Geo-Semantic Event-POI Matching of Large Mobility Datasets

Ndiouma Bame

, Ibrahima Gueye

and Hubert Naacke

Département de Mathématiques-Informatique, Université Cheikh Anta Diop, Dakar, Senegal

LTISI, Ecole Polytechnique de Thiès, Thiès, Senegal

LIP6, Sorbonne Université, Paris, France

Keywords:

Geographic Similarity, Semantic Similarity, Event-Poi Matching, Sentence Embedding, Open Data.

Abstract:

Users often share data about their daily activities through social networks. These event data are very useful for

a variety of uses cases such as points of interest (POI) recommendation. However, event data often lack infor-

mation about POIs. Thus, enriching event data with POI information is of upmost importance. This implies to

know the POI in which an event took place before completing the data. We face the problem of aligning two

types of data sources, event data and POI data, which is difﬁcult because they do not have a common identiﬁer

or the same descriptive attributes. This work proposes and implements a complete methodology for the en-

richment of a large dataset of geolocated data on user events with POI using both geographical and semantic

properties. This effective methodology for matching POIs with geo-located events comprises four steps: (i)

in a ﬁrst step, we cross-reference the data using spatial proximity to deﬁne the geographical neighborhood of

each event; (ii) in a second step, we deﬁne the semantic neighborhood of each event based on a threshold on

the semantic similarity. The semantic similarity exploits events data such as their contextual description and

the tags by crossing them with those of the POI. (iii) these two types of similarity are combined for each POI

of the event semantic neighborhood, to evaluate a geo-semantic similarity score; (iv) subsequently, each event

is matched with the POI of the semantic neighborhood which maximizes the geo-semantic similarity score.

We propose a robust modeling of our methodology and evaluate the effectiveness of our approach.

1 INTRODUCTION

In the big data era, social networks and connected

objects, IoT or smart sensors generate an increas-

ing amount of data in real time on the web in gen-

eral. More and more data are produced in the form

of events describing the daily life activities of users.

These data are often available for analysis. The ex-

ploitation of these big data is in great demand by dif-

ferent actors in order to make more informed deci-

sions in various ﬁelds. For example, on the one hand,

the analysis of these data would allow to a particular

user to customize his activities over time according to

his interests and the possibilities that arise (Beladev

et al., 2016) (Behera et al., 2020) (Zhou et al., 2022).

On the other hand, for a company (through its sales or

marketing services), these data would allow to adapt

its service offers by anticipating the needs of potential

customers according to their interests(Fayyaz et al.,

2020) (Esmaeili et al., 2020). Thus, to suggest, guide

or advise users, several works have been carried out

as part of the recommendation of services to adapt

the offers according to the activities of the user on the

web (Esmaeili et al., 2020) (Hassannia et al., 2019)

(Hernández-Nieves et al., 2020): as a popular exam-

ple, we can cite the suggestion of articles to a user

based on the articles that he has consulted recently

or the most sold in his location. A point of interest

(POI) is generally a spatial object or event that can

attract user interest. It is characterized at least by

a name, a type and geographical coordinates. Sev-

eral works are being carried out within the context of

POI recommendation for users (Mishra et al., 2020)

(Massimo and Ricci, 2022) (Ricci, 2022). These

works generally consist of recommendation models

that are designed based on datasets about users ac-

tivities check-in (or events) through photos, videos,

posts,etc. The geo-location of these activities makes

it possible to match them with the POI often used by

relatives. This matching aims to complete the infor-

mation of user activities with those of the POIs. In

the same way that there are datasets on user activities

such as YFCC100M, YELP, Foursquare. There are

also datasets on POI such as Wikipedia, Geonames,

496

Bame, N., Gueye, I. and Naacke, H.

Geo-Semantic Event-POI Matching of Large Mobility Datasets.

DOI: 10.5220/0012132700003541

In Proceedings of the 12th International Conference on Data Science, Technology and Applications (DATA 2023), pages 496-503

ISBN: 978-989-758-664-4; ISSN: 2184-285X

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

Google Places, OpenStreetMap. These two types of

datasets are independent in the sense that they have

no common identiﬁer. They only share geolocation

information which is not sufﬁcient to associate a user

event with a POI. Thus, a good matching of these two

types of datasets is necessary to obtain complete data

allowing to design effective recommendation models.

