Content uploaded by Giorgia Di Tommaso
Author content
All content in this area was uploaded by Giorgia Di Tommaso on Sep 24, 2018
Content may be subject to copyright.
Wiki-MID: a very large Multi-domain Interests
Dataset of Twitter users with mappings to
Wikipedia
Giorgia Di Tommaso, Stefano Faralli*, Giovanni Stilo, and Paola Velardi
Department of Computer Science, University of Rome,
*Unitelma-Sapienza, Italy
{ditommaso,stilo,velardi}@di.uniroma1.it
stefano.faralli@unitelmasapienza.it
Abstract. This paper presents Wiki-MID, a LOD compliant multi-
domain interests dataset to train and test Recommender Systems, and
the methodology to create the dataset from Twitter messages in English
and Italian. Our English dataset includes an average of 90 multi-domain
preferences per user on music, books, movies, celebrities, sport, politics
and much more, for about half million users traced during six months
in 2017. Preferences are either extracted from messages of users who
use Spotify, Goodreads and other similar content sharing platforms, or
induced from their ”topical” friends, i.e., followees representing an in-
terest rather than a social relation between peers. In addition, preferred
items are matched with Wikipedia articles describing them. This unique
feature of our dataset provides a mean to categorize preferred items,
exploiting available semantic resources linked to Wikipedia such as the
Wikipedia Category Graph, DBpedia, BabelNet and others.
Keywords: semantic recommenders, Twitter, Wikipedia, users’ interest
Permanent URL: https://doi.org/10.6084/m9.figshare.6231326
1 Introduction
Recommender systems are widely integrated in online services to provide sugges-
tions and personalize the on-line store for each customer. Recommenders identify
preferred items for individual users based on their past behaviors or on other
similar users. Popular examples are Amazon [1] and Youtube [2]. Other sites
that incorporate recommendation engines include Facebook, Netflix, Goodreads,
Pandora and many others.
Despite the vast amount of proposed algorithms, the evaluation of recom-
mender systems is very difficult [3]. In particular, if the system is not operational
and no real users are available, the quality of recommendations must be evalu-
ated on existing datasets, whose number is limited and what is more, they are
focused on specific domains (i.e, music, movies, etc.). Since different algorithms
may be better or worse depending on the specific purpose of the recommender,
2 Giorgia Di Tommaso, Stefano Faralli, Giovanni Stilo and Paola Velardi
the availability of multi-domain datasets could be greatly beneficial. Unfortu-
nately, real-life cross-domain datasets are quite scarce, mostly gathered by ”big
players” such as Amazon and eBay, and they not available to the research com-
munity1.
In this paper we present a methodology for extracting from Twitter a large
dataset of user preferences - that we call Wiki-MID - in multiple domains and in
two languages, Italian and English. To reliably extract preferences from users’
messages, we exploit popular services such as Spotify, Goodreads and others. Fur-
thermore, we infer many other preferences from users’ friendship lists, identifying
those followees representing an interest rather than a peer friendship relation.
In this way we learn, for any user, several interests concerning books, movies,
music, actors, politics, sport, etc. The other unique feature of our dataset, in
addition to multiple languages and domains, is that preferred items are matched
with corresponding Wikipedia pages, thus providing the possibility to generalize
users’ interests exploiting available semantic resources linked to Wikipedia, such
as the Wikipedia Category Graph, BabelNet, DBpedia, and others.
The paper is organized as follows: Section 2 summarizes previous research
on creating datasets for recommender systems, Sections 3, 4 and 5 present the
methodology to create Wiki-MID, Section 6 is dedicated to dataset statistics
and evaluation, and Section 7 describes the released resource, which has been
designed on top of the Semantically-Interlinked Online Communities (SIOC)
core ontology. Finally, in Section 8 we provide a summary of distinctive features
of our resource and some directions for future work.
2 Related work
Recommender systems are based on one of three basic approaches [4]: collabora-
tive filtering [5] generates recommendations collecting preferences of many users,
content-based filtering [6] suggests items similar to those already chosen by the
users, and knowledge-based recommendation [7] identifies a semantic correlation
between user’s preferences and existing items. Hybrid approaches are also widely
adopted, e.g., [8]. All approaches share the need of sufficiently large datasets to
learn preferences and to evaluate the system, a problem that is one of the main
obstacles to a wider diffusion of recommenders [9], since only a small number of
researchers can access real users data, due to privacy issues.
To overcome the lack of datasets, challenges as RecSys have been launched2,
and dedicated web sites have been created (e.g., SNAP3or Kaggle4), where re-
searchers can upload their datasets and make them available to the community.
However it is still difficult to find appropriate data for novel types of recom-
menders, as the majority is focused on a single topic, like music [10], [11], ) food
([12], [13]), travel ([14], [15]) and more [16]. Furthermore, while a small number
1https://recsys.acm.org/wp-content/uploads/2014/10/recsys2014- tutorial-cross_domain.pdf
2https://recsys.acm.org/
3http://snap.stanford.edu/data
4https://www.kaggle.com/datasets/?sortBy=hottest&group=all
Multi-domain Interests Dataset of Twitter users with mappings to Wikipedia 3
of large datasets are available, such as Movielens [17], Million song dataset [18]
and Netflix Prize Dataset [19], many others are quite small and based on very
focused experiments.
