=Paper=
{{Paper
|id=None
|storemode=property
|title=Social Sifter: An Agent-Based Recommender System to Mine the Social Web
|pdfUrl=https://ceur-ws.org/Vol-966/STIDS2012_P02_NachawatiEtAl_SocialSifter.pdf
|volume=Vol-966
|dblpUrl=https://dblp.org/rec/conf/stids/NachawatiRYKB12
}}
==Social Sifter: An Agent-Based Recommender System to Mine the Social Web==
https://ceur-ws.org/Vol-966/STIDS2012_P02_NachawatiEtAl_SocialSifter.pdf
Social Sifter: An Agent-Based Recommender System
to Mine the Social Web
M. Omar Nachawati, Rasheed Rabbi, Genong (Eugene) Yu, Larry Kerschberg and Alexander Brodsky
Dept. of Computer Science
George Mason University
Fairfax, VA, USA
{mnachawa, rrabbi, gyu, kersch, brodsky} at gmu.edu
Abstract— With the recent growth of the Social Web, an information that is made available in structured, machine
emerging challenge is how we can integrate information from readable formats, such as RDF and OWL .
the heterogeneity of current Social Web sites to improve
semantic access to the information and knowledge across the Conventionally, finding answers to questions and learning
from the knowledge mine existed on the Social Web has
entire World Wide Web, the Web. Interoperability across the
Social Web sites make the simplest of inferences based on data primarily been a manual process. It requires a lot of
intelligence in sifting through the mountains of Social Web
from different sites challenging. Even if such data were
interoperable across multiple Social Web sites, the ability of pages using only a keyword-based Web search engine, which is
akin to a primitive pitch-fork in Semantic Web terms. More
meaningful inferences of a collective intelligence [1] system
depends on both its ability to marshal such semantic data, as recently, however, Social Web sites have begun to embrace
Semantic Web technologies such as RDF and OWL, and have
well as its ability to accurately understand and precisely
respond to queries from its users. This paper presents the been offering much more machine-friendly data, such as geo-
tagged images on Flickr, Friend Of A Friend (FOAF) exports
architecture for Social Sifter, an agent-based, collective
intelligence system for assimilating information and knowledge in FaceBook and hCalendar [7] tagged events on Blogger. Such
developments have sparked the evolution of the Social Web
across the Social Web. A health recommender system
prototype was developed using the Social Sifter architecture, into a collective knowledge system [1], where the contributions
of the user community are aggregated and marshaled with
which recommends treatments, prevention advice, therapies for
ailments, and doctors and hospitals based on shared knowledge from other heterogeneous sources (e.g., web pages,
news and encyclopedia articles, and academic journals) in a
experiences available on the Social Web.
synergy dubbed the Social Semantic Web.
Keywords: social semantic search; collective knowledge
systems; recommender systems, OWL; RDF; SPARQL While the Semantic Web focuses on data to enable
interoperability among heterogeneous semi-structured web
I. INTRODUCTION pages, the focus of the Social Semantic Web vision is to create
a system of collective intelligence by improving the way
Since its inception, the World Wide Web has always
people share and explore their own and others knowledge and
overwhelmed users with its vast quantity of information. The
experience [1]. Work on the Social Sifter promotes that grand
advent of Social Webs, coined Web 2.0, has placed an
vision and expands on the research done on the patented
additional burden on Web search engines. While the
Knowledge Sifter architecture [7, 8, 9], as well as the Personal
established algorithms that Web search engines employ are
Health Explorer [11], undertaken at George Mason University.
effective in surfacing the most popular results through
As a proof of concept, we have designed a social health
hyperlink analysis, as demonstrated by the Hubs and
knowledge and recommender system based on the Social Sifter
Authorities algorithm [2] and the PageRank algorithm [3],
platform that utilizes the Social Semantic Web to provide
those results are not necessarily relevant despite popularity and
precise search results and recommendations.
these algorithms have fallen short of solving the problem of
information overload [1, 2, 3] on the World Wide Web. The rest of this paper is organized as follows: section II
discusses related work, section III describes the Social Sifter
The research into natural language understanding [4]
architecture and a brief description of the prototype system.
attempts to close that gap. However the quality of machine
Section IV highlights the experimental results, and Section V
generated semantics still pales in comparison to that of humans.
identifies the possible future work on the Social Sifter platform.
