=Paper=
{{Paper
|id=None
|storemode=property
|title=Urban Information Integration in MI-Search: Results and Future Research Activities
|pdfUrl=https://ceur-ws.org/Vol-915/paper_8.pdf
|volume=Vol-915
|dblpUrl=https://dblp.org/rec/conf/invit/MontanelliGC12
}}
==Urban Information Integration in MI-Search: Results and Future Research Activities==
https://ceur-ws.org/Vol-915/paper_8.pdf
Urban Information Integration in MI-Search: Results
and Future Research Activities*
Stefano Montanelli, Lorenzo Genta, and Silvana Castano
Università degli Studi di Milano
DI - Via Comelico, 39 - 20135 Milano
{stefano.montanelli,lorenzo.genta,silvana.castano}@unimi.it
Abstract. In this paper, we present the main achievements of the MI-Search
project for “multi-web” information integration around topics relevant for urban
users like for example city events and points of interest. In particular, we dis-
cuss the results of our experimental evaluation over a considered case study
about the city of Milan as well as ongoing/future research activities in the
framework of MI-Search.
1 Introduction
The recent success of mobile urban applications like point-of-interest exploration
apps and thematic event-publishing walls has produced a new attention on informa-
tion integration issues in pervasive and highly-dynamic scenarios. Existing tools in
this field are mainly focused on exploiting conventional geo-local information ex-
tracted from pre-organized integrated maps [5,8,9]. User-generated contents taken
from Social Web platforms (e.g., microblogging posts, RSS news) and/or semantic
web data taken from Linked Data repositories (e.g., Freebase, DBpedia) are mostly
ignored by such a kind of mobile applications. The MI-Search project aims at provid-
ing the capability to dynamically mix up and integrate “multi-web” information
around topics relevant for urban users like for example city events and points of inter-
est [6].
In this paper, we present the main achievements of the MI-Search project. First, we
overview the MI-Search solutions for urban-oriented, event-centric surfing of web
contents (Section 2). Then, we discuss the results of our experimental evaluation over
a considered case study about the city of Milan (Section 3). In particular, the experi-
mentation is aimed at assessing the effectiveness of the MI-Search techniques in re-
trieving pertinent and integrated information about a given event of interest. A discus-
sion about the MI-Search effectiveness from the scalability point of view is also pro-
vided. Finally, we outline ongoing/future research activities in the framework of MI-
Search (Section 4) and we provide our concluding remarks (Section 5).
*
This work is funded by Regione Lombardia and Fastweb S.p.A. in the framework of the
Dote in Ricerca project.
2 Overview of MI-Search
MI-Search is featured by an approach for urban information integration based on the
notions of smart city view and similarity cluster (see Figure 1).
Fig. 1. The MI-Search approach for urban information integration
By smart city view, we mean a set of similarity clusters about the events of a con-
sidered urban space that satisfies the selection criterion specified by an interested
urban user, like a citizen, a tourist, or a business agent. The idea of smart city views is
defined in MI-Search to enforce on-the-fly filtering operations over the available data
about the events of a considered urban space (e.g., a city, a metropolitan area, a re-
gion). The selection criterion consists in one or more user requirements featuring
keyword-, location-, and/or time-based constraints to use for choosing the pertinent
similarity clusters. The creation of a smart city view can be manually triggered by an
interested user that specifies a set of keywords. In a more realistic scenario, smart city
views are automatically generated by MI-Search to show potentially interesting urban
events based on the current user position or a planned appointment in the personal
user agenda. For instance, consider an user that plans to attend the musical Sister Act
at Teatro Nazionale, Milan, Italy on September, 20th 2012 8-11PM. When an entry is
inserted in the user personal agenda, MI-Search generates a smart city view featured
by pertinent similarity clusters that contain events/web resources concerned with Sis-
th
ter Act, nearby Teatro Nazionale, in a time period compatible with September, 20
2012 8-11PM. The user can explore the cluster contents to discover additional events
(e.g., a special menu offer for musical attendants by a restaurant next to Teatro Nazio-
nale) or useful information of interest (e.g., curiosities about the Sister Act musical).
