User-sentiment based Evaluation for Market Fitness Trackers

Evaluation of Fitbit One, Jawbone Up and Nike+ Fuelband based on Amazon.com

Customer Reviews

Hassan Issa, Alaa Shafaee, Stefan Agne, Stephan Baumann and Andreas Dengel

German Research Center for Artiﬁcial Intelligence (DFKI), Kaiserslautern, Germany

Keywords:

Activity Monitoring, Fitness Applications, Life Logging, Quantiﬁed Self, Sentiment Analysis, Fitness

Trackers Evaluation, Fitness, Fitbit One, Jawbone Up, Nike+ Fuelband.

Abstract:

Wearable ﬁtness and health trackers have been in an accelerated growth since their recent introduction to

consumer markets. Given the growth potential of this market sector, much more devices, with unbounded

set of features, are being introduced to the consumer at a fast pace. This makes the task of evaluating ﬁtness

trackers extremely challenging knowing that the results of an evaluation will quickly become obsolete. In this

paper, a user-sentiment based evaluation of ﬁtness trackers is demonstrated on market leading ﬁtness trackers.

The used approach relies on the crowd, expressed by Amazon.com product reviews, to present an aspect-based

evaluation of any market ﬁtness tracker. Utilizing the crowd knowledge acquired, a personalized recommender

system for ﬁtness trackers is also presented.

1 INTRODUCTION

Ever since the recent introduction of wearable com-

puting devices, the market size for these devices has

been rapidly growing and expected to attain a market

size of $30 billion as forecasted by (BCCResearch,

2014). Accordingly, ﬁtness and health trackers, which

correspond to a big part of this growth, are expecting

a 96 million units shipment in 2018. Fitness trackers

are always-on devices that provide basic pedometer

functionalities including estimating steps taken, dis-

tance traveled and calories burned. In addition, these

devices connect and sync their data easily to smart-

phone apps and web portals and enable social sharing

and friendly competitions. Fitness and health track-

ers are constantly incorporating new sensors and al-

gorithms to provide further health and ﬁtness insights

and functionalities for users. Samsung Simband 2

notably collects real-time biometric data including

heart rate, blood ﬂow and pressure, skin temperature,

CO2 and oxygen levels as well as EKG

levels. Like-

wise, machine learning techniques are employed by

some trackers like AMIIGO

ﬁtness tracker which is

http://www.samsung.com/us/globalinnovation/

Electrocardiogram; spelled with a ’K’ because in German

it is spelled Elektrokardiogramm

https://amiigo.com/

able to detect over 100 exercises along with the num-

ber of sets and repetitions and the ability to learn new

exercises by the user.

The increasing public interest in ﬁtness trackers as

consumer products, in addition to their potential uses

in health care necessitate the presence of proper eval-

uations for all the aspects of these devices. The huge

number of devices available on market and the speed

in which new devices with new features are released,

makes it very difﬁcult to come up with a proper eval-

uation for market products that won’t quickly turn ob-

solete especially when taking longer device usage-

time as a factor in evaluation. In (Shafaee et al.,

2014), we introduced a new framework for evaluat-

ing ﬁtness trackers based on the public opinion ex-

pressed through Amazon.com product reviews. This

framework follows an aspect-based sentiment analy-

sis approach that dynamically identiﬁes aspects of ev-

ery ﬁtness tracker and evaluates all sentiment-bearing

mentions of an aspect in all Amazon.com reviews of

the tracker.

In this paper, we present an extensive evaluation of

the top three market devices from manufacturers that

together dominated over 97% of the ﬁtness trackers

market in 2013 according to NPD Group

, namely;

https://www.npd.com/latest-reports/consumer-

technology-reports/

171

Issa H., Shafaee A., Agne S., Baumann S. and Dengel A..

User-sentiment based Evaluation for Market Fitness Trackers - Evaluation of Fitbit One, Jawbone Up and Nike+ Fuelband based on Amazon.com

Customer Reviews.