Concerning the source of data for extracting preferences, social networks are
often used (mainly Twitter, Facebook, Google+, LinkedIn or a combination of
sources, such as in [20]), since their content is freely available with more or less
severe restrictions. The interested reader can refer to [21] for a detailed survey
of methods adopted in literature to collect social data for the purpose of infer-
ring and enhancing users’ interests profiles. Preferences are induced from users’
profiles (e.g, [22]), authoritative (topical) friendship relations [23], followee bi-
ographies [24], and messages ([25], [26], [27] and many others).
Data extraction from Twitter messages is a popular strategy, however, it is also
computationally expensive and error-prone, since it requires natural language
processing techniques to analyze the text. To overcome this difficulty, a number
of studies exploited platforms (e.g., Youtube, Spotify) that integrate among their
services the ability to post the user’s personal content on the most popular social
network sites, such as movies that users are watching. Sharing this information
is done in a simple and predefined way. Depending on the social network chosen,
the content, for example a Youtube video, will be shared with a pre-formatted
message formed by the video name, a link, a self-generated text and, if provided, a
numerical rating (eg. ”How It’s Made: Bread” https://youtu.be/3UjUWfwWAC4
via YouTube). The message can also be enriched and personalized by the user.
In [25] these types of messages are extracted from Twitter, to detect music in-
terests. The dataset is based on 100,000,000 tweets with the #nowplaying main
tag. Tweets are extracted via Twitter APIs over 3-years and next, MusicBrainz
and Spotify are used to add more details. Other studies extract data about music
[27] or sport [28] events. However, all the datasets generated in this way concern
only one domain of interest.
To the best of our knowledge, the only really multi-domain dataset is presented
in [29], where pre-structured tweets about three domains - movies, books and
video-clips - are extracted respectively from IMDb (Internet Movies Database),
Youtube and Goodreads. With respect to this work, we collect a much wider
number of interests, since in addition to pre-formatted messages based on a
number of available services, we reliably extract many additional types of in-
terests exploiting users’ followees lists. Furthermore, as shown in Section 7, we
collected many interest types for each user, while the dataset released in [29]
includes only 7 users with at least 3 types of interests.
3 Workflow
This section summarizes the data sources and workflow to create the Wiki-
MID multi-domain resource. We extract preferences (with unary ratings) from a
user’s messages and from his/her friendship list, identifying those followees who
represent an interest rather than a peer friendship relationship. The process is
in three steps:
4 Giorgia Di Tommaso, Stefano Faralli, Giovanni Stilo and Paola Velardi
1. Extracting interests from users’ textual communications. Using textual fea-
tures extracted from users’ communications, profiles or lists seems a natural
way for modeling their interests. However, this information source has sev-
eral drawbacks when applied to large data streams, such as the set of Twitter
users. First, it is computationally very demanding to process millions of daily
tweets in real time; secondly, the extraction process is error prone, given the
highly ungrammatical nature of micro-blogs. To reliably extract preferences
from Twitter users’ messages, in line with other works surveyed in Section
2, we use a number of available services, described hereafter, that allow to
share activities and preferences in different domains - movies, books etc. -
using pre-formatted expressions (e.g, for Spotify: #NowPlaying) followed by
the url of a web site, from which we can extract information without errors.
The drawback is that a relatively small number of users access these services
and in addition, preferences are extracted only in few domains.
2. Extracting interests from users’ friendship lists. In [30] the authors argue that
users’ interests can also be implicitly represented by the authoritative (topi-
cal) friends they are linked to. This information is available in users’ profiles
and does not require additional textual processing. Furthermore, interests
inferred from topical friends are less volatile since, as shown in [31], ”com-
mon” users tend to be rather stable in their relationships. Topical friends are
therefore both relatively stable and readily accessible indicators of a user’s
interest. Another advantage is that average Twitter users have hundreds of
followees, many of which, rather than genuine friends, are indicators of a
variety of interests in different domains, such as entertainment, sport, art
and culture, politics, etc.
3. Mapping interests onto Wikipedia pages. The final step is to associate each
interest, either extracted from messages or inferred from friendship relations,
with a corresponding Wikipedia page, e.g.:
@nytimes ⇒WIKI:EN:The New York Times
(in this example, @nytimes is a Twitter account extracted from a user’s
friendship list). Although not all interests can be mapped on Wikipedia,
our experiments show that this is possible in a large number of cases, since
Wikipedia articles are created almost in real-time in correspondence with
virtually any popular entity, either book, or song, actor, event, etc.
We applied this workflow to two Twitter streams in two languages, English
and Italian, as explained in the next Sections.
4 Extraction of users’ interests
4.1 Extracting interests from messages
Everyday a huge number of people uses on-line platforms (eg. Yelp, Foursquare,
Spotify, etc.) that allow to share activities and preferences on different domains
on a social network in a standard way. Among the most popular services accessed
by Twitter users, we selected those providing pre-formatted messages:
Multi-domain Interests Dataset of Twitter users with mappings to Wikipedia 5
– Spotify: Spotify is a music service offering on-demand streaming of music,
both desktop and mobile. Users can also create playlists, share and edit them
in collaboration with other users. In addition to accessing the Spotify web
site, users can retrieve additional information such as the record label, song
releases, date of release etc.. Since 2014, Spotify is widely used in America,
Europe and Australia. Spotify is among the services allowing to generate self-
generated content shares in Twitter. An example of these tweets is: ”#Now-
Playing The Sound Of Silence by Disturbed https://t.co/d8Sib5EDVf”.
