CoMoRea'18 - 14th Workshop on Context and Activity Modeling and Recognition

Purpose-of-Visit-Driven Semantic Similarity

Analysis on Semantic Trajectories for Enhancing
The Future Location Prediction
Antonios Karatzoglou Dominik Koehler Michael Beigl
Robert Bosch GmbH Karlsruhe Institute of Technology Karlsruhe Institute of Technology
Corporate Sector Research and Germany Pervasive Computing Systems
Advance Engineering, and Email: [email protected] Germany
Karlsruhe Institute of Technology Email: [email protected]
Germany
Email: [email protected],
[email protected]

Abstract—The number of people that are using or are even conducting an ongoing online survey1 . Based on their results
dependent on Location Based Services (LBS) is growing rapidly so far, 62% of the participants feel that an A.I. should fulfil
every year. In order to offer timely and user-tailored services, their needs before even asking and 74% state that an A.I.
providers rely increasingly on forward-looking algorithms. For
this reason, location prediction plays a key role in LBS. Recent should prevent them from making mistakes, thus confirming
approaches in location prediction leverage semantics in order the significance of anticipatory behaviour. Consequently, the
to overcome drawbacks that characterise conventional non- significance and value of location prediction with respect to
semantic systems. However, when it comes to modelling locations, the domain of location-based services and of context-aware
the majority of them constrain themselves to static semantical systems in general is self-explanatory.
constructs and hierarchies, without taking the current situation,
and most importantly, the users’ varying personal perception into Location information reveals to us humans more than just
account. In this work, we introduce a novel dynamic approach the whereabouts. It gives indirectly insight about the what
that aims at taking the variation of the users’ perception explicitly and the when. For instance the location “night club” is put
into consideration when describing locations, in order to elevate usually in context with some overall, high semantic level
the overall prediction performance. For this purpose, we consider
purposes, such as “socializing” and “having fun”, and a set
explicitly time and purpose of visit by building so called Purpose-
of-Visit-Dependent Frames (PoVDF). Our framework is hybrid of elementary, lower semantic level activities, like “drinking”,
and combines both a data-driven, as well as a knowledge-driven “dancing”, and “meeting friends”. Moreover, a human would
model. To fuse these two models, we define a Purpose-of-Visit- additionally associate some corresponding temporal informa-
Driven Semantic Similarity (PoVDSS) metric and use it as a fusing tion, such as “night”, “weekend”, or “once a week” to it. In
component between the two models. We conducted a user study
order for us humans to be capable of interpreting locations at
to evaluate our approach on a real data set and compared it with
two state of the art semantic and non-semantic algorithms. Our such a high level and associating them with all this additional
evaluation shows that our approach yields a location prediction information, we rely both on a broad framework of semantics
accuracy of up to 80%. hidden behind them, as well as on a large portion of world
and common sense knowledge. At the same time, each human
I. I NTRODUCTION takes his/her own personal experience and knowledge into
consideration.
According to recent statistics published by eMarketer [1], A generic semantic framework together with a common
242 million people in the USA are expected to use location- sense knowledge base provide a mutual basis among different
based services (LBS) in 2018, almost twice the number of people for interpreting things similarly. In comparison, per-
LBS users in 2013 (123 millions). While typical context-aware sonal experience can rather lead to different interpretations
applications, such as LBS, are focusing on reacting to users’ among people. Let us clarify this in the location scenario by
location-dependent needs, current approaches strive to behave going back to the example mentioned before. In the particular
in a forward-looking manner in order to raise the quality of example, the location “night club” was interpreted from the
their service. This Proactivity is a key feature, which leads to perspective of a guest, which is the most common one. But
a great number of benefits including higher efficiency and an for the barkeeper, who may probably have to open the place at
improved human-machine-interaction. Space10, the innovation noon in order for the beverage suppliers to replace the empty
lab of IKEA, is currently attempting to determine the way
Artificial Intelligence (A.I.) should interfere in our lives by 1 http://www.doyouspeakhuman.com/