This paper aims to propose and implement a new ap-

proach for matching user events (such as photos) with

POIs using both the geographical proximity and the

semantic similarity of their labels. In the remainder of

the paper, the research problem is posed in section 2.

Then, in section 3, the different steps of our event-

poi matching approach are explained. After that, the

validation of the proposal is presented in section 4 be-

fore we ﬁnish with the conclusion and future work in

section 5.

2 PROBLEM STATEMENT

For POI recommendation studies, researchers use

datasets that are made available by sources such as

social networks where users share their activities

(events) via publications in the form of photos, videos

or textual content. These events contain data about lo-

cation (GPS coordinates), time (date of the event) and

also labels (title, description, keywords,...). In order

to have enriched datasets, researchers complete these

event data with the data about POI. This completion

aims to ﬁnd the POI where each event takes place and

therefore to strengthen the data about event in order

to have more effective recommendation systems.

Existing work on POI recommendation (Chen

et al., 2023)(Wang et al., 2018)(Gottapu and Sri-

ram Monangi, 2017)(Lim et al., 2015) completes the

events data by matching them with the geographi-

cally closest POI, often ignoring the semantic links

between the events and the POI. This can lead to in-

consistencies, such as the user taking a photo (event)

in front of the visited POI , while standing closer to

another POI. Using the event labels would make it

possible to better identify the POI to be matched with

the event. Thus, matching must take into account the

semantic features of events and POIs. Other work has

matched events (Farazi and Raﬁei, 2022) with the se-

mantically closest POI, ignoring geographical prox-

imity. Again, in some cases, the user could describe

his event using the terms of a popular POI that would

be in the area of the (less known) POI that is vis-

ited. For example, a tourist taking a photo in a restau-

rant next to the Eiffel Tower would talk more about

the Eiffel Tower than the restaurant in his photo la-

bels. However, by integrating the distance between

the event and the candidates POI, we could have a

better match. The matching process must therefore

also take into account the geographical proximity be-

tween POI and events. To take into account seman-

tic characteristics, some works (Lim et al., 2015) of-

ten target speciﬁc cities and try to manually ﬁnd the

POI where each event takes place. A problem of this

data matching approach is that only the most popular

POIs are considered. As a result, many other POIs

were ignored, resulting in some events in the photo

dataset being ignored nearby. Another problem is that

the produced datasets cannot be used for studies on

other cities. Thus, these datasets are not interesting

for studies in a broader context. It is important to note

that in addition to these limited datasets for POI rec-

ommendation, there are larger datasets on POIs such

as Gowalla, Geonames, Wikipedia, Google Places,

which contain information such as type, location, etc.

about POIs on a global scale. An effective matching

of these POI datasets with user events datasets from

social networks would make it possible to provide re-

searchers with more complete and reliable datasets for

POI recommendation in a broader context.

Several challenges appear when we want to carry

out this matching work. (i) Indeed, the POIs data of-

ten come from sources which are different from the

source of the users events data. (ii) In large cities,

POIs (museums, monuments, restaurants, and other

sites) are often concentrated on relatively small ar-

eas. Their concentration is much stronger in urban

areas or large (Yannan et al., 2020), cities(Deng et al.,

2019). It then becomes very difﬁcult to match an

event to a single POI. This work propose to match

events data that are published on social networks with

large datasets on POI to produce a larger and enriched

dataset.

3 SPATIO-SEMANTIC

MATCHING OF EVENT TO POI

3.1 Preliminary Deﬁnitions: POI, User,

Event

We consider a set P of POIs: P = (p

, . . . , p

Each POI p

∈ P is characterized by an identiﬁer,

a location (latitude, longitude), a category, and

labels such as the name, alternative names and its

description. We deﬁne the function terms(p) which

allows us to produce a single text containing, for

each POI, its name, its category and its descrip-

tion. This function will be useful later when we

calculate the semantic similarity between POIs

Geo-Semantic Event-POI Matching of Large Mobility Datasets

497

and events. terms(p) = {{p.name, p.category} ∪

p.AlternativeName ∪ p.description}

Let U be a set of users: U = (u

, . . . , u

). Each

user, u

∈ U, is deﬁned by an identiﬁer and a set E

events: E

= (e

i,1

, . . . , e

i,k

). Each event e

i, j

is charac-

terized by a date, a geo-location (latitude, longitude)

and text labels assigned by the user. We deﬁne the

function terms(e) which, as for terms(p), will serve

us later in the calculations on the semantic similarity

between POI and events.

terms(e) = {{e.title} ∪ e.description ∪ e.Usertags}

3.2 Geographical Similarity

Let d

geo

(e, p) be the geodetic distance between an

event e and a POI p.

