We develop a new method, Stars' Galactic Origin (StarGO), to identify the galactic origins of halo stars using their kinematics. Our method is based on a self-organizing map (SOM), which is one of the most popular unsupervised learning algorithms. STARGO combines SOM with a novel adaptive group identification algorithm with essentially no free parameters. To evaluate our model, we build a synthetic stellar halo from mergers of nine satellites in the Milky Way. We construct the mock catalog by extracting a heliocentric volume of 10 kpc from our simulations and assigning expected observational uncertainties corresponding to bright stars from Gaia DR2 and LAMOST DR5. We compare the results from STARGO against those from a friends-of-friends-based method in the space of orbital energy and angular momentum. We show that STARGO is able to systematically identify more satellites and achieve higher number fraction of identified stars for most of the satellites within the extracted heliocentric volume. When applied to data from Gaia DR2, STARGO will enable us to reveal the origins of the inner stellar halo in unprecedented detail.