Authorized licensed use limited to: Silpakorn University provided by UniNet. Downloaded on February 15,2024 at 03:52:41 UTC from IEEE Xplore. Restrictions apply.
978-1-5386-3227-7/18/$31.00 ©2018 IEEE 100
 CoMoRea'18 - 14th Workshop on Context and Activity Modeling and Recognition

bottles and carries the responsibility that everything works on the other hand indirectly the compliance with the data
alright during the night, the club is a working location, bound protection regulations. Finally and not unimportant in the age
to a completely different set of high and low level activities. A of the Internet of Things (IoT), semantics and particularly the
similar ambiguous effect would arise in the case of a restaurant associated knowledge graphs and ontologies, which are meant
between a guest and the cook or the waiter working there. as tools for sharing knowledge in first place, provide a robust
Let us now consider another example. People, visiting a basis for machine-to-machine (M2M) communication. This is
conference or having a business lunch at a hotel, would vital for creating independent, fully autonomous and intelligent
experience the hotel during that time from a perspective similar environments.
to that of the working receptionist, since their visit are both However, to the best of our knowledge, none of the se-
of professional nature. However, if the same people would mantic trajectory based approaches up to this point have been
enjoy a drink at the bar of the same hotel after their business taking the varying human perception of locations into account.
is over, they would rather associate the hotel more with a Instead, they constrain themselves to static semantic location
night life location, like the “night club” mentioned before. types and inflexible associations between locations and users
This example highlights another important issue, namely that as we will see in the related work chapter (section II). In
people tend to perceive, interpret and associate locations to this paper, we introduce a semantic trajectory based location
each other dynamically, depending on the situation, in which prediction approach that considers explicitly the dynamic,
they find themselves. Nathan et al. underpins this idea by purpose-of-visit-driven varying interpretation and clustering
interpreting movement between locations as the outcome of of locations in order to achieve a higher performance. Our
the synergy of four components [2]: the individual’s internal approach relies on the hypothesis that locations resemble one
state, its motion capacity, its navigation capacity and a group another from the point of view of the user in relation to the
of potential external factors. purpose of visit and that similar location show also similar
Location prediction algorithms that utilize semantics and transitions as well. For this purpose, we combine two different
rely on so called semantic trajectories go beyond plain nu- modeling techniques, a data-driven and a knowledge-driven
merical data, like GPS tracks and Cell-ID sequences. Fig. 1 one, by using semantic similarity as a fusing component. We
illustrates an example of such a semantic trajectory. The use designed and carried out a user study, in order to investigate
up to what degree users associate locations to more than
one purpose of visit and thus derive an indication about the
semantic ambiguity of locations in a real world scenario.
Furthermore, we used the same collected real data set to
evaluate our approach. We can show that our framework is
able to converge more towards human movement patterns and
can lead to a higher performance compared to other semantic
trajectory based approaches.
This paper is structured as follows. Section II provides
a brief summary of the most relevant semantic-enhanced
location prediction approaches. Next, section III describes in
Fig. 1: Example of an 1-day long Semantic Trajectory. (Image based detail our approach, while in IV we both give insight into our
on: https:// commons.wikimedia.org/ wiki/ File:San Francisco downtown.jpg) user study and the resulted data, as well as discuss thoroughly
our outcomes and performance of our framework. At last, in
of semantics provides a number of advantages. Spaccapietra et section V, we provide a short overview about our work and a
al. highlighted in [3] how important it is to utilize semantics summarization of our major results.
when analyzing trajectories, while at the same time pointing
out the variation of the underlying purpose and semantical II. R ELATED W ORK
meaning of them. The most significant advantage, is the fact Our focus in this paper lies on the semantic-enhanced
that semantic trajectories are capable of capturing the essence location prediction. However, in this section, we will go very
of human movement patterns. This can be particularly helpful briefly through some non-semantic, but well-known work on
for a location predictor in places that have not been visited location prediction as well.
before by the users and for which there are therefore no Ashbrook et al. introduce in [4] a method for learning
GPS recordings available on which the predictor could be semantically significant locations out from the users’ GPS
trained. Beyond that, semantic enhanced location prediction trajectories, like home, work and others, on the top of which
systems gain transparency through the use of semantics. Due a next place predictor is proposed using a simple 1st-degree
to the fact that the data collected and processed by the Markov Chain model. Focusing solely on next place predic-
respective systems are human-understandable, the user has the tion, Gambs et al. [5] use higher degree Markov Chains,
opportunity to better understand how such systems work and whereby the last n places are used to make a prediction about
why certain predictions come into effect. This leads on the the next one. Gao et al. [6] extends the Markov Chains by
one hand to a better human-machine relationship and assists adding temporal context.

