A Systematic Literature Review of Preprocessing Frameworks forTransactional Datasets: Data Quality, Temporal Integrity, FeatureEngineering, Imbalance Handling, Leakage Prevention, andReproducibility

Abstract

Transactional datasets constitute an important source of information for machine-learning applications in finance, retail, banking, e-commerce, cybersecurity, insurance, telecommunications, and digital services. Unlike many static tabular datasets, transactional data are characterized by temporal ordering, repeated entities, heterogeneous attributes, high cardinality, evolving behavior, missing and inconsistent records, duplicated events, skewed numerical variables, categorical information, and frequently severe class imbalance. These characteristics make preprocessing a critical determinant of downstream model validity. Nevertheless, existing studies often treat preprocessing as a collection of isolated operations rather than as a systematic, leakage-aware, and reproducible framework. This systematic literature review examines preprocessing frameworks for transactional datasets and synthesizes evidence across data acquisition, quality assessment, cleaning, deduplication, missing-value treatment, temporal normalization, categorical encoding, numerical transformation, feature engineering, aggregation, dimensionality reduction, imbalance handling, validation, leakage prevention, and reporting. The review is structured according to PRISMA 2020 principles and draws on literature concerning machine-learning preprocessing, data-mining process frameworks, transaction analytics, fraud detection, and methodological reliability. PRISMA 2020 provides a 27-item reporting framework and revised flow diagrams for systematic reviews. The synthesis indicates that transactional preprocessing differs fundamentally from generic tabular preprocessing because temporal and entity relationships must be preserved throughout the pipeline. The literature consistently identifies data cleaning, transformation, feature engineering, and reduction as major preprocessing activities, while transaction-oriented studies additionally emphasize duplicate removal, normalization, temporal features, behavioral aggregation, and heterogeneous-source integration. The review further identifies temporal leakage, entity leakage, preprocessing leakage, and target-derived feature leakage as major threats to validity. General machine-learning guidance emphasizes that preprocessing parameters should be learned only from training data and that test data should remain isolated from model-development decisions. Based on the synthesis, this paper proposes a Transactional Data Preprocessing Framework (TDPF) consisting of nine stages: data acquisition and provenance, structural validation, data-quality management, temporal and entity integrity, transformation, transaction-level and behavioral feature engineering, imbalance management, leakage-aware validation, and reproducible deployment monitoring.