With　the　development　of　e-sports,　Multiplayer　Online　Battle　Arena　(MOBA)　has　become　the　most　popular　game　type　in　the　world.　In　addition　to　bringing　huge　commercial　value,　it　also　produces　a　large　amount　of　rich　public　game　data,　which　attracts　researchers　to　deeply　combine　deep　learning　technology　with　MOBA.　Among　them,　item　recommendation　is　the　core　task,　including　pre-game　item　recommendation　and　in-game　item　recommendation.　Compared　with　in-game　item　recommendation,　pre-game　recommendation　faces　serious　data　sparsity　and　missing　data　labels,　which　makes　pre-game　item　recommendation　a　hard　challenge.　Existing　methods　pay　more　attention　to　conventional　indicators　such　as　precision　and　recall,　but　ignore　the　usability　of　results　in　MOBA,　and　have　not　fully　considered　the　relationship　between　items.　In　this　paper,　we　propose　a　pre-game　item　recommendation　method　based　on　self-supervised　learning.　While　repairing　the　missing　labels,　the　Transformer　encoder　is　used　to　learn　the　relationship　between　items.　Experimental　results　on　a　public　dataset　show　that　our　method　outperforms　previous　methods　in　terms　of　precision　and　usability　indicators.　©　2023　IEEE.