This paper presents a novel approach for evaluating road trafic usage using multi-type Geographical Cellular Trafic (GCT). Working with a major telecom company, we propose a new prediction task for transportation trafic using GCT data. To accurately tackle this task, we propose a model that efectively integrates multivariate relation exploration and spatio-temporal modeling across multiple regions. Furthermore, we develop a new core as the foundation of each modeling component, eficiently improving the incorporation of attention mechanisms in the CNN-based architecture. Extensive experiments demonstrate the superior performance of our model in successfully handling the prediction task and reveal the influence of various GCT combinations. It is worth noting that our proposed data and model can pave a new path for intelligent transportation systems and urban planning.
ally, we present a novel core for each View Modeling, designed to enhance the eficiency of the attention mechRecently, trafic prediction has become increasingly im- anism when processing convolution-encoded represenportant for intelligent transportation systems [1, 2]. Ac- tations. Our experiments demonstrate the superior percurate trafic prediction can help alleviate trafic conges- formance of our model compared to representative and tion [3] and optimize trafic signal control [ 4]. However, state-of-the-art baselines, underscoring the importance traditional trafic prediction approaches rely on dedicated of incorporating multi-type GCT data. Overall, this work sensors, which require costly maintenance and develop- makes the following key contributions: ment, have limited deployment coverage, and are susceptible to insuficient usable information. • Novel data: We collected over 30 million GCT
To tackle the limitations of traditional trafic predic- records from diverse road segments, analyzing tion, we leverage large-scale and widely-distributed mo- spatial correlations, relationships among GCT bile user data integrated with road network information types, and their evolution over time. to analyze trafic usage conditions. Collaborating with • Prediction task and model: Our novel task Taiwan’s major telecom provider, Chunghwa Telecom, predicts V-GCT, which provides valuable transwe utilize geolocated cellular trafic, named Geographi- portation insights for city authorities and is becal Cellular Trafic (GCT) , which is further classified into ing employed in a proof-of-concept area. Meanvehicle, pedestrian, and stationary types. Accumulating while, our multivariate spatio-temporal model efGCT at fixed intervals ofers insights into human ac- fectively captures dependencies and relationships tivity patterns and road network usage, defined as GCT between GCT types for accurate predictions. lfow. Consequently, we propose a new task of forecasting • Experiments and analysis: Extensive evaluaspecific vehicular GCT (V-GCT) flow in various regions, tions demonstrate our model’s superior perforwhich is highly related to road trafic conditions and mance against baselines for diverse prediction difers from predicting cellular trafic usage for mobile intervals. Ablation and sensitivity analyses highnetworks in previous studies [5, 6, 7, 8]. Hence, our pro- light the importance of model components and posed task and dataset using mobile user data ofer new GCT flow combinations for adaptability and realinsights into road network usage and trafic conditions. world potential.
To address this new task, we propose a model with Multivariate, Temporal, and Spatial View Modelings that integrates multi-type GCT and utilizes spatio-temporal 2. Data Processing correlations to predict V-GCT flow accurately. Addition
Type vehicle pedestrian stationary
Multi-Type Geographical Cellular Trafic (GCT). GCT is cellular trafic with estimated GPS coordinates obtained from triangulation, indicating where the trafifc was generated. Each GCT is classified 1 into three categories: vehicle, pedestrian, and stationary. GCT Flow. We define GCT flow as the total quantity of GCT within a fixed interval (e.g., 5 min) as in previous vehicular flow studies [ 9]. With multi-type GCT, there are various GCT flows, including vehicle (V-GCT), pedestrian (P-GCT), and stationary (S-GCT) flows.
Road Segments. Road segments are defined as 20m x 20m areas, based on the average road size in our proof-ofconcept (POC) area in Hsinchu, Taiwan. These segments geographically interconnect, forming a road network. to maintain consistency, and assigned unique V-GCT, P-GCT, and S-GCT flows to each road segment. Data Privacy Protection. In compliance with strict personal data protection laws, we established a collaboration agreement with the company to outline data sharing terms and ensure adherence to privacy regulations. We processed GCT data in a secure intranet environment and hashed IMEI numbers to protect the privacy users.
