Artificial intelligence and machine learning have recently made outstanding contributions to the performance of information system and cyber--physical system security. There has been a plethora of research in this area, resulting in an outburst of publications over the past two years. Choosing the right algorithm to solve a complex security problem in a very precise industrial context is a challenging task. Therefore, in this paper, we propose a Learning Algorithm Recommendation Framework that, for a clearly defined situation, guides the selection of learning algorithm and scientific discipline (e.g. RNN, GAN, RL, CNN,...) which have sparked great interest to the scientific community and which therefore offers preponderant elements and benefits for further deployments. This framework has the advantage of having been generated from an extensive analysis of the literature, as illustrated by this paper for the recurrent neural networks and their variations.