Session-based　recommendation　aims　to　model　users＇　interests　based　on　user　sequences.　Most　models　can＇t　fully　capture　user　preferences.　In　this　paper,　we　propose　a　Complex　Transitions　Learning　model　with　Graph　Attentive　Network,　which　captures　item　transitions　in　neighbor　sessions　comprehensively.　Global　and　local　session　graphs　are　constructed　from　multiple　sessions.　Item　representations　are　learned　by　GNNs.　Specially,　representation　of　session　and　final　item　are　combined　to　model　association　between　items　and　sessions.　Experimental　results　indicate　that　our　method　outperforms　previous　models.　©　2023　IEEE.