3.2.1 Geographical Neighborhood

The geographical neighborhood of the event e is de-

noted Ngeo(e) and contains the POI p located within

a radius less than r of e :

geo

(e) = {p ∈ P|d

geo

(e, p) ≤ r}

In (Gueye et al., 2020), authors proposed an ef-

ﬁcient method to compute the geographical matching

between an event and a set of POI. This matching con-

sists in measuring the geodetic distance between the

event and each POI in its surroundings based on the

GPS coordinates (latitude and longitude).

3.2.2 Geographical Similarity Computation

Let d

Max

(e) be : the maximum distance of POIs which

are in the neighborhood N

geo

(e) of the event e.

Max

(e) = {Max(d

geo

(e, p))|p ∈ N

geo

(e)}

The geographical similarity s

geo

(e, p) between

an event e and a POI p is a bounded quantiﬁcation

(in [0,1]) of the distance d

geo

(e, p) with respect to

the maximum distance d

Max

(e) between e and its

neighborhood Ngeo(e).

geo

(e, p) =

Max

(e) − d

geo

(e, p)

Max

(e) + 0.01

This similarity s

geo

(e, p) varies inversely propor-

tional to the distance d

geo

(e, p). Thus, we ob-

tain a geographical similarity at 1 (maximum) when

geo

(e, p) = 0 and a geographical similarity at 0 (min-

imum) when d

geo

(e, p) = d

Max

(e).

3.3 Semantic Similarity

The semantic similarity s

sem

(e, p) is a quantiﬁcation

of the semantic link between a POI p and an event e.

It is based on the textual similarity between the meta-

data of the POI p (terms(p)) and those of the event e

(terms(e)).

3.3.1 Semantic Similarity Computation

We rely on a pre-trained language model to statisti-

cally represent the semantic information of a textual

description. A language model encodes the content of

a sentence by a dense vector of large dimension. Note

that the dimension of the vectors is much smaller than

the size of the vocabulary, hence the name of embed-

ding of a sentence in a denser space. In other words,

a language model M associates the vector v

= M(p)

with a sentence p.

Language models are based on machine learning

methods. They are trained for classiﬁcation tasks in

such a way that semantically similar sentences have

similar vectors. Thus, the cosine between the vectors

represents a measure of similarity between two sen-

tences. We use Bert sentence transformer (Reimers

and Gurevych, 2019), called sBert, because its efﬁ-

ciency, in terms of precision for a sentence classiﬁca-

tion task, exceeds that of other language models such

as ((Cer et al., 2018), (Pennington et al., 2014)). One

of the advantages of the sBert model is that it can de-

termine a vector representation of words on which the

model has not been trained. This makes it possible to

consider the textual content of events and POIs with-

out doing linguistic pre-processing or denoising.

The semantic similarity s

sem

(e, p) is therefore de-

ﬁned by :

sem

(e, p) = cosinus(M(terms(e)), M(terms(p)))

3.3.2 Semantic Neighborhood

The semantic neighborhood N

sem

(e) of an event e cor-

responds to the set of POIs (p) of its geographical

neighborhood N

geo

(e) which have a semantic simi-

larity S

sem

(e, p) with the event which is greater than

or equal to a local similarity threshold of the event

min

(e). This local similarity threshold is obtained

from the maximum between the average of the sim-

ilarities between the event and the POIs of its geo-

graphical neighborhood and a predeﬁned global sim-

ilarity threshold G

sem

. Thus, we have:

min

(e) = max



sem

, avg(S

sem

(e, p)|p ∈ N

geo

(e))



from where:

sem

(e) = {p ∈ N

geo

(e)|S

sem

(e, p) ≥ S

min

(e)}

DATA 2023 - 12th International Conference on Data Science, Technology and Applications

498

3.4 Geo-Semantic Similarity

The geo-semantic similarity is a quantiﬁcation both

spatial and semantic between an event and a POI by a

normalized score. This score denoted S

geosem

(e, p) for

a couple (event e, POI p) is a weighted linear combi-

nation of the geographical similarity and the semantic

similarity between the event and the POI.