2.3.1. Time Evolving Spatial Correlations Data sourcing. We extracted essential data fields from Spatial Correlation. Building on the approach used in the telecom company’s Geographical Cellular Trafic Stor- [10, 11], we utilized the Pearson correlation coeficient age Database to reduce storage and computational re- to assess the spatial correlation between road segments quirements, as shown in Table 1. Each row represents one in the POC area. Specifically, when using the historical GCT, comprising five data fields: International Mobile one-hour V-GCT as the series variable for each segment Station Equipment Identity (IMEI, a unique mobile phone in Figure 2, a Pearson correlation coeficient is assigned identifier), latitude and longitude coordinates, recording to each road segment pair, ranging from -1 to 1. Road segtime, and GCT type. ments in closer proximity tend to exhibit similar V-GCT Road Segment Selection for GCT Collection. We flow patterns, leading to higher Pearson correlation coefselected geographically connected road segments as the ifcients. This highlights the spatial correlation between scope for GCT data collection to capture the mobility of road segments based on V-GCT. mobile users across diferent areas, as shown in Figure Time Evolving Correlation. We use the Pearson cor1. We collaborated with transportation authorities to relation coeficient to explore spatial correlations over identify 21 road segments for analysis, each with unique time, as shown in Figure 2. Notable observations include: functional locations nearby (e.g., universities, science - At 18:00, road segments near the Hsinchu Science parks, and shopping areas). To ensure relevance to road Park (ID: 57, 59, 62) exhibit high Pearson coeficients, inusage conditions, we extracted GCTs located within these dicating similar movement patterns as commuting users road segments and included them in Table 1. leave the workspace.
Preprocessing for GCT Flow. We preprocessed the - At 19:00, highly correlated regions emerge along the data to ensure accuracy by eliminating inconsistencies, route (ID: 44, 35, 6, 30, 43, 45) from the work area to outliers, and missing values. Next, we removed dupli- residential areas. cate GCT record with identical IMEI and timestamps - By 20:00, users gradually return home or dine out, to prevent distortion and ensure reliability. Finally, we resulting in high Pearson coeficients for road segments calculated the total GCTs for every 5-minute interval near residential-commercial mixed areas (ID: 1, 16, 41, 43) reflecting similar V-GCT patterns. Road segments 1The algorithm is the telecom company’s confidential trade secret. near the Hsinchu Science Park (ID: 54, 56, 57, 59, 62) also exhibit high Pearson coeficients again, as users who ifnish work later leave the area.
Overall, the time-evolving Pearson correlation analysis not only reveals spatial correlations between road segments by V-GCT flow but also indicates changes in population activity patterns during diferent periods. Concentrated regions with high Pearson coeficients may shift over time, providing a new insight for understanding user flow and identifying congestion points in trafic management over time. 2.3.2. Understanding Regional Functionality
through Multi-Type GCT Flows
tasks [12], but traditionally involves time-consuming manual labeling. By analyzing variations between three types of GCT flows, we can uncover hidden interactions of user groups in diferent urban areas.
Implicit Interactions among Multi-Type GCT Flows.
Figure 3(a) and 3(b) reveal distinct patterns in diferent urban areas, with commuting areas displaying dominant V-GCT flow and residential-commercial areas showing significant P-GCT and S-GCT flows. Inspired by [ 13], we subtracted V-GCT from P-GCT and S-GCT to obtain (V-P)-GCT and (V-S)-GCT, respectively, using one day of 2.4. Potential Applications data, and calculated the Pearson correlation coeficient between V-GCT and these flows. The subtraction types, Multi-type GCT flows provide new insights for urban capturing relative diferences between the GCT flows, planning, with possible future applications including: exhibit higher correlations with V-GCT, highlighting dis- Transportation Management. GCT flow assists autinct patterns with unique characteristics in diferent thorities in developing efective trafic strategies, improvareas. For instance, a high (V-P)-GCT value may indi- ing flow and reducing travel times. cate a region with more vehicular trafic. These relative Public Safety. Real-time monitoring systems using GCT diferences better capture GCT flow interactions. lfow aid in understanding crowd density and mobility, Deriving Insights into Model Design. The higher cor- ensuring safety during critical incidents. relations for subtraction types in multi-type GCT flows Urban Planning. Analyzing GCT flow helps planners provide valuable insights for our model. Incorporating identify infrastructure needs and optimize city layouts these relative diferences improves capturing interactions, to accommodate growing populations. and distinguishing area functionality.
at fixed intervals. Each historical GCT flow is represented by = {1 , 1 , ..., }, where corresponds to V-GCT, S-GCT, or P-GCT, and ∈ R denotes the values at time step . Our objective is to predict V-GCT for all segments over the next steps using one or multiple GCT flows from the past steps. We denote the predicted values as { +1, +2, . . . , + }, where + ∈ R .
basis for the multivariate, spatial, and temporal modeling
components. This core aims to improve the eficiency
roefpartetseenntitoantiomneacfhtearniCsmNNsiennhcaonddinlign.g the multi-channel ˆ = ‖=1( ( ∑︁ ℎ)), (1)
Preliminary. Graph Convolutional Network(GCN)- ∈
based models [
The goal of Multivariate View Modeling is to investigate Temporal View Modeling, as shown in the TS-Module of
the relationships among multi-type GCT flows before Figure 4, converts the multi-channel representation into
the TS-Module, and extract implicit relations to enhance the shape [, , ] for GCSAT processing. GCSAT treats
V-GCT prediction. Although existing GAT-based models input as nodes with dimensions, and in Equation
have shown improved task accuracy by exploring fea- 2 represents a complete graph.
ture relationships [20, 21], attention mechanisms may Time series data in practical scenarios often exhibit
not fully utilize the potential diferences within various both short-term and long-term dependencies.
