Multivariate and Propagation Graph Attention Network for ...
Transcript of Multivariate and Propagation Graph Attention Network for ...
![Page 1: Multivariate and Propagation Graph Attention Network for ...](https://reader030.fdocuments.in/reader030/viewer/2022012022/6169c27911a7b741a34b0d3d/html5/thumbnails/1.jpg)
Multivariate and Propagation Graph Attention Network forSpatial-Temporal Prediction with Outdoor Cellular Traffic
Chung-Yi Lin1,2, Hung-Ting Su1, Shen-Lung Tung2, Winston H. Hsu11National Taiwan University, 2Chunghwa Telecom Laboratories
ABSTRACTSpatial-temporal prediction is a critical problem for intelligent trans-portation, which is helpful for tasks such as traffic control and ac-cident prevention. Previous studies rely on large-scale traffic datacollected from sensors. However, it is unlikely to deploy sensorsin all regions due to the device and maintenance costs. This paperaddresses the problem via outdoor cellular traffic distilled fromover two billion records per day in a telecom company, becauseoutdoor cellular traffic induced by user mobility is highly relatedto transportation traffic. We study road intersections in urban andaim to predict future outdoor cellular traffic of all intersectionsgiven historic outdoor cellular traffic. Furthermore, We proposea new model for multivariate spatial-temporal prediction, mainlyconsisting of two extending graph attention networks (GAT). FirstGAT is used to explore correlations among multivariate cellulartraffic. Another GAT leverages the attention mechanism into graphpropagation to increase the efficiency of capturing spatial depen-dency. Experiments show that the proposed model significantlyoutperforms the state-of-the-art methods on our dataset.
CCS CONCEPTSβ’ Information systemsβ Data mining.
KEYWORDSmultivariate spatial-temporal prediction, graph attention network,outdoor cellular trafficACM Reference Format:Chung-Yi Lin1,2, Hung-Ting Su1, Shen-Lung Tung2, Winston H. Hsu1.2021. Multivariate and Propagation Graph Attention Network for Spatial-Temporal Prediction with Outdoor Cellular Traffic. In Proceedings of the 30thACM International Conference on Information and Knowledge Management(CIKM β21), November 1β5, 2021, Virtual Event, QLD, Australia. ACM, NewYork, NY, USA, 5 pages. https://doi.org/10.1145/3459637.3482152
1 INTRODUCTIONRecently, spatial-temporal prediction becomes one of the funda-mental techniques in building intelligent transportation. Based onlarge-scale data (e.g., traffic speed, volume) collected from sensors,previous studies [7, 16] achieve successful performance. However,it is unlikely to deploy sensors in all regions (e.g., road intersections,
Permission to make digital or hard copies of all or part of this work for personal orclassroom use is granted without fee provided that copies are not made or distributedfor profit or commercial advantage and that copies bear this notice and the full citationon the first page. Copyrights for components of this work owned by others than theauthor(s) must be honored. Abstracting with credit is permitted. To copy otherwise, orrepublish, to post on servers or to redistribute to lists, requires prior specific permissionand/or a fee. Request permissions from [email protected] β21, November 1β5, 2021, Virtual Event, QLD, AustraliaΒ© 2021 Copyright held by the owner/author(s). Publication rights licensed to ACM.ACM ISBN 978-1-4503-8446-9/21/11. . . $15.00https://doi.org/10.1145/3459637.3482152
rural areas) due to the device and maintenance costs, which givesrise to the task of data collection for spatial-temporal prediction.
We seek an alternative approach to evaluate the traffic statewithout sensors. With billions of mobile devices entering the in-ternet, massive records of cellular traffic [1] are collected at celltowers. Many studies have been made for its application, such ascellular vehicle probes for traffic state estimation [3, 9], and cellulartraffic prediction modeling [1, 11, 12]. Nevertheless, existing stud-ies rarely consider transportation traffic induced by user mobility.According to the analyses from over billions of records, outdoorcellular traffic is found to be induced by user mobility. Therefore,we leverage outdoor cellular traffic representing the traffic state forspatial-temporal forecasting.
