Understanding the drivers that lead to interaction between target species in a fishery and marine mammals is a critical aspect in efforts to reduce bycatch. In the California drift gillnet fishery static management approaches and gear changes have reduced bycatch but neither measure ascertains the underlying dynamics causing bycatch events. To avoid further potentially drastic measures such as hard caps, dynamic management approaches that consider the scales relevant to physical dynamics, animal movement and human use could be implemented. A key component to this approach is determining the factors that lead to fisheries interactions. Using 25 years (1990-2014) of National Oceanic and Atmospheric Administration fisheries' observer data from the California drift gillnet fishery, we model the relative probability of bycatch (presence--absence) of four cetacean species in the California Current System (short-beaked common dolphin Delphinus delphis, northern right whale dolphins Lissodelphis borealis, Risso's dolphins Grampus griseus, and Pacific white-sided dolphins Lagenorhynchus obliquidens ). Due to the nature of protected species bycatch, these are rare-events, which cause a large amount of absences (zeros) in each species' dataset. Using a data-assimilative configuration of the Regional Ocean Modeling System, we determined the capabilities of a flexible machine-learning algorithm to handle these zero-inflated datasets in order to explore the physical drivers of cetacean bycatch in the California drift gillnet fishery. Results suggest that cetacean bycatch probability has a complex relationship with the physical environment, with mesoscale variability acting as a strong driver. Through the modeling process, we observed varied responses to the range of sample sizes in the zero-inflated datasets, determining the minimum number of presences capable of building an accurate model. The selection of predictor variables and model evaluation statistics were found to play an important role in assessing the biological significance of our species distribution models. These results highlight the statistical capability (and incapability) of modeling techniques to predict the complex nature driving fishery interaction of cetacean bycatch in the California drift gillnet fishery. By determining where fisheries interactions are most likely to occur, we can inform near real-time management approaches to reduce bycatch while still allowing fishermen to meet their catch quotas.