A Data-Driven Unified Taxonomy and Benchmarking Framework for Machine Learning-Based Task Scheduling in Edge–Cloud Computing

Abstract Machine learning (ML)- based task scheduling has become a fundamental research focus in edge-cloud computing and seeks to optimize latency, energy, and resource usage in distributed systems. However, as algorithms have gotten rapidly better, recent studies do not engage in consistent, rigorous benchmarking practices, so comparisons across studies are not reliable and cumulative scientific advancement cannot be made. This paper is a descriptive, quantitative review of 103 exemplary articles on ML-based task scheduling in edge clouds. It is established that there are significant methodological differences: 82% of the studies employ synthetic workloads only, and only 35% perform rigorous benchmarking. Although 95/103 studies claim to use multi-objective optimization, they rely extensively on performance-based metrics (e.g., latency and QoS), whereas reliability and SLA compliance are addressed in fewer than 12% of studies. In addition, baseline comparison practices are uneven, and fewer than 10% of studies provide paradigm-balanced comparisons. We perform inferential statistical tests to go beyond descriptive observations and show significant structural dependencies among learning paradigms, dataset types, and benchmarking rigour (chi-square tests, p < 0.001, large effect sizes). The findings suggest that the identified differences are not random but systemic, revealing underlying issues in contemporary evaluation practices. These weaknesses will be addressed by providing a Unified Taxonomy and Benchmarking Framework (UTBF), including a hierarchical method classification and a standardised, reproducibility-aware evaluation protocol. UTBF transforms taxonomy into a working assessment tool by formalizing the strengths of benchmarking and cross-paradigm comparisons and promoting uniform experimental design, unlike traditional survey-based taxonomies. The proposed framework will be used to make fair comparisons, enhance methodological rigour, and perform more deployment-relevant validation in future research on ML-based edge-cloud task scheduling.