Supporting the user in nding relevant information is one of the major tasks in Information Retrieval. The user-centered approach to IR is based on a user pro le that represents the user's interests and preferences, to obtain relevant information beyond the formulation of a query. Nowadays, the content di used by people in the Social Web represents a source of primary information to build a user pro le. For this reason, the aim of this paper is to propose an event-based approach to user pro ling in social media, where an event is identi ed as a namedentity. Furthermore, by the proposed approach, it is possible to consider the user's interests evolution over time.
The main objective of Information Retrieval is to support users in the process
of nding information relevant to their needs and interests, which are usually
formulated by means of a query. Originally, Information Retrieval Systems (IRS)s
relied on the so called system-centered approach, where the IRS produces the
same results to the same query, independently from the user context (the \one
size ts all" paradigm). In addition to this, by employing this approach, for a
particular user it could not be easy to formulate a need or a complex request using
only few words. For these reasons, the user-centered approach to Information
Retrieval is by now the most frequently applied, since it is aimed at identifying
the information relevant to the user beyond the classic formulation of a query,
by considering the user's context [
This paper presents a preliminary approach for the de nition of dynamic user
pro les that exploit the content generated and di used by users through social
media, the so called User-Generated Content (UGC). In fact, UGC can be
considered as a sort of repository from which the user's interests and preferences can
be acquired, and which also allows to capture their dynamic nature [
In the literature, several models to represent user pro les have been proposed,
such as bag-of-words [
Another important issue that concerns the de nition of a user pro le is its
updating [
The proposed approach focuses on the idea that a topical user pro le can be
centered on the concept of event. An event has been de ned in the literature as
\a speci c thing that happens at a speci c time and place along with all necessary
preconditions and unavoidable consequences" [
The term named-entity was coined by Grishman and Sundheim in [
Therefore, the proposed approach can be described according to the following three phases: (i) the UGC acquisition phase, (ii) the named-entity extraction and user pro le formal representation phase, and (iii) the user pro le updating phase. 3.1
In this phase, the user's social media posts gathered from both Facebook (when available) and Twitter are collected and grouped according to a time-window, the granularity of which can be variable, i.e., one week, 15 days, etc.
In general, each post can contain both text and other multimedia data, but in this paper only textual information is considered. In Facebook, a post can represent each user's status update without particular limits, while in Twitter a post is known as tweet, and there are limits in the tweet length; originally, the maximum length was xed to 140 characters, although this has been made more exible over time. Currently, the length of a tweet cannot exceed 280 characters, and each tweet can be equipped with a picture and a location.
As previously illustrated, in this work a user pro le is composed of
namedentities extracted from the user's posts in a speci c time-window. Formally, let
us denote by Dui the feed generated by a user u in a time-window ti. From Dui, a
collection Ei of named-entities is extracted. The extraction can be performed by
exploiting a Named-Entity Recognition (NER) tool; in the proposed approach,
DBpedia Spotlight [
P Ti(u) = fhe; !eiije 2 Eig: Given a feed Dui, the weight !ei is computed as follows: !ei =
Nei ; jDuij where Nei is the number of posts in Dui in which the entity e is cited, and jDuij represents the total number of posts in Dui.
To provide a topical representation of an entity e, a bag-of-words BoWei is associated with e, which contains the words extracted from the posts of the feed Dui in which e is cited. Formally:
e ! BoWei = fhw1; noccw1 i; hw2; noccw2 i; : : : ; hwj ; noccwj i; : : :g; where the symbol ! expresses the association between e and the bag-of-words, and noccwj represents the number of occurrences of the word wj in the posts referring to e. 3.3
The Temporal User Pro le de ned in the previous phase allows to capture the user's interests based on the content s/he published in a given time period. In order to be able to capture the dynamicity of the user's interests over time, a General User Pro le P (u) is de ned in this phase. Speci cally, at each time window ti, an updated instance Pi(u) of the General User Pro le is created, which takes into account both the current interests of the user and the previous ones. The rst instance, i.e., P1(u), coincides with the Temporal User Pro le P T1(u) associated with the rst time-window t1, the second instance, i.e., P2(u), considers both the Temporal User Pro le P T2(u) associated with time-window t2 and P1(u) related to t1, and so on, as illustrated in Figure 1. From a formal point of view:
Pi(u) = fhe; eiije 2 Eig; where ei is the interest value of e at the time-window ti, and Ei is the set of all entities extracted from the user's feed until time-window ti (included). The value of ei in Pi(u) and the bag-of-words associated with e are obtained as follows. PT1(u)
PT2(u) P1(u) t1
P2(u) t2 … … …
PTi(u) Pi(u) ti … … …
PTj(u) Pj(u) tj
P(u)
Computing the interest weights At each time-window ti, the value of ei can be computed according to the following simple formula, which have been de ned to preliminarily evaluate the approach: ei = ( !ei e(i 1)+!ei 2 i = 1 or e(i 1) = 0 otherwise (1) where e(i 1) is the interest value of the entity e at time t(i 1).