The score S

geosem

(e, p) is :

geosem

(e, p) = a.s

geo

(e, p) + (1 −a).s

sem

(e, p)

where the parameter a which is bounded between 0

and 1, weights the contributions of the semantic link

and the geographical link. The choice of its value

depends on the link (semantic or geographical) to be

highlighted more in relation to the other. Thus, a zero

value of a (a = 0) considers only the semantic simi-

larity, while a maximum value of a (a = 1) considers

only the geographical similarity. So, to calculate the

geo-semantic similarity, a non-extreme value must be

chosen.

3.5 Event-POI Matching Algorithm

In the rest of this section, we present the different

steps needed to match each event e with its most geo-

semantically similar POI.

1. Filtering based on the geographical neighbor-

hood: consists in determining the geographical

neighborhood N

geo

(e) of each event e.

2. Filtering based on the semantic neighborhood:

consists in determining the semantic neighbor-

hood N

sem

(e) of each event e.Indeed, among the

POIs in the geographical neighborhood of the

event (N

geo

(e)), only those that are in its semantic

neighborhood N

sem

(e) are considered for the rest

of the matching.

3. Matching based on geo-semantic similarity: con-

sists in associating the event, with the POI of

its semantic neighborhood N

sem

(e) which is clos-

est based on the geo-semantic similarity. Let

geosem

(e) be the POI that matches e. The POI

geosem

(e) in the semantic neighborhood of the

event maximizes the geo-semantic similarity be-

tween a neighbor and the event. We have:

geosem

(e) = arg max

p∈N

sem

(e)

geosem

(e, p))

At the end of this matching, we have a set M of event-

poi associations (e, m

geosem

(e)).

Moreover, in order to compare the proposed

matching algorithm with existing methods we deﬁne

the geographic-only matching. Let m

geo

(e) be the ge-

ographically nearest POI to event e:

geo

(e) = arg max

p∈N

geo

(e)

geo

(p, e)

Respectively, we deﬁne the semantic-only match-

ing. Let m

sem

(e) be the semantically most similar POI

to event e,

sem

(e) = arg max

p∈N

sem

(e)

sem

(p, e)

4 EXPERIMENTAL VALIDATION

4.1 Implementation

The proposed approach has been implemented with

scalability objectives in mind, since we target a solu-

tion that can handle large datasets. To this end, we

relied both on Apache spark for data preparation and

SBert

python package to get POIs and events em-

beddings.

4.2 Experimental Setup

We ran our experiments on a 64 cores machine

equiped with 376GB RAM running Debian 4.19

Linux.

4.3 Dataset Description

We used two publicly available datasets. The ﬁrst

dataset is Yahoo Flickr Creative Commons 100M

(YFCC) (Thomee et al., 2016) that contains user

events about photos they took. Each event caries in-

formation about the user id, the date at which the

photo was taken, the photo geo position, user tags,

a title, a description and other attributes related to the

picture itself such as the camera charactetistics.

The second dataset we used is geonames (Geon-

ames, ) that contains world wide POI data. Each POI

has a name, an alternate name, a description, a cate-

gory, a geo position. We preprocess the YFCC dataset

to remove redundancy: we ensure that only one photo

has been taken for a given user, location and date.

Due to the high computional cost of our approach and

since we observed that a large part of the users took

all their photos in only one country, we restrict our

analysis to France.

www.sbert.net/

Geo-Semantic Event-POI Matching of Large Mobility Datasets

499

4.4 Experimental Methodology

Computing Geographical Neigborhood. The goal

is to ﬁnd the POIs geographically closest to each

event, and thus the geographic neighborhood of the

event. Indeed, in most cases, a given event is closely

surrounded by several points of interest. We assigned

to each photo event, a set of geographically close

POIs located inside a ﬁxed radius circle centered at

the photo location. We set a radius equal to 500m

to delimit the neighborhood of each event. To deter-

mine such geographic neighborhood for each event,

we computed a spatial similarity join between the

Geoname POI dataset and photos data. We used the

method described in (Gueye et al., 2020) to perform

this join efﬁciently.

