cross-context multi-level interpretations to generate personalized recommendations

Post on 06-Jul-2018

212 Views

Category:

Documents

0 Downloads

Preview:

Click to see full reader

TRANSCRIPT

8/17/2019 Cross-Context Multi-level Interpretations to Generate Personalized Recommendations

http://slidepdf.com/reader/full/cross-context-multi-level-interpretations-to-generate-personalized-recommendations 1/3

(심사용 작성양식) // 온라인 접수시, 본 라인은 삭제한 후 제출해 주세요

간 문맥 멀티 레 해!은 "# $%& '성()

Cross-Context Multi-level Interpretations to generate

Personalized RecommendationsSyed Imran Ali and Sungyoung Lee

Department of Computer Science and Engineering Kyung Hee University

Seocheon-dong !iheung-gu "ongin-si !yeonggi-do #epu$lic of Korea %%&-'()

 * imran+ali sylee ,osla$+.hu+ac+.rA$stract

A huge array of personalized healthcare and wellness systems are introduced for digital health and

Quantified Self movement in recent years. These systems share common features which include

tasks such as self-tracking, self-quantifications, and self-monitoring ased on the raw sensory data.These features assist users to e more aware of their health. !owever, such systems cannot e

considered sufficient enough for changing the unhealthy haits of users. To induce healthy haits in

users, system needs to provide situation aware personalized recommendations. "ne ig challenge

in this regard is the careful interpretations of user conte#t, environmental variales, and user

preferences. $e propose an innovative method for cross-conte#t multi-level interpretations to

generate personalized health and wellness recommendations. $ith such interpretations, the self-

quantification systems can e advanced to generate personalized recommendations in addition to

tracking, quantifying, and monitoring. %y achieving this, we can increase the ownership of the

application y the user, thus increase the chances of healthy haits promotion and induction.

1. IntroductionA huge array of systems for personalized healthcare and

wellness management has rapidly grown during the recentyears due to the rapid increase in wearale and moile

technologies &'(. These systems and applications have

augmented the portfolio of digital health and

quantification-self movement. $ith advances in Self-

Quantification, systems capturing and recording data of 

human health and fitness are now feasile. $ith the

availaility of this data, the systems provide etter

understanding of users) health status and their

relationship to the world around them &*(. Self-

quantification systems and applications are of high

importance to educate users to e more aware of their

health status, ut alone this cannot e consideredsufficient to change unhealthy haits in users. +hanging

unhealthy haits of users, demands to provide actionale

recommendations. #isting systems such as itit le# &(,

/awone 0p &1(, and 2isfit Shine &3( are some e#amples

of providing some asic recommendations ased on the

measured steps and slept hours. Samsung S !ealth &4(

and 5oogle it &6( are working as personal fitness coach

and health-tracking platform respectively on the asis of 

capturing user steps counting.

7reviously, we proposed an innovative digital health

framework called 2ining 2ind &', 8( for personalized

healthcare and wellness support. 9t provides personalized

recommendations on the asis of e#pert knowledge and

domain guidelines. The main limitation with e#isting

recommendation is that they are ased on user profile

information without taking into account user)s current

conte#t, preferences, and environmental variales. To get

personalized recommendation services, a granular level of user)s preference information is also required along with

conte#tual information and environmental variales. 9n this

paper, we proposed an innovative method to interpret the

user situation on three parameters: user conte#t ;low level

and high level<, physical conte#t ;weather<, and user

preferences.

. Met!odAs shown in ig. ', the high level architecture of the

proposed method depicts the overall functional workflow.

At first level, the data is collected from wearale sensors

and smart phone. The data is curated and persisted with alogical division of user profile data, conte#tual data, and

environmental data. The user profile data includes user

demographics ;gender, age< and physiological factors

;height and weight<. +onte#tual data is categorized into

two type of conte#ts: low level and high level conte#ts. =ow

level conte#t includes location, emotion, and physical

activities. !igh level conte#t includes the conte#t

determined from the low level conte#t, such as office-

works, commuting, eating, and others. nvironmental

variales include weather conte#t as rainy, sunny, etc.

>nowledge %uilder component reasons over user profile

data and physical activity data along with rules of 

knowledge ase to initial level of recommendations. or an

instance, let the system generate recommendation as? @an

hour e#ercise is required to urn # amount of calories.

8/17/2019 Cross-Context Multi-level Interpretations to Generate Personalized Recommendations

http://slidepdf.com/reader/full/cross-context-multi-level-interpretations-to-generate-personalized-recommendations 2/3

This is a generic level recommendation statement that can

e interpreted differently, such as? walking m' minutes or

running m* minutes or hiking m1 minutes or others. rom

the guidelines, a set of possile recommendations can e

inferred for an hour e#ercise to urn # amount of calories.

As per the guidelines, all the aove recommendations are

correct, however, a suset of them is applied to the user

according to hisBher current situation. Also, it is possile

that the recommendation uilder generates one

recommendation, e.g. @hiking # minutes, ut hiking

recommendation may not e permitted ecause of the

weather condition is rainy. 5enerate the personalized

recommendations we need to interpret the

recommendation on the asis of user situations.