DOI: 10.5220/0005447401710179

In Proceedings of the 1st International Conference on Information and Communication Technologies for Ageing Well and e-Health (ICT4AgeingWell-

2015), pages 171-179

ISBN: 978-989-758-102-1

 2015 SCITEPRESS (Science and Technology Publications, Lda.)

Fitbit One

, Jawbone Up

and Nike+ Fuelband

. It is

worth mentioning that the choice of these three prod-

ucts is strictly for their leading position in the con-

sumer market and that there are no limitations by our

framework to include any other ﬁtness/health tracker

available for sale on Amazon.com. Not only an an-

swer to the general question ”which is the best ﬁt-

ness tracker?” is given in this work, but also to the

more personal question ”which ﬁtness trackers suits

me best?” is given in Section 5.4. A personalized

recommender system for suggesting the best ﬁtness

tracker for each individual based on her needs and

preferences is thus presented.

The rest of this paper is organized as follows. Sec-

tion 2 discusses the related work for this paper. Sec-

tion 3 introduces the devices and the dataset used for

the evaluation. An overview of the back-end senti-

ment analysis framework used as basis for the pre-

sented sentiment summarizer is given in Section 4.

Results and analysis of the Evaluation are presented

in Section 5, and ﬁnally the conclusions and future

work are discussed in Section 6

2 RELATED WORK

The sentiment analysis framework introduced in Sec-

tion 4 is further detailed and evaluated in (Shafaee

et al., 2014). The system components are proven to

top state-of-the art alternatives. Studies that address

the accuracy of ﬁtness trackers are relatively few. A

study (Dannecker KL, 2014) aimed at evaluating the

accuracy of energy expenditure estimation of con-

sumer physical activity monitors found out that Fitbit

sensors tend to signiﬁcantly underestimate energy ex-

penditure. In (Guo et al., 2013), an extensive physical

testing for several ﬁtness trackers and pedometers in-

cluding Nike+ Fuelband and Fitbit One is presented.

The study shows that Fitbit One is the most accurate

in counting steps, which veriﬁes the results presented

in Section 5.2.1.

3 DEVICES AND DATASET

3.1 Devices

In this section a brief introduction to the three devices

used in evaluation is given. The devices considered

are shown in Figure 1.

http://www.ﬁtbit.com/one

https://jawbone.com/up

http://www.nike.com/us/en us/c/nikeplus-fuelband

Figure 1: From left to right: Fitbit One, Jawbone Up, Nike+

Fuelband.

3.1.1 Fitbit One

Fitbit one is a clip-on tracker that was announced on

September 17, 2012. It features a digital display and it

is the ﬁrst tracker to sync through Bluetooth 4.0. Fit-

bit One supports Android, iOS and Windows Phone

smartphone operating systems. It can record several

daily activities including number of steps taken, dis-

tance travelled on foot, number of ﬂoors climbed,

calories burned, vigorously active minutes, as well as

quality of sleep. It also incorporates a ”silent alarm”

that wakes the user up through gentle vibrations. The

launch price for Fitbit One was 99.99 U.S. dollar.

3.1.2 Jawbone Up

Jawbone up was initially announced in November

2011. Due to manufacturing problems, the product

was relaunched in November 2012. The Up comes in

the form of a waterproof wristband. It can track num-

ber of steps taken, distance travelled on foot, calo-

ries burned as well as quality of sleep. It incorporates

a vibration alarm feature and communicates to many

3rd party lifestyle and ﬁtness apps and services. Jaw-

bone Up supports Android and iOS through a native

smartphone app. The launch price for Jawbone Up

was 129.99 U.S. dollar.

3.1.3 Nike+ Fuelband

Nike+ Fuelband is a wristband tracker that was an-

nounced on January 19, 2012. Fuelband uses a set of

LEDs as a screen. It tracks the number of steps taken

and calories burned, it displays time and computes

a proprietary measure of ﬁtness activity, NikeFuel.

Users set their daily goal in terms of NikeFuel, and

the band displays the progress of the user in achiev-

ing their goal through an array of colored LEDs. The

Fuelband syncs wirelessly to smartphones and both

Android and iOS are supported through a native app.

The launch price for Nike+ Fuelband was 149.99 U.S.

dollar.

