Social networks provide rich information about user interests and activities representing a valuable source for search personalization. However, social information is typically large and dynamic making its exploitation to obtain relevant search results a very challenging task. This work presents a PhD project plan that investigates Social Information Retrieval. The goal is threefolds: (1) create confidence area for information search by community detection based on tags similarity (2) introduce a new notion of Social Document Profile based on user activities, and (3) propose a novel ranking model based on social relevance.
1.1
Social networks are becoming one of the predominant sources of information.
Users of such networks publish documents that can take different forms,
including text, image, audio, and video. Additionally, they can perform different
types of actions around published documents. These actions can be classified as
descriptive or reactive. Descriptive actions, mainly tagging, reflect the content
of documents, while reactive actions such as like, dislike, rate, favorite, share,
and comment reflect users’ feedbacks regarding documents. This rich repository
of users’ actions triggered many research works to exploit social information
for search personalization [
We propose to provide tailored answers to users’ needs by exploiting social information in two different stages. First, we use descriptive actions to create, for each user, a confidence search area according to his profile. Second, we use both descriptive and reactive actions to define a social profile, per confidence area, for each document. The novel contribution by this paper has the following salient properties: 1. We model a social information retrieval framework as an undirected graph of social entities (User, Document, Tags and Clicks) where links represent entities relations generated in a social context, Tags represent descriptive actions, and Clicks represent reactive actions. 2. We exploit user profile as a tool for community detection based on Tags similarity. The goal is to establish a confidence search area for each user. 3. We propose a novel Social Document Profile based on a tripartite graph (Content, Tags, Clicks) that represents documents not only using their content but also their social profile given by Tags and Clicks. 4. We propose a novel scoring model that combines content relevance based on user profile and social relevance based on social document profile. Our proposed approach goes beyond existing IR personalization techniques in several ways. First, it combines two areas: community detection in social networks and information retrieval. Second, unlike existing approaches, we define personalization approach based not only on user profile but also on document social profile. Third, none of the existing approaches takes into account clicks as social information defining document profile. 2
Search personalization using social information has been investigated extensively.
The first class of approaches limits social information to annotations or tags
[
The second class of approaches exploits, in addition to tags, social
relationships between users [
All approaches described above focus on how to generate user profile using
social information but none of them takes into account social document profile.
In our work, we exploit user profile not at query time but to detect interest
communities as confidence search areas. Moreover, we build a social document
profile based on clicks which was not considered in related work. A work that
went beyond using only tags and user relationships is by Wang et. al., [
Another research area related to our work is community detection where
various methods have been proposed [
We define the Social Graph SG as a tuple SG = fU, D, T, C, A1, A2g where U = fu1, . . . , ukg, D = fd1, . . . , dlg, T = ft1, . . . , tmg and C = fc1, . . . , ceg are respectively the set of Users, Documents, Tags and Clicks. A1 = {ui, dj , tf } 2 U D T is a set of annotations reflecting each user ui tagging document dj with tag tf and A2 = {ui, dj , cr} 2 U D C is a set of clicks reflecting each user ui reacting to document dj using click cr (see Figure 1). Our personalized search strategy consists in the following steps. First, we extract users’ communities from social networks based on users’ profiles. The profile of a user is defined by the set of tags he used to annotate documents. Thus, the community detection problem is reduced to computing tags similarity by using the subgraph G = (U, T) of the social graph SG. Second, upon receiving a search query Q = fq1; :::; qng from a user u, we proceed as follows: 1. We retrieve the topk relevant results to the query. Each result is associated with a content relevance score; the more relevant and important a result is with respect to the query, the higher its relevance score. 2. For each of the topk results, we compute its social score based on how popular it is in user u’s community. This popularity is defined by related clicks (share, favourite, comment, etc). 3. The results are then re-ranked based on the combination of the content relevance score and the social relevance score of each result. 3.2
Social User Profile Our proposed model for social information retrieval is
based on a central phase of community detection. Our aim is to detect community
of interest to personalize IR processes. We propose to use the subgraph G = (U,
T) of the social graph SG to detect similar users based on the tags they use. Note
that, we take into account the time factor s since users’ interest change over time.