Authorized licensed use limited to: Silpakorn University provided by UniNet. Downloaded on February 15,2024 at 03:52:41 UTC from IEEE Xplore. Restrictions apply.
101
 CoMoRea'18 - 14th Workshop on Context and Activity Modeling and Recognition

In contrast, Ying et al. [7] use their own Geographic Seman- Context & User Data

tic Information Database (GSID) to enrich semantically their

Raw Data/Trajectory Preprocessing
recorded (GPS- or Cell-ID-)trajectories. From the resulting
Data Cleaning Stay Location Extraction Consistency Checking
semantic trajectories, patterns are extracted, which in turn are
converted into Semantic Pattern Trees to finally provide the
Semantic Data/Trajectory Enrichment
next place prediction. Samaan et al. use spatial conceptual
Spatial Information & Location Type Purpose-of-Visit (PoV)/Activity Temporal Information
maps to describe buildings and road network elements also
semantically [8]. In addition, a user context knowledge base
containing the users’ preferences among other is used to Model Generation

support the location prediction. Ridhawi et al. apply a similar Ontology N-dim Markov Chains

algorithm for tracking and predicting users indoors in order to Propagation Training

support location aware services in [9]. In their work, other than

Optimization
Samaan et al., the semantic information is stored in an onto- PoV-Dependent Frame Generation PoV-Semantic Similarity Analysis
logical model. Long et al. propose in [10] a location clustering
algorithm, which makes use of Latent Dirichlet Allocation, a Fig. 2: Layer diagram of the Purpose-of-Visit-Driven Semantic Sim-
probabilistic model used normally to cluster documents based ilarity Analysis based location prediction framework (PoVDSSA).
on the topics contained in them. They define so called geo-
graphic topics based on popularity in an unsupervised manner
from the check-ins and corresponding venues of Foursquare The particular annotation occurs in a semi-supervised manner,
users. These topics replace static location categories like the partly by the user (through an Android app running on her
one provided by Foursquare. Additionally, they investigate smartphone (see section IV)), and partly by using the Google
how location clusters behave depending on whether it is a Places API. In this way, both public as well as private
weekday or weekend. In [11], Krishnamurthy et al. exploit locations, like the users’ home or work, can be correctly
Twitter to predict the location of its users. They introduce identified among the recorded data. The resulting data are then
the concept of localness to express how “close” certain terms used to propagate our Ontology-based Knowledge Base (OKB)
appearing in a tweet are to particular cities. To this effect, described in section III-A and to build our so called Purpose-
they investigate several different semantic similarity metrics of-Visit-Dependent Frame objects (PoVDF). A reasoner as-
such as the Jaccard and the Tversky Indices. After determining sists additionally the creation and/or the extension of our
the localness scores for each city of the corresponding local ontology (e.g. subsumption reasoning based on Description
entities in the users’ tweets, it is possible to estimate their Logic (DL)) in case of lacking data by making use of the
location. Ye et al. in [12] also use data from Location Based existing assertions. PoVDSSA is the core component of our
Social Networks (LBSN). Their method uses the semantic approach and refers to the Purpose-of-Visit-Driven Semantic
annotations of check-ins i.e. location categories rather then Similarity Analysis that takes place in order to cluster locations
GPS coordinates. It is based on a mixed hidden Markov Chain, dynamically and in a personalized manner depending on the
which rather than predicting the concrete next place, it predicts purpose of visit by the user. PoVDSSA is described in detail
the category of the next one to increase accuracy. in section III-C.
All existing approaches constraint themselves to fixed lo- The bottom branch of our framework comprises a proba-
cation categories and types without considering the dynamic bilistic graph, a multi-dimensional Markov Chain model in
purpose-of-visit-dependent variation of the meaning of loca- particular, which represents our actual location prediction
tions to the users. This lack of flexibility is carried over to model. First, the prediction model is being trained with the
their methods of representing associations between locations available raw, non-semantic data. Next, the trained model
and users through static and unalterable axioms or rules. Solely expects an optimization through the updating of its (previously
Long et al.’s work investigates a dynamic approach but it is learned) parameters by taking the current semantic similarity
based alone on popularity and not on the semantics behind the scores of the current locations into account. Finally, the
locations. optimized PoVDSSA prediction model provides an estimation
about the future location, which the user intent to visit next.
III. P URPOSE - OF -V ISIT-D RIVEN A PPROACH
Fig. 2 summarizes the principal functions of our approach into
Our porposed framework architecture is given in Fig. 3. Our a set of three layers: the preprocessing layer, the semantic
approach merges semantics with machine learning and consists trajectory enrichment layer and the modeling and optimization
consequently of two main parts: the semantic enrichment part layer.
and the actual predictor (top and bottom branch respectively).
A. Ontology-based Knowledge Base
In the first branch, sensed data like location and time In order to be able to represent both the taxonomic relations
are being semantically annotated and together with further of locations and the annotated set of high and low level
semantical information, such as the purpose of visit, these activities, as well as the time, and their interrelations, we
are stored in the Semantically Annotated Database (SADB). have chosen to create an ontology-based knowledge base. We

