Access the full text.

Sign up today, get DeepDyve free for 14 days.

Journal of Advanced Transportation
, Volume 2023 – Jan 5, 2023

/lp/hindawi-publishing-corporation/pgdrt-prediction-demand-based-on-graph-convolutional-network-for-KKr7xnDPJ1

- Publisher
- Hindawi Publishing Corporation
- ISSN
- 0197-6729
- eISSN
- 2042-3195
- DOI
- 10.1155/2023/7152010
- Publisher site
- See Article on Publisher Site

Hindawi Journal of Advanced Transportation Volume 2023, Article ID 7152010, 13 pages https://doi.org/10.1155/2023/7152010 Research Article PGDRT: Prediction Demand Based on Graph Convolutional Network for Regional Demand-Responsive Transport 1 1 2 Eunkyeong Lee , Hosik Choi , and Do-Gyeong Kim Department of Urban Big Data Convergence, University of Seoul, Seoul 02504, Republic of Korea Department of Transportation Engineering & Graduate School, Department of Urban Big Data Convergence, University of Seoul, Seoul 02504, Republic of Korea Correspondence should be addressed to Do-Gyeong Kim; dokkang@uos.ac.kr Received 3 June 2022; Accepted 13 September 2022; Published 5 January 2023 Academic Editor: Yanming Shen Copyright © 2023 Eunkyeong Lee et al. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. To provide an eﬃcient demand-responsive transport (DRT) service, we established a model for predicting regional movement demand that reﬂects spatiotemporal characteristics. DRT facilitates the movement of restricted passengers. However, passengers with restrictions are highly dependent on transportation services, and there are large ﬂuctuations in travel demand based on the region, time, and intermittent demand constraints. Without regional demand predictions, the gaps between the desired boarding times of passengers and the actual boarding times are signiﬁcantly increased, resulting in ineﬃcient transportation services with minimal movement and maximum costs. Therefore, it is necessary to establish a regional demand generation prediction model that reﬂects temporal features for eﬃcient demand response service operations. In this study, a graph convolutional network model that performs demand prediction using spatial and temporal information was developed. The proposed model considers a region’s unique characteristics and the inﬂuence between regions through spatial information, such as the proximity between regions, convenience of transportation, and functional similarity. In addition, three types of temporal characteristics—adjacent visual characteristics, periodic characteristics, and representative characteristics—were deﬁned to reﬂect past demand patterns. With the proposed demand forecasting model, measures can be taken, such as having empty vehicles move to areas where demand is expected or encouraging adjustment of the vehicle’s rest time to avoid congestion. Thus, fast and eﬃcient transportation satisfying the movement demand of passengers with restrictions can be achieved, resulting in sustainable transportation. 1. Introduction in a decrease in use, because the service does not adequately satisfy the requirements of passenger travel. If this phenom- enon persists, supply is concentrated on major routes, which The right to travel refers to citizens’ right to move freely and safely. Because it is a fundamental right that is indispensable can create barriers to passengers’ travel rights. Moreover, socially disadvantaged people (elderly people, disabled peo- to human life, eﬀorts to ensure the right to move continu- ously are needed [1]. Although transportation patterns have ple, residents of vulnerable areas, etc.) can experience severe changed over the past few decades, mainstream passenger isolation. Demand-responsive transport (DRT) services have emerged to solve this problem. transport (e.g., buses and taxis) has not changed suﬃciently to meet these changes. In particular, timed route methods, A DRT service refers to a transportation service that responds to the movement demand of passengers without such as buses, incur ﬁxed operating costs. If a passenger is not picked up, a loss occurs; if the passenger’s demand a predetermined route or operation plan. It combines low changes, the utilization rate decreases, and eventually, the fares, which are the advantages of buses running ﬁxed routes, and convenient boarding and disembarking and ﬁxed cost increases. This leads to a vicious cycle that results 2 Journal of Advanced Transportation speed, which are the advantages of taxis. Therefore, relative (iii) Finally, a graph convolutional network (GCN) was to buses and taxis, DRT services achieve a tradeoﬀ in terms used for the ﬁrst time to reﬂect spatial factors in of eﬃciency and cost. DRT services have the following demand prediction according to the region of the advantages over ﬁxed-route operations. First, the demand DRT service resolution is optimized. For DRT services, the driving dis- The remainder of this paper is organized as follows. tance of a ﬁxed-route vehicle divided by the number of pas- Section 2 presents related research and basic deep- sengers onboard is approximately half that for a ﬁxed-route learning models related to DRT service demand predic- operation. Additionally, DRT services have the advantage of tion. Section 3 presents the proposed method. Section 4 eﬃcient operational cost management. DRT services are presents the results of applying the proposed method to economical because the ﬁxed cost incurred when there is actual data. Section 5 presents conclusions and suggestions no demand is low. Another advantage of DRT services is for further research. their environmental superiority. They have a shorter toler- ance distance than ﬁxed-route vehicles. They are ecofriendly with regard to greenhouse-gas emissions and fuel consump- 2. Related Works and Preliminaries tion because they use small vehicles. Finally, passengers are This section introduces DRT service demand prediction highly satisﬁed with DRT services. DRT services operating research and deep-learning methods. in a door-to-door manner achieve higher levels of passenger satisfaction than ﬁxed-route operations, where passengers must travel directly to the station [2]. 