The　prediction　accuracy　rate　of　user　response　prediction　tasks　has　significant　practical　implications.　A　slight　improvement　in　this　metrics　can　lead　to　a　significant　gain　of　commercial　value　for　real-world　applications.　Factorization　machine　(FM)　is　a　popular　prediction　model　that　has　been　used　widely　to　predict　user　response,　such　as　click,　purchase,　and　browse.　However,　since　FM　learns　a　generalized　representation　for　each　feature　but　ignores　the　importance　of　a　given　feature　in　different　instances,　its　prediction　accuracy　is　compromised.　In　order　to　address　this　problem,　a　recent　research　proposes　an　input-aware　factorization　machine　(IFM)　that　refines　the　representation　by　weighting　the　importance　of　a　given　feature　in　the　different　instances.　Unfortunately,　IFM　overemphasizes　the　effect　of　instances,　while　ignoring　the　generalized　representation　for　each　feature,　which　limits　the　prediction　accuracy.　In　this　paper,　we　propose　an　instance-weight　balanced　factorization　machine　called　IBFM,　which　considers　the　importance　of　instances　and　retains　the　fundamental　effectiveness　of　generalized　representation　by　properly　balancing　the　instance　weight　for　the　representation　of　each　feature.　Comprehensive　experimental　results　show　that　the　IBFM　model　obtains　a　higher　prediction　accuracy　rate　over　FM,　IFM,　and　three　other　state-of-the-art　models.　©　2021　IEEE.