In this paper, we propose a new spatial-temporal dataset1 viaoutdoor cellular traffic distilled from over a billion records per dayin a telecom company, which is extremely valuable for surveillingtraffic states in regions without sensors. For instance, accumulatedoutdoor cellular traffic in a unit time step exceeds the threshold,the region might occur certain events (e.g., traffic congestion orparade). We study the road intersections of a major city forming aroad network and collect corresponding outdoor cellular traffic intime steps. This paper aims to predict the future outdoor cellulartraffic of all intersections given historic outdoor cellular traffic as anew spatial-temporal task.
Spatial-temporal prediction with a road network has been widelystudied. Recent researches [7, 8, 13, 14] have achieved great suc-cess in prediction by conducting propagating information on thegraph data within graph neural networks (GNN). However, thedataset of previous studies is the traffic speed detected from sen-sors, which is usually less varied between time steps. While in ourdataset, the quantity of outdoor cellular traffic can drastically vari-ous in different times, especially the peak (e.g., 18:00) is 200 timesgreater than the least active times (e.g., 03:00) in the same road in-tersection, such significant difference is more challenging than thatprior dataset. Therefore, We argue that expanding the uni-historicdata into multivariate consisting of various temporal-periodic datais critical to capturing more hidden correlations in the complextemporal pattern, which has not been explored in previous meth-ods. As more complicated temporal features fed into the predictivemodel, we notice that the propagation process of GNN does notconsider the dynamic attention between nodes during updatingnode information, which might reduce the capacity of modeling.
To address the challenges, we propose a new framework forspatial-temporal prediction, namely MPGAT (Multivariate andPropagation Graph Attention Network), mainly consisting of twoextending graph attention networks (GAT) [10] modules. (1) Multi-variate GAT (M-GAT) explores the correlations amongmultivariateinput, which can be effectively adapted to multivariate time series.
1the dataset is accessible on Github: https://github.com/cylin-cmlab/CIKM21-MPGAT
arX
iv:2
108.
0830
7v2
[cs
.LG
] 2
0 A
ug 2
021
![Page 2: Multivariate and Propagation Graph Attention Network for ...](https://reader030.fdocuments.in/reader030/viewer/2022012022/6169c27911a7b741a34b0d3d/html5/thumbnails/2.jpg)
(2) Propagation GAT (P-GAT) incorporates the propagation strat-egy into GAT to captures the spatial dependency between regions,benefited from the attention mechanism and spatial closeness ofthe graph. We evaluate the proposed framework on our dataset.Experimental results with statistic analysis show that MPGAT out-performs significantly several state-of-the-art models.
2 DATASET DESCRIPTIONDataCollection: The data was collected from a large-scale cellular-geographic system in a telecom company, which receives over twobillion records from mobile devices in one day. Each record con-tains International Mobile Station Equipment Identification (IMEI,a unique number of a mobile phone), International Mobile Sub-scriber Identity (IMSI, a unique number of a subscriber), the cre-ation time, the GPS location, and the location type (categorized bythe telecom company, e.g., outdoor and indoor). As privacy consid-erations, we use IMEI to represent each mobile subscriber due toIMSI being able to identify the mobile subscriber particularly, andIMEI is anonymized by hashing.Spatial-temporal Dataset: To explicitly discover the spatial de-pendency, we study six road intersections (e.g., around the trainstation and college) as a road network, as shown in Figure 1. Then,we individually aggregated the quantity of outdoor IMEI located atintersections in the unit time step (i.e., 5-minutes) as age spatial-temporal dataset. Specifically, the aggregated IMEI quantity overtime steps demonstrates strong temporal correlations, and geo-graphically connected intersections contain spatial dependence.