Computing Semantic Neighborhood. After hav-

ing computed the geographical neighborhoods of the

events, we then evaluate the semantic similarity be-

tween each event and each POI of its geographical

neighborhood. We thus obtain for each event, the se-

mantic similarity with each POI of its geographical

neighborhood. Then, the semantic neighborhood con-

tains the subset of POIs that have a semantic similarity

greater than a given threshold.

We consider a globally ﬁxed threshold and inves-

tigate in our experiments several threshold values.

Geo-Semantic Matching. For each event, we con-

sidered the POIs of its semantic neighborhood and

computed the geo-semantic similarity of each POI

with the event from the geographic and semantic sim-

ilarities and the parameter a. The values of the param-

eter a used for this calculation are deﬁned according

to the experiments presented in the following section.

Thus, each event is associated with the POI that max-

imizes the geo-semantic similarity.

4.5 Experiments

The main goal of our experiments is to show that tak-

ing into account both geo position and content allows

for better matching between photo events and POIs.

The matching is rather obvious in case the closest

POIs are also the closest ones semantically. However

this is not always the case. Often, POIs semantically

similar with a photo are not the closest ones, while

they are still located in the geographic neighborhood

of the photo. In order to observe the beneﬁt of using

both geographic and semantic similarities, we exper-

iment various matching conﬁgurations depending on

two parameters:

• The minimum semantic similarity threshold that

each candidate POI must exceed. We vary its

value from 0 to 100% using a 10% step.

• a the relative weight of the geographic similarity

taken into account in the combined geo-semantic

similarity. The parameter a is a geo-weight ra-

tio. By varying its value from 0 to 100% using a

10% step, we can generalize geo-semantic simi-

larity over geographic and semantic ones: when a

equals 0, similarity is purely semantic. Reversely,

when a equals to 1 similarity is purely geographic.

We ran the matching algorithm for each combina-

tion of the two parameters (threshold, a) . Then, we

report the average similarities of the matched (photo,

POI) couples. We also report the number of matched

events and compute the relative matching ratio com-

pared to the total number of events to match.

In addition, to assess the beneﬁt of the combined

geo-semantic similarity strategy, we count the number

of matched events that would also have been matched

by either a geographic-only or a semantic-only strat-

egy.

4.6 Quantitative Validation Results

Effect of Semantic Threshold on the Number of

Matched Events. Figure 1 shows the number of

matched events when the semantic threshold is vary-

ing from 0 to 100%.

Figure 1: Matched events according to the semantic thresh-

old.

Obviously, when the threshold is set to 0, every

event is matched since the dataset has been prepared

so that every event has at least one POI in its neigh-

borhood. More interestingly, when the threshold is in-

creased, the ratio of matched events decreases rather

slowly, i.e., the ratio remains higher than a linear de-

crease for a threshold in the range of 0% to 40%.

This means that POIs located in the event neighbor-

hood (i.e., selected only based on their geo-position)

do have some semantic similarity with this event. We

can therefore recommend a threshold of up to 40% in

order to keep matching more than half of the events.

DATA 2023 - 12th International Conference on Data Science, Technology and Applications

500

4.6.1 Effect of the Geo-Weight Ratio

We aim to assess the effect of the geo-weight ratio

on the matching performance: what value of the geo-

weight ratio would allow to preserve both a good ge-

ographic and semantic similarities ? In other words,

we highlight the compromise between semantic and

geographical matching. The higher the geo-weight ra-

tio, the closer the points will be geographically to the

event, but the further they will be semantically. To

this end, for several pairs (a, T ) of (geo-weight, se-

mantic threshold) values, we ﬁrst consider the set E

of events such that each e ∈ E

has at least one POI,

denoted n, in its neighborhood for which the semantic

similarity between n and e is greater than T . Then, we

run our geo-semantic matching algorithm on E

and

get the set M

of (e, p) matches such that each event

e is matched with p which equals to m

geosem

(e)) the

most geo-semantic similar POI among the neighbor-

ing POIs of e: M

= {(e, m

geosem

(e))|∀e ∈ E

} Then

we report, in blue, the average geographic similarity

of the matched POIs geoSim =

∑

(e, p)∈M

geo

(e, p)

and

respectively, in orange, their average semantic simi-

larity semSim =

∑

(e, p)∈M

sem

(e, p)

Figure 2: Similarities evolution according to geographical-

weight.

