S C I E N C E P A S S I O N T E C H N O L O G Y
http://know-center.tugraz.at/
Understanding the Impact of Weather forPOI RecommendationsChristoph Trattner, Alex Oberegger, Lukas Eberhard, Denis Parra, Le-andro Marinho, Know-Center@Graz University of Technology
September 15, 2016
2
Understanding the Impact of Weather for POI Recommendations,
Content
1. Motivation
2. Dataset
3. Empirical Analysis
4. Weather Aware POI Recommender (WPOI)
5. Conclusions & Future Work
Christoph Trattner, Alex Oberegger, Lukas Eberhard, Denis Parra, Leandro Marinho, Know-CenterSeptember 15, 2016
3
Motivation
Motivation
Geolocation services on portable devicesLocation-based social networks (LBSN)
YelpFoursquare
Vast amount of check-in data
Comments on placesRecommendations on placesRatings on places
Use data to assist users
Christoph Trattner, Alex Oberegger, Lukas Eberhard, Denis Parra, Leandro Marinho, Know-CenterSeptember 15, 2016
4
Motivation
Context aware recommender systems (CARS)concentrated on
Social contextGeographical contextTime context
No research on the impact of weather
Pool not popular at rainy conditionsIce cream shop not popular at cold temperatures
Christoph Trattner, Alex Oberegger, Lukas Eberhard, Denis Parra, Leandro Marinho, Know-CenterSeptember 15, 2016
5
Motivation
Problem Statement
Definition
Given a user u, the user’s check-in history Lu, i.e., thePOIs that the user has visited in the past, and the currentweather context c the aim is to predict the POIsL̂u = {l1, . . . , l|L|} that the user will likely visit in the futurethat are not in Lu.
Christoph Trattner, Alex Oberegger, Lukas Eberhard, Denis Parra, Leandro Marinho, Know-CenterSeptember 15, 2016
6
Motivation
Research Questions
RQ1 Are the users’ mobility patterns influenced byweather?
RQ2 How can weather context information beincorporated into existent recommender systems?
RQ3 To what extent can weather information be usedto increase recommender accuracy?
RQ4 Which weather feature provides the highestimpact on the recommender accuracy?
Christoph Trattner, Alex Oberegger, Lukas Eberhard, Denis Parra, Leandro Marinho, Know-CenterSeptember 15, 2016
7
Dataset
Dataset
Foursquare Check-in Dataset from Dingqi Yang [3][4]
Worldwide check-ins from April 2012 to September2013
Filtered by U.S. cities
∼3,000,000 Check-ins∼500,000 Venues∼50,000 Users60 Cities
Christoph Trattner, Alex Oberegger, Lukas Eberhard, Denis Parra, Leandro Marinho, Know-CenterSeptember 15, 2016
8
Dataset
Dataset
Weather Data from forecast.io weather API [1]
One API call per <venue, city, time> triple∼27,000 API callsEight weather attributes
Visibility, Precipitation intensity, Humidity, Cloud cover,Pressure, Windspeed, Temperature, Moonphase
Outdated category information needed recrawling
Christoph Trattner, Alex Oberegger, Lukas Eberhard, Denis Parra, Leandro Marinho, Know-CenterSeptember 15, 2016
9
Dataset
Check-in Distributions
Categories
100 101 102 103 104 105 106