Simultafeatures during a limited number of epochs [13]. neously capturing these patterns using attention among
Deriving Insights from Multi-type GCT Flows. We time nodes is challenging due to entangled dependencies,
have verified that there is a more robust correlation be- making it dificult to identify valuable signals [23, 17].
tween V-GCT and the relative diferences of P-GCT and Extracting Diferent Time Scales To address the above
S-GCT, as shown by the Pearson coeficients displayed issue, we adopt two kernels with diferent sizes, inspired
in Figure 3. Inspired by anomaly detection [20, 13, 22], by [
GCSAT for Multivariate Diference Modeling. After establishing the diference representation, Δ is applied to GCSAT in Equation 2, as follows: where the outputs for diferent time scale representations in Equation 5 are truncated to match the temporal nodes of the representation with the largest kernel size and then concatenated accordingly.
Gating Mechanism. Leveraging the gating mechanism’s benefits in [ 15, 17], which controls the amount of information passed to the next module, we process the outputs of two Equation 5 separately with distinct activation functions and perform element-wise multiplication: = (1) ⊙ (2),
(6) where 1 and 2 are separately generated in Equation 5, denotes the sigmoid function, denotes the tangent hyperbolic function, and ⊙ represents the Hadamard product. is the output of the Gating Mechanism and will be fed into the next modeling.
As the analysis in Figure 2 demonstrates the existence of spatial correlations between V-GCT flows among road segments. Thus, it is crucial to explore the relationships between road segments. Spatial View Modeling, depicted ˆΔ = (Δ, ), where each element in Δ is considered as the feature node, is a complete graph, and GCSAT is stacked with two layers for more detail extraction.
For the output of GCSAT at each time step, we only extract the first node of ˆΔ, ˆΔ[:, 0, :], which is the aggregated V-GCT representation with a shape [, 1, ]. Then, we concatenate the outputs for each time step in Equation 3 along the second dimension, resulting in a shape [, , ]: (3) (4) = ‖=1(ˆΔ[:, 0, :]),
By capturing complex interactions and relationships
between GCT flows, our approach can achieve more
accurate V-GCT flow predictions. The output is then
forwarded to the next Temporal View Modeling.
in the TS-Module of Figure 4, aims to model these rela- Modeling and a Spatial View Modeling component. We
tionships to improve our understanding and prediction also followed [
GCSATs for modeling bidirectional flow. Leveraging GCSAT’s flexibility, we extract spatial correlations among 4.2. V-GCT Prediction Evaluation road segments. First, we transform the representation from the previous temporal modeling output into a We evaluated our model and various baselines for predictshape of [,, ], with road segments as nodes and ing future V-GCT flows at 15 (3 steps), 30 (6 steps), and features. 60-minute (12 steps) intervals, and the results are shown
To account for bidirectional V-GCT flow among road in Table 2, including the average MAE, RMSE, and MAPE
segments, we utilize two diferent GCSATs to explore over 10 repetitions for each method. Our observations
propagations in both directions. The adjacency matrix are as follows:
of road segments is constructed using road network Performance comparison across prediction steps.
distance and a thresholded Gaussian kernel [24]. Fol- Our model consistently outperformed various baselines,
lowing [15], we define the forward transition matrix including TCN-based [26], GCN-based [
(MAE), Root Mean Squared Error (RMSE), and Mean Absolute Percentage Error (MAPE).
Train/Valid/Test data processing. Each type of GCT lfow was processed in 5-minute intervals, from August 28, 2022, to September 29, 2022, across 21 road segments. We followed [25], splitting data 70%-20%-10% for training, testing, and validation. Each sequence sample had 24 time steps; the first 12 ( ) as historical input and the remaining 12 () as ground truth.
Baselines. We have selected seven representative trafic
prediction baselines for this task. These baselines are
categorized as follows, with overviews in Appendix A:
We conducted an ablation study to assess the impact of
our model’s components on trafic prediction tasks.