Figure 1: The dataset of outdoor cellular traffic is collectedaround the six road intersections(left fig.), and the IMEIquantities for every time step are shown in the right fig.
3 PRELIMINARIESIntersection Network: Each road intersection is geographicallyconnected to adjacent neighbors, creating an intersection network.Thus, we define the network as a directed graph πΊπππ‘πππ πππ‘πππ =(π ,πΈ), whereπ is a set of π intersections of the network, πΈ is the setof edges representing the connectivity between the intersections.Multivariate Input: According to our dataset, IMEI quantity hasa drastic change in different time steps, therefore we expand IMEIquantity of unit time step as different temporal-periodic series to
capture more temporal patterns, forming a multivariate input:(1)The IMEI quantity adjacent to predictions:
ππ = {π₯π,π‘0βπππ+1, ..., π₯π,π‘0 }βπ πΓπππ denotes historical IMEI quan-tity with πππ time steps adjacent to the predictions, revealing theshort-term factor, where π₯π,π‘0 is the IMEI quantity of π intersec-tions at time step π‘0.(2)The moving average of IMEI quantity adjacent to predictions:
πππππ= {π₯ππππ ,π‘0βπππ+1, ..., π₯ππππ ,π‘0 }βπ πΓπππ is a set of mov-
ing average (MA) of IMEI quantity over πππ time steps, where eachvalue of πππππ
is calculated by accumulating the IMEI quantityover ππ time steps and then dividing the sum by ππ . Moving aver-age is commonly used with time-series data to smooth out short-term fluctuations and highlight longer-term trends [15]. The fea-tures of πππ5 and πππ20 are adopted in this paper.(3)The daily intervals adjacent to prediction: [2]
Suppose the sampling frequency is p time steps per day, ππ =
{π₯π,π‘0βπππΓπ , π₯π,π‘0βπππβ1Γπ , ..., π₯π,π‘0 }βπ πΓπππ denotes a set of daily-interval data, consisting of IMEI quantity at π‘0 time step in differentdays closed to predictions. For example, there are π(=288) time stepsin one day by 5-minute time step. We consider the daily-intervalquantity as one of the features to capture the sequence regularpattern.Problem: Given π = {π1, ..., ππΉ }βπ πΉΓπΓπππ to denote the mul-tivariate input of all the intersections over observed time steps,predict future IMEI quantity set π = {π 1, ..., ππ }βπ πΓπππ’π‘ overthe coming time steps, where πΉ is the number of features in mul-tivariate input, π is the number of intersections in πΊπππ‘πππ πππ‘πππ ,πππ and πππ’π‘ is the length of time steps, and ππ is the future IMEIquantity of intersection π .
4 METHODOLOGYFigure 2 illustrates the framework of our proposed Multivariate andPropagation Graph Attention Network (MPGAT). The multivari-ate input π , such as π = {ππ, πππ5 , πππ20 , ππ }, is fed to M-GATto capture correlations between multivariate input π . Then, thedistilled outputs from M-GAT are forwarded to the connected TCNlayer. For capturing the spatial-temporal dependency, TCN is in-terleaved with P-GAT as a spatial-temporal block, where P-GAT isto model the spatial dependency between intersections. By stack-ing multiple spatial-temporal blocks, the capacity of modeling thespatial-temporal dependencies increases [14]. To avoid gradientvanishing, Residual Layers [4] is appended from the top of each in-terleaved layer to its end, and the Skipped Layers are concatenatedafter each TCN to an output layer.
4.1 Multivariate Graph Attention (M-GAT)As we aforementioned, the IMEI quantity series has been expandedinto multivariate input. Due to the strong feature-extraction ca-pability of GAT, M-GAT contains multiple GAT layers to capturethe associations between multivariate input. Inspired by [5], M-GAT treats each component of multivariate input as one node in acomplete graph and contains two stacked GAT layers.