2.1. DRT Service Demand Prediction. Because demand pre- DRT is applied to the movement of passengers with diction must precede the eﬃcient operation of DRT services, restrictions, e.g., in areas where demand is intermittent or many studies have recently been conducted using various transportation services are insuﬃcient and vulnerable [3]. methodologies. For example, in [8], after the entire region Real-time response to the travel demand is crucial for eﬃcient was divided into grids, the demand for a DRT service was DRT service operations, requiring a system and demand fore- predicted using a convolutional neural network (CNN), casting model to allocate requests to vehicles quickly and eﬃ- LSTM, and ConvLSTM, along with exogenous variables ciently when passengers receive travel requests [4]. The such as weather. In [9], an appropriate DRT type was iden- demand forecasting ﬁeld for mainstream passenger transport tiﬁed by estimating the average number of people getting on continues to improve with the development of deep-learning and oﬀ at bus stops in a regular pattern identiﬁed through technologies such as long short-term memory (LSTM). For cluster classiﬁcation of time-by-time boarding points for example, in [5], LSTM was utilized to predict future demand the eﬃcient placement of DRT. according to past demand through traﬃc card data analysis. Recent studies focus on spatial dependence, traveler per- However, DRT services are designed for the movement of pas- sonal heterogeneous, sparse uncertainty, and demand pre- sengers with restrictions; therefore, they exhibit a diﬀerent diction quality requirements. Reference [10] mentioned demand pattern from general mainstream passenger transpor- that variables representing factors related to the characteris- tation. Because the existing liquor passenger transportation tics of service supply, demographic characteristics, land use, model cannot be applied, a model that reﬂects the movement and accessibility should be discovered and fused to reﬂect characteristics of passengers with restrictions is required. the direct impact and ripple eﬀect on demand. Their It is crucial to consider the demand at previous times in research uses a model structure (Attention, ConvLSTM) that the region, but it is also essential to reﬂect spatial character- can demonstrate demand patterns of call taxis for the dis- istics. Each region has spatial characteristics, such as com- abled as a service supply characteristic. In addition, to reﬂect mercial districts and suburban areas [6, 7]. Because spatial demographic characteristics, the administrative region, characteristics aﬀect temporal trends, spatiotemporal factors which is a division of a population-based area, was used as must be considered. In this study, three types of components variables representing factors related to land use and accessi- that reﬂect spatial, temporal, and spatiotemporal character- bility were discovered and utilized as a functional similarity istics were constructed and reﬂected in the model. Because adjacency matrix of the GCN method. However, this paper DRT services are subject to spatiotemporal inﬂuences, the is aimed at developing an optimal bus route rather than a data are sparse. LSTM aﬃliation is not well suited for sparse DRT service. Call taxi for the disabled is a short-distance data. To solve this problem, we used channel-wise attention transportation service for people who cannot go to the and temporal means to alleviate the sparsity of the data to appropriate stop due to severe disabilities. There is a sepa- the greatest extent possible and then used ConvLSTM. rate long-distance customized bus service for the disabled The main contributions of this study are as follows. in Seoul. Therefore, the use of the call taxi for the disabled is diﬀerent. Reference [11] is a thesis that studies the error (i) First, we improved the interpretability of the model distribution rather than speciﬁc parameters, learning by identifying the cause of spatiotemporal demand methods, and hyperparameter adjustments for a transporta- and reﬂecting it in the model tion demand prediction model for adequate public transpor- tation (PT) operation. To build an accurate model, it is (ii) Second, we used channel-wise attention and tempo- necessary to study the error distribution considered in the ral means to maximize the demand for sparse study. References [12, 13] utilized H-ConvLSTM that applies demand response convolution based on a hexagonal shape rather than a Journal of Advanced Transportation 3 conventional pixel standard. We improved the performance in time was proposed. The ﬁrst part is a long-term encoder by using the ensemble for postaggregation, like bagging. To for encoding the past moving demand. The second part is reﬂect the interregional relationship between hexagons, they a short-term encoder for deriving next-step predictions from used the GCN additionally. generated multistep predictions. The third part is an In particular, the traﬃc demand was predicted using call attention-based output