The standard form of these tweets is:
#NowPlaying <title>by <artist > <URL>
By filtering the tweets stream and using Twitter APIs for hashtag detection,
we generated a stream of all the users who listened music using Spotify.
– Goodreads and aNobii: Similarly to Spotify, a number of platforms al-
lows to share opinions and reviews on books. In these platforms, users can
share both titles and ratings. Similarly to Spotify, generated tweets have
a predefined structure and point to an URL. In the book domain, we use
Goodreads (10 million users and 300 million books in the database) and for
Italian, the more popular aNobii service.
– IMDb and TVShowTime: In the domain of movies, currently there are no
dominant services. Popular platforms in this area are Flixter, themoviedb.org
and iCheckMovies. However, many of these platforms use the IMDb database,
owned by Amazon, which handles information about movies, actors, direc-
tors, TV shows, and video games. We also use the TvShowTime service for
Italian users.
In order to extract users’ preferences from these services, we first collect in a
Twitter stream T S all messages including a hashtag related to one of the above
mentioned services (#NowPlaying, #IMDb ..). Then, we extract from T S the
music, movie and book preferences for the set of users Uwho accessed these ser-
vices. Unlike [29], we avoid parsing tweets using specific regular expressions, since
users are free to insert additional text in the pre-formatted message. Rather, in
line with [32], we exploit an element that most pre-formatted tweets have: the
URL, e.g., #NowPlaying High by James Blunt https://t.co/7EiepE2Bvz
Accordingly, we collect all tweets containing the selected hashtags and discard
those which do not include an URL. The reason for extracting the informa-
tion from the URL (which is computationally more demanding) rather than
from the tweet itself is twofold: i) Tweets can be ambiguous or malformed,
and furthermore, users can insert additional text in the pre-formatted message,
e.g, ”#NowPlaying Marty. This guy is amazing.http: // t. co/ jwxvLiNenW ”.
Scraping the html page at the URL address ensures that we extract data with-
out errors, even for complex items such as book and movie titles; ii) The URL
includes additional information (e.g., not only the title of a song, but also the
singer and the record label), which provide us a context to reliably match the
extracted entity (song, book, movie) with a Wikipedia article, as detailed in
Section 5.1. Since the URL in tweets is a short URL, we first extend the origi-
nal URL so that all URLs belonging to a given platform can be identified (for
6 Giorgia Di Tommaso, Stefano Faralli, Giovanni Stilo and Paola Velardi
example, https://t.co/oShYDc6DeL →http://spoti.fi/2cTPn0U). Next, we
access the web site and scrape its content. For each platform we obtain the fol-
lowing data:
Music:<Title, Author (eg. singer, band)>
Books:<Title, Author>
Movie:<Title, Year of production, Type (eg. movie, tv series)>
4.2 Extracting interests from users’ ”topical” friends
In addition to preferences extracted from users’ messages, we also induce inter-
ests from their topical friends, a notion that we first introduced in [23]. We denote
as topical friends those Twitter accounts in a user’s followees list representing
popular entities (celebrities, products, locations, events . . .). For example, if a
user follows @David Lynch, this means that he/she likes his movies, rather than
being a genuine friend of the director. There are several clues to identify topical
friends in a friendship list: first, topical relations are mostly not reciprocated, sec-
ond, popular users have a high in-degree. However, these two clues alone do not
allow to distinguish e.g., bloggers or very social users from truly popular entities.
To learn a classification model to distinguish between topical and peer friends,
we first collected a network of Verified Twitter Accounts. Verified accounts 5
are authentic accounts of public interest. We started from a set of seed verified
contemporary accounts in 2016, and we then crawled the network following only
verified friends, until no more verified accounts could be found. This left us with
a network of 107,018 accounts of verified contemporary users (V), representing
a ”training set” to identify authoritative users’ profiles. To learn a model of au-
thoritativeness, we used the set Vand a random balanced set of ¬Vusers. For
each account in Vand ¬V, we extracted three structural features (in degree,
out degree and their ratio) and one binary textual feature (presence in the user’s
account profile of role words such as singer, artist, musicians, writer..). Then, we
used 80% of these accounts to train a SVM classifier with Laplacian kernel and
the remaining 20% for testing with cross-validation, obtaining a total accuracy
of 0.88 (true positive rate 0.95 and true negative rate 0.82).
Next, from the set Uof users in our Twitter datasets (separately for the
English and Italian streams), we collected the set Fof Twitter accounts such
that, for any f∈Fthere is at least one u∈Usuch that ufollows f. The
previously learned classifier was used to select a subset Ft⊆Fof authoritative
users representing ”candidate” topical friends.
Finally, an additional filtering step is applied to identify ”true” topical friends in
Ft, i.e., genuine users’ interests, which consists in determining which members
of the set Fthave a matching Wikipedia page. This step is described in Section
5. The intuition is that, if one such match exists, the entity to which the Twitter
account belongs is indeed ”topical”6. Although this filtering step may affect the
recall of the method, it provides high accuracy, as demonstrated in Section 6.
5https://developer.twitter.com/en/docs/api-reference-index
6we do not directly attempt a match of all f∈Fwith Wikipedia, since it is very
computationally demanding and has a reduced precision.
Multi-domain Interests Dataset of Twitter users with mappings to Wikipedia 7
5 Mapping interests to Wikipages
The last step of our methodology consists in mapping the collected users’ inter-
ests to Wikipages. This step has both the advantage of improving the precision
of detected users’ interests, and providing a mean to categorize them. We use
different mapping methodologies for interests extracted from messages and those
induced from users’ friendship lists.