A similarity cluster is built by collecting web resources that can have a different
nature (e.g., official sites, event walls, user posts/comments), but are similar in con-
tent (e.g., they refer to the same event, such as an art exhibition). Each similarity clus-
ter is characterized by a set of keyword-, location-, and time-based descriptors that are
exploited by MI-Search for matching with the user requirements to generate smart city
views.
The approach for urban information integration of MI-Search has the goal to define
a set of similarity clusters through the execution of three main phases, that are web
content acquisition, web content matching, and web content classification. In the fol-
lowing, we provide a summary overview of the MI-Search approach for urban infor-
mation integration. Further technical details can be found in [4,6].
Web content acquisition. MI-Search is based on a support repository called MI-
Search-DB capable of storing through a uniform representation all the different kinds
of web contents considered for advertisement of urban events. In MI-Search-DB, web
contents are distinguished in events that represent official initiatives like art exhibi-
tions or concerts and other resources that are acquired from web and then classified in
Tagged Resource, Microdata Resource, and Semantic Web Resource. An event is
acquired from electronic publishing walls and they are characterized by attributes that
describe its temporal frame (i.e., from-date, to-date, time, and frequency) and other
features, like description and price (where needed). Events are also associated with
information about contact-points (e.g., Phone, Facebook page, Twitter channel) and
geo-coordinates where the event takes place, respectively. Tagged resources are tradi-
tional web resources (i.e., web pages) and they are characterized by a raw structure
with few metadata. Microdata resources are posts/comments coming from news feeds
and microblogging systems (e.g., Facebook, Twitter posts). A microdata resource is
characterized by a short textual content and a set of metadata/properties, like title,
author, and creation date, that are commonly employed to describe publishing items.
Semantic web resources are instances/individuals coming from RDF(S) knowledge
repositories of the web-of-data (e.g., Freebase, DBpedia). These resources are charac-
terized by a structured description composed of a set of assertions denoting their spe-
cification in the web document of origin.
Each web content, either event or resource, is associated with a set of tags denoting
the keywords that more prominently characterize the event/resource.
Web content matching. This step has the goal to evaluate the degree of similarity
between each pair of web contents stored in the MI-Search-DB}. Given two web con-
tents wci and wcj, the similarity coefficient (wci,wcj) [0,1] denotes the level of
similarity of wci and wcj based on their common tags. We define Tagwc = {tag1, …,
tagm} as the set of tags associated with the web content wc in MI-Search-DB.
The similarity coefficient (wci,wcj) is calculated as follows:
where tagx ~ tagy denotes that tagx and tagy are matching tags
according to a string matching metric that considers the syntax of tagx and tagy. For
calculation, we employ our matching system HMatch 2.0, where state-of-the-art me-
trics for string matching (e.g., I-Sub, Q-Gram, Edit-Distance, and Jaro-Winkler) are
implemented [3].
Web content classification. Similarity clusters are built by relying on a clique
percolation method (CPM) [7]. This method receives in input a graph G where nodes
are the web contents stored in the MI-Search-DB repository and edges are established
between any pair (wci, wcj) of similar contents for which (wci,wcj) ≥ thm (0,1],
where thm is a matching threshold denoting the minimum level of similarity required
to consider two web contents as matching contents. The CPM returns a set of similari-
ty clusters where each cluster collects a region of nodes in G that are more densely
connected to each other than to the nodes outside the region.
