Deep Multi-View Spatial-Temporal Network for Taxi Demand Prediction

Authors: Huaxiu Yao, Fei Wu, Jintao Ke, Xianfeng Tang, Yitian Jia, Siyu Lu, Pinghua Gong, Jieping Ye, Zhenhui Li

AAAI 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on large-scale real taxi demand data demonstrate effectiveness of our approach over state-of-the-art methods.
Researcher Affiliation Collaboration Huaxiu Yao, Fei Wu Pennsylvania State University {huaxiuyao, fxw133}@ist.psu.edu Hong Kong University of Science and Technology jke@connect.ust.hk Xianfeng Tang Pennsylvania State University xianfeng@ist.psu.edu Yitian Jia, Siyu Lu, Pinghua Gong, Jieping Ye, Didi Chuxing {jiayitian, lusiyu, gongpinghua, yejieping}@didichuxing.com Zhenhui Li Pennsylvania State University jessieli@ist.psu.edu
Pseudocode Yes Algorithm 1: Training Pipeline of DMVST-Net
Open Source Code No The paper mentions using TensorFlow and Keras for implementation but does not provide a link or explicit statement about the availability of their own source code.
Open Datasets No The paper uses a "large-scale online taxi request dataset collected from Didi Chuxing" but does not provide any public access information (link, DOI, or citation for public availability).
Dataset Splits Yes The data from 02/01/2017 to 03/19/2017 is used for training (47 days), and the data from 03/20/2017 to 03/26/2017 (7 days) is used for testing. ... The first 90% of the training samples were selected for training each model and the remaining 10% were in the validation set for parameter tuning.
Hardware Specification Yes All these experiments were run on a cluster with four NVIDIA P100 GPUs.
Software Dependencies No The paper states "We use Tensorflow and Keras (Chollet and others 2015) to implement our proposed model" but does not specify version numbers for these software components.
Experiment Setup Yes The size of each neighborhood considered was set as 9 9 (i.e., S = 9), which corresponds to 6km 6km rectangles. For spatial view, we set K = 3 (number of layers), τ = 3 3 (size of filter), λ = 64 (number of filters used), and d = 64 (dimension of the output). For the temporal component, we set the sequence length h = 8 (i.e., 4 hours) for LSTM. The output dimension of graph embedding is set as 32. The output dimension for the semantic view is set to 6. We used Sigmoid function as the activation function for the fully connected layer in the final prediction component. Activation functions in other fully connected layers are Re LU. Batch normalization is used in the local CNN component. The batch size in our experiment was set to 64. The first 90% of the training samples were selected for training each model and the remaining 10% were in the validation set for parameter tuning. We also used early-stop in all the experiments. The early-stop round and the max epoch were set to 10 and 100 in the experiment, respectively.