Checkin Frequency (log)
0
2
4
6
8
10
12
Fre
quen
cy
∼200 main categories
User
100 101 102 103 104
Checkin Frequency (log)
100
101
102
103
104
Fre
quen
cy (
log)
∼11,000 user with >100check-ins
Christoph Trattner, Alex Oberegger, Lukas Eberhard, Denis Parra, Leandro Marinho, Know-CenterSeptember 15, 2016
10
Empirical Analysis
Weather Distributions
Temperature
30 20 10 0 10 20 30 40 50Temperature
0.000
0.005
0.010
0.015
0.020
0.025
0.030
0.035
0.040
0.045
Che
ckin
pro
babi
lity
σ = 9.36 µ = 16.69 Data
Normal distributed
Moonphase
0.0 0.2 0.4 0.6 0.8 1.0Moonphase
0.0
0.2
0.4
0.6
0.8
1.0
1.2
1.4
Che
ckin
pro
babi
lity
σ = 0.29 µ = 0.53 Data
Uniform distributed
Christoph Trattner, Alex Oberegger, Lukas Eberhard, Denis Parra, Leandro Marinho, Know-CenterSeptember 15, 2016
11
Empirical Analysis
Correlation between weather features
Christoph Trattner, Alex Oberegger, Lukas Eberhard, Denis Parra, Leandro Marinho, Know-CenterSeptember 15, 2016
12
Empirical Analysis
Impact on Travel Distance
Cloud Cover
100 101 102
Distance(log(KM))
106
105
104
103
102
101
Pro
babi
lity
(log)
Low Cloud coverHigh Cloud cover
Pressure
100 101 102
Distance(log(KM))
106
105
104
103
102
101
Pro
babi
lity
(log)
Low PressureHigh Pressure
Christoph Trattner, Alex Oberegger, Lukas Eberhard, Denis Parra, Leandro Marinho, Know-CenterSeptember 15, 2016
13
Empirical Analysis
Impact on Category Popularity
Ice Cream Shop
30 20 10 0 10 20 30 40 50Temperature
0.02
0.01
0.00
0.01
0.02
0.03
0.04
0.05
0.06
0.07
Che
ckin
pro
babi
lity
Ski Area
20 10 0 10 20 30 40Temperature
0.02
0.00
0.02
0.04
0.06
0.08
0.10
0.12
Che
ckin
pro
babi
lity
Christoph Trattner, Alex Oberegger, Lukas Eberhard, Denis Parra, Leandro Marinho, Know-CenterSeptember 15, 2016
14
Empirical Analysis
User Mobility Patterns
Mobility pattern of a“frosty” user
20 10 0 10 20 30 40Temperature
0.02
0.00
0.02
0.04
0.06
0.08
0.10
Che
ckin
pro
babi
lity
Mobility pattern of a“heated” user
20 10 0 10 20 30 40Temperature
0.02
0.00
0.02
0.04
0.06
0.08
0.10
Che
ckin
pro
babi
lity
Christoph Trattner, Alex Oberegger, Lukas Eberhard, Denis Parra, Leandro Marinho, Know-CenterSeptember 15, 2016
15
Empirical Analysis
Seasonal Impact on Categories
All Cate
gorie
s
Nation
al Par
k
Ski Are
a
Austria
n Res
taura
nt
Hawaii
an R
estau
rant
Hotel P
ool
Zoo
Ice C
ream
Sho
p
Baseb
all S
tadium
Mexica
n Res
taura
ntGym
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7C
heck
in p
roba
bilit
yWinterSpringSummerAutumn
Christoph Trattner, Alex Oberegger, Lukas Eberhard, Denis Parra, Leandro Marinho, Know-CenterSeptember 15, 2016
16
Empirical Analysis
Regional Weather Variability
Hon
olul
uO
akla
ndP
hoen
ixS
an D
iego
Mia
mi
Los
Ang
eles
San
Jos
eT
ampa
El P
aso
Sea
ttle
Por
tland
Mem
phis
New
Orle
ans
Sac
ram
ento
Nas
hvill
eS
an A
nton
ioD
alla
sC
harlo
tteA
tlant
aJa
ckso
nvill
eR
alei
ghT
alla
hass
eeH
oust
onK
ansa
s C
ityS
t. Lo
uis
Den
ver
Ric
hmon
dN
orfo
lkB
altim
ore
Aus
tinW
ashi
ngto
n D
.C.
