Human-Centered Clustering for Time Series Data - Archive ouverte HAL Accéder directement au contenu
Communication Dans Un Congrès Année : 2021

Human-Centered Clustering for Time Series Data

Angela Bonifati
Raymond Ng
  • Fonction : Auteur

Résumé

Clustering is a fundamental step of several data science pipelines leading to compute groups of objects with high similarity. In this paper, we focus on the problem of clustering time series and we show how we can make this process human-centered. In particular, by relying on FeatTS, a feature-based semi-supervised framework, we show how parameters can be tuned by the users in an intuitive fashion. The considered parameters are: (1) the learning threshold leading to choose how much labeled data are needed as input; (2) the cutting threshold leading to prune time series whose distance is below this threshold; (3) the number of clusters desired as output leading to create different clusters than the number of classes reflected by the labeled data. In particular, while the first two parameters allow to tune the amount of the supervision and tweak the number of significant features, the third parameter leads to showcase the robustness of the clustering process with respect to the number of clusters provided as input. Finally, the significance of the entire multi-step clustering pipeline is empirically demonstrated through a handful of ablation tests.
Fichier principal
Vignette du fichier
DashAccepted (1).pdf (674.54 Ko) Télécharger le fichier

Dates et versions

hal-03548284 , version 1 (30-01-2022)

Identifiants

  • HAL Id : hal-03548284 , version 1

Citer

Donato Tiano, Angela Bonifati, Raymond Ng. Human-Centered Clustering for Time Series Data. 3rd Workshop on Data Science with Human in the Loop @ KDD 2021, Aug 2021, Singapore (virtual), Singapore. ⟨hal-03548284⟩
60 Consultations
116 Téléchargements

Partager

Gmail Facebook X LinkedIn More