1. Urban Water Quality Prediction based on Multi-task Multi-view Learning
Presented by: Tyler Pietri, Dennis Silva, Michelle Lin

2. Content Background: Problem & Motivation Methodology Evaluation
Conclusion

3. Background: Problem & Motivation

4. Background
Urban water quality consists of: Physical (debris) Chemical (residual chlorine, turbidity, pH Level) Biological (bacteria, algae) Radiological (iodine-131)
Typically, water quality is determined by comparing the physical and chemical characteristics of a water sample with water quality guidelines or standards.
Turbidity is the cloudiness or haziness of a fluid caused by large numbers of individual particles that are generally invisible to the naked eye, similar to smoke in air
Nuclear power plant meltdown in Japan a few years ago.

5. Background
Water quality issues related to urban development: Urban Runoff (Nitrogen & Phosphorus) Sediment Buildup Population Growth Sewage Overflows Waterborne Pathogens
Population Growth: Need more facilities (housing, developments, roads, shopping centers, etc). Disturb land stress water sources
Nitrogen and phosphorus can come from fertilizers in heavy rains or storms, can increase aquatic plants/algae growth
Storm sewers (storm runoff), sanitary sewers (raw sewage), combined sewers (combination raw and stormwater)

6. USA equivalent water is about $0.02 on average, although can vary

8. Background Contaminated water = significant issue
131M people drink contaminated water in USA

9. Motivation Existing Solutions Fail to address complexity
Localized predictions
Lack of global applicability
Healthy populace
Informed policy
Urban projects
Personal motivation and existing solutions that are problematic

10. Existing Solutions
Can you more accurately predict water quality by assuming water quality is correlated?

11. Methodology

12. Data Spatial Heterogeneous Road network Multiple sources
Temporal
Shenzhen, China
Spatial
Road network
Pipeline network
POI
Temporal
Water quality
Weather

13. Framework

14. Temporal View - a single feature vector
The latest 12 hours temporal data in a station. Time Series Signal.
Water quality (RC, Turbidity, pH)
Water hydraulic data (flow, pressure)
Meteorological features: temperature, humidity, barometer pressure, wind speed, weather

15. Spatial View Water pipe network structures (length, diameter and age)
Road network structures (road segment density, road length)
POIs (distribution)
k nearest neighbors (water quality and hydraulic characteristics)
i.e. Neighbors’ temporal features via the geographical similarity (the sum of top-k shortest paths between stations)

16. Prediction Model Spatial prediction Temporal prediction
W is the linear mapping function for station l
Assume both contribute equally, the prediction model

17. Objective function Considering the least-squares loss function
Inherent characteristics of the same node from various aspects
This penalty enforces the agreement on the prediction results

18. Objective function - includes the global impact on a station
This Graph Laplacian penalty ensures a small deviation between two nodes that are near in the pipeline system.
Sl,m measures the spatial autocorrelation between l and m. If it is large, the penalty will force similarity between l and m.

19. L2,1-norm of W(the weight matrix over nodes)
D: # of features
M: # of nodes
Group Lasso penalty.
It encourages all tasks to select a common set of features and thereby plays the role of group feature selection.

20. Optimization The objective function can be written as
h(W) is smooth, g(W) is non-smooth. Use the Fast Iterative Shrinkage Thresholding Algorithm (FISTA) or Accelerated Gradient Descent.

21. Model

22. Evaluation

23. Evaluation Focus Model Performance RSME Varied prediction intervals
Assumption Soundness
Single task
Multitask

24. stMTMV vs. Alternatives
Model Performance
stMTMV = ↓ Error
↑ Interval = ↑ Error
Close approximate of known values
stMTMV vs. Alternatives
stMTMV vs. True Output

25. stMTMV vs. stMTMV Varients
Assumption Validity
Varied inputs
Inclusive model = best model
Proved
water quality = interconnected
Spatial-temporal element
stMTMV vs. stMTMV Varients

26. Assumption Validity Varied Views Temporal (t) Spatial (s)
t+s without alignment
t+s with alignment
Proved
Spatial-temporal element
Alignment is necessary
stMTMV vs. views

27. Future Work

28. Conclusions and Future Work
Explore stMTVT model using different prediction functions
Nonlinear, convex, etc.
Apply model to varied problem (traffic)

29. Q&A

30. References
Liu, Ye, et al. "Urban water quality prediction based on multi-task multi-view learning." Proceedings of the International Joint Conference on Artificial Intelligence "Safe Water Is a Scarce Commodity Worldwide." Statista. N.p., n.d. Web. < Wright, Paul. "5 Infographics That Show The Need For Drinking Water Testing." SCIEX. N.p., n.d. Web. < Survey, U.S. Geological. "Water Resources of the United States." Water Resources of the United States: U.S. Geological Survey. N.p., n.d.. <