Figure 2 shows four ﬁgures for different thresh-

old values T ∈ {0, 0.3, 0.5, 0.6}. Each ﬁgure reports

two plots when the geo-weight value (x-axis) varies

from 0 to 100%. Let’s focus on the ﬁrst ﬁgure (thresh-

old:0). As expected, the semantic similarity (orange

plot) is rather low around 0.4, and gets even lower

down to 0.3 when POI geolocations are taken into ac-

count i.e., when the geo-weight increases along the

x-axis.

One the second ﬁgure, the T threshold has been

increased to 0.3 so the neighborhood of each event

has been restricted to more semantically similar POIs.

We can observe that the difference between the orange

and blue plots is diminishing. This means that target-

ing more similar POIs helps to improve the overall

matching quality. This trend can still be seen in the

third ﬁgure (T=0.5): the orange and blue curves ap-

proach each other as soon as the geo-weight thresh-

old exceeds 50%, reaching a similarity of around 0.6.

However, for a higher threshold value (T=0.6 in the

last ﬁgure) the reverse trend appears where the or-

ange and blue curves are moving away from each

other which is explained by the fact that E

0.6

contains

too few events with highly semantically similar neigh-

bors.

Overall, observing the trend on the four ﬁgures,

we can conclude that an interesting tradeoff holds for

a threshold T of 0.5 (ﬁgure where curves are the clos-

est) and a geo_weight of 50%.

4.6.2 Ability to Use Geo-Semantic Matching to

Associate a Larger Number of Events

We investigate whether our proposed geo-semantic

matching can match more POIs than the baseline

geographic-only or semantic-only methods.

Figure 3: Relative matching capabilities (Gain in %) ac-

cording to semantic threshold.

Given a semantic threshold T , we deﬁne the rela-

tive capability C

geo

(T ) (also named Geo_gain on the

Figure on Figure 3) as the number of events for which

the matched POI differs from the closest one geo-

graphically, divided by the number of events.

geo

(T ) =

|{(p, e) ∈ M

|p ̸= m

geo

(e)}|

|Match

Respectively, C

sem

(T ), also named Sem_gain on

Figure 3, is the number of events for which the

matched POI is not the most similar one semantically,

divided by the number of events.

sem

(T ) =

|{(p, e) ∈ M

|p ̸= m

sem

(e)}|

|Match

Geo-Semantic Event-POI Matching of Large Mobility Datasets

501

Figure 3 reports C

geo

(T ) in blue and C

sem

(T ) in

orange for the semantic threshold T varying from

0 to 100%, for three different geo_weight values

{0.3, 0.4, 0.6}.

We observe that the capabilities decrease when the

semantic threshold is increasing. This is because the

set of candidate POIs becomes more and more re-

stricted as the threshold increases, thus the matched

POI is more often either the closest geographically

or the most similar semantically. More signiﬁcantly,

theses ﬁgure show that our proposed matching ﬁnds

matches that could not have been found by simpler

methods. Overall, up to a threshold of 0.4, our solu-

tion yields a relative capability of at least 10% more

than the geographic-only matching and 35% more

that semantic-only matching.