5.1 Mapping movies, songs and books
Mapping interests extracted from users’ messages to Wikipedia pages is a very
reliable process, given the additional contextual information extracted from the
URL (see previous Section 4.1). Wikipedia mapping is obtained by a cascade of
weighted boolean query on a Lucene Index, as in the example below, used to
search the Wikipage of an item:
< T I T LE ∈W ik iT itle >w1∧< AU T H OR ∈W ikiGl oss >w2∧((< W ORD S ∈
W ikiT itle >w3∨< AU T H OR ∈W ikiT itl e >w4∨¬(< W OR DS ∈W ik iT itle >w3
∨< AUT H OR ∈W ikiT itle >w4∨< W ORDS ∈W ikiT ext >w5))
∨< W ORDS ∈W ik iT ext >w5< W O RDS > for music = {”song”}< W ORDS >
for books = {”books”, ”novel”, ”saga” .. . }< W ORDS > for movie = {”film”, ”se-
ries”, ”TV series”, ”episode” . . . }
where wiis a weight assigned to a query. When the page doesn’t exist or is not
available, we search the page of the item’s author, using similar queries.
5.2 Mapping topical friends
Matching interests extracted from a user’s friendship list with corresponding
Wikipedia pages is far more complex, because of homonymy, polysemy and am-
biguity. Furthermore, the information included in a user’s Twitter profile is very
sketchy and in some case misleading, therefore it may not provide sufficient
context to detect a similarity with the correspondent Wikipedia article. For ex-
ample, Bill Gate’s description field7in his Twitter profile is: ”Sharing things I’m
learning through my foundation work and other interests...” which has little in
common with his Wikipedia page: ”William Henry Gates III (born October 28,
1955) is an American business magnate, investor, author, philanthropist, hu-
manitarian and co-founder of the Microsoft Corporation along with Paul Allen.”
We note that other studies have considered this task. For example, the authors in
[33]) use an heuristics based on the overlap coefficient of last 20 topical followees’
tweets and the Wikipedia article summary, which is rather data demanding. In
[34] the authors use a methodology which is similar to the one we firstly pre-
sented in [23], based on a comparison between Twitter description fields and the
content of a Wikipage. As previously noted (see the Bill Gates example), this
might not be sufficient in many cases. In the present work, to reliably assign a
Wikipedia page to a large fragment of users in the set Ftof U0sauthoritative
7as retrieved on January 2018
8 Giorgia Di Tommaso, Stefano Faralli, Giovanni Stilo and Paola Velardi
friends, we use an ensemble of methods, with adjudication by majority voting.
The methodology is described in what follows.
1. Task Description and data - Given a set Ft={f1, f2, ..., fn}of candidate
”topical” Twitter profiles and a set of Wikipages W={w1, w2, ..., wm}we define
a mapping function M:Ft⇒W∪ {λ}where the value of the function Mfor a
given Twitter profile fiis a Wikipage wj, which is the corresponding Wikipage
of the entity having the twitter profile fior λ, where λmeans ”no match”.
We define an ensemble of three mappers exploiting the information included
in Twitter profiles and in DBpedia entities associated to Wikipedia.
Profiles of Twitter users provide, among the others, the following information:
– profile address: e.g., https://twitter.com/katyperry
– user ID: a numeric value to uniquely identify a user (not visible on the
rendered web page);
– screen name a string that can be used to refer to a user when posting a
message (e.g. @katyperry);
– name the extensive name of the owner of the profile (e.g. ”Katy Perry”);
– url: the link to a profile-relevant homepage (e.g. ”katyperry.com”). Only a
fragment of profiles have an URL to a homepage.
– description: a short description to describe the user and welcome profile
visitors.
Furthermore, from each wikipage wj(e.g., Figure 1, upper right, shows the
Wikipedia page of the singer ”Katy Perry”) it is possible - thanks to DBpedia -
to collect additional information, here is a small subset:
– title: the title of the page (e.g. ”Katy Perry”);
– content: the textual content of the page;
– homepage: a property (collected and included on DBpedia from infoboxes)
which (when present) links to a web page (homepage) related to the main
entity described in wj(e.g., ”katyperry.com”)
– links extracted from the homepage: are those links included on the
source html of the above mentioned homepage, e.g., in the html of the web-
page at katyperry.com we find: https://facebook.com/katyperry,https:
//twitter.com/katyperry, . . .
2. Mapping methods - We rely on an ensemble of three different methodolo-
gies (M1,M2and M3) of association between the set Ftof Twitter profiles and
Wikipages. The first is based on text mining and structural properties of the so-
cial network, the other two are based on finding direct correspondences between
the field url in a Twitter profile and the property homepage in a DBpedia entity.
1. M1- Context Based mapping: We use the methodology that we first
presented in [35], summarized in what follows:
a) Selection of candidate senses: For any fiin Ft, find a (possibly empty) list
Multi-domain Interests Dataset of Twitter users with mappings to Wikipedia 9
Wikipedia2Twitter
Twitter2Wikipedia
Webpage wj.homepage
wj.homepage
katyperry.com
wj.address
https://en.wikipedia.org/wiki/Katy_Perry
wj.title
Katy Perry
links_extracted_from_the_homepage
https:twitter.com/katyperry, ...
fi.url
katyperry.com
fi.description
(empty)
fi.profile_address
https:twitter.com/katyperry
fi.screen_name
katyperry
fi.name
”Katy Perry”
Wikipage wj
Twitter Profile fi
wj.homepage
katyperry.com
-Wj.homepage equals to fi.profile_address ?
