1. Inferring Air Quality for Station Location Recommendation Based on Urban Big Data
Author: Hsun-Ping Hsieh, Shou-De Lin, Yu Zheng
Date:2016/07/26 Source: KDD ’15
Speaker:Chen-Wei Hsu

2. Outline
Introduction
Method
Experiment
Conclusion

3. Introduction Motivation
how to infer realtime air quality of any arbitrary location given environmental data and data from very sparse monitoring locations
if one needs to establish few new monitor stations to improve the inference quality, how to determine the best locations for such purpose?

4. Introduction
Goal
Given a set of existing air monitoring stations, where to establish the next ones?
try to create an AQI inference mechanism that not only can infer the AQI values of any arbitrary unobserved location but also reveal the confidence of its inference
we propose to establish new stations at the locations that can minimize the uncertainty of the inference model

5. Introduction
Framework

6. Outline
Introduction
Method
Experiment
Conclusion

7. Method Data and Features Air Quality Records Meteorological Data
Point-Of-Interests (POIs)
Road Networks

8. Method INFERRING ARBITRARY AQI VALUES
Constructed to model the spatial-temporal correlation between grids
try to learn the weights of the edges, assuming they represent the correlations between nodes based on their features
emphasizes on inferring the AQI values for each location, which presumes those grids whose features are close to each other tend to share similar AQI values
The feature weights are adjusted to minimize the uncertainty of the model on inferring the unobserved locations

9. The construction of an AG consists of four parts:
Connecting to Station Location
Connecting to Near-by Locations
Connecting to Recent Layers
Connecting to Similar Layers
connect every unobserved node
to all the observed nodes in the same time stamp

10. Method affinity-based AQI inference (AQInf) model
a graph-based semi-supervised learning solution
The observed AQI on labeled nodes v <- V are utilized to infer the AQI distributions P(u) of unlabeled nodes u <-U
we assume that nodes with similar features should have similar AQI distributions. This relationship is modeled by edge weights through the combined affinity function
we propose to tune the parameters to minimize the model uncertainty

11. Method

12. Method Combined Affinity Function
KL : measure the difference between two AQI distributions

13. Method BUILDING MEASUREMENT STATIONS
Greedy-based Entropy Minimization(GEM)
a method called greedy-based entropy minimization (GEM) that aims at ranking locations based on their capability to reduce uncertainty

14. Method

15. Outline
Introduction
Method
Experiment
Conclusion

16. Experiments The Effectiveness of AQInf Beijing area : 50*50 grids
22 of the grids have the monitoring stations
period spans from 8/24/2012 to 10/31/2013, containing time stamps

17. Experiments To evaluate the usefulness of the proposed features
Geographical distances features plus three recent and three similar time layers as features (denoted by D+T3)
features in (1) plus the meteorology data (denoted by D+T3+M)
features in (2) plus Roadnet and POI features (denoted by ALL)

18. Experiments

19. Experiments
The Effectiveness of GEM

20. Experiments

21. Experiments

22. Experiments
Evaluating GEM with different time spans

23. Experiments

24. Outline
Introduction
Method
Experiment
Conclusion

25. Conclusion
This paper proposes a model to recommend the most proper locations in which building new air quality monitoring stations can lead to the largest accuracy improvement on air quality inference
In the future, we will focus on improving the efficiency of this model through parallelization
we will seek for more applications of our model, in particular in the area of traffic monitoring and surveillance in urban areas