The　accurate　prediction　of　origin-destination　(OD)　flows　is　essential　for　advancing　sustainable　urban　mobility　and　supporting　resilient　urban　planning.　However,　the　inherent　heterogeneity　of　mobility　patterns　results　in　complex　geographic　unit　relations,　diverse　spatial　organizational　structures,　and　the　long-tailed　effect　on　OD　flow　distribution.　This　study　proposes　a　novel　OD　flow　prediction　method　based　on　graph-based　deep　learning　(named　as　HMCG-LGBM).　Specifically,　1)　a　modularity-based　graph　reconstruction　strategy　is　presented　for　geographic　unit　relation　augmentation　by　eliminating　weak　connections;　2)　the　heterogeneous　spatial　organization　of　OD　flows　is　captured　by　combining　the　community　detection　approach　and　graph　attention　mechanism　with　the　introduction　of　socio-economic　and　spatial　features;　and　3)　a　weighted　loss　function　with　distribution　smoothing　paradigm　is　developed　to　enhance　the　prediction　for　low-probability　mobility　events,　addressing　the　challenges　posed　by　long-tailed　distributions.　Extensive　experiments　conducted　on　real-world　datasets　show　that　the　predictive　performance　of　the　proposed　method　is　significantly　improved,　with　the　RMSE　and　MAE　reduced　from　the　baselines　by　11.1%-33.3%　and　14.1%-22.2%,　respectively.　The　results　also　demonstrate　the　robustness　of　the　proposed　method　for　dealing　with　imbalanced　OD　flow　distributions,　providing　valuable　insights　for　spatial　interaction　predictive　modeling　in　the　context　of　sustainable　urban　systems.