Authorized licensed use limited to: Silpakorn University provided by UniNet. Downloaded on February 15,2024 at 03:52:41 UTC from IEEE Xplore. Restrictions apply.
102
 Context & User LODB
Data

RawDB

Purpose-of-Visit-
Ontology OKB Driven Semantic
SADB Learning/
Similarity Analysis
Propagation
(PoVDSSA)
Reasoner

PoVDSSA - Intention/
Probabilistic Training/ Customization/
based Future Location
Graph/Model Identification Optimization
Graph/Model Prediction

Fig. 3: Purpose-of-Visit-Driven Semantic Similarity Analysis based location prediction (PoVDSSA) framework, whereby rawDB refers to
the raw data Database, SADB refers to the Semantically Annotated Database, LODB to a Linked Open Database, OKB to the Ontology-based
Knowledge Base and PoVDSSA to the Purpose-of-Visit-Driven Semantic Similarity Analysis Component respectively.

implemented our ontology in OWL using the Protege tool2 and

it comprises four major entities (classes): the Locations, the
Purpose of Visit, the Actions, and the Temporal & Event class.
In order to build the Locations class, we oriented ourselves on
the Foursquare location type categorization. The Purpose of
Visit entity represents the reasons of visiting certain locations.
It includes complex, high-level activities that may take place in
a location. The Actions class includes the elementary actions
of which high-level activities are composed. For instance the
activity celebrate a birthday covers the actions meet friends,
meet family, eat, drink, etc.. Finally, the Temporal & Event
entity describes time from a human point of view, considering
a human-like time granularity. For describing duration (time
intervals with start and end) and time in general we made use
of the standard OWL Time Ontology3 . This is necessary, inter
alia, for defining timeslots and blocks which in turn provide
the temporal granularity of our semantic trajectories. Fig. 4
illustrates indicative a part of the aforementioned Locations Fig. 4: Part of Location entity
class.
For achieving our goal in this work and to go beyond simple
memoryless stochastic process; a process which additionally
semantic labels whilst keeping a personal, experience-based
satisfies the Markov property, where predictions for the future
perception of locations at the same time, we want to associate
based on a short history yields similar results to those based
locations with the rest of the concepts of our ontology, such
on the whole history. Markov Chains are categorized by their
as time, purpose of visit, and actions.This conforms with the
order depending on how far back history is taken into account.
so called n-ary relations, which can be found in Description
A 1st-order Markov Chain is determined by the following
Logic. For handling this problem, we define a new entity,
conditional (Markov) property [14]:
which we name Purpose-of-Visit-dependent Frame (PoVdF).
The name refers partly to Minsky’s Frames in the 70s’, who p(z (m+1) |z (1) , z (2) , ..., z (m) ) = p(z (m+1) |z (m) ) (1)
used this term to encapsulate situation- and experience-based
knowledge [13]. , whereby z(1) , z(2) , ... is a series of random variables. A
2nd-order Markov Chain would analogously consider both
B. Multi-dimensional Markov Chain Model the current, as well as the previous state, etc. In our work,
We have chosen the probabilistic Markov Chain modeling we use Markov Chains to model semantic trajectories and
method to be our predictor’s base. A Markov Chain defines a their semantic locations L = {l1 , l2 , ...} and to predict
consecutively the future movement behavior of the users U
2 http://protege.stanford.edu = {u1 , u2 , ...}. In particular, we built a multi-dimensional
3 https://www.w3.org/TR/owl-time/#toc construct that comprises a set of transition matrices for each