3 Experimental results
An experimental evaluation has been performed to assess the effectiveness of the MI-
Search approach for urban information integration. To this end, two datasets called
MI-DS-focused and MI-DS-large have been defined for experimentation. These data-
sets are built by exploiting well-known publishing walls related to events about the
city of Milan, Italy. In particular, MI-DS-focused is fully based on the
http://www.milanodabere.it/ publishing wall, while MI-DS-large stores events extracted
from heterogeneous sources (e.g., http://www.milanodabere.it/, http://eventi-milano.it/,
http://www.eventiesagre.it/). The two datasets mainly differ in the number of stored
events (i.e., 134 events in MI-DS-focused vs. 253 events in MI-DS-large) and in the
number of tags associated with events (i.e., 1115 tags in MI-DS-focused vs. 247 tags
in MI-DS-large). The choice of having two datasets storing a strongly different num-
ber of events is motivated by the idea to provide a basic measurement of scalability
performance of MI-Search (see below). Moreover, we note that tags are associated
with events in two modalities: i) they are extracted from the considered publishing
walls by exploiting predefined wall categories (e.g., art-exhibition, entertainment,
theater), and ii) they are manually inserted by the wall staff. The difference of the two
datasets in the number of tags depends on the fact that most of the existing publishing
walls about Milan provide a poor event categorization, and thus events are associated
to a small number of tags (usually only one in MI-DS-large). The publishing wall
http://www.milanodabere.it/ (that has been exploited to generate MI-DS-focused) pro-
vides a more accurate event categorization that is also coupled with a manual event
annotation by the wall staff. As a consequence, in MI-DS-focused, each event is asso-
ciated with 8 tags on average (1115/134 8) that are appropriate to ensure interesting
matching results.
Capability of MI-Search to retrieve pertinent events. This experiment aims at
evaluating the quality of the similarity clusters on top of which smart city views are
built. In particular, given a dataset of web resources, we apply matching and classifi-
cation techniques and we analyze the resulting similarity clusters to assess whether
the events therein contained have been correctly clustered. In other words, we are
interested in measuring the capability of the matching techniques of MI-Search in
detecting similarities among events and web contents. Besides the basic similarities
that can be detected through the conventional search functionalities of publishing
walls, we want to evaluate the capability of MI-Search to discover non-trivial map-
pings among the dataset elements. The MI-DS-focused dataset has been employed in
this experiment due to the high number of tags-per-event that characterizes this data-
set. In this experiment, we used the event classification of http://www.milanodabere.it/
as baseline where two events are set to be similar if they are placed in the same cate-
gory by the wall staff. We executed matching over the events of MI-DS-focused using
HMatch 2.0 and we analyzed the overlap between the mappings detected by HMatch
2.0 and the baseline. We observed that with low values of matching threshold (e.g.,
0.2 thm 0.5) the MI-Search techniques are capable of detecting most of the map-
pings in the baseline (recall 80%). With higher values of similarity threshold (e.g.,
0.6 thm 1.0), recall decreases but interesting values of precision are obtained (pre-
cision 90%). Furthermore, we note that a number of mappings that are not con-
tained in the baseline are found by HMatch 2.0. In some cases, these additional map-
pings are false positives and they cause a precision decrease. This side-effect mostly
depends on the quality of the tags associated with the events in the dataset. Tags are
manually provided by the wall staff and usually they lack of accuracy in the sense that
they are too much generic for actually describing the event features. However, in
some other cases, these additional mappings represent non-trivial similarity mappings
between pairs of events that are somehow related. The pair of events in Figure 2 is an
example of non-trivial mapping between two events differently categorized in the
baseline but actually similar since they refer to events about the same historical cha-
racter (i.e., Gian Giacomo Poldi Pezzoli)1.
Appetizer at Poldi Pezzoli, Milan (EventID:267)
Il Museo Poldi Pezzoli propone un aperitivo, orchestrato dal Ristorante Don
Lisander, in tandem con l'attuale mostra dedicata a Gian Giacomo Poldi Pez-
zoli, noto collezionista risorgimentale, oltre che uno dei protagonisti delle Cin-
que Giornate di Milano…
tag: appetizer, centre, historical, Manzoni, Milan, muse, Pezzoli, Poldi, street…
(a)
Exhibition Gian Giacomo Poldi Pezzoli, Milan (eventID:195)
Nobiluomo colto e raffinato, Gian Giacomo Poldi Pezzoli fu uno dei protagonisti
delle Cinque Giornate di Milano. Una mostra allestita nelle stesse sale in cui il
collezionista visse e lavorò, ne ricorda passioni e impegno civico…
tag: centre, exhibition, Giacomo, Gian, historical, house, Manzoni, Milan, mu-
se, painting, Pezzoli, Poldi, Risorgimento, street…
(b)
Fig. 2. Example of a non-trivial mapping detected by HMatch 2.0 in the MI-DS-focused dataset
1
In this example, the web contents are left in the original Italian language while tags and
other metadata have been translated into English for reader convenience.
