Deep slow slip events (SSEs) at subduction zones have significantly contributed to refining our understanding of the megathrust earthquake cycle at the brittle-ductile transition. However, the specific combination of factors that determine their occurrence has not yet been fully explored. Here we evaluate the contribution of several of these characteristics using globally mapped geophysical data that are used as proxies for physical properties of the subducting plate. This is performed by classifying 25 km-wide, trench-parallel segments into binary classes based on the observation (or lack thereof) of deep, short- or long-term SSEs. The five characteristics explored here include subducting plate age, sediment thickness, relative plate velocity, slab dip, and plate surface roughness. We use these characteristics to train six Machine Learning models based on different learning algorithms: Gaussian Naïve Bayes, Logistic Regression, Linear Discriminant Analysis, Random Forest, Support Vector Machine, and K-Nearest Neighbour. Short-term SSE models show that subducting plate age, relative velocity, and sediment thickness have the strongest predictive power with the first two characteristics negatively correlating and sediment thickness positively correlating with SSE occurrence, respectively. These results are consistent with a conceptual model where slow slip is controlled by conditions favoring the enduring release (and possible storage) of fluids near the source region. However, the relationship between these features and elevated pore fluid pressures is not established here and further evidence is needed to validate this hypothesis. We then use a final model constructed as a weighted average of the best performing models to make predictions on the probability of SSE occurrence, with predicted short-term SSE occurrence in South America, the Aleutians, Sumatra, Vanuatu and Solomon, as well as long-term SSE occurrence in the Aleutians, Izu-Bonin, Kuril-Kamchatka, Mariana, and Tonga-Kermadec. Overall, long-term SSE models do not perform as well as the short-term SSE models which may indicate that long-term SSEs are controlled by a different and/or extended set of physical characteristics than the short-term SSEs.