We consider a single layer of M-GAT as an example. For intersec-tion π , the set of multivariate inputππ = {π π
1 , ..., π ππΉ
} β π πΉΓπππ isfirst projected as a set of latent representationsππ = {π» π
1 , ..., π» ππΉ
}
![Page 3: Multivariate and Propagation Graph Attention Network for ...](https://reader030.fdocuments.in/reader030/viewer/2022012022/6169c27911a7b741a34b0d3d/html5/thumbnails/3.jpg)
Figure 2: The framework ofMPGAT, whereM-GAT explorescorrelations amongmultivariate input, and P-GAT capturesspatial dependency with multiple directions separately.
β π πΉΓπππΓπ·β²by a convolution layer before M-GAT, whereπππ is the
length of historical time steps, and π·β²is the dimension of latent
representation. Afterward, ππ is fed to M-GAT, forming a completegraph with πΉ nodes. The attention score between nodes representsthe importance from node π to node π , where π , π β {1, 2, . . . , πΉ }, andcan be computed as follows:
π ππ π =
ππ₯π (πΏππππ¦π ππΏπ (π€ππ (π» π
πβ₯ π» π
π)))βπΉ
π=1 ππ₯π (πΏππππ¦π ππΏπ (π€ππ (π» π
πβ₯ π» π
π)))
(1)
, where β₯ is the concatenation operation, π€ππ βπ 2π·
β²is the weight
vector with transposition, and the attention score is normalized bya SoftMax function with LeakyReLU.
With the normalized attention scores, the output of one M-GATlayer for node π is given by:
οΏ½ΜοΏ½ ππ = π (
πΉβοΈπ=1
(π ππ π π»
ππ )) (2)
, where π denotes a non-linear activation function. οΏ½ΜοΏ½ ππ
is the ag-gregated latent representation for feature π of intersection π , whichcontains the implicit influence from others. For π intersections,the output set is presented as οΏ½ΜοΏ½ = {οΏ½ΜοΏ½1, ..., οΏ½ΜοΏ½πΉ } β π πΉΓπΓπππΓπ·
β². To
reduce the model complexity, we only distill the latent representa-tions of IMEI quantity outputted at the last layer of M-GAT, denotesas οΏ½ΜοΏ½π β π πΓπππΓπ·
β², and feeding οΏ½ΜοΏ½π into the next layer of MPGAT.
4.2 Propagation Graph Attention (P-GAT)P-GAT consists of two directions Attention Propagation and oneGlobal learning to capture the spatial dependency between inter-sections. To our best knowledge, we are the first work to apply theattention mechanism on graph propagation.Attention Propagation: The advantage of the propagation pro-cess [13] is that it aggregated node information through the graph
Figure 3: Multivariate input ππ of each intersection π is pro-jected into latent representations π»π , and M-GAT capturescorrelations among multivariate with graph attention.
structure recursively and preserves a proportion of nodesβ statesduring the process, eliminating the smoothing problem with dif-fusion method in [7]. With the benefit of propagating the statesthrough graph structure, we generalize the attention mechanisminto the propagation process, as shown in Figure 4. Consideringπ nodes with the set of initial representations πππ = {π1, . . . ,ππ }from the previous layer, the propagation and output is defined as:
π ` = (1 β π½)πππ + π½ (π΄πππ π π`β1) (3)
π = Ξ(β₯π`=0π` ) (4)
, where π 0 = πππ , π΄πππ π is the masked attention matrix for uni-direction, ` represents the propagation steps, π½ is a hyperparameterbetween 0 and 1 to control the ratio of nodeβs propagation, Ξ is amultilayer perceptron layer (MLP) to transform the channel dimen-sion of concatenation results, and π is the final output of π nodes.In practice, we employ the CNN as MLP, and setπ is 2.
