A small library for loading and downloading relational datasets.

pip install relational-datasets

This API and the datasets at https://github.com/srlearn/datasets/ are currently being experimented with.

• Prefer Julia? Check out RelationalDatasets.jl.

Open enhancements and bugs are tracked here:

• Issues: relational-datasets package
• Issues: datasets

But here is a short-term Roadmap:

• Modes: srlearn/datasets: Issue 11
• Converting propositional->relational
  • Problem Settings
    • Binary Classification
      • Classification: (0, 1)
      • Classification: (-1, 1)
      • Classification: maybe recommend sklearn.preprocessing.LabelBinarizer
    • Regression
      • Regression: y ∈ float
    • Multiclass Classification: When target is int and in [0, 1, 2, ...]
  • Categorical datatype support in X matrix.
  • Dataframes: pandas

Running the fetch method downloads a version of a datset to your local cache:

import relational_datasets relational_datasets.fetch("toy_cancer") relational_datasets.fetch("toy_father", "v0.0.3") relational_datasets.fetch("cora")

Resulting in:

~/relational_datasets/ ├── toy_cancer_v0.0.4.zip <--- latest ├── toy_father_v0.0.3.zip <--- specific version └── cora_v0.0.4.zip <--- latest

The load method returns train and test folds—each with pos, neg, and facts. Internally it uses fetch, so it will automatically download a dataset if it is not available.

For example: "Load fold-2 of webkb"

from relational_datasets import load train, test = load("webkb", "v0.0.4", fold=2) len(train.facts) # 1344

The relational_datasets.convert module has functions for converting vector-based datasets into relational/ILP-style datasets:

When y is a vector of 0/1

from relational_datasets.convert import from_numpy import numpy as np data, modes = from_numpy( np.array([[0, 1, 1], [0, 1, 2], [1, 2, 2]]), np.array([0, 0, 1]), ) data, modes

(RelationalDataset(pos=['v4(id3).'], neg=['v4(id1).', 'v4(id2).'], facts=['v1(id1,0).', 'v1(id2,0).', 'v1(id3,1).', 'v2(id1,1).', 'v2(id2,1).', 'v2(id3,2).', 'v3(id1,1).', 'v3(id2,2).', 'v3(id3,2).']), ['v1(+id,#varv1).', 'v2(+id,#varv2).', 'v3(+id,#varv3).', 'v4(+id).'])

When y is a vector of floats

from relational_datasets.convert import from_numpy import numpy as np data, modes = from_numpy( np.array([[0, 1, 1], [0, 1, 2], [1, 2, 2]]), np.array([1.1, 0.9, 2.5]), ) data, modes

(RelationalDataset(pos=['regressionExample(v4(id1),1.1).', 'regressionExample(v4(id2),0.9).', 'regressionExample(v4(id3),2.5).'], neg=[], facts=['v1(id1,0).', 'v1(id2,0).', 'v1(id3,1).', 'v2(id1,1).', 'v2(id2,1).', 'v2(id3,2).', 'v3(id1,1).', 'v3(id2,2).', 'v3(id3,2).']), ['v1(+id,#varv1).', 'v2(+id,#varv2).', 'v3(+id,#varv3).', 'v4(+id).'])

load_breast_cancer is based on the Breast Cancer Wisconsin dataset.

Here we: (1) load the data and class labels, (2) split into training and test sets, (3) bin the continuous features to discrete, and (4) convert to the relational format.

from sklearn.datasets import load_breast_cancer from sklearn.model_selection import train_test_split from sklearn.preprocessing import KBinsDiscretizer # (1) Load X, y = load_breast_cancer(return_X_y=True) # (2) Split X_train, X_test, y_train, y_test = train_test_split(X, y) # (3) Discretize disc = KBinsDiscretizer(n_bins=5, encode="ordinal") X_train = disc.fit_transform(X_train) X_test = disc.transform(X_test) X_train = X_train.astype(int) X_test = X_test.astype(int) # (4) Convert from relational_datasets.convert import from_numpy train, modes = from_numpy(X_train, y_train) test, _ = from_numpy(X_test, y_test)

pip install relational-datasets

git clone https://github.com/srlearn/relational-datasets.git cd relational-datasets pip install -e .

• Alexander Hayes - Indiana University, Bloomington

This package was partially based on datasets from the Starling Lab Datasets Collection, which included specific contributions by Harsha Kokel and Devendra Singh Dhami. Tushar Khot converted many to the ILP format from Alchemy 2 format, but that occurred before versions were tracked. Some inspiration was drawn from the "RelationalDatasets" list that Jonas Schouterden collected.