![7
METHODOLOGY
Channel-Specific Graph Attention Layer (CGATL)
• For multi-channel representation H (size [C × N × D]), employ C independent GATs.
o Examine distinct correlations among N nodes of each channel c.
where 𝑖 ∈ {1, 2, . . . , 𝑍}, 𝑁𝑖 is the neighbors of node 𝑖, and 𝜎(∙) is a nonlinear
function.
Where H with dimension [C × N × D] and G is a graph structure indicating
connections among nodes in H.
o Adopt 1×1 convolution layers for Encoder and Decoder to compute efficiently:
Reduces the channel count from C to C′ (C′< C) in encoder and restore to C in decoder.](https://image.slidesharecdn.com/20240701labseminarhuyteltrans-240702022853-b137d3d6/85/20240701_LabSeminar_Huy-TelTrans-Applying-Multi-Type-Telecom-Data-to-Transportation-Evaluation-and-Prediction-via-Multifaceted-Graph-Modeling-pptx-7-320.jpg)
![8
METHODOLOGY
Multivariate Facet Modeling
• Insight: Modeling with subtracted type flows to emphasize
differences between V-GCT and P-GCT, hence revealing regional
attributes.
• Each time step t, concatenate 𝐻𝑣𝑔𝑐𝑡
𝑡
with ∆𝑝𝑔𝑐𝑡
𝑡
, ∆𝑠𝑔𝑐𝑡
𝑡
along third
dimension
o Then, reshape into [C × 3 × N] from [C × N × 3] and apply CGATL.
Where ‘Extract’ function retrieves the enhanced representation of 𝐻𝑣𝑔𝑐𝑡
𝑡
, the first element of 𝐻∆
𝑡
from CGATL’s output.
o Concatenate the output of all time step along second dimension [C × T × N]:
𝐻𝑣𝑔𝑐𝑡
𝑡
: multi-channel representations of VGCT flow and either P-GCT or S-GCT flow. f: either be P-GCT or S-GCT flow.
𝐺M: complete graph that reveals the connections among various GCT flow types.](https://image.slidesharecdn.com/20240701labseminarhuyteltrans-240702022853-b137d3d6/85/20240701_LabSeminar_Huy-TelTrans-Applying-Multi-Type-Telecom-Data-to-Transportation-Evaluation-and-Prediction-via-Multifaceted-Graph-Modeling-pptx-8-320.jpg)
![9
METHODOLOGY
Temporal Facet Modeling
• Insight: emphasize short-term and long-term from heightened fluctuations in V-GCT flow.
• Use two CNNs with kernel sizes (2 × 1) and (5 × 1) to extract short- and long-term temporal
patterns from multichannel representation H(H2 and H5).
• Apply CGATL separate and concatenate:
o Reshape H2 from [C × (T − 1) × S] to [C × (T − 4) × S] to align of H5.
o Implement the Gating Mechanism to manage ratio of information passed to the next module
Where ‘Concat’ function merges two outputs along the second dimension of
[C × T × N].
where 𝐻𝑇
1
and 𝐻𝑇
2
are generated from CGATL. 𝜇 denotes the tangent
hyperbolic function, and ⨀ represents the Hadamard product.](https://image.slidesharecdn.com/20240701labseminarhuyteltrans-240702022853-b137d3d6/85/20240701_LabSeminar_Huy-TelTrans-Applying-Multi-Type-Telecom-Data-to-Transportation-Evaluation-and-Prediction-via-Multifaceted-Graph-Modeling-pptx-9-320.jpg)
![10
METHODOLOGY
Spatial Facet Modeling
• Insight: GCT flow of an upstream segment influences the downstream GCT flow.
o Users’ mobility may cause them to move back and forth due to their activity behaviors.
o Accounting for bidirectional variations among road segments can lead to a comprehensive
understanding of user mobility.
• Assign specific directional graphs, 𝐺𝐹 and 𝐺𝐵, to each CGATL and then combine:
where H is multi-channel representation reshaped to [C × S ×T], regarded as
S road segment with T observations.