Indi
anap
olis
Pitt
sbur
ghS
alt L
ake
City
Cle
vela
ndR
oche
ster
Chi
cago
Okl
ahom
a C
ityC
inci
nnat
iD
etro
itB
uffa
loP
hila
delp
hia
Col
umbu
sC
olum
bia
Milw
auke
eC
hest
erN
ew Y
ork
Lans
ing
Bos
ton
Har
risbu
rgS
aint
Pau
lH
artfo
rdD
es M
oine
sP
rovi
denc
eN
ewar
kT
rent
onM
inne
apol
isB
rook
lyn
Mad
ison
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9S
tand
ard
devi
atio
nTotal variability normed over all weather features
South citiesNorth cities
Christoph Trattner, Alex Oberegger, Lukas Eberhard, Denis Parra, Leandro Marinho, Know-CenterSeptember 15, 2016
17
Weather Aware POI Recommender (WPOI)
Weather Aware POI Recommender (WPOI)Many basic recommender algorithms available
Most Popular (MP)KNN-Algorithms (User, Item)Matrix Factorization (MF)
Enrich existing recommender system with weathercontextTest recommender accuracy in four citiesrepresenting variety of climate (cardinal directions)
MinneapolisBostonMiamiHonolulu
Christoph Trattner, Alex Oberegger, Lukas Eberhard, Denis Parra, Leandro Marinho, Know-CenterSeptember 15, 2016
18
Weather Aware POI Recommender (WPOI)
Rank-GeoFM by Li et al. (2015) [2]
Based on MFContext aware (Geo and Time)Easy to extendHigh accuracy
Based on iterative learning of latent modelparameters (weights)
Learns model parameters with stochastic gradientdescent (SGD)
Christoph Trattner, Alex Oberegger, Lukas Eberhard, Denis Parra, Leandro Marinho, Know-CenterSeptember 15, 2016
19
Weather Aware POI Recommender (WPOI)
yult = u(1)u · l (1)
l︸ ︷︷ ︸user preference score
+ t (1)t · l (2)
l︸ ︷︷ ︸time preference score
+
+ u(2)u ·
∑l∗∈Nk (l)
wll∗ l(1)l∗︸ ︷︷ ︸
geographical influence score
+ l (3)l ·
∑t∗∈T
mtt∗t(1)t∗︸ ︷︷ ︸
temporal influence score
U(1,2),L(1,...,3),T (1) → latent model parameters to learn
wll is the probability that l is visited given that l∗ hasbeen visited in terms of geographical position and mttis the probability that the popularity of a POI in timeslot t is influenced by those in time slot t∗.
Christoph Trattner, Alex Oberegger, Lukas Eberhard, Denis Parra, Leandro Marinho, Know-CenterSeptember 15, 2016
20
Weather Aware POI Recommender (WPOI)
Replace time with weather feature
Start algorithm for each weather feature separately
Measure accuracy and compare with RankGeoFMand base algorithms
yulc = u(1)u · l (1)
l︸ ︷︷ ︸user preference score
+ f (1)c · l (2)
l︸ ︷︷ ︸weather preference score
+
+ u(2)u ·
∑l∗∈Nk (l)
wll∗ l(1)l∗︸ ︷︷ ︸
geographical influence score
+ l (3)l ·
∑c∗∈FC
wicc∗f(1)c∗︸ ︷︷ ︸
weather influence score
Christoph Trattner, Alex Oberegger, Lukas Eberhard, Denis Parra, Leandro Marinho, Know-CenterSeptember 15, 2016
21
Weather Aware POI Recommender (WPOI)
Christoph Trattner, Alex Oberegger, Lukas Eberhard, Denis Parra, Leandro Marinho, Know-CenterSeptember 15, 2016
22
Weather Aware POI Recommender (WPOI)
Algorithm 1: WPOIInput: check-in data D, hyperparameters that steer context influence and learning rate
1 Init: Initialize Θ withN (0, 0.01); Shuffle D2 repeat3 for (u, l, c) ∈ D do4 Compute yulc and n = 05 repeat6 Sample a POI l′ and feature class c′, Compute yul′c′ and set n++7 until (u,l’,c’) ranked inccorrectly || everything ranked correct8 if (u,l’,c’) ranked inccorrectly then9 update latent model parameters Θ
10 according to the gradient of the error function.11 end12 end13 until convergence14 return Θ = {L(1), L(2), L(3),U(1),U(2), F (1)}
Christoph Trattner, Alex Oberegger, Lukas Eberhard, Denis Parra, Leandro Marinho, Know-CenterSeptember 15, 2016
23
Weather Aware POI Recommender (WPOI)Boston
0 1000 2000 3000 4000 5000Iteration
0.00
0.01
0.02
0.03
0.04
0.05
0.06M
easu
reBoston, WPOI (Temperature)
NDCG@20
Minneapolis
0 1000 2000 3000 4000 5000Iteration
0.00
0.01
0.02
0.03
0.04
0.05
0.06
Mea
sure
Minneapolis, WPOI (Temperature)NDCG@20
Miami
0 1000 2000 3000 4000 5000Iteration
0.00
0.01
0.02
0.03
0.04
0.05
Mea
sure
Miami, WPOI (Temperature)NDCG@20
Honolulu
0 1000 2000 3000 4000 5000Iteration
0.00
0.01
0.02
0.03
0.04
0.05
0.06
0.07
Mea
sure
Honolulu, WPOI (Temperature)NDCG@20
Christoph Trattner, Alex Oberegger, Lukas Eberhard, Denis Parra, Leandro Marinho, Know-CenterSeptember 15, 2016
24
Weather Aware POI Recommender (WPOI)Boston
RankG
eoFM
WPOI (
Press
.)
