Towards Automated Algorithm Selection for Link Prediction

EasyChair Preprint 13986

Abstract

Link prediction represents an important field within network science which involves the systematic (i.e. algorithmic) prediction of missing edges within a graph-based representation. In this study, the seminal algorithm selection problem is formally proposed and formulated for the domain of link prediction so as to facilitate the automated selection of algorithms. More specifically, a regression-based meta-learning approach is proffered, the aim of which is to approximate the relationship between algorithmic performance and graph-based features, thereby facilitating data-driven algorithm selection. The study’s contributions include an appropriate formalisation of the algorithm selection problem for link prediction, together with the extraction of informative graph-based features as well as the generation of algorithmic performance in respect of various prominent link prediction approaches. A suitable meta-learner is trained with respect to the aforementioned meta-data in order to induce automated algorithm selection. Feature importance is also carried out so as to identify pertinent graph-based features in respect of the predictive task at hand. It may be inferred from the results that the meta-learner showcases admirable predictive capabilities in respect of diverse network data sets. Decision support in respect of link prediction algorithm selection may be induced — a novel and significant contribution to the domain of link prediction.

Keyphrases: algorithm selection, link prediction, meta-learning

@booklet{EasyChair:13986,
  author    = {Lienke Brown and Stephan Nel},
  title     = {Towards Automated Algorithm Selection for Link Prediction},
  howpublished = {EasyChair Preprint 13986},
  year      = {EasyChair, 2024}}