隨著第五代（5G）和深度學習（DL）的快速發展，對數據傳輸容量的需求越來越大。基於大數據的分析和預測是智能地管理小區站點的趨勢。為了充分利用ML，強大的ML模型：3D CNN，RNN和CNN-RNN非常可靠。在我們之前的工作中，我們已經掌握了基於一種數據，互聯網流量預測移動互聯網流量的技術\cite{previous_paper}。但是，互聯網流量不僅受到一個因素的影響。這可能是由於外部條件造成的，這反過來影響了我們想要預測的目標。在這項工作中，我們使用2013年11月和12月期間來自米蘭市的互聯網流量，定期數據，天氣，新聞和社交數據來捕捉多數據與深度多模態學習模型之間的關係，稱為Multi-Modal CNN-RNN（MMCR）比僅有一種數據帶來更精確的預測。通過融合方法組合不同的數據，以便相關數據可以幫助我們預測目標。我們還使用學習方法通過深度學習模型調整當前時間的數據。實驗結果表明，使用有助於網絡流量的數據可以提高預測精度。我們還與其他工作設計的架構進行了比較。我們的方法也可以得到很好的結果。

With the fast development of the fifth-generation (5G) and deep learning (DL), the demand for data transmission capacity is getting more and more. Analyzing and forecasting based on big data is a tendency to manage cell sites intelligently. To have full use of ML, the powerful ML model: 3D CNN, RNN, and CNN-RNN are very reliable. We have already grasped the technique of predicting mobile internet traffic based on one kind of data, internet traffic, in our previous work \cite{previous_paper}. However, internet traffic is not only affected by one factor. It may be due to external conditions, which in turn affects the goals we want to predict. In this work, we use internet traffic, periodic data, weather, news and social data from the city of Milan during November and December in 2013 to catch the relationship between multi-data with a deep multi-modal learning model, called Multi-modal CNN-RNN (MMCR), bring on more precise forecasting than only one kind of data. Combine different data through a fused approach, so that relevant data can help us to predict the target. We also use the learning method to adjust the data at the current time through the deep learning model. The experimental results show that using data that is helpful for network traffic can improve prediction accuracy. And we also compare with the architecture designed by other work. Our method can also get good results.

[1] Pradeep K.Atrey,M.AnwarHossain,AbdulmotalebElSaddik,andMohanS.
Kankanhalli. Multimodalfusionformultimediaanalysis:asurvey. Multimedia Sys-
tems, 16(6):345–379,Nov2010.
[2] Gianni Barlacchi,MarcoDeNadai,RobertoLarcher,AntonioCasella,Cristiana
Chitic, GiovanniTorrisi,FabrizioAntonelli,AlessandroVespignani,AlexPentland,
and BrunoLepri.Amulti-sourcedatasetofurbanlifeinthecityofMilanandthe
ProvinceofTrentino. Scientific data, 2:150055,2015.
[3] Cisco VisualNetworkingIndexCisco.Globalmobiledatatrafficforecastupdate,
2016–2021. white paper, 2016.
[4] PauloCortez,MiguelRio,MiguelRocha,andPedroSousa.Multi-scaleInternet
trafficforecastingusingneuralnetworksandtimeseriesmethods. Expert Systems,
29(2):143–155, 2012.
[5] A. Damnjanovic,J.Montojo,YongbinWei,TingfangJi,TaoLuo,M.Vajapeyam,Tae-
sang Yoo,OsokSong,andD.Malladi.Asurveyon3GPPheterogeneousnetworks.
WirelessCommunications,IEEE, 18(3):10–21,June2011.
[6] DennyBritz.Recurrentneuralnetworkstutorial,part1–introductiontornns.
[7] JeffDonahue,LisaAnneHendricks,SergioGuadarrama,MarcusRohrbach,Sub-
hashini Venugopalan,KateSaenko,andTrevorDarrell.Long-termrecurrentconvo-
lutional networksforvisualrecognitionanddescription. CoRR, abs/1411.4389,2014.
[8] G.E. HintonandR.R.Salakhutdinov.Reducingthedimensionalityofdatawithneural
networks. Science (NewYork,N.Y.), 313:504–7,082006.
[9] C. Huang,C.Chiang,andQ.Li.Astudyofdeeplearningnetworksonmobiletraffic
forecasting. In 2017 IEEE28thAnnualInternationalSymposiumonPersonal,Indoor,
and MobileRadioCommunications(PIMRC), pages1–6,Oct2017.
[10] C. W.Huang,C.T.Chiang,andQ.Li.Astudyofdeeplearningnetworksonmobile
trafficforecasting.In 2017 IEEE28thAnnualInternationalSymposiumonPersonal,
Indoor,andMobileRadioCommunications(PIMRC), pages1–6,Oct2017.