ICT4AgeingWell2015-InternationalConferenceonInformationandCommunicationTechnologiesforAgeingWelland

e-Health

172

3.2 Dataset

For the evaluation, a total of 3, 241 Amazon.com re-

views of the three ﬁtness trackers Fitbit One (FO),

Jawbone Up (JU) and Nike+ Fuelband (NF) are

used. The distribution of the reviews (REV), sen-

tences (SEN), aspect mentions (MEN) and sentiment-

bearing aspect mentions (SBM) among the three de-

vices is shown in Table 1. The signiﬁcantly lower

number of reviews available for Nike+ Fuelband does

not necessarily reﬂect its market share with respect

to the two other trackers. It is probably due to sales

through Nike’s channels rather than Amazon.com.

Table 1: Details about the dataset used for evaluating the

devices.

Device REV SEN MEN SBM

FO 2,504 15,154 7,682 6,372

JU 696 4,397 2,138 1,645

NF 41 246 818 612

Total 3,241 19,797 10,638 8,629

4 OVERVIEW OF THE

SENTIMENT ANALYSIS

SYSTEM USED

Figure 2: System Overview.

A general overview listing the main components of

the sentiment analysis system used for the evalua-

tions is shown in Figure 2. For each product, its

name or Amazon ID(s) are fed into the system. Since

Amazon no longer offers reviews through its API, a

script is utilized to extract reviews using regular ex-

pressions. Any noise caused by mismatched HTML

tags is removed and the reviews are then inserted into

a database. The Stanford CoreNLP (Manning et al.,

2014) is then used to extract sentences in each review,

tokenize them and get the part-of-speech and base

form for each token. The output of the previous step is

then fed into three other components, namely, product

name extractor, aspect extractor and sentiment extrac-

tor which are deﬁned below.

• Product-Name Extractor: The product name ex-

tractor has two main aims. First, it automatically

extracts the names of competing products. Sec-

ond, it identiﬁes mentions of the given product or

competing products in text. For example, if an

author says: ”I love my Fitbit One. I got it after I

broke my Ultra 2 months ago.”, the system identi-

ﬁes that the ﬁrst sentence discusses ”Fitbit One”,

whereas the latter discusses ”Fitbit Ultra”. This

component is essential to assign the sentiment to

the proper product and is useful in the other com-

ponents.

• Aspect Extractor: This component extracts the

aspects of the product that reviewers commented

on. For ﬁtness trackers, the aspects can be price,

customer service, steps counter, mobile apps, etc.

Aspects are classiﬁed into static aspects and dy-

namic aspects. Static Aspects are features that

are common to all products such as price and cus-

tomer service while dynamic aspects are features

that may differ from a product to another. For ﬁt-

ness trackers, heart rate monitoring is an example

of a dynamic aspect since not all ﬁtness trackers

monitor heart rate.

• Grouping Aspects into Categories: After iden-

tifying the aspects of the product, sentences dis-

cussing each aspect are extracted. Given that an

aspect may be discussed by different people using

different terms, all the different mentions of the

same aspect are grouped. To illustrate, consider

the direct and indirect mention of the ”price” as-

pect in these two sentences: ”Nike+ Fuelband is

quite expensive.” and ”The Fuelband has a high

price tag!”

• Sentiment Extractor: The sentiment extractor

component aims at detecting the opinion of a

whole sentence or part of it. That is, given a

phrase, it returns whether it is positive, negative or

neutral. For instance, ”I love this device” is posi-

tive, ”This device is really useless” is negative and

”I bought this device yesterday” is neutral.

The next step is to assign the sentiment to the

proper aspect. Finally, the data collected from the

User-sentimentbasedEvaluationforMarketFitnessTrackers-EvaluationofFitbitOne,JawboneUpandNike+Fuelband

basedonAmazon.comCustomerReviews

173

Table 2: Fitbit One Results.

Aspect Positive Negative Neutral Total

Steps Counter 605 183 492 1280

Sleep Assessment 570 199 373 1142

Syncing 408 146 263 817

Calories Burned 315 113 245 673

Floors Climbed 276 99 267 642

App Functionality 332 114 182 628

Website 332 85 176 593

Battery 206 100 201 507

Price 186 102 124 412

Pocket/Clipping 109 67 112 288

Customer Service 148 50 60 258

Distance Traveled 90 23 88 201

Badges 71 12 32 115

MyFitnessPal 38 9 21 68

Activity Indicator (Flower) 36 8 14 58

Total 3722 1310 2650 7682

sentiment analysis of the reviews for each product is

presented through an interactive web-based summa-

rizer that provides the user with statistics and insights

about the products. Further details on the components

presented are discussed in (Shafaee et al., 2014).

