Graph,　as　a　powerful　data　structure,　has　shown　superior　capability　on　modeling　complex　systems.　Since　real-world　objects　and　their　interactions　are　often　multi-modal　and　multi-typed,　compared　with　traditional　homogeneous　graphs,　heterogeneous　graphs　can　represent　real-world　objects　more　effectively.　Meanwhile,　rich　semantic　information　brings　great　challenges　for　learning　heterogeneous　graph　representation　(HGR).　Most　existing　HGR　methods　are　based　on　the　concept　of　meta-path,　which　is　constructed　based　on　direct　neighbors　and　define　composite　semantic　relations　in　heterogeneous　graph.　However,　when　the　direct　neighbor　information　is　inadequate,　which　always　happens　due　to　insufficient　observation,　the　quality　of　meta-paths　cannot　be　guaranteed.　Therefore,　we　propose　a　novel　HGR　framework　based　on　latent　direct　neighbors.　Specifically,　random　walks　are　first　utilized　to　discover　the　potential　candidates　from　indirect　neighbors.　Then　HodgeRank　is　introduced　to　determine　the　latent　direct　neighbors　according　to　their　importance　to　the　target.　After　that,　neighborhood　relationships　are　augmented　with　the　selected　latent　direct　neighbors,　and　the　adjacency　tensor　of　the　heterogeneous　graph　is　refactored　correspondingly.　Finally,　Graph　Transformer　Network　is　adopted　to　construct　semantic　meta-paths　automatically　and　generate　HGR.　Numerical　experiments　on　different　real-world　heterogeneous　networks　show　that　our　new　approach　can　produce　more　meta-path　instances　and　introduce　more　complex　and　diverse　semantic　information,　and　consequently　achieves　more　accurate　predictions　compared　with　several　state-of-the-art　baselines.　©　2022　Elsevier　Ltd