Authorized licensed use limited to: Silpakorn University provided by UniNet. Downloaded on February 15,2024 at 03:52:41 UTC from IEEE Xplore. Restrictions apply.
 CoMoRea'18 - 14th Workshop on Context and Activity Modeling and Recognition

user, one for each timeslot and day of week. Fig. 5 illustrates an example to clarify this statement. The locations “park”,
a part of the location transition matrix. Each row and column “gym” and “restaurant” would probably land in three different
represent a single semantic location. Fig. 6 and 7 elucidate categories if we tried to cluster them by taking only the
vividly the multi-dimensionality of our predictor. location type into account. In contrast, our approach provides
a more dynamic clustering by considering additionally the
C. Purpose-of-Visit Semantic Similarity Analysis (PoVDSSA) purpose of visit. In this case, “park” and “gym” would be
Humans tend to employ cognitive frames, that is certain considered temporarily similar if the person visits the park for
mental constructs, in order to interpret things, entities and their jogging due to the fact that jogging is a fitness activity and thus
experiences about them [15]. Taxonomization and the build- common to the gym’s overall purpose of visit. Analogously,
ing of groups and relations between them and the included “park” would be found similar to the “restaurant” if the person
entities help clarifying concepts and are therefore of high has a pick-nick at that park.
importance to us. In order to build groups, humans draw on The PoVD similarity analysis takes place each time when
the fundamental notion of Similarity. Two objects are similar, a prediction is to be made, during which is investigated
when they share the same characteristics. While this definition how similar the current location is to each of the other
refers rather to the similarity between two physical objects, locations/location types found in our propagated knowledge
it can analogously be extended to a more general one that base. At each time, we weight the Markov matrix’ transition
expresses a characteristic-based similarity between two objects probability row of the location LmaxSim with the highest
in a knowledge graph or an ontology. This kind of similarity similarity to the current location Lcur (disregarding the current
can then consequently be referred to as semantic similarity. location itself, because it yields the absolut similarity of
Likavec et al. inspired from Tversky’s work [16], defined and 1.0) by multiplying it with the maximum similarity score
investigated such a property-based semantic similarity among SimScore itself. Finally, we use the resulted row to update
ontological objects in [17]. In our work, we adopt Likavec’s the transition probability row that corresponds to the current
method and define a similar equation to cluster semantically location by applying the following formula:
the visited locations of the users based on the purpose of visit
and the corresponding time. So, we treat both the time, and T P (Lcur )i,new =T P (LmaxSim ) × SimScore
(4)
the purpose of being at a location as characteristic features + of f set × T P (Lcur )i,old
of that particular location, which in turn reflects the PoVDF
The updating algorithm is described in detail in Algo. 1. So,
concept mentioned in section III-A. Equation 2 illustrates the
property-based semantic similarity adapted to our use case:
Algorithm 1: Markov transition probability updating process.
CP (l1 , l2 ) Data: Stay at current location l ∈ K, Current Context C (Purpose of Visit,
Sim(l1 , l2 ) = (2) Time, ..), Multi-Markov-Chain model M , Number of all locations
DP (l1 ) + DP (l2 ) + CP (l1 , l2 )
K = [k1 , . . . kn ]
Result: Updated transition probabilities for location l
, whereby l1 and l2 represent two different locations, CP 1 minsim ← 0.1, . . . , 0.9;
refers to the common purposes of visiting the particular loca- 2 probabilities[l→∗] ← M.getP robabilities(l, C);
3 probabilities ← probabilities[l→∗] ;
tions and DP gives the distinctive purposes that are associated 4 siml,∗ [siml,k1 , . . . , siml,kn ] ← getSimilarities(l, C);
only to the one location and do not appear in conjunction with 5 siml,∗ .sortReverse();
6 while siml,∗ .hasN ext() do
the other. However, after looking formula 3 closer, it becomes // location k ∈ K, shows highest similarity score to
apparent that it describes solely the similarity between two l
7 siml,k ← siml,∗ .next();
particular stays l1 and l2 at the locations L1 and L2 and 8 if siml,k ≥ minsim then
not the overall location similarity. In order to calculate the 9 probabilities[k→∗] ← M.getP robabilities(k, C);
10 probabilities ←
overall semantic similarity, we compute the average pairwise updateP robabilities(probabilities[l→∗] , probabilities[k→∗] ,
similarity of all existing stays at L1 and L2 as shown in 11 siml,k );
formula 3. This reflects our definition of a Purpose-of-Visit- 12 break;
13 end
Driven Semantic Similarity (PoVDSS). 14 end
15 return probabilities;
PN PM
i=1 j=1 sim(li , lj )
P oV DSS(L1 , L2 )pair∅ = (3)
M ∗N the updated transition probabilities for the current location
, whereby M and N provide the number of stays at the depend on the one hand of the transition probabilities on the
location L1 and L2 respectively. It must be noted here that most similar location and the corresponding similarity score.
our PoV-driven semantic similarity goes through all different On the other hand, they still depend on the old values, however
location and purpose-of-visit hierarchy levels that are mod- much less now, due to the offset factor. This provides us with
elled in our ontology and calculates a max-min normalized a smooth and adaptable rewarding-penalizing function.
aggregated value between 0 and 1.
By clustering locations in this way we are able to go beyond IV. E VALUATION AND D ISCUSSION
a simple type-specific categorization of locations and cluster A well known issue in research, and especially in the LBS
even locations of different type together. Let us consider community is the lack of appropriate training and evaluation

