Purpose: Data augmentation (DA) is increasingly used to overcome limited training samples and improve computational learning, yet its role in scientometrics and knowledge-network analysis remains conceptually unsettled. In this field, apparently incomplete or sparse data may reflect database coverage, inconsistent metadata, choices made during network construction, constraints on labelled training examples, or genuine characteristics of scientific activity. Treating these distinct conditions as interchangeable risks altering the phenomena that scientometric indicators are intended to measure. Moreover, methods for predicting links, recommending citations, or learning network representations are not necessarily augmentation methods. This study critically assesses the available evidence, distinguishes DA from data cleaning, recovery, integration, imputation, and computational inference, and develops an evidence-informed framework for deciding whether intervention is warranted, which approach is defensible, and how resulting data should be interpreted and validated.
Methodology: A systematic literature review was combined with directed qualitative content analysis using a deductive- inductive coding strategy. The Web of Science Core Collection corpus comprised 7,000 bibliographic records archived in seven export files, with publication years examined against the 2018–2026 review window. Temporal screening excluded 393 records, leaving 6,607; title and abstract screening excluded 6,507, and 100 records proceeded to full-text assessment. Historical screening records documented 43 initially included studies and 57 not retained for synthesis. Selection was reported using PRISMA 2020, with PRISMA principles guiding transparent discussion of search documentation. Because the complete original database query and individual screening decisions were not preserved, the historical search cannot be fully replicated. A subsequent recovery audit revisited the fixed corpus without undertaking a new database search. Of the original 43 studies, 31 remained in the core set and 12 were reassigned to supplementary methodological evidence; 12 additional eligible studies were recovered from the archived records. The revised core therefore comprised 43 studies, though its composition differed. These were classified by evidentiary role into explicit scholarly-context DA applications (A, six studies), computational mechanisms in scientific data and networks (B, 33), and data-quality or conceptual-boundary studies (C, four). Thirteen supplementary studies (D), including one newly identified, informed interpretation but were excluded from the core denominator; four sources found outside the archived corpus provided contextual background only. A study- evidence matrix recorded bibliographic identifiers, data contexts, mechanisms, analytical tasks, validation approaches, findings, and limitations. Coding distinguished directly reported results, cross-study synthesis, and conceptual inference. Because designs and outcomes were heterogeneous, the synthesis was qualitative rather than a meta-analysis, and no aggregate study-quality score was inferred. No causal effects were quantitatively estimated.
Findings: Only six of the 43 core studies documented DA in scholarly or closely related scientific-data settings. 76.7% concerned computational mechanisms, such as citation or article recommendation, link and collaboration prediction, network embedding, and knowledge-graph representations, which do not by themselves establish that augmentation occurred. Four studies addressed data quality and conceptual boundaries. Direct applications involved heterogeneous tasks, including patent-based technology analysis, synthetic publication time series, paper recommendation, and peer-review text. Their results supported task-specific possibilities rather than a general improvement in scientometric validity. Five recurring problem patterns emerged: uneven database coverage, metadata inconsistency, network-construction effects, class imbalance in particular learning tasks, and restricted training samples. Importantly, sparse collaboration structures or skewed citation distributions may be genuine properties of scientific systems, not defects requiring artificial correction. An inferred citation or predicted collaboration becomes DA only when its output is subsequently used as additional data or representation; predictive accuracy alone does not make that relationship historically observed. Accordingly, four epistemic states were differentiated: observed data, independently verified recovered data, inferred or predicted outputs, and synthetic or modified data. The synthesis identified four broad DA families- statistical or synthetic generation, deep generative methods, graph-based augmentation, and language-based augmentation—while emphasizing that evidence from adjacent domains cannot be transferred uncritically. Validation must extend beyond predictive metrics to statistical, structural, and semantic fidelity, alongside temporal validity where relevant. For scientific networks, changes in degree distributions, centrality, communities, and rankings require particular scrutiny. Provenance records, baseline comparisons, out-of-sample testing, sensitivity analysis, and expert review, where appropriate, support decisions about acceptance, restricted use, refinement, or rejection.
Conclusion: DA is a conditional methodological option, not a universal remedy for missing or imperfect scientometric data. The proposed eight-stage, decision-oriented conceptual framework moves from problem identification and causal diagnosis to specification of data and analytical units, selection of an intervention pathway and suitable mechanism, independent output validation, explicit recording of epistemic status, and sensitivity-informed decisions about use. Three principles govern these decisions: distinguish data operations conceptually, preserve evidence provenance, and maintain the epistemic status of every generated or recovered relationship. The framework does not claim empirical validation or predictive superiority; rather, it makes the assumptions and evidentiary thresholds governing interventions auditable. Future research should test it against reference datasets and examine whether augmented outputs preserve meaningful citation, collaboration, and thematic structures before their inclusion in substantive scientometric analyses.