In case of the directed graph πΊπΌππ‘πππ πππ‘πππ , P-GAT models twodirections, the forward and the backward propagations with ownmasked attentionmatrix. Consider each intersection ofπΊπΌππ‘πππ πππ‘πππ
as one node, the attention score ππ π in uni-directionπ΄πππ π betweennodes π and π , where π , π β {1, 2, . . . , π }, is computed as:
ππ π = π€ππ (ππ β₯ ππ ) (5)
ππ π =ππ₯π (πΏππππ¦π ππΏπ (ππ π ))βππ
π=1 ππ₯π (πΏππππ¦π ππΏπ (ππ π ))(6)
, where ππ is the latent representation of node π from the previouslayer,π€π
π β π 2π·β²β²is a weight vector with transposition, π·
β²β²is the
dimension of ππ , and ππ denotes the adjacent neighbors of node π .The attention coefficient ππ π in Equation 5 is masked according tothe adjacency matrices in the corresponding direction, where ππ πis masked with a large negative value (e.g., -9e15) if node π and π
are not adjacent; otherwise, the coefficient ππ π would be preserved.In this way, the attention score ππ π between adjacent nodes wouldincrease after SoftMax normalization, while the non-adjacent isset to zero. The matrix π΄πππ π can be represented as [ππ π ], whichimplies the relationship between adjacent nodes but excludes thenon-adjacent.Global Learning: To discover hidden correlations among nodes,we employ another attention module to enhance the learning ca-pacity. Similar to M-GAT, Global Learning treats each intersection
![Page 4: Multivariate and Propagation Graph Attention Network for ...](https://reader030.fdocuments.in/reader030/viewer/2022012022/6169c27911a7b741a34b0d3d/html5/thumbnails/4.jpg)
Table 1: The average performance of MAPE comparison on outdoor cellular traffic dataset
Models Mean Β±std. Dev. h1 Mean Β±std. Dev. h Mean Β±std. Dev. h Mean Β±std. Dev. h
5min 15min 30min 60min
GWNET(β19)[14] 0.1576Β±0.0027 -1/1 0.1781Β±0.0034 0/1 0.1934Β±0.0028 0/1 0.2194Β±0.0037 0/1MTGNN(β20)[13] 0.1629Β±0.0032 1/1 0.1811Β±0.0027 1/1 0.1951Β±0.0031 0/1 0.2198Β±0.0033 0/1STAWNET(β21)[8] 0.1611Β±0.0026 1/1 0.1823Β±0.0037 1/1 0.1947Β±0.0029 0/1 0.2180Β±0.0026 -1/1MPGAT-1 0.1582Β±0.0024 0.1778Β±0.0019 0.1940Β±0.0025 0.2213Β±0.0029MPGAT 0.1511Β±0.0013 0.1720Β±0.0012 0.1876Β±0.0016 0.2149Β±0.0026
h1 means whether the result of MPGAT-1/MPGAT is significant according to Wilcoxon rank-sum test compared to the baseline method.
Figure 4: The function of attention propagation in P-GAT.
as a node and builds a complete graph to capture the hidden rela-tionships of spatial dependency. We use Equation 6 to construct anattention matrix, where ππ π is not masked, ππ denotes all the nodesofπΊπΌππ‘πππ πππ‘πππ . All nodes would learn globally the attention scoretoward others and updates the node information directly throughthe global attention.Output: P-GAT explores the bi-directional correlation with direc-tional modeling and captures the implicit associations adaptivelywith global modeling for spatial dependency learning. Finally, P-GAT fuses two propagation modeling and global modeling outputsas an input fed to the next layer.