𝐺𝐹, 𝐺𝐵: Forward and backward graph structure, created by threshold of
Gaussian Kernel 𝐴/𝑟𝑜𝑤𝑠𝑢𝑚(𝐴) and 𝐴𝑇
/𝑟𝑜𝑤𝑠𝑢𝑚(𝐴𝑇
).](https://image.slidesharecdn.com/20240701labseminarhuyteltrans-240702022853-b137d3d6/85/20240701_LabSeminar_Huy-TelTrans-Applying-Multi-Type-Telecom-Data-to-Transportation-Evaluation-and-Prediction-via-Multifaceted-Graph-Modeling-pptx-10-320.jpg)
![11
EXPERIMENT AND RESULT
EXPERIMENT SETTINGs
• Dataset:
o Geographical Cellular Traffic in Hsinchu, Taiwan.
• Baselines:
o Deep Learning: Temporal Convolution (TCN).
o STGNN: Graph WaveNet(GWNet)[1], MTGNN [2], Gman[3], MPNet [4], DMGCN [5] and ESG[6].
[1] Wu, Z., Pan, S., Long, G., Jiang, J., & Zhang, C. (2019). Graph wavenet for deep spatial-temporal graph modeling. arXiv preprint arXiv:1906.00121.
[2] Wu, Z., Pan, S., Long, G., Jiang, J., Chang, X., & Zhang, C. (2020, August). Connecting the dots: Multivariate time series forecasting with graph neural networks. In Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining (pp. 753-763).
[3] Zheng, C., Fan, X., Wang, C., & Qi, J. (2020, April). Gman: A graph multi-attention network for traffic prediction. In Proceedings of the AAAI conference on artificial intelligence (Vol. 34, No. 01, pp. 1234-1241).
[4] Lin, C. Y., Su, H. T., Tung, S. L., & Hsu, W. H. (2021, October). Multivariate and propagation graph attention network for spatial-temporal prediction with outdoor cellular traffic. In Proceedings of the 30th ACM International Conference on Information & Knowledge Management (pp. 3248-
3252).
[5] Han, L., Du, B., Sun, L., Fu, Y., Lv, Y., & Xiong, H. (2021, August). Dynamic and multi-faceted spatio-temporal deep learning for traffic speed forecasting. In Proceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining (pp. 547-555).
[6] Ye, J., Liu, Z., Du, B., Sun, L., Li, W., Fu, Y., & Xiong, H. (2022, August). Learning the evolutionary and multi-scale graph structure for multivariate time series forecasting. In Proceedings of the 28th ACM SIGKDD conference on knowledge discovery and data mining (pp. 2296-2306).
• Measurement:
o Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Mean Absolute Percentage Error
(MAPE).](https://image.slidesharecdn.com/20240701labseminarhuyteltrans-240702022853-b137d3d6/85/20240701_LabSeminar_Huy-TelTrans-Applying-Multi-Type-Telecom-Data-to-Transportation-Evaluation-and-Prediction-via-Multifaceted-Graph-Modeling-pptx-11-320.jpg)
![[20240701_LabSeminar_Huy]TelTrans: Applying Multi-Type Telecom Data to Transportation Evaluation and Prediction via Multifaceted Graph Modeling.pptx](https://image.slidesharecdn.com/20240701labseminarhuyteltrans-240702022853-b137d3d6/85/20240701_LabSeminar_Huy-TelTrans-Applying-Multi-Type-Telecom-Data-to-Transportation-Evaluation-and-Prediction-via-Multifaceted-Graph-Modeling-pptx-15-320.jpg)
![[20240701_LabSeminar_Huy]TelTrans: Applying Multi-Type Telecom Data to Transportation Evaluation and Prediction via Multifaceted Graph Modeling.pptx](https://image.slidesharecdn.com/20240701labseminarhuyteltrans-240702022853-b137d3d6/85/20240701_LabSeminar_Huy-TelTrans-Applying-Multi-Type-Telecom-Data-to-Transportation-Evaluation-and-Prediction-via-Multifaceted-Graph-Modeling-pptx-16-320.jpg)
[20240701_LabSeminar_Huy]TelTrans: Applying Multi-Type Telecom Data to Transportation Evaluation and Prediction via Multifaceted Graph Modeling.pptx
- 1. Quang-Huy Tran Network Science Lab Dept. of Artificial Intelligence The Catholic University of Korea E-mail: huytran1126@gmail.com 2024-07-01 TelTrans: Applying Multi-Type Telecom Data to Transportation Evaluation and Prediction via Multifaceted Graph Modeling ChungYi Lin et al. AAAI-2024: Proceedings of the Thirty-Eighth Conference on Artificial Intelligence
- 2. 2 OUTLINE • MOTIVATION • METHODOLOGY • EXPERIMENT & RESULT • CONCLUSION
- 3. 3 MOTIVATION • Accurate traffic prediction can alleviate congestion and enhance traffic signal optimization in many application fields for intelligent transportation systems. o But they rely on dedicated sensors – cost and maintenance issue. Overview • Leveraging the extensive mobile network coverage. o Geographically inherent cellular traffic offers a unique insight into transportation dynamics.