4.7 Validation of the Quality of the

Geo-Semantic Matching

In the previous section we measured the similarity of

an event with its best matched POI, using either geo-

positioning or semantics or a combination of both,

whereas here the main goal is different. We aim to

measure the interest of our approach for the end-user:

To ﬁnd matches that better ﬁt what the user would

have pointed to, is it more interesting to consider only

the geolocation of events and POIs or to also con-

sider their semantic data? For this purpose, we have

collected 301 events in the city of Paris. Then, us-

ing the event photo as a gold standard, we manually

matched each event to one or more POIs related to the

event photo. We call this database the Ground Truth

Database (GTD). Then, for each event in the GTD we

computed the most similar POI neighbor based on our

combined geographic and semantic similarity s

geosem

matching method. We then compared them with those

of the GTD to obtain a precision and recall value. The

precision is deﬁned as the number of matches that

were computed and are actually true with respect to

the GTD, divided by the number of matches that were

computed. In other words, it measures the percent-

age of correct matches. The recall, on the other hand,

is deﬁned as the number of matches that were com-

puted and are actually true in the GTD, divided by the

number of true matches. In other words, it measures

the completeness of the matching algorithm. On Fig-

ure 4, we report the precision and recall values for

several geometric weight (denoted a) values ranging

from 0, meaning that only the semantic similarity is

taken into account, to 1, meaning that only the geo-

graphic similarity is taken into account.

We observe that the precision is maximized for

a = 20%. We can explain that the precision remains

Figure 4: Precision and Recall of our Geo-semantic simi-

larity measures.

below 60% by the fact that our matching method

may assign more than one POI to an event in case

of ties, whereas in most cases only one POI is ac-

tually a true match. Furthermore, we observe that

semantic-only matching performs with less precision

(13% drop down) than our approach. Respectively,

the geographic-only matching is 28% less precise. As

far as recall is concerned, it reaches a maximum value

of 87% when a is 20%. More precisely, the relative in-

crease is of 27% compared to the geographic-only ap-

proach and 8.7% compared to the semantic-only ap-

proach. This conﬁrms the beneﬁt of our approach. We

can conclude that the combination of both geographic

and semantic similarity allows for working in a more

efﬁcient way.

5 CONCLUSIONS AND FUTURE

WORKS

The matching of the user activity data with those

of the POIs is necessary insofar as it enriches these

data with information on the properties of the visited

POIs. This would improve the performance of recom-

mendation models by considering user preferences.

This work proposed and implemented an approach for

matching POI to events by using both geographical

positioning and semantic metadata.

We deﬁned a geo-semantic similarity function that

combines geographical and semantic similarity. Then

we associate with each event, the POI which maxi-

mizes that geo-semantic similarity score. We made

some experimental validation using real open datasets

such as Flickr’s YFCC100M (Thomee et al., 2016) for

user events and the Geonames (Geonames, ) dataset

for points of interest. Through quantitative validation,

DATA 2023 - 12th International Conference on Data Science, Technology and Applications

502

we studied statistics on the evolution of semantics

and geographical similarities and we then were able

to recommend semantic similarity threshold and geo-

graphical weighting values in order to improve both

geographical and semantic similarities. Moreover,

we conducted qualitative validation with a Ground

Truth Database which conﬁrms the effectiveness of

our proposal with satisfactory performances. In fu-

ture work, we plan to address the problem related to

events that have too low a semantic similarity with

POI in their geographical neighborhood. We intend

to deﬁne the concept of user visits by grouping suc-

cessive and close events from the same user, which

will allow deducing the next most likely POI to at-

tach to an event. Another direction is to consider the

trajectories of a user composed of a sequence of suc-

cessively visited POI. A criterion for the consistency

of a trajectory could be speciﬁed. As a perspective,

we also plan to study event-POI matching in real-time

mobility data streams.

REFERENCES

Behera, R. K., Gunasekaran, A., Gupta, S., Kamboj, S.,

and Bala, P. K. (2020). Personalized digital market-

ing recommender engine. Journal of Retailing and

Consumer Services, 53:101799.

Beladev, M., Rokach, L., and Shapira, B. (2016). Recom-

mender systems for product bundling. Knowledge-

Based Systems, 111:193–206.

Cer, D., Yang, Y., Kong, S., Hua, N., Limtiaco, N., John,

R. S., Constant, N., Guajardo-Cespedes, M., Yuan, S.,

Tar, C., Strope, B., and Kurzweil, R. (2018). Univer-

sal sentence encoder for english. In Proceedings of

the 2018 Conference on Empirical Methods in Natu-

ral Language Processing, pages 169–174.