RankG
eoFMT
WPOI (
Hum.)
WPOI (
Wind
sp.)
WPOI (
Vis.)
WPOI (
Cld.C
ov.)
WPOI (
Temp.)
WPOI (
Moonp
h.)
WPOI(P
rec.
Int.)
0.00
0.01
0.02
0.03
0.04
0.05
0.06
0.07
ND
CG
@20
Minneapolis
RankG
eoFM
WPOI (
Wind
sp.)
WPOI (
Hum.)
WPOI (
Temp.)
WPOI (
Press
.)
WPOI(P
rec.
Int.)
WPOI (
Moonp
h.)
RankG
eoFMT
WPOI (
Cld.C
ov.)
WPOI (
Vis.)
0.00
0.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08
ND
CG
@20
Miami
RankG
eoFM
WPOI (
Hum.)
RankG
eoFMT
WPOI (
Cld.C
ov.)
WPOI (
Moonp
h.)
WPOI (
Wind
sp.)
WPOI (
Press
.)
WPOI (
Temp.)
WPOI (
Vis.)
WPOI(P
rec.
Int.)
0.00
0.01
0.02
0.03
0.04
0.05
0.06
ND
CG
@20
Honolulu
RankG
eoFM
WPOI (
Press
.)
WPOI (
Hum.)
WPOI (
Moonp
h.)
WPOI (
Wind
sp.)
WPOI (
Temp.)
RankG
eoFMT
WPOI (
Cld.C
ov.)
WPOI (
Vis.)
WPOI(P
rec.
Int.)
0.00
0.02
0.04
0.06
0.08
0.10
ND
CG
@20
Christoph Trattner, Alex Oberegger, Lukas Eberhard, Denis Parra, Leandro Marinho, Know-CenterSeptember 15, 2016
25
Conclusions & Future Work
Conclusions
Empirical Analysis
Weather features have little impact on travel distance
Seasonality has an impact on category popularity
Category popularity is dependent on the region
Higher weather variability in the north
Christoph Trattner, Alex Oberegger, Lukas Eberhard, Denis Parra, Leandro Marinho, Know-CenterSeptember 15, 2016
26
Conclusions & Future Work
WPOI
Weather context more useful than time
WPOI better in regions closer to tropical zone
Precipitation intensity and visibility improve accuracybest
Weather context is indeed a useful contextualinformation in POI recommender systems
Christoph Trattner, Alex Oberegger, Lukas Eberhard, Denis Parra, Leandro Marinho, Know-CenterSeptember 15, 2016
27
Conclusions & Future Work
Future Work
Only one weather feature at a time
Incorporate travel distance probabilities underdifferent weather conditions
Incorporate user sensitivity to weather conditions
Christoph Trattner, Alex Oberegger, Lukas Eberhard, Denis Parra, Leandro Marinho, Know-CenterSeptember 15, 2016
28
Conclusions & Future Work
Questions?
Christoph Trattner, Alex Oberegger, Lukas Eberhard, Denis Parra, Leandro Marinho, Know-CenterSeptember 15, 2016
29
Conclusions & Future Work
References I
[1] FORECAST.IO.
forecast.io weather api, 2015.
[2] LI, X., CONG, G., LI, X.-L., PHAM, T.-A. N., AND KRISHNASWAMY, S.
Rank-geofm: A ranking based geographical factorization method for point ofinterest recommendation.
In Proceedings of the 38th International ACM SIGIR Conference onResearch and Development in Information Retrieval (New York, NY, USA,2015), SIGIR ’15, ACM, pp. 433–442.
Christoph Trattner, Alex Oberegger, Lukas Eberhard, Denis Parra, Leandro Marinho, Know-CenterSeptember 15, 2016
30
Conclusions & Future Work
References II
[3] YANG, D., ZHANG, D., CHEN, L., AND QU, B.