- 4. 4 INTRODUCTION • Present the Geographical Cellular Traffic (GCT) Flows: o Define the GCT flow as the accumulation of GCTs over specific time intervals in road segments. o Categorize into three distinct types: Vehicle (V-GCT), Pedestrian(P-GCT), and Stationary (S-GCT). Contribution • Propose a novel model with three facets for predicting V-GCT: o Multivariate: discerning the interplay among multi-type GCT flows to uncover hidden regional functionality. o Temporal: separating short-term and long-term dynamics to avoid entangled dependencies. o Spatial: capturing bidirectional user mobility to understand the spatial dependencies.
- 5. 5 METHODOLOGY Problem Definition • Given a dataset: o 𝑁 road segments, D observations. o Feature matrix from historical 𝑇𝑖𝑛 steps of multi-type GCT flow. o 𝑓′ stands for V-GCT, S-GCT, or P-GCT. • Problem: forecast V-GCT in the upcoming 𝑇𝑜𝑢𝑡 steps
- 6. 6 METHODOLOGY Main Architecture • Three facets from CGAT: o Channel-Specific Graph Attention (CGAT). o Multivariate Facet: Capturing interactions among multitype GCT flows reveals implicit regional functionality. o Temporal Facet: Extracts short-term and long-term patterns to discern sudden and regular patterns separately. o Spatial Facet: Captures bidirectional spatial dependencies among road segments due to user mobility. • Overall: o Begins with a 32-channel CNN layer to encode all GCT flow types. Vehicle (V-GCT), Pedestrian(P-GCT), and Stationary (S-GCT). o Output Layer: use 64-channel CNN layer as skip connections at every Temporal Facet Modeling.
- 7. 7 METHODOLOGY Channel-Specific Graph Attention Layer (CGATL) • For multi-channel representation H (size [C × N × D]), employ C independent GATs. o Examine distinct correlations among N nodes of each channel c. where 𝑖 ∈ {1, 2, . . . , 𝑍}, 𝑁𝑖 is the neighbors of node 𝑖, and 𝜎(∙) is a nonlinear function. Where H with dimension [C × N × D] and G is a graph structure indicating connections among nodes in H. o Adopt 1×1 convolution layers for Encoder and Decoder to compute efficiently: Reduces the channel count from C to C′ (C′< C) in encoder and restore to C in decoder.
- 8. 8 METHODOLOGY Multivariate Facet Modeling • Insight: Modeling with subtracted type flows to emphasize differences between V-GCT and P-GCT, hence revealing regional attributes. • Each time step t, concatenate 𝐻𝑣𝑔𝑐𝑡 𝑡 with ∆𝑝𝑔𝑐𝑡 𝑡 , ∆𝑠𝑔𝑐𝑡 𝑡 along third dimension o Then, reshape into [C × 3 × N] from [C × N × 3] and apply CGATL. Where ‘Extract’ function retrieves the enhanced representation of 𝐻𝑣𝑔𝑐𝑡 𝑡 , the first element of 𝐻∆ 𝑡 from CGATL’s output. o Concatenate the output of all time step along second dimension [C × T × N]: 𝐻𝑣𝑔𝑐𝑡 𝑡 : multi-channel representations of VGCT flow and either P-GCT or S-GCT flow. f: either be P-GCT or S-GCT flow. 𝐺M: complete graph that reveals the connections among various GCT flow types.
