geoman: multi-level attention networks for geo-sensory ... · geoman: multi-level attention...
TRANSCRIPT
GeoMAN: Multi-level Attention Networks for
Geo-sensory Time Series Prediction
Yuxuan Liang, Songyu Ke, Junbo Zhang, Xiuwen Yi, Yu Zheng
Geo-sensory Time Series
• Massive sensors deployed in physical world
Sensors
Roads
Volume: 32
Speed: 50km/h
• Properties
• Each sensor has a unique geospatial location
• Constantly reporting time series readings
• With geospatial correlation between readings
• Prediction on geo-sensory time series
• Motivation: traffic control, air quality forecast…
• Goal: predict target series at a sensor
SensorsPipelines
RC: 0.84
pH: 7.1
Turbidity: 0.54
Sensors
PM2.5: 74
PM10: 90
NO2: 57
𝑡1
𝑡2
𝑡3
Challenges
• Affected by many factors• Readings of previous time interval
• Readings of other sensors in nearby regions
• External factors: weather, time and land use
• Dynamic Inter-sensor correlations
• Dynamic temporal correlation
DAY 1DAY 2
t1
t2
tT
t3
Tim
e
Historical Readings
TimeWeather
POIs Sensor Network
LSTM
Spatial Attn
LSTM
Spatial Attn
LSTM
Spatial Attn
h0
Encoder
Fram
ewo
rk
Spatial Attention
Local Global
Concat
LSTMDecoder
1tc1
ˆ i
ty
Temporal Attn
LSTMtc ˆ i
ty
Co
nca
t
Time Features Embed
Weather Forecasts Embed
SensorID Embed
POIs & Sensor Networks
External Factor Fusion Multi-level Attention Network
LSTMc ˆ iy
𝑿1 𝑿2 𝑿𝑇
t1 t2 tTtT-1
O3
PM10
PM2.5
SO2
CO
???
??
Local spatial attention Global spatial attention External factors fusion
t1
t2
tT
t3
Tim
e
Historical Readings
TimeWeather
POIs Sensor Network
Spatial Attention
𝒉𝑡−1
Local features of a given sensor
Importance of each
local feature at 𝑡
New local features at 𝑡
tanh tanh tanh
softmax
𝒙𝑖,1 𝒙𝑖,𝑘
... ...
𝒙𝑖,𝑁𝑙
𝑎𝑡1 𝑎𝑡
𝑘 𝑎𝑡𝑁𝑙
𝑥𝑡𝑖,1
𝑥𝑡𝑖,𝑘
... ...
... ...
𝑥𝑡𝑖,𝑁𝑙
concat
𝒙𝑡𝑙𝑜𝑐𝑎𝑙
tanh tanh tanh
softmax
𝑿1
... ...
𝛽𝑡1 𝛽𝑡
𝑘 𝛽𝑡𝑁𝑙
𝒉𝑡−1
𝑦𝑡1 𝑦𝑡
𝑘
... ...
... ...
𝑦𝑡𝑁𝑔
concat
𝒙𝑡𝑔𝑙𝑜𝑏𝑎𝑙
𝑿𝑘 𝑿𝑵𝒈
Similarity
matrix
Historical readings of each sensor
• Local spatial attention• Local features target series
• Global spatial attention• Select relevant sensors
Importance of
each sensor at 𝑡
New global features at 𝑡
Local features
at time 𝑡Target series of
all sensors at 𝑡
Temporal Attention
• Sequence-to-sequence architecture
• Select relevant previous time slots to make predictions
𝑿1
𝒉1
𝑿2
𝒉2
𝑿3
𝒉3
𝑿5
𝒉5
𝑿6
𝒉6
𝑿7
𝒉7
𝑿8
𝒉8
𝑿4
𝒉4
𝑿9
𝒉9
Encoder Decoder
Evaluation
• Task 1 - water quality prediction
• Water quality data
• Residual chlorine
• 10 kinds of time series
• From 14 sensors in Shenzhen
• Update each 5 minutes
• Meteorology data
• POIs data
• Task 2 - air quality prediction
• Air quality data
• PM2.5
• 19 kinds of time series
• From 35 sensors in Beijing
• Hourly updates
• Meteorology data
• POIs data
Results
Results
Visualization: Dynamic Correlation
• Case study over air quality dataset
• Discuss on sensor 𝑆0
• 4:00 to 16:00 on Feb. 28, 2017
(a) Air quality stations in Beijing
S0
S1
S11
S6
S13
S17 S32
S23
S27 S26
S3
S4S16
Target sensor Discussed sensor
(b) Plot of local spatial attention weights
0
3
6
9
Wind speed towards different directionsAir pollutants
Southeast
wind
Hu
mid
ity
NO
2
Tem
per
ature
Enco
der
Ste
p
0
3
6
9
S6 S11 S16 S23S1
(c) Plot of global spatial attention weights
Remote sensors
En
cod
er
Ste
pS13
Conclusion
• A very fundamental but challenging task• Dynamic inter-sensor correlation
• Dynamic temporal correlation
• External factors
• Our method• Multi-level attention network
• Spatial attention: captures the dynamic inter-sensor correlation
• Temporal attention: captures the dynamic temporal correlation
• External factor fusion
• Results• More accurate
• Easily interpreted
Data & Code
Thanks!
We are Hiring!
Mail to: [email protected]
Liang, Yuxuan, et al. “GeoMAN: Multi-level Attention Networks for Geo-sensory Time Series Prediction." IJCAI. 2018.