module for modeling dynamic tem- taxi data for people with disabilities in Seoul. In [14], the poral and channel-wise information. In [21], ST-Conv waiting times for disabled people in Seoul were predicted block—a combination of temporal-gated convolution and using SARIMA and LSTM and compared. In [15], the call spatial graph convolution—was used to predict the traﬃc taxi latency for the disabled was predicted using several speed at the next point in time. In this study, we predicted hyperparameters of LSTM. However, in these studies, only the demand for a DRT service using the graph-based past temporal characteristics were considered; spatial char- method. acteristics were omitted or reﬂected only in the Euclidean 2.3. GCN. The GCN applies to graph G = ðV, AÞ, where V space. Furthermore, because the spatial relationship is not jVj×jVj refers to vertices and A ∈ ℝ is a matrix with edges based only on the location in Euclidean distance, it is neces- expressing the relationships between the vertices. The GCN sary to reﬂect various spatial structures based on non- can extract a local feature from a non-Euclidean structure Euclidean distance in the model. in another receptive ﬁeld. For example, to utilize convolu- tion in the graph structure, the Fourier transform [22] can 2.2. Spatiotemporal Prediction. Demand prediction and be used. To share the basis of the Fourier transform, we urban traﬃc prediction ﬁelds, such as traﬃc volume predic- compute the Laplacian matrix tion and congestion distribution estimation, exist in tasks that reﬂect spatiotemporal factors. Previous studies on urban −1/2 −1/2 L = I − D AD , ð1Þ traﬃc prediction can be classiﬁed into two categories according to the input data format. Grid-based inﬂow and outﬂow prediction is based on images, whereas graph- where D denotes the degree matrix. We denote X as the fea- th k k based traﬃc speed prediction is based on graphs. tures of the l layer, α as trainable coeﬃcients, L as the k -order multiplier of the graph Laplacian matrix, and σ as 2.2.1. Grid-Based Inﬂow and Outﬂow Prediction. The an activation function. The graph convolution operation demand forecasts for DRT services and taxis are highly [23] using a Laplacian matrix is deﬁned as follows: similar [8]. Therefore, to predict the general taxi demand, the entire area is converted into an image set to a grid of a K−1 speciﬁc size and utilized. In [16], exogenous variables such X = σ 〠 α L X : ð2Þ l+1 k l as weather and weekend availability were added in a fully k=0 connected layer. The values before a certain point, such as the distant, near, recent of the grid, are learned through We learn the relationships between adjacent vertices by convolution. In [6], predetermined point-of-time values updating feature X through multiple layers. Moreover, and point-of-interest (POI) characteristics were learned because the GCN has the characteristics of learning weight by a grid through convolution, such as the time, day, sharing and local features, which are characteristics of the and week of the set grid, and combined through ResPlus CNN, it is possible to obtain a node feature reﬂecting the to predict the regional taxi demand at the next time. Col- connection information of the adjacent (hop) nodes of each lecting exogenous variables that may be related to future node. demand can improve the predictive performance. W×H×C 2.4. Channel-Wise Attention. Given an input X ∈ ℝ , Although weather, POI, or traﬃc ﬂow was used in the channel-wise attention [17, 18] learns the weights for each foregoing studies, the performance improvement was channel to ﬁnd and highlight the most important frame with insigniﬁcant relative to the increase in the number of larger weights. Here, H, W, and C refer to the height, width, parameters, because the improvement through exogenous and channel number of the image, respectively. The variables was orthogonal to capturing complex spatiotem- channel-wise attention is deﬁned as follows. A summary of poral dependence in the data [7]. each channel 2.2.2. Graph-Based Traﬃc Speed Prediction. Graph struc- W H tures—not images—are used to solve various urban prob- z = F ðÞ X = 〠〠 X for c =1, ⋯, C ð3Þ c pool :,:,c i,j,c lems. In contrast to the grid-based method, research is WH i=0 j=0 focused on solving various urban problems, such as predict- ing traﬃc speed, rather than predicting movement demand. is obtained. Then, we obtain the attention s = σðW δðW zÞÞ. 2 1 For example, in [17], a GCN with three adjacency matrices The algorithm learns to assign a large weight to the impor- was used. Spatial characteristics were adopted, along with a tant channels. The attention value to the original input contextual gated recurrent neural network [14, 18, 19] and values is channel-wise as follows: temporal characteristics with values prior to a certain point in time of closeness, period, and trend. In [20], a three- X = X ⊙ s , for c = 1, ⋯, C: ð4Þ :,:,c :,:,c c part model that predicts the travel demand at the next point 4 Journal of Advanced Transportation 37.70°N 37.65°N 37.60°N 37.55°N 37.50°N 37.45°N 126.8°E 126.9°E 127.0°E 127.1°E 127.2°E 0 4 1 5 2 6 (a) 37.70°N 37.65°N 37.60°N 37.55°N 37.50°N 37.45°N 126.8°E 126.9°E 127.0°E 127.1°E 127.2°E NA (b) Figure 1: Distributions of demands at 5 p.m. on November 1, 2019. (a) Distribution of demands (continuous) at 5 p.m. (b) Distribution of demands (0/1) at 5 p.m. Journal of Advanced Transportation 5 (i) Spatial dependency (ii) Temporal dependency (iii) Spatial-temporal encoding relationships time J observations with dependency by region contextual gated