-links_extracted_from_the_homepage contains
fi.profile_address ?
-links_extracted_from fi.description contains
wj.homepage or wj.address ?
-fi.url equals wj.homepage or wj.address ?
Fig. 1. Example of Twitter2Wikipedia and Wikipedia2Twitter mapping
of candidate wikipages, using BabelNet [36] synonym sets (in BabelNet, each
”BabelSynset” points to a unique Wikipedia entry ). For example, @kathy-
perry has candidates Katy Perry and Katy Perry discography, but there are
cases with dozens of candidates (e.g., https://en.wikipedia.org/wiki/
John_Williams_(disambiguation));
b) BoW Disambiguation: Compute the bag-of-words (BoW) similarity be-
tween the user description in fi’s Twitter account and each candidate wikipage.
The BoW representation for each wikipage is obtained from its associated
BabelNet relations (relations are described in [37]);
c) Structural Similarity: If no Wikipages can be found with a sufficient level
of similarity (as for the previous example of Bill Gates description field),
select from fi’s friendship list those friends already mapped to a wikipage
-if any- and compute the similarity between those wikipages and candidate
wikipages. For example, to correctly map the Twitter account of Bill Gates to
Wikipedia, profile information of the following Twitter users in his friendship
list are used: Paul Allen, Melinda Gates, TechCrunch, Microsoft Foundation,
and more. Note that Paul Allen is explicitly mentioned in the first sentence
of Bill Gate’s wikipage.
2. M2- Twitter2Wikipedia: as sketched in Figure 1, we first collect a set
of URL from a given profile fiincluding: the link (if any) in the field url
and all the links extracted from the profile description filed (In the ex-
ample of Figure 1, since the profile description is empty, we collect only
10 Giorgia Di Tommaso, Stefano Faralli, Giovanni Stilo and Paola Velardi
the link katyperry.com). Second, we search a Wikipage wj(if any) for
which one of the links collected for fi(in our example we collected the link
katyperry.com) matches with the link provided in the homepage property,
(in our example the Wikipage with title ”Katy Perry” has a property home-
page whose value matches exactly the link katyperrry.com), or directly with
the address of the page itself (e.g. https://en.wikipedia.org/wiki/Katy_
Perry). Note that this mapping method is error prone: for example, from the
Twitter profile of Paul Gilmour we extract the following url: skysports.com
matching with the homepage property of the following wikipage: https:
//en.wikipedia.org/wiki/Sky_Sports. Although related, this is not Paul
Gilmour’s page https://en.wikipedia.org/wiki/Paul_Gilmour.
3. M3- Wikipedia2Twitter:M3is symmetric to M2. As shown in Figure 1,
we map a given fito a Wikipage wjif the homepage property, or one of the
links extracted from the source html of the homepage in wj, matches the
Twitter profile address. Like for Twitter2Wikipedia, this mapping method
is error prone.
For each of the above three approaches we add three additional mapping
functions ES M1,ES M2,ESM3where each mapping function is defined as :
ESMk(ti) = (wj,if Mk(ti) = wjand ti.name =wj.title
λ, otherwise
In other words, ESMk”reinforces” the result of Mkif the name field in the
Twitter profile perfectly matches with the title of the Wikipedia page. Note that
this is often not the case, as for @realDonaldTrump.
3. Ensemble Voting - For a given Twitter profile fithe ensemble voting mech-
anism selects the Wikipage wjfor which there is maximum agreement among the
6 mapping functions (M1,M2,M3,ES M1,ES M2,ESM3), and there are at least 2
Mj,Mkin agreement (j6=k). The threshold 2 has been empirically selected to
obtain the best compromise between number of mapped interests and precision,
as detailed in Section 6.
6 Wiki-MID statistics and evaluation
The outlined process has been applied to two streams of Twitter data, in English
and Italian, extracted during 6 months (April-September 2017) using Twitter
APIs. We collected the maximum allowed Twitter traffic of English users men-
tioning service-related hashtags (e.g., #NowPlaying for Spotify), and the full
stream of messages in Italian, since they do not exceed the maximum. As a final
result, we obtained for a large number of users a variety of interests along with
their corresponding Wikipedia pages. An excerpt of a Twitter user’s interests is
shown in Table 1. In the example, we selected two interests from each of the four
sources from which they have been induced: IMDb (movies), Goodreads (books),
Multi-domain Interests Dataset of Twitter users with mappings to Wikipedia 11
USER ID:787930***
Source Interest Wikipage
IMDb
Eyes Wide Open - 2009 - movie WIKI:EN:Eyes Wide Open (2009 film)
Okja - 2017 - movie WIKI:EN:Okja
Goodreads
The Beautifull Cassandra - Jane Austen WIKI:EN:Jane Austen
The Beach - Alex Garland WIKI:EN:The Beach (novel)
Spotify
I Don’t Know What I Can Save You From - Kings of Convenience WIKI:EN:Kings of Convenience!