Nationtelescope: Monitoring and visualizing large-scale collective behaviorin lbsns.
Journal of Network and Computer Applications 55 (2015), 170–180.
[4] YANG, D., ZHANG, D., AND QU, B.
Participatory cultural mapping based on collective behavior in locationbased social networks.
ACM Transactions on Intelligent Systems and Technology (2015).
in press.
Christoph Trattner, Alex Oberegger, Lukas Eberhard, Denis Parra, Leandro Marinho, Know-CenterSeptember 15, 2016
31
Conclusions & Future Work
Weather feature Properties Range in dataset
Precipitation intensity Precipitation inten-sity measured inmilimeters of liquidwater/hour.
0mm/h − 34, 29mm/h
Temperature Temperature mea-sured in degreeCelsius
−24, 48◦ − 46, 58◦
Wind speed Wind speedmeasured in me-ters/second
0m/s − 19, 13m/s
Cloud cover Value between 0 and1 displaying the per-centage of the skycovered by clouds.
0− 1
Christoph Trattner, Alex Oberegger, Lukas Eberhard, Denis Parra, Leandro Marinho, Know-CenterSeptember 15, 2016
32
Conclusions & Future Work
Humidity Value between 0 and 1representing the “Per-centage relative humidity”is defined as the partialpressure of water vapor inair divided by the vaporpressure of water at thegiven temperature.”
0, 02φ− 1, 00φ
Pressure Atmospheric pressuremeasured in hectopas-cals.
957, 11hPa− 1046, 05hPa
Visibility Value representing the av-erage visibility in kilome-ters capped at 16,09
0km − 16, 09km
Moonphase Value from 0 to 1 rep-resenting the range be-tween new moon and fullmoon
0− 1
Christoph Trattner, Alex Oberegger, Lukas Eberhard, Denis Parra, Leandro Marinho, Know-CenterSeptember 15, 2016
33
Conclusions & Future Work
Correlations
Humidity Temperature -0.99 Relative humidity repre-sents the saturation ofmoisture in the air andcold air does not need thatmuch moisture to be satu-rated.
Windspeed Temperature 0.93 Foehn-effect, Windchillfactor not included
Humidity Windspeed -0.93 See positive correlationbetween windspeed andtemperature and nega-tive between humidity andtemperature.
Christoph Trattner, Alex Oberegger, Lukas Eberhard, Denis Parra, Leandro Marinho, Know-CenterSeptember 15, 2016
34
Conclusions & Future Work
Correlations
Temperature Visibility 0.77 At colder temperaturessaturation is reached ear-lier and therefore fog oc-curs more often that blursvisibility.
Humidity Visibility -0.77 High humidity blurs visibil-ity
Cloud cover Visibility -0.67 Clouds cover the sun →visibility is low
Christoph Trattner, Alex Oberegger, Lukas Eberhard, Denis Parra, Leandro Marinho, Know-CenterSeptember 15, 2016
35
Conclusions & Future Work
0 20 40 60 80 100 120 140
Category
0.51
0.52
0.53
0.54
0.55
0.56M
oonp
hase
Christoph Trattner, Alex Oberegger, Lukas Eberhard, Denis Parra, Leandro Marinho, Know-CenterSeptember 15, 2016
36
Conclusions & Future Work
Algorithm 2: WPOIInput: check-in data D, geographical influence matrix W , weather influence matrix WI, hyperparameters
ε,C, α, β and learning rate γg and γw
Output: parameters of the model Θ = {L(1), L(2), L(3),U(1),U(2), F (1)}1 init: Initialize Θ withN (0, 0.01); Shuffle D2 repeat3 for (u, l, c) ∈ D do4 Compute yulc and n = 05 repeat6 Sample a POI l′ and feature class c′, Compute yul′c′ and set n++7 until I(xulc > xul′c′ )I(yulc < yul′c′ + ε) = 1 or n > |L|8 if I(xulc > xul′c′ )I(yulc < yul′c′ + ε) = 1 then9 η = E
(⌊|L|n
⌋)δull′
10 g =(∑
c∗∈FCfwic′c∗ f (1)
c∗ −∑
c+∈FCfwicc+ f (1)
c+
)11 f (1)
c ← f (1)c − γwη(l(2)
l′− l(2)
l )
12 l(3)l ← l(3)