This became a core challenge for the Semantic Web or Web
3.0, where information is made available in structured, II. RELATED WORK
machine-friendly formats allowing machines not only to sort
and filter such data, but also to combine data from multiple A. Knowledge Sifter and Personal Health Explorer
Web sites in a meaningful way and allow inferences to be made Semantic systems belong to a class of systems that make use
upon that data. While semantic query languages, such as of ontologies, context awareness and other semantic methods to
SPARQL, can provide a database-like interface to the World make informed recommendations. Such research in semantic
Wide Web, it is only as good as the quantity and quality of search at George Mason University began with WebSifter [8, 9,
1
10], an agent-based multi-criteria ranking system to select of [11] to leverage an integrated semantic search engine and
semantically meaningful Web pages from multiple search recommender system.
engines such as Google, Yahoo, etc. The work further led to a
patent [8]. Knowledge Sifter (KS) [8] is motivated by III. THE SOCIAL SIFTER ARCHITECTURE
WebSifter [7,8], but is augmented with the advanced use of Social Sifter, an enhancement of the existing Knowledge
semantic web ontologies, authoritative sources, and a service- Sifter (KS), is a collection of cooperating agents that are
oriented plug-and-play architecture. Knowledge Sifter is a exposed through web services and exhibits a Service-Oriented
scalable agent-based web services framework that is aimed to Architecture (SOA)-based framework.
support i) ontology guided semantic searches, ii) refine
searches based on relevant feedback, and iii) accessing
heterogeneous data sources via agent-based knowledge
services. Personal Health Explorer (PHE) is an enhancement of
KS to perform semantic search in biomedical domain. PHE
leverages additional features of a personal health graph to be
identified, categorized, and reconstituted by providing links to
the user to rate individual results and return to previous queries
and update information through a semantically supported path.
KS and PHE are able to obtain more relevant search results
than classic search engines; while the result is very general, it
leaves room to make it more personalized. Both KS and PHE
make multifaceted efforts towards realizing the Semantic Web
vision, primarily focusing on the formal ontological sources.
PHE provides facilities to include a user’s Personal Health
Record (PHR), which entails additional permission and access
control which may be constrained by HIPAA regulations.
Interestingly, both of these systems did not use the data
available on the Social Web, namely Wikipedia, YouTube,
Flickr, Facebook, LinkedIn, etc. This is where Social Sifter Figure 1. Social Sifter Architecture – Tiers and Components
makes its contribution.
Depending on the functionality, agents are allocated into
B. BLISS and Cobot three different architecture layers – i) the User Layer, ii) the
Other attempts to utilize Web 2.0 technology to enhance the Knowledge Management Layer, and iii) Data Layer. The User
quality and relevance of health recommendation systems Layer consists of the User and Preferences agents, and
include bookmarking, crowd sourcing, crowd tagging and manages all user interaction and data preferences. The
harvesting user recommendations. The Biological Literature Knowledge Management Layer handles the support for
Social Ranking System (BLISS) is one such prototype system semantic search, access to data sources, and the ranking of
that allows users to bookmark and promote their search results using technologies like the Ontology, Social Web
recommendation to communities of special interest, facilitate Crawling, Ranking, Query Formulation, and Web Services
the annotation and ranking by the community, and present the agents. The Data Layer consists of the data repositories that
results to allow other users to get the recommendations based provide authoritative information and documents. The
on community ranking [6]. The bookmarking approach is hierarchy of the architecture layers is already defined in KS;
useful in establishing the authoritativeness of information over three additional agents were added, with an alteration of the
the long term because it uses social voting or ranking [5]. underlying algorithm to perform the execution flow into the
Social Sifter.
The Cobot system uses social conversation and social
tagging (preference) to enhance the health recommendations. Social Web agent basically collaborates with following two
Three techniques are noteworthy: (1) user-initiative dialogue in agents to manipulate social web information.
capturing user’s intent, (2) social tagging in establishing the Open SW agent performs open search within the blogs,
authoritativeness of social information, and (3) case-based related support groups etc.
semantic reasoning in utilizing social knowledge for
recommendation [5]. User Specific SW agent identifies user social identities
across the web and conducts Collaborative Filtering by
C. Semantic Analytics on Social Networks processing social tags, user participation and responses
A multi-step engineering process is described in [9] to utilize available on the social webs.
social knowledge. These steps are common procedure to across
IV. HEALTH RECOMMENDER SYSTEM
the initiatives to transform the social web information to
semantic knowledge. As a proof-of-concept, we are building a health
recommender system using our Social Sifter architecture that
Social Sifter adheres to the underlying framework of provides health recommendations for any type of sickness,
Knowledge Sifter [9], the knowledge manipulation mechanism disease or disorder. The present system does not do any natural
of PHE [10], and engineering process for semantic association language processing on user queries, and therefore is limited as
2
to what it can accept as a valid query. Currently, the system friendship and affiliation information to generate the
accepts a comma delimited list of words that relate to a specific user’s Social Graph.
ailment and returns a list of relevant descriptions of the
iii) User Agent passes the SPARQL query and the collected
ailment, therapy options, doctors, and treatment centers as
collected from the Social Semantic Web from our knowledge User Profile information to the Query Formulation Agent.