This means that, if the user had no previous interest in the entity e, the interest value ei in Pi(u) assumes the value of !ei. Otherwise, the value of ei is the arithmetic mean between the values e(i 1) and !ei.
Updating the bag-of-words The bag-of-words related to a given entity has to be updated at each time-window, in order to have, within each instance of the General User Pro le, an up-to-date topical representation for the entity. Speci cally, for time-window ti, the bag-of-words associated with an entity e in Pi(u) is constituted by the union of the terms related to e from the bag-ofwords in the Temporal User Pro le P Ti(u) and those from the bag-of-words in P(i 1)(u). If a term appears in the bags-of-words of both P Ti(u) and P(i 1)(u), the term frequency associated with the term in Pi(u) can be computed as the arithmetic mean of the term frequency in P Ti(u) and the term frequency in P(i 1)(u); otherwise, if a term does not appear in P Ti(u), the frequency of the term in Pi(u) can be obtained by applying a decay function to the frequency of the term in P(i 1)(u); a simple solution could be represented by the function f (tf ) = 1 tf , where is a decay parameter, and tf is the term frequency in P(i 1)(u). In this way, within the updated bag-of-words representing e, the terms that have the higher frequency are the ones that better describe the entity, and the terms that are no more employed in recent posts will assume less and less importance over time.
The prototype implementing the approach described in the previous section has been developed in Java 8. In order to crawl both Facebook and Twitter posts, the Twitter4J,1 version 4.04, and Facebook4J,2 version 2.4.8, libraries have been employed. Each feed has been indexed by exploiting the Apache Lucene library,3 version 6.4.0. DBpedia Spotlight4 has been used for the automatic extraction of named-entities from unstructured text.
In order to preliminarily evaluate if the proposed approach is able to track the evolution of the user's interests over time, a simple analysis on two users's interests has been conducted. Table 1 illustrates the real users that have been considered, i.e., Barack Obama and Coldplay; their feeds are publicly accessible and their interests easily analyzable. Data refer to the period ranging from March 2013 to January 2017. For the two users, a couple of entities (interests) that is worth mentioning are commented in the following.
In Figures 2 and 3, the bar chart illustrates the value of the weight !ei associated with the entity in each Temporal User Pro le at each time-window ti, while the line chart illustrates the evolution of the interest value ei of the entity in the instances of the General User Pro le.
In relation to the user Barack Obama, two entities are analyzed in Figure 2; it emerges that the entity Patient Protection and A ordable Care Act represents a long-term interest, since its evolution graph in the left line chart shows a growing curve, while the entity United States Senate only appears during the last part of the tracking period, as illustrated in the right line chart. Since it never occurred before, it is plausible to state that it represents a short-term interest, related to a particular phase of his political career.
Figure 3 illustrates two entities that might represent two short-term interests referred to Coldplay. The interesting observations that emerge from this gure are that the entity iTunes occurs especially during the rst part of the overall period of social activity analyzed; instead, the evolution trend of the entity Spotify is almost the opposite of the iTunes one. Considering that Coldplay 1 http://twitter4j.org/ 2 http://facebook4j.org/ 3 http://lucene.apache.org/core/ 4 http://www.dbpedia-spotlight.org/ use their Twitter account mainly for advertisement purposes, it is possible to suppose that for iTunes an interest drift occurred because Coldplay decided, at a certain point, to focus more on Spotify instead of iTunes as a music provider. In this paper, a preliminary approach for modeling a dynamic user pro le in the Social Web by exploiting the User-Generated Content (UGC) spreading through social media has been proposed. The approach is based on an event-based strategy, where the user's interests are represented as named-entities extracted from Facebook and Twitter feeds related to a given period of time. The proposed approach allows to consider the dynamic nature of the user's interests, and to update the user pro le accordingly; it could be useful for example to search or recommend social media contents, in a scenario where opinions and interests change very quickly. In this paper, only preliminary results connected to the evolution of the user's interests over time have been presented. In the future, the model will be re ned by considering for example the normalization of interest weights across entities, and extended evaluations will be performed.