-
Figure 1.
Structure of convolutional neural network.
-
Figure 2.
Structure of graph convolutional neural network.
-
Figure 3.
Structure of recurrent neural network.
-
Figure 4.
Structure of LSTM network and GRU network.
-
Figure 5.
Structure of Transformer network.
-
Figure 6.
Structure of Temporal Convolutional Network.
-
Figure 7.
Encoder-Decoder structure diagram.
-
Figure 8.
Structure of ResNet network.
-
Figure 9.
Structure of multi-task learning.
-
Model Spatial topology construction Spatial dependency Temporal dependency Data set Other factors CSTN[71] Raster 3DCNN Conv LSTM NYC-TOD Local spatial context, meteorological information, globally relevant context MultiConvLSTM[72] Raster MultiConv ConvLSTM NYC taxi None CLTS[73] Raster Conv2D ConvLSTM Beijing Taxi None GEML[74] Raster (Geographic/
semantic nodes)SGCN (Grid embedding) LSTM NYC-taxi /DiDi ChengDu Multi-task learning FL-GCN[75] Graph Graph convolution (nodes, edges) Kalman filtering New Jersey Highway None CAS-CNN[76] Raster Split CNN URT Channel-wise attention MPGCN[77] Graph 2D-GCN LSTM DIDI Beijing /DiDi shanghai None GCN-SBULSTM[78] Graph GCN Stacking bidirectional unidirectional LSTMs SZ Metro None ST-ED-RMGC[79] Graph Multi-graph convolutional networks LSTM NYC taxi Encoder decoder DNEAT[80] Dynamic node
topologyGCN (k-TNEAT) k-hop temporal encoder DiDi ChengDu/ NYC taxi None Spatial OD-BiConvLSTM[81] Raster Conv2D BiLSTM NYC taxi None CMOD[82] Graph (Event) The graph represents
learningCTDG (Continuous-time evolution representation) BJ Subway/NYC-Taxi Multi- Head Attention HMOD[83] Graph (Event) Graph embedded /Random walk GRU/CTDG NYC-Taxi/ Beijing Metro None SIZINB-GNN[84] Graph GNN TCN CDP dataset None ODformer[86] Graph (Event) 2DGCN ODformer NYC taxi OD attention SI-GCN[87] Graph (Event) GCN (graph embedding) Encoder-decoder DIDI Beijing a mapping function STGDL[88] Graph (road) S-GCN ResNet-based block ST-Conv CNN NYC taxi/DIDI Haikou both short-term and long-term OD predictions CWGAN-div[89] Graph (road network) GAN ResNet NYC taxi network-wide OD demand DMGC-GAN[90] Graph (neighbor/
mutual attraction/
passengers' mobility
association mode)GCN TMGCN NYC taxi None Hex D-GCN[91] Graph (hexagon-based path) GCN CNN Taxi Shanghai None OD-TGAT[65] Graph (grid map) GAT GRU NYC Taxi None TFF[92] Graph GCN ST-Attention block Chongqing A modified Kalman filter (KF) CSGCN[93] Graph GCN CNN Taxi Beijing Shifted Graph Clustering gHMC-STA[94] Graph GCN Multi-Head Convolution Taxi Beijing Graph multi-head convolution for spatio-temporal aggregation HSTN[95] Raster Separable 2D-CNN ResNet Taxi Shanghai None CTBGCN[96] Graph 2DGCN Conv-LSTM NYC Taxi None CT-GCN[97] Graph GCN ST block DIDI Haikou None Table 1.
Summary of deep learning models in taxi origin-destination prediction.
