The problem of Point-Of-Interest (POI) recommendation, based on the user’s historical check-in records, determines whether a user checks in at specific POI. However, the user-POI data have a long-tail distribution phenomenon. To mitigate the sparsity of check-in data, it is a good idea to exploit the sufficient attributes of POI and recommend POIs in both geography wise and category wise. Generally, this problem can be treated as two specific tasks with feature combination, ignoring cross-task dependencies and feature disentanglement. To address the aforementioned problems, this paper proposes a novel joint framework named InteractPOI, enabling two-stage interaction bewteen geography-wise and category-wise POI recommendations. Specifically, this paper comprehensively considers the sequence effect and the neighbor effect both from geography wise and category wise. For the first-stage interaction, we design a disentangled graph embedding model to distinguish different influencing factors from geography wise and category wise. For the second-stage interaction, we integrate a gating mechanism for feature fusion with a complementary algorithm for interactive optimization. Extensive experiments on two datasets demonstrate the superiority of the proposed model.