We utilized relative polarity measurements and machine learning techniques to better resolve focal mechanisms and stress orientations considering a catalog of ∼29,000 relocated earthquakes that occurred during 1984–2021 in the southeastern San Francisco Bay Area. Earthquake focal mechanisms are commonly produced using P wave first motion polarities, which traditionally requires events to be well-recorded across a seismic network with good focal sphere coverage. We adapted recently developed approaches that are less dependent on high signal-to-noise records and exploit similar waveforms to produce relative polarity and amplitude measurements between earthquake pairs. These techniques were previously only applied on localized earthquake sequences, and we further developed these approaches so that they can be utilized for regional catalogs. We validated or corrected manually identified polarities by performing polarity consensuses using earthquake pairs. Missing and unreliable polarity measurements were imputed using iterative random forests, an unsupervised ensemble machine learning method. Relative P and S wave amplitude measurements were made between earthquakes, constraining S/P ratios for low signal-to-noise waveforms. Using these techniques, we were able to reduce focal mechanism uncertainties by an average of ∼13° and produced well-constrained focal mechanisms for ∼6 times as many earthquakes than those produced using only the traditionally derived polarities. We performed stress inversions using the focal mechanisms by grouping the focal mechanism results into a quadtree structure. Our stress results are consistent with previous work, albeit at a higher spatial resolution, and demonstrate these techniques can aid our understanding of fault structures and kinematics in more detail than was previously possible.