Due　to　the　complex　crystal　structures　and　interatomic　interactions,　the　prediction　of　magnetic　properties　and　effective　composition　design　of　rare　earth　permanent　magnets　are　quite　difficult.　As　the　most　promising　permanent　magnets　for　high-temperature　applications,　Sm-Co　alloys　have　been　developed　for　several　decades　by　intuition,　experience　and　trial-and-error　methods.　In　this　work,　rapid　and　accurate　prediction　of　saturation　magnetization　of　Sm-Co　alloys　was　realized　by　machine　learning　integrated　with　selection　of　characteristics　of　constituent　elements,　such　as　pseudopotential　core　radius,　heat　of　fusion,　boiling　point,　valence　electron　number　and　covalent　radius.　Based　on　the　data-driven　strategy　and　the　proposed　criteria　for　elements　selection,　new-type　Sm-Co　based　alloys　with　excellent　comprehensive　magnetic　performance　were　prepared.　The　methods　of　feature　construction　and　optimal　multistep　feature　selection　in　machine　learning　loops　developed　in　this　study　are　applicable　for　properties　prediction　and　composition　design　of　a　series　of　multicomponent　alloys.