5 EVALUATION RESULTS AND

ANALYSIS

In this section, the results of the sentiment analysis

process introduced in Section 4 applied to the three

devices chosen for this study are presented. Table 2,

Table 3 and Table 4 show the aspects detected by the

system and the distribution of the sentiments among

these aspects for Fitbit One, Jawbone Up and Nike+

Fuelband respectively. Before discussing the out-

put of the evaluation, a precision/recall analysis of

the system is ﬁrst presented to estimate the credibil-

ity of the evaluations following in this section. For

this sake, human annotators were asked to annotate

all sentiments in 106 Amazon reviews distributed al-

most equally among the three products. The results

for our system were compared to two state-of the-art

supervised methods, namely Support Vector Machine

(SVM) trained on unigrams and Naive Bayes classi-

ﬁers trained on unigrams(Pang et al., 2002). Figure 3

shows that our system has the highest recall with very

high precision for the positive sentiment class. Sim-

ilarly, Figure 4 shows our system has signiﬁcantly

higher precision and a fairly good recall when com-

pared to the other approaches for the negative senti-

ment class. A big advantage of the proposed system

is that it is lexicon-based which makes it possible to

generalize to different domains. This is not the case

with the supervised approaches which require a large

set of annotated data and can therefore poorly gener-

alize.

Figure 3: Recall and precision of sentiment detection for

the positive sentiment class.

5.1 Controversial Aspects

Regardless of the sentiment, the frequency of as-

pect mentions in reviews is very signiﬁcant, espe-

cially for evaluating distinguished aspects that set a

product apart from its market competitors. For ex-

ample, the NikeFuel activity measure introduced in

the Nike+ Fuelband is the most discussed aspect for

this device with 22.7% of the Fuelband’s aspect men-

tions discussing this unique aspect. In contrast, the

ﬂower activity indicator in the Fitbit One, which re-

sembles NikeFuel for the Fuelband, takes a share of

0.7% only from the Fitbit’s aspect mentions and it is

ICT4AgeingWell2015-InternationalConferenceonInformationandCommunicationTechnologiesforAgeingWelland

e-Health

174

Figure 4: Recall and precision of sentiment detection for

the negative sentiment class.

the least discussed aspect for the device. This indi-

cates that Nike’s distinguished activity measure has

successfully captured the interest of consumers where

as Fitbit falls short in this regard. The Power Nap

aspect of Jawbone Up, which is a feature that calcu-

lates the optimal nap duration for the user and wakes

them up through gentle vibrations, is barely discussed

with 0.7% of aspect mentions despite being an inno-

vative and positively received feature. This signiﬁes

that the usability of this feature can be questioned.

However, Sleep assessment, in general, has proven to

be an important feature to users grabbing over 20% of

Jawbone’s aspect mentions and about 15% of Fitbit’s

aspect mentions. This indicates that Nike is miss-

ing a highly demanded feature by not including it in

their device. Also, Fitbit’s ability to compute ﬂoors

climbed holds over 8% of its aspect mentions almost

equal to the share of ”Calories Burned” feature and

thus it also represents a feature demanded by con-

sumers.

5.2 Static Aspects Evaluation

Static aspects describe essentially the basic features

and capabilities of a ﬁtness tracker. In this section,

we assess the public satisfaction of the performance-

related aspects in all three devices. Namely: Steps

Counter, Distance Traveled and Calories Burned. We

point here, that similar insight and ﬁgures can be ex-

ported by the sentiment summarizer for all the other

aspects. However, due to space restrictions and to

avoid repetitions, detailed comparison for the other

aspects are omitted. Still, useful insights for each de-

vice are given in Section 5.3.

5.2.1 Steps Counter Evaluation

All three devices seem to be satisfactory to consumers

with respect to steps counting as shown in Figure 5

despite a signiﬁcant advantage for the Fitbit One.

This actually supports the physical testing of accu-

racy of several ﬁtness trackers including Fitbit One

and Nike+ Fuelband by (Guo et al., 2013) in which

the One outperformed all other devices.

Figure 5: Results of the public sentiment evaluation for as-

pect ”Steps Counter” as displayed by the sentiment summa-

rizer.