Cecommendation 9nterpreter in ig. ', gets the initial

recommendation uilt y recommendation uilder

component and interpret with the strategy discussed in

following section.

2.1. Design Strategy of Multi-level Interpretations of recommendations And Results$e devised a matri# representation for each conte#t and

environmental variale to map against each type of 

recommendations such as running, walking, hiking,

 Dogging, and others. An e#ample of one of the low level

conte#ts location is shown in Tale '.

Figure 1: High level architecture of proposed approach of 

generating personalized recommendations on the basis of userprofile, context, and environmental data.

Similar matrices are created for all possile conte#ts and

environmental variales. "n the asis of current instance

of user conte#t and environmental variale;s<, we generate

a conte#tual matri#. 9n the conte#tual matri#, the current

instance of conte#t B environmental variale is crossed

with generated recommendation. Suppose current

instance has the values as loc E home, high level conte#t

;!+=< E amusement, emotion E happy, and weather E

rainy and the recommendation generated y recommendation uilder is hiking. 7rior to interpretations,

we first look at the column of recommended activity and

draw the interpretations on the asis of AFG operation. or

instance, one of the possile interpretations is? the

recommendation in the current situation is not applicale

and we need to find alternative recommendations. $e

apply the methods y selecting 3 potential users and

validate the results for different situations.

Table 1: Location context values mapping with differentrecommendations 1 shows applicable, ! shows not applicable"

Loc#$ec Hi%ing &ogging 'itting 'tretching ((

)*m ! 1 1 1 ((

Home ! 1 1 1 ((

+ffice ! ! 1 1 ((

((( (( (( (( (( ((

Tale * shows an e#ample of conte#tual matri# and

aggregate vector creation. +onte#tual matri# is composed

of current conte#ts of the user. All the current conte#ts

e.g. locations, !=+, weather, emotion, etc. are evaluatedin terms of recommended activities. =ogical AFG-operation

is performed on each of the physical activity ;column-wise

manner< to generate aggregate vector. This aggregate

vector contains a list of physical activities which can e

performed in the given conte#t of the user.Table : 'ample contextual matrix and aggregate vector 1 showsapplicable, ! shows not applicable"

-ontext#$ec

Hi%ing 

&ogging 'itting 'tretching ((

Loc Home ! 1 1 1 (

HL-/nactivit* 

1 1 ! 1 (

 0eather$ain* 

! 1 1 1 (

((( (( (( (( (( (

 Aggregate

 Vector

! 1 ! 1

". Conclusion and #uture $or% The coming era in health and wellness services is all

aout inducing healthy haits in users. $e presented our

work on situation-aware recommendation interpretation onthe asis of user situation including conte#t,

environmental variale, and user preferences. $e foresee

enormous application of the proposed approach in health

and wellness applications. The current systems and

applications can enhance their capailities to employ our

interpretation service.

Acknowledgement. This research was supported y the

2S97;2inistry of Science, 9+THuture 7lanning<, >orea,

under the 9TC+;9nformation Technology Cesearch +enter<

support program supervised y the F97A;Fational 9T

9ndustry 7romotion Agency<I ;F97A-*J'1-;!JJ'-'1-'JJ<

Re&erences12.3anos +, 3ilal(4min 5, 4li(6han 0, 4fzel 5, 4li T, 6ang 

8/17/2019 Cross-Context Multi-level Interpretations to Generate Personalized Recommendations

http://slidepdf.com/reader/full/cross-context-multi-level-interpretations-to-generate-personalized-recommendations 3/3

3H, Lee '. 5ining 5inds: 4n innovative framewor% forpersonalized health and wellness support. /n:7roceedings of the /nternational -onference on7ervasive -omputing Technologies for Healthcare,/stanbul, Tur%e*8 !19".

2.4lmal%i 5, 5artin('anchez F, )ra* 6. 'elf(uantification: the informatics of personal datamanagement for health and fitness. !1;".

;2.<Fitbit Flex,= http:##www.fitbit.com#flex, accessed:!19(1!(9. !19".

>2.<&awbone ?p,= https:##@awbone.com#up, accessed:!19(1!(A. !19".

92.<5isfit 'hine,=

http:##www.misfitwearables.com#products#shine,accessed: !19(1!(B. !19".

B2.C' Health, C http:##shealth.samsung.com#, accessed:!19(11(!1. !19".

D2.C)oogle Fit, C https:##fit.google.com#, accessed: !19(

11(!. !19".A2. 3anos +, 4min 53, 6han 04, 4fzel 5, 4hmad 5, 4li 5,

 4li T, 4li $, 3ilal 5, Han 5. 4n /nnovative 7latform for7erson(-entric Health and 0ellness 'upport. /n:3ioinformatics and 3iomedical Engineering: 'pringer8p.1;1(1>!, !19".

top related