Authorized licensed use limited to: Silpakorn University provided by UniNet. Downloaded on February 15,2024 at 03:52:41 UTC from IEEE Xplore. Restrictions apply.
104
 CoMoRea'18 - 14th Workshop on Context and Activity Modeling and Recognition

Fig. 5: Part of location transition matrix. Fig. 6: Timeslot-specific transition matrices. Fig. 7: Day-specific sets of transition matrices.

datasets. The few good ones that exist, lack either in con- 250
#Aufenthalte
# Semantic (alle
location labels (allNutzer)
users)

sistency, or in providing the necessary thorough semantical 200

information. For this reason, we were forced to collect our

own data and conducted our own user study. We tracked the 150

movement of 6 mobile users for a period of 5 weeks and 100

came to an average of approximately 4000 GPS entries per

user. A tracking and annotation Android app was designed and 50

implemented in order to collect our data. We used the AWARE 0

context logging framework4 , which brings many benefits, such

Fr ilyH me

ar b
H e
e

G t
O m

W e
op Ce k

Bu Sta a
ng r

U tat l

ta r

aP h

fe

H Ba r

Ca icS r

D ur r
nc ch l

n
ht y
Ci lub

Ci erJ n
Tr yC int

Pi hu n

ce

rD ite

El rdw nP er
In on eS e
st sS re
lE re

Ke ugS nt
bP re
Li ce

D y
er
Re Do te
o
la S al
ke

pi nte

nS te

Ba
to be

st cto
Sh cal or

ig it

ar
as em
nd m
om

r c
rm Pu

ni io

rg tio

zz rc
C io
eS o

Ca
y

du ic to

ba to
ria to

a
a ia l

in
ffi

la

ec a la

la
M

N ers

t o

H As ea
ai en

t
c
m Ho
ie o

br
st
is r
G n

a
r
v

r
pe
as an activity-based energy efficient and phone battery saving

i
Fa

ed
Su

t r
bu
M

Am
tracking algorithm. The users were able to pause, or close
completely the app at any time. Each user was assigned an Fig. 8: Semantic label distribution of locations.
anonymous and random generated identification number (ID)
in order to preserve the users’ privacy. The data were collected 20 100
#Standorte
# Locations 90 #Seman
and stored first locally, on the mobile phone of the user, and #Standorte mit multi.
# Multi-purpose Zweck
locations
80
15
sent encrypted to our server at the end of the study. Three 70
60
Amazon vouchers were raffled among the participants in order 10 50
to increase their motivation. During the study we asked the 40
30
users to: 5
20

• Label their location 10

0 0
• Enter the purpose of visit
1

10

4
User1 User2 User3 User4 User5 User6
r

r
ze

ze

ze

ze

ze

ze

ze

ze
r
ze
ut

ut

ut

ut

ut

ut

ut

ut
• Define whether the particular location had been visited
ut
N

N
for another purpose up to that point Fig. 9: Number of multipurpose locations proportional to the total N