5.2.2 Distance Travelled Evaluation

The good performance in counting the steps by the

devices leads to a good estimation of the distance

travelled by users as shown in Figure 6. Since all

three devices do not include location sensors, the dis-

tance value is probably calculated through estimations

based on the individual’s personal data provided to the

device during the set up process.

Figure 6: Results of the public sentiment evaluation for

aspect ”Distance Travelled” as displayed by the sentiment

summarizer.

5.2.3 Calories Burned Evaluation

In a similar manner, the devices estimate calories

burned based on the performance of the user and

his/her entered physical data (e.g. height, age, weight,

User-sentimentbasedEvaluationforMarketFitnessTrackers-EvaluationofFitbitOne,JawboneUpandNike+Fuelband

basedonAmazon.comCustomerReviews

175

Table 3: Jawbone Up Results.

Aspect Positive Negative Neutral Total

Sleep Assessment 225 82 128 435

Battery 113 113 123 349

App Functionality 159 57 95 311

Steps Counter 90 30 97 217

Syncing 78 65 66 209

Price 63 55 30 148

Customer Service 51 31 45 127

Alarm 51 15 24 90

Calories Burned 38 13 27 78

Food Logging 28 18 15 61

Website 19 5 16 40

Distance Traveled 17 2 16 35

Powernap 6 2 8 16

Pocket 7 4 3 14

Activity Logging 5 1 2 8

Total 950 493 695 2138

Table 4: Nike+ Fuelband Results.

Aspect Positive Negative Neutral Total

NikeFuel 70 49 67 186

Price 41 29 38 108

Battery 35 21 41 97

Calories Burned 33 21 38 92

Steps Counter 29 22 31 82

App Functionality 38 15 22 75

Syncing 29 20 18 67

Website 19 10 23 52

Customer Service 10 13 9 32

Distance Traveled 8 6 13 27

Total 312 206 300 818

gender, etc.). Both Fitbit One and Jawbone Up seem

to be equally satisfactory for users estimating the

calories burned where as the Fuelband has a signif-

icantly less positive-to-negative mention ratio com-

pared to the two other devices as shown in Figure 7.

5.3 Further Discussion and Insights

In addition to the statistics provided, the sentiment

summarizer allows for easy tracking of positive, neg-

ative and neutral mentions of each aspect of every de-

vice through out the reviews. For example (Figure 8),

a user can request all the negative mentions of ’Cus-

tomer Service’ for Jawbone Up and the summarizer

will return a list of all sentences with these negative

mentions sorted by the degree of negativity estimated

by the system. This gives a new dimension in the eval-

uation by highlighting pros and cons based on the ex-

perience of customers. Many useful conclusions can

Figure 7: Results of the public sentiment evaluation for as-

pect ”Calories Burned” as displayed by the sentiment sum-

marizer.

be drawn from people’s opinion. Some of them are

given below.

ICT4AgeingWell2015-InternationalConferenceonInformationandCommunicationTechnologiesforAgeingWelland

e-Health

176

Figure 8: Snapshot of a part of the list of negative Customer

Service aspect mentions for Jawbone Up as shown in the

sentiment summarizer.

5.3.1 Fitbit One

Many proponents of Fitbit commented on how it

helped them greatly pay attention to their sleep qual-

ity and how it encouraged better sleeping habits. The

ratio between the number of users in favor of the sleep

assessment made by the tracker to those against it

is almost 3:1. Some customers however, found the

sleeping band to be uncomfortable or the sleeping

data to be inaccurate. One of them said: ”Tossing

in the middle of the night or pushing covers off your

body will register as waking up, and watching an hour

or two of television without moving your hand before

falling asleep will count as sleep time. Unfortunately,

this makes the sleep tracker a good idea that is not

very accurate.”

Some reviewers commented that it holds a good

charge (e.g., charging it only on weekends) and some

were pleased that it emails them a reminder when the

charge is low. Quoting a user ”The charging cord and

Bluetooth dongle are better executed than the stand

used in the previous version, and it’s MUCH easier

to travel with.”. However, some said that their Fitbits

did not hold a charge for more than 8-12 hours and

others experienced the total death of the device.