The example of Figure 2 represents the positive impact of using similarity match-
ing techniques in the construction of event clusters. This result further highlights that
the choice of appropriate tags for describing web contents is a key aspect for enforc-
ing an effective event classification (see Section 4 for further details about this topic).
Scalability performance of MI-Search. This experiment aims at evaluating the
scalability of MI-Search when the number of elements to consider for clustering in-
creases. In this respect, the scalability performances of MI-Search mostly depend on
the efficiency of the CPM techniques employed for generating similarity clusters. The
MI-DS-large dataset has been exploited in this experiment by executing CPM with a
progressively-increasing number of considered events stored in MI-DS-large. Fur-
thermore, we also executed CPM by varying the matching threshold thm used for de-
tecting similar web contents. This way, we can observe the scalability performances
of MI-Search on change of the degree of interconnection for the graph G used by
CPM (see Table 1).
Table 1. Scalability results obtained with the MI-DS-large dataset
#nodes thm=0.4 thm=0.6 thm=0.9 thm=0.97
30 ~250ms ~50ms ~25ms ~25ms
40 ~1s ~130ms ~30ms ~25ms
50 ~6s ~200ms ~30ms ~25ms
100 ~800s ~11s ~50ms ~30ms
250 ~800s ~200ms ~40ms
In the results, we note that CPM scales very well with high values of matching
threshold thm ≥ 0.9. Performances become critical when low/intermediate values of
thm are employed and more than 100 nodes belong to the graph G. In general, an in-
termediate value of matching threshold (e.g., 0.55 thm 0.65) is suggested to ensure
a good trade-off between precision and recall (see the experiment above). For this
reason, we observe that CPM can be suitably employed when small urban spaces are
considered (with less than 100 nodes to consider at a time). For larger urban spaces,
clustering solutions more efficient than CPM need to be enforced (see Section 4 for
further details about this topic).
4 Ongoing and future research activities
The following research activities about MI-Search are ongoing and/or planned in the
next future.
Periodic refresh of similarity clusters. The similarity clusters of MI-Search need
to be periodically updated to refresh the information about existing events and to in-
clude new events in the system. On this topic, two different research activities are
planned. On one side, we are working on a strategy for the incremental, on-the-fly
update of the similarity clusters. The basic idea is to refresh the acquisition of each
stored web content and to compare past and present descriptions to detect possible
changes. If the tag descriptions are changed, we consider to re-place the web content
in a different similarity cluster by evaluating the degree of similarity between the new
associated tags and the cluster descriptors. This strategy is adequate for contents cha-
racterized by a low obsolescence such as semantic web resources. On the other side,
we plan to work on a strategy for the batch, from-scratch reconstruction of the entire
set of similarity clusters. The basic idea is to start a new session of acquisition, match-
ing, and classification when the current similarity clusters are becoming obsolete.
This strategy is adequate for contents characterized by a high obsolescence such as
event descriptions on electronic walls and tagged/microdata resources. We also note
that these two strategies can be used in combination to reduce the overall computa-
tional effort. For instance, the incremental, on-the-fly refresh of existing clusters can
be employed for rapid update of the current classification, while the batch, from-
scratch reconstruction of similarity clusters can be executed when a sufficient number
of new events are found to be included in the MI-Search system.
