Nuclear fusion is one of the best options to achieve a virtually limitless energy source in the future. However, sustaining burning plasma reactions is very challenging because disruptive events cause the loss of plasma confinement and damages to the tokamak machine. Thus, reproducing a nuclear fusion reaction on Earth represents an actual scientific and technical challenge. Next generation machines must be able to predict, mitigate or avoid the disruptions but the present understanding of the phenomenon has not yet gone so far as to provide a physical model describing their root cause. On the contrary, a large quantity of experimental disruption data exists. For these reason, in the last decades, machine learning techniques have been applied to design alarm systems both for mitigating or to actively avoid approaching disruptions. More recently, the disruption predictor concept has been evolving toward a more complex system, able to detect the proximity of the plasma state to the boundaries of the operational space free from disruptions in order to schedule avoidance strategies rather than mitigation actions. This system should also be able to determine which type of disruption is about to occur in order to efficiently take suitable disruption avoidance actions. This paper reports an overview of machine learning algorithms as tools for disruption prediction and classification at JET, the world's largest operational plasma physics experiment, located at Culham Centre for Fusion Energy in Oxfordshire, UK. Both traditional neural network models for disruption prediction and manifold learning approaches as a tool for plasma operational space mapping and visualization, disruption prediction and classification, are described.