Graph neural networks (GNNs) have been widely applied in traffic demand prediction, and transportation modes can be divided into station-based mode and free-floating traffic mode. Existing research in traffic graph construction primarily relies on map matching to construct graphs based on the road network. However, the complexity and inhomogeneity of data distribution in free-floating traffic demand forecasting make road network matching inflexible. To tackle these challenges, this paper introduces a novel graph construction method tailored to free-floating traffic mode. We propose a novel density-based clustering algorithm (HDPC-L) to determine the flexible positioning of nodes in the graph, overcoming the computational bottlenecks of traditional clustering algorithms and enabling effective handling of large-scale datasets. Furthermore, we extract valuable information from ridership data to initialize the edge weights of GNNs. Comprehensive experiments on two real-world datasets, the Shenzhen bike-sharing dataset and the Haikou ride-hailing dataset, show that the method significantly improves the performance of the model. On average, our models show an improvement in accuracy of around 25\% and 19.5\% on the two datasets. Additionally, it significantly enhances computational efficiency, reducing training time by approximately 12% and 32.5% on the two datasets. We make our code available at https://github.com/houjinyan/HDPC-L-ODInit.