Skip to content

Active query strategies for semi-supervised clustering on top of scikit-learn and SciPy

License

Notifications You must be signed in to change notification settings

aymericb213/scikit-query

Repository files navigation

Documentation Status version Python codecov license Downloads

scikit-query

Clustering aims to group data into clusters without the help of labels, unlike classification algorithms. A well-known shortcoming of clustering algorithms is that they rely on an objective function geared toward specific types of clusters (convex, dense, well-separated), and hyperparameters that are hard to tune. Semi-supervised clustering mitigates these problems by injecting background knowledge in order to guide the clustering. Active clustering algorithms analyze the data to select interesting points to ask the user about, generating constraints that allow fast convergence towards a user-specified partition.

scikit-query is a library of active query strategies for constrained clustering inspired by scikit-learn and the now inactive active-semi-supervised-clustering library by Jakub Švehla.

It is focused on algorithm-agnostic query strategies, i.e. methods that do not rely on a particular clustering algorithm. From an input dataset, they produce a set of constraints by making insightful queries to an oracle. A variant for incremental constrained clustering is provided for applicable algorithms, taking a data partition into account.

In typical scikit way, the library is used by instanciating a class and using its fit method.

from skquery.pairwise import AIPC
from skquery.oracle import MLCLOracle

qs = AIPC()
oracle = MLCLOracle(truth=labels, budget=10)
constraints = qs.fit(dataset, oracle)

Algorithms

Algorithm Description Constraint type Works in incremental setting ? Source Date
Random sampling ML/CL, triplet ✔️
FFQS Neighborhood-based ML/CL ✔️ Basu et al. 2004
MMFFQS (MinMax) Neighborhood-based, similarity ML/CL ✔️ Mallapragada et al. 2008
NPU Neighborhood-based, information theory ML/CL ✔️ Xiong et al. 2013
SASC SVDD, greedy approach ML/CL Abin & Beigy 2014
AIPC Fuzzy clustering, information theory ML/CL Zhang et al. 2019

Dependencies

scikit-query is developed on Python >= 3.10, and requires the following libraries :

  • pandas>=2.0.1
  • matplotlib>=3.7.1
  • numpy>=1.24.3
  • scikit-learn>=1.2.2
  • cvxopt>=1.3.1
  • scikit-fuzzy>=0.4.2
  • scipy>=1.10.1
  • plotly>=5.14.1

Contributors

FFQS, MinMax and NPU are based upon Jakub Švehla's implementation. Other algorithms have been implemented by Aymeric Beauchamp or his students from the University of Orléans :

  • Salma Badri, Elis Ishimwe, Brice Jacquesson, Matthéo Pailler (2023)