This document discusses exploring hubness properties in oceanographic sensor data. It begins by explaining why hubness matters for time series classification using nearest neighbor methods, which are considered the state of the art approach. The document then outlines analyzing a dataset from ocean sensors measuring various metrics like temperature, pressure, and wind speed over 20 days. Visualizing the hubness maps revealed some sensors exhibited unusually high hubness, indicating potentially erroneous measurements. The conclusion is that detecting bad hubness may help identify faulty sensors, and hubness-aware classification methods could benefit analyzing sensor stream data.

1. EXPLORING THE
HUBNESS-RELATED Nenad
Tomašev
PROPERTIES OF Dunja
OCEANOGRAPHIC Mladenić
SENSOR DATA

2. PRESENTATION OUTLINE
Hubness and why it matters
Oceanographic data: overview
Bad hubs in the measurements
Visualizing the problematic sensors

3. WHY IT MATTERS
 Hubness is the skewness (asymmetry) in the distribution of k-
occurrences: some points ( Hubs) become neighbors very VERY
often
 This often happens in high dimensional data
 It is, however, a phenomenon only of importance for nearest-
neighbor methods
 So, why should we care, in general?

4. WHY IT MATTERS
 Sensor data = streams, time series
 The state of the art for time series data: 1 -NN classifier
coupled with an appropriate metric for comparing the time
series
 In other words: nearest neighbor methods are not only
occasionally used for time series classification, they are
considered the state of the art!
 So, hubness matters.

5. RELATED WORK
 Radovanovic, Nanopulous, Ivanovic: Time series classification
in many intrinsic dimensions, SDM 2010
 Due to the correlation between subsequent values, not all
time series are inherently very high dimensional
 Some, however – are. These time series have been shown to
exhibit hubness. Also – bad hubness.
 It was shown that in such cases, bad -hubness-based weighting
is helpful (the hw -kNN algorithm)

6. ANALYSIS GOALS
 Explore the k-nearest neighbor structure of the oceanographic
sensor data
 Explore the bad hubness in the data
 Visualize the results

7. TEST CASE: OCEANOGRAPHIC DATA
 Integrated Ocean Observing System data
(http://www.ioos.gov/)
 Nodes spread across the Pacific, Atlantic and Great lakes…
 Several sensors at each node, measuring various quantities
 air temperature, barometric pressure, wind, water level
observation, water level prediction, salinity, water
temperature and conductivity

8. TEST CASE: OCEANOGRAPHIC DATA
 20 days worth of measurements
 10.11 .-30.11.2010.
 Sampled every 6 minutes (10 measurements an hour)
 4801 measurements total for each sensor
 Missing values: replaced by the average of the closest known
values

9. THE EXPERIMENTAL SETUP
 Tested under two dif ferent metrics
 Manhattan, Variance of between-series differences
 Future work: perform the experiments with DTW (Dynamic Time
Warping)
 Defined “Pacific”, “Atlantic” and “Lakes” as location-based
labels = 3 categories

10. SKEWNESS, BAD HUBNESS

11. CLASS TO CLASS HUBNESS MATRIX, K=3,
WIND MEASUREMENTS
0.772 0.186 0.042
0.013 0.987 0.0
0.027 0.014 0.959
Atlantic = 1. Pacific = 2. Lakes = 3

12. WOULD THE HUBNESS-AWARE METHODS
HELP?

13. WIND MEASUREMENTS: SENSOR
HUBNESS MAP

14. WIND MEASUREMENTS: SENSOR
HUBNESS MAP

15. WATER TEMPERATURE: SENSOR
HUBNESS MAP

16. WATER TEMPERATURE: SENSOR
HUBNESS MAP

17. BAROMETRIC PRESSURE: SENSOR
HUBNESS MAP

18. AIR TEMPERATURE: THE BERMUDA
TRIANGLE 

19. CONCLUSIONS:
 Bad hubness may be useful to detect potentially erroneous
measurement devices
 Some measurement type stream apparently do exhibit
hubness, so hubness is a phenomenon of interest for dealing
with sensor data
 Hubness-aware methods could be potentially helpful when
working with sensor data

20. AKNOWLEDGEMENTS
This work was supported by the ICT
Programme of the EC PlanetData (ICTNoE-
257641).

21. THANK YOU FOR YOUR ATTENTION