Nothing Matters When We’re Dancing - The Magnetic Fields WIKI:EN:The Magnetic Fields
Topical friends
@IMDb WIKI:EN:IMDb
@UNICEF uk WIKI:EN:UNICEF UK
@TheMagFields WIKI:EN:The Magnetic Fields
@BarackObama WIKI:EN:Barack Obama
@Spotify WIKI:EN:Spotify
Table 1. Excerpt of a Twitter user’s interests
Spotify (music) and the user’s topical friends. Although a detailed analytics of
interest categories is deferred to further studies, the example shows the common
trend that a user’s interests, either extracted from his/her messages or from top-
ical friends, are strongly related, and in same case identical. For example, the
user in Table 1 frequently accesses the IMDb and Spotify services, and he/she is
also a follower of the IMDb and Spotify Twitter accounts. Furthermore, his/her
interest in the band The Magnetic Field emerges from both source types.
Overall, we followed 444,744 English-speaking and 25,135 italian-speaking
users (the set U) who accessed at least one of the services mentioned in Section
4.1. Tables 2 and 3 show general statistics of interests extracted from users’ mes-
sages respectively, for English and Italian speaking user. In the English dataset
we crawled more than 20M tweets from these users, of which, about 2.7M could
be associated to the URL of a corresponding book, movie or song. On average,
we collected 6 interests per user. What is more, several users have interests in at
least two of the three domains. Figure 2 compares the Venn diagram of interest
types in our dataset (left) with that reported in [29] (right), to demonstrate
the superior coverage of our dataset, even when considering only preferences ex-
tracted from users’ messages. The last line of Tables 2 and 3 (precision) shows
that the methodology to extract and map preferences from messages is very re-
liable. We evaluated the precision (two judges with adjudication) on a randomly
selected balanced sample of 1200 songs, books, and movies in English, obtain-
ing a precision of 96% with a k-Fleiss Inter Annotator Agreement (IAA) of 18.
For the Italian dataset, we evaluated 750 songs, books, and movies, obtaining a
precision of 98%, and a k-Fleiss of 0.97.
The number and variety of extracted preferences is mostly determined by
the interests induced from users’ topical friends, as shown in Table 4 (Table 5
for the Italian dataset). The average number of interests induced for each user
is as high as 82, and the distribution is shown in Figure 3, left (English stream),
and right (Italian stream). Figure 3 (left) shows, e.g., that there are 100,000
users in Uwith ≥100 interests induced from their topical friends. As far as the
topical interests mapping performance is concerned, in [35] we estimated that
inducing interests from topical friends and subsequent mapping to Wikipedia
with mapping method M1 has an accuracy of 84%. Since our aim in this work
8The evaluation is rather straightforward, as readers may verify inspecting the re-
leased dataset and mappings.
12 Giorgia Di Tommaso, Stefano Faralli, Giovanni Stilo and Paola Velardi
message-based interests (|U|=444,744 English speaking users)
Music Books Movie Total
Platform: Spotify Goodreads IMDb All
#crawled tweets (tweets with selected hash-
tags)
19,941,046 693,975 97,772 20,732,793
#cleaned tweets (tweets fro which an URL was
extracted)
2,519,166 139,882 88,355 2,747,403
# of unique interests with a mapping to a
Wikipage
253,311 20,710 8,282 282,303
average #interests per user 6 8 6 6
average #users per interest 7 3 7 6
precision of Wikipedia mapping (on 3 samples
of 400 items each)
94% 96% 97% 96%
Table 2. 6-months (April-September 2017) statistics on message-based interests
extracted from English-speaking users
message-based interests (|U|= 25,135 Italian speaking users)
Music Books Movie Total
platform Spotify ANobii IMDb
TVShowTime
All
#crawled tweets (tweets with selected hash-
tags)
273,256 12,198 2,229 287,683
#cleaned tweets (tweets for which an URL was
extracted)
70,330 12,193 2,119 84,642
# of unique interests with a mapping to a
Wikipage
9,926 4,690 279 14,895
average #interests per user 3 9 7 6
average #users per interest 5 2 5 4
precision of Wikipedia mapping (on 3 samples
of 250 items each)
96% 98% 100% 98%
Table 3. 6-months (April-September 2017) statistics on message-based interests
extracted from Italian-speaking users
Fig. 2. Venn Diagram of message-based interest types for our English dataset (left)
and the dataset in Dooms et al. (right)
is to generate a highly accurate dataset, first, we used an ensemble of methods,
as detailed in Section 5.2, and furthermore, we considered only the subset F0
t
in Ftwith indegree (with respect to our population U) higher than 40. In fact,
we noted that less popular topical friends may still include bloggers or Twitter
Multi-domain Interests Dataset of Twitter users with mappings to Wikipedia 13
Interests induced from topical friends (|U|=444,744 English speaking users)
# of topical friends F0
twith indegree ≥40 in U409,743
# of unique interests with a mapping to a Wikipage 58,789
average #interests per user 82
precision of Wikipedia mapping (tested on a sample of 1,250 items in F0
t) 90%
Table 4. 6-months (April-September 2017) statistics on interests induced from topical
friends of English-speaking users
Interests induced from topical friends (|U|=25,135 Italian speaking users)
# of topical friends F0
twith indegree ≥42 in U29,075
# of unique interests with a mapping to a Wikipage 4,580
average #interests per user 41.96
precision of Wikipedia mapping (tested on a sample of 1,250 items in F0
t) 90%
Table 5. 6-months (April-September 2017) statistics on interests induced from topical
friends of Italian-speaking users
users for which, despite some popularity, a Wikipage does not exist. In these
cases, our methodology may suggest false positives. When applying the indegree
filter, the precision - manually evaluated with adjudication on 1250 accounts
randomly chosen in this restricted population F0
t- is as high as 90%, as shown
in the last line of Tables 4 and 5. The k-Fleiss IAA are 0.95 and 0.92, respectively.
