Personalized　recommender　systems　are　playing　an　increasingly　critical　role　in　a　variety　of　online　applications.　In　recent　years,　advancements　in　graph-structured　deep　neural　networks　have　attracted　considerable　interest　and　achieved　state-of-the-art　performance　on　recommender　system　benchmarks.　However,　existing　graph-based　recommendation　methods　generally　characterize　each　user　with　just　one　representation　vector,　which　is　insufficient　to　convey　diverse　preferences　of　users.　To　address　this　issue,　in　this　paper,　we　approach　the　learning　of　user　representations　from　a　different　perspective,　by　modeling　users　based　on　multiple　representation　embeddings.　We　propose　a　Multiple　Preferences　with　Item　Attributes　for　Graph　Convolutional　Collaborative　Filtering　(MPIA)　framework　built　upon　the　message-aggregation　concept　of　graph　neural　networks,　which　can　generate　preference-specific　user　representations　to　better　model　the　diverse　preferences　of　users.　By　taking　advantage　of　graph　representation　learning　techniques,　MPIA　learns　preference-specific　embeddings　for　users　and　attribute-specific　embeddings　for　items.　Moreover,　we　utilize　shared　embeddings　for　user　and　item　representations　to　obtain　the　commonalities　in　multiple　networks.　Specifically,　we　construct　a　user-item　bipartite　graph　for　each　preference,　and　enrich　the　representation　of　each　node　with　the　topological　structure　and　features　of　its　neighbors.　We　also　design　a　preference-attribute　fusion　method　to　acquire　more　accurate　item　retrievals　for　every　aspect　of　interest.　Extensive　experiments　conducted　on　three　real-world　datasets　demonstrate　the　effectiveness　of　the　proposed　MPIA　framework.