ConvLSTM GCN (CGConvLSTM) Various time values (closeness, period, ConvLSTM trend) Neighborhood Prediction ConvLSTM Transportation ConvLSTM Functional similarity Figure 2: Model overview. Require: jVj−1 ðtrendÞ ðperiodÞ ðclosenessÞ 1: Past demands X = ðx , x , x Þ i i i i=1 jVj−1 ðt+1Þ 2: Future true demand y = ðx Þ i i i=0 3: Adjacency matrix A = ðA , A , A Þ, degree matrix D = ðD , D , D Þ, hop: K N T F N T F jVj−1 ðt+1Þ Ensure: future prediction demand yb = ðx̂ Þ i i i=0 4: While training do 5: For all A do 6: (1) Spatial dependency: apply Chebyshev to each adjacency matrix A 7: L⟵ rescale (normalize ðLÞ) 8: For all K do 9: T ⟵ Chebyshev ðL, T Þ k+1 k 10: End for 11: (2) Temporal dependency: apply with contextual gating (CG) and ConvLSTM 12: H ⟵ ConvLSTMðCGðX, T ÞÞ i k+1 13: End for 14: (3) Spatial-temporal dependency: apply with FC (fully connected) and GCN (graph convolution network) 15: y ⟵ FCðGCNðHiÞÞ 16: Compute loss: L = BCELossðyb , y Þ i i 17: End while Algorithm 1: Training procedure of the proposed method. Here, F is a global average pooling operation, and W used. The call taxis were primarily operated in Seoul but pool 1 sometimes moved to areas adjacent to Seoul, depending on and W are the corresponding weights. δ and σ are nonlin- the passenger demand. However, we limited the spatial ear functions for each ReLU, i.e., rectiﬁed linear unit and sig- range to Seoul. Therefore, we included data from both moid function. departure and destination sets within Seoul. The call taxi data for the disabled included the following information. 3. Method For each call, the variables were the type of call (regular 3.1. Description of Dataset. In this study, DRT service data of reception, full-day reservation, and direct call), reception, call taxis for the disabled in Seoul for two years (from 00:00 hope, dispatch, boarding, departure, destination, departure coordinates, customer number, purpose of use, and number on January 1, 2018, to 24:00 on December 31, 2019) were 6 Journal of Advanced Transportation 1.0 1.0 400 400 350 350 0.8 0.8 300 300 0.6 0.6 250 250 200 200 0.4 0.4 150 150 100 100 0.2 0.2 50 50 0 0.0 0 0.0 0 50 100 150 200 250 300 350 400 0 50 100 150 200 250 300 350 400 (a) (b) 1.0 0.8 0.6 0.4 0.2 0 0.0 0 50 100 150 200 250 300 350 400 (c) Figure 3: Adjacency matrices for neighborhood, transportation, and functional similarity. (a) Neighborhood adjacency. (b) Transportation adjacency. (c) Functional-similarity adjacency. of boarding vehicles. Of the 424 administrative districts, 3.2. Proposed Method. The proposed method consists of three steps. The ﬁrst step is to encode the spatial dependency, the Wirye-dong, which had no passenger demand in 2018 or 2019, was excluded. In addition, we excluded data correspond- second step is to use ConvLSTM [24] to reﬂect the temporal ing to hours other than the primary operating hours. There dependency, and the third step is to use a GCN [23] to reﬂect were 1,699,614 data points within 11 h, including 7–17 h. the temporal dependency. Figure 2 illustrates the overall pro- The data contained one row of demand consisting of a cess. Furthermore, pseudocode is presented in Algorithm 1. three-dimensional matrix with 8030 rows and 423 (adminis- 3.3. Encoding Spatial Dependency. The proposed method uti- trative districts) columns in 365 days × 2 ðyearsÞ ×11 (time lizes several types of adjacency matrices to reﬂect the spatial zones) by aggregating the number of demands by administra- dependency. The adjacency matrix A reﬂects the neighbor- tive district in the date-time period. The number of demand hood between administrative districts. cases was continuous data; however, as mentioned previously, the number of demand cases had an extensive and intermit- tent distribution. Zero accounted for 62% of the cases 1, v and v are adjacent, i j (1,064,141 of 1,699,614), one accounted for 21%, and the A = ð5Þ N,i,j others accounted for only 17%. The class imbalance problem 0, otherwise was alleviated by treating multiple demands as one demand (0/1). For example, for 5 p.m. on November 1, 2019, the data exhibited a wide variety of demands, as shown in Figure 1(a). Figure 3(a) shows a heat map of the adjacency matrix for However, the number of demands was changed according to adjacent connections between administrative districts. The whether there was demand, as shown in Figure 1(b). second adjacency matrix A was designed to reﬂect the real T Journal of Advanced Transportation 7 J×|V|×P F (X) ∈ ℝ pool (1) (J) J×|V|×P [X , … , X ] = X ∈ ℝ (1) (J) J×1×P [z , … , z ] = z ∈ ℝ s ∈ ℝ ∀ |V| F attention pool J observations (1) (J) J×|V|×P J observations [H , … , H ] = H ∈ ℝ J×|V|×P F (GCN(X)) ∈ ℝ pool J reweighted observations Figure 4: Contextual gating mechanism of the proposed method. quotient (LQ) measures the dispersion of a speciﬁc industry. We calculate the satisfaction of medical care and disability facilities in administrative districts by com- paring them with Seoul city. It can be interpreted that the higher the coeﬃcient, the higher the satisfaction of the owned facilities compared to other administrative districts, and vice versa—the lower the coeﬃcient, the Conv insuﬃcient. LQ, a quantitative indicator, was used to LSTM compare the functional similarity between the two administrative districts. The adjacency matrix and nor- malized Laplacian matrix for the functional similarity (1) (J) J×|V|×P [H , … , H ] = H ∈ ℝ between the two administrative districts were expressed in a heat map, as shown in Figure 3(c). Chebyshev polynomials [25] were used to embed the conﬁgured adjacency matrix. We