Seismic data collected before volcanic eruptions holds important clues for forecasting future eruptions. However, many volcanoes have short monitoring histories or have infrequent eruptions. Here, we aim to enhance volcanic eruption forecasting through machine learning, tackling challenges in generalizing eruption precursors across diverse volcanoes, including those with limited seismic data. Additionally, we investigate similarities and differences in pre-eruptive seismic activity. We combine records from 24 volcanoes to train machine learning (ML) models that recognize eruption precursors and tested the accuracy by forecasting on volcanoes whose data had been withheld during training. These forecasts had nearly the same accuracy as other models that were given the advantage of training on past eruptions at the target volcano. We find that numerous eruption precursors are applicable across various volcanoes, challenging the notion that volcanoes exhibit individualistic behaviour. The accuracy of the ML models is about 50 to 70% higher than traditional methods based on averaged seismic amplitude. Our ML models should be a tool for volcano observatories to provide timely probabilistic forecasts suitable to address volcanic risk, especially where prior eruption data is limited. This methodology and workflow can be applied to other hazard types that suffer from regional data-scarcity, e.g., wildfires and floods.