l − γwηg
13 l(2)
l′← l(2)
l′− γwηfc
14 l(2)l ← l(2)
l + γwηfc15 end16 end17 Project updated factors to accomplish constraints18 until convergence19 return Θ = {L(1), L(2), L(3),U(1),U(2), F (1)}
Christoph Trattner, Alex Oberegger, Lukas Eberhard, Denis Parra, Leandro Marinho, Know-CenterSeptember 15, 2016
37
Conclusions & Future Work
Symbol Description
U set of users u1, u2, ..., u|U|
L set of POIs l1, l2, ..., l|L|
FCf set of classes for feature f
F set of weather feature classes f1, f2, ..., f|FCf |
Θ latent model parameters containing the learned weights{L(1), L(2), L(3),U(1),U(2),F (1)} for locations, users and weatherfeatures
Xul |U|x |L| matrix containing the check-ins of users at POIs
Xulc |U|x |L|x |FCf | matrix containing the check-ins of users at POIs ata specific feature class c
D1 user-POI pairs: (u, l)|xul > 0
D2 user-POI-feature class triples: (u, l, c)|xulc > 0
d(l, l ′) geo distance between the latitude and longitude of l and l ′
Christoph Trattner, Alex Oberegger, Lukas Eberhard, Denis Parra, Leandro Marinho, Know-CenterSeptember 15, 2016
38
Conclusions & Future Work
W geographical probability matrix of size |L|x |L| where wll′
contains the probability of l ′ being visited after l has beenvisited according to their geographical distance. wll′ =(0.5 + d(l, l ′))−1)
WI probability that a weather feature class c is influenced by
feature class c′. wicc′ =∑
u∈U∑
l∈L xulc xulc′√∑u∈U
∑l∈L x2
ulc
√∑u∈U
∑l∈L x2
ulc′
Nk (l) set of k nearest neighbors of POI l
yul the recommendation score of user u and POI l
yulc the recommendation score of user u, POI l and weatherfeature class c
I(·) indicator function returning I(a) = 1 when a is true and 0otherwise
ε margin to soften ranking incompatibility
Incomp(yulc , ε) a function that counts the number of locations l ′ ∈ L thatshould be ranked lower than l at the current weather contextc and user u but are ranked higher by the model.
γw learning rate for updates on weather latent parameters.Christoph Trattner, Alex Oberegger, Lukas Eberhard, Denis Parra, Leandro Marinho, Know-CenterSeptember 15, 2016
39
Conclusions & Future Work
γg learning rate for updates on latent parameters from base ap-proach.
E(·) a function that turns the rating incompatibility Incomp(yulc , ε)
into a loss. E(r) =∑r
i=11i
O objective function to minimize during the iterative learning.O =
∑(u,l,c)∈D2 E(Incomp(yulc , ε))
s(a) sigmoid function s(a) = 11+exp(−a)
δucll′ function to approximate the indicator function with a continu-ous sigmoid function. δucll′ = s(yul′c + ε− yulc)(1− s(yul′c +ε− yulc))
b |L|n c if the nth location l ′ was ranked incorrect by the model the ex-pactation is that overall b |L|n c locations are ranked incorrect.
∂E∂Θ
calculation of stochastic gradient =
E(b |L|n c
)δucll′
∂(yucl′+ε−yucl )
∂Θ
Θ← Θ− ∂E∂Θ
SGD based optimization of the latent model parameters
Christoph Trattner, Alex Oberegger, Lukas Eberhard, Denis Parra, Leandro Marinho, Know-CenterSeptember 15, 2016
40
Conclusions & Future Work
Dataset Statistics
City #Users #Venues #Check-ins Sparsity
Minneapolis 436 797 37,737 89.1%
Boston 637 1141 42,956 94.3%
Miami 410 796 29,222 91.0%
Honolulu 173 410 16,042 77.4%
Christoph Trattner, Alex Oberegger, Lukas Eberhard, Denis Parra, Leandro Marinho, Know-CenterSeptember 15, 2016
41
Conclusions & Future Work
Formulas
SEx =SD√
n(1)
Christoph Trattner, Alex Oberegger, Lukas Eberhard, Denis Parra, Leandro Marinho, Know-CenterSeptember 15, 2016