Exploring the Data Universe with Semantic Signatures: Plous Lecture 2015

Analogies Observatories Semantic Signatures Challenges Next Steps
Exploring the Data Universe with
Semantic Signatures
Plous Lecture 2015
Krzysztof Janowicz
STKO Lab, University of California, Santa Barbara, USA
Exploring the Data Universe with Semantic Signatures K. Janowicz

Puddingand planets

Analogies & Atoms
Plum Pudding

Analogies & Atoms
Thomson’s Plum Pudding Model (1904)
Positive charge distributed equally in the atom, electrons embedded as raisins

Analogies & Atoms
Rutherford(-Bohr) Solar System Model (1911/13)
Small nucleus with a high mass and electrons that revolve around it

Analogies & Atoms
Analogies
‘And I cherish more than anything the Analogies,
my most trustworthy masters. They know all the
secrets of Nature, and they ought to be least
neglected in Geometry.’ (Johannes Kepler)
Analogies enable us to explore a new
domain (target) by mapping its structure
to another, more familiar domain (source).
They allow us to ask new questions which
only become meaningful in the new domain.

Observatoriesand sensors

Astronomical Observatories And Their Sensors
The Griffith Observatory
Griﬃth donated funds and land to build the observatory to make astronomy accessible to
the public. This was in clear contrast to the prevailing idea of locating observatories on
remote mountaintops and restrict them to scientists. Today, our society is willing to invest
billions to study phenomena that may not even exist anymore (e.g., the Pillars of Creation).

Observatories and Their Sensors
Whether on land or in space, observatories and their sensors serve
diﬀerent purposes and are most useful when they work together.

Spectral Signatures, Bands, and Remote Sensing
Spectral signatures are the combination of emitted, reﬂected, or absorbed
electromagnetic radiation at varying wavelengths (bands) that uniquely
identify a feature type.
Spectral libraries, the idea of sharing spectral signatures, has
revolutionized remote sensing.

A Universe of Data?
The Data Universe: Synthesis Is The New Analysis
What is the common core of the digital universe, physical-cyber-social
systems, digital earth, 4th paradigm, big data, social machines, and so forth?
Synthesis is the new analysis
Observational science versus experimental science
(Unintended)reuse of existing data, semantic interoperability
Heterogeneity: multi-thematic, multi-perspective, multi-resolution

A Universe of Data?
Towards Data Observatories
Web Science Trust: ‘A web observatory is a system that gives public access
to some speciﬁc aspects of the WWW and provides the infrastructure and
visualization techniques to support monitoring, analysis, and experiments.’
Web Science Trust wants to establish a network of observatories.
New questions: are there laws of the data universe?

Constructing the
Analogy

Semantic Signatures and Bands
Semantic Signatures As Analogy To Spectral Signatures
Geospatial bands
based on geographic location
ANND
Ripley’s K Bins
J Measure
Dzero
Temporal bands
based on geo-social check-ins
24 Hours
7 Days
Seasons
Thematic bands
based on venue tips and reviews
LDA topics
TF-IDF
Makes use of data
heterogeneity

Thematic Bands
Thematic Bands & Geo-Indicativeness
Places at geographic location 34.43, -119.71 are:
of types city, county seat,...
at the coastline, near the mountains, have Mediterranean climate,...
described in terms of urban area, economy, tourism, government, employment,...
Interesting observation: some of these terms will co-occur by type, others per region.

Thematic Bands
Thematic Bands & Geo-Indicativeness
A thematic band can be
computed out of unstructured
text from sources such as
Wikipedia, travel blogs, news
articles, and so forth.
Non-georeferenced plain text
is often still geo-indicative
Diﬀerent types of geographic
features have diﬀerent,
diagnostic topics associated to
them (out of 500 topics)
Indicative topics and be lifted to
the type-level.
Here, we modeled topics using
latent Dirichlet allocation (LDA)

Thematic Bands
Thematic Bands & Geographic Feature Types
City topics: 204>450>104>282>267>497>443>484>277>97>...
Town topics: 425>450>419>367>104>429>266>69>204>308>...
Mountain topics: 27>110>5>172>208>459>232>398>453>183>...

Temporal Bands
Temporal Bands
Study geo-social
check-in data to
location-based social
networks.
Aggregate them to the
feature type level and
clean them.
Intuitively, people visit
wineries in the
after-noon and evening
and bakeries in the
mornings.
Combining weekly and
hourly bands to create
place type signatures.

