U-Air: When Urban Air Quality Inference Meets Big Data - Intro

Yung-Sheng Lu

Oct 14, 2016

@NCKU CSIE

Yu Zheng, Furui Liu, Hsun-Ping Hsieh @ Microsoft Research Asia, Beijing China

Outline

• Abstract

• Introduction

• Overview

• TODO

U-Air: When Urban Air Quality Inference Meets Big Data

Abstract

U-Air: When Urban Air Quality Inference Meets Big Data

Abstract

• real-time and fine-grained air quality information

• semi-supervised learning approach based on a co-training framework

  • ​spatial classifier

  • temporal classifier

• ​​Exp: Beijing and Shanghai

Introduction

U-Air: When Urban Air Quality Inference Meets Big Data

Motivation

• expensive cost of building and maintaining

• ​urban air quality varies by locations non-linearly

• hard to portable and need a relatively long sensing period

Challenges

• identify discriminative features from a variety of data sources
  

• incorporate heterogeneous features into a data analytics model effectively​
  

• the labeled data is insufficient

Overview

U-Air: When Urban Air Quality Inference Meets Big Data

Preliminary

• Air quality index (AQI)

United States Environmental Protection Agency

Preliminary

• Trajectory (τ)

• Point of Interests (POI)

• Road Network (RN)

• Grid and Affecting Region (G)

Framework

Framework (cont.)

• Pre-processing data flow

  • ​receive spatial trajectories

  • stored in a trajectory database

  • geo-indexed​

Framework (cont.)

• Learning data flow​

  • spatio-temporal properties (ST)

  • semi-supervised learning approach based on co-training​

    • temporal classifier (TC)

    • spatial classifier (SC)

  • individual model for each kind of pollutant

Framework (cont.)

• semi-supervised learning approach based on
  co-training​

  • temporal classifier (TC)

    • linear-chain conditional random field (CRF)

    • temporally-related features

  • spatial classifier (SC)

    • artificial neural network (ANN)

    • spatially-related features

Framework (cont.)

• Inference data flow

  • offline - spatially-related features

  • online - temporally-related features

  • conduct the inference every hour

TODO

U-Air: When Urban Air Quality Inference Meets Big Data

TODO

• Parse data
  http://nrl.iis.sinica.edu.tw/LASS/AirBox/

• Classify data

  • Spatial-related and temporal-related

  • Other data resources

• Framework

  • Follow this paper