5 EXPERIMENTSDataset: This paper verifies the proposed MPGAT and the com-pared models with the outdoor cellular traffic dataset, which con-tains the IMEI quantity in 5-minutes time steps of six road intersec-tions ranging from Jan.1, 2020, to Jun.30, 2020. The dataset is splitwith 70% for training, 10% for validation, and 20% for testing.Comparison Methods: Our dataset is a spatial-temporal datasetwith a road network, similar tasks such as traffic speed have gener-ally been addressed better by graph-based models [16]. Due to thepage limitation, we focus on the latest graph-based models as base-lines in our evaluation: GWENT [14], MTGNN [13], STAWNET [8].To verify the effectiveness of the correlation among multivariate,we build two models, MPGAT-1 and MPGAT, where MPGAT-1 isremoving M-GAT and use univariate IMEI quantity as input.Experimental Settings: Our experiments are conducted with aNVIDIA DGX-1. In practice, the length of historical input πππ is12 for all models including ours and baselines. MPGAT uses eightspatial-temporal blocks to cover the input sequence, where eachblock has a TCN interleaved with a P-GAT. We adapt the Adamoptimizer with a learning rate of 0.001 to train our model. The eval-uation metrics we choose mean absolute percentage error (MAPE).
Main Results: We implement our model and compare it with thebaselines for 5 min,15 min, 30 min, and 60 min predictions on ourdataset. We run 30 testing times and report the average values, asshown in Table 1. To determine whether MPGAT-1 and the MPGAToutperform in statistical significance with the compared models,we conducted the Wilcoxon rank-sum test inspired by [6], where h= 1 is the proposed model significantly outperformed the comparedmodel, h = -1 is the compared model significantly outperformedthe proposed model, and h = 0 is not significantly different. Forexample, h = -1/1 means that MPGAT-1 is worse, but MPGAT isbetter than the compared model.
According to Table 1, as the prediction time increases, MAPEof the same method rises for all, showing that the longer the pre-diction time, the more challenging the prediction task. Second,with the statistical analysis, MPGAT significantly achieves the bestperformance in the dataset. Third, MPGAT-1 is slightly underper-formed than GWNET while outperforming others in the short-termprediction. It is more beneficial to practical applications on short-term factors, e.g., transportation agencies can immediately optimizetraffic congestion.
Figure 5 shows the changes in the prediction performance ofvarious methods as the prediction steps increase. We observe thatMPGAT consistently outperforms MPGAT-1 and baseline models,indicating it is critical to explore the correlation amongmultivariate,especially that our dataset has a drastic change between neighbor-ing time steps. Moreover, we notice that the MAPE value of thesame method conducted on our dataset is two times larger thanthe traffic speed dataset [7], which indicates that spatial-temporalprediction with cellular traffic data is more challenging. We are op-timistic that our proposed task and dataset would pave a new pathfor spatial-temporal prediction and urban computing applications.
6 CONCLUSIONIn this paper, we propose a new spatial-temporal dataset via out-door cellular traffic and a model MPGAT for multivariate spatial-temporal prediction. Experimental results show that the proposedMPGAT significantly outperforms other models on the dataset. Inthe future, we intend to expand the dataset more robust to cover theentering city and use transfer learning to employ on other cities.
ACKNOWLEDGMENTSThis work was supported in part by the Ministry of Science andTechnology, Taiwan, under Grant MOST 110-2634-F-002-026. We
![Page 5: Multivariate and Propagation Graph Attention Network for ...](https://reader030.fdocuments.in/reader030/viewer/2022012022/6169c27911a7b741a34b0d3d/html5/thumbnails/5.jpg)
Figure 5:MAPEof comparedmethods in different predictiontimes shows MPGAT achieves better results in all cases.
benefit from NVIDIA DGX-1 AI Supercomputer and are grateful tothe National Center for High-performance Computing.
REFERENCES[1] Luoyang Fang, Xiang Cheng, Haonan Wang, and Liuqing Yang. 2018. Mobile
demand forecasting via deep graph-sequence spatiotemporal modeling in cellularnetworks. IEEE Internet of Things Journal 5, 4 (2018), 3091β3101.
[2] Shengnan Guo, Youfang Lin, Ning Feng, Chao Song, and Huaiyu Wan. 2019.Attention based spatial-temporal graph convolutional networks for traffic flowforecasting. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 33.922β929.