• Rate how important the particular location is to them, and number of locations per user.
• Provide additional descriptive information, like “sitting in
a coffee shop with a friend after work”, etc.
After analyzing our data, we came all in all to 431 differ- semantic trajectory-based approach of Ying [7], which we used
ent purpose-of-visit entries with respect to the following 9 here as reference. Fig. 10 shows how our PoVDSSA-based
high level location types: “home”, “transportation”, “night approach performs in comparison to Ying’s approach among
life”, “shopping”, “services”, “food”, “free time”, “educa- various k values with respect to accuracy, precision, recall
tion/university” and “work”. Fig. 8 shows the distribution of and f-score. We can see that our approach outperforms Ying’s
all annotated locations during our user study among all users. framework almost every time in all four metrics. It provides
Fig. 9 shows the proportion between the total number of the a clearly higher accuracy performance and recall value than
various (unique) semantic locations per user to the number Ying’s. This means that our system is not only more accurate,
of the locations at which more than one purpose of visit was but its estimations scatter less, especially when the k is high.
filled out. This can be attributed to the fact that our approach is able
We used k-fold cross validation as our evaluation method to replace and fill out missing current location transitions
and tested the following k-parameter values: [5, 10, 15, 20]. through existing transitions coming from the corresponding
We compared our approach to a multi-dimensional semantic similar location.
1st order Markov Chain Model, as well as to the well known Furthermore, on the hand, we can see that our purpose-of-
visit based approach shows an overall better performance than
4 http://www.awareframework.com Markov in most of the cases. Fig. 11 gives the performance of

Authorized licensed use limited to: Silpakorn University provided by UniNet. Downloaded on February 15,2024 at 03:52:41 UTC from IEEE Xplore. Restrictions apply.
105
 CoMoRea'18 - 14th Workshop on Context and Activity Modeling and Recognition

1
YingYing
1
Ying Ying
study. We evaluated our framework in contrast to Ying’s well
P
r
0.8
semMarkov
PoVDSSA
0.8
semMarkov
PoVDSSA
known approach and to a conventional Markov Chain model.
R
e
We show that our approach outperforms both Ying’s and

Trefferquote
e
Präzision

c 0.6 0.6
c
i
s
0.4
a
l 0.4
the Markov-based approach. However, some drawbacks, due
i
o
0.2
l
0.2
mostly to the lack of a bigger dataset, can also be identified.
n
In the future we plan to investigate further semantic similar-
0 0
5 10 15 20 5 10 15 20 ity algorithms and their integration in the overall prediction
k k

1 1
model, while investigating other machine learning models at
YingYing Ying Ying

A 0.8
semMarkov
PoVDSSA
0.8
semMarkov
PoVDSSA the same time.
F
Treffergenauigkeit

c
-
c
0.6 S 0.6 R EFERENCES
F-Maß
u
c
r
o
a 0.4 0.4
c
r [1] eMarketer. (2015) Key trends in mobile advertis-
e
y
0.2 0.2 ing. [Online]. Available: https://www.statista.com/statistics/436071/
location-based-service-users-usa/
0 0
5 10 15 20 5 10 15 20 [2] R. Nathan, W. M. Getz, E. Revilla, M. Holyoak, R. Kadmon, D. Saltz,
k k and P. E. Smouse, “A movement ecology paradigm for unifying or-
ganismal movement research,” Proceedings of the National Academy of
Fig. 10: Comparison of our approach (PoVSSA) to Ying’s approach Sciences, vol. 105, no. 49, pp. 19 052–19 059, 2008.
with regard to accuracy, precision, recall and f-score. [3] S. Spaccapietra, C. Parent, M. L. Damiani, J. A. de Macedo, F. Porto,
and C. Vangenot, “A conceptual view on trajectories,” Data Knowl. Eng.,
Mittelwert
Mean value vol. 65, no. 1, pp. 126–146, Apr. 2008.
1 [4] D. Ashbrook and T. Starner, “Learning significant locations and pre-
Markov Markov dicting user movement with gps,” in Wearable Computers, 2002.(ISWC
PoVDSSA 1dim semMarkov
0.8 2002). Proceedings. Sixth International Symposium on. IEEE, 2002,
semMultiMarkov-(Aktivität)
PoVDSSA Multi-dim - Activity
semMultiMarkov-(Aktivität,Tageszeit)
Trefferquote