-
Model Dynamic/Static Directed/Undirected Time window Sparse data Spatial-temporal correlation GEML[74] Static Fluid relationship but undirected Discrete-time snapshots with the same time granularity Multi-granularity level mesh embedding/pre-weighted aggregator None MultiConvLSTM[72] Dynamic and Static Undirected Discrete-time snapshots with the same time granularity Self-attention None CLTS[73] Dynamic Undirected Discrete-time snapshots with the same time granularity None None CSTN[71] Static Undirected Discrete-time snapshots with the same time granularity None None CAS-CNN[76] Static Undirected Discrete-time snapshots with the same time granularity Split the CNN/masking loss function None MPGCN[77] Two static adjacency matrices and one dynamic adjacency Undirected Discrete-time snapshots with the same time granularity None Spatiotemporal dynamic adjacency matrix GCN-SBULSTM[78] Static Undirected Discrete-time snapshots with the same time granularity None None ST-ED-RMGC[79] Static Undirected Discrete-time snapshots with the same time granularity None Parallel prediction DNEAT[80] Dynamic and Static Directed Discrete-time snapshots of different time granularities The loss function $ \mathcal{l} $ $ {\mathcal{l}}_{reg} $ $ {\mathcal{l}}_{mask} $ None Spatial OD-BiConvLSTM[81] Dynamic Undirected Discrete-time snapshots (sliding window) The loss function $ \mathcal{l} $ None CMOD[82] Dynamic Undirected Continuous dynamic time The loss function focuses on non-zero demand Node embedding HMOD[83] Dynamic and Static Directed and semantically differentiated Continuous dynamic time None Node embedding SIZINB-GNN[84] Static Undirected Discrete-time snapshots with the same time granularity Zero expansion negative binomial distribution/probability of learning input being zero additive parameter $ \pi $ None ODformer[86] Static Directed Long sequence time window Transformer Spatial-Temporal Transformers SI-GCN[87] Dynamic Directed Continuous dynamic time Negative sampling data imputation, STGDL[88] Static Directed Discrete-time snapshots (sliding window) Two gate mechanisms ST-GDL model CWGAN-div[89] Dynamic and Static Undirected Discrete-time snapshots (moving average) Residual blocks Interpretable conditional information DMGC-GAN[90] Dynamic and Static Directed Discrete-time snapshots (sliding window) GAN TMGCN Hex D-GCN[91] Dynamic Undirected Discrete-time snapshots (moving average) Filling the lower triangle matrix None OD-TGAT[65] Static Directed Discrete-time snapshots (sliding window) GAT None TFF[92] Dynamic Directed Discrete-time snapshots (moving average) Two-step design performs None CSGCN[93] Static Directed Discrete-time snapshots (sliding window) None None gHMC-STA[94] Static Directed Discrete-time snapshots (sliding window) None None HSTN[95] Static Directed Discrete-time snapshots (moving average) None None CTBGCN[96] Static Directed Discrete-time snapshots (sliding window) None None CT-GCN[97] Static Directed Discrete-time snapshots (sliding window) None None Table 2.
Summary of problem solving in taxi origin-destination prediction.
-
Dataset Links NYC-Taxi https://www.nyc.gov/site/tlc/about/tlc-trip-record-data.page Porto − T-drive (Beijing) https://www.microsoft.com/en-us/research/publication/t-drive-trajectory-data-sample/ Taxi-Shanghai − Taxi-Shenzhen https://opendata.sz.gov.cn/ Taxi-Chengdu https://tianchi.aliyun.com/dataset/39384 Table 3.
Open taxi datasets.
-
Model Github GEML https://github.com/Zekun-Cai/GEML-Origin-Destination-Matrix-Prediction-via-Graph-Convolution CSTN https://github.com/liulingbo918/CSTN FL-GCN https://github.com/alzmxx/OD_Prediction MPGCN https://github.com/underdoc-wang/MPGCN ST-ED-RMGC https://github.com/kejintao/ST-ED-RMGC/tree/main/od_prediction CMOD https://github.com/liangzhehan/cmod HMOD https://github.com/Rising0321/HMOD SIZINB-GNN https://github.com/zhuangdingyi/stzinb Table 4.
Open-source codes.
Figures
(9)
Tables
(4)