[3] Ayalew Belay Habtie, Ajith Abraham, and Dida Midekso. 2017. Artificial neu-ral network based real-time urban road traffic state estimation framework. InComputational Intelligence in Wireless Sensor Networks. Springer, 73β97.
[4] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. Deep residuallearning for image recognition. In Proceedings of the IEEE conference on computervision and pattern recognition. 770β778.
[5] Yingfan Huang, Huikun Bi, Zhaoxin Li, Tianlu Mao, and Zhaoqi Wang. 2019.Stgat: Modeling spatial-temporal interactions for human trajectory prediction.In Proceedings of the IEEE/CVF International Conference on Computer Vision. 6272β6281.
[6] Yau-Tarng Juang, Shen-Lung Tung, and Hung-Chih Chiu. 2011. Adaptive fuzzyparticle swarm optimization for global optimization of multimodal functions.Information Sciences 181, 20 (2011), 4539β4549.
[7] Yaguang Li, Rose Yu, Cyrus Shahabi, and Yan Liu. 2017. Diffusion convolu-tional recurrent neural network: Data-driven traffic forecasting. arXiv preprintarXiv:1707.01926 (2017).
[8] Chenyu Tian and Wai Kin Chan. 2021. Spatial-temporal attention wavenet:A deep learning framework for traffic prediction considering spatial-temporaldependencies. IET Intelligent Transport Systems 15, 4 (2021), 549β561.
[9] Amir Hosein Valadkhani, Yang Hong, and Mohsen Ramezani. 2017. Integration ofloop and probe data for traffic state estimation on freeway and signalized arteriallinks. In 2017 IEEE 20th International Conference on Intelligent TransportationSystems (ITSC). IEEE, 1β6.
[10] Petar VeliΔkoviΔ, Guillem Cucurull, Arantxa Casanova, Adriana Romero, PietroLio, and Yoshua Bengio. 2017. Graph attention networks. arXiv preprintarXiv:1710.10903 (2017).
[11] Jing Wang, Jian Tang, Zhiyuan Xu, Yanzhi Wang, Guoliang Xue, Xing Zhang, andDejun Yang. 2017. Spatiotemporal modeling and prediction in cellular networks: Abig data enabled deep learning approach. In IEEE INFOCOM 2017-IEEE Conferenceon Computer Communications. IEEE, 1β9.
[12] Xu Wang, Zimu Zhou, Fu Xiao, Kai Xing, Zheng Yang, Yunhao Liu, and ChunyiPeng. 2018. Spatio-temporal analysis and prediction of cellular traffic in metrop-olis. IEEE Transactions on Mobile Computing 18, 9 (2018), 2190β2202.
[13] ZonghanWu, Shirui Pan, Guodong Long, Jing Jiang, Xiaojun Chang, and ChengqiZhang. 2020. Connecting the dots: Multivariate time series forecasting with graphneural networks. In Proceedings of the 26th ACM SIGKDD International Conferenceon Knowledge Discovery & Data Mining. 753β763.
[14] Zonghan Wu, Shirui Pan, Guodong Long, Jing Jiang, and Chengqi Zhang. 2019.Graph WaveNet for Deep Spatial-Temporal Graph Modeling.. In IJCAI.
[15] Hao Yu, Xiaofeng Chen, Zhenning Li, Guohui Zhang, Pan Liu, Jinfu Yang, andYin Yang. 2019. Taxi-based mobility demand formulation and prediction usingconditional generative adversarial network-driven learning approaches. IEEETransactions on Intelligent Transportation Systems 20, 10 (2019), 3888β3899.
[16] Chuanpan Zheng, Xiaoliang Fan, Cheng Wang, and Jianzhong Qi. 2020. Gman: Agraph multi-attention network for traffic prediction. In Proceedings of the AAAIConference on Artificial Intelligence, Vol. 34. 1234β1241.