Abstract:
The proliferation of web data and the advent of extensive knowledge graphs have led to the creation
of vast volumes of disconnected data, resulting in data silos. Integrating and sharing web data across diverse
domains require that equivalent instances across different sources are correctly identified. Instance matching,
often referred to as Entity Resolution, encompasses the task of determining whether two instances correspond
to same resource or entity. This process poses significant challenges, particularly in distinguishing between
identical entities and those with similar attributes. Candidate generation has a pivotal role in facilitating
appropriate comparisons between entities across disparate datasets. This paper employs an inverted index
based approach to identify candidates for the matching task and a query likelihood model based selection to
further reduce the candidate set. This paper proposes a novel system architecture employing hybrid ensemble
classifiers and a methodology for identifying equivalent instances despite the challenges posed by diverse data
representations in instance matching. Through experimental evaluation on real-world datasets, we demonstrate
that our hybrid ensemble learning approach consistently outperforms standalone matchers in terms of accuracy
and F1score. A comprehensive literature review on instance matching, discussing practical considerations and
challenges for data interlinking is also presented here. Future research directions aimed at contributing to
seamless data integration and knowledge sharing across disparate domains are also outlined.