transformed the adjacency Temporal mean matrix into a Laplacian matrix as follows: |V| nodes −1/2 −1/2 J reweighted observations L = I − D AD , ð8Þ Figure 5: Contextual gating mechanism. where D is degree matrix, L is normalized graph Lapl- caian matrix, and I is identity matrix. travel distance between the administrative districts. Using k-order Chebyshev polynomials [25], ÀÁ ÀÁ A = max 0, conn v , v − A ∈fg 0, 1 : ð6Þ ÀÁ ÀÁ ÀÁ T,i,j i j N,i,j ~ ~ ~ ~ fAðÞ ; θ = T L =2xT L − T L with T = I, T = L, i k k−1 k−2 0 1 ð9Þ Figure 3(b) shows a heat map of the adjacency matrix for the transportation convenience connection between administrative districts. According to the third encoding. adjacency matrix A , for administrative districts that 3.4. Learning Temporal Dependency. Contextual gates and are more functionally similar, the demand patterns are ConvLSTM deploy temporal dependencies. We use input more similar. values based on closeness, period, and trend. For closeness, we consider the demands from 1, 2, and 3 h in the past. A = I sim P , P >0:9 − A − A ∈ 0, 1 : ð7Þ fg F,i,j v v T,i,j N,i,j i j The period is the same as that of 1, 2, and 3 d in the past. The trend is the demand a week in the past. As shown in Figure 4, contextual gating is performed. Here, simð·Þ denotes cosine similarity. P is a vector of the medical location quotient (LQ), disability LQ, We ﬁrst compute GCN (X) applying GCN to the original number of resident registration disabilities, and demand value. In the model, GCN is applied as follows. Multigraph movements for each administrative district. Location convolution is used, such as equation (10), to reﬂect spatial 8 Journal of Advanced Transportation –10 –20 –30 –20 –10 0 10 20 z1 Cluster 0 Cluster 3 Cluster 1 Cluster 4 Cluster 2 (a) 37.70°N 37.65°N 37.60°N 37.55°N 37.50°N 37.45°N 126.8°E 126.9°E 127.0°E 127.1°E 127.2°E Cluster 0 3 (b) Figure 6: Results for the functional similarity adjacency matrix obtained using t-SNE: (a) t-SNE results for the functional similarity; (b) t-SNE results for the Seoul map. jV j×jVj dependency by utilizing several graphs conﬁgured. Multi- ∈ ℝ is the aggregation matrix of other samples. If W P ×P l l+1 graph convolution is used to reﬂect spatial dependency. X ∈ ℝ is the feature transformation matrix, X is l+1 jVj×P jVj×P l l+1 updated to ∈ ℝ , X ∈ ℝ is feature vectors of region V layer l+1 in l and l +1:σ is activation function and ⊔ is aggregation X = σðÞ ⊔ fAðÞ ; θ X W : ð10Þ function, where is sum. A is a set of graphs, and f ðA ; θ Þ l+1 A∈A i l l z2 Journal of Advanced Transportation 9 7 8 9 10 11 12 13 14 15 16 17 Time of day Demand No demand Figure 7: Percentage of demand occurrences by hour. Table 1: Mean feature vectors by clusters. Cluster Medical LQ Disabled LQ Garage Return home Treatment Rehabilitation Religion Commute Shopping Business 0 1.57 1.07 0.08 1259.99 616.59 416.56 78.55 148.14 12.31 0.33 1 1.46 0.96 0.11 1075.18 814.89 474.19 119.67 138.13 11.78 0.33 2 1.22 0.94 0.07 1218.28 553.75 454.73 90.22 89.83 10.15 0.23 3 1.53 1.00 0.09 1200.80 1200.80 717.32 59.14 141.73 21.75 0.55 4 1.61 1.08 0.10 1319.93 1319.93 405.94 68.83 97.03 12.47 0.49 Then, we apply global average pooling to all regions. is as follows: ðÞ j ðÞ closeness ðÞ period ðÞ trend ~ ~ ~ ~ X = mean X , mean X , mean X : jj V ðÞ j ðÞ j ðÞ z = 〠 F X + F GCN X , for j =1, ⋯, J : pool pool i i V ð14Þ jj i=1 ð11Þ We learn the temporal characteristics of each region using ConvLSTM in the temporal mean reweighted obser- vations. Owing to intermittent demand, we convert sparse Let σ be a sigmoid function and δ be the GeLU, i.e., Ga data into dense data. Therefore, we average features for linear unit function. Equation (11) produces the following closeness, period, and trend and then apply ConvLSTM. summary: Across all regions, ConvLSTM is applied to the values of the reweighted observations. This results in a single vector ðÞ 1 ðÞ 2 ðÞ J that aggregates the learned spatiotemporal information. s = s , s ⋯,s = σ W δ W z , ð12Þ ðÞ ðÞ 2 1 1 j ðÞ ðÞ ~ ~ H = convlstm X ,⋯,X , for i =1, ⋯, V : ð15Þ jj for each of the temporary observation periods. We multi- plied the calculated summary by the original value. Finally, a multigraph GCN is applied to the result of the ConvLSTM to learn spatiotemporal characteristics simulta- ðÞ j ðÞ j ðÞ j neously. We then apply a fully connected layer for aggregation. X = X ⊙ s , for j =1, ⋯, J : ð13Þ ̂y =FCððÞ GCNðÞ H : ð16Þ Through the contextual gating mechanism, we obtain reweighted observations with weights over time. 4. Spatiotemporal Characteristics of DRT Service However, the LSTM architecture may not be well learned from sparse data. To resolve this, we applied In the proposed method, the regional demand for DRT ser- ConvLSTM after the temporal mean, as shown in vices is predicted via graph-based deep learning using the Figure 5. For each of the three inputs, the temporal mean spatiotemporal characteristics of the demand in the past Percentage of demand by hour (%) 10 Journal of Advanced Transportation 7 8 9 101112 13 14 15 16 17 789 10 11 12 13 14 15 16 17 Time of day Time of day Purpose Call type Return home Rehabiltation Direct call Etc. Religion Full-day reservation Shopping Treatment Regular reservation Business Commute (a) (b) Figure 8: Percentage of demand for diﬀerent (a) purposes of use and (b) call types based on hours. Table 2: Comparison of various methods. two years. Therefore, it is necessary to investigate the cause of the presence or absence of demand. Accessibility to the Model Accuracy Precision Recall F1 score DRT service is inﬂuenced mainly