- 9. 9 METHODOLOGY Temporal Facet Modeling • Insight: emphasize short-term and long-term from heightened fluctuations in V-GCT flow. • Use two CNNs with kernel sizes (2 × 1) and (5 × 1) to extract short- and long-term temporal patterns from multichannel representation H(H2 and H5). • Apply CGATL separate and concatenate: o Reshape H2 from [C × (T − 1) × S] to [C × (T − 4) × S] to align of H5. o Implement the Gating Mechanism to manage ratio of information passed to the next module Where ‘Concat’ function merges two outputs along the second dimension of [C × T × N]. where 𝐻𝑇 1 and 𝐻𝑇 2 are generated from CGATL. 𝜇 denotes the tangent hyperbolic function, and ⨀ represents the Hadamard product.
- 10. 10 METHODOLOGY Spatial Facet Modeling • Insight: GCT flow of an upstream segment influences the downstream GCT flow. o Users’ mobility may cause them to move back and forth due to their activity behaviors. o Accounting for bidirectional variations among road segments can lead to a comprehensive understanding of user mobility. • Assign specific directional graphs, 𝐺𝐹 and 𝐺𝐵, to each CGATL and then combine: where H is multi-channel representation reshaped to [C × S ×T], regarded as S road segment with T observations. 𝐺𝐹, 𝐺𝐵: Forward and backward graph structure, created by threshold of Gaussian Kernel 𝐴/𝑟𝑜𝑤𝑠𝑢𝑚(𝐴) and 𝐴𝑇 /𝑟𝑜𝑤𝑠𝑢𝑚(𝐴𝑇 ).
- 11. 11 EXPERIMENT AND RESULT EXPERIMENT SETTINGs • Dataset: o Geographical Cellular Traffic in Hsinchu, Taiwan. • Baselines: o Deep Learning: Temporal Convolution (TCN). o STGNN: Graph WaveNet(GWNet)[1], MTGNN [2], Gman[3], MPNet [4], DMGCN [5] and ESG[6]. [1] Wu, Z., Pan, S., Long, G., Jiang, J., & Zhang, C. (2019). Graph wavenet for deep spatial-temporal graph modeling. arXiv preprint arXiv:1906.00121. [2] Wu, Z., Pan, S., Long, G., Jiang, J., Chang, X., & Zhang, C. (2020, August). Connecting the dots: Multivariate time series forecasting with graph neural networks. In Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining (pp. 753-763). [3] Zheng, C., Fan, X., Wang, C., & Qi, J. (2020, April). Gman: A graph multi-attention network for traffic prediction. In Proceedings of the AAAI conference on artificial intelligence (Vol. 34, No. 01, pp. 1234-1241). [4] Lin, C. Y., Su, H. T., Tung, S. L., & Hsu, W. H. (2021, October). Multivariate and propagation graph attention network for spatial-temporal prediction with outdoor cellular traffic. In Proceedings of the 30th ACM International Conference on Information & Knowledge Management (pp. 3248- 3252). [5] Han, L., Du, B., Sun, L., Fu, Y., Lv, Y., & Xiong, H. (2021, August). Dynamic and multi-faceted spatio-temporal deep learning for traffic speed forecasting. In Proceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining (pp. 547-555). [6] Ye, J., Liu, Z., Du, B., Sun, L., Li, W., Fu, Y., & Xiong, H. (2022, August). Learning the evolutionary and multi-scale graph structure for multivariate time series forecasting. In Proceedings of the 28th ACM SIGKDD conference on knowledge discovery and data mining (pp. 2296-2306). • Measurement: o Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Mean Absolute Percentage Error (MAPE).
- 12. 12 EXPERIMENT AND RESULT RESULT – Overall Performance
- 13. 13 EXPERIMENT AND RESULT RESULT – Sensitivity Analysis of Multi-Type GCT Flows
- 14. 14 CONCLUSION • Presented multi-type GCT flows as a novel data source for transportation and proposed MFGM to predict V-GCT flows. o Integrating multi-type GCT flows. o Accuracy is improved. • Integrated V-GCT into transportation systems, presenting new applications for telecom data in transportation. Summarization