groups="all")
mfe.fit(X, y)
ft = mfe.extract()
print(ft)

You can simply omit the target attribute for unsupervised tasks while fitting the data into the MFE model. The pymfe package automatically finds and extracts only the metafeatures suitable for this type of task. Examples are given next:

# Load a dataset
from sklearn.datasets import load_iris
from pymfe.mfe import MFE

data = load_iris()
y = data.target
X = data.data

# Extract default unsupervised measures
mfe = MFE()
mfe.fit(X)
ft = mfe.extract()
print(ft)

# Extract all available unsupervised measures
mfe = MFE(groups="all")
mfe.fit(X)
ft = mfe.extract()
print(ft)

Several measures return more than one value. To aggregate the returned values, summarization function can be used. This method can compute min, max, mean, median, kurtosis, standard deviation, among others. The default methods are the mean and the sd. Next, it is possible to see an example of the use of this method:

## Extract default measures using min, median and max 
mfe = MFE(summary=["min", "median", "max"])
mfe.fit(X, y)
ft = mfe.extract()
print(ft)
                          
## Extract default measures using quantile
mfe = MFE(summary=["quantiles"])
mfe.fit(X, y)
ft = mfe.extract()
print(ft)

You can easily list all available metafeature groups, metafeatures, summary methods and metafeatures filtered by groups of interest:

from pymfe.mfe import MFE

# Check all available meta-feature groups in the package
print(MFE.valid_groups())

# Check all available meta-features in the package
print(MFE.valid_metafeatures())

# Check available meta-features filtering by groups of interest
print(MFE.valid_metafeatures(groups=["general", "statistical", "info-theory"]))

# Check all available summary functions in the package
print(MFE.valid_summary())

It is possible to pass custom arguments to every metafeature using MFE extract method kwargs. The keywords must be the target metafeature name, and the value must be a dictionary in the format {argument: value}, i.e., each key in the dictionary is a target argument with its respective value. In the example below, the extraction of metafeatures min and max happens as usual, but the metafeatures sd, nr_norm and nr_cor_attr will receive user custom argument values, which will interfere in each metafeature result.

# Extract measures with custom user arguments
mfe = MFE(features=["sd", "nr_norm", "nr_cor_attr", "min", "max"])
mfe.fit(X, y)
ft = mfe.extract(
    sd={"ddof": 0},
    nr_norm={"method": "all", "failure": "hard", "threshold": 0.025},
    nr_cor_attr={"threshold": 0.6},
)
print(ft)

If you want to extract metafeatures from a pre-fitted machine learning model (from sklearn package), you can use the extract_from_model method without needing to use the training data:

import sklearn.tree
from sklearn.datasets import load_iris
from pymfe.mfe import MFE

# Extract from model
iris = load_iris()
model = sklearn.tree.DecisionTreeClassifier().fit(iris.data, iris.target)
extractor = MFE()
ft = extractor.extract_from_model(model)
print(ft)

# Extract specific metafeatures from model
extractor = MFE(features=["tree_shape", "nodes_repeated"], summary="histogram")

ft = extractor.extract_from_model(
    model,
    arguments_fit={"verbose": 1},
    arguments_extract={"verbose": 1, "histogram": {"bins": 5}})

print(ft)

You can also extract your metafeatures with confidence intervals using bootstrap. Keep in mind that this method extracts each metafeature several times, and may be very expensive depending mainly on your data and the number of metafeature extract methods called.

# Extract metafeatures with confidence interval
mfe = MFE(features=["mean", "nr_cor_attr", "sd", "max"])
mfe.fit(X, y)

ft = mfe.extract_with_confidence(
    sample_num=256,
    confidence=0.99,
    verbose=1,
)

print(ft)

Documentation

We write a great Documentation to guide you on how to use the pymfe library. You can find in the documentation interesting pages like:

Developer notes

License

This project is licensed under the MIT License - see the License file for details.

Cite Us

If you use the pymfe in scientific publication, we would appreciate citations to the following paper:

Edesio Alcobaça, Felipe Siqueira, Adriano Rivolli, Luís P. F. Garcia, Jefferson T. Oliva, & André C. P. L. F. de Carvalho (2020). MFE: Towards reproducible meta-feature extraction. Journal of Machine Learning Research, 21(111), 1-5.

You can also use the bibtex format:

@article{JMLR:v21:19-348,
  author  = {Edesio Alcobaça and
             Felipe Siqueira and
             Adriano Rivolli and
             Luís P. F. Garcia and
             Jefferson T. Oliva and
             André C. P. L. F. de Carvalho
  },
  title   = {MFE: Towards reproducible meta-feature extraction},
  journal = {Journal of Machine Learning Research},
  year    = {2020},
  volume  = {21},
  number  = {111},
  pages   = {1-5},
  url     = {http://jmlr.org/papers/v21/19-348.html}
}

Acknowledgments

We would like to thank every Contributor that directly or indirectly has make this project to happen. Thank you all.

References

  1. Brazdil, P., Carrier, C. G., Soares, C., & Vilalta, R. (2008). Metalearning: Applications to data mining. Springer Science and Business Media.
  2. Pinto, F., Soares, C., & Mendes-Moreira, J. (2016, April). Towards automatic generation of metafeatures. In Pacific-Asia Conference on Knowledge Discovery and Data Mining (pp. 215-226). Springer, Cham.
  3. Rivolli, A., Garcia, L. P. F., Soares, C., Vanschoren, J., and de Carvalho, A. C. P. L. F. (2018). Characterizing classification datasets: a study of meta-features for meta-learning. arXiv:1808.10406.

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

pymfe-0.4.3.tar.gz (138.2 kB view hashes)

Uploaded Source

Built Distribution

pymfe-0.4.3-py3-none-any.whl (156.0 kB view hashes)

Uploaded Python 3

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page