by time and space, as HA 65.18 36.74 55.29 44.15 shown in Figures 6 and 7. In this section, the factors that aﬀect the demand for transportation services are identiﬁed Logistic regression 68.04 30.92 66.07 42.13 through temporal and spatial characteristic analyses. XGboost [28] 69.82 42.47 65.19 51.43 Our method 78.13 75.41 62.26 68.21 4.1. Analysis of Spatial Characteristics. To visually validate the spatial dependency embedded vectors of the functional similarity adjacency matrix, we used t-distributed stochastic cial to predict the demand in the period when the demand is plummeting, as most administrative districts exhibited a neighbor embedding (t-SNE) [26] over a low-dimensional space. Then, we applied k-means clustering [27] to the lower demand of >50% at 7 a.m. These results are attributed to the purpose of passenger use. dimensions. We performed dimension reduction with t-SNE Figure 8 shows the usage purpose pattern: the number of for visualization and observed ﬁve clusters, as shown in Figure 6. Table 1 presents the mean feature vectors for each people returning home increased by 12 p.m., and the demand for treatment, rehabilitation, and commuting/work increased group. Group 0 shows the residential area with the most passen- in the morning. In the case of movement for this purpose, gers boarding to commute. Meanwhile, there is a moderate because the movement is often constant, it is possible to pre- dict the demand position using this pattern. A functionally demand for the rest of the purposes. In the case of group 1, the number of garages is relatively large, and it is a resi- similar adjacency matrix can explain this pattern. According to the ratio of call types by time, direct calls dential area where people board the most for returning home and religious purposes. In the case of group 2, the and full-day reservations were inversely proportional in the case of full-day reservations. Therefore, we infer that dis- medical LQ and disability LQ are low, and they do not board abled call taxis operate regularly. We make three policy sug- well for business work and treatment purposes. In the case of group 3, many people used DRT service for returning home, gestions. First, the demand should be checked on the previous date by expanding the operating time zone of the rehabilitation, and shopping, and the pursuit of work was relatively high. Finally, in the case of group 4, the medical full-day reservation. Currently, the service is only operated at 7 a.m., 8 a.m., and 10 a.m. However, the demand should LQ and disability LQ are high, and the residential area tends be predicted by expanding the operating hours or establish- to have the highest purpose of returning home. ing a system that can be ﬂexibly received the reservation before anytime. Second, movement should be encouraged 4.2. Analysis of Temporal Characteristics. As shown in Figure 7, aggregating the demand status for the two years by utilizing measures such as deploying additional temporal vehicles at 7 a.m., when the demand is the highest. Third, by the hour revealed that 7 a.m. was the most in demand maximum movement should be achieved at the minimum and shows a decreasing trend at 8 a.m. and 9 a.m. However, it increases again from 10 a.m. and then to decrease to 20% cost by avoiding and adjusting the driver’s rest time between 10 a.m. and 12 p.m., when the demand increases again. from 1 p.m. to 5 p.m. Because of this characteristic, it is cru- Percentage by call type (%) Percentage by purpose (%) Journal of Advanced Transportation 11 Table 3: Eﬀect of adding components to the spatial correlation modeling on the performance. Component Accuracy Precision Recall F1 score Neighborhood 77.89 76.50 59.65 67.03 Neighborhood+transportation 77.75 75.54 60.56 67.22 Neighborhood+transportation+functional 78.13 75.41 62.26 68.21 Table 4: Eﬀects of temporal correlation modeling. Table 7: Mean waiting time depending on whether there is a vacant vehicle that exists or not (min). Temporal Accuracy Precision Recall F1 score Time Exist Nonexist Diﬀerence (nonexist-exist) Average pooling 76.60 71.46 63.11 67.02 7 49.23 57.96 8.73 Max pooling 77.56 72.71 64.72 68.49 8 46.58 54.34 7.76 LSTM 77.45 73.40 62.99 67.80 9 42.14 40.32 -1.82 ConvLSTM 78.13 75.41 62.26 68.21 10 26.95 48.24 21.29 11 28.81 44.96 16.15 Table 5: Eﬀects of time combinations. 12 30.16 42.07 11.91 13 32.91 35.63 2.72 (# closeness, # period, # F1 J Accuracy Precision Recall temporal) score 14 33.77 34.78 1.01 7 (3, 3, 1) 78.13 75.41 62.26 68.21 15 31.74 29.38 -2.36 5 (2, 2, 1) 77.83 74.07 63.31 68.27 16 34.50 30.36 -4.14 3 (1, 1, 1) 77.12 73.14 62.08 67.16 17 28.17 34.55 6.38 2 (0, 1, 1) 70.02 62.17 52.76 57.08 Total 33.91 49.71 15.8 2 (1, 0, 1) 70.24 62.69 52.48 57.13 2 (1, 1, 0) 71.25 65.15 51.40 57.46 were input as data conﬁgurations, and the characteristics of each administrative district (medical LQ, disability LQ, Table 6: Measures according to K. etc.) were added. Three adjacency matrices were used, and the results of the experiment are presented in Table 3. The K Accuracy Precision Recall F1 score ﬁrst row presents the results obtained using only the neigh- 2 77.95 73.96 64.02 68.63 borhood adjacency matrix. The second row presents the 3 78.13 75.41 62.26 68.21 results obtained using two transportation adjacency matri- 4 77.87 73.38 64.77 68.81 ces: the neighborhood and transportation adjacency matri- ces. The third row presents the results obtained using all three functional adjacency matrices, i.e., neighborhood, 4.3. Model Performance Comparison. In this section, we transportation, and functional