[11] WenhaoHuang,GuojieSong,HaikunHong,andKunqingXie.Deeparchitecturefor
trafficflowprediction:Deepbeliefnetworkswithmultitasklearning. IEEE Transac-
tions onIntelligentTransportationSystems, 15(5):2191–2201,2014.
[12] WenweiJin,YoufangLin,ZhihaoWu,andHuaiyuWan.Spatio-temporalrecurrent
convolutionalnetworksforcitywideshort-termcrowdflowsprediction.In Proceed-
ings ofthe2NdInternationalConferenceonComputeandDataAnalysis, ICCDA
2018, pages28–35,NewYork,NY,USA,2018.ACM.
[13] Mingyu KimJuhanNamHonglakLeeAndrewY.NgJiquanNgiam,AdityaKhosla.
Multimodal deeplearning. Proc.28thInt.Conf.MachineLearning(ICML-11), 2011.
[14] Bahador Khaleghi,AlaaM.Khamis,FakhriKarray,andSaiedehN.Razavi.Multi-
sensor datafusion:Areviewofthestate-of-the-art. Information Fusion, 14:28–44,
2013.
[15] D. Lahat,T.Adali,andC.Jutten.Multimodaldatafusion:Anoverviewofmethods,
challenges, andprospects. ProceedingsoftheIEEE, 103(9):1449–1477,Sept2015.
[16] Y.Lecun,L.Bottou,Y.Bengio,andP.Haffner.Gradient-basedlearningappliedto
document recognition. ProceedingsoftheIEEE, 86(11):2278–2324,Nov1998.
[17] YannLeCun,BernhardEBoser,JohnSDenker,DonnieHenderson,RichardE
Howard,WayneEHubbard,andLawrenceDJackel.Handwrittendigitrecogni-
tion withaback-propagationnetwork.In Advances inneuralinformationprocessing
systems, pages396–404,1990.
[18] LISA lab.Convolutionalneuralnetworks(lenet).
[19] Jonathan Masci,MichaelM.Bronstein,AlexanderM.Bronstein,andJ¨urgenSchmid-
huber.Multimodalsimilarity-preservinghashing. CoRR, abs/1207.1522,2012.
[20] TiagoPradoOliveira,JamilSalemBarbar,andAlexsandroSantosSoares.Computer
networktrafficprediction:acomparisonbetweentraditionalanddeeplearningneural
networks. International JournalofBigDataIntelligence, 3(1):28–37,2016.
[21] SoujanyaPoria,ErikCambria,andAlexanderF.Gelbukh.Deepconvolutionalneural
networktextualfeaturesandmultiplekernellearningforutterance-levelmultimodal
sentiment analysis.In EMNLP, 2015.
[22] D. RamachandramandG.W.Taylor.Deepmultimodallearning:Asurveyonrecent
advancesandtrends. IEEE SignalProcessingMagazine, 34(6):96–108,Nov2017.
[23] Nitish SrivastavaandRuslanSalakhutdinov.Multimodallearningwithdeepboltz-
mann machines. JournalofMachineLearningResearch, 15:2949–2980,2014.
[24] D. Wang,P.Cui,M.Ou,andW.Zhu.Learningcompacthashcodesformultimodal
representations usingorthogonaldeepstructure. IEEE TransactionsonMultimedia,
17(9):1404–1416, Sep.2015.
[25] D. Wu,L.Pigou,P.Kindermans,N.D.Le,L.Shao,J.Dambre,andJ.Odobez.
Deep dynamicneuralnetworksformultimodalgesturesegmentationandrecognition.
IEEE TransactionsonPatternAnalysisandMachineIntelligence, 38(8):1583–1597,
Aug 2016.
[26] D. Wu,L.Pigou,P.J.Kindermans,N.D.H.Le,L.Shao,J.Dambre,andJ.M.
Odobez. Deepdynamicneuralnetworksformultimodalgesturesegmentationand
recognition. IEEE TransactionsonPatternAnalysisandMachineIntelligence,
38(8):1583–1597, Aug2016.
[27] Jing Xu,JiangWang,YuanpingZhu,YangYang,XiaojinZheng,ShuangdieWang,
LigangLiu,KariHorneman,andYongTeng.Cooperativedistributedoptimizationfor
the hyper-densesmallcelldeployment. IEEE CommunicationsMagazine, 52(5):61–
67, 2014.
[28] C. Zhang,H.Zhang,D.Yuan,andM.Zhang.Citywidecellulartrafficprediction
based ondenselyconnectedconvolutionalneuralnetworks. IEEE Communications
Letters, 22(8):1656–1659,Aug2018.
[29] Junbo Zhang,YuZheng,andDekangQi.Deepspatio-temporalresidualnetworksfor
citywide crowdflowsprediction.In ProceedingsoftheThirty-FirstAAAIConference
on ArtificialIntelligence(AAAI-17), pages1655–1661,2017.
38