AI+时空大数据赋能城市精细化决策_周骁

1 .AI+ 时空大数据赋能 城市精细化决策 中国人民大学高瓴人工智能学院 周骁 XIAO ZHOU | 2021.04.01 

2 .Multidisciplinary 跨学科研究 

3 .Urban Planning + Computer Science The World’s Cities -------- United Nations The proportion of the world’s population living in urban areas 城市计算 Urban Computing 1950 --------- 30% 2050 --------- 68% 2018 --------- 55% The number of megacities globally (≥10 million people) 1950 --------- 2 New York & Tokyo 2018 --------- 33 1975 --------- 4 + Shanghai & Mexico City https://esa.un.org/unpd/wup/Publications 

4 .Urban Planning + Computer Science 城市计算 Urban Computing 

5 .Ø Large-scale Ø Low-cost Ø Real-time Geo-social Network Data DIGITAL PHYSICAL WORLD WORLD In everyday life …… GSNs D T Ø Spatial Information Raw Data Dat Ø Temporal Information Ø Feedback Information 

6 . Thesis The exploration of geo-social network data with various analytical techniques can advance our understanding of the nature of urban phenomena & the development of intelligent urban computing applications. 

7 . Cultural Investment & Socio-Economic Deprivation 01 Title: Cultural investment and urban socio-economic development: A geosocial network approach. Royal Society Open Science 2017 Main Cultural Patterns Extraction & Urban Planning Contributions 02 Title: Discovering Latent Patterns of Urban Cultural Interactions in WeChat for Modern City Planning. KDD 2018 03 New Venues Recommendation Title: Topic-Enhanced Memory Networks for Personalised Point-of-Interest Recommendation. KDD 2019 

8 .Cultural Investment & Socio-Economic Deprivation 城市文化投资 & 城市地区社会经济落后情况 

9 . Royal Society Open Science Cultural To make Investment it better & Socio-Economic Deprivation n Less extreme form, there is no ‘best’ plan for every situation n Procedures need to be open to change. n Existing relations of power need to be both exposed and challenged. “Bilbao n Planners miracle” should think about their existing roles and be more active Guggenheim and creative Museum Frank Gehry 1997 

10 . Cultural Investment & Royal Society Open Science Socio-Economic Deprivation Datasets Foursquare data • Transition data • Venue profile Socio-Economic data • Index of Multiple Deprivation (IMD) • 2010 & 2015 Cultural expenditure data t duration |V| |E| <C> <k> • DCLG Local Authority 1 Jan 2011 - Dec 2011 15 832 469 229 0.221 59 Revenue Expenditure 2 Jan 2012 - Dec 2012 16 189 715 113 0.228 70 & Financing 3 Jan 2013 - Dec 2013 17 684 742 017 0.240 84 https://www.communities.gov.uk/archived/publications/localgovernment/regenerationthroughculture. 

11 . Royal Society Open Science Initial Status Cultural Expenditure & Network Feature Changes of London Boroughs Expenditure 2010-2012 IMD Score 2010 CEA 2010 CVA 2010 Venues Created 2011-2013 In-Degree 2011-2013 Out-Degree 2011-2013 Relationships between IMD, cultural expenditure & network features Group 1 Group 2 Group 3 Group 4 Initial Less More More Less IMD Deprived Deprived Deprived Deprived More Less More Less CEA Advantaged Advantaged Advantaged Advantaged Groups of London Wards in Analyses. Cultural Investment & Socio-Economic Deprivation 

12 . Royal Society Open Science Network features between groups in independent Cultural Expenditure &one-way ANOVAChanges of London Boroughs Network Feature Group 1 Group 2 Group 3 Group 4 Initial Less More More Less IMD Deprived Deprived Deprived Deprived More Less More Less CEA Advantaged Advantaged Advantaged Advantaged Network Properties between Groups Cultural Investment & Socio-Economic Deprivation 

13 . Royal Society Open Science Network features with statistically significant Cultural effects Expenditure in factorial & Network Featurerepeated Changesmeasures of London ANOVA Boroughs Group 1 Group 2 Group 3 Group 4 Initial Less More More Less IMD Deprived Deprived Deprived Deprived More Less More Less CEA Advantaged Advantaged Advantaged Advantaged Network Properties between Groups & Years Cultural Investment & Socio-Economic Deprivation 

14 . Royal Society Open Science Cultural Investment & Socio-Economic Deprivation Evaluation for Supervised Prediction Methods on Different Ward Sets ACCURACY PRECISION Supervised learning model AUC using network features together with cultural expenditure and geographical features to predict the IMD change for urban areas Ward Set Based on IMD Rank Change Criterion Evaluation for Supervised Prediction Methods on Different Feature Sets Network features make the largest contribution to ACCURACY PRECISION prediction models, as the AUC reduction of prediction effectiveness is biggest when they are removed. Classification Methods Prediction for Local Deprivation Changes 