Many users ﬁnd setting up the Fitbit One and

syncing the data very easy and smooth. They also

love syncing the Fitbit with many third-party apps

especially to MyFitnessPal app. In fact, the system

identiﬁed MyFitnessPal as one of the aspects of Fitbit

because people comment a lot on syncing their Fitbits

to it. For the Fitbit One app itself, some people claim

that it is buggy for most Androids. Others complained

that it requires a dongle to sync to PC or MAC say-

ing that the dongle is very small and can be easily lost.

Some could not pair it to Android. A frequent traveler

said that the time shown after syncing is not accurate.

Although people are pleased that it is small and

light, many people lost it because, according to them,

it can easily fall out of the holder and the holder does

not attach ﬁrmly to clothes. Therefore, they highly

recommend having it inside the pocket rather than

clipped onto the pocket or belt. Despite some who

are motivated by the growth of the ”activity-progress

ﬂower”, many others believe that Fitbit should change

this visualization.

5.3.2 Jawbone Up

The most popular feature of Jawbone Up is sleep as-

sessment. Most of the customers who reviewed this

device expressed how they like its accompanying fea-

tures, namely the silent vibrating alarm, Power Nap

which wakes a user up at the optimal time, and the

Idle Alert which reminds users to move when they

have been inactive for too long. Nonetheless, many

people complained of having to change it manually

to sleep mode before sleeping and changing it back

to awake mode later, which one can forget. Some

customers also questioned the accuracy of the sleep

information the band provides.

A major problem many users faced is the complete

death of the band after a period of use. Nonetheless,

many users said that the customer service replaced

their bands. On the other hand, some described the

customer service as poor and one of the complaints

said that they could not exchange the item or help with

its problems because it was not bought directly from

the Jawbone company.

Many people recommend calibrating the band to

record the distance accurately. Some sentences anal-

ysed by the summarizer stated that wearing the Jaw-

bone on ankle is said to increase the accuracy of the

band, yet, some commented that it is uncomfortable.

Some customers said that they did not experience

any problems with Syncing. However, the Jawbone

UP does not sync wirelessly which made some com-

plain about having to take the band cap off and in-

serting the band in the headphone’s jack of the phone

every time. Last but not least, the Jawbone Up app

was generally positively rated by most consumers.

User-sentimentbasedEvaluationforMarketFitnessTrackers-EvaluationofFitbitOne,JawboneUpandNike+Fuelband

basedonAmazon.comCustomerReviews

177

5.3.3 Nike+ Fuelband

NikeFuel score is perhaps the most remarkable fea-

ture of the Fuelband being, by far, the most reviewed

aspect of the band as shown in Table 4. By read-

ing some of the sentences discussing this aspect, the

Fuel score seems very motivating for many. How-

ever, some are skeptical of how it is calculated. A

user said ”I’ve noticed that the Fuel really adds up

when running and walking but not as much with other

workouts that are way more intense than a jog on the

treadmill.”. Another said ”Of course, if you shake

your wrist while on the elliptical or jogging, you can

increase your points by 50%.”. Some people said that

it is very motivational, however, they did not recom-

mend it for people who want to accurately track their

steps. With respect to syncing, some people found

syncing over the phone or USB to be smooth and

easy. On the other hand, it failed to sync with some

and was described as slow. Some users complained

that the screws in the band rust. One suggested us-

ing a ﬁngernail polish to stop the rust. Many users

liked the accompanying app which was described as

motivational, easy-to-use and well-designed. In fact,

the number of positive reviews on the Nike+ Fuelband

app is more than twice the negative reviews on it. Yet,

some were disappointed that it did not come with an

Android app.

5.4 Personalized Fitness Tracker

Recommender

Given the vast knowledge the presented sentiment

summarizer acquired by analysing the sentiments in

Amazon.com reviews of ﬁtness trackers, the second

step is to utilize this knowledge to help each potential

customer choose the device that suits her needs best

based on her own preferences. Figure 9 shows the

user interface for the device recommender in the sen-

timent summarizer. A user can select using 5-value

sliders, how important an aspect is to her, and the rec-

ommender will dynamically upgrade score gauges for

each device thereby helping each customer making

the right purchase decision based on all the acquired

knowledge of owners of these devices.