Chen, L., Cao, J., Tao, H., and Wu, J. (2023). Trip rein-

forcement recommendation with graph-based repre-

sentation learning. 17(4).

Deng, Y., Jiping, Liu, Y., and Luo, A. (2019). Detecting

urban polycentric structure from poi data. ISPRS In-

ternational Journal of Geo-Information, 8(6).

Esmaeili, L., Mardani, S., Golpayegani, S. A. H., and

Madar, Z. Z. (2020). A novel tourism recommender

system in the context of social commerce. Expert Sys-

tems with Applications, 149:113301.

Farazi, S. and Raﬁei, D. (2022). Tracking where events take

place: Reverse spatial term queries on streaming data.

CoRR, abs/2201.07378.

Fayyaz, Z., Ebrahimian, M., Nawara, D., Ibrahim, A., and

Kashef, R. (2020). Recommendation systems: Algo-

rithms, challenges, metrics, and business opportuni-

ties. Applied Sciences, 10(21).

Geonames. The geonames dataset. "http://www.geonames.

org/export/". Online; Accessed: June 20, 2022.

Gottapu, R. D. and Sriram Monangi, L. V. (2017). Point-of-

interest recommender system for social groups. Pro-

cedia Computer Science, 114:159–164.

Gueye, I., Naacke, H., and Gançarski, S. (2020). Enrich-

ing geolocalized dataset with pois descriptions at large

scale. In Innovations and Interdisciplinary Solutions

for Underserved Areas (4th , InterSol), pages 264–

273.

Hassannia, R., Vatankhah Barenji, A., Li, Z., and Alipour,

H. (2019). Web-based recommendation system for

smart tourism: Multiagent technology. Sustainability,

11(2).

Hernández-Nieves, E., Hernández, G., Gil-González, A.-B.,

Rodríguez-González, S., and Corchado, J. M. (2020).

Fog computing architecture for personalized recom-

mendation of banking products. Expert Systems with

Applications, 140:112900.

Lim, K. H., Chan, J., Leckie, C., and Karunasekera, S.

(2015). Personalized tour recommendation based on

user interests and points of interest visit durations. In

International Conference on Artiﬁcial Intelligence, IJ-

CAI’15, pages 1778–1784.

Massimo, D. and Ricci, F. (2022). Building effective recom-

mender systems for tourists. AI Magazine, 43(2):209–

224.

Mishra, R. K., Urolagin, S., and Jothi, J. A. A. (2020).

Sentiment analysis for poi recommender systems. In

2020 Seventh International Conference on Informa-

tion Technology Trends (ITT), pages 174–179.

Pennington, J., Socher, R., and Manning, C. D. (2014).

Glove: Global vectors for word representation. In

Conference on Empirical Methods in Natural Lan-

guage Processing, EMNLP, pages 1532–1543.

Reimers, N. and Gurevych, I. (2019). Sentence-BERT: Sen-

tence embeddings using Siamese BERT-networks. In

Empirical Methods in Natural Language Processing

and International Joint Conference on Natural Lan-

guage Processing (EMNLP-IJCNLP), pages 3982–

3992.

Ricci, F. (2022). Recommender Systems in Tourism, pages

457–474. Springer.

Thomee, B., Shamma, D. A., Friedland, G., Elizalde, B.,

Ni, K., Poland, D., Borth, D., and Li, L.-J. (2016).

Yfcc100m: The new data in multimedia research.

Commun. ACM, 59(2):64–73.

Wang, H., Shen, H., Ouyang, W., and Cheng, X. (2018). Ex-

ploiting poi-speciﬁc geographical inﬂuence for point-

of-interest recommendation. In International Joint

Conference on Artiﬁcial Intelligence, IJCAI’18, pages

3877–3883.

Yannan, Z., Lu, Z., and Xinhuan, Z. (2020). The spatial dis-

tribution of retail outlets in urumqi: The application

of points of interest. Open Geosciences, 12(1):1541–

1556.

Zhou, C., Leng, M., Liu, Z., Cui, X., and Yu, J.

(2022). The impact of recommender systems and

pricing strategies on brand competition and consumer

search. Electronic Commerce Research and Applica-

tions, 53:101144.

Geo-Semantic Event-POI Matching of Large Mobility Datasets

503