We remark that we are not concerned here with measuring the recall, since the
objective is to release a dataset with high precision and high coverage, in terms
of number of interests per user, over the considered populations. To this end, the
indegree threshold 40 was selected upon repeated experiments to obtain the best
trade-off between the distribution of interests in the population Uand precision
of Wikipedia mapping.
Concerning coverage, when merging the two sources of information, our English
dataset includes an average of 90 interests per user for about 450k users, and
a total of 282,303 + 58,789 = 341,092 unique interests in a large variety of
domains. As a comparison, even when considering single domains, the largest
available datasets9, like MovieLens and Bookcrossing, do not exceed 150,000
users and 250,000 items, with a much lower density in terms of interests per user
-although these resources provide ranked preferences rather than unary, as in
WikiMED. Even the popular Million Songs Dataset Challenge [18] consists of a
larger set of users (1.2 million users) but a comparable number of unique interests
in a single domain (380,000 songs). To the best of our knowledge, this is the
largest freely available multi-domain interest dataset reported in literature, and
furthermore, we provide the unique feature of a reliable mapping to Wikipedia.
14 Giorgia Di Tommaso, Stefano Faralli, Giovanni Stilo and Paola Velardi
Fig. 3. Distribution of interests induced from users’ topical friends: English dataset
(left) and Italian dataset (right).
sioc:UserAccount
sioc:follows
interest
skos:relatedMatch
sioc:likes
resource
skos:relatedMatch
Fig. 4. The data model adopted for the design of our resource.
7 The Wiki-MID resource
Our resource is designed on top of the Semantically-Interlinked Online Commu-
nities (SIOC) core ontology.10 The SIOC ontology favors the inclusion of data
mined from social networks communities into the Linked Open Data (LOD)
cloud. As shown in Figure 4 we represent Twitter users as instances of the
SIOC UserAccount class. Topical users and message based user interests are
then associated, through the usage of the Simple Knowledge Organization Sys-
tem Namespace Document (SKOS)11 predicate relatedMatch, to a corresponding
Wikipedia page as a result of our automated mapping methodology. We release at
http://wikimid.tweets.di.uniroma1.it/wikimid/ both the dataset and the
related software under Creative Commons Attribution-Non Commercial-Share
Alike 4.0 License.
8 Concluding Remarks
In this paper we presented Wiki-MID, a LOD-compliant resource that captures
Twitter users’ interests in multiple domains. With respect to other available
datasets for Recommender Systems, our resource has several unique features:
1) users’ interests are induced from their messages and authoritative (”topical”)
9https://www.kdnuggets.com/2016/02/nine-datasets- investigating-recommender- systems.html
10 http://rdfs.org/sioc/spec/sioc.html
11 http://www.w3.org/2004/02/skos/core.html
Multi-domain Interests Dataset of Twitter users with mappings to Wikipedia 15
friends, and associated with corresponding Wikipedia articles, thus providing a
mean to derive a semantic categorization of interests through the exploitation of
available resources linked to Wikipedia, such as the Wikipedia Category Graph,
DBPedia, BabelNet, and others;
2) for every user, we are hence able to extract in two languages (English and
Italian) a variety of interests in multiple categories, such as art, science, enter-
tainment, politics, sport and more;
3) the dimension of the dataset is comparable with the largest single-domain
interest datasets in literature, and the average number of multi-domain interests
per user is far more large than other multi-domain datasets.
Further note, as shown in Section 6, that extracting interests from messages
and topical friends, and subsequent mapping to Wikipedia, is a very reliable
process (4% error rate for message-induced interests and 10% for friendship-
induced). In addition, the availability of semantic resources linked to Wikipedia
offers the possibility to identify for each user the ”dominant” interest categories,
on which recommenders could rely when suggesting new items. We leave to
future research the exploitation of these features.
Acknowledgments. This work has been supported by the IBM Faculty Award
#2305895190 and by the MIUR under grant ”Dipartimenti di eccellenza 2018-
2022” of the Department of Computer Science of Sapienza University.