The scores are computed using Formula 1; where

d denotes a device, k denotes the total number of as-

pects for which the user have assigned a non-zero

weight using the slider, a

denotes a single aspect

(i ∈ [1, k]), q

∈ [0, 4] denotes the weight assigned by

the user to aspect a

, pos(a

) and neg(a

) denote the

number of positive (res. negative) mentions of the as-

pect a

for device d, and A

denotes the set of all as-

pects for device d. Intuitively, when the user gives a

non-zero weight for a certain aspect using the slider,

this aspect is added to the set of user-requested as-

pects. For each aspect a

of these k aspects, the system

checks if it is included in the aspects of the considered

device. If yes, a value

pos(a

)−neg(a

)

pos(a

)+neg(a

)

∈ (−1, 1) is re-

turned indicating how recommended (res. unrecom-

mended) the device is when considering this aspect

alone. Otherwise, if the aspect requested by user is

not available for the device, −1 is returned indicating

that the device is not recommended at all when con-

sidering this aspect. The result is adjusted for each

aspect based on a user-selected weight q

. The ﬁ-

nal score Score

is then computed as the normalized

mean for the returned weighted values for each as-

pect.

Figure 9: Snapshot of the personalized device recommender

functionality in the sentiment summarizer.

Score

100

∑

i=1

(

pos(a

)−neg(a

)

pos(a

)+neg(a

)

i f a

∈ A

−1 otherwise

(1)

6 CONCLUSIONS AND FUTURE

WORK

In this paper, the novel ﬁtness-trackers’ evaluation ap-

proach presented in (Shafaee et al., 2014) is demon-

strated to evaluate three market leading ﬁtness track-

ers. The approach relies on aspect-based sentiment

analysis of Amazon.com reviews of these products.

The results of the evaluation are presented to the user

in an easy-to-use web-based summarizer. Statistics

and insights resulting from the evaluation are pre-

sented in Section 5 where also a personalized ﬁtness

tracker recommender system is presented. In addi-

tion to the ongoing research on enhancing the un-

derlying algorithms used in the back-end sentiment

analysis system introduced in Section 4, our efforts

are also oriented on automating the whole process

starting from fetching the Amazon.com reviews until

reaching an always up-to-date summarizer that adds

ICT4AgeingWell2015-InternationalConferenceonInformationandCommunicationTechnologiesforAgeingWelland

e-Health

178

even new measures to its insights, including sentiment

variation over time, and per different countries.

ACKNOWLEDGEMENTS

This research is supported by the Ministry for Educa-

tion, Sciences, Further Education and Culture of the

State of Rhineland-Palatinate, Germany (MBWWK)

and is part of the project MyCustomer.

REFERENCES

BCCResearch (2014). Wearable computing: Technologies,

applications and global markets. Technical report,

BCC Research.

Dannecker KL, Sazonova NA, M. E. S. E. B. R. (2014). A

comparison of energy expenditure estimation of sev-

eral physical activity monitors. Medicine and science

in sports and exercise.

Guo, F., Li, Y., Kankanhalli, M. S., and Brown, M. S.

(2013). An evaluation of wearable activity monitoring

devices. In Proceedings of the 1st ACM International

Workshop on Personal Data Meets Distributed Multi-

media, PDM ’13, pages 31–34, New York, NY, USA.

ACM.

Manning, C. D., Surdeanu, M., Bauer, J., Finkel, J.,

Bethard, S. J., and McClosky, D. (2014). The Stan-

ford CoreNLP natural language processing toolkit. In

Proceedings of 52nd Annual Meeting of the Associa-

tion for Computational Linguistics: System Demon-

strations, pages 55–60.

Pang, B., Lee, L., and Vaithyanathan, S. (2002). Thumbs

up?: Sentiment classiﬁcation using machine learning

techniques. In Proceedings of the ACL-02 Conference

on Empirical Methods in Natural Language Process-

ing - Volume 10, EMNLP ’02, pages 79–86, Strouds-

burg, PA, USA. Association for Computational Lin-

guistics.

Shafaee, A., Issa, H., Agne, S., Baumann, S., and Dengel,

A. (”2014”). Aspect-based sentiment analysis of ama-

zon reviews for ﬁtness tracking devices. In PAKDD

Workshops, pages ”50–61”.

User-sentimentbasedEvaluationforMarketFitnessTrackers-EvaluationofFitbitOne,JawboneUpandNike+Fuelband

basedonAmazon.comCustomerReviews

179