Skip to main content

Table 1 Spatiotemporal FSE methods

From: Feature selection and extraction in spatiotemporal traffic forecasting: a systematic literature review

Method Short description Number of studies
Spatial Temporal
Exogenous filtering methods
 All All spatial locations within the research road segment 26
 All upstream All upstream spatial locations within the research road segment 13
 Upstream Only direct upstream neighbour(s) 44
 Upstream + downstream Only direct upstream and downstream neighbours 34
 Downstream Only direct downstream neighbour(s) 4
 Higher order Higher order neighbours (neighbours of neighbours), starting from the second order 7
 Window Several upstream and downstream neighbours 8
 Predefined maximum lag A set of lags {1, 2,  … , T}, where T is a predefined maximum time lag 95
 Travel time With one of the dimensions (spatial or temporal) fixed, the other can be limited by travel time between spatial locations 6 6
 Micro-simulation Estimate spatiotemporal relationships using individual cars’ routes 2
 Network characteristics Use network characteristic (i.e. betweenness centrality and vulnerability) to discover complementary links 3
Endogenous filtering methods
 CCF Cross-correlation function between traffic at different spatial locations 32 26
 Graphical LASSO Graphical least absolute shrinkage and selection operator 4 1
 Granger causality Granger causality tests, incl. Vector autoregressive model 2 2
 LARS Least-angle regression 3 2
 MARS Multivariate adaptive regression splines 3 3
 Custom Authors’ custom formulas (e.g. a combination of physical distance and correlation between spatial locations) 9 2
Wrapper methods
 Empirical Empirical feature selection based on the forecasting model characteristics (information criterion, RMSE, permutation feature importance, etc.) 12 50
 GA Genetic algorithm with spatiotemporal links in a chromosome and the model performance is based on a fitness function 5 3
 PSO Particle swarm optimisation with spatiotemporal links in a candidate solution 2 2
 PSO-GA Combination of genetic algorithm and particle swarm optimisation 1 1
Embedded methods
 LASSO Least absolute shrinkage and selection operator (L1-norm loss function) regularisation 7 3
 MCP, SCAD Maximum concave penalty regularisation
Smoothly clipped absolute deviation regularisation
1 1
 SRM Structural Risk Minimisation 1 1
 Regularised kernel Regularised kernel function (i.e. Laplacian) 1 1
 RBM Restricted Boltzmann machine, usually as part of a deep learning network 2 2
 Sparse AE Sparse autoencoders, usually as part of a deep learning network 1 1
 LSTM Long short-term memory unit stores temporal information for either long or short time periods 5
 Internal Other methodology-specific regularisation 2
Dimension reduction
 Spatial clustering/ Temporal aggregation Clustering of spatial locations using different methods (self-organising maps, empirical grouping, etc.)
For temporal dimension – selection of an appropriate temporal aggregation level
12 1
 PCA-EVD Principal component analysis, based on eigenvalue decomposition 9 8
 PCA-SVD Principal component analysis, based on singular-value decomposition 4 2
 LSDA Local shrunk discriminant analysis 1 1
 NMF Non-negative matrix factorisation 2 1
 SSA Singular spectrum analysis 3 3