similarity. compare the two aforementioned models. Let ̂y =Pr(X ) i i As shown, the method exhibited the best performance be the conditional probability given an input x . For a loss i when all three adjacency matrices were used. However, in of observation, we used the binary-cross entropy loss. the case of the second row, the performance was inferior to that achieved using only the neighborhood adjacency matrix. Table 4 presents the performance with respect to the L = − 〠½Š y · logðÞ ̂y +1ðÞ − y · logðÞ 1 − ̂y : ð17Þ BCE type of temporal correlation. ConvLSTM outperformed i i i i i=1 vanilla LSTM, which did not reﬂect the spatial information. Also, max pooling shows lower performance. The training dataset included data from January 1, 2018, The performance diﬀerences for diﬀerent combinations to October 31, 2019. Twenty percent of the data were used of closeness, period, and trend are presented in Table 5. As for the validation. Data from November 1, 2019, to Decem- time was used more, performance increased. In the case of ber 31, 2019, were used as test data. To maintain chronolog- call taxi data for the disabled, the demand is very intermit- ical order, the data were not shuﬄed. ConvLSTM had four tent, so the less time is used, the greater the sparse value will hidden sizes and three layers, and the GCN had 64 hidden be aﬀected. In addition, in the case of the demand a week sizes. ago, the actual past information is excessively required; The performance of the proposed method was compared therefore, the demand was ﬁxed to 1. The performance dif- with that of other methods, and the results are presented in ference when using the performance diﬀerence according Table 2. Compared with the existing time series and classiﬁ- to the use of K is presented in Table 6. cation model, the proposed method achieved signiﬁcantly In the GCN, problems such as oversmoothing occur as better performance. In contrast to the other methodologies, the number of layers K increases excessively [29]. Similarly, previous time zones, e.g., the closeness, period, and trend, in this study, when K increased by four or more, the 12 Journal of Advanced Transportation performance was degraded. Finally, in the case of Seoul, if it [2] F. M. Coutinho, N. van Oort, Z. Christoforou, M. J. Alonso- González, O. Cats, and S. Hoogendoorn, “Impacts of replacing is inﬂuenced by too many hops, the performance is a ﬁxed public transport line by a demand responsive transport degraded, reﬂecting irrelevant administrative district system: Case study of a rural area in Amsterdam,” Research in relationships. Transportation Economics, vol. 83, article 100910, 2020. At the time of demand generation, we investigate the [3] K. M. Nahiduzzaman, T. Campisi, A. M. Shotorbani, K. Assi, average diﬀerence time in waiting time between the case K. Hewage, and R. Sadiq, “Inﬂuence of Socio-Cultural Attri- where the empty car is waiting and the case where there is butes on Stigmatizing Public Transport in Saudi Arabia,” Sus- no waiting. Table 7 shows the mean waiting time depending tainability, vol. 13, no. 21, article 12075, 2021. on whether there is a vacant vehicle that exists or not. When [4] S. Jain, N. Ronald, R. Thompson, and S. Winter, “Predicting an empty car is on standby, we expect that we could reduce susceptibility to use demand responsive transport using demo- waiting time by about 16 minutes on average. graphic and trip characteristics of the population,” Travel Behaviour and Society, vol. 6, pp. 44–56, 2017. 5. Conclusions [5] F. Toqué, M. Khouadjia, E. Come, M. Trepanier, and L. Oukhellou, “Short & long term forecasting of multimodal The proposed method can resolve unequal waiting times transport passenger ﬂows with machine learning methods,” between regions by predicting the demand location for eﬃ- in In2017 IEEE 20th International Conference on Intelligent cient operation of DRT services, which can support mini- Transportation Systems (ITSC), pp. 560–566, Yokohama, mum cost–maximum movement. The objective of this Japan, 2017. study was to reduce the waiting time by eﬃciently rearran- [6] Z. Lin, J. Feng, Z. Lu, Y. Li, and D. Jin, “Deepstn+: context- ging nearby empty cars by predicting the regional demand aware spatial- temporal neural network for crowd ﬂow predic- for Seoul’s call taxi service for the disabled, which has inter- tion in metropolis,” Proceedings of the AAAI Conference on Artiﬁcial Intelligence, vol. 33, no. 1, pp. 1020–1027, 2019. mittent call characteristics. After conﬁguring various sub- graphs, the GCN was used to reﬂect the spatial [7] D. Lee, S. Jung, Y. Cheon, D. Kim, and S. You, “Demand fore- characteristics between regions, and the model was con- cast- ing from spatiotemporal data with graph networks and temporal-guided embedding,” 2019, http://arxiv.org/abs/ structed using the temporal mean and ConvLSTM to reﬂect 1905.10709. temporal characteristics. Using various subgraphs from real [8] E. Chandakas, “On demand forecasting of demand-responsive data analysis showed alleviated results in terms of accuracy paratransit services with prior reservations,” Transportation and interpretation. We expect improved convenience of Research Part C: Emerging Technologies, vol. 120, article movement and satisfaction with public transportation by 102817, 2020. reducing the waiting time. In addition, DRT services can [9] J. Choi, J. Song, M. Kang, and K. Hwang, A clustering analysis replace public buses, increase the eﬃciency of subsidies for on the machine learning-based bus routes types and bus stations various types of public transportation, and