Therefore, the social user profile Pi of user ui is defined by Pi = ft1; : : : ; tmgs.
To detect community between users it is then to compute tags similarity.
Community Detection We propose to adopt Katz coefficient for community
detection. Katz coefficient is a similarity index proposed in the field of social
science and was recently rediscovered in the context of collaborative
recommendation and Kernel methods where they are known as Von Neuman Kernel.
Katz proposed a method of calculating similarity taking into account not only
the number of direct links between elements, but also the number of indirect
links [
N Katz := X l=1 lpathsli;j where l is the length of the path and l is the appropriate weight to path l. 3.3
Each document has a social profile defined by annotations (Tags) and Clicks in addition to its content. Therefore, a document D is defined by the threefold fCt, T, Cg where Ct, T and C respectively correspond to Content, Tag and Click. Therefore, a document is evaluated through two measures: content relevance and social relevance.
Content relevance. To compute the relevance of a document dx to user query, we use BM25 (or Okapi) scoring function given by :
BM 25(dx; qi) = IDF (qi):
f (qi; dx):(k1 + 1) f (qi; dx) + k1:(1 b + b: ajvdgxdjl ) where f(qi, dx) is the count of term qi in document dx, |dx| is the length of document dx, avgdl is the average document length in the collection of documents, k1 = 1.2 and b = 0.75, IDF(qi) is the inverse document frequency weight of the query term qi which is computed as : where N is the total number of documents in the collection, and n(qi) is the number of documents containing n(qi). Thus, the content relevance score of a document x is given by:
IDF (qi) = log
N
n(qi) + 0:5 n(qi) + 0:5 n Rel(dx; Q) = X BM 25(dx; qi)
i=1 Social relevance. To compute social relevance, we use the tripartite graph (User, Document, Click) from the Social Graph SG. We consider the Clicks C = fc1, . . . , ceg to estimate the social popularity of a document in a given community. For the same query by two different users returned results are ordered differently depending on the social context of each user. Our idea for the social relevance computation is to to find a social score for clicks which is the weighted sum of clicks weighted scores. We consider the following click score of document dx clicked by click ci in the community of user u: cs(dx; ci; u) = count(ci; dx; u) count(dx; u) where: count(ci; dx; u) is the number of users, in the community of user u, who used click ci for document dx, and count(dx; u) is the total number of users, in the community of user u, who clicked on document dx. By combining the click scores, we obtain the social score of document dx in the community of user u given by:
e SS(dx; u) = X i=1 ics(dx; ci; u) where e is the number of clicks types (For example, in Facebook we have e=3 because we have 3 clicks types : like, share and comment) and Pin=1 i = 1 where i is a weighted coefficient selected by the query initiator. 3.4
We use a linear combination of the content score Rel(dx; Q) and the social score ss(dx; u) to obtain the final score of a document dx returned as a result for query Q initiated by user u:
S(dx; u) = Rel(dx; Q) + (1 )SS(dx; u) where 0
As a short term objective, we plan to implement our personalized search approach and perform experiments on real-world data to evaluate its performance focusing on the following tasks: . 1. Compare our click-based personalization with tag-based personalization 2. Study closely the impact of the social document model on search results 3. Analyze how our technique performs depending on the level of activities in different communities. 4.1
We will test our personalized search approach using data crawled from YouTube 2 which has the main characteristics needed for our solution. This dataset have been crawled during the period between October, 15th, 2012 and December, 25th, 2012. It contains 890682 videos, 282074 users and 1014190 information about social clicks (comment, favourite and rated). 2 www.youtube.com
Acknowledgements. This research was supported by the RARE project at KRDB research centre for knowledge and data at Free University of BozenBolzano.