Multi-view　learning　has　garnered　significant　attention　due　to　its　ability　to　learn　more　comprehensive　representations.　However,　a　critical　challenge　lies　in　the　presence　of　redundant　information　within　multi-view　data.　This　redundancy　not　only　dramatically　increases　computational　complexity　but　also　introduces　substantial　noise,　severely　disrupting　the　learning　process　and　undermining　the　effectiveness　of　the　model.　To　address　this　issue,　we　propose　a　learnable　framework　termed　dual-channel　optimized　multi-view　learning　via　Information　bottleneck.　This　framework　aims　to　capture　more　comprehensive　consistency　and　complementary　information　while　minimizing　the　influence　of　redundant　information.　Specifically,　we　design　a　dual-channel　architecture　to　thoroughly　explore　both　feature　and　topology　information　from　multi-view　data.　For　the　feature　space,　we　utilize　deep　neural　networks　to　approximate　matrix　factorization,　better　capturing　latent　commonalities　across　multiple　views.　In　the　topology　space,　we　adopt　a　prediction-driven　approach,　emphasizing　the　importance　of　views　with　higher　prediction　confidence　in　the　fused　view　while　reducing　the　contribution　of　views　with　lower　confidence.　Furthermore,　we　employ　information　bottleneck　on　both　features　and　topology　to　minimize　redundant　information　and　connections　while　preserving　the　most　relevant　information　for　downstream　tasks.　The　experiments　confirm　the　robustness　of　the　model　and　its　effectiveness　in　classification　tasks.　©　2025　Elsevier　B.V.