Scalability of content classification techniques. The CPM techniques are inade-
quate for cluster aggregation when a high number of web contents needs to be ma-
naged. To overcome this limitation, we aim to equip MI-Search with a suite of differ-
ent aggregation methods to be dynamically activated according to the number of web
contents stored in the MI-Search-DB. On one side, we plan to investigate the use of
hierarchical clustering techniques [2]. This choice has a twofold motivation. First, the
complexity of the algorithm is quadratic in the worst case under the assumption that
an agglomerative strategy is adopted. Second, agglomerative hierarchical clustering
enforces a bottom-up approach which allows to stop the cluster computation once that
a desired level of aggregation is obtained. As a result, we argue that the adoption of
hierarchical clustering techniques enables to improve the content classification tech-
niques of MI-Search in terms of efficiency and flexibility at the same time. On the
other side, we plan to study supervised clustering techniques with predefined seeds
[1]. This kind of clustering techniques are based on a predefined set of cluster repre-
sentatives around which all the other elements are then aggregated. In MI-Search,
cluster representatives are urban events taken from publishing walls, while other web
contents like tagged, microdata, and semantic web resources are the elements to be
aggregated with events according to similarity coefficients. We argue that, the use of
supervised clustering techniques allows to further improve scalability performances of
web content classification due to the fact that the identification of cluster representa-
tives is immediate and similarity coefficients can be exploited for choosing the most
appropriate cluster to place the other contents.
Extraction of effective tag descriptions for web contents. For the web contents
of the MI-Search-DB, the set of associated tags are extracted from the websites of
origin. Usually, manually inserted tags and/or basic event categorizations of the elec-
tronic walls are exploited to derive the tag descriptions to use for populating MI-
Search-DB. From the experimentation, we observe that this strategy provides a low
number of associated tags, which are insufficient to effectively support the matching
operations. For this reason, we plan to integrate the current tag-extraction techniques
with tag-mining techniques derived from the literature on information retrieval. We
set a threshold that expresses the minimum number of tags to be associated with each
web content stored in the MI-Search-DB. Thus, tag-mining techniques are invoked to
complement the results of tag-extraction techniques when the required number of
extracted tags is not reached. State-of-the-art techniques in the field of text analysis
will be employed to this end, such as stop-word dropping, term tokeniza-
tion/normalization, and word stemming/lemmatization.
5 Concluding remarks
In this paper, we presented the main features and results of the MI-Search project for
the construction of smart city views based on web-content integration techniques.
Ongoing and future research activities in the framework of the MI-Search project have
been also outlined. The design of a complete MI-Search prototype for the Android
platform is currently under development. Further details are available at
http://islab.di.unimi.it/misearch/.
References
1. Basu, S., Banerjee, A., Mooney, R.: Semi-supervised Clustering by Seeding. In: Proc. of
the 19th Int. Conference on Machine Learning (ICML 2002). Sydney, Australia (2002)
2. Castano, S., De Antonellis, V., De Capitani Di Vimercati, S.: Global Viewing of Hetero-
geneous Data Sources. IEEE Transactions on Knowledge and Data Engineering 13(2)
(2001)
3. Castano, S., Ferrara, A., Montanelli, S.: Matching Ontologies in Open Networked Sys-
tems: Techniques and Applications. Journal on Data Semantics V (2006)
4. Castano, S., Ferrara, A., Montanelli, S.: Thematic Exploration of Linked Data. In: Proc. of
the 1st VLDB Int. Workshop on Searching and Integrating New Web Data Sources (VLDS
2011). Seattle, USA (2011)
5. Kanhere, S.S.: Participatory Sensing: Crowdsourcing Data from Mobile Smartphones in
Urban Spaces. In: Proc. of the 12th IEEE Int. Conference on Mobile Data Management
(MDM 2011). Luleå, Sweden (2011)
6. Montanelli, S., Castano, S.: MI-Search: a Smart Approach for Urban Information Cloud-
ing. In: Proc. of the 4th Interop-Vlab.it Workshop. Rome, Italy (2011)
7. Palla, G., Derényi, I., Farkas, I., Vicsek, T.: Uncovering the Overlapping Community
Structure of Complex Networks in Nature and Society. Nature 435 (2005)
8. Papadopoulos, S., Zigkolis, C., Kompatsiaris, Y., Vakali, A.: Cluster-based Landmark and
Event Detection on Tagged Photo Collections. IEEE Multimedia Magazine 18(1) (2011)
9. Schmeiß, D., Scherp, A., Staab, S.: Integrated Mobile Visualization and Interaction of
Events and POIs. In: Proc. of the 18th Int. ACM Conference on Multimedia. Firenze, Italy
(2010)