Long-period (LP), hybrid and volcano-tectonic (VT) seismicity are important indicators for tracking the evolution of volcanic processes. However, classification of a large number of these events remains a challenging task. Here, we propose an unsupervised-learning classification method and apply it to 5,949 seismic events in Kilauea volcano, Hawai’i, during a 4-month period before the collapse of Pu’u’ O’o on 30 April, 2018. We successfully separate the LPs, VTs and hybrids and show three episodes of LPs and hybrids. The last episode in 9-27 April before the collapse event shows unusually high rate and more shallow origins, coincident with the rapid changes of near-caldera deformation and lava lake elevation in Halema’uma’u. The hybrids are found to be closely associated with LPs and magma movement but mixed with shear-failure or near-surface resonance. Our method can be used to construct catalogs of diverse seismicity and contribute to volcano monitoring and eruption forecasting.