R PoVDSSA Multi-dim - Activity, Time

e 0.6 semMultiMarkov-(Aktivität,Wochentag)
pp. 101–108.
c PoVDSSA Multi-dim - Activity, Day
a
semMultiMarkov-(Aktivität,Tageszeit,Wochentag)
PoVDSSA Multi-dim - Activity, Day, Time [5] S. Gambs, M.-O. Killijian, and M. N. del Prado Cortez, “Next place
l 0.4 prediction using mobility markov chains,” in Proceedings of the First
l
Workshop on Measurement, Privacy, and Mobility. ACM, 2012, p. 3.
0.2
[6] H. Gao, J. Tang, and H. Liu, “Mobile location prediction in spatio-
0 temporal context,” in Nokia mobile data challenge workshop, vol. 41,
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 2012, p. 44.
min. Ähnlichkeit
Similarity Threshhold
[7] J. J.-C. Ying, W.-C. Lee, T.-C. Weng, and V. S. Tseng, “Semantic
trajectory mining for location prediction,” in Proceedings of the 19th
Fig. 11: Comparison of various different configurations of our ACM SIGSPATIAL International Conference on Advances in Geographic
approach (PoVSSA) to the Markov model. Information Systems, ser. GIS ’11. New York, NY, USA: ACM, 2011,
pp. 34–43.
[8] N. Samaan and A. Karmouch, “A mobility prediction architecture
based on contextual knowledge and spatial conceptual maps,” IEEE
different configurations of our approach to Markov in relation Transactions on Mobile Computing, vol. 4, no. 6, pp. 537–551, Nov.
to the similarity threshold value, which determines whether 2005.
[9] Y. A. Ridhawi, I. A. Ridhawi, A. Karmouch, and A. Nayak, “A context-
an updating should occur or not (see Algo. 1). However, aware and location prediction framework for dynamic environments,”
on the other hand, we can see that temporal information in 2011 IEEE 7th International Conference on Wireless and Mobile
actually impairs our system instead of improving it. This Computing, Networking and Communications (WiMob), Oct 2011, pp.
172–179.
can be explained by the fact that, when it comes to human [10] X. Long, L. Jin, and J. Joshi, “Exploring trajectory-driven local geo-
movement, it is the sequence itself, which plays the most graphic topics in foursquare,” in Proceedings of the 2012 ACM Confer-
significant role and not the absolute information of time or ence on Ubiquitous Computing, ser. UbiComp ’12. New York, NY,
USA: ACM, 2012, pp. 927–934.
day. What also stands out in Fig. 11, is that the higher the [11] R. Krishnamurthy, P. Kapanipathi, A. P. Sheth, and K. Thirunarayan,
similarity threshold is chosen, the worst the performance. Knowledge Enabled Approach to Predict the Location of Twitter Users.
In tangible terms, this describes the situation in which our Cham: Springer International Publishing, 2015, pp. 187–201.
[12] J. Ye, Z. Zhu, and H. Cheng, “What’s your next move: User activity
Markov model gets updated only when highly similar (to prediction in location-based social networks,” in Proceedings of the 2013
the current) locations can be found. This limits however the SIAM International Conference on Data Mining. SIAM, 2013, pp. 171–
number of the existing transitions, which in turn downgrades 179.
[13] M. Minsky, “A framework for representing knowledge,” Cambridge,
the performance. MA, USA, Tech. Rep., 1974.
[14] C. M. Bishop, Pattern Recognition and Machine Learning (Information
V. C ONCLUSION AND F UTURE W ORK Science and Statistics). Secaucus, NJ, USA: Springer-Verlag New York,
Inc., 2006.
In this paper, we propose and investigate Purpose-of-Visit- [15] B. Heine, H. Narrog, C. J. Fillmore, and C. Baker,
Driven Semantic Similarity (PoVDSSA) as part of a semantic “A frames approach to semantic analysis.” [Online]. Avail-
location prediction method in combination with a multidi- able: //www.oxfordhandbooks.com/10.1093/oxfordhb/9780199544004.
001.0001/oxfordhb-9780199544004-e-013
mensional Markov Chain model. We were driven by our hy- [16] A. Tversky, “Features of similarity.” Psychological review, vol. 84, no. 4,
pothesis that a dynamic situation-dependent location clustering p. 327, 1977.
could enhance the overall performance. We implemented our [17] S. Likavec, F. Osborne, and F. Cena, “Property-based semantic similarity
and relatedness for improving recommendation accuracy and diversity,”
PoVDSSA-based algorithm and tested it on a real life data Int. J. Semant. Web Inf. Syst., vol. 11, no. 4, pp. 1–40, Oct. 2015.
set, which we collected ourselves by conducting a mobile user

Authorized licensed use limited to: Silpakorn University provided by UniNet. Downloaded on February 15,2024 at 03:52:41 UTC from IEEE Xplore. Restrictions apply.
106