Openly　accessible　map　images　have　become　a　valuable　source　for　geographic　information　and　cartographic　design.　Map　layout　becomes　a　natural　leverage　point　to　unlock　fine-grained　map　information　as　it　facilitates　the　expression　of　map　content　through　the　spatial　coordination　of　map　elements.　Current　layout　studies　focus　on　documents,　user　interfaces,　and　floor　plans,　leading　to　models　that　poorly　recognize　complex　map　layouts.　Inspired　by　the　strong　performance　of　graph　neural　networks　(GNNs)　in　modeling　element-wise　layout　relationships,　this　paper　proposes　MapLayNet,　a　GNN-based　model　for　map　layout　representation　learning.　The　model　introduces　the　structural　features　and　proposes　a　dual-branch　network　architecture　to　enhance　the　model　representation　of　complex　map　layouts　at　global　and　local　scales.　To　ensure　model　generalization　for　diverse　map　layout　designs,　we　employed　a　weakly　supervised　learning　strategy　and　trained　the　model　using　a　heuristic-based　triplet　loss　of　map　layout　similarity,　along　with　a　mutual　information　loss　that　effectively　utilizes　the　dual-branch　architecture.　For　model　evaluation,　MapLayNet　has　outperformed　all　baseline　methods　in　similarity　and　stability　by　3.2%　and　10.3%,　respectively,　while　showing　better　consistency　against　human　evaluation　on　a　map　layout　retrieval　task.　With　the　effective　model　design,　the　embedding　learned　by　MapLayNet　is　capable　of　generating　a　concept　hierarchy　with　compactness　in　lower-level　layout　patterns　and　partial-ordered　concept　relations　in　the　higher-level　layout　abstraction.　Potential　applications　such　as　map　layout　retrieval　and　design　recommendation　can　be　envisioned　with　the　resulting　interpretable　and　manipulable　map　layout　embedding.