15 .Cultural Patterns Extraction & Urban Planning 

16 . KDD’18 Cultural Patterns Extraction & Modern Urban Planning A Brief Overview 

17 . KDD’18 Cultural Patterns Extraction & Modern Urban Planning WeChat Moments Check-ins Beijing Geo-Social Network Data 97.4% Residents in Beijing are WeChat monthly active users A low-cost, real-time and fine-grained new digital data source that provides us 56.2 Million Check-ins unprecedented insights into spatio-temporal patterns of human mobility in cities. 9.5 Million Users 2.4 Million Venues 

18 . KDD’18 Cultural Patterns Extraction & Modern Urban Planning 1 2 4 3 Research Framework 

19 . KDD’18 Temporal Latent Dirichlet Allocation TLDA 𝜑# 𝛽 LDA TLDA 𝛼 𝜃! 𝑧!" 𝑣!" ¡ Topics Cultural Patterns ¡ 𝛾 𝜙" ¡ Document ¡ User Cultural Check-ins ¡ Words ¡ Cultural POI Categories & Time Slots Innovation of TLDA Topic modeling approach to patterns extraction Cultural Patterns Extraction Improves LDA by integrating temporal factors & Modern Urban Planning 

20 . KDD’18 Cultural Patterns Extraction & Modern Urban Planning TLDA ‘Golden Week’ of Spring Festival National Day New Year’s Day Performance of TLDA Compared with Standard LDA 

21 . KDD’18 Innovation of TCV A novel evaluation method for TLDA model to measure word-time coherence value (CV) within each topic. Patten z Typical venue set 𝑉 ∗ = 𝑍𝑜𝑜, 𝐴𝑞𝑢𝑎𝑟𝑖𝑢𝑚 Typical time set 𝑇 ∗ = 𝑆𝑢𝑚𝑆𝑎𝑡11, 𝑆𝑢𝑚𝑆𝑢𝑛14, 𝑆𝑝𝑟𝑆𝑢𝑛14 Equations TLDA Temporal Coherence Value (TCV) Cultural Patterns Extraction & Modern Urban Planning 

22 . KDD’18 Cultural Patterns Extraction & Modern Urban Planning TLDA Performance of TLDA Compared with Standard LDA 

23 . KDD’18 Cultural Patterns Extraction & Modern Urban Planning All within-group similarities are higher than 0.9. Most of the inter-group similarities are lower than 0.1. 03 04 05 01 02 06 TLDA Probabilities of Venues Categories over Patterns 

24 . KDD’18 Cultural Patterns Extraction & Modern Urban Planning Pattern 1 2 3 4 5 6 Key word Scenic spot & Park Music Plant & Animal Museum Sports Gym TLDA Temporal Distribution over Patterns 

25 . Personalised Ordering Points to Identify the Clustering Structure POPTICS DEMAND ¡ Users 𝝁 ¡ 𝑫 𝒙 =∑ 𝟏 𝒆𝒙𝒑 − 𝒙'𝝁 𝟐 DSR 𝟐𝝅𝒓 𝟐 𝟐𝒓𝟐 ¡ Demand-Supply Ratio 𝑫(𝒙) ¡ 𝑫𝑺𝑹 𝒙 = 𝑺(𝒙) SUPPLY ¡ Venues 𝛎 𝟏 𝒙'𝝂 𝟐 ¡ 𝑺 𝒙 =∑ 𝒆𝒙𝒑 − 𝟐𝝅𝝈𝟐 𝟐𝝈𝟐 KDD’18 Cultural Patterns Extraction Demand-Supply Interaction (DSI) Model & Modern Urban Planning 

26 . KDD’18 Cultural Patterns Extraction & Modern Urban Planning TLDA Results of the DSI Model 

27 .New Venues Recommendation 城市新场所推荐系统 KDD 2019 

28 . Topic-Enhanced Memory Networks for Personalised Point-of-Interest Recommendation. KDD’19 New Urban Venues Recommendation TLDA A Video Introducing the TEMN 

29 . KDD’19 New Urban Venues Recommendation ;<%&= 5& ?!@A 0+'B1C <+=/B1E F+BG1H%IJ6H7 'BK+7J<G Check-ins z %2& !$ 2 -. & 4 32 , / 0 6& u 1&) … 0 1 … 0 %' & !" ! 8& +&) 6) v >&=%&= 0 0 … 1 () 8) 7&) :9&) 6* j !# 9 1 0 … 0 3 (* 8* 7&* :9&* TLDA Architecture of Topic-Enhanced Memory Network (TEMN) for Personalised Point-of-Interest Recommendation