Management Layer. We intend for future versions of the health Query Refinement: The Query Formulation Agent then
recommender system to allow for unrestricted language queries attempts to enrich the original SPARQL query by:
by performing natural language processing to transform the
unstructured query input into a more structured format, i) Semantic Query Decomposition: It will generate multiple
acceptable by the Social Sifter architecture. sub-queries that generalize and specialize the term
pancreatic cancer based on the health-domain ontology
from The National Center for Biomedical Ontologies
(NCBO), a BioPortal and MedLine (Medical Literature
Analysis and Retrieval System Online), which is a
Parsing Key
Words
bibliographic database of life sciences and biomedical
information.
Query Enrichment with Semantics ii) Marshalling: selected data will be marshaled with the
Key Words
amassed folksonomy from the Social Web Agent. The
inference engine will also generate queries based on the
results of any cluster analysis from data crawled from the
Ontology RDF Social Media Social Web, which may pick up, for instance, other
ailments that people have discussed together with
pancreatic cancer.
Decomposing into Multiple Sub Queries
iii) Ranking: The end result of this meta-search is a weighted
tree of sub-queries, where weights are assigned based,
among other features, on the static nature of the sub-
Perform Search
query generated (heuristically) as well as the importance
with Existing
Search Engine
of the source (back-reference analysis).
Post Query Processing: Once all sub-queries have been
Analyze, Rank and
organize search result defined, the Web Service Agent passes them to the Data Layer,
which accordingly runs the queries and itself ranks each result,
based on many factors, including relevance (ontological),
Display
importance (back-reference based) and belief (Bayesian-based
inference from Social Semantic Web).
Figure 2. Social Sifter work flow diagram
Result Scrutinizing: The results are then returned to the
A. Scenario for Pancreatic Cancer Integration Agent, which combines different classes (based on
Consider the case when a user is exploring recommendations the results from the classifier) of results based on a total
for pancreatic cancer. According to the NIH, treatment options ordering derived from the aggregated ontology, and back-
include surgery and biliary stents. The NIH also lists links to reference analysis. The agent also performs a clustering
support groups, among which CancerCare.org features a social analysis on the result set to further group the results and
question-answer forum that is categorized by topic. Our perform statistical calculations on the groups of results before
inference agent for health recommendations takes advantage of passing them to the User Layer.
this domain knowledge in attempting to provide better quality Result displaying: The User Layer then displays the grouped
recommendations than what would be available from a general and ranked results according to the preferences selected by the
Web search engine. Let us walk through the steps of the health user.
recommender system for this particular query.
B. Query life cycle for Pancreatic Cancer in Social sifter
Query Submission: User logs into the health recommender
system website and enters the following query terms The life cycle of a query in Social Sifter, e.g., searching for
“pancreatic cancer.” “pancreatic cancer”, is as follows: (1) a user allows access to
his profile, (2) Sifter culls information from his social
Query String Preparation: networks, (3) Sifter initiates targeted information harvesting,
(4) Sifter conducts semantic inference and reasoning, and (5)
i) The User Agent parses the query string to identify key
Sifter presents socially- and semantically-renked results are to
words.
the user.
ii) The Preference Agent collects context information,
C. Social Sifter Prototype
including the user’s IP address, a query session identifier,
and the best geographic location estimate available for The Social Sifter prototype has been implemented to use
that user. It tries to create a User Profile by indexing information retrievable from Facebook using Graph API in
3
gathering the information about the users. In Facebook, each and Social Sifter. Social Sifter provided integrated results and
user can have feeds, likes, activities, interests, music, books, used social ranking to rearrange the categories depending on
videos, events, groups, checkins, games, and his personal users profile information. Location is determined based on user
information, like hometown and related locations. These provided current living locations. More testing is being carried
provide a very rich base for understanding the intension of a out to determine metrics to assess the quality of social semantic
user when he is searching on the Web. search recommendations.
Social Sifter combined both semantic reasoning and social VI. CONCLUSIONS
ranking to better understand user’s intention and present the
results to users, based on initial search keywords or phrases Social semantic search is an integration of social networks
provided. The algorithm for the currently implemented search and semantic search. Semantic search provides rich means in
is described as follows. enhancing search, especially the user’s intent and semantic
reasoning. Social search involves people and links to their
(1) Login: User logs into his Facebook using OAuth social graphs. In this paper, a prototype social semantic search
authentication. The program gets the authorized token and engine, Social Sifter, has been presented. The lessons learned
uses it to access user’s information with user’s from the implementation showed two areas for improving
concurrence. search accuracy: social contextual information (user intent
understanding) and social semantic ranking (results relevance).