Average prediction metrics were calculated for prediction
steps 1 (5 min.) through 12 (60 min.). Table 3 compares
the full model with three ablated versions: without
Spatial View Modeling (w/o S), without Temporal View
Modeling (w/o T), and without Multivariate View Modeling
(w/o M). Our observations are as follows:
• Temporal Convolution (TCN) [26] Impact of w/o S. Omitting Spatial View Modeling had
• GCN-based models: Graph Wavenet (GWNet) the most significant impact on the model’s performance,
[15], MTGNN [
15 min. 30 min. 60 min.
MAE RMSE MAPE MAE RMSE MAPE MAE RMSE MAPE 5.55± 0.02 8.82± 0.04 34.5%± 0.6 5.74± 0.02 9.38± 0.06 36.9%± 0.9 6.58± 0.05 11.22± 0.13 38.7%± 1.4 5.46± 0.01 8.72± 0.04 32.5%± 1.3 5.62± 0.03 9.08± 0.11 32.9%± 1.3 6.08± 0.06 10.31± 0.19 34.6%± 1.6 5.37± 0.01 8.61± 0.05 32.6%± 1.7 5.51± 0.04 8.99± 0.14 34.1%± 1.2 5.77± 0.02 9.68± 0.11 34.4%± 1.0 5.29± 0.02 8.52± 0.03 32.2%± 1.4 5.45± 0.01 8.86± 0.01 34.2%± 1.6 5.74± 0.03 9.66± 0.13 35.3%± 1.9 5.30± 0.01 8.46± 0.04 32.9%± 2.0 5.44± 0.03 8.84± 0.09 34.7%± 1.6 5.73± 0.03 9.68± 0.06 34.8%± 2.1 5.28± 0.04 8.48± 0.11 31.8%± 1.7 5.46± 0.01 8.81± 0.03 33.6%± 1.9 5.82± 0.02 9.56± 0.08 34.6%± 1.1 5.26± 0.03 8.43± 0.03 31.1%± 1.1 5.40± 0.02 8.76± 0.06 31.8%± 1.3 5.65± 0.02 9.46± 0.09 32.9%± 1.9 5.23± 0.01 8.27± 0.06 29.8%± 0.7 5.33± 0.02 8.54± 0.06 30.5%± 0.8 5.54± 0.03 9.25± 0.07 31.8%± 0.7 0 denotes the TCN-based methods, 1 denotes the GCN-based methods, and 2 denotes the attention-based methods.
accounting for the diferences in magnitude between Figure 5: The combinations of V-GCT with all subtraction pedestrian, vehicular, and stationary GCT flows to make types consistently yields the lowest prediction error for multiaccurate trafic predictions. step predictions.
Impact of w/o T. The model’s performance significantly decreased when Temporal View Modeling was removed, as evidenced by lower metrics across all prediction steps. V-GCT and (V-S)-GCT model performs better, indicating This result highlights the crucial role of Temporal View that the relative diferences between vehicle and pedesModeling in capturing temporal patterns and dependen- trian or stationary GCT flows may vary in importance cies, which contribute to enhanced prediction accuracy. with the prediction horizon.
The complete model consistently outperforms ablated Performance of model with V-GCT and all subtracversions, emphasizing the significance of Spatial, Tempo- tion types. The model incorporating V-GCT and all ral, and Multivariate View Modeling. Each component subtraction types consistently achieves the lowest MAE plays a critical role, and their integration leads to en- error over all prediction steps, demonstrating the efechanced V-GCT flow predictions. tiveness of improving prediction by exploring hidden relative diferences among multi-type GCT flows, and our model’s capability in modeling complex relations.
tivariate features, namely V-GCT, (V-S)-GCT, and (V-P)GCT. Thus, we conducted a parameter sensitivity analy- We proposed and analyzed a multi-type GCT approach sis for diferent combinations of V-GCT and subtraction that overcomes limitations in trafic prediction. Our pretypes to assess their impact on prediction performance. dictive model efectively combined multivariate spatioFigure 5 shows the MAE for three feature combination temporal modeling for V-GCT prediction across multiple models across 12 prediction steps: V-GCT with (V-S)- road segments, outperforming the baselines. Our experiGCT, V-GCT with (V-P)-GCT, and a combination of all ments highlighted the importance of model components types including V-GCT, (V-P)-GCT, and (V-S)-GCT. Key and GCT flow combinations for prediction accuracy. By observations are: initially validating the efectiveness of predicting V-GCT, Impact on Prediction Step: For prediction steps 1 to we can further explore other types of GCT predictive 4, the model with V-GCT and (V-P)-GCT outperforms results, ofering potential applications for improving inthe one with V-GCT and (V-S)-GCT. Beyond step 4, the telligent transportation.