understanding the impact of weather for poi recommendations

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


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


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.


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?


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


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Outdated category information needed recrawling


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


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


11

Empirical Analysis

Correlation between weather features


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


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


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


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


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


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


18


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)


19


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∗.


20


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


21



22


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)}


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


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


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


26


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


27


Future Work

Only one weather feature at a time

Incorporate travel distance probabilities underdifferent weather conditions

Incorporate user sensitivity to weather conditions


28


Questions?


29


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.


30


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.


31


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


32


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


33


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.


34


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


35


0 20 40 60 80 100 120 140

Category

0.51

0.52

0.53

0.54

0.55

0.56M

oonp

hase


36


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)}


37


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 ′


38


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


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


40


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%


41


Formulas

SEx =SD√

n(1)


understanding the impact of weather for poi recommendations

Education