Spatial Bands
Spatial Bands
POI plotted by similarity to bar and post oﬃce in OpenStreetMap data (London)
Similarity measured as association strength in OSM change history
Bars (and similar features) tend to clump together
Post Oﬃces (and similar features) are rather uniformly distributed

Spatial Bands
Spatial Bands
Dzero measures the likelihood of features of a certain type to co-occur
within a speciﬁc semantic and spatial range.
General idea: generate recommendations and clean up data based on
type likelihood. ’How likely is a post oﬃce directly next to an existing one?’

Sensor Resolution & Social Sensing
(Remote sensing) sensors can be characterized by their resolution
Spatial resolution: smallest feature that can be detected, i.e., the pixel size.
Temporal resolution: smallest time interval between a repeated observation.
Spectral resolution: number, position, and width of spectral bands.
Radiometric resolution: small distinguishable diﬀerences in radiation magnitude.
Analogous social sensor resolutions, e.g., types of bands, number of topics.

Platial Resolution of Termporal Signatures
Circular temporal signatures histograms for Theme Park (a,b,c) and Drugstore (d,e,f).
About 50% of ≈ 400 Point Of Interest (POI) types are regionally invariant in the USA.

Temporal Resolution of Termporal Signatures
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
The ’Foursquare-day’
How and when do people check-in at places, manually, automatically?
Do they check-out? If not, after what time are they checked-out automatically?

Distinguishable Feature Types For Thematic Signatures From 500-Topics
Which classes in a feature type schema can be meaningfully distinguished?

SpatialSearch Challenges

From Space to Place Through Time
1. Challenge: Mapping User Locations from Spaces to Places

1. Challenge: Mapping User Locations from Spaces to Places
Estimate the place visited by a user from the user’s spatial location
(e.g., as measured by their smartphone).

Baseline: Google Place API
Marker Category Distance (m)
A Bakery 39.2
B Nightclub 41.4
C Nightclub 69.9
D American Restaurant 62.7
E Bakery 73.7
F Fast Food 65.0
G Apparel Store 85.8
H Ice Cream Shop 82.6
I Movie Theater 94.2
J Pub 88.9
K Cosmetics Shop 60.9
L Diner 70.0
M Italian Restaurant 45.7
N Furniture / Home Store 114.9
O Grocery Store 147.8
P BBQ Joint 82.3
Q Burrito Place 88.1
R Italian Restaurant 93.6
Geolocation APIs map geographic coordinates, e.g., from a user’s
smartphone, to an ordered sets of nearby candidate POI.
These services typically return the n nearest POI within a certain radius and
use spatial distance to the provided coordinates to determine their order.

Our Approach: Distort POI Locations Using Temporal Signatures
Marker Category Distance (m) Monday 10AM (10−3) Saturday 11PM (10−3)
A Bakery 39.2 6.28 4.08
B Nightclub 41.4 0.26 44.16
C Nightclub 69.9 0.26 44.16
D American Restaurant 62.7 1.61 9.50
E Bakery 73.7 6.28 4.08
F Fast Food 65.0 4.80 5.78
G Apparel Store 85.8 2.51 1.09
H Ice Cream Shop 82.6 0.84 15.88
I Movie Theater 94.2 1.44 11.00
J Pub 88.9 0.53 22.66
K Cosmetics Shop 60.9 3.87 1.57
L Diner 70.0 5.49 7.56
M Italian Restaurant 45.7 1.42 7.96
N Furniture / Home Store 114.9 4.79 5.01
O Grocery Store 147.8 4.53 1.38
P BBQ Joint 82.3 0.43 9.35
Q Burrito Place 88.1 0.54 3.16
R Italian Restaurant 93.6 1.42 7.96
The likelihood of visiting a coﬀee shop, university, bakery, etc at 7pm is
rather low, while it is a peak hour for restaurants.
In analogy to scale distortion in cartography, we can modify the purely spatial
ranking by pulling and pushing places based on the check-in probability
of their temporal type signatures.
Diﬀerent distortion models: linear, non-linear, symmetrical, non-symmetric

