BlueEdge: Edge computing-based framework for efficient text matching and duplicate detection
Year: 2026
Type: article
Abstract: Despite advances in NLP libraries and cloud-based pipelines, memory usage, latency, network connectivity, and privacy issues still make it difficult to use these tools directly on edge and mobile devices. This paper introduces BlueEdge, a memory-bound string-matching system for text normalization and name-based duplicate detection, with a proposed mobile-edge architecture comprising NLTK (Natural Language Toolkit) preprocessing, Kivy (Python GUI Library) user interaction, and Firebase (Cloud-based Database) synchronization. The algorithm's main contribution is an optimized implementation of the Levenshtein distance that stores only two rows of the matrix at any given time, reducing space complexity to O(min(m, n)). For the reported benchmarking protocol, the space requirement is measured by the peak memory footprint, which is below 9 KB for the dependency-free core comparison algorithm. Evaluation of BlueEdge is based on a real-world-informed benchmark dataset that was synthetically expanded to enable controlled, repeatable evaluation across six common name-matching error types: spelling variation, misspelling, name abbreviation, honorific prefix, common nickname, and split-name errors. The error taxonomy is derived from an initial collection of real-world errors from the previous BlueEdge prototype, and the extended benchmark provides well-balanced coverage of duplicate, non-duplicate, Arabic-script, and hard-negative cases. These hard negatives involve cases where two distinct identities share the same full name by chance, as well as cases where distinct identities share the same family name, to evaluate performance beyond trivially separable duplicate pairs. A lightweight classification model based on a Levenshtein-based edit-distance feature and five complementary similarity measures, with a leakage-aware group-based evaluation protocol, achieves an F1 score of 0.988 ± 0.006 in 10-fold cross-validation. A 50-tree ensemble achieves accuracy statistically indistinguishable from that of the full 200-tree model across 50 paired fold-seed evaluations (Wilcoxon p = 1.000), while reducing inference cost. In addition, augmenting the engineered feature set with multilingual Sentence-BERT does not improve the mean cross-validated F1 score; a separate paired held-out comparison also finds no statistically significant difference (McNemar exact p = 1.000). These results indicate that engineered similarity features remain competitive for this short, structured name-matching task under edge-oriented constraints. The experiments also show a limitation of text-only duplicate detection: names with accidental or familial resemblance cannot be reliably resolved without secondary non-textual verification. Overall, the findings support BlueEdge as a validated, reproducible framework for research and development of memory-efficient duplicate-detection methods and point to full on-device deployment and larger-scale evaluation in real-world settings as future research directions.
Source: Array
Institution: Mansoura University
Language: English
Topic: Data Quality and Management
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $1,680
OpenAlex ID: W7212373660