Relational Data Imputation with Graph Neural Networks - Archive ouverte HAL Accéder directement au contenu
Communication Dans Un Congrès Année : 2024

Relational Data Imputation with Graph Neural Networks

Riccardo Cappuzzo
  • Fonction : Auteur
  • PersonId : 1332367
Paolo Papotti

Résumé

Performing data analysis over incomplete data produces biased results and sub-par performance. Imputation over relational datasets that contain both categorical and continuous variables is challenging. The challenges are accentuated when the missingness proportion of dataset is high, wherein a large fraction of the relation contain missing values, or if missing values occur in multiple attributes of a single tuple. In this paper, we propose GRIMP, a novel approach for imputation that tackles these challenges. GRIMP achieves high imputation accuracy through a combination of three novel ideas. First, it represents relational data as a heterogeneous graph, encoding sophisticated relationships between tuples, attributes and cell values. Second, it uses graph representation learning based on message passing to combine and aggregate the representations from appropriate neighborhoods. This allows GRIMP to leverage information from other cell values of the same tuple and that of similar tuples for imputation. Finally, it uses a self-supervised multi-task learning paradigm for training imputation models. In other words, GRIMP does not need any explicit training data as it uses the existing relational data, even when it has missing values. GRIMP trains an imputation model for each attribute using a two-stage approach consisting of a task agnostic section, where the parameters are shared across all attributes, and an attribute specific imputation model. Experiments over ten datasets and seven baselines show that GRIMP performs accurate imputation and provides new insights about the limitations of data imputation systems.
Fichier principal
Vignette du fichier
publi-7503.pdf (1.01 Mo) Télécharger le fichier
Origine : Fichiers produits par l'(les) auteur(s)

Dates et versions

hal-04378971 , version 1 (08-01-2024)

Identifiants

  • HAL Id : hal-04378971 , version 1

Citer

Riccardo Cappuzzo, Saravanan Thirumuruganathan, Paolo Papotti. Relational Data Imputation with Graph Neural Networks. EDBT/ICDT 2024, 27th International Conference on Extending Database Technology, Mar 2024, Paestum, Italy. ⟨hal-04378971⟩

Collections

EURECOM ANR
101 Consultations
34 Téléchargements

Partager

Gmail Facebook X LinkedIn More