Our Approach: Distort POI Locations Using Temporal Signatures
Marker Actual
Dist.(m)
Distorted
Dist.(m)
A 39.2 25.8
B 41.4 71.4
C 69.9 99.9
D 62.7 79.8
E 73.7 60.3
F 65.0 59.5
G 85.8 95.6
H 82.6 106.7
I 94.2 112.8
J 88.9 116.1
K 60.9 61.1
L 70.0 60.6
M 45.7 64.5
N 114.9 109.5
O 147.8 143.9
P 82.3 110.5
Q 88.1 115.2
R 93.6 112.4
Method MRR SRR nDCG 1st Pos.
Distance-Only 0.359 443.8 0.583 211
Temporally Adjusted 0.453 793.5 0.711 423

Vague Cognitive Regions: Where is SoCal?
2. Challenge: Vague Cognitive Regions
Where is SoCal and NorCal?

Baseline: Tests With Human Participants
44 participants, 90 hexagon tessellation (≈ 4920km2
each)
Google Maps search for SoCal
[More on the extraction of polygons at spatial@ucsb.local2015]

Data and Correlations
Source SoCal NorCal Total
Flickr 22132 19706 41838
Instagram 169648 116984 286632
Twitter 10376 3294 13670
Travel Blogs 107 78 185
Wikipedia 1450 700 2150
0 1000 2000 3000 4000
0.00.20.40.60.81.0
Empirical Cumluative Distribution
Flickr photo counts per user
CDF
Source ρ (M1) τ (M1)
Flickr 0.881 0.721
Instagram 0.867 0.711
Twitter 0.874 0.714
TravelBlogs & Wikipedia 0.897 0.74
Means 0.870 0.712

Vague Cognitive Regions and Inter-rater Agreement
5
5
5 6.25
5.75
6.75
5.25
5.75
5.25
6.75
4.25
6.75
6.75
5.75
5.25
6.75
5.75
5.25
5.75
4.75
6.75
6.75
4.25 4.75
5.25
4
4
5
4
4
4
5
5.5
5.5
3.5
6.5
1.5
6.5
6.5 5.5
5.5
3.5
4.5
4.253
3
3
3
3.5
2.5
3.5
2.5
2.5
2.5
3.5
2.5
3.25
3.25
3.75
1.75
2.75
3
2
2
2
2.25 2.25
2.25
2.25
0 80 160 240 32040
Miles
®
Legend
Insufficient Data
Very Northern Californian
Moderately Northern Californian
Slightly Northern Californian
Equally Northern and Southern Californian
Slightly Southern Californian
Moderately Southern Californian
Very Southern Californian
Standard Deviations
< 0.01
0.01 - 0.50
0.51 - 1.00
1.01- 1.73
> 1.73
Source Four Raters Five Raters
Kendall’s W 0.953 0.929
p-value < 0.001 < 0.001
Key idea: Data sources become
raters/ participants.

Vague Cognitive Regions and Thematic Signatures
Do you even have to mine for the Socal and Norcal term directly?

Vague Cognitive Regions and Self-Similarity
0 5 10 15
0.000.050.100.150.20
KLD Divergence
Northern California
Southern California
Both Northern & Southern California
Based on 60 topics, the similarity between SoCal (and NorCal) cells is
higher than between SoCal and NorCal cells.

Is the Data Universe Homogenous amd Isotropic?
Limits Of The Data Universe Analogy
At large scale, the physical universe is homogenous and isotropic

Is the Data Universe Homogenous amd Isotropic?
Limits Of The Data Universe Analogy
In terms of geospatial distribution the Social Media Web is neither homogenous
nor isotropic. If you direct your social sensing instrument to a certain region,
there will be no signal.

NextSteps

POI Pulse Observatory
POI Pulse Observatory: Explore the Pulse of Los Angeles Using Signatures
http://poipulse.com/

POI Pulse Observatory
A Public Data Observatory at UCSB?
A tangible & public observatory at
UCSB; remember Griffith’s will.
Show & stream data from different
sources and show analysis results
Visualize privacy implications of data
Show citizens how their everyday data
is used for scientific discoveries

The Right Place

Exploring the Data Universe with Semantic Signatures: Plous Lecture 2015

More Related Content

Similar to Exploring the Data Universe with Semantic Signatures: Plous Lecture 2015 (20)

More from kjanowicz (15)

Recently uploaded (20)

Exploring the Data Universe with Semantic Signatures: Plous Lecture 2015