Predicting　users＇　actions　based　on　anonymous　sessions　is　a　challenging　problem　due　to　the　uncertainty　of　user　behavior　and　limited　information.　Recent　advances　in　graph　neural　networks　(GNN)　have　led　to　a　promising　approach　for　addressing　this　problem.　However,　existing　methods　have　three　major　issues.　First,　they　are　incapable　of　modeling　the　transitions　between　inconsecutive　items.　Second,　they　are　infeasible　for　learning　the　cross-feature　interactions　when　learning　the　item　relationships.　Third,　very　few　models　can　adapt　to　the　improvement　of　embedding　quality　to　help　improve　recommendation　performance.　Therefore,　to　address　these　issues,　we　propose　a　novel　model　named　Multilevel　Feature　Interactions　Learning　(MFIL)　that　effectively　learns　item　and　session　representation　using　GNN.　By　leveraging　item　side　information,　e.g.,　brands　and　categories,　MFIL　can　model　transitions　between　inconsecutive　items　in　the　session　graph　(session-level).　We　further　design　hierarchical　structures　to　learn　the　feature　interactions,　which　is　effective　to　estimate　the　importance　weights　of　different　neighboring　items　in　the　global　graph　(global-level).　In　addition,　an　effective　learning　strategy　is　employed　to　enhance　MFIL＇s　capability,　and　it　performs　better　than　the　classic　regularization　methods.　Extensive　experiments　conducted　on　real-world　datasets　demonstrate　that　MFIL,　significantly　outperforms　existing　state-of-the-art　graph-based　methods.