(2) Information Retrieval: The system retrieves the
information about the user (Feeds, Likes, Activities, The current implemented prototype system is limited in the
Interests, Music, Books, Photos, Videos etc.) and uses use of the semantic reasoning. The crawling of data should be
them in supporting the targeted harvesting of information expanded to other social media and social networks. Integration
and formulating the social ranking of results in categories. of these results into a standard semantic data store is necessary
to realize the power of semantic reasoning. Further study
(3) Social ranking – A simple algorithm is used to calculate
directions are: (1) to integrate mature ontologies, (2) to define
the social weights of the harvested information in each
customized actions to demonstrate the approach in health
category. The algorithm is basically counting the
domain, and (3) to use the reasoning power of semantics.
occurrences of keywords or phrases in each category.
(4) Social context – The user’s background information is REFERENCES
used in refining the search results or filtering the results. [1] T. Gruber, “Collective knowledge systems: Where the Social Web meets
One specific example is the location information. The the Semantic Web,” Web Semantics: Science, Services and Agents on the
World Wide Web, vol. 6, no. 1, pp. 4–13, Feb. 2008.
home location of the person is generally used to limit the
places to be searched and returned. [2] J. M. Kleinberg, “Authoritative sources in a hyperlinked environment,” J.
ACM, vol. 46, no. 5, pp. 604–632, Sep. 1999.
(5) Semantic result presentation – The results are presented to
[3] L. Page, S. Brin, R. Motwani, and T. Winograd, The PageRank Citation
users in groups: people, groups, events, places, events, Ranking: Bringing Order to the Web. 1999.
pages, or posts. The current implementation is limited to
use the categories or semantics of Facebook. The actions [4] A. Ntoulas, G. Chao, and J. Cho, “The infocious web search engine:
improving web searching through linguistic analysis,” in Special interest tracks
in Facebook link objects and people. They are the bases and posters of the 14th international conference on World Wide Web, New
for our search engine in weighing the harvesting strategies. York, NY, USA, 2005, pp. 840–849.
They are also important in ranking the results and the [5] J. M. Gomez, G. Alor-Hernandez, R. Posada-Gomez, M. A. Abud-
categories when presenting the search results to users. The Figueroa, and A. Garcia-Crespo, “SITIO: A Social Semantic Recommendation
current implementation used the same social ranking Platform,” in 17th International Conference on Electronics, Communications
strategy described in (3). and Computers, 2007. CONIELECOMP ’07, 2007, p. 29–29
[6] “hCalendar 1.0 · Microformats Wiki.” [Online]. Available:
D. Proactive Social Search http://microformats.org/wiki/hcalendar. [Accessed: 15-Apr-2012].
The existing Facebook semantics do not capture the [7] L. Kerschberg, W. Kim, and A. Scime, “WebSifter II: A Personalizable
semantic of health queries. For health problems, users may be Meta-Search Agent Based on Weighted Semantic Taxonomy Tree,” in
interested in finding out the cure of certain diseases, which is International Conference on Internet Computing, Las Vegas, NV, 2001
not captured by the current set of actions available in [8] L. Kerschberg, W. Kim, and A. Scime, “Personalizable semantic taxonomy-
Facebook. Customized actions can be implemented using the based search agent,” U.S. Patent 7117207Oct-2006
Facebook Open Graph, but it is beyond the scope of this paper.
[9] L. Kerschberg, H. Jeong, Y. Song, and W. Kim, “A Case-Based Framework
for Collaborative Semantic Search in Knowledge Sifter,” Case-Based
V. EXPERMENTAL FINDINGS Reasoning Research and Development, vol. 4626/2007, pp. 16–30, 2007.
The Social Sifter prototype has been implemented. The [10] T. G. Morrell and L. Kerschberg, “Personal Health Explorer: A Semantic
Facebook Graph API was used as the basis for harvesting Health Recommendation System," workshop on Data-Driven Decision Support
social network information about the user. Social information and Guidance System (DGSS), 28th IEEE International Conference on Data
was used in two aspects – understanding the user’s intention Engineering, Arlington, VA April 1, 2012.
(context) and ranking results (social semantic ranking). The [11] Boanerges Aleman-Meza, Meenakshi Nagarajan, Cartic Ramakrishnan,
two aspects showed improved search results. For example, the Li Ding, Pranam Kolari, Amit P. Sheth, I. Budak Arpinar, Anupam Joshi, Tim
searching case using phrase – “pancreatic cancer” can be Finin. Semantic Analytics on Social Networks: Experiences in Addressing the
compared using three different engines – Google, Facebook, Problem of Conflict of Interest Detection. WWW 2006, May 23–26, 2006,
Edinburgh, Scotland. ACM 1-59593-323-9/06/0005.
4