References
1. Linden, G., Smith, B., York, J.: Amazon. com recommendations: Item-to-item
collaborative filtering. IEEE Internet computing 7(1) (2003) 76–80
2. Davidson, J., Liebald, B., Liu, J., et al.: The youtube video recommendation
system. In: Proc. of the 4th RecSys, ACM (2010) 293–296
3. Fouss, F., Saerens, M.: Evaluating performance of recommender systems: An ex-
perimental comparison. In: WI-IAT 2008. Int. Conf. on. Volume 1., IEEE 735–738
4. Felfernig, A., Jeran, M., Ninaus, G., Reinfrank, F., Reiterer, S., Stettinger, M.:
Basic approaches in recommendation systems. In: RSSE. Springer (2014) 15–37
5. Schafer, J.B., Frankowski, D., Herlocker, J., Sen, S.: Collaborative filtering recom-
mender systems. In: The adaptive web. Springer (2007) 291–324
6. Pazzani, M.J., Billsus, D.: Content-based recommendation systems. In: The adap-
tive web. Springer (2007) 325–341
7. Trewin, S.: Knowledge-based recommender systems. Encyclopedia of library and
information science 69(Supplement 32) (2000) 180
8. Burke, R.: Hybrid recommender systems: Survey and experiments. User modeling
and user-adapted interaction 12(4) (2002) 331–370
9. Gunawardana, A., Shani, G.: A survey of accuracy evaluation metrics of recom-
mendation tasks. JMLR 10(Dec) (2009) 2935–2962
10. Dror, G., Koenigstein, N., Koren, Y., Weimer, M.: The yahoo! music dataset and
kdd-cup’11. In: Proc. of KDD Cup 2011. (2012) 3–18
11. Shepitsen, A., Gemmell, J., Mobasher, B., Burke, R.: Personalized recommendation
in social tagging systems using hierarchical clustering. In: RecSys, 2008, ACM
16 Giorgia Di Tommaso, Stefano Faralli, Giovanni Stilo and Paola Velardi
12. Kamishima, T., Akaho, S.: Nantonac collaborative filtering: A model-based ap-
proach. In: Proc. of the 4th RecSys, ACM (2010) 273–276
13. Sawant, S., Pai, G.: Yelp food recommendation system (2013)
14. Wang, H., Lu, Y., Zhai, C.: Latent aspect rating analysis on review text data: a
rating regression approach. In: Proc. of the 16th ACM SIGKDD. (2010) 783–792
15. Mavalankar, A.A., et al.: Hotel recommendation system. Internal Report (2017)
16. C¸ ano, E., Morisio, M.: Characterization of public datasets for recommender sys-
tems. In: RTSI, IEEE 1st International Forum on, IEEE (2015) 249–257
17. Harper, F.M., Konstan, J.A.: The movielens datasets: History and context. TiiS16
18. McFee, B., Bertin-Mahieux, T., Ellis, D.P., Lanckriet, G.R.: The million song
dataset challenge. In: Proc. of the 21st WWW, ACM (2012) 909–916
19. Bennett, J., Lanning, S., et al.: The netflix prize. In: Proc. of KDD, NY (2007)
20. Yan, M., Sang, J., Xu, C.: Mining cross-network association for youtube video
promotion. In: Proc. of the 22nd ACMMM, ACM (2014) 557–566
21. Piao, G., Breslin, J.G.: Inferring user interests in microblogging social networks:
A survey. In: arXiv:1712.07691v3. (2017)
22. Chaabane, A., Acs, G., Kaafar, M.A., et al.: You are what you like! information
leakage through users? interests. In: Proc. of the 19th NDSS Symposium. (2012)
23. Faralli, S., Stilo, G., Velardi, P.: Large scale homophily analysis in twitter using a
twixonomy. In: Proc. of 24th IJCAI, 2015, Buenos Aires,Jul. 25-31, 2015. (2015)
2334–2340
24. Piao, G., Breslin, J.G.: Inferring user interests for passive users on twitter by
leveraging followee biographies. In: Proc. of ECIR. (2017)
25. Pichl, M., Zangerle, E., Specht, G.: # nowplaying on# spotify: Leveraging spotify
information on twitter for artist recommendations. In: ICWE. (2015) 163–174
26. P., K., P., J., C., V., A., S.: User interests identification on twitter using a hierar-
chical knowledge base. LNCS: The Semantic Web: Trends and Challenges (2014)
27. Schinas, E., Papadopoulos, S., Diplaris, S., Kompatsiaris, Y., Mass, Y., Herzig, J.,
Boudakidis, L.: Eventsense: Capturing the pulse of large-scale events by mining
social media streams. In: Proc. of the 17th PCI, ACM (2013) 17–24
28. Nichols, J., Mahmud, J., Drews, C.: Summarizing sporting events using twitter.
In: Proc. of the 2012 Int. Conf. on Intelligent User Interfaces, ACM (2012) 189–198
29. Dooms, S., De Pessemier, T., Martens, L.: Mining cross-domain rating datasets
from structured data on twitter. In: Proc. of the 23rd WWW, ACM (2014) 621–624
30. Barbieri, N., Bonchi, F., Manco, G.: Who to follow and why: link prediction with
explanations. In: Proc. of the 20th ACM SIGKDD, ACM (2014) 1266–1275
31. Myers, S.A., Leskovec, J.: The bursty dynamics of the twitter information network.
In: Proc. of the 23rd WWW, ACM (2014) 913–924
32. Pichl, M., Zangerle, E., Specht, G.: Combining spotify and twitter data for gener-
ating a recent and public dataset for music recommendation. In: Grundlagen von
Datenbanken. (2014) 35–40
33. Besel, C., Schl¨otterer, J., Granitzer, M.: Inferring semantic interest profiles from
twitter followees: Does twitter know better than your friends? SAC ’16 (2016)
34. Nechaev, Y., Corcoglioniti, F., Giuliano, C.: Sociallink: Linking dbpedia entities
to corresponding twitter accounts. In: The Semantic Web – ISWC 2017. (2017)
35. Faralli, S., Stilo, G., Velardi, P.: Automatic acquisition of a taxonomy of microblogs
users’ interests. Journal of Web Semantics (2017)
36. Navigli, R., Ponzetto, S.P.: BabelNet: The automatic construction, evaluation and
application of a wide-coverage multilingual semantic network. AI (2012) 217–250
37. Delli Bovi, L., Telesca, L., Navigli, R.: Large-scale information extraction from
textual definitions through deep syntactic and semantic analysis. TACL 3(2015)
