Um Arcabouço Incremental de Resolução de Entidades Multivisão para Perfis Dinâmicos de Usuários Acadêmicos em Ambientes de Big Data
Year: 2026
Type: article
Abstract: Entity Resolution (ER) in streaming Big Data environments demands dynamic, low-latency indexing architectures capable of resolving high-dimensional user profiles without incurring the computational overhead of complete index reconstruction. Conventional tree-based incremental indexing systems are fundamentally constrained by their inability to operate over semantic vector spaces, while Locality-Sensitive Hashing (LSH) blocking paradigms exhibit pronounced sensitivity to hash collisions, generating noisy candidate blocks that degrade downstream matching quality. This article proposes a unified, real-time incremental entity resolution framework engineered for multi-view academic user profiles. The proposed architecture integrates three synergistic components: (i) an incremental LSH-SimHash blocking layer that projects multi-view profile representations into candidate buckets while preserving high-dimensional angular relationships within a compact binary signature space; (ii) independent Hierarchical Navigable Small World (HNSW) graphs maintained within individual LSH buckets, functioning as localized approximate nearest neighbor (ANN) search engines that eliminate global structural reconstruction upon profile arrivals; and (iii) an Adaptive Candidate Selection (ACS) strategy that prunes retrieved neighbor sets using a multi-criteria scoring function jointly evaluating vector cosine similarity, graph neighborhood structural consistency, and cross-view semantic alignment. Comprehensive empirical evaluations conducted across three large-scale benchmark datasets—the Academic Profiles Dataset, the North Carolina Voter Registration Dataset, and the OZ Synthetic Dataset—demonstrate that the proposed framework achieves a peak F1-score of 96.84% while maintaining sub-4 ms end-to-end query resolution latencies and a stable, predictable memory footprint, consistently outperforming established tree-based and dynamic approximate blocking baselines across all evaluation dimensions.
Source: Revista Coleta Científica
Institution: University of Boumerdes
Language: English
Topic: Data Quality and Management
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7217137030