<p>With the development of society, Large-scale global optimization (LSGO) problems are widely occurring in the real world. Meta-heuristics, due to its effectiveness and easy implementation, has become a popular method for solving LSGO problems. However, each algorithm has strengths for a particular type of problem and there is no single algorithm that performs well on all problem instances. To improve the efficiency of solving LSGO problems, this paper proposes a two-layer data-driven meta-heuristics recommendation model (TD-MRM) combining meta-learning and reinforcement learning. The optimization process of TD-MRM consists of two parts: “off-line optimization based on meta-learning” and “on-line optimization based on reinforcement learning”. Firstly, TD-MRM achieves adaptive meta-heuristics recommendation on the process of “off-line optimization based on meta-learning”; secondly, based on the recommended meta-heuristics obtained by off-line optimization, TD-MRM achieves automated algorithm scheduling on the process of “on-line optimization based on reinforcement learning”. To test the performance of the proposed TD-MRM, CEC’2010 and CEC’2013 large-scale benchmark functions are used as LSGO problem instances and seven meta-heuristics for LSGO problems are selected as the candidate algorithms. Comprehensive experiments demonstrate the significantly efficiency of the TD-MRM, it can provide the optimal solution in 83% of the LSGO problems. Two real-world LSGO problems are adopted to evaluate the practicality of TD-MRM. The experimental results indicate that TD-MRM achieved the optimal solution in both cases, with an average performance improvement of 8.57%. In addition, four meta-features are enriched to fully characterize LSGO problems and the effectiveness of them is validated.</p>