Graph　Convolutional　Network　(GCN)　has　become　a　hotspot　in　graph-based　machine　learning　due　to　its　powerful　graph　processing　capability.　Most　of　the　existing　GCN-based　approaches　are　designed　for　single-view　data.　In　numerous　practical　scenarios,　data　is　expressed　through　multiple　views,　rather　than　a　single　view.　The　ability　of　GCN　to　model　homogeneous　graphs　is　indisputable,　while　it　is　insufficient　in　facing　the　heterophily　property　of　multi-view　data.　In　this　paper,　we　revisit　multi-view　learning　to　propose　an　implicit　heterogeneous　graph　convolutional　network　that　efficiently　captures　the　heterogeneity　of　multi-view　data　while　exploiting　the　powerful　feature　aggregation　capability　of　GCN.　We　automatically　assign　optimal　importance　to　each　view　when　constructing　the　meta-path　graph.　High-order　cross-view　meta-paths　are　explored　based　on　the　obtained　graph,　and　a　series　of　graph　matrices　are　generated.　Combining　graph　matrices　with　learnable　global　feature　representation　to　obtain　heterogeneous　graph　embeddings　at　various　levels.　Finally,　in　order　to　effectively　utilize　both　local　and　global　information,　we　introduce　a　graph-level　attention　mechanism　at　the　meta-path　level　that　allocates　private　information　to　each　node　individually.　Extensive　experimental　results　convincingly　support　the　superior　performance　of　the　proposed　method　compared　to　other　state-of-the-art　approaches.