generate proﬁts for adoption of demand responsive transport, Korean Society and labor inducement eﬀects for transportation companies, Of Transportation, 2021. revitalizing the local economy and increasing the sharing [10] J. Wang, T. Yamamoto, and K. Liu, “Spatial dependence and rate of public transportation. spillover eﬀects in customized bus demand: empirical evidence using spatial dynamic panel models,” Transport Policy, Data Availability vol. 105, pp. 166–180, 2021. [11] I. Peled, K. Lee, J. Yu, J. Dauwels, and F. C. Pereira, “On the Data used to support the ﬁndings of this study are available quality requirements of demand prediction for dynamic public from the corresponding author upon request. transport,” Communications in Trans- portation Research, vol. 1, article 100008, 2021. Conflicts of Interest [12] Z. Chen, K. Liu, J. Wang, and T. Yamamoto, “H-convlstm- based bagging learning approach for ride-hailing demand pre- The authors declare no conﬂicts of interest regarding the diction considering imbalance problems and sparse uncer- publication of this paper. tainty,” Transportation Research Part C: Emerging Technologies, vol. 140, p. 103709, 2022. [13] Z. Chen, K. Liu, and T. Feng, “Examine the prediction error of Acknowledgments ride-hailing travel demands with various ignored sparse demand eﬀects,” Journal of Advanced Transportation, This study was supported by the Basic Study and Interdisci- vol. 2022, Article ID 7690309, 11 pages, 2022. plinary R&D Foundation Fund of the University of Seoul (2021). The authors express their gratitude for this support. [14] G. Han, S. Ha, J. J. Hong, and C. Lee, A study on the application of demand forecasting for call taxi for the disabled in Seoul using deep learning, Korean Society Of Transportation, 2017. References [15] D. Hong, G. Han, S. Ha, and C. Lee, Optimum hyperparameter selection of lstm model for call taxi waiting time for persons with [1] J.-H. Son, D.-G. Kim, E. Lee, and H. Choi, “Investigating the disabilities in Korea, Korean Society Of Transportation, 2018. spatiotemporal imbalance of accessibility to demand respon- sive transit (drt) service for people with disabilities: explana- [16] J. Zhang, Z. Yu, and D. Qi, “Deep spatio-temporal residual net- tory case study in South Korea,” Journal of Advanced works for citywide crowd ﬂows prediction,” 2016, http://arxi- Transportation, vol. 2022, Article ID 6806947, 9 pages, 2022. v.org/abs/1610.00081. Journal of Advanced Transportation 13 [17] G. Xu, Y. Li, L. Wang et al., “Spa- tiotemporal multi-graph convolution network for ride-hailing demand forecasting,” Proceedings of the AAAI Conference on Artiﬁcial Intelligence, vol. 33, no. 1, pp. 3656–3663, 2019. [18] J. Hu, L. Shen, and G. Sun, “Squeeze-and-excitation networks,” 2017, http://arxiv.org/abs/1709.01507. [19] L. Chen, H. Zhang, J. Xiao, L. Nie, J. Shao, and T.-S. Chua, “SCA-CNN: spatial and channel-wise attention in convolu- tional networks for image captioning,” 2016, http://arxiv.org/ abs/1611.05594. [20] L. Bai, L. Yao, S. S. Kanhere, X. Wang, and Q. Z. Sheng, “Stg2seq: spatial- temporal graph to sequence model for multi-step passenger demand forecasting,” 2019, http://arxi- v.org/abs/1905.10069. [21] B. Yu, H. Yin, and Z. Zhu, “Spatio-temporal graph convolu- tional neural net- work: a deep learning framework for traﬃc forecasting,” 2017, http://arxiv.org/abs/1709.04875. [22] D. K. Hammond, P. Vandergheynst, and R. Gribonval, “Wave- lets on graphs via spectral graph theory,” Applied and Compu- tational Harmonic Analysis, vol. 30, no. 2, pp. 129–150, 2011. [23] T. N. Kipf and M. Welling, “Semi-supervised classiﬁcation with graph convolutional networks,” 2016, http://arxiv.org/ abs/1609.02907. [24] X. Shi, Z. Chen, H. Wang, D.-Y. Yeung, W.-K. Wong, and W.- c. Woo, “Convolutional LSTM network: a machine learning approach for precipitation nowcasting,” 2015, http://arxi- v.org/abs/1506.04214. [25] M. Deﬀerrard, X. Bresson, and P. Vandergheynst, “Convolu- tional neural networks on graphs with fast localized spectral ﬁltering,” Advances In Neural Information Processing Systems, vol. 29, 2016. [26] L. van der Maaten and G. Hinton, “Visualizing data using t- sne,” Journal of Machine Learning Research, vol. 9, no. 86, pp. 2579–2605, 2008. [27] S. Lloyd, “Least squares quantization in pcm,” IEEE Transac- tions on Information Theory, vol. 28, no. 2, pp. 129–137, 1982. [28] T. Chen and C. Guestrin, “Xgboost: A scalable tree boosting system,” InProceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining, vol. 13, pp. 785–794, 2016. [29] U. Alon and E. Yahav, “On the bottleneck of graph neural net- works and its practical implications,” 2020, http://arxiv.org/ abs/2006.05205.

Journal of Advanced Transportation – Hindawi Publishing Corporation

**Published: ** Jan 5, 2023

Loading...

You can share this free article with as many people as you like with the url below! We hope you enjoy this feature!

Read and print from thousands of top scholarly journals.

System error. Please try again!

Already have an account? Log in

Bookmark this article. You can see your Bookmarks on your DeepDyve Library.

To save an article, **log in** first, or **sign up** for a DeepDyve account if you don’t already have one.

Copy and paste the desired citation format or use the link below to download a file formatted for EndNote

Access the full text.

Sign up today, get DeepDyve free for 14 days.

All DeepDyve websites use cookies to improve